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.)
91 * If the merge_across_nodes tunable is unset, then KSM maintains multiple
92 * stable trees and multiple unstable trees: one of each for each NUMA node.
96 * struct mm_slot - ksm information per mm that is being scanned
97 * @link: link to the mm_slots hash list
98 * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
99 * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
100 * @mm: the mm that this information is valid for
103 struct hlist_node link;
104 struct list_head mm_list;
105 struct rmap_item *rmap_list;
106 struct mm_struct *mm;
110 * struct ksm_scan - cursor for scanning
111 * @mm_slot: the current mm_slot we are scanning
112 * @address: the next address inside that to be scanned
113 * @rmap_list: link to the next rmap to be scanned in the rmap_list
114 * @seqnr: count of completed full scans (needed when removing unstable node)
116 * There is only the one ksm_scan instance of this cursor structure.
119 struct mm_slot *mm_slot;
120 unsigned long address;
121 struct rmap_item **rmap_list;
126 * struct stable_node - node of the stable rbtree
127 * @node: rb node of this ksm page in the stable tree
128 * @head: (overlaying parent) &migrate_nodes indicates temporarily on that list
129 * @list: linked into migrate_nodes, pending placement in the proper node tree
130 * @hlist: hlist head of rmap_items using this ksm page
131 * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid)
132 * @nid: NUMA node id of stable tree in which linked (may not match kpfn)
136 struct rb_node node; /* when node of stable tree */
137 struct { /* when listed for migration */
138 struct list_head *head;
139 struct list_head list;
142 struct hlist_head hlist;
150 * struct rmap_item - reverse mapping item for virtual addresses
151 * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
152 * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
153 * @nid: NUMA node id of unstable tree in which linked (may not match page)
154 * @mm: the memory structure this rmap_item is pointing into
155 * @address: the virtual address this rmap_item tracks (+ flags in low bits)
156 * @oldchecksum: previous checksum of the page at that virtual address
157 * @node: rb node of this rmap_item in the unstable tree
158 * @head: pointer to stable_node heading this list in the stable tree
159 * @hlist: link into hlist of rmap_items hanging off that stable_node
162 struct rmap_item *rmap_list;
164 struct anon_vma *anon_vma; /* when stable */
166 int nid; /* when node of unstable tree */
169 struct mm_struct *mm;
170 unsigned long address; /* + low bits used for flags below */
171 unsigned int oldchecksum; /* when unstable */
173 struct rb_node node; /* when node of unstable tree */
174 struct { /* when listed from stable tree */
175 struct stable_node *head;
176 struct hlist_node hlist;
181 #define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */
182 #define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */
183 #define STABLE_FLAG 0x200 /* is listed from the stable tree */
185 /* The stable and unstable tree heads */
186 static struct rb_root one_stable_tree[1] = { RB_ROOT };
187 static struct rb_root one_unstable_tree[1] = { RB_ROOT };
188 static struct rb_root *root_stable_tree = one_stable_tree;
189 static struct rb_root *root_unstable_tree = one_unstable_tree;
191 /* Recently migrated nodes of stable tree, pending proper placement */
192 static LIST_HEAD(migrate_nodes);
194 #define MM_SLOTS_HASH_BITS 10
195 static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
197 static struct mm_slot ksm_mm_head = {
198 .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
200 static struct ksm_scan ksm_scan = {
201 .mm_slot = &ksm_mm_head,
204 static struct kmem_cache *rmap_item_cache;
205 static struct kmem_cache *stable_node_cache;
206 static struct kmem_cache *mm_slot_cache;
208 /* The number of nodes in the stable tree */
209 static unsigned long ksm_pages_shared;
211 /* The number of page slots additionally sharing those nodes */
212 static unsigned long ksm_pages_sharing;
214 /* The number of nodes in the unstable tree */
215 static unsigned long ksm_pages_unshared;
217 /* The number of rmap_items in use: to calculate pages_volatile */
218 static unsigned long ksm_rmap_items;
220 /* Number of pages ksmd should scan in one batch */
221 static unsigned int ksm_thread_pages_to_scan = 100;
223 /* Milliseconds ksmd should sleep between batches */
224 static unsigned int ksm_thread_sleep_millisecs = 20;
227 /* Zeroed when merging across nodes is not allowed */
228 static unsigned int ksm_merge_across_nodes = 1;
229 static int ksm_nr_node_ids = 1;
231 #define ksm_merge_across_nodes 1U
232 #define ksm_nr_node_ids 1
235 #define KSM_RUN_STOP 0
236 #define KSM_RUN_MERGE 1
237 #define KSM_RUN_UNMERGE 2
238 #define KSM_RUN_OFFLINE 4
239 static unsigned long ksm_run = KSM_RUN_STOP;
240 static void wait_while_offlining(void);
242 static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
243 static DEFINE_MUTEX(ksm_thread_mutex);
244 static DEFINE_SPINLOCK(ksm_mmlist_lock);
246 #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
247 sizeof(struct __struct), __alignof__(struct __struct),\
250 static int __init ksm_slab_init(void)
252 rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
253 if (!rmap_item_cache)
256 stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
257 if (!stable_node_cache)
260 mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
267 kmem_cache_destroy(stable_node_cache);
269 kmem_cache_destroy(rmap_item_cache);
274 static void __init ksm_slab_free(void)
276 kmem_cache_destroy(mm_slot_cache);
277 kmem_cache_destroy(stable_node_cache);
278 kmem_cache_destroy(rmap_item_cache);
279 mm_slot_cache = NULL;
282 static inline struct rmap_item *alloc_rmap_item(void)
284 struct rmap_item *rmap_item;
286 rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL);
292 static inline void free_rmap_item(struct rmap_item *rmap_item)
295 rmap_item->mm = NULL; /* debug safety */
296 kmem_cache_free(rmap_item_cache, rmap_item);
299 static inline struct stable_node *alloc_stable_node(void)
301 return kmem_cache_alloc(stable_node_cache, GFP_KERNEL);
304 static inline void free_stable_node(struct stable_node *stable_node)
306 kmem_cache_free(stable_node_cache, stable_node);
309 static inline struct mm_slot *alloc_mm_slot(void)
311 if (!mm_slot_cache) /* initialization failed */
313 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
316 static inline void free_mm_slot(struct mm_slot *mm_slot)
318 kmem_cache_free(mm_slot_cache, mm_slot);
321 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
323 struct mm_slot *slot;
325 hash_for_each_possible(mm_slots_hash, slot, link, (unsigned long)mm)
332 static void insert_to_mm_slots_hash(struct mm_struct *mm,
333 struct mm_slot *mm_slot)
336 hash_add(mm_slots_hash, &mm_slot->link, (unsigned long)mm);
340 * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
341 * page tables after it has passed through ksm_exit() - which, if necessary,
342 * takes mmap_sem briefly to serialize against them. ksm_exit() does not set
343 * a special flag: they can just back out as soon as mm_users goes to zero.
344 * ksm_test_exit() is used throughout to make this test for exit: in some
345 * places for correctness, in some places just to avoid unnecessary work.
347 static inline bool ksm_test_exit(struct mm_struct *mm)
349 return atomic_read(&mm->mm_users) == 0;
353 * If the mm isn't the one associated with the current
354 * ksm_scan.mm_slot ksm_exit() will not down_write();up_write() and in
355 * turn the ksm_test_exit() check run inside a mm->mmap_sem critical
356 * section, will not prevent exit_mmap() to run from under us. In
357 * turn, in those cases where we could work with an "mm" that isn't
358 * guaranteed to be associated with the current ksm_scan.mm_slot,
359 * ksm_get_mm() is needed instead of the ksm_test_exit() run inside
360 * the mmap_sem. Return true if the mm_users was incremented or false
361 * if it we failed at taking the mm because it was freed from under
362 * us. If it returns 1, the caller must take care of calling mmput()
363 * after it finishes using the mm.
