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/sched/mm.h>
23 #include <linux/sched/coredump.h>
24 #include <linux/rwsem.h>
25 #include <linux/pagemap.h>
26 #include <linux/rmap.h>
27 #include <linux/spinlock.h>
28 #include <linux/jhash.h>
29 #include <linux/delay.h>
30 #include <linux/kthread.h>
31 #include <linux/wait.h>
32 #include <linux/slab.h>
33 #include <linux/rbtree.h>
34 #include <linux/memory.h>
35 #include <linux/mmu_notifier.h>
36 #include <linux/swap.h>
37 #include <linux/ksm.h>
38 #include <linux/hashtable.h>
39 #include <linux/freezer.h>
40 #include <linux/oom.h>
41 #include <linux/numa.h>
43 #include <asm/tlbflush.h>
48 #define DO_NUMA(x) do { (x); } while (0)
51 #define DO_NUMA(x) do { } while (0)
55 * A few notes about the KSM scanning process,
56 * to make it easier to understand the data structures below:
58 * In order to reduce excessive scanning, KSM sorts the memory pages by their
59 * contents into a data structure that holds pointers to the pages' locations.
61 * Since the contents of the pages may change at any moment, KSM cannot just
62 * insert the pages into a normal sorted tree and expect it to find anything.
63 * Therefore KSM uses two data structures - the stable and the unstable tree.
65 * The stable tree holds pointers to all the merged pages (ksm pages), sorted
66 * by their contents. Because each such page is write-protected, searching on
67 * this tree is fully assured to be working (except when pages are unmapped),
68 * and therefore this tree is called the stable tree.
70 * In addition to the stable tree, KSM uses a second data structure called the
71 * unstable tree: this tree holds pointers to pages which have been found to
72 * be "unchanged for a period of time". The unstable tree sorts these pages
73 * by their contents, but since they are not write-protected, KSM cannot rely
74 * upon the unstable tree to work correctly - the unstable tree is liable to
75 * be corrupted as its contents are modified, and so it is called unstable.
77 * KSM solves this problem by several techniques:
79 * 1) The unstable tree is flushed every time KSM completes scanning all
80 * memory areas, and then the tree is rebuilt again from the beginning.
81 * 2) KSM will only insert into the unstable tree, pages whose hash value
82 * has not changed since the previous scan of all memory areas.
83 * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
84 * colors of the nodes and not on their contents, assuring that even when
85 * the tree gets "corrupted" it won't get out of balance, so scanning time
86 * remains the same (also, searching and inserting nodes in an rbtree uses
87 * the same algorithm, so we have no overhead when we flush and rebuild).
88 * 4) KSM never flushes the stable tree, which means that even if it were to
89 * take 10 attempts to find a page in the unstable tree, once it is found,
90 * it is secured in the stable tree. (When we scan a new page, we first
91 * compare it against the stable tree, and then against the unstable tree.)
93 * If the merge_across_nodes tunable is unset, then KSM maintains multiple
94 * stable trees and multiple unstable trees: one of each for each NUMA node.
98 * struct mm_slot - ksm information per mm that is being scanned
99 * @link: link to the mm_slots hash list
100 * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
101 * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
102 * @mm: the mm that this information is valid for
105 struct hlist_node link;
106 struct list_head mm_list;
107 struct rmap_item *rmap_list;
108 struct mm_struct *mm;
112 * struct ksm_scan - cursor for scanning
113 * @mm_slot: the current mm_slot we are scanning
114 * @address: the next address inside that to be scanned
115 * @rmap_list: link to the next rmap to be scanned in the rmap_list
116 * @seqnr: count of completed full scans (needed when removing unstable node)
118 * There is only the one ksm_scan instance of this cursor structure.
121 struct mm_slot *mm_slot;
122 unsigned long address;
123 struct rmap_item **rmap_list;
128 * struct stable_node - node of the stable rbtree
129 * @node: rb node of this ksm page in the stable tree
130 * @head: (overlaying parent) &migrate_nodes indicates temporarily on that list
131 * @hlist_dup: linked into the stable_node->hlist with a stable_node chain
132 * @list: linked into migrate_nodes, pending placement in the proper node tree
133 * @hlist: hlist head of rmap_items using this ksm page
134 * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid)
135 * @chain_prune_time: time of the last full garbage collection
136 * @rmap_hlist_len: number of rmap_item entries in hlist or STABLE_NODE_CHAIN
137 * @nid: NUMA node id of stable tree in which linked (may not match kpfn)
141 struct rb_node node; /* when node of stable tree */
142 struct { /* when listed for migration */
143 struct list_head *head;
145 struct hlist_node hlist_dup;
146 struct list_head list;
150 struct hlist_head hlist;
153 unsigned long chain_prune_time;
156 * STABLE_NODE_CHAIN can be any negative number in
157 * rmap_hlist_len negative range, but better not -1 to be able
158 * to reliably detect underflows.
160 #define STABLE_NODE_CHAIN -1024
168 * struct rmap_item - reverse mapping item for virtual addresses
169 * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
170 * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
171 * @nid: NUMA node id of unstable tree in which linked (may not match page)
172 * @mm: the memory structure this rmap_item is pointing into
173 * @address: the virtual address this rmap_item tracks (+ flags in low bits)
174 * @oldchecksum: previous checksum of the page at that virtual address
175 * @node: rb node of this rmap_item in the unstable tree
176 * @head: pointer to stable_node heading this list in the stable tree
177 * @hlist: link into hlist of rmap_items hanging off that stable_node
180 struct rmap_item *rmap_list;
182 struct anon_vma *anon_vma; /* when stable */
184 int nid; /* when node of unstable tree */
187 struct mm_struct *mm;
188 unsigned long address; /* + low bits used for flags below */
189 unsigned int oldchecksum; /* when unstable */
191 struct rb_node node; /* when node of unstable tree */
192 struct { /* when listed from stable tree */
193 struct stable_node *head;
194 struct hlist_node hlist;
199 #define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */
200 #define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */
201 #define STABLE_FLAG 0x200 /* is listed from the stable tree */
202 #define KSM_FLAG_MASK (SEQNR_MASK|UNSTABLE_FLAG|STABLE_FLAG)
203 /* to mask all the flags */
205 /* The stable and unstable tree heads */
206 static struct rb_root one_stable_tree[1] = { RB_ROOT };
207 static struct rb_root one_unstable_tree[1] = { RB_ROOT };
208 static struct rb_root *root_stable_tree = one_stable_tree;
209 static struct rb_root *root_unstable_tree = one_unstable_tree;
211 /* Recently migrated nodes of stable tree, pending proper placement */
212 static LIST_HEAD(migrate_nodes);
213 #define STABLE_NODE_DUP_HEAD ((struct list_head *)&migrate_nodes.prev)
215 #define MM_SLOTS_HASH_BITS 10
216 static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
218 static struct mm_slot ksm_mm_head = {
219 .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
