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>
44 #ifdef CONFIG_OPTIMIZE_KSM
45 #include <asm/checksum.h>
50 #define DO_NUMA(x) do { (x); } while (0)
53 #define DO_NUMA(x) do { } while (0)
57 * A few notes about the KSM scanning process,
58 * to make it easier to understand the data structures below:
60 * In order to reduce excessive scanning, KSM sorts the memory pages by their
61 * contents into a data structure that holds pointers to the pages' locations.
63 * Since the contents of the pages may change at any moment, KSM cannot just
64 * insert the pages into a normal sorted tree and expect it to find anything.
65 * Therefore KSM uses two data structures - the stable and the unstable tree.
67 * The stable tree holds pointers to all the merged pages (ksm pages), sorted
68 * by their contents. Because each such page is write-protected, searching on
69 * this tree is fully assured to be working (except when pages are unmapped),
70 * and therefore this tree is called the stable tree.
72 * In addition to the stable tree, KSM uses a second data structure called the
73 * unstable tree: this tree holds pointers to pages which have been found to
74 * be "unchanged for a period of time". The unstable tree sorts these pages
75 * by their contents, but since they are not write-protected, KSM cannot rely
76 * upon the unstable tree to work correctly - the unstable tree is liable to
77 * be corrupted as its contents are modified, and so it is called unstable.
79 * KSM solves this problem by several techniques:
81 * 1) The unstable tree is flushed every time KSM completes scanning all
82 * memory areas, and then the tree is rebuilt again from the beginning.
83 * 2) KSM will only insert into the unstable tree, pages whose hash value
84 * has not changed since the previous scan of all memory areas.
85 * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
86 * colors of the nodes and not on their contents, assuring that even when
87 * the tree gets "corrupted" it won't get out of balance, so scanning time
88 * remains the same (also, searching and inserting nodes in an rbtree uses
89 * the same algorithm, so we have no overhead when we flush and rebuild).
90 * 4) KSM never flushes the stable tree, which means that even if it were to
91 * take 10 attempts to find a page in the unstable tree, once it is found,
92 * it is secured in the stable tree. (When we scan a new page, we first
93 * compare it against the stable tree, and then against the unstable tree.)
95 * If the merge_across_nodes tunable is unset, then KSM maintains multiple
96 * stable trees and multiple unstable trees: one of each for each NUMA node.
100 * struct mm_slot - ksm information per mm that is being scanned
101 * @link: link to the mm_slots hash list
102 * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
103 * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
104 * @mm: the mm that this information is valid for
107 struct hlist_node link;
108 struct list_head mm_list;
109 struct rmap_item *rmap_list;
110 struct mm_struct *mm;
114 * struct ksm_scan - cursor for scanning
115 * @mm_slot: the current mm_slot we are scanning
116 * @address: the next address inside that to be scanned
117 * @rmap_list: link to the next rmap to be scanned in the rmap_list
118 * @seqnr: count of completed full scans (needed when removing unstable node)
120 * There is only the one ksm_scan instance of this cursor structure.
123 struct mm_slot *mm_slot;
124 unsigned long address;
125 struct rmap_item **rmap_list;
130 * struct stable_node - node of the stable rbtree
131 * @node: rb node of this ksm page in the stable tree
132 * @head: (overlaying parent) &migrate_nodes indicates temporarily on that list
133 * @list: linked into migrate_nodes, pending placement in the proper node tree
134 * @hlist: hlist head of rmap_items using this ksm page
135 * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid)
136 * @nid: NUMA node id of stable tree in which linked (may not match kpfn)
140 struct rb_node node; /* when node of stable tree */
141 struct { /* when listed for migration */
142 struct list_head *head;
143 struct list_head list;
146 struct hlist_head hlist;
154 * struct rmap_item - reverse mapping item for virtual addresses
155 * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
156 * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
157 * @nid: NUMA node id of unstable tree in which linked (may not match page)
158 * @mm: the memory structure this rmap_item is pointing into
159 * @address: the virtual address this rmap_item tracks (+ flags in low bits)
160 * @oldchecksum: previous checksum of the page at that virtual address
161 * @node: rb node of this rmap_item in the unstable tree
162 * @head: pointer to stable_node heading this list in the stable tree
163 * @hlist: link into hlist of rmap_items hanging off that stable_node
166 struct rmap_item *rmap_list;
168 struct anon_vma *anon_vma; /* when stable */
170 int nid; /* when node of unstable tree */
173 struct mm_struct *mm;
174 unsigned long address; /* + low bits used for flags below */
175 unsigned int oldchecksum; /* when unstable */
177 struct rb_node node; /* when node of unstable tree */
178 struct { /* when listed from stable tree */
179 struct stable_node *head;
180 struct hlist_node hlist;
185 #define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */
186 #define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */
187 #define STABLE_FLAG 0x200 /* is listed from the stable tree */
189 /* The stable and unstable tree heads */
190 static struct rb_root one_stable_tree[1] = { RB_ROOT };
191 static struct rb_root one_unstable_tree[1] = { RB_ROOT };
192 static struct rb_root *root_stable_tree = one_stable_tree;
193 static struct rb_root *root_unstable_tree = one_unstable_tree;
195 /* Recently migrated nodes of stable tree, pending proper placement */
196 static LIST_HEAD(migrate_nodes);
198 #define MM_SLOTS_HASH_BITS 10
199 static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
201 static struct mm_slot ksm_mm_head = {
202 .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
204 static struct ksm_scan ksm_scan = {
205 .mm_slot = &ksm_mm_head,
208 static struct kmem_cache *rmap_item_cache;
209 static struct kmem_cache *stable_node_cache;
210 static struct kmem_cache *mm_slot_cache;
212 /* The number of nodes in the stable tree */
213 static unsigned long ksm_pages_shared;
215 /* The number of page slots additionally sharing those nodes */
216 static unsigned long ksm_pages_sharing;
218 /* The number of nodes in the unstable tree */
219 static unsigned long ksm_pages_unshared;
221 /* The number of rmap_items in use: to calculate pages_volatile */
222 static unsigned long ksm_rmap_items;
224 /* Number of pages ksmd should scan in one batch */
225 static unsigned int ksm_thread_pages_to_scan = 100;
227 /* Milliseconds ksmd should sleep between batches */
228 static unsigned int ksm_thread_sleep_millisecs = 20;
231 /* Zeroed when merging across nodes is not allowed */
232 static unsigned int ksm_merge_across_nodes = 1;
233 static int ksm_nr_node_ids = 1;
235 #define ksm_merge_across_nodes 1U
236 #define ksm_nr_node_ids 1
239 #define KSM_RUN_STOP 0
240 #define KSM_RUN_MERGE 1
241 #define KSM_RUN_UNMERGE 2
242 #define KSM_RUN_OFFLINE 4
243 static unsigned long ksm_run = KSM_RUN_STOP;
244 static void wait_while_offlining(void);
246 static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
247 static DEFINE_MUTEX(ksm_thread_mutex);
248 static DEFINE_SPINLOCK(ksm_mmlist_lock);
250 #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
251 sizeof(struct __struct), __alignof__(struct __struct),\
254 static int __init ksm_slab_init(void)
256 rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
257 if (!rmap_item_cache)
260 stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
261 if (!stable_node_cache)
264 mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
271 kmem_cache_destroy(stable_node_cache);
273 kmem_cache_destroy(rmap_item_cache);
278 static void __init ksm_slab_free(void)
280 kmem_cache_destroy(mm_slot_cache);
281 kmem_cache_destroy(stable_node_cache);
282 kmem_cache_destroy(rmap_item_cache);
283 mm_slot_cache = NULL;
286 static inline struct rmap_item *alloc_rmap_item(void)
288 struct rmap_item *rmap_item;
290 rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL);
296 static inline void free_rmap_item(struct rmap_item *rmap_item)
299 rmap_item->mm = NULL; /* debug safety */
300 kmem_cache_free(rmap_item_cache, rmap_item);
303 static inline struct stable_node *alloc_stable_node(void)
305 return kmem_cache_alloc(stable_node_cache, GFP_KERNEL);
308 static inline void free_stable_node(struct stable_node *stable_node)
310 kmem_cache_free(stable_node_cache, stable_node);
313 static inline struct mm_slot *alloc_mm_slot(void)
315 if (!mm_slot_cache) /* initialization failed */
317 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
320 static inline void free_mm_slot(struct mm_slot *mm_slot)
322 kmem_cache_free(mm_slot_cache, mm_slot);
325 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
327 struct mm_slot *slot;
329 hash_for_each_possible(mm_slots_hash, slot, link, (unsigned long)mm)
336 static void insert_to_mm_slots_hash(struct mm_struct *mm,
337 struct mm_slot *mm_slot)
340 hash_add(mm_slots_hash, &mm_slot->link, (unsigned long)mm);
344 * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
345 * page tables after it has passed through ksm_exit() - which, if necessary,
346 * takes mmap_sem briefly to serialize against them. ksm_exit() does not set
347 * a special flag: they can just back out as soon as mm_users goes to zero.