365 static __always_inline bool ksm_get_mm(struct mm_struct *mm)
367 return likely(atomic_inc_not_zero(&mm->mm_users));
371 * We use break_ksm to break COW on a ksm page: it's a stripped down
373 * if (get_user_pages(current, mm, addr, 1, 1, 1, &page, NULL) == 1)
376 * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
377 * in case the application has unmapped and remapped mm,addr meanwhile.
378 * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP
379 * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
381 static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
388 page = follow_page(vma, addr, FOLL_GET | FOLL_MIGRATION);
389 if (IS_ERR_OR_NULL(page))
392 ret = handle_mm_fault(vma->vm_mm, vma, addr,
395 ret = VM_FAULT_WRITE;
397 } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV | VM_FAULT_OOM)));
399 * We must loop because handle_mm_fault() may back out if there's
400 * any difficulty e.g. if pte accessed bit gets updated concurrently.
402 * VM_FAULT_WRITE is what we have been hoping for: it indicates that
403 * COW has been broken, even if the vma does not permit VM_WRITE;
404 * but note that a concurrent fault might break PageKsm for us.
406 * VM_FAULT_SIGBUS could occur if we race with truncation of the
407 * backing file, which also invalidates anonymous pages: that's
408 * okay, that truncation will have unmapped the PageKsm for us.
410 * VM_FAULT_OOM: at the time of writing (late July 2009), setting
411 * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
412 * current task has TIF_MEMDIE set, and will be OOM killed on return
413 * to user; and ksmd, having no mm, would never be chosen for that.
415 * But if the mm is in a limited mem_cgroup, then the fault may fail
416 * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
417 * even ksmd can fail in this way - though it's usually breaking ksm
418 * just to undo a merge it made a moment before, so unlikely to oom.
420 * That's a pity: we might therefore have more kernel pages allocated
421 * than we're counting as nodes in the stable tree; but ksm_do_scan
422 * will retry to break_cow on each pass, so should recover the page
423 * in due course. The important thing is to not let VM_MERGEABLE
424 * be cleared while any such pages might remain in the area.
426 return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
429 static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm,
432 struct vm_area_struct *vma;
433 vma = find_vma(mm, addr);
434 if (!vma || vma->vm_start > addr)
436 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
441 static void break_cow(struct rmap_item *rmap_item)
443 struct mm_struct *mm = rmap_item->mm;
444 unsigned long addr = rmap_item->address;
445 struct vm_area_struct *vma;
448 * It is not an accident that whenever we want to break COW
449 * to undo, we also need to drop a reference to the anon_vma.
451 put_anon_vma(rmap_item->anon_vma);
454 * The "mm" of the unstable tree rmap_item isn't necessairly
455 * associated with the current ksm_scan.mm_slot, it could be
456 * any random mm. So we need ksm_get_mm here to prevent the
457 * exit_mmap to run from under us in mmput().
462 down_read(&mm->mmap_sem);
463 vma = find_mergeable_vma(mm, addr);
465 break_ksm(vma, addr);
466 up_read(&mm->mmap_sem);
470 static struct page *page_trans_compound_anon(struct page *page)
472 if (PageTransCompound(page)) {
473 struct page *head = compound_head(page);
475 * head may actually be splitted and freed from under
476 * us but it's ok here.
484 static struct page *get_mergeable_page(struct rmap_item *rmap_item)
486 struct mm_struct *mm = rmap_item->mm;
487 unsigned long addr = rmap_item->address;
488 struct vm_area_struct *vma;
492 * The "mm" of the unstable tree rmap_item isn't necessairly
493 * associated with the current ksm_scan.mm_slot, it could be
494 * any random mm. So we need ksm_get_mm here to prevent the
495 * exit_mmap to run from under us in mmput().
500 down_read(&mm->mmap_sem);
501 vma = find_mergeable_vma(mm, addr);
505 page = follow_page(vma, addr, FOLL_GET);
506 if (IS_ERR_OR_NULL(page))
508 if (PageAnon(page) || page_trans_compound_anon(page)) {
509 flush_anon_page(vma, page, addr);
510 flush_dcache_page(page);
515 up_read(&mm->mmap_sem);
521 * This helper is used for getting right index into array of tree roots.
522 * When merge_across_nodes knob is set to 1, there are only two rb-trees for
523 * stable and unstable pages from all nodes with roots in index 0. Otherwise,
524 * every node has its own stable and unstable tree.
526 static inline int get_kpfn_nid(unsigned long kpfn)
528 return ksm_merge_across_nodes ? 0 : NUMA(pfn_to_nid(kpfn));
531 static void remove_node_from_stable_tree(struct stable_node *stable_node)
533 struct rmap_item *rmap_item;
535 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
536 if (rmap_item->hlist.next)
540 put_anon_vma(rmap_item->anon_vma);
541 rmap_item->address &= PAGE_MASK;
545 if (stable_node->head == &migrate_nodes)
546 list_del(&stable_node->list);
548 rb_erase(&stable_node->node,
549 root_stable_tree + NUMA(stable_node->nid));
550 free_stable_node(stable_node);
554 * get_ksm_page: checks if the page indicated by the stable node
555 * is still its ksm page, despite having held no reference to it.
556 * In which case we can trust the content of the page, and it
557 * returns the gotten page; but if the page has now been zapped,
558 * remove the stale node from the stable tree and return NULL.
559 * But beware, the stable node's page might be being migrated.
561 * You would expect the stable_node to hold a reference to the ksm page.
562 * But if it increments the page's count, swapping out has to wait for
563 * ksmd to come around again before it can free the page, which may take
564 * seconds or even minutes: much too unresponsive. So instead we use a
565 * "keyhole reference": access to the ksm page from the stable node peeps
566 * out through its keyhole to see if that page still holds the right key,
567 * pointing back to this stable node. This relies on freeing a PageAnon
568 * page to reset its page->mapping to NULL, and relies on no other use of
569 * a page to put something that might look like our key in page->mapping.
570 * is on its way to being freed; but it is an anomaly to bear in mind.
572 static struct page *get_ksm_page(struct stable_node *stable_node, bool lock_it)
575 void *expected_mapping;
578 expected_mapping = (void *)stable_node +
579 (PAGE_MAPPING_ANON | PAGE_MAPPING_KSM);
581 kpfn = READ_ONCE(stable_node->kpfn);
582 page = pfn_to_page(kpfn);
585 * page is computed from kpfn, so on most architectures reading
586 * page->mapping is naturally ordered after reading node->kpfn,
587 * but on Alpha we need to be more careful.
589 smp_read_barrier_depends();
590 if (READ_ONCE(page->mapping) != expected_mapping)
594 * We cannot do anything with the page while its refcount is 0.
595 * Usually 0 means free, or tail of a higher-order page: in which
596 * case this node is no longer referenced, and should be freed;
597 * however, it might mean that the page is under page_freeze_refs().
598 * The __remove_mapping() case is easy, again the node is now stale;
599 * but if page is swapcache in migrate_page_move_mapping(), it might
600 * still be our page, in which case it's essential to keep the node.
602 while (!get_page_unless_zero(page)) {
604 * Another check for page->mapping != expected_mapping would
605 * work here too. We have chosen the !PageSwapCache test to
606 * optimize the common case, when the page is or is about to
607 * be freed: PageSwapCache is cleared (under spin_lock_irq)
608 * in the freeze_refs section of __remove_mapping(); but Anon
609 * page->mapping reset to NULL later, in free_pages_prepare().
611 if (!PageSwapCache(page))
616 if (READ_ONCE(page->mapping) != expected_mapping) {
623 if (READ_ONCE(page->mapping) != expected_mapping) {
633 * We come here from above when page->mapping or !PageSwapCache
634 * suggests that the node is stale; but it might be under migration.