221 static struct ksm_scan ksm_scan = {
222 .mm_slot = &ksm_mm_head,
225 static struct kmem_cache *rmap_item_cache;
226 static struct kmem_cache *stable_node_cache;
227 static struct kmem_cache *mm_slot_cache;
229 /* The number of nodes in the stable tree */
230 static unsigned long ksm_pages_shared;
232 /* The number of page slots additionally sharing those nodes */
233 static unsigned long ksm_pages_sharing;
235 /* The number of nodes in the unstable tree */
236 static unsigned long ksm_pages_unshared;
238 /* The number of rmap_items in use: to calculate pages_volatile */
239 static unsigned long ksm_rmap_items;
241 /* The number of stable_node chains */
242 static unsigned long ksm_stable_node_chains;
244 /* The number of stable_node dups linked to the stable_node chains */
245 static unsigned long ksm_stable_node_dups;
247 /* Delay in pruning stale stable_node_dups in the stable_node_chains */
248 static int ksm_stable_node_chains_prune_millisecs = 2000;
250 /* Maximum number of page slots sharing a stable node */
251 static int ksm_max_page_sharing = 256;
253 /* Number of pages ksmd should scan in one batch */
254 static unsigned int ksm_thread_pages_to_scan = 100;
256 /* Milliseconds ksmd should sleep between batches */
257 static unsigned int ksm_thread_sleep_millisecs = 20;
259 /* Checksum of an empty (zeroed) page */
260 static unsigned int zero_checksum __read_mostly;
262 /* Whether to merge empty (zeroed) pages with actual zero pages */
263 static bool ksm_use_zero_pages __read_mostly;
266 /* Zeroed when merging across nodes is not allowed */
267 static unsigned int ksm_merge_across_nodes = 1;
268 static int ksm_nr_node_ids = 1;
270 #define ksm_merge_across_nodes 1U
271 #define ksm_nr_node_ids 1
274 #define KSM_RUN_STOP 0
275 #define KSM_RUN_MERGE 1
276 #define KSM_RUN_UNMERGE 2
277 #define KSM_RUN_OFFLINE 4
278 static unsigned long ksm_run = KSM_RUN_STOP;
279 static void wait_while_offlining(void);
281 static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
282 static DEFINE_MUTEX(ksm_thread_mutex);
283 static DEFINE_SPINLOCK(ksm_mmlist_lock);
285 #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
286 sizeof(struct __struct), __alignof__(struct __struct),\
289 static int __init ksm_slab_init(void)
291 rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
292 if (!rmap_item_cache)
295 stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
296 if (!stable_node_cache)
299 mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
306 kmem_cache_destroy(stable_node_cache);
308 kmem_cache_destroy(rmap_item_cache);
313 static void __init ksm_slab_free(void)
315 kmem_cache_destroy(mm_slot_cache);
316 kmem_cache_destroy(stable_node_cache);
317 kmem_cache_destroy(rmap_item_cache);
318 mm_slot_cache = NULL;
321 static __always_inline bool is_stable_node_chain(struct stable_node *chain)
323 return chain->rmap_hlist_len == STABLE_NODE_CHAIN;
326 static __always_inline bool is_stable_node_dup(struct stable_node *dup)
328 return dup->head == STABLE_NODE_DUP_HEAD;
331 static inline void stable_node_chain_add_dup(struct stable_node *dup,
332 struct stable_node *chain)
334 VM_BUG_ON(is_stable_node_dup(dup));
335 dup->head = STABLE_NODE_DUP_HEAD;
336 VM_BUG_ON(!is_stable_node_chain(chain));
337 hlist_add_head(&dup->hlist_dup, &chain->hlist);
338 ksm_stable_node_dups++;
341 static inline void __stable_node_dup_del(struct stable_node *dup)
343 VM_BUG_ON(!is_stable_node_dup(dup));
344 hlist_del(&dup->hlist_dup);
345 ksm_stable_node_dups--;
348 static inline void stable_node_dup_del(struct stable_node *dup)
350 VM_BUG_ON(is_stable_node_chain(dup));
351 if (is_stable_node_dup(dup))
352 __stable_node_dup_del(dup);
354 rb_erase(&dup->node, root_stable_tree + NUMA(dup->nid));
355 #ifdef CONFIG_DEBUG_VM
360 static inline struct rmap_item *alloc_rmap_item(void)
362 struct rmap_item *rmap_item;
364 rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL |
365 __GFP_NORETRY | __GFP_NOWARN);
371 static inline void free_rmap_item(struct rmap_item *rmap_item)
374 rmap_item->mm = NULL; /* debug safety */
375 kmem_cache_free(rmap_item_cache, rmap_item);
378 static inline struct stable_node *alloc_stable_node(void)
381 * The allocation can take too long with GFP_KERNEL when memory is under
382 * pressure, which may lead to hung task warnings. Adding __GFP_HIGH
383 * grants access to memory reserves, helping to avoid this problem.
385 return kmem_cache_alloc(stable_node_cache, GFP_KERNEL | __GFP_HIGH);
388 static inline void free_stable_node(struct stable_node *stable_node)
390 VM_BUG_ON(stable_node->rmap_hlist_len &&
391 !is_stable_node_chain(stable_node));
392 kmem_cache_free(stable_node_cache, stable_node);
395 static inline struct mm_slot *alloc_mm_slot(void)
397 if (!mm_slot_cache) /* initialization failed */
399 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
402 static inline void free_mm_slot(struct mm_slot *mm_slot)
404 kmem_cache_free(mm_slot_cache, mm_slot);
407 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
409 struct mm_slot *slot;
411 hash_for_each_possible(mm_slots_hash, slot, link, (unsigned long)mm)
418 static void insert_to_mm_slots_hash(struct mm_struct *mm,
419 struct mm_slot *mm_slot)
422 hash_add(mm_slots_hash, &mm_slot->link, (unsigned long)mm);
426 * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
427 * page tables after it has passed through ksm_exit() - which, if necessary,
428 * takes mmap_sem briefly to serialize against them. ksm_exit() does not set
429 * a special flag: they can just back out as soon as mm_users goes to zero.
430 * ksm_test_exit() is used throughout to make this test for exit: in some
431 * places for correctness, in some places just to avoid unnecessary work.
433 static inline bool ksm_test_exit(struct mm_struct *mm)
435 return atomic_read(&mm->mm_users) == 0;
439 * We use break_ksm to break COW on a ksm page: it's a stripped down
441 * if (get_user_pages(addr, 1, 1, 1, &page, NULL) == 1)
444 * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
445 * in case the application has unmapped and remapped mm,addr meanwhile.
446 * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP
447 * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
449 * FAULT_FLAG/FOLL_REMOTE are because we do this outside the context
450 * of the process that owns 'vma'. We also do not want to enforce
451 * protection keys here anyway.
453 static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
460 page = follow_page(vma, addr,
461 FOLL_GET | FOLL_MIGRATION | FOLL_REMOTE);
462 if (IS_ERR_OR_NULL(page))
465 ret = handle_mm_fault(vma, addr,
466 FAULT_FLAG_WRITE | FAULT_FLAG_REMOTE);
468 ret = VM_FAULT_WRITE;
470 } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV | VM_FAULT_OOM)));
472 * We must loop because handle_mm_fault() may back out if there's
473 * any difficulty e.g. if pte accessed bit gets updated concurrently.
475 * VM_FAULT_WRITE is what we have been hoping for: it indicates that
476 * COW has been broken, even if the vma does not permit VM_WRITE;
477 * but note that a concurrent fault might break PageKsm for us.
479 * VM_FAULT_SIGBUS could occur if we race with truncation of the
480 * backing file, which also invalidates anonymous pages: that's
481 * okay, that truncation will have unmapped the PageKsm for us.