348 * ksm_test_exit() is used throughout to make this test for exit: in some
349 * places for correctness, in some places just to avoid unnecessary work.
351 static inline bool ksm_test_exit(struct mm_struct *mm)
353 return atomic_read(&mm->mm_users) == 0;
357 * We use break_ksm to break COW on a ksm page: it's a stripped down
359 * if (get_user_pages(current, mm, addr, 1, 1, 1, &page, NULL) == 1)
362 * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
363 * in case the application has unmapped and remapped mm,addr meanwhile.
364 * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP
365 * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
367 static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
374 page = follow_page(vma, addr, FOLL_GET | FOLL_MIGRATION);
375 if (IS_ERR_OR_NULL(page))
378 ret = handle_mm_fault(vma->vm_mm, vma, addr,
381 ret = VM_FAULT_WRITE;
383 } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_OOM)));
385 * We must loop because handle_mm_fault() may back out if there's
386 * any difficulty e.g. if pte accessed bit gets updated concurrently.
388 * VM_FAULT_WRITE is what we have been hoping for: it indicates that
389 * COW has been broken, even if the vma does not permit VM_WRITE;
390 * but note that a concurrent fault might break PageKsm for us.
392 * VM_FAULT_SIGBUS could occur if we race with truncation of the
393 * backing file, which also invalidates anonymous pages: that's
394 * okay, that truncation will have unmapped the PageKsm for us.
396 * VM_FAULT_OOM: at the time of writing (late July 2009), setting
397 * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
398 * current task has TIF_MEMDIE set, and will be OOM killed on return
399 * to user; and ksmd, having no mm, would never be chosen for that.
401 * But if the mm is in a limited mem_cgroup, then the fault may fail
402 * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
403 * even ksmd can fail in this way - though it's usually breaking ksm
404 * just to undo a merge it made a moment before, so unlikely to oom.
406 * That's a pity: we might therefore have more kernel pages allocated
407 * than we're counting as nodes in the stable tree; but ksm_do_scan
408 * will retry to break_cow on each pass, so should recover the page
409 * in due course. The important thing is to not let VM_MERGEABLE
410 * be cleared while any such pages might remain in the area.
412 return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
415 static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm,
418 struct vm_area_struct *vma;
419 if (ksm_test_exit(mm))
421 vma = find_vma(mm, addr);
422 if (!vma || vma->vm_start > addr)
424 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
429 static void break_cow(struct rmap_item *rmap_item)
431 struct mm_struct *mm = rmap_item->mm;
432 unsigned long addr = rmap_item->address;
433 struct vm_area_struct *vma;
436 * It is not an accident that whenever we want to break COW
437 * to undo, we also need to drop a reference to the anon_vma.
439 put_anon_vma(rmap_item->anon_vma);
441 down_read(&mm->mmap_sem);
442 vma = find_mergeable_vma(mm, addr);
444 break_ksm(vma, addr);
445 up_read(&mm->mmap_sem);
448 static struct page *page_trans_compound_anon(struct page *page)
450 if (PageTransCompound(page)) {
451 struct page *head = compound_head(page);
453 * head may actually be splitted and freed from under
454 * us but it's ok here.
462 static struct page *get_mergeable_page(struct rmap_item *rmap_item)
464 struct mm_struct *mm = rmap_item->mm;
465 unsigned long addr = rmap_item->address;
466 struct vm_area_struct *vma;
469 down_read(&mm->mmap_sem);
470 vma = find_mergeable_vma(mm, addr);
474 page = follow_page(vma, addr, FOLL_GET);
475 if (IS_ERR_OR_NULL(page))
477 if (PageAnon(page) || page_trans_compound_anon(page)) {
478 flush_anon_page(vma, page, addr);
479 flush_dcache_page(page);
484 up_read(&mm->mmap_sem);
489 * This helper is used for getting right index into array of tree roots.
490 * When merge_across_nodes knob is set to 1, there are only two rb-trees for
491 * stable and unstable pages from all nodes with roots in index 0. Otherwise,
492 * every node has its own stable and unstable tree.
494 static inline int get_kpfn_nid(unsigned long kpfn)
496 return ksm_merge_across_nodes ? 0 : NUMA(pfn_to_nid(kpfn));
499 static void remove_node_from_stable_tree(struct stable_node *stable_node)
501 struct rmap_item *rmap_item;
503 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
504 if (rmap_item->hlist.next)
508 put_anon_vma(rmap_item->anon_vma);
509 rmap_item->address &= PAGE_MASK;
513 if (stable_node->head == &migrate_nodes)
514 list_del(&stable_node->list);
516 rb_erase(&stable_node->node,
517 root_stable_tree + NUMA(stable_node->nid));
518 free_stable_node(stable_node);
522 * get_ksm_page: checks if the page indicated by the stable node
523 * is still its ksm page, despite having held no reference to it.
524 * In which case we can trust the content of the page, and it
525 * returns the gotten page; but if the page has now been zapped,
526 * remove the stale node from the stable tree and return NULL.
527 * But beware, the stable node's page might be being migrated.
529 * You would expect the stable_node to hold a reference to the ksm page.
530 * But if it increments the page's count, swapping out has to wait for
531 * ksmd to come around again before it can free the page, which may take
532 * seconds or even minutes: much too unresponsive. So instead we use a
533 * "keyhole reference": access to the ksm page from the stable node peeps
534 * out through its keyhole to see if that page still holds the right key,
535 * pointing back to this stable node. This relies on freeing a PageAnon
536 * page to reset its page->mapping to NULL, and relies on no other use of
537 * a page to put something that might look like our key in page->mapping.
538 * is on its way to being freed; but it is an anomaly to bear in mind.
540 static struct page *get_ksm_page(struct stable_node *stable_node, bool lock_it)
543 void *expected_mapping;
546 expected_mapping = (void *)stable_node +
547 (PAGE_MAPPING_ANON | PAGE_MAPPING_KSM);
549 kpfn = ACCESS_ONCE(stable_node->kpfn);
550 page = pfn_to_page(kpfn);
553 * page is computed from kpfn, so on most architectures reading
554 * page->mapping is naturally ordered after reading node->kpfn,
555 * but on Alpha we need to be more careful.
557 smp_read_barrier_depends();
558 if (ACCESS_ONCE(page->mapping) != expected_mapping)
562 * We cannot do anything with the page while its refcount is 0.
563 * Usually 0 means free, or tail of a higher-order page: in which
564 * case this node is no longer referenced, and should be freed;
565 * however, it might mean that the page is under page_freeze_refs().
566 * The __remove_mapping() case is easy, again the node is now stale;
567 * but if page is swapcache in migrate_page_move_mapping(), it might
568 * still be our page, in which case it's essential to keep the node.
570 while (!get_page_unless_zero(page)) {
572 * Another check for page->mapping != expected_mapping would
573 * work here too. We have chosen the !PageSwapCache test to
574 * optimize the common case, when the page is or is about to
575 * be freed: PageSwapCache is cleared (under spin_lock_irq)
576 * in the freeze_refs section of __remove_mapping(); but Anon
577 * page->mapping reset to NULL later, in free_pages_prepare().