635 * We need smp_rmb(), matching the smp_wmb() in ksm_migrate_page(),
636 * before checking whether node->kpfn has been changed.
639 if (READ_ONCE(stable_node->kpfn) != kpfn)
641 remove_node_from_stable_tree(stable_node);
646 * Removing rmap_item from stable or unstable tree.
647 * This function will clean the information from the stable/unstable tree.
649 static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
651 if (rmap_item->address & STABLE_FLAG) {
652 struct stable_node *stable_node;
655 stable_node = rmap_item->head;
656 page = get_ksm_page(stable_node, true);
660 hlist_del(&rmap_item->hlist);
664 if (!hlist_empty(&stable_node->hlist))
669 put_anon_vma(rmap_item->anon_vma);
670 rmap_item->address &= PAGE_MASK;
672 } else if (rmap_item->address & UNSTABLE_FLAG) {
675 * Usually ksmd can and must skip the rb_erase, because
676 * root_unstable_tree was already reset to RB_ROOT.
677 * But be careful when an mm is exiting: do the rb_erase
678 * if this rmap_item was inserted by this scan, rather
679 * than left over from before.
681 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
684 rb_erase(&rmap_item->node,
685 root_unstable_tree + NUMA(rmap_item->nid));
686 ksm_pages_unshared--;
687 rmap_item->address &= PAGE_MASK;
690 cond_resched(); /* we're called from many long loops */
693 static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
694 struct rmap_item **rmap_list)
697 struct rmap_item *rmap_item = *rmap_list;
698 *rmap_list = rmap_item->rmap_list;
699 remove_rmap_item_from_tree(rmap_item);
700 free_rmap_item(rmap_item);
705 * Though it's very tempting to unmerge rmap_items from stable tree rather
706 * than check every pte of a given vma, the locking doesn't quite work for
707 * that - an rmap_item is assigned to the stable tree after inserting ksm
708 * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing
709 * rmap_items from parent to child at fork time (so as not to waste time
710 * if exit comes before the next scan reaches it).
712 * Similarly, although we'd like to remove rmap_items (so updating counts
713 * and freeing memory) when unmerging an area, it's easier to leave that
714 * to the next pass of ksmd - consider, for example, how ksmd might be
715 * in cmp_and_merge_page on one of the rmap_items we would be removing.
717 static int unmerge_ksm_pages(struct vm_area_struct *vma,
718 unsigned long start, unsigned long end)
723 for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
724 if (ksm_test_exit(vma->vm_mm))
726 if (signal_pending(current))
729 err = break_ksm(vma, addr);
736 * Only called through the sysfs control interface:
738 static int remove_stable_node(struct stable_node *stable_node)
743 page = get_ksm_page(stable_node, true);
746 * get_ksm_page did remove_node_from_stable_tree itself.
751 if (WARN_ON_ONCE(page_mapped(page))) {
753 * This should not happen: but if it does, just refuse to let
754 * merge_across_nodes be switched - there is no need to panic.
759 * The stable node did not yet appear stale to get_ksm_page(),
760 * since that allows for an unmapped ksm page to be recognized
761 * right up until it is freed; but the node is safe to remove.
762 * This page might be in a pagevec waiting to be freed,
763 * or it might be PageSwapCache (perhaps under writeback),
764 * or it might have been removed from swapcache a moment ago.
766 set_page_stable_node(page, NULL);
767 remove_node_from_stable_tree(stable_node);
776 static int remove_all_stable_nodes(void)
778 struct stable_node *stable_node;
779 struct list_head *this, *next;
783 for (nid = 0; nid < ksm_nr_node_ids; nid++) {
784 while (root_stable_tree[nid].rb_node) {
785 stable_node = rb_entry(root_stable_tree[nid].rb_node,
786 struct stable_node, node);
787 if (remove_stable_node(stable_node)) {
789 break; /* proceed to next nid */
794 list_for_each_safe(this, next, &migrate_nodes) {
795 stable_node = list_entry(this, struct stable_node, list);
796 if (remove_stable_node(stable_node))
803 static int unmerge_and_remove_all_rmap_items(void)
805 struct mm_slot *mm_slot;
806 struct mm_struct *mm;
807 struct vm_area_struct *vma;
810 spin_lock(&ksm_mmlist_lock);
811 ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
812 struct mm_slot, mm_list);
813 spin_unlock(&ksm_mmlist_lock);
815 for (mm_slot = ksm_scan.mm_slot;
816 mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
818 down_read(&mm->mmap_sem);
819 for (vma = mm->mmap; vma; vma = vma->vm_next) {
820 if (ksm_test_exit(mm))
822 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
824 err = unmerge_ksm_pages(vma,
825 vma->vm_start, vma->vm_end);
830 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
832 spin_lock(&ksm_mmlist_lock);
833 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
834 struct mm_slot, mm_list);
835 if (ksm_test_exit(mm)) {
836 hash_del(&mm_slot->link);
837 list_del(&mm_slot->mm_list);
838 spin_unlock(&ksm_mmlist_lock);
840 free_mm_slot(mm_slot);
841 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
842 up_read(&mm->mmap_sem);
845 spin_unlock(&ksm_mmlist_lock);
846 up_read(&mm->mmap_sem);
850 /* Clean up stable nodes, but don't worry if some are still busy */
851 remove_all_stable_nodes();
856 up_read(&mm->mmap_sem);
857 spin_lock(&ksm_mmlist_lock);
858 ksm_scan.mm_slot = &ksm_mm_head;
859 spin_unlock(&ksm_mmlist_lock);
862 #endif /* CONFIG_SYSFS */
864 static u32 calc_checksum(struct page *page)
867 void *addr = kmap_atomic(page);
868 checksum = jhash2(addr, PAGE_SIZE / 4, 17);
873 static int memcmp_pages(struct page *page1, struct page *page2)
878 addr1 = kmap_atomic(page1);
879 addr2 = kmap_atomic(page2);
880 ret = memcmp(addr1, addr2, PAGE_SIZE);
881 kunmap_atomic(addr2);
882 kunmap_atomic(addr1);
886 static inline int pages_identical(struct page *page1, struct page *page2)
888 return !memcmp_pages(page1, page2);
891 static int write_protect_page(struct vm_area_struct *vma, struct page *page,
894 struct mm_struct *mm = vma->vm_mm;
900 unsigned long mmun_start; /* For mmu_notifiers */
901 unsigned long mmun_end; /* For mmu_notifiers */
903 addr = page_address_in_vma(page, vma);
907 BUG_ON(PageTransCompound(page));
910 mmun_end = addr + PAGE_SIZE;
911 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
913 ptep = page_check_address(page, mm, addr, &ptl, 0);
917 if (pte_write(*ptep) || pte_dirty(*ptep)) {
920 swapped = PageSwapCache(page);
921 flush_cache_page(vma, addr, page_to_pfn(page));
923 * Ok this is tricky, when get_user_pages_fast() run it doesn't
924 * take any lock, therefore the check that we are going to make
925 * with the pagecount against the mapcount is racey and
926 * O_DIRECT can happen right after the check.
927 * So we clear the pte and flush the tlb before the check
928 * this assure us that no O_DIRECT can happen after the check
929 * or in the middle of the check.
931 entry = ptep_clear_flush_notify(vma, addr, ptep);
933 * Check that no O_DIRECT or similar I/O is in progress on the
936 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
937 set_pte_at(mm, addr, ptep, entry);
940 if (pte_dirty(entry))
941 set_page_dirty(page);
942 entry = pte_mkclean(pte_wrprotect(entry));
943 set_pte_at_notify(mm, addr, ptep, entry);
949 pte_unmap_unlock(ptep, ptl);
951 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
957 * replace_page - replace page in vma by new ksm page
958 * @vma: vma that holds the pte pointing to page
959 * @page: the page we are replacing by kpage
960 * @kpage: the ksm page we replace page by
961 * @orig_pte: the original value of the pte
963 * Returns 0 on success, -EFAULT on failure.