483 * VM_FAULT_OOM: at the time of writing (late July 2009), setting
484 * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
485 * current task has TIF_MEMDIE set, and will be OOM killed on return
486 * to user; and ksmd, having no mm, would never be chosen for that.
488 * But if the mm is in a limited mem_cgroup, then the fault may fail
489 * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
490 * even ksmd can fail in this way - though it's usually breaking ksm
491 * just to undo a merge it made a moment before, so unlikely to oom.
493 * That's a pity: we might therefore have more kernel pages allocated
494 * than we're counting as nodes in the stable tree; but ksm_do_scan
495 * will retry to break_cow on each pass, so should recover the page
496 * in due course. The important thing is to not let VM_MERGEABLE
497 * be cleared while any such pages might remain in the area.
499 return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
502 static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm,
505 struct vm_area_struct *vma;
506 if (ksm_test_exit(mm))
508 vma = find_vma(mm, addr);
509 if (!vma || vma->vm_start > addr)
511 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
516 static void break_cow(struct rmap_item *rmap_item)
518 struct mm_struct *mm = rmap_item->mm;
519 unsigned long addr = rmap_item->address;
520 struct vm_area_struct *vma;
523 * It is not an accident that whenever we want to break COW
524 * to undo, we also need to drop a reference to the anon_vma.
526 put_anon_vma(rmap_item->anon_vma);
528 down_read(&mm->mmap_sem);
529 vma = find_mergeable_vma(mm, addr);
531 break_ksm(vma, addr);
532 up_read(&mm->mmap_sem);
535 static struct page *get_mergeable_page(struct rmap_item *rmap_item)
537 struct mm_struct *mm = rmap_item->mm;
538 unsigned long addr = rmap_item->address;
539 struct vm_area_struct *vma;
542 down_read(&mm->mmap_sem);
543 vma = find_mergeable_vma(mm, addr);
547 page = follow_page(vma, addr, FOLL_GET);
548 if (IS_ERR_OR_NULL(page))
550 if (PageAnon(page)) {
551 flush_anon_page(vma, page, addr);
552 flush_dcache_page(page);
558 up_read(&mm->mmap_sem);
563 * This helper is used for getting right index into array of tree roots.
564 * When merge_across_nodes knob is set to 1, there are only two rb-trees for
565 * stable and unstable pages from all nodes with roots in index 0. Otherwise,
566 * every node has its own stable and unstable tree.
568 static inline int get_kpfn_nid(unsigned long kpfn)
570 return ksm_merge_across_nodes ? 0 : NUMA(pfn_to_nid(kpfn));
573 static struct stable_node *alloc_stable_node_chain(struct stable_node *dup,
574 struct rb_root *root)
576 struct stable_node *chain = alloc_stable_node();
577 VM_BUG_ON(is_stable_node_chain(dup));
579 INIT_HLIST_HEAD(&chain->hlist);
580 chain->chain_prune_time = jiffies;
581 chain->rmap_hlist_len = STABLE_NODE_CHAIN;
582 #if defined (CONFIG_DEBUG_VM) && defined(CONFIG_NUMA)
583 chain->nid = -1; /* debug */
585 ksm_stable_node_chains++;
588 * Put the stable node chain in the first dimension of
589 * the stable tree and at the same time remove the old
592 rb_replace_node(&dup->node, &chain->node, root);
595 * Move the old stable node to the second dimension
596 * queued in the hlist_dup. The invariant is that all
597 * dup stable_nodes in the chain->hlist point to pages
598 * that are wrprotected and have the exact same
601 stable_node_chain_add_dup(dup, chain);
606 static inline void free_stable_node_chain(struct stable_node *chain,
607 struct rb_root *root)
609 rb_erase(&chain->node, root);
610 free_stable_node(chain);
611 ksm_stable_node_chains--;
614 static void remove_node_from_stable_tree(struct stable_node *stable_node)
616 struct rmap_item *rmap_item;
618 /* check it's not STABLE_NODE_CHAIN or negative */
619 BUG_ON(stable_node->rmap_hlist_len < 0);
621 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
622 if (rmap_item->hlist.next)
626 VM_BUG_ON(stable_node->rmap_hlist_len <= 0);
627 stable_node->rmap_hlist_len--;
628 put_anon_vma(rmap_item->anon_vma);
629 rmap_item->address &= PAGE_MASK;
634 * We need the second aligned pointer of the migrate_nodes
635 * list_head to stay clear from the rb_parent_color union
636 * (aligned and different than any node) and also different
637 * from &migrate_nodes. This will verify that future list.h changes
638 * don't break STABLE_NODE_DUP_HEAD.
640 #if GCC_VERSION >= 40903 /* only recent gcc can handle it */
641 BUILD_BUG_ON(STABLE_NODE_DUP_HEAD <= &migrate_nodes);
642 BUILD_BUG_ON(STABLE_NODE_DUP_HEAD >= &migrate_nodes + 1);
645 if (stable_node->head == &migrate_nodes)
646 list_del(&stable_node->list);
648 stable_node_dup_del(stable_node);
649 free_stable_node(stable_node);
653 * get_ksm_page: checks if the page indicated by the stable node
654 * is still its ksm page, despite having held no reference to it.
655 * In which case we can trust the content of the page, and it
656 * returns the gotten page; but if the page has now been zapped,
657 * remove the stale node from the stable tree and return NULL.
658 * But beware, the stable node's page might be being migrated.
660 * You would expect the stable_node to hold a reference to the ksm page.
661 * But if it increments the page's count, swapping out has to wait for
662 * ksmd to come around again before it can free the page, which may take
663 * seconds or even minutes: much too unresponsive. So instead we use a
664 * "keyhole reference": access to the ksm page from the stable node peeps
665 * out through its keyhole to see if that page still holds the right key,
666 * pointing back to this stable node. This relies on freeing a PageAnon
667 * page to reset its page->mapping to NULL, and relies on no other use of
668 * a page to put something that might look like our key in page->mapping.
669 * is on its way to being freed; but it is an anomaly to bear in mind.
671 static struct page *get_ksm_page(struct stable_node *stable_node, bool lock_it)
674 void *expected_mapping;
677 expected_mapping = (void *)((unsigned long)stable_node |
680 kpfn = READ_ONCE(stable_node->kpfn);
681 page = pfn_to_page(kpfn);
684 * page is computed from kpfn, so on most architectures reading
685 * page->mapping is naturally ordered after reading node->kpfn,
686 * but on Alpha we need to be more careful.
688 smp_read_barrier_depends();
689 if (READ_ONCE(page->mapping) != expected_mapping)
693 * We cannot do anything with the page while its refcount is 0.
694 * Usually 0 means free, or tail of a higher-order page: in which
695 * case this node is no longer referenced, and should be freed;
696 * however, it might mean that the page is under page_freeze_refs().
697 * The __remove_mapping() case is easy, again the node is now stale;
698 * but if page is swapcache in migrate_page_move_mapping(), it might
699 * still be our page, in which case it's essential to keep the node.
701 while (!get_page_unless_zero(page)) {
703 * Another check for page->mapping != expected_mapping would
704 * work here too. We have chosen the !PageSwapCache test to
705 * optimize the common case, when the page is or is about to
706 * be freed: PageSwapCache is cleared (under spin_lock_irq)
707 * in the freeze_refs section of __remove_mapping(); but Anon
708 * page->mapping reset to NULL later, in free_pages_prepare().