579 if (!PageSwapCache(page))
584 if (ACCESS_ONCE(page->mapping) != expected_mapping) {
591 if (ACCESS_ONCE(page->mapping) != expected_mapping) {
601 * We come here from above when page->mapping or !PageSwapCache
602 * suggests that the node is stale; but it might be under migration.
603 * We need smp_rmb(), matching the smp_wmb() in ksm_migrate_page(),
604 * before checking whether node->kpfn has been changed.
607 if (ACCESS_ONCE(stable_node->kpfn) != kpfn)
609 remove_node_from_stable_tree(stable_node);
614 * Removing rmap_item from stable or unstable tree.
615 * This function will clean the information from the stable/unstable tree.
617 static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
619 if (rmap_item->address & STABLE_FLAG) {
620 struct stable_node *stable_node;
623 stable_node = rmap_item->head;
624 page = get_ksm_page(stable_node, true);
628 hlist_del(&rmap_item->hlist);
632 if (stable_node->hlist.first)
637 put_anon_vma(rmap_item->anon_vma);
638 rmap_item->address &= PAGE_MASK;
640 } else if (rmap_item->address & UNSTABLE_FLAG) {
643 * Usually ksmd can and must skip the rb_erase, because
644 * root_unstable_tree was already reset to RB_ROOT.
645 * But be careful when an mm is exiting: do the rb_erase
646 * if this rmap_item was inserted by this scan, rather
647 * than left over from before.
649 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
652 rb_erase(&rmap_item->node,
653 root_unstable_tree + NUMA(rmap_item->nid));
654 ksm_pages_unshared--;
655 rmap_item->address &= PAGE_MASK;
658 cond_resched(); /* we're called from many long loops */
661 static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
662 struct rmap_item **rmap_list)
665 struct rmap_item *rmap_item = *rmap_list;
666 *rmap_list = rmap_item->rmap_list;
667 remove_rmap_item_from_tree(rmap_item);
668 free_rmap_item(rmap_item);
673 * Though it's very tempting to unmerge rmap_items from stable tree rather
674 * than check every pte of a given vma, the locking doesn't quite work for
675 * that - an rmap_item is assigned to the stable tree after inserting ksm
676 * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing
677 * rmap_items from parent to child at fork time (so as not to waste time
678 * if exit comes before the next scan reaches it).
680 * Similarly, although we'd like to remove rmap_items (so updating counts
681 * and freeing memory) when unmerging an area, it's easier to leave that
682 * to the next pass of ksmd - consider, for example, how ksmd might be
683 * in cmp_and_merge_page on one of the rmap_items we would be removing.
685 static int unmerge_ksm_pages(struct vm_area_struct *vma,
686 unsigned long start, unsigned long end)
691 for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
692 if (ksm_test_exit(vma->vm_mm))
694 if (signal_pending(current))
697 err = break_ksm(vma, addr);
704 * Only called through the sysfs control interface:
706 static int remove_stable_node(struct stable_node *stable_node)
711 page = get_ksm_page(stable_node, true);
714 * get_ksm_page did remove_node_from_stable_tree itself.
719 if (WARN_ON_ONCE(page_mapped(page))) {
721 * This should not happen: but if it does, just refuse to let
722 * merge_across_nodes be switched - there is no need to panic.
727 * The stable node did not yet appear stale to get_ksm_page(),
728 * since that allows for an unmapped ksm page to be recognized
729 * right up until it is freed; but the node is safe to remove.
730 * This page might be in a pagevec waiting to be freed,
731 * or it might be PageSwapCache (perhaps under writeback),
732 * or it might have been removed from swapcache a moment ago.
734 set_page_stable_node(page, NULL);
735 remove_node_from_stable_tree(stable_node);
744 static int remove_all_stable_nodes(void)
746 struct stable_node *stable_node;
747 struct list_head *this, *next;
751 for (nid = 0; nid < ksm_nr_node_ids; nid++) {
752 while (root_stable_tree[nid].rb_node) {
753 stable_node = rb_entry(root_stable_tree[nid].rb_node,
754 struct stable_node, node);
755 if (remove_stable_node(stable_node)) {
757 break; /* proceed to next nid */
762 list_for_each_safe(this, next, &migrate_nodes) {
763 stable_node = list_entry(this, struct stable_node, list);
764 if (remove_stable_node(stable_node))
771 static int unmerge_and_remove_all_rmap_items(void)
773 struct mm_slot *mm_slot;
774 struct mm_struct *mm;
775 struct vm_area_struct *vma;
778 spin_lock(&ksm_mmlist_lock);
779 ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
780 struct mm_slot, mm_list);
781 spin_unlock(&ksm_mmlist_lock);
783 for (mm_slot = ksm_scan.mm_slot;
784 mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
786 down_read(&mm->mmap_sem);
787 for (vma = mm->mmap; vma; vma = vma->vm_next) {
788 if (ksm_test_exit(mm))
790 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
792 err = unmerge_ksm_pages(vma,
793 vma->vm_start, vma->vm_end);
798 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
800 spin_lock(&ksm_mmlist_lock);
801 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
802 struct mm_slot, mm_list);
803 if (ksm_test_exit(mm)) {
804 hash_del(&mm_slot->link);
805 list_del(&mm_slot->mm_list);
806 spin_unlock(&ksm_mmlist_lock);
808 free_mm_slot(mm_slot);
809 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
810 up_read(&mm->mmap_sem);
813 spin_unlock(&ksm_mmlist_lock);
814 up_read(&mm->mmap_sem);
818 /* Clean up stable nodes, but don't worry if some are still busy */
819 remove_all_stable_nodes();
824 up_read(&mm->mmap_sem);
825 spin_lock(&ksm_mmlist_lock);
826 ksm_scan.mm_slot = &ksm_mm_head;
827 spin_unlock(&ksm_mmlist_lock);
830 #endif /* CONFIG_SYSFS */
832 static u32 calc_checksum(struct page *page)
835 void *addr = kmap_atomic(page);
836 #ifdef CONFIG_OPTIMIZE_KSM
837 checksum = csum_partial(addr, PAGE_SIZE / 4, 17);
839 checksum = jhash2(addr, PAGE_SIZE / 4, 17);
845 static int memcmp_pages(struct page *page1, struct page *page2)
850 addr1 = kmap_atomic(page1);
851 addr2 = kmap_atomic(page2);
852 #ifdef CONFIG_KSM_ASSEMBLY_MEMCMP
853 ret = memcmpksm(addr1, addr2, PAGE_SIZE);
855 ret = memcmp(addr1, addr2, PAGE_SIZE);
857 kunmap_atomic(addr2);
858 kunmap_atomic(addr1);
862 static inline int pages_identical(struct page *page1, struct page *page2)
864 return !memcmp_pages(page1, page2);
867 static int write_protect_page(struct vm_area_struct *vma, struct page *page,
870 struct mm_struct *mm = vma->vm_mm;
876 unsigned long mmun_start; /* For mmu_notifiers */
877 unsigned long mmun_end; /* For mmu_notifiers */
879 addr = page_address_in_vma(page, vma);
883 BUG_ON(PageTransCompound(page));
886 mmun_end = addr + PAGE_SIZE;
887 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
889 ptep = page_check_address(page, mm, addr, &ptl, 0);
893 if (pte_write(*ptep) || pte_dirty(*ptep)) {
896 swapped = PageSwapCache(page);
897 flush_cache_page(vma, addr, page_to_pfn(page));
899 * Ok this is tricky, when get_user_pages_fast() run it doesn't
900 * take any lock, therefore the check that we are going to make
901 * with the pagecount against the mapcount is racey and
902 * O_DIRECT can happen right after the check.
903 * So we clear the pte and flush the tlb before the check
904 * this assure us that no O_DIRECT can happen after the check
905 * or in the middle of the check.
907 entry = ptep_clear_flush(vma, addr, ptep);
909 * Check that no O_DIRECT or similar I/O is in progress on the
912 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
913 set_pte_at(mm, addr, ptep, entry);
916 if (pte_dirty(entry))
917 set_page_dirty(page);
918 entry = pte_mkclean(pte_wrprotect(entry));
919 set_pte_at_notify(mm, addr, ptep, entry);
925 pte_unmap_unlock(ptep, ptl);
927 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
933 * replace_page - replace page in vma by new ksm page
934 * @vma: vma that holds the pte pointing to page
935 * @page: the page we are replacing by kpage
936 * @kpage: the ksm page we replace page by
937 * @orig_pte: the original value of the pte
939 * Returns 0 on success, -EFAULT on failure.