965 static int replace_page(struct vm_area_struct *vma, struct page *page,
966 struct page *kpage, pte_t orig_pte)
968 struct mm_struct *mm = vma->vm_mm;
974 unsigned long mmun_start; /* For mmu_notifiers */
975 unsigned long mmun_end; /* For mmu_notifiers */
977 addr = page_address_in_vma(page, vma);
981 pmd = mm_find_pmd(mm, addr);
986 mmun_end = addr + PAGE_SIZE;
987 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
989 ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
990 if (!pte_same(*ptep, orig_pte)) {
991 pte_unmap_unlock(ptep, ptl);
996 page_add_anon_rmap(kpage, vma, addr);
998 flush_cache_page(vma, addr, pte_pfn(*ptep));
999 ptep_clear_flush_notify(vma, addr, ptep);
1000 set_pte_at_notify(mm, addr, ptep, mk_pte(kpage, vma->vm_page_prot));
1002 page_remove_rmap(page);
1003 if (!page_mapped(page))
1004 try_to_free_swap(page);
1007 pte_unmap_unlock(ptep, ptl);
1010 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1015 static int page_trans_compound_anon_split(struct page *page)
1018 struct page *transhuge_head = page_trans_compound_anon(page);
1019 if (transhuge_head) {
1020 /* Get the reference on the head to split it. */
1021 if (get_page_unless_zero(transhuge_head)) {
1023 * Recheck we got the reference while the head
1024 * was still anonymous.
1026 if (PageAnon(transhuge_head))
1027 ret = split_huge_page(transhuge_head);
1030 * Retry later if split_huge_page run
1034 put_page(transhuge_head);
1036 /* Retry later if split_huge_page run from under us. */
1043 * try_to_merge_one_page - take two pages and merge them into one
1044 * @vma: the vma that holds the pte pointing to page
1045 * @page: the PageAnon page that we want to replace with kpage
1046 * @kpage: the PageKsm page that we want to map instead of page,
1047 * or NULL the first time when we want to use page as kpage.
1049 * This function returns 0 if the pages were merged, -EFAULT otherwise.
1051 static int try_to_merge_one_page(struct vm_area_struct *vma,
1052 struct page *page, struct page *kpage)
1054 pte_t orig_pte = __pte(0);
1057 if (page == kpage) /* ksm page forked */
1060 if (PageTransCompound(page) && page_trans_compound_anon_split(page))
1062 BUG_ON(PageTransCompound(page));
1063 if (!PageAnon(page))
1067 * We need the page lock to read a stable PageSwapCache in
1068 * write_protect_page(). We use trylock_page() instead of
1069 * lock_page() because we don't want to wait here - we
1070 * prefer to continue scanning and merging different pages,
1071 * then come back to this page when it is unlocked.
1073 if (!trylock_page(page))
1076 * If this anonymous page is mapped only here, its pte may need
1077 * to be write-protected. If it's mapped elsewhere, all of its
1078 * ptes are necessarily already write-protected. But in either
1079 * case, we need to lock and check page_count is not raised.
1081 if (write_protect_page(vma, page, &orig_pte) == 0) {
1084 * While we hold page lock, upgrade page from
1085 * PageAnon+anon_vma to PageKsm+NULL stable_node:
1086 * stable_tree_insert() will update stable_node.
1088 set_page_stable_node(page, NULL);
1089 mark_page_accessed(page);
1091 } else if (pages_identical(page, kpage))
1092 err = replace_page(vma, page, kpage, orig_pte);
1095 if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
1096 munlock_vma_page(page);
1097 if (!PageMlocked(kpage)) {
1100 mlock_vma_page(kpage);
1101 page = kpage; /* for final unlock */
1111 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
1112 * but no new kernel page is allocated: kpage must already be a ksm page.
1114 * This function returns 0 if the pages were merged, -EFAULT otherwise.
1116 static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
1117 struct page *page, struct page *kpage)
1119 struct mm_struct *mm = rmap_item->mm;
1120 struct vm_area_struct *vma;
1124 * The "mm" of the unstable tree rmap_item isn't necessairly
1125 * associated with the current ksm_scan.mm_slot, it could be
1126 * any random mm. So we need ksm_get_mm() here to prevent the
1127 * exit_mmap to run from under us in mmput(). Otherwise
1128 * rmap_item->anon_vma could point to an anon_vma that has
1129 * already been freed (i.e. get_anon_vma() below would run too
1132 if (!ksm_get_mm(mm))
1135 down_read(&mm->mmap_sem);
1136 vma = find_mergeable_vma(mm, rmap_item->address);
1140 err = try_to_merge_one_page(vma, page, kpage);
1144 /* Unstable nid is in union with stable anon_vma: remove first */
1145 remove_rmap_item_from_tree(rmap_item);
1147 /* Must get reference to anon_vma while still holding mmap_sem */
1148 rmap_item->anon_vma = vma->anon_vma;
1149 get_anon_vma(vma->anon_vma);
1151 up_read(&mm->mmap_sem);
1157 * try_to_merge_two_pages - take two identical pages and prepare them
1158 * to be merged into one page.
1160 * This function returns the kpage if we successfully merged two identical
1161 * pages into one ksm page, NULL otherwise.
1163 * Note that this function upgrades page to ksm page: if one of the pages
1164 * is already a ksm page, try_to_merge_with_ksm_page should be used.
1166 static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
1168 struct rmap_item *tree_rmap_item,
1169 struct page *tree_page)
1173 err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
1175 err = try_to_merge_with_ksm_page(tree_rmap_item,
1178 * If that fails, we have a ksm page with only one pte
1179 * pointing to it: so break it.
1182 break_cow(rmap_item);
1184 return err ? NULL : page;
1188 * stable_tree_search - search for page inside the stable tree
1190 * This function checks if there is a page inside the stable tree
1191 * with identical content to the page that we are scanning right now.
1193 * This function returns the stable tree node of identical content if found,
1196 static struct page *stable_tree_search(struct page *page)
1199 struct rb_root *root;
1200 struct rb_node **new;
1201 struct rb_node *parent;
1202 struct stable_node *stable_node;
1203 struct stable_node *page_node;
1205 page_node = page_stable_node(page);
1206 if (page_node && page_node->head != &migrate_nodes) {
1207 /* ksm page forked */
1212 nid = get_kpfn_nid(page_to_pfn(page));
1213 root = root_stable_tree + nid;
1215 new = &root->rb_node;
1219 struct page *tree_page;
1223 stable_node = rb_entry(*new, struct stable_node, node);
1224 tree_page = get_ksm_page(stable_node, false);
1227 * If we walked over a stale stable_node,
1228 * get_ksm_page() will call rb_erase() and it
1229 * may rebalance the tree from under us. So
1230 * restart the search from scratch. Returning
1231 * NULL would be safe too, but we'd generate
1232 * false negative insertions just because some
1233 * stable_node was stale which would waste CPU
1234 * by doing the preparation work twice at the
1239 ret = memcmp_pages(page, tree_page);
1240 put_page(tree_page);
1244 new = &parent->rb_left;
1246 new = &parent->rb_right;
1249 * Lock and unlock the stable_node's page (which
1250 * might already have been migrated) so that page
1251 * migration is sure to notice its raised count.
1252 * It would be more elegant to return stable_node
1253 * than kpage, but that involves more changes.
1255 tree_page = get_ksm_page(stable_node, true);
1257 unlock_page(tree_page);
1258 if (get_kpfn_nid(stable_node->kpfn) !=
1259 NUMA(stable_node->nid)) {
1260 put_page(tree_page);
1266 * There is now a place for page_node, but the tree may
1267 * have been rebalanced, so re-evaluate parent and new.