710 if (!PageSwapCache(page))
715 if (READ_ONCE(page->mapping) != expected_mapping) {
722 if (READ_ONCE(page->mapping) != expected_mapping) {
732 * We come here from above when page->mapping or !PageSwapCache
733 * suggests that the node is stale; but it might be under migration.
734 * We need smp_rmb(), matching the smp_wmb() in ksm_migrate_page(),
735 * before checking whether node->kpfn has been changed.
738 if (READ_ONCE(stable_node->kpfn) != kpfn)
740 remove_node_from_stable_tree(stable_node);
745 * Removing rmap_item from stable or unstable tree.
746 * This function will clean the information from the stable/unstable tree.
748 static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
750 if (rmap_item->address & STABLE_FLAG) {
751 struct stable_node *stable_node;
754 stable_node = rmap_item->head;
755 page = get_ksm_page(stable_node, true);
759 hlist_del(&rmap_item->hlist);
763 if (!hlist_empty(&stable_node->hlist))
767 VM_BUG_ON(stable_node->rmap_hlist_len <= 0);
768 stable_node->rmap_hlist_len--;
770 put_anon_vma(rmap_item->anon_vma);
771 rmap_item->head = NULL;
772 rmap_item->address &= PAGE_MASK;
774 } else if (rmap_item->address & UNSTABLE_FLAG) {
777 * Usually ksmd can and must skip the rb_erase, because
778 * root_unstable_tree was already reset to RB_ROOT.
779 * But be careful when an mm is exiting: do the rb_erase
780 * if this rmap_item was inserted by this scan, rather
781 * than left over from before.
783 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
786 rb_erase(&rmap_item->node,
787 root_unstable_tree + NUMA(rmap_item->nid));
788 ksm_pages_unshared--;
789 rmap_item->address &= PAGE_MASK;
792 cond_resched(); /* we're called from many long loops */
795 static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
796 struct rmap_item **rmap_list)
799 struct rmap_item *rmap_item = *rmap_list;
800 *rmap_list = rmap_item->rmap_list;
801 remove_rmap_item_from_tree(rmap_item);
802 free_rmap_item(rmap_item);
807 * Though it's very tempting to unmerge rmap_items from stable tree rather
808 * than check every pte of a given vma, the locking doesn't quite work for
809 * that - an rmap_item is assigned to the stable tree after inserting ksm
810 * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing
811 * rmap_items from parent to child at fork time (so as not to waste time
812 * if exit comes before the next scan reaches it).
814 * Similarly, although we'd like to remove rmap_items (so updating counts
815 * and freeing memory) when unmerging an area, it's easier to leave that
816 * to the next pass of ksmd - consider, for example, how ksmd might be
817 * in cmp_and_merge_page on one of the rmap_items we would be removing.
819 static int unmerge_ksm_pages(struct vm_area_struct *vma,
820 unsigned long start, unsigned long end)
825 for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
826 if (ksm_test_exit(vma->vm_mm))
828 if (signal_pending(current))
831 err = break_ksm(vma, addr);
838 * Only called through the sysfs control interface:
840 static int remove_stable_node(struct stable_node *stable_node)
845 page = get_ksm_page(stable_node, true);
848 * get_ksm_page did remove_node_from_stable_tree itself.
854 * Page could be still mapped if this races with __mmput() running in
855 * between ksm_exit() and exit_mmap(). Just refuse to let
856 * merge_across_nodes/max_page_sharing be switched.
859 if (!page_mapped(page)) {
861 * The stable node did not yet appear stale to get_ksm_page(),
862 * since that allows for an unmapped ksm page to be recognized
863 * right up until it is freed; but the node is safe to remove.
864 * This page might be in a pagevec waiting to be freed,
865 * or it might be PageSwapCache (perhaps under writeback),
866 * or it might have been removed from swapcache a moment ago.
868 set_page_stable_node(page, NULL);
869 remove_node_from_stable_tree(stable_node);
878 static int remove_stable_node_chain(struct stable_node *stable_node,
879 struct rb_root *root)
881 struct stable_node *dup;
882 struct hlist_node *hlist_safe;
884 if (!is_stable_node_chain(stable_node)) {
885 VM_BUG_ON(is_stable_node_dup(stable_node));
886 if (remove_stable_node(stable_node))
892 hlist_for_each_entry_safe(dup, hlist_safe,
893 &stable_node->hlist, hlist_dup) {
894 VM_BUG_ON(!is_stable_node_dup(dup));
895 if (remove_stable_node(dup))
898 BUG_ON(!hlist_empty(&stable_node->hlist));
899 free_stable_node_chain(stable_node, root);
903 static int remove_all_stable_nodes(void)
905 struct stable_node *stable_node, *next;
909 for (nid = 0; nid < ksm_nr_node_ids; nid++) {
910 while (root_stable_tree[nid].rb_node) {
911 stable_node = rb_entry(root_stable_tree[nid].rb_node,
912 struct stable_node, node);
913 if (remove_stable_node_chain(stable_node,
914 root_stable_tree + nid)) {
916 break; /* proceed to next nid */
921 list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
922 if (remove_stable_node(stable_node))
929 static int unmerge_and_remove_all_rmap_items(void)
931 struct mm_slot *mm_slot;
932 struct mm_struct *mm;
933 struct vm_area_struct *vma;
936 spin_lock(&ksm_mmlist_lock);
937 ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
938 struct mm_slot, mm_list);
939 spin_unlock(&ksm_mmlist_lock);
941 for (mm_slot = ksm_scan.mm_slot;
942 mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
944 down_read(&mm->mmap_sem);
945 for (vma = mm->mmap; vma; vma = vma->vm_next) {
946 if (ksm_test_exit(mm))
948 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
950 err = unmerge_ksm_pages(vma,
951 vma->vm_start, vma->vm_end);
956 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
957 up_read(&mm->mmap_sem);
959 spin_lock(&ksm_mmlist_lock);
960 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
961 struct mm_slot, mm_list);
962 if (ksm_test_exit(mm)) {
963 hash_del(&mm_slot->link);
964 list_del(&mm_slot->mm_list);
965 spin_unlock(&ksm_mmlist_lock);
967 free_mm_slot(mm_slot);
968 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
971 spin_unlock(&ksm_mmlist_lock);
974 /* Clean up stable nodes, but don't worry if some are still busy */
975 remove_all_stable_nodes();
980 up_read(&mm->mmap_sem);
981 spin_lock(&ksm_mmlist_lock);
982 ksm_scan.mm_slot = &ksm_mm_head;
983 spin_unlock(&ksm_mmlist_lock);
986 #endif /* CONFIG_SYSFS */
988 static u32 calc_checksum(struct page *page)
991 void *addr = kmap_atomic(page);
992 checksum = jhash2(addr, PAGE_SIZE / 4, 17);
997 static int memcmp_pages(struct page *page1, struct page *page2)
1002 addr1 = kmap_atomic(page1);
1003 addr2 = kmap_atomic(page2);
1004 ret = memcmp(addr1, addr2, PAGE_SIZE);
1005 kunmap_atomic(addr2);
1006 kunmap_atomic(addr1);
1010 static inline int pages_identical(struct page *page1, struct page *page2)
1012 return !memcmp_pages(page1, page2);
1015 static int write_protect_page(struct vm_area_struct *vma, struct page *page,
1018 struct mm_struct *mm = vma->vm_mm;
1019 struct page_vma_mapped_walk pvmw = {
1025 unsigned long mmun_start; /* For mmu_notifiers */
1026 unsigned long mmun_end; /* For mmu_notifiers */
1028 pvmw.address = page_address_in_vma(page, vma);
1029 if (pvmw.address == -EFAULT)
1032 BUG_ON(PageTransCompound(page));
1034 mmun_start = pvmw.address;
1035 mmun_end = pvmw.address + PAGE_SIZE;
1036 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1038 if (!page_vma_mapped_walk(&pvmw))
1040 if (WARN_ONCE(!pvmw.pte, "Unexpected PMD mapping?"))