941 static int replace_page(struct vm_area_struct *vma, struct page *page,
942 struct page *kpage, pte_t orig_pte)
944 struct mm_struct *mm = vma->vm_mm;
950 unsigned long mmun_start; /* For mmu_notifiers */
951 unsigned long mmun_end; /* For mmu_notifiers */
953 addr = page_address_in_vma(page, vma);
957 pmd = mm_find_pmd(mm, addr);
960 BUG_ON(pmd_trans_huge(*pmd));
963 mmun_end = addr + PAGE_SIZE;
964 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
966 ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
967 if (!pte_same(*ptep, orig_pte)) {
968 pte_unmap_unlock(ptep, ptl);
973 page_add_anon_rmap(kpage, vma, addr);
975 flush_cache_page(vma, addr, pte_pfn(*ptep));
976 ptep_clear_flush(vma, addr, ptep);
977 set_pte_at_notify(mm, addr, ptep, mk_pte(kpage, vma->vm_page_prot));
979 page_remove_rmap(page);
980 if (!page_mapped(page))
981 try_to_free_swap(page);
984 pte_unmap_unlock(ptep, ptl);
987 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
992 static int page_trans_compound_anon_split(struct page *page)
995 struct page *transhuge_head = page_trans_compound_anon(page);
996 if (transhuge_head) {
997 /* Get the reference on the head to split it. */
998 if (get_page_unless_zero(transhuge_head)) {
1000 * Recheck we got the reference while the head
1001 * was still anonymous.
1003 if (PageAnon(transhuge_head))
1004 ret = split_huge_page(transhuge_head);
1007 * Retry later if split_huge_page run
1011 put_page(transhuge_head);
1013 /* Retry later if split_huge_page run from under us. */
1020 * try_to_merge_one_page - take two pages and merge them into one
1021 * @vma: the vma that holds the pte pointing to page
1022 * @page: the PageAnon page that we want to replace with kpage
1023 * @kpage: the PageKsm page that we want to map instead of page,
1024 * or NULL the first time when we want to use page as kpage.
1026 * This function returns 0 if the pages were merged, -EFAULT otherwise.
1028 static int try_to_merge_one_page(struct vm_area_struct *vma,
1029 struct page *page, struct page *kpage)
1031 pte_t orig_pte = __pte(0);
1034 if (page == kpage) /* ksm page forked */
1037 if (!(vma->vm_flags & VM_MERGEABLE))
1039 if (PageTransCompound(page) && page_trans_compound_anon_split(page))
1041 BUG_ON(PageTransCompound(page));
1042 if (!PageAnon(page))
1046 * We need the page lock to read a stable PageSwapCache in
1047 * write_protect_page(). We use trylock_page() instead of
1048 * lock_page() because we don't want to wait here - we
1049 * prefer to continue scanning and merging different pages,
1050 * then come back to this page when it is unlocked.
1052 if (!trylock_page(page))
1055 * If this anonymous page is mapped only here, its pte may need
1056 * to be write-protected. If it's mapped elsewhere, all of its
1057 * ptes are necessarily already write-protected. But in either
1058 * case, we need to lock and check page_count is not raised.
1060 if (write_protect_page(vma, page, &orig_pte) == 0) {
1063 * While we hold page lock, upgrade page from
1064 * PageAnon+anon_vma to PageKsm+NULL stable_node:
1065 * stable_tree_insert() will update stable_node.
1067 set_page_stable_node(page, NULL);
1068 mark_page_accessed(page);
1070 } else if (pages_identical(page, kpage))
1071 err = replace_page(vma, page, kpage, orig_pte);
1074 if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
1075 munlock_vma_page(page);
1076 if (!PageMlocked(kpage)) {
1079 mlock_vma_page(kpage);
1080 page = kpage; /* for final unlock */
1090 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
1091 * but no new kernel page is allocated: kpage must already be a ksm page.
1093 * This function returns 0 if the pages were merged, -EFAULT otherwise.
1095 static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
1096 struct page *page, struct page *kpage)
1098 struct mm_struct *mm = rmap_item->mm;
1099 struct vm_area_struct *vma;
1102 down_read(&mm->mmap_sem);
1103 if (ksm_test_exit(mm))
1105 vma = find_vma(mm, rmap_item->address);
1106 if (!vma || vma->vm_start > rmap_item->address)
1109 err = try_to_merge_one_page(vma, page, kpage);
1113 /* Unstable nid is in union with stable anon_vma: remove first */
1114 remove_rmap_item_from_tree(rmap_item);
1116 /* Must get reference to anon_vma while still holding mmap_sem */
1117 rmap_item->anon_vma = vma->anon_vma;
1118 get_anon_vma(vma->anon_vma);
1120 up_read(&mm->mmap_sem);
1125 * try_to_merge_two_pages - take two identical pages and prepare them
1126 * to be merged into one page.
1128 * This function returns the kpage if we successfully merged two identical
1129 * pages into one ksm page, NULL otherwise.
1131 * Note that this function upgrades page to ksm page: if one of the pages
1132 * is already a ksm page, try_to_merge_with_ksm_page should be used.
1134 static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
1136 struct rmap_item *tree_rmap_item,
1137 struct page *tree_page)
1141 err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
1143 err = try_to_merge_with_ksm_page(tree_rmap_item,
1146 * If that fails, we have a ksm page with only one pte
1147 * pointing to it: so break it.
1150 break_cow(rmap_item);
1152 return err ? NULL : page;
1156 * stable_tree_search - search for page inside the stable tree
1158 * This function checks if there is a page inside the stable tree
1159 * with identical content to the page that we are scanning right now.
1161 * This function returns the stable tree node of identical content if found,
1164 static struct page *stable_tree_search(struct page *page)
1167 struct rb_root *root;
1168 struct rb_node **new;
1169 struct rb_node *parent;
1170 struct stable_node *stable_node;
1171 struct stable_node *page_node;
1173 page_node = page_stable_node(page);
1174 if (page_node && page_node->head != &migrate_nodes) {
1175 /* ksm page forked */
1180 nid = get_kpfn_nid(page_to_pfn(page));
1181 root = root_stable_tree + nid;
1183 new = &root->rb_node;
1187 struct page *tree_page;
1191 stable_node = rb_entry(*new, struct stable_node, node);
1192 tree_page = get_ksm_page(stable_node, false);
1196 ret = memcmp_pages(page, tree_page);
1197 put_page(tree_page);
1201 new = &parent->rb_left;
1203 new = &parent->rb_right;
1206 * Lock and unlock the stable_node's page (which
1207 * might already have been migrated) so that page
1208 * migration is sure to notice its raised count.
1209 * It would be more elegant to return stable_node
1210 * than kpage, but that involves more changes.
1212 tree_page = get_ksm_page(stable_node, true);
1214 unlock_page(tree_page);
1215 if (get_kpfn_nid(stable_node->kpfn) !=
1216 NUMA(stable_node->nid)) {
1217 put_page(tree_page);
1223 * There is now a place for page_node, but the tree may
1224 * have been rebalanced, so re-evaluate parent and new.
1235 list_del(&page_node->list);
1236 DO_NUMA(page_node->nid = nid);
1237 rb_link_node(&page_node->node, parent, new);
1238 rb_insert_color(&page_node->node, root);
1244 list_del(&page_node->list);
1245 DO_NUMA(page_node->nid = nid);
1246 rb_replace_node(&stable_node->node, &page_node->node, root);
1249 rb_erase(&stable_node->node, root);
1252 stable_node->head = &migrate_nodes;
1253 list_add(&stable_node->list, stable_node->head);
1258 * stable_tree_insert - insert stable tree node pointing to new ksm page
1259 * into the stable tree.