1278 list_del(&page_node->list);
1279 DO_NUMA(page_node->nid = nid);
1280 rb_link_node(&page_node->node, parent, new);
1281 rb_insert_color(&page_node->node, root);
1287 list_del(&page_node->list);
1288 DO_NUMA(page_node->nid = nid);
1289 rb_replace_node(&stable_node->node, &page_node->node, root);
1292 rb_erase(&stable_node->node, root);
1295 stable_node->head = &migrate_nodes;
1296 list_add(&stable_node->list, stable_node->head);
1301 * stable_tree_insert - insert stable tree node pointing to new ksm page
1302 * into the stable tree.
1304 * This function returns the stable tree node just allocated on success,
1307 static struct stable_node *stable_tree_insert(struct page *kpage)
1311 struct rb_root *root;
1312 struct rb_node **new;
1313 struct rb_node *parent;
1314 struct stable_node *stable_node;
1316 kpfn = page_to_pfn(kpage);
1317 nid = get_kpfn_nid(kpfn);
1318 root = root_stable_tree + nid;
1321 new = &root->rb_node;
1324 struct page *tree_page;
1328 stable_node = rb_entry(*new, struct stable_node, node);
1329 tree_page = get_ksm_page(stable_node, false);
1332 * If we walked over a stale stable_node,
1333 * get_ksm_page() will call rb_erase() and it
1334 * may rebalance the tree from under us. So
1335 * restart the search from scratch. Returning
1336 * NULL would be safe too, but we'd generate
1337 * false negative insertions just because some
1338 * stable_node was stale which would waste CPU
1339 * by doing the preparation work twice at the
1344 ret = memcmp_pages(kpage, tree_page);
1345 put_page(tree_page);
1349 new = &parent->rb_left;
1351 new = &parent->rb_right;
1354 * It is not a bug that stable_tree_search() didn't
1355 * find this node: because at that time our page was
1356 * not yet write-protected, so may have changed since.
1362 stable_node = alloc_stable_node();
1366 INIT_HLIST_HEAD(&stable_node->hlist);
1367 stable_node->kpfn = kpfn;
1368 set_page_stable_node(kpage, stable_node);
1369 DO_NUMA(stable_node->nid = nid);
1370 rb_link_node(&stable_node->node, parent, new);
1371 rb_insert_color(&stable_node->node, root);
1377 * unstable_tree_search_insert - search for identical page,
1378 * else insert rmap_item into the unstable tree.
1380 * This function searches for a page in the unstable tree identical to the
1381 * page currently being scanned; and if no identical page is found in the
1382 * tree, we insert rmap_item as a new object into the unstable tree.
1384 * This function returns pointer to rmap_item found to be identical
1385 * to the currently scanned page, NULL otherwise.
1387 * This function does both searching and inserting, because they share
1388 * the same walking algorithm in an rbtree.
1391 struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1393 struct page **tree_pagep)
1395 struct rb_node **new;
1396 struct rb_root *root;
1397 struct rb_node *parent = NULL;
1400 nid = get_kpfn_nid(page_to_pfn(page));
1401 root = root_unstable_tree + nid;
1402 new = &root->rb_node;
1405 struct rmap_item *tree_rmap_item;
1406 struct page *tree_page;
1410 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
1411 tree_page = get_mergeable_page(tree_rmap_item);
1416 * Don't substitute a ksm page for a forked page.
1418 if (page == tree_page) {
1419 put_page(tree_page);
1423 ret = memcmp_pages(page, tree_page);
1427 put_page(tree_page);
1428 new = &parent->rb_left;
1429 } else if (ret > 0) {
1430 put_page(tree_page);
1431 new = &parent->rb_right;
1432 } else if (!ksm_merge_across_nodes &&
1433 page_to_nid(tree_page) != nid) {
1435 * If tree_page has been migrated to another NUMA node,
1436 * it will be flushed out and put in the right unstable
1437 * tree next time: only merge with it when across_nodes.
1439 put_page(tree_page);
1442 *tree_pagep = tree_page;
1443 return tree_rmap_item;
1447 rmap_item->address |= UNSTABLE_FLAG;
1448 rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1449 DO_NUMA(rmap_item->nid = nid);
1450 rb_link_node(&rmap_item->node, parent, new);
1451 rb_insert_color(&rmap_item->node, root);
1453 ksm_pages_unshared++;
1458 * stable_tree_append - add another rmap_item to the linked list of
1459 * rmap_items hanging off a given node of the stable tree, all sharing
1460 * the same ksm page.
1462 static void stable_tree_append(struct rmap_item *rmap_item,
1463 struct stable_node *stable_node)
1465 rmap_item->head = stable_node;
1466 rmap_item->address |= STABLE_FLAG;
1467 hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
1469 if (rmap_item->hlist.next)
1470 ksm_pages_sharing++;
1476 * cmp_and_merge_page - first see if page can be merged into the stable tree;
1477 * if not, compare checksum to previous and if it's the same, see if page can
1478 * be inserted into the unstable tree, or merged with a page already there and
1479 * both transferred to the stable tree.
1481 * @page: the page that we are searching identical page to.
1482 * @rmap_item: the reverse mapping into the virtual address of this page
1484 static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1486 struct rmap_item *tree_rmap_item;
1487 struct page *tree_page = NULL;
1488 struct stable_node *stable_node;
1490 unsigned int checksum;
1493 stable_node = page_stable_node(page);
1495 if (stable_node->head != &migrate_nodes &&
1496 get_kpfn_nid(stable_node->kpfn) != NUMA(stable_node->nid)) {
1497 rb_erase(&stable_node->node,
1498 root_stable_tree + NUMA(stable_node->nid));
1499 stable_node->head = &migrate_nodes;
1500 list_add(&stable_node->list, stable_node->head);
1502 if (stable_node->head != &migrate_nodes &&
1503 rmap_item->head == stable_node)
1507 /* We first start with searching the page inside the stable tree */
1508 kpage = stable_tree_search(page);
1509 if (kpage == page && rmap_item->head == stable_node) {
1514 remove_rmap_item_from_tree(rmap_item);
1517 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
1520 * The page was successfully merged:
1521 * add its rmap_item to the stable tree.
1524 stable_tree_append(rmap_item, page_stable_node(kpage));
1532 * If the hash value of the page has changed from the last time
1533 * we calculated it, this page is changing frequently: therefore we
1534 * don't want to insert it in the unstable tree, and we don't want
1535 * to waste our time searching for something identical to it there.
1537 checksum = calc_checksum(page);
1538 if (rmap_item->oldchecksum != checksum) {
1539 rmap_item->oldchecksum = checksum;
1544 unstable_tree_search_insert(rmap_item, page, &tree_page);
1545 if (tree_rmap_item) {
1546 kpage = try_to_merge_two_pages(rmap_item, page,
1547 tree_rmap_item, tree_page);
1548 put_page(tree_page);
1551 * The pages were successfully merged: insert new
1552 * node in the stable tree and add both rmap_items.
1555 stable_node = stable_tree_insert(kpage);
1557 stable_tree_append(tree_rmap_item, stable_node);
1558 stable_tree_append(rmap_item, stable_node);
1563 * If we fail to insert the page into the stable tree,
1564 * we will have 2 virtual addresses that are pointing
1565 * to a ksm page left outside the stable tree,
1566 * in which case we need to break_cow on both.