1043 if (pte_write(*pvmw.pte) || pte_dirty(*pvmw.pte) ||
1044 (pte_protnone(*pvmw.pte) && pte_savedwrite(*pvmw.pte)) ||
1045 mm_tlb_flush_pending(mm)) {
1048 swapped = PageSwapCache(page);
1049 flush_cache_page(vma, pvmw.address, page_to_pfn(page));
1051 * Ok this is tricky, when get_user_pages_fast() run it doesn't
1052 * take any lock, therefore the check that we are going to make
1053 * with the pagecount against the mapcount is racey and
1054 * O_DIRECT can happen right after the check.
1055 * So we clear the pte and flush the tlb before the check
1056 * this assure us that no O_DIRECT can happen after the check
1057 * or in the middle of the check.
1059 entry = ptep_clear_flush_notify(vma, pvmw.address, pvmw.pte);
1061 * Check that no O_DIRECT or similar I/O is in progress on the
1064 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
1065 set_pte_at(mm, pvmw.address, pvmw.pte, entry);
1068 if (pte_dirty(entry))
1069 set_page_dirty(page);
1071 if (pte_protnone(entry))
1072 entry = pte_mkclean(pte_clear_savedwrite(entry));
1074 entry = pte_mkclean(pte_wrprotect(entry));
1075 set_pte_at_notify(mm, pvmw.address, pvmw.pte, entry);
1077 *orig_pte = *pvmw.pte;
1081 page_vma_mapped_walk_done(&pvmw);
1083 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1089 * replace_page - replace page in vma by new ksm page
1090 * @vma: vma that holds the pte pointing to page
1091 * @page: the page we are replacing by kpage
1092 * @kpage: the ksm page we replace page by
1093 * @orig_pte: the original value of the pte
1095 * Returns 0 on success, -EFAULT on failure.
1097 static int replace_page(struct vm_area_struct *vma, struct page *page,
1098 struct page *kpage, pte_t orig_pte)
1100 struct mm_struct *mm = vma->vm_mm;
1107 unsigned long mmun_start; /* For mmu_notifiers */
1108 unsigned long mmun_end; /* For mmu_notifiers */
1110 addr = page_address_in_vma(page, vma);
1111 if (addr == -EFAULT)
1114 pmd = mm_find_pmd(mm, addr);
1119 mmun_end = addr + PAGE_SIZE;
1120 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1122 ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
1123 if (!pte_same(*ptep, orig_pte)) {
1124 pte_unmap_unlock(ptep, ptl);
1129 * No need to check ksm_use_zero_pages here: we can only have a
1130 * zero_page here if ksm_use_zero_pages was enabled alreaady.
1132 if (!is_zero_pfn(page_to_pfn(kpage))) {
1134 page_add_anon_rmap(kpage, vma, addr, false);
1135 newpte = mk_pte(kpage, vma->vm_page_prot);
1137 newpte = pte_mkspecial(pfn_pte(page_to_pfn(kpage),
1138 vma->vm_page_prot));
1140 * We're replacing an anonymous page with a zero page, which is
1141 * not anonymous. We need to do proper accounting otherwise we
1142 * will get wrong values in /proc, and a BUG message in dmesg
1143 * when tearing down the mm.
1145 dec_mm_counter(mm, MM_ANONPAGES);
1148 flush_cache_page(vma, addr, pte_pfn(*ptep));
1149 ptep_clear_flush_notify(vma, addr, ptep);
1150 set_pte_at_notify(mm, addr, ptep, newpte);
1152 page_remove_rmap(page, false);
1153 if (!page_mapped(page))
1154 try_to_free_swap(page);
1157 pte_unmap_unlock(ptep, ptl);
1160 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1166 * try_to_merge_one_page - take two pages and merge them into one
1167 * @vma: the vma that holds the pte pointing to page
1168 * @page: the PageAnon page that we want to replace with kpage
1169 * @kpage: the PageKsm page that we want to map instead of page,
1170 * or NULL the first time when we want to use page as kpage.
1172 * This function returns 0 if the pages were merged, -EFAULT otherwise.
1174 static int try_to_merge_one_page(struct vm_area_struct *vma,
1175 struct page *page, struct page *kpage)
1177 pte_t orig_pte = __pte(0);
1180 if (page == kpage) /* ksm page forked */
1183 if (!PageAnon(page))
1187 * We need the page lock to read a stable PageSwapCache in
1188 * write_protect_page(). We use trylock_page() instead of
1189 * lock_page() because we don't want to wait here - we
1190 * prefer to continue scanning and merging different pages,
1191 * then come back to this page when it is unlocked.
1193 if (!trylock_page(page))
1196 if (PageTransCompound(page)) {
1197 if (split_huge_page(page))
1202 * If this anonymous page is mapped only here, its pte may need
1203 * to be write-protected. If it's mapped elsewhere, all of its
1204 * ptes are necessarily already write-protected. But in either
1205 * case, we need to lock and check page_count is not raised.
1207 if (write_protect_page(vma, page, &orig_pte) == 0) {
1210 * While we hold page lock, upgrade page from
1211 * PageAnon+anon_vma to PageKsm+NULL stable_node:
1212 * stable_tree_insert() will update stable_node.
1214 set_page_stable_node(page, NULL);
1215 mark_page_accessed(page);
1217 * Page reclaim just frees a clean page with no dirty
1218 * ptes: make sure that the ksm page would be swapped.
1220 if (!PageDirty(page))
1223 } else if (pages_identical(page, kpage))
1224 err = replace_page(vma, page, kpage, orig_pte);
1227 if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
1228 munlock_vma_page(page);
1229 if (!PageMlocked(kpage)) {
1232 mlock_vma_page(kpage);
1233 page = kpage; /* for final unlock */
1244 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
1245 * but no new kernel page is allocated: kpage must already be a ksm page.
1247 * This function returns 0 if the pages were merged, -EFAULT otherwise.
1249 static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
1250 struct page *page, struct page *kpage)
1252 struct mm_struct *mm = rmap_item->mm;
1253 struct vm_area_struct *vma;
1256 down_read(&mm->mmap_sem);
1257 vma = find_mergeable_vma(mm, rmap_item->address);
1261 err = try_to_merge_one_page(vma, page, kpage);
1265 /* Unstable nid is in union with stable anon_vma: remove first */
1266 remove_rmap_item_from_tree(rmap_item);
1268 /* Must get reference to anon_vma while still holding mmap_sem */
1269 rmap_item->anon_vma = vma->anon_vma;
1270 get_anon_vma(vma->anon_vma);
1272 up_read(&mm->mmap_sem);
1277 * try_to_merge_two_pages - take two identical pages and prepare them
1278 * to be merged into one page.
1280 * This function returns the kpage if we successfully merged two identical
1281 * pages into one ksm page, NULL otherwise.