1261 * This function returns the stable tree node just allocated on success,
1264 static struct stable_node *stable_tree_insert(struct page *kpage)
1268 struct rb_root *root;
1269 struct rb_node **new;
1270 struct rb_node *parent = NULL;
1271 struct stable_node *stable_node;
1273 kpfn = page_to_pfn(kpage);
1274 nid = get_kpfn_nid(kpfn);
1275 root = root_stable_tree + nid;
1276 new = &root->rb_node;
1279 struct page *tree_page;
1283 stable_node = rb_entry(*new, struct stable_node, node);
1284 tree_page = get_ksm_page(stable_node, false);
1288 ret = memcmp_pages(kpage, tree_page);
1289 put_page(tree_page);
1293 new = &parent->rb_left;
1295 new = &parent->rb_right;
1298 * It is not a bug that stable_tree_search() didn't
1299 * find this node: because at that time our page was
1300 * not yet write-protected, so may have changed since.
1306 stable_node = alloc_stable_node();
1310 INIT_HLIST_HEAD(&stable_node->hlist);
1311 stable_node->kpfn = kpfn;
1312 set_page_stable_node(kpage, stable_node);
1313 DO_NUMA(stable_node->nid = nid);
1314 rb_link_node(&stable_node->node, parent, new);
1315 rb_insert_color(&stable_node->node, root);
1321 * unstable_tree_search_insert - search for identical page,
1322 * else insert rmap_item into the unstable tree.
1324 * This function searches for a page in the unstable tree identical to the
1325 * page currently being scanned; and if no identical page is found in the
1326 * tree, we insert rmap_item as a new object into the unstable tree.
1328 * This function returns pointer to rmap_item found to be identical
1329 * to the currently scanned page, NULL otherwise.
1331 * This function does both searching and inserting, because they share
1332 * the same walking algorithm in an rbtree.
1335 struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1337 struct page **tree_pagep)
1339 struct rb_node **new;
1340 struct rb_root *root;
1341 struct rb_node *parent = NULL;
1344 nid = get_kpfn_nid(page_to_pfn(page));
1345 root = root_unstable_tree + nid;
1346 new = &root->rb_node;
1349 struct rmap_item *tree_rmap_item;
1350 struct page *tree_page;
1354 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
1355 tree_page = get_mergeable_page(tree_rmap_item);
1356 if (IS_ERR_OR_NULL(tree_page))
1360 * Don't substitute a ksm page for a forked page.
1362 if (page == tree_page) {
1363 put_page(tree_page);
1367 ret = memcmp_pages(page, tree_page);
1371 put_page(tree_page);
1372 new = &parent->rb_left;
1373 } else if (ret > 0) {
1374 put_page(tree_page);
1375 new = &parent->rb_right;
1376 } else if (!ksm_merge_across_nodes &&
1377 page_to_nid(tree_page) != nid) {
1379 * If tree_page has been migrated to another NUMA node,
1380 * it will be flushed out and put in the right unstable
1381 * tree next time: only merge with it when across_nodes.
1383 put_page(tree_page);
1386 *tree_pagep = tree_page;
1387 return tree_rmap_item;
1391 rmap_item->address |= UNSTABLE_FLAG;
1392 rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1393 DO_NUMA(rmap_item->nid = nid);
1394 rb_link_node(&rmap_item->node, parent, new);
1395 rb_insert_color(&rmap_item->node, root);
1397 ksm_pages_unshared++;
1402 * stable_tree_append - add another rmap_item to the linked list of
1403 * rmap_items hanging off a given node of the stable tree, all sharing
1404 * the same ksm page.
1406 static void stable_tree_append(struct rmap_item *rmap_item,
1407 struct stable_node *stable_node)
1409 rmap_item->head = stable_node;
1410 rmap_item->address |= STABLE_FLAG;
1411 hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
1413 if (rmap_item->hlist.next)
1414 ksm_pages_sharing++;
1420 * cmp_and_merge_page - first see if page can be merged into the stable tree;
1421 * if not, compare checksum to previous and if it's the same, see if page can
1422 * be inserted into the unstable tree, or merged with a page already there and
1423 * both transferred to the stable tree.
1425 * @page: the page that we are searching identical page to.
1426 * @rmap_item: the reverse mapping into the virtual address of this page
1428 static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1430 struct rmap_item *tree_rmap_item;
1431 struct page *tree_page = NULL;
1432 struct stable_node *stable_node;
1434 unsigned int checksum;
1437 stable_node = page_stable_node(page);
1439 if (stable_node->head != &migrate_nodes &&
1440 get_kpfn_nid(stable_node->kpfn) != NUMA(stable_node->nid)) {
1441 rb_erase(&stable_node->node,
1442 root_stable_tree + NUMA(stable_node->nid));
1443 stable_node->head = &migrate_nodes;
1444 list_add(&stable_node->list, stable_node->head);
1446 if (stable_node->head != &migrate_nodes &&
1447 rmap_item->head == stable_node)
1451 /* We first start with searching the page inside the stable tree */
1452 kpage = stable_tree_search(page);
1453 if (kpage == page && rmap_item->head == stable_node) {
1458 remove_rmap_item_from_tree(rmap_item);
1461 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
1464 * The page was successfully merged:
1465 * add its rmap_item to the stable tree.
1468 stable_tree_append(rmap_item, page_stable_node(kpage));
1476 * If the hash value of the page has changed from the last time
1477 * we calculated it, this page is changing frequently: therefore we
1478 * don't want to insert it in the unstable tree, and we don't want
1479 * to waste our time searching for something identical to it there.
1481 checksum = calc_checksum(page);
1482 if (rmap_item->oldchecksum != checksum) {
1483 rmap_item->oldchecksum = checksum;
1488 unstable_tree_search_insert(rmap_item, page, &tree_page);
1489 if (tree_rmap_item) {
1490 kpage = try_to_merge_two_pages(rmap_item, page,
1491 tree_rmap_item, tree_page);
1492 put_page(tree_page);
1495 * The pages were successfully merged: insert new
1496 * node in the stable tree and add both rmap_items.
1499 stable_node = stable_tree_insert(kpage);
1501 stable_tree_append(tree_rmap_item, stable_node);
1502 stable_tree_append(rmap_item, stable_node);
1507 * If we fail to insert the page into the stable tree,
1508 * we will have 2 virtual addresses that are pointing
1509 * to a ksm page left outside the stable tree,
1510 * in which case we need to break_cow on both.
1513 break_cow(tree_rmap_item);
1514 break_cow(rmap_item);
1520 static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
1521 struct rmap_item **rmap_list,
1524 struct rmap_item *rmap_item;
1526 while (*rmap_list) {
1527 rmap_item = *rmap_list;
1528 if ((rmap_item->address & PAGE_MASK) == addr)
1530 if (rmap_item->address > addr)
1532 *rmap_list = rmap_item->rmap_list;
1533 remove_rmap_item_from_tree(rmap_item);
1534 free_rmap_item(rmap_item);
1537 rmap_item = alloc_rmap_item();
1539 /* It has already been zeroed */
1540 rmap_item->mm = mm_slot->mm;
1541 rmap_item->address = addr;
1542 rmap_item->rmap_list = *rmap_list;
1543 *rmap_list = rmap_item;
1548 static struct rmap_item *scan_get_next_rmap_item(struct page **page)
1550 struct mm_struct *mm;
1551 struct mm_slot *slot;
1552 struct vm_area_struct *vma;
1553 struct rmap_item *rmap_item;
1556 if (list_empty(&ksm_mm_head.mm_list))
1559 slot = ksm_scan.mm_slot;
1560 if (slot == &ksm_mm_head) {
1562 * A number of pages can hang around indefinitely on per-cpu
1563 * pagevecs, raised page count preventing write_protect_page
1564 * from merging them. Though it doesn't really matter much,
1565 * it is puzzling to see some stuck in pages_volatile until
1566 * other activity jostles them out, and they also prevented
1567 * LTP's KSM test from succeeding deterministically; so drain
1568 * them here (here rather than on entry to ksm_do_scan(),
1569 * so we don't IPI too often when pages_to_scan is set low).
1571 lru_add_drain_all();
1574 * Whereas stale stable_nodes on the stable_tree itself
1575 * get pruned in the regular course of stable_tree_search(),
1576 * those moved out to the migrate_nodes list can accumulate:
1577 * so prune them once before each full scan.