1569 break_cow(tree_rmap_item);
1570 break_cow(rmap_item);
1576 static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
1577 struct rmap_item **rmap_list,
1580 struct rmap_item *rmap_item;
1582 while (*rmap_list) {
1583 rmap_item = *rmap_list;
1584 if ((rmap_item->address & PAGE_MASK) == addr)
1586 if (rmap_item->address > addr)
1588 *rmap_list = rmap_item->rmap_list;
1589 remove_rmap_item_from_tree(rmap_item);
1590 free_rmap_item(rmap_item);
1593 rmap_item = alloc_rmap_item();
1595 /* It has already been zeroed */
1596 rmap_item->mm = mm_slot->mm;
1597 rmap_item->address = addr;
1598 rmap_item->rmap_list = *rmap_list;
1599 *rmap_list = rmap_item;
1604 static struct rmap_item *scan_get_next_rmap_item(struct page **page)
1606 struct mm_struct *mm;
1607 struct mm_slot *slot;
1608 struct vm_area_struct *vma;
1609 struct rmap_item *rmap_item;
1612 if (list_empty(&ksm_mm_head.mm_list))
1615 slot = ksm_scan.mm_slot;
1616 if (slot == &ksm_mm_head) {
1618 * A number of pages can hang around indefinitely on per-cpu
1619 * pagevecs, raised page count preventing write_protect_page
1620 * from merging them. Though it doesn't really matter much,
1621 * it is puzzling to see some stuck in pages_volatile until
1622 * other activity jostles them out, and they also prevented
1623 * LTP's KSM test from succeeding deterministically; so drain
1624 * them here (here rather than on entry to ksm_do_scan(),
1625 * so we don't IPI too often when pages_to_scan is set low).
1627 lru_add_drain_all();
1630 * Whereas stale stable_nodes on the stable_tree itself
1631 * get pruned in the regular course of stable_tree_search(),
1632 * those moved out to the migrate_nodes list can accumulate:
1633 * so prune them once before each full scan.
1635 if (!ksm_merge_across_nodes) {
1636 struct stable_node *stable_node;
1637 struct list_head *this, *next;
1640 list_for_each_safe(this, next, &migrate_nodes) {
1641 stable_node = list_entry(this,
1642 struct stable_node, list);
1643 page = get_ksm_page(stable_node, false);
1650 for (nid = 0; nid < ksm_nr_node_ids; nid++)
1651 root_unstable_tree[nid] = RB_ROOT;
1653 spin_lock(&ksm_mmlist_lock);
1654 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1655 ksm_scan.mm_slot = slot;
1656 spin_unlock(&ksm_mmlist_lock);
1658 * Although we tested list_empty() above, a racing __ksm_exit
1659 * of the last mm on the list may have removed it since then.
1661 if (slot == &ksm_mm_head)
1664 ksm_scan.address = 0;
1665 ksm_scan.rmap_list = &slot->rmap_list;
1669 down_read(&mm->mmap_sem);
1670 if (ksm_test_exit(mm))
1673 vma = find_vma(mm, ksm_scan.address);
1675 for (; vma; vma = vma->vm_next) {
1676 if (!(vma->vm_flags & VM_MERGEABLE))
1678 if (ksm_scan.address < vma->vm_start)
1679 ksm_scan.address = vma->vm_start;
1681 ksm_scan.address = vma->vm_end;
1683 while (ksm_scan.address < vma->vm_end) {
1684 if (ksm_test_exit(mm))
1686 *page = follow_page(vma, ksm_scan.address, FOLL_GET);
1687 if (IS_ERR_OR_NULL(*page)) {
1688 ksm_scan.address += PAGE_SIZE;
1692 if (PageAnon(*page) ||
1693 page_trans_compound_anon(*page)) {
1694 flush_anon_page(vma, *page, ksm_scan.address);
1695 flush_dcache_page(*page);
1696 rmap_item = get_next_rmap_item(slot,
1697 ksm_scan.rmap_list, ksm_scan.address);
1699 ksm_scan.rmap_list =
1700 &rmap_item->rmap_list;
1701 ksm_scan.address += PAGE_SIZE;
1704 up_read(&mm->mmap_sem);
1708 ksm_scan.address += PAGE_SIZE;
1713 if (ksm_test_exit(mm)) {
1714 ksm_scan.address = 0;
1715 ksm_scan.rmap_list = &slot->rmap_list;
1718 * Nuke all the rmap_items that are above this current rmap:
1719 * because there were no VM_MERGEABLE vmas with such addresses.
1721 remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
1723 spin_lock(&ksm_mmlist_lock);
1724 ksm_scan.mm_slot = list_entry(slot->mm_list.next,
1725 struct mm_slot, mm_list);
1726 if (ksm_scan.address == 0) {
1728 * We've completed a full scan of all vmas, holding mmap_sem
1729 * throughout, and found no VM_MERGEABLE: so do the same as
1730 * __ksm_exit does to remove this mm from all our lists now.
1731 * This applies either when cleaning up after __ksm_exit
1732 * (but beware: we can reach here even before __ksm_exit),
1733 * or when all VM_MERGEABLE areas have been unmapped (and
1734 * mmap_sem then protects against race with MADV_MERGEABLE).
1736 hash_del(&slot->link);
1737 list_del(&slot->mm_list);
1738 spin_unlock(&ksm_mmlist_lock);
1741 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1742 up_read(&mm->mmap_sem);
1745 spin_unlock(&ksm_mmlist_lock);
1746 up_read(&mm->mmap_sem);
1749 /* Repeat until we've completed scanning the whole list */
1750 slot = ksm_scan.mm_slot;
1751 if (slot != &ksm_mm_head)
1759 * ksm_do_scan - the ksm scanner main worker function.
1760 * @scan_npages - number of pages we want to scan before we return.
1762 static void ksm_do_scan(unsigned int scan_npages)
1764 struct rmap_item *rmap_item;
1765 struct page *uninitialized_var(page);
1767 while (scan_npages-- && likely(!freezing(current))) {
1769 rmap_item = scan_get_next_rmap_item(&page);
1772 cmp_and_merge_page(page, rmap_item);
1777 static int ksmd_should_run(void)
1779 return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
1782 static int ksm_scan_thread(void *nothing)
1785 set_user_nice(current, 5);
1787 while (!kthread_should_stop()) {
1788 mutex_lock(&ksm_thread_mutex);
1789 wait_while_offlining();
1790 if (ksmd_should_run())
1791 ksm_do_scan(ksm_thread_pages_to_scan);
1792 mutex_unlock(&ksm_thread_mutex);
1796 if (ksmd_should_run()) {
1797 schedule_timeout_interruptible(
1798 msecs_to_jiffies(ksm_thread_sleep_millisecs));
1800 wait_event_freezable(ksm_thread_wait,
1801 ksmd_should_run() || kthread_should_stop());
1807 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1808 unsigned long end, int advice, unsigned long *vm_flags)
1810 struct mm_struct *mm = vma->vm_mm;
1814 case MADV_MERGEABLE:
1816 * Be somewhat over-protective for now!
1818 if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE |
1819 VM_PFNMAP | VM_IO | VM_DONTEXPAND |
1820 VM_HUGETLB | VM_MIXEDMAP))
1821 return 0; /* just ignore the advice */
1824 if (*vm_flags & VM_SAO)
1828 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
1829 err = __ksm_enter(mm);
1834 *vm_flags |= VM_MERGEABLE;
1837 case MADV_UNMERGEABLE:
1838 if (!(*vm_flags & VM_MERGEABLE))
1839 return 0; /* just ignore the advice */
1841 if (vma->anon_vma) {
1842 err = unmerge_ksm_pages(vma, start, end);
1847 *vm_flags &= ~VM_MERGEABLE;
1854 int __ksm_enter(struct mm_struct *mm)
1856 struct mm_slot *mm_slot;
1859 mm_slot = alloc_mm_slot();
1863 /* Check ksm_run too? Would need tighter locking */
1864 needs_wakeup = list_empty(&ksm_mm_head.mm_list);
1866 spin_lock(&ksm_mmlist_lock);
1867 insert_to_mm_slots_hash(mm, mm_slot);
1869 * When KSM_RUN_MERGE (or KSM_RUN_STOP),
1870 * insert just behind the scanning cursor, to let the area settle
1871 * down a little; when fork is followed by immediate exec, we don't
1872 * want ksmd to waste time setting up and tearing down an rmap_list.