1283 * Note that this function upgrades page to ksm page: if one of the pages
1284 * is already a ksm page, try_to_merge_with_ksm_page should be used.
1286 static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
1288 struct rmap_item *tree_rmap_item,
1289 struct page *tree_page)
1293 err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
1295 err = try_to_merge_with_ksm_page(tree_rmap_item,
1298 * If that fails, we have a ksm page with only one pte
1299 * pointing to it: so break it.
1302 break_cow(rmap_item);
1304 return err ? NULL : page;
1307 static __always_inline
1308 bool __is_page_sharing_candidate(struct stable_node *stable_node, int offset)
1310 VM_BUG_ON(stable_node->rmap_hlist_len < 0);
1312 * Check that at least one mapping still exists, otherwise
1313 * there's no much point to merge and share with this
1314 * stable_node, as the underlying tree_page of the other
1315 * sharer is going to be freed soon.
1317 return stable_node->rmap_hlist_len &&
1318 stable_node->rmap_hlist_len + offset < ksm_max_page_sharing;
1321 static __always_inline
1322 bool is_page_sharing_candidate(struct stable_node *stable_node)
1324 return __is_page_sharing_candidate(stable_node, 0);
1327 struct page *stable_node_dup(struct stable_node **_stable_node_dup,
1328 struct stable_node **_stable_node,
1329 struct rb_root *root,
1330 bool prune_stale_stable_nodes)
1332 struct stable_node *dup, *found = NULL, *stable_node = *_stable_node;
1333 struct hlist_node *hlist_safe;
1334 struct page *_tree_page, *tree_page = NULL;
1336 int found_rmap_hlist_len;
1338 if (!prune_stale_stable_nodes ||
1339 time_before(jiffies, stable_node->chain_prune_time +
1341 ksm_stable_node_chains_prune_millisecs)))
1342 prune_stale_stable_nodes = false;
1344 stable_node->chain_prune_time = jiffies;
1346 hlist_for_each_entry_safe(dup, hlist_safe,
1347 &stable_node->hlist, hlist_dup) {
1350 * We must walk all stable_node_dup to prune the stale
1351 * stable nodes during lookup.
1353 * get_ksm_page can drop the nodes from the
1354 * stable_node->hlist if they point to freed pages
1355 * (that's why we do a _safe walk). The "dup"
1356 * stable_node parameter itself will be freed from
1357 * under us if it returns NULL.
1359 _tree_page = get_ksm_page(dup, false);
1363 if (is_page_sharing_candidate(dup)) {
1365 dup->rmap_hlist_len > found_rmap_hlist_len) {
1367 put_page(tree_page);
1369 found_rmap_hlist_len = found->rmap_hlist_len;
1370 tree_page = _tree_page;
1372 /* skip put_page for found dup */
1373 if (!prune_stale_stable_nodes)
1378 put_page(_tree_page);
1383 * nr is counting all dups in the chain only if
1384 * prune_stale_stable_nodes is true, otherwise we may
1385 * break the loop at nr == 1 even if there are
1388 if (prune_stale_stable_nodes && nr == 1) {
1390 * If there's not just one entry it would
1391 * corrupt memory, better BUG_ON. In KSM
1392 * context with no lock held it's not even
1395 BUG_ON(stable_node->hlist.first->next);
1398 * There's just one entry and it is below the
1399 * deduplication limit so drop the chain.
1401 rb_replace_node(&stable_node->node, &found->node,
1403 free_stable_node(stable_node);
1404 ksm_stable_node_chains--;
1405 ksm_stable_node_dups--;
1407 * NOTE: the caller depends on the stable_node
1408 * to be equal to stable_node_dup if the chain
1411 *_stable_node = found;
1413 * Just for robustneess as stable_node is
1414 * otherwise left as a stable pointer, the
1415 * compiler shall optimize it away at build
1419 } else if (stable_node->hlist.first != &found->hlist_dup &&
1420 __is_page_sharing_candidate(found, 1)) {
1422 * If the found stable_node dup can accept one
1423 * more future merge (in addition to the one
1424 * that is underway) and is not at the head of
1425 * the chain, put it there so next search will
1426 * be quicker in the !prune_stale_stable_nodes
1429 * NOTE: it would be inaccurate to use nr > 1
1430 * instead of checking the hlist.first pointer
1431 * directly, because in the
1432 * prune_stale_stable_nodes case "nr" isn't
1433 * the position of the found dup in the chain,
1434 * but the total number of dups in the chain.
1436 hlist_del(&found->hlist_dup);
1437 hlist_add_head(&found->hlist_dup,
1438 &stable_node->hlist);
1442 *_stable_node_dup = found;
1446 static struct stable_node *stable_node_dup_any(struct stable_node *stable_node,
1447 struct rb_root *root)
1449 if (!is_stable_node_chain(stable_node))
1451 if (hlist_empty(&stable_node->hlist)) {
1452 free_stable_node_chain(stable_node, root);
1455 return hlist_entry(stable_node->hlist.first,
1456 typeof(*stable_node), hlist_dup);
1460 * Like for get_ksm_page, this function can free the *_stable_node and
1461 * *_stable_node_dup if the returned tree_page is NULL.
1463 * It can also free and overwrite *_stable_node with the found
1464 * stable_node_dup if the chain is collapsed (in which case
1465 * *_stable_node will be equal to *_stable_node_dup like if the chain
1466 * never existed). It's up to the caller to verify tree_page is not
1467 * NULL before dereferencing *_stable_node or *_stable_node_dup.
1469 * *_stable_node_dup is really a second output parameter of this
1470 * function and will be overwritten in all cases, the caller doesn't
1471 * need to initialize it.
1473 static struct page *__stable_node_chain(struct stable_node **_stable_node_dup,
1474 struct stable_node **_stable_node,
1475 struct rb_root *root,
1476 bool prune_stale_stable_nodes)
1478 struct stable_node *stable_node = *_stable_node;
1479 if (!is_stable_node_chain(stable_node)) {
1480 if (is_page_sharing_candidate(stable_node)) {
1481 *_stable_node_dup = stable_node;
1482 return get_ksm_page(stable_node, false);
1485 * _stable_node_dup set to NULL means the stable_node
1486 * reached the ksm_max_page_sharing limit.
1488 *_stable_node_dup = NULL;
1491 return stable_node_dup(_stable_node_dup, _stable_node, root,
1492 prune_stale_stable_nodes);
1495 static __always_inline struct page *chain_prune(struct stable_node **s_n_d,
1496 struct stable_node **s_n,
1497 struct rb_root *root)
1499 return __stable_node_chain(s_n_d, s_n, root, true);
1502 static __always_inline struct page *chain(struct stable_node **s_n_d,
1503 struct stable_node *s_n,
1504 struct rb_root *root)
1506 struct stable_node *old_stable_node = s_n;
1507 struct page *tree_page;
1509 tree_page = __stable_node_chain(s_n_d, &s_n, root, false);
1510 /* not pruning dups so s_n cannot have changed */
1511 VM_BUG_ON(s_n != old_stable_node);
1516 * stable_tree_search - search for page inside the stable tree
1518 * This function checks if there is a page inside the stable tree
1519 * with identical content to the page that we are scanning right now.