1579 if (!ksm_merge_across_nodes) {
1580 struct stable_node *stable_node;
1581 struct list_head *this, *next;
1584 list_for_each_safe(this, next, &migrate_nodes) {
1585 stable_node = list_entry(this,
1586 struct stable_node, list);
1587 page = get_ksm_page(stable_node, false);
1594 for (nid = 0; nid < ksm_nr_node_ids; nid++)
1595 root_unstable_tree[nid] = RB_ROOT;
1597 spin_lock(&ksm_mmlist_lock);
1598 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1599 ksm_scan.mm_slot = slot;
1600 spin_unlock(&ksm_mmlist_lock);
1602 * Although we tested list_empty() above, a racing __ksm_exit
1603 * of the last mm on the list may have removed it since then.
1605 if (slot == &ksm_mm_head)
1608 ksm_scan.address = 0;
1609 ksm_scan.rmap_list = &slot->rmap_list;
1613 down_read(&mm->mmap_sem);
1614 if (ksm_test_exit(mm))
1617 vma = find_vma(mm, ksm_scan.address);
1619 for (; vma; vma = vma->vm_next) {
1620 if (!(vma->vm_flags & VM_MERGEABLE))
1622 if (ksm_scan.address < vma->vm_start)
1623 ksm_scan.address = vma->vm_start;
1625 ksm_scan.address = vma->vm_end;
1627 while (ksm_scan.address < vma->vm_end) {
1628 if (ksm_test_exit(mm))
1630 *page = follow_page(vma, ksm_scan.address, FOLL_GET);
1631 if (IS_ERR_OR_NULL(*page)) {
1632 ksm_scan.address += PAGE_SIZE;
1636 if (PageAnon(*page) ||
1637 page_trans_compound_anon(*page)) {
1638 flush_anon_page(vma, *page, ksm_scan.address);
1639 flush_dcache_page(*page);
1640 rmap_item = get_next_rmap_item(slot,
1641 ksm_scan.rmap_list, ksm_scan.address);
1643 ksm_scan.rmap_list =
1644 &rmap_item->rmap_list;
1645 ksm_scan.address += PAGE_SIZE;
1648 up_read(&mm->mmap_sem);
1652 ksm_scan.address += PAGE_SIZE;
1657 if (ksm_test_exit(mm)) {
1658 ksm_scan.address = 0;
1659 ksm_scan.rmap_list = &slot->rmap_list;
1662 * Nuke all the rmap_items that are above this current rmap:
1663 * because there were no VM_MERGEABLE vmas with such addresses.
1665 remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
1667 spin_lock(&ksm_mmlist_lock);
1668 ksm_scan.mm_slot = list_entry(slot->mm_list.next,
1669 struct mm_slot, mm_list);
1670 if (ksm_scan.address == 0) {
1672 * We've completed a full scan of all vmas, holding mmap_sem
1673 * throughout, and found no VM_MERGEABLE: so do the same as
1674 * __ksm_exit does to remove this mm from all our lists now.
1675 * This applies either when cleaning up after __ksm_exit
1676 * (but beware: we can reach here even before __ksm_exit),
1677 * or when all VM_MERGEABLE areas have been unmapped (and
1678 * mmap_sem then protects against race with MADV_MERGEABLE).
1680 hash_del(&slot->link);
1681 list_del(&slot->mm_list);
1682 spin_unlock(&ksm_mmlist_lock);
1685 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1686 up_read(&mm->mmap_sem);
1689 spin_unlock(&ksm_mmlist_lock);
1690 up_read(&mm->mmap_sem);
1693 /* Repeat until we've completed scanning the whole list */
1694 slot = ksm_scan.mm_slot;
1695 if (slot != &ksm_mm_head)
1703 * ksm_do_scan - the ksm scanner main worker function.
1704 * @scan_npages - number of pages we want to scan before we return.
1706 static void ksm_do_scan(unsigned int scan_npages)
1708 struct rmap_item *rmap_item;
1709 struct page *uninitialized_var(page);
1711 while (scan_npages-- && likely(!freezing(current))) {
1713 rmap_item = scan_get_next_rmap_item(&page);
1716 cmp_and_merge_page(page, rmap_item);
1721 static int ksmd_should_run(void)
1723 return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
1726 static int ksm_scan_thread(void *nothing)
1729 set_user_nice(current, 5);
1731 while (!kthread_should_stop()) {
1732 mutex_lock(&ksm_thread_mutex);
1733 wait_while_offlining();
1734 if (ksmd_should_run())
1735 ksm_do_scan(ksm_thread_pages_to_scan);
1736 mutex_unlock(&ksm_thread_mutex);
1740 if (ksmd_should_run()) {
1741 schedule_timeout_interruptible(
1742 msecs_to_jiffies(ksm_thread_sleep_millisecs));
1744 wait_event_freezable(ksm_thread_wait,
1745 ksmd_should_run() || kthread_should_stop());
1751 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1752 unsigned long end, int advice, unsigned long *vm_flags)
1754 struct mm_struct *mm = vma->vm_mm;
1758 case MADV_MERGEABLE:
1760 * Be somewhat over-protective for now!
1762 if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE |
1763 VM_PFNMAP | VM_IO | VM_DONTEXPAND |
1764 VM_HUGETLB | VM_NONLINEAR | VM_MIXEDMAP))
1765 return 0; /* just ignore the advice */
1768 if (*vm_flags & VM_SAO)
1772 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
1773 err = __ksm_enter(mm);
1778 *vm_flags |= VM_MERGEABLE;
1781 case MADV_UNMERGEABLE:
1782 if (!(*vm_flags & VM_MERGEABLE))
1783 return 0; /* just ignore the advice */
1785 if (vma->anon_vma) {
1786 err = unmerge_ksm_pages(vma, start, end);
1791 *vm_flags &= ~VM_MERGEABLE;
1798 int __ksm_enter(struct mm_struct *mm)
1800 struct mm_slot *mm_slot;
1803 mm_slot = alloc_mm_slot();
1807 /* Check ksm_run too? Would need tighter locking */
1808 needs_wakeup = list_empty(&ksm_mm_head.mm_list);
1810 spin_lock(&ksm_mmlist_lock);
1811 insert_to_mm_slots_hash(mm, mm_slot);
1813 * When KSM_RUN_MERGE (or KSM_RUN_STOP),
1814 * insert just behind the scanning cursor, to let the area settle
1815 * down a little; when fork is followed by immediate exec, we don't
1816 * want ksmd to waste time setting up and tearing down an rmap_list.
1818 * But when KSM_RUN_UNMERGE, it's important to insert ahead of its
1819 * scanning cursor, otherwise KSM pages in newly forked mms will be
1820 * missed: then we might as well insert at the end of the list.
1822 if (ksm_run & KSM_RUN_UNMERGE)
1823 list_add_tail(&mm_slot->mm_list, &ksm_mm_head.mm_list);
1825 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
1826 spin_unlock(&ksm_mmlist_lock);
1828 set_bit(MMF_VM_MERGEABLE, &mm->flags);
1829 atomic_inc(&mm->mm_count);
1832 wake_up_interruptible(&ksm_thread_wait);
1837 void __ksm_exit(struct mm_struct *mm)
1839 struct mm_slot *mm_slot;
1840 int easy_to_free = 0;
1843 * This process is exiting: if it's straightforward (as is the
1844 * case when ksmd was never running), free mm_slot immediately.
1845 * But if it's at the cursor or has rmap_items linked to it, use
1846 * mmap_sem to synchronize with any break_cows before pagetables
1847 * are freed, and leave the mm_slot on the list for ksmd to free.
1848 * Beware: ksm may already have noticed it exiting and freed the slot.