1874 * But when KSM_RUN_UNMERGE, it's important to insert ahead of its
1875 * scanning cursor, otherwise KSM pages in newly forked mms will be
1876 * missed: then we might as well insert at the end of the list.
1878 if (ksm_run & KSM_RUN_UNMERGE)
1879 list_add_tail(&mm_slot->mm_list, &ksm_mm_head.mm_list);
1881 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
1882 spin_unlock(&ksm_mmlist_lock);
1884 set_bit(MMF_VM_MERGEABLE, &mm->flags);
1885 atomic_inc(&mm->mm_count);
1888 wake_up_interruptible(&ksm_thread_wait);
1893 void __ksm_exit(struct mm_struct *mm)
1895 struct mm_slot *mm_slot;
1896 int easy_to_free = 0;
1899 * This process is exiting: if it's straightforward (as is the
1900 * case when ksmd was never running), free mm_slot immediately.
1901 * But if it's at the cursor or has rmap_items linked to it, use
1902 * mmap_sem to synchronize with any break_cows before pagetables
1903 * are freed, and leave the mm_slot on the list for ksmd to free.
1904 * Beware: ksm may already have noticed it exiting and freed the slot.
1907 spin_lock(&ksm_mmlist_lock);
1908 mm_slot = get_mm_slot(mm);
1909 if (mm_slot && ksm_scan.mm_slot != mm_slot) {
1910 if (!mm_slot->rmap_list) {
1911 hash_del(&mm_slot->link);
1912 list_del(&mm_slot->mm_list);
1915 list_move(&mm_slot->mm_list,
1916 &ksm_scan.mm_slot->mm_list);
1919 spin_unlock(&ksm_mmlist_lock);
1922 free_mm_slot(mm_slot);
1923 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1925 } else if (mm_slot) {
1926 down_write(&mm->mmap_sem);
1927 up_write(&mm->mmap_sem);
1931 struct page *ksm_might_need_to_copy(struct page *page,
1932 struct vm_area_struct *vma, unsigned long address)
1934 struct anon_vma *anon_vma = page_anon_vma(page);
1935 struct page *new_page;
1937 if (PageKsm(page)) {
1938 if (page_stable_node(page) &&
1939 !(ksm_run & KSM_RUN_UNMERGE))
1940 return page; /* no need to copy it */
1941 } else if (!anon_vma) {
1942 return page; /* no need to copy it */
1943 } else if (anon_vma->root == vma->anon_vma->root &&
1944 page->index == linear_page_index(vma, address)) {
1945 return page; /* still no need to copy it */
1947 if (!PageUptodate(page))
1948 return page; /* let do_swap_page report the error */
1950 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1952 copy_user_highpage(new_page, page, address, vma);
1954 SetPageDirty(new_page);
1955 __SetPageUptodate(new_page);
1956 __set_page_locked(new_page);
1962 int rmap_walk_ksm(struct page *page, struct rmap_walk_control *rwc)
1964 struct stable_node *stable_node;
1965 struct rmap_item *rmap_item;
1966 int ret = SWAP_AGAIN;
1967 int search_new_forks = 0;
1969 VM_BUG_ON_PAGE(!PageKsm(page), page);
1972 * Rely on the page lock to protect against concurrent modifications
1973 * to that page's node of the stable tree.
1975 VM_BUG_ON_PAGE(!PageLocked(page), page);
1977 stable_node = page_stable_node(page);
1981 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
1982 struct anon_vma *anon_vma = rmap_item->anon_vma;
1983 struct anon_vma_chain *vmac;
1984 struct vm_area_struct *vma;
1987 anon_vma_lock_read(anon_vma);
1988 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
1992 if (rmap_item->address < vma->vm_start ||
1993 rmap_item->address >= vma->vm_end)
1996 * Initially we examine only the vma which covers this
1997 * rmap_item; but later, if there is still work to do,
1998 * we examine covering vmas in other mms: in case they
1999 * were forked from the original since ksmd passed.
2001 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
2004 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2007 ret = rwc->rmap_one(page, vma,
2008 rmap_item->address, rwc->arg);
2009 if (ret != SWAP_AGAIN) {
2010 anon_vma_unlock_read(anon_vma);
2013 if (rwc->done && rwc->done(page)) {
2014 anon_vma_unlock_read(anon_vma);
2018 anon_vma_unlock_read(anon_vma);
2020 if (!search_new_forks++)
2026 #ifdef CONFIG_MIGRATION
2027 void ksm_migrate_page(struct page *newpage, struct page *oldpage)
2029 struct stable_node *stable_node;
2031 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
2032 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
2033 VM_BUG_ON_PAGE(newpage->mapping != oldpage->mapping, newpage);
2035 stable_node = page_stable_node(newpage);
2037 VM_BUG_ON_PAGE(stable_node->kpfn != page_to_pfn(oldpage), oldpage);
2038 stable_node->kpfn = page_to_pfn(newpage);
2040 * newpage->mapping was set in advance; now we need smp_wmb()
2041 * to make sure that the new stable_node->kpfn is visible
2042 * to get_ksm_page() before it can see that oldpage->mapping
2043 * has gone stale (or that PageSwapCache has been cleared).
2046 set_page_stable_node(oldpage, NULL);
2049 #endif /* CONFIG_MIGRATION */
2051 #ifdef CONFIG_MEMORY_HOTREMOVE
2052 static void wait_while_offlining(void)
2054 while (ksm_run & KSM_RUN_OFFLINE) {
2055 mutex_unlock(&ksm_thread_mutex);
2056 wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE),
2057 TASK_UNINTERRUPTIBLE);
2058 mutex_lock(&ksm_thread_mutex);
2062 static void ksm_check_stable_tree(unsigned long start_pfn,
2063 unsigned long end_pfn)
2065 struct stable_node *stable_node;
2066 struct list_head *this, *next;
2067 struct rb_node *node;
2070 for (nid = 0; nid < ksm_nr_node_ids; nid++) {
2071 node = rb_first(root_stable_tree + nid);
2073 stable_node = rb_entry(node, struct stable_node, node);
2074 if (stable_node->kpfn >= start_pfn &&
2075 stable_node->kpfn < end_pfn) {
2077 * Don't get_ksm_page, page has already gone:
2078 * which is why we keep kpfn instead of page*
2080 remove_node_from_stable_tree(stable_node);
2081 node = rb_first(root_stable_tree + nid);
2083 node = rb_next(node);
2087 list_for_each_safe(this, next, &migrate_nodes) {
2088 stable_node = list_entry(this, struct stable_node, list);
2089 if (stable_node->kpfn >= start_pfn &&
2090 stable_node->kpfn < end_pfn)
2091 remove_node_from_stable_tree(stable_node);
2096 static int ksm_memory_callback(struct notifier_block *self,
2097 unsigned long action, void *arg)
2099 struct memory_notify *mn = arg;
2102 case MEM_GOING_OFFLINE:
2104 * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items()
2105 * and remove_all_stable_nodes() while memory is going offline:
2106 * it is unsafe for them to touch the stable tree at this time.
2107 * But unmerge_ksm_pages(), rmap lookups and other entry points
2108 * which do not need the ksm_thread_mutex are all safe.
2110 mutex_lock(&ksm_thread_mutex);
2111 ksm_run |= KSM_RUN_OFFLINE;
2112 mutex_unlock(&ksm_thread_mutex);
2117 * Most of the work is done by page migration; but there might
2118 * be a few stable_nodes left over, still pointing to struct
2119 * pages which have been offlined: prune those from the tree,
2120 * otherwise get_ksm_page() might later try to access a
2121 * non-existent struct page.