1521 * This function returns the stable tree node of identical content if found,
1524 static struct page *stable_tree_search(struct page *page)
1527 struct rb_root *root;
1528 struct rb_node **new;
1529 struct rb_node *parent;
1530 struct stable_node *stable_node, *stable_node_dup, *stable_node_any;
1531 struct stable_node *page_node;
1533 page_node = page_stable_node(page);
1534 if (page_node && page_node->head != &migrate_nodes) {
1535 /* ksm page forked */
1540 nid = get_kpfn_nid(page_to_pfn(page));
1541 root = root_stable_tree + nid;
1543 new = &root->rb_node;
1547 struct page *tree_page;
1551 stable_node = rb_entry(*new, struct stable_node, node);
1552 stable_node_any = NULL;
1553 tree_page = chain_prune(&stable_node_dup, &stable_node, root);
1555 * NOTE: stable_node may have been freed by
1556 * chain_prune() if the returned stable_node_dup is
1557 * not NULL. stable_node_dup may have been inserted in
1558 * the rbtree instead as a regular stable_node (in
1559 * order to collapse the stable_node chain if a single
1560 * stable_node dup was found in it). In such case the
1561 * stable_node is overwritten by the calleee to point
1562 * to the stable_node_dup that was collapsed in the
1563 * stable rbtree and stable_node will be equal to
1564 * stable_node_dup like if the chain never existed.
1566 if (!stable_node_dup) {
1568 * Either all stable_node dups were full in
1569 * this stable_node chain, or this chain was
1570 * empty and should be rb_erased.
1572 stable_node_any = stable_node_dup_any(stable_node,
1574 if (!stable_node_any) {
1575 /* rb_erase just run */
1579 * Take any of the stable_node dups page of
1580 * this stable_node chain to let the tree walk
1581 * continue. All KSM pages belonging to the
1582 * stable_node dups in a stable_node chain
1583 * have the same content and they're
1584 * wrprotected at all times. Any will work
1585 * fine to continue the walk.
1587 tree_page = get_ksm_page(stable_node_any, false);
1589 VM_BUG_ON(!stable_node_dup ^ !!stable_node_any);
1592 * If we walked over a stale stable_node,
1593 * get_ksm_page() will call rb_erase() and it
1594 * may rebalance the tree from under us. So
1595 * restart the search from scratch. Returning
1596 * NULL would be safe too, but we'd generate
1597 * false negative insertions just because some
1598 * stable_node was stale.
1603 ret = memcmp_pages(page, tree_page);
1604 put_page(tree_page);
1608 new = &parent->rb_left;
1610 new = &parent->rb_right;
1613 VM_BUG_ON(page_node->head != &migrate_nodes);
1615 * Test if the migrated page should be merged
1616 * into a stable node dup. If the mapcount is
1617 * 1 we can migrate it with another KSM page
1618 * without adding it to the chain.
1620 if (page_mapcount(page) > 1)
1624 if (!stable_node_dup) {
1626 * If the stable_node is a chain and
1627 * we got a payload match in memcmp
1628 * but we cannot merge the scanned
1629 * page in any of the existing
1630 * stable_node dups because they're
1631 * all full, we need to wait the
1632 * scanned page to find itself a match
1633 * in the unstable tree to create a
1634 * brand new KSM page to add later to
1635 * the dups of this stable_node.
1641 * Lock and unlock the stable_node's page (which
1642 * might already have been migrated) so that page
1643 * migration is sure to notice its raised count.
1644 * It would be more elegant to return stable_node
1645 * than kpage, but that involves more changes.
1647 tree_page = get_ksm_page(stable_node_dup, true);
1648 if (unlikely(!tree_page))
1650 * The tree may have been rebalanced,
1651 * so re-evaluate parent and new.
1654 unlock_page(tree_page);
1656 if (get_kpfn_nid(stable_node_dup->kpfn) !=
1657 NUMA(stable_node_dup->nid)) {
1658 put_page(tree_page);
1668 list_del(&page_node->list);
1669 DO_NUMA(page_node->nid = nid);
1670 rb_link_node(&page_node->node, parent, new);
1671 rb_insert_color(&page_node->node, root);
1673 if (is_page_sharing_candidate(page_node)) {
1681 * If stable_node was a chain and chain_prune collapsed it,
1682 * stable_node has been updated to be the new regular
1683 * stable_node. A collapse of the chain is indistinguishable
1684 * from the case there was no chain in the stable
1685 * rbtree. Otherwise stable_node is the chain and
1686 * stable_node_dup is the dup to replace.
1688 if (stable_node_dup == stable_node) {
1689 VM_BUG_ON(is_stable_node_chain(stable_node_dup));
1690 VM_BUG_ON(is_stable_node_dup(stable_node_dup));
1691 /* there is no chain */
1693 VM_BUG_ON(page_node->head != &migrate_nodes);
1694 list_del(&page_node->list);
1695 DO_NUMA(page_node->nid = nid);
1696 rb_replace_node(&stable_node_dup->node,
1699 if (is_page_sharing_candidate(page_node))
1704 rb_erase(&stable_node_dup->node, root);
1708 VM_BUG_ON(!is_stable_node_chain(stable_node));
1709 __stable_node_dup_del(stable_node_dup);
1711 VM_BUG_ON(page_node->head != &migrate_nodes);
1712 list_del(&page_node->list);
1713 DO_NUMA(page_node->nid = nid);
1714 stable_node_chain_add_dup(page_node, stable_node);
1715 if (is_page_sharing_candidate(page_node))
1723 stable_node_dup->head = &migrate_nodes;
1724 list_add(&stable_node_dup->list, stable_node_dup->head);
1728 /* stable_node_dup could be null if it reached the limit */
1729 if (!stable_node_dup)
1730 stable_node_dup = stable_node_any;
1732 * If stable_node was a chain and chain_prune collapsed it,
1733 * stable_node has been updated to be the new regular
1734 * stable_node. A collapse of the chain is indistinguishable
1735 * from the case there was no chain in the stable
1736 * rbtree. Otherwise stable_node is the chain and
1737 * stable_node_dup is the dup to replace.
1739 if (stable_node_dup == stable_node) {
1740 VM_BUG_ON(is_stable_node_chain(stable_node_dup));
1741 VM_BUG_ON(is_stable_node_dup(stable_node_dup));
1742 /* chain is missing so create it */
1743 stable_node = alloc_stable_node_chain(stable_node_dup,
1749 * Add this stable_node dup that was
1750 * migrated to the stable_node chain
1751 * of the current nid for this page
1754 VM_BUG_ON(!is_stable_node_chain(stable_node));
1755 VM_BUG_ON(!is_stable_node_dup(stable_node_dup));
1756 VM_BUG_ON(page_node->head != &migrate_nodes);
1757 list_del(&page_node->list);
1758 DO_NUMA(page_node->nid = nid);
1759 stable_node_chain_add_dup(page_node, stable_node);
1764 * stable_tree_insert - insert stable tree node pointing to new ksm page
1765 * into the stable tree.
1767 * This function returns the stable tree node just allocated on success,
1770 static struct stable_node *stable_tree_insert(struct page *kpage)
1774 struct rb_root *root;
1775 struct rb_node **new;
1776 struct rb_node *parent;
1777 struct stable_node *stable_node, *stable_node_dup, *stable_node_any;
1778 bool need_chain = false;
1780 kpfn = page_to_pfn(kpage);
1781 nid = get_kpfn_nid(kpfn);
1782 root = root_stable_tree + nid;
1785 new = &root->rb_node;
1788 struct page *tree_page;
1792 stable_node = rb_entry(*new, struct stable_node, node);
1793 stable_node_any = NULL;
1794 tree_page = chain(&stable_node_dup, stable_node, root);
1795 if (!stable_node_dup) {
1797 * Either all stable_node dups were full in
1798 * this stable_node chain, or this chain was
1799 * empty and should be rb_erased.