1851 spin_lock(&ksm_mmlist_lock);
1852 mm_slot = get_mm_slot(mm);
1853 if (mm_slot && ksm_scan.mm_slot != mm_slot) {
1854 if (!mm_slot->rmap_list) {
1855 hash_del(&mm_slot->link);
1856 list_del(&mm_slot->mm_list);
1859 list_move(&mm_slot->mm_list,
1860 &ksm_scan.mm_slot->mm_list);
1863 spin_unlock(&ksm_mmlist_lock);
1866 free_mm_slot(mm_slot);
1867 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1869 } else if (mm_slot) {
1870 down_write(&mm->mmap_sem);
1871 up_write(&mm->mmap_sem);
1875 struct page *ksm_might_need_to_copy(struct page *page,
1876 struct vm_area_struct *vma, unsigned long address)
1878 struct anon_vma *anon_vma = page_anon_vma(page);
1879 struct page *new_page;
1881 if (PageKsm(page)) {
1882 if (page_stable_node(page) &&
1883 !(ksm_run & KSM_RUN_UNMERGE))
1884 return page; /* no need to copy it */
1885 } else if (!anon_vma) {
1886 return page; /* no need to copy it */
1887 } else if (anon_vma->root == vma->anon_vma->root &&
1888 page->index == linear_page_index(vma, address)) {
1889 return page; /* still no need to copy it */
1891 if (!PageUptodate(page))
1892 return page; /* let do_swap_page report the error */
1894 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1896 copy_user_highpage(new_page, page, address, vma);
1898 SetPageDirty(new_page);
1899 __SetPageUptodate(new_page);
1900 __set_page_locked(new_page);
1906 int page_referenced_ksm(struct page *page, struct mem_cgroup *memcg,
1907 unsigned long *vm_flags)
1909 struct stable_node *stable_node;
1910 struct rmap_item *rmap_item;
1911 unsigned int mapcount = page_mapcount(page);
1913 int search_new_forks = 0;
1915 VM_BUG_ON(!PageKsm(page));
1916 VM_BUG_ON(!PageLocked(page));
1918 stable_node = page_stable_node(page);
1922 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
1923 struct anon_vma *anon_vma = rmap_item->anon_vma;
1924 struct anon_vma_chain *vmac;
1925 struct vm_area_struct *vma;
1927 anon_vma_lock_read(anon_vma);
1928 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
1931 if (rmap_item->address < vma->vm_start ||
1932 rmap_item->address >= vma->vm_end)
1935 * Initially we examine only the vma which covers this
1936 * rmap_item; but later, if there is still work to do,
1937 * we examine covering vmas in other mms: in case they
1938 * were forked from the original since ksmd passed.
1940 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1943 if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
1946 referenced += page_referenced_one(page, vma,
1947 rmap_item->address, &mapcount, vm_flags);
1948 if (!search_new_forks || !mapcount)
1951 anon_vma_unlock_read(anon_vma);
1955 if (!search_new_forks++)
1961 int try_to_unmap_ksm(struct page *page, enum ttu_flags flags)
1963 struct stable_node *stable_node;
1964 struct rmap_item *rmap_item;
1965 int ret = SWAP_AGAIN;
1966 int search_new_forks = 0;
1968 VM_BUG_ON(!PageKsm(page));
1969 VM_BUG_ON(!PageLocked(page));
1971 stable_node = page_stable_node(page);
1975 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
1976 struct anon_vma *anon_vma = rmap_item->anon_vma;
1977 struct anon_vma_chain *vmac;
1978 struct vm_area_struct *vma;
1980 anon_vma_lock_read(anon_vma);
1981 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
1984 if (rmap_item->address < vma->vm_start ||
1985 rmap_item->address >= vma->vm_end)
1988 * Initially we examine only the vma which covers this
1989 * rmap_item; but later, if there is still work to do,
1990 * we examine covering vmas in other mms: in case they
1991 * were forked from the original since ksmd passed.
1993 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1996 ret = try_to_unmap_one(page, vma,
1997 rmap_item->address, flags);
1998 if (ret != SWAP_AGAIN || !page_mapped(page)) {
1999 anon_vma_unlock_read(anon_vma);
2003 anon_vma_unlock_read(anon_vma);
2005 if (!search_new_forks++)
2011 #ifdef CONFIG_MIGRATION
2012 int rmap_walk_ksm(struct page *page, int (*rmap_one)(struct page *,
2013 struct vm_area_struct *, unsigned long, void *), void *arg)
2015 struct stable_node *stable_node;
2016 struct rmap_item *rmap_item;
2017 int ret = SWAP_AGAIN;
2018 int search_new_forks = 0;
2020 VM_BUG_ON(!PageKsm(page));
2021 VM_BUG_ON(!PageLocked(page));
2023 stable_node = page_stable_node(page);
2027 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
2028 struct anon_vma *anon_vma = rmap_item->anon_vma;
2029 struct anon_vma_chain *vmac;
2030 struct vm_area_struct *vma;
2032 anon_vma_lock_read(anon_vma);
2033 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
2036 if (rmap_item->address < vma->vm_start ||
2037 rmap_item->address >= vma->vm_end)
2040 * Initially we examine only the vma which covers this
2041 * rmap_item; but later, if there is still work to do,
2042 * we examine covering vmas in other mms: in case they
2043 * were forked from the original since ksmd passed.
2045 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
2048 ret = rmap_one(page, vma, rmap_item->address, arg);
2049 if (ret != SWAP_AGAIN) {
2050 anon_vma_unlock_read(anon_vma);
2054 anon_vma_unlock_read(anon_vma);
2056 if (!search_new_forks++)
2062 void ksm_migrate_page(struct page *newpage, struct page *oldpage)
2064 struct stable_node *stable_node;
2066 VM_BUG_ON(!PageLocked(oldpage));
2067 VM_BUG_ON(!PageLocked(newpage));
2068 VM_BUG_ON(newpage->mapping != oldpage->mapping);
2070 stable_node = page_stable_node(newpage);
2072 VM_BUG_ON(stable_node->kpfn != page_to_pfn(oldpage));
2073 stable_node->kpfn = page_to_pfn(newpage);
2075 * newpage->mapping was set in advance; now we need smp_wmb()
2076 * to make sure that the new stable_node->kpfn is visible
2077 * to get_ksm_page() before it can see that oldpage->mapping
2078 * has gone stale (or that PageSwapCache has been cleared).
2081 set_page_stable_node(oldpage, NULL);
2084 #endif /* CONFIG_MIGRATION */
2086 #ifdef CONFIG_MEMORY_HOTREMOVE
2087 static int just_wait(void *word)
2093 static void wait_while_offlining(void)
2095 while (ksm_run & KSM_RUN_OFFLINE) {
2096 mutex_unlock(&ksm_thread_mutex);
2097 wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE),
2098 just_wait, TASK_UNINTERRUPTIBLE);
2099 mutex_lock(&ksm_thread_mutex);
2103 static void ksm_check_stable_tree(unsigned long start_pfn,
2104 unsigned long end_pfn)
2106 struct stable_node *stable_node;
2107 struct list_head *this, *next;
2108 struct rb_node *node;
2111 for (nid = 0; nid < ksm_nr_node_ids; nid++) {
2112 node = rb_first(root_stable_tree + nid);
2114 stable_node = rb_entry(node, struct stable_node, node);
2115 if (stable_node->kpfn >= start_pfn &&
2116 stable_node->kpfn < end_pfn) {
2118 * Don't get_ksm_page, page has already gone:
2119 * which is why we keep kpfn instead of page*
2121 remove_node_from_stable_tree(stable_node);
2122 node = rb_first(root_stable_tree + nid);
2124 node = rb_next(node);
2128 list_for_each_safe(this, next, &migrate_nodes) {
2129 stable_node = list_entry(this, struct stable_node, list);
2130 if (stable_node->kpfn >= start_pfn &&
2131 stable_node->kpfn < end_pfn)
2132 remove_node_from_stable_tree(stable_node);
2137 static int ksm_memory_callback(struct notifier_block *self,
2138 unsigned long action, void *arg)
2140 struct memory_notify *mn = arg;
2143 case MEM_GOING_OFFLINE:
2145 * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items()
2146 * and remove_all_stable_nodes() while memory is going offline:
2147 * it is unsafe for them to touch the stable tree at this time.
2148 * But unmerge_ksm_pages(), rmap lookups and other entry points
2149 * which do not need the ksm_thread_mutex are all safe.