2123 ksm_check_stable_tree(mn->start_pfn,
2124 mn->start_pfn + mn->nr_pages);
2127 case MEM_CANCEL_OFFLINE:
2128 mutex_lock(&ksm_thread_mutex);
2129 ksm_run &= ~KSM_RUN_OFFLINE;
2130 mutex_unlock(&ksm_thread_mutex);
2132 smp_mb(); /* wake_up_bit advises this */
2133 wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE));
2139 static void wait_while_offlining(void)
2142 #endif /* CONFIG_MEMORY_HOTREMOVE */
2146 * This all compiles without CONFIG_SYSFS, but is a waste of space.
2149 #define KSM_ATTR_RO(_name) \
2150 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2151 #define KSM_ATTR(_name) \
2152 static struct kobj_attribute _name##_attr = \
2153 __ATTR(_name, 0644, _name##_show, _name##_store)
2155 static ssize_t sleep_millisecs_show(struct kobject *kobj,
2156 struct kobj_attribute *attr, char *buf)
2158 return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
2161 static ssize_t sleep_millisecs_store(struct kobject *kobj,
2162 struct kobj_attribute *attr,
2163 const char *buf, size_t count)
2165 unsigned long msecs;
2168 err = kstrtoul(buf, 10, &msecs);
2169 if (err || msecs > UINT_MAX)
2172 ksm_thread_sleep_millisecs = msecs;
2176 KSM_ATTR(sleep_millisecs);
2178 static ssize_t pages_to_scan_show(struct kobject *kobj,
2179 struct kobj_attribute *attr, char *buf)
2181 return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
2184 static ssize_t pages_to_scan_store(struct kobject *kobj,
2185 struct kobj_attribute *attr,
2186 const char *buf, size_t count)
2189 unsigned long nr_pages;
2191 err = kstrtoul(buf, 10, &nr_pages);
2192 if (err || nr_pages > UINT_MAX)
2195 ksm_thread_pages_to_scan = nr_pages;
2199 KSM_ATTR(pages_to_scan);
2201 static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
2204 return sprintf(buf, "%lu\n", ksm_run);
2207 static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
2208 const char *buf, size_t count)
2211 unsigned long flags;
2213 err = kstrtoul(buf, 10, &flags);
2214 if (err || flags > UINT_MAX)
2216 if (flags > KSM_RUN_UNMERGE)
2220 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
2221 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
2222 * breaking COW to free the pages_shared (but leaves mm_slots
2223 * on the list for when ksmd may be set running again).
2226 mutex_lock(&ksm_thread_mutex);
2227 wait_while_offlining();
2228 if (ksm_run != flags) {
2230 if (flags & KSM_RUN_UNMERGE) {
2231 set_current_oom_origin();
2232 err = unmerge_and_remove_all_rmap_items();
2233 clear_current_oom_origin();
2235 ksm_run = KSM_RUN_STOP;
2240 mutex_unlock(&ksm_thread_mutex);
2242 if (flags & KSM_RUN_MERGE)
2243 wake_up_interruptible(&ksm_thread_wait);
2250 static ssize_t merge_across_nodes_show(struct kobject *kobj,
2251 struct kobj_attribute *attr, char *buf)
2253 return sprintf(buf, "%u\n", ksm_merge_across_nodes);
2256 static ssize_t merge_across_nodes_store(struct kobject *kobj,
2257 struct kobj_attribute *attr,
2258 const char *buf, size_t count)
2263 err = kstrtoul(buf, 10, &knob);
2269 mutex_lock(&ksm_thread_mutex);
2270 wait_while_offlining();
2271 if (ksm_merge_across_nodes != knob) {
2272 if (ksm_pages_shared || remove_all_stable_nodes())
2274 else if (root_stable_tree == one_stable_tree) {
2275 struct rb_root *buf;
2277 * This is the first time that we switch away from the
2278 * default of merging across nodes: must now allocate
2279 * a buffer to hold as many roots as may be needed.
2280 * Allocate stable and unstable together:
2281 * MAXSMP NODES_SHIFT 10 will use 16kB.
2283 buf = kcalloc(nr_node_ids + nr_node_ids, sizeof(*buf),
2285 /* Let us assume that RB_ROOT is NULL is zero */
2289 root_stable_tree = buf;
2290 root_unstable_tree = buf + nr_node_ids;
2291 /* Stable tree is empty but not the unstable */
2292 root_unstable_tree[0] = one_unstable_tree[0];
2296 ksm_merge_across_nodes = knob;
2297 ksm_nr_node_ids = knob ? 1 : nr_node_ids;
2300 mutex_unlock(&ksm_thread_mutex);
2302 return err ? err : count;
2304 KSM_ATTR(merge_across_nodes);
2307 static ssize_t pages_shared_show(struct kobject *kobj,
2308 struct kobj_attribute *attr, char *buf)
2310 return sprintf(buf, "%lu\n", ksm_pages_shared);
2312 KSM_ATTR_RO(pages_shared);
2314 static ssize_t pages_sharing_show(struct kobject *kobj,
2315 struct kobj_attribute *attr, char *buf)
2317 return sprintf(buf, "%lu\n", ksm_pages_sharing);
2319 KSM_ATTR_RO(pages_sharing);
2321 static ssize_t pages_unshared_show(struct kobject *kobj,
2322 struct kobj_attribute *attr, char *buf)
2324 return sprintf(buf, "%lu\n", ksm_pages_unshared);
2326 KSM_ATTR_RO(pages_unshared);
2328 static ssize_t pages_volatile_show(struct kobject *kobj,
2329 struct kobj_attribute *attr, char *buf)
2331 long ksm_pages_volatile;
2333 ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
2334 - ksm_pages_sharing - ksm_pages_unshared;
2336 * It was not worth any locking to calculate that statistic,
2337 * but it might therefore sometimes be negative: conceal that.
2339 if (ksm_pages_volatile < 0)
2340 ksm_pages_volatile = 0;
2341 return sprintf(buf, "%ld\n", ksm_pages_volatile);
2343 KSM_ATTR_RO(pages_volatile);
2345 static ssize_t full_scans_show(struct kobject *kobj,
2346 struct kobj_attribute *attr, char *buf)
2348 return sprintf(buf, "%lu\n", ksm_scan.seqnr);
2350 KSM_ATTR_RO(full_scans);
2352 static struct attribute *ksm_attrs[] = {
2353 &sleep_millisecs_attr.attr,
2354 &pages_to_scan_attr.attr,
2356 &pages_shared_attr.attr,
2357 &pages_sharing_attr.attr,
2358 &pages_unshared_attr.attr,
2359 &pages_volatile_attr.attr,
2360 &full_scans_attr.attr,
2362 &merge_across_nodes_attr.attr,
2367 static struct attribute_group ksm_attr_group = {
2371 #endif /* CONFIG_SYSFS */
2373 static int __init ksm_init(void)
2375 struct task_struct *ksm_thread;
2378 err = ksm_slab_init();
2382 ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
2383 if (IS_ERR(ksm_thread)) {
2384 pr_err("ksm: creating kthread failed\n");
2385 err = PTR_ERR(ksm_thread);
2390 err = sysfs_create_group(mm_kobj, &ksm_attr_group);
2392 pr_err("ksm: register sysfs failed\n");
2393 kthread_stop(ksm_thread);
2397 ksm_run = KSM_RUN_MERGE; /* no way for user to start it */
2399 #endif /* CONFIG_SYSFS */
2401 #ifdef CONFIG_MEMORY_HOTREMOVE
2402 /* There is no significance to this priority 100 */
2403 hotplug_memory_notifier(ksm_memory_callback, 100);
2412 subsys_initcall(ksm_init);