2151 mutex_lock(&ksm_thread_mutex);
2152 ksm_run |= KSM_RUN_OFFLINE;
2153 mutex_unlock(&ksm_thread_mutex);
2158 * Most of the work is done by page migration; but there might
2159 * be a few stable_nodes left over, still pointing to struct
2160 * pages which have been offlined: prune those from the tree,
2161 * otherwise get_ksm_page() might later try to access a
2162 * non-existent struct page.
2164 ksm_check_stable_tree(mn->start_pfn,
2165 mn->start_pfn + mn->nr_pages);
2168 case MEM_CANCEL_OFFLINE:
2169 mutex_lock(&ksm_thread_mutex);
2170 ksm_run &= ~KSM_RUN_OFFLINE;
2171 mutex_unlock(&ksm_thread_mutex);
2173 smp_mb(); /* wake_up_bit advises this */
2174 wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE));
2180 static void wait_while_offlining(void)
2183 #endif /* CONFIG_MEMORY_HOTREMOVE */
2187 * This all compiles without CONFIG_SYSFS, but is a waste of space.
2190 #define KSM_ATTR_RO(_name) \
2191 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2192 #define KSM_ATTR(_name) \
2193 static struct kobj_attribute _name##_attr = \
2194 __ATTR(_name, 0644, _name##_show, _name##_store)
2196 static ssize_t sleep_millisecs_show(struct kobject *kobj,
2197 struct kobj_attribute *attr, char *buf)
2199 return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
2202 static ssize_t sleep_millisecs_store(struct kobject *kobj,
2203 struct kobj_attribute *attr,
2204 const char *buf, size_t count)
2206 unsigned long msecs;
2209 err = strict_strtoul(buf, 10, &msecs);
2210 if (err || msecs > UINT_MAX)
2213 ksm_thread_sleep_millisecs = msecs;
2217 KSM_ATTR(sleep_millisecs);
2219 static ssize_t pages_to_scan_show(struct kobject *kobj,
2220 struct kobj_attribute *attr, char *buf)
2222 return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
2225 static ssize_t pages_to_scan_store(struct kobject *kobj,
2226 struct kobj_attribute *attr,
2227 const char *buf, size_t count)
2230 unsigned long nr_pages;
2232 err = strict_strtoul(buf, 10, &nr_pages);
2233 if (err || nr_pages > UINT_MAX)
2236 ksm_thread_pages_to_scan = nr_pages;
2240 KSM_ATTR(pages_to_scan);
2242 static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
2245 return sprintf(buf, "%lu\n", ksm_run);
2248 static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
2249 const char *buf, size_t count)
2252 unsigned long flags;
2254 err = strict_strtoul(buf, 10, &flags);
2255 if (err || flags > UINT_MAX)
2257 if (flags > KSM_RUN_UNMERGE)
2261 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
2262 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
2263 * breaking COW to free the pages_shared (but leaves mm_slots
2264 * on the list for when ksmd may be set running again).
2267 mutex_lock(&ksm_thread_mutex);
2268 wait_while_offlining();
2269 if (ksm_run != flags) {
2271 if (flags & KSM_RUN_UNMERGE) {
2272 set_current_oom_origin();
2273 err = unmerge_and_remove_all_rmap_items();
2274 clear_current_oom_origin();
2276 ksm_run = KSM_RUN_STOP;
2281 mutex_unlock(&ksm_thread_mutex);
2283 if (flags & KSM_RUN_MERGE)
2284 wake_up_interruptible(&ksm_thread_wait);
2291 static ssize_t merge_across_nodes_show(struct kobject *kobj,
2292 struct kobj_attribute *attr, char *buf)
2294 return sprintf(buf, "%u\n", ksm_merge_across_nodes);
2297 static ssize_t merge_across_nodes_store(struct kobject *kobj,
2298 struct kobj_attribute *attr,
2299 const char *buf, size_t count)
2304 err = kstrtoul(buf, 10, &knob);
2310 mutex_lock(&ksm_thread_mutex);
2311 wait_while_offlining();
2312 if (ksm_merge_across_nodes != knob) {
2313 if (ksm_pages_shared || remove_all_stable_nodes())
2315 else if (root_stable_tree == one_stable_tree) {
2316 struct rb_root *buf;
2318 * This is the first time that we switch away from the
2319 * default of merging across nodes: must now allocate
2320 * a buffer to hold as many roots as may be needed.
2321 * Allocate stable and unstable together:
2322 * MAXSMP NODES_SHIFT 10 will use 16kB.
2324 buf = kcalloc(nr_node_ids + nr_node_ids,
2325 sizeof(*buf), GFP_KERNEL | __GFP_ZERO);
2326 /* Let us assume that RB_ROOT is NULL is zero */
2330 root_stable_tree = buf;
2331 root_unstable_tree = buf + nr_node_ids;
2332 /* Stable tree is empty but not the unstable */
2333 root_unstable_tree[0] = one_unstable_tree[0];
2337 ksm_merge_across_nodes = knob;
2338 ksm_nr_node_ids = knob ? 1 : nr_node_ids;
2341 mutex_unlock(&ksm_thread_mutex);
2343 return err ? err : count;
2345 KSM_ATTR(merge_across_nodes);
2348 static ssize_t pages_shared_show(struct kobject *kobj,
2349 struct kobj_attribute *attr, char *buf)
2351 return sprintf(buf, "%lu\n", ksm_pages_shared);
2353 KSM_ATTR_RO(pages_shared);
2355 static ssize_t pages_sharing_show(struct kobject *kobj,
2356 struct kobj_attribute *attr, char *buf)
2358 return sprintf(buf, "%lu\n", ksm_pages_sharing);
2360 KSM_ATTR_RO(pages_sharing);
2362 static ssize_t pages_unshared_show(struct kobject *kobj,
2363 struct kobj_attribute *attr, char *buf)
2365 return sprintf(buf, "%lu\n", ksm_pages_unshared);
2367 KSM_ATTR_RO(pages_unshared);
2369 static ssize_t pages_volatile_show(struct kobject *kobj,
2370 struct kobj_attribute *attr, char *buf)
2372 long ksm_pages_volatile;
2374 ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
2375 - ksm_pages_sharing - ksm_pages_unshared;
2377 * It was not worth any locking to calculate that statistic,
2378 * but it might therefore sometimes be negative: conceal that.
2380 if (ksm_pages_volatile < 0)
2381 ksm_pages_volatile = 0;
2382 return sprintf(buf, "%ld\n", ksm_pages_volatile);
2384 KSM_ATTR_RO(pages_volatile);
2386 static ssize_t full_scans_show(struct kobject *kobj,
2387 struct kobj_attribute *attr, char *buf)
2389 return sprintf(buf, "%lu\n", ksm_scan.seqnr);
2391 KSM_ATTR_RO(full_scans);
2393 static struct attribute *ksm_attrs[] = {
2394 &sleep_millisecs_attr.attr,
2395 &pages_to_scan_attr.attr,
2397 &pages_shared_attr.attr,
2398 &pages_sharing_attr.attr,
2399 &pages_unshared_attr.attr,
2400 &pages_volatile_attr.attr,
2401 &full_scans_attr.attr,
2403 &merge_across_nodes_attr.attr,
2408 static struct attribute_group ksm_attr_group = {
2412 #endif /* CONFIG_SYSFS */
2414 static int __init ksm_init(void)
2416 struct task_struct *ksm_thread;
2419 err = ksm_slab_init();
2423 ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
2424 if (IS_ERR(ksm_thread)) {
2425 printk(KERN_ERR "ksm: creating kthread failed\n");
2426 err = PTR_ERR(ksm_thread);
2431 err = sysfs_create_group(mm_kobj, &ksm_attr_group);
2433 printk(KERN_ERR "ksm: register sysfs failed\n");
2434 kthread_stop(ksm_thread);
2438 ksm_run = KSM_RUN_MERGE; /* no way for user to start it */
2440 #endif /* CONFIG_SYSFS */
2442 #ifdef CONFIG_MEMORY_HOTREMOVE
2443 /* There is no significance to this priority 100 */
2444 hotplug_memory_notifier(ksm_memory_callback, 100);
2453 module_init(ksm_init)