1801 stable_node_any = stable_node_dup_any(stable_node,
1803 if (!stable_node_any) {
1804 /* rb_erase just run */
1808 * Take any of the stable_node dups page of
1809 * this stable_node chain to let the tree walk
1810 * continue. All KSM pages belonging to the
1811 * stable_node dups in a stable_node chain
1812 * have the same content and they're
1813 * wrprotected at all times. Any will work
1814 * fine to continue the walk.
1816 tree_page = get_ksm_page(stable_node_any, false);
1818 VM_BUG_ON(!stable_node_dup ^ !!stable_node_any);
1821 * If we walked over a stale stable_node,
1822 * get_ksm_page() will call rb_erase() and it
1823 * may rebalance the tree from under us. So
1824 * restart the search from scratch. Returning
1825 * NULL would be safe too, but we'd generate
1826 * false negative insertions just because some
1827 * stable_node was stale.
1832 ret = memcmp_pages(kpage, tree_page);
1833 put_page(tree_page);
1837 new = &parent->rb_left;
1839 new = &parent->rb_right;
1846 stable_node_dup = alloc_stable_node();
1847 if (!stable_node_dup)
1850 INIT_HLIST_HEAD(&stable_node_dup->hlist);
1851 stable_node_dup->kpfn = kpfn;
1852 set_page_stable_node(kpage, stable_node_dup);
1853 stable_node_dup->rmap_hlist_len = 0;
1854 DO_NUMA(stable_node_dup->nid = nid);
1856 rb_link_node(&stable_node_dup->node, parent, new);
1857 rb_insert_color(&stable_node_dup->node, root);
1859 if (!is_stable_node_chain(stable_node)) {
1860 struct stable_node *orig = stable_node;
1861 /* chain is missing so create it */
1862 stable_node = alloc_stable_node_chain(orig, root);
1864 free_stable_node(stable_node_dup);
1868 stable_node_chain_add_dup(stable_node_dup, stable_node);
1871 return stable_node_dup;
1875 * unstable_tree_search_insert - search for identical page,
1876 * else insert rmap_item into the unstable tree.
1878 * This function searches for a page in the unstable tree identical to the
1879 * page currently being scanned; and if no identical page is found in the
1880 * tree, we insert rmap_item as a new object into the unstable tree.
1882 * This function returns pointer to rmap_item found to be identical
1883 * to the currently scanned page, NULL otherwise.
1885 * This function does both searching and inserting, because they share
1886 * the same walking algorithm in an rbtree.
1889 struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1891 struct page **tree_pagep)
1893 struct rb_node **new;
1894 struct rb_root *root;
1895 struct rb_node *parent = NULL;
1898 nid = get_kpfn_nid(page_to_pfn(page));
1899 root = root_unstable_tree + nid;
1900 new = &root->rb_node;
1903 struct rmap_item *tree_rmap_item;
1904 struct page *tree_page;
1908 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
1909 tree_page = get_mergeable_page(tree_rmap_item);
1914 * Don't substitute a ksm page for a forked page.
1916 if (page == tree_page) {
1917 put_page(tree_page);
1921 ret = memcmp_pages(page, tree_page);
1925 put_page(tree_page);
1926 new = &parent->rb_left;
1927 } else if (ret > 0) {
1928 put_page(tree_page);
1929 new = &parent->rb_right;
1930 } else if (!ksm_merge_across_nodes &&
1931 page_to_nid(tree_page) != nid) {
1933 * If tree_page has been migrated to another NUMA node,
1934 * it will be flushed out and put in the right unstable
1935 * tree next time: only merge with it when across_nodes.
1937 put_page(tree_page);
1940 *tree_pagep = tree_page;
1941 return tree_rmap_item;
1945 rmap_item->address |= UNSTABLE_FLAG;
1946 rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1947 DO_NUMA(rmap_item->nid = nid);
1948 rb_link_node(&rmap_item->node, parent, new);
1949 rb_insert_color(&rmap_item->node, root);
1951 ksm_pages_unshared++;
1956 * stable_tree_append - add another rmap_item to the linked list of
1957 * rmap_items hanging off a given node of the stable tree, all sharing
1958 * the same ksm page.
1960 static void stable_tree_append(struct rmap_item *rmap_item,
1961 struct stable_node *stable_node,
1962 bool max_page_sharing_bypass)
1965 * rmap won't find this mapping if we don't insert the
1966 * rmap_item in the right stable_node
1967 * duplicate. page_migration could break later if rmap breaks,
1968 * so we can as well crash here. We really need to check for
1969 * rmap_hlist_len == STABLE_NODE_CHAIN, but we can as well check
1970 * for other negative values as an undeflow if detected here
1971 * for the first time (and not when decreasing rmap_hlist_len)
1972 * would be sign of memory corruption in the stable_node.
1974 BUG_ON(stable_node->rmap_hlist_len < 0);
1976 stable_node->rmap_hlist_len++;
1977 if (!max_page_sharing_bypass)
1978 /* possibly non fatal but unexpected overflow, only warn */
1979 WARN_ON_ONCE(stable_node->rmap_hlist_len >
1980 ksm_max_page_sharing);
1982 rmap_item->head = stable_node;
1983 rmap_item->address |= STABLE_FLAG;
1984 hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
1986 if (rmap_item->hlist.next)
1987 ksm_pages_sharing++;
1993 * cmp_and_merge_page - first see if page can be merged into the stable tree;
1994 * if not, compare checksum to previous and if it's the same, see if page can
1995 * be inserted into the unstable tree, or merged with a page already there and
1996 * both transferred to the stable tree.
1998 * @page: the page that we are searching identical page to.
1999 * @rmap_item: the reverse mapping into the virtual address of this page
2001 static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
2003 struct mm_struct *mm = rmap_item->mm;
2004 struct rmap_item *tree_rmap_item;
2005 struct page *tree_page = NULL;
2006 struct stable_node *stable_node;
2008 unsigned int checksum;
2010 bool max_page_sharing_bypass = false;
2012 stable_node = page_stable_node(page);
2014 if (stable_node->head != &migrate_nodes &&
2015 get_kpfn_nid(READ_ONCE(stable_node->kpfn)) !=
2016 NUMA(stable_node->nid)) {
2017 stable_node_dup_del(stable_node);
2018 stable_node->head = &migrate_nodes;
2019 list_add(&stable_node->list, stable_node->head);
2021 if (stable_node->head != &migrate_nodes &&
2022 rmap_item->head == stable_node)
2025 * If it's a KSM fork, allow it to go over the sharing limit
2028 if (!is_page_sharing_candidate(stable_node))
2029 max_page_sharing_bypass = true;
2032 /* We first start with searching the page inside the stable tree */
2033 kpage = stable_tree_search(page);
2034 if (kpage == page && rmap_item->head == stable_node) {
2039 remove_rmap_item_from_tree(rmap_item);
2042 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
2045 * The page was successfully merged:
2046 * add its rmap_item to the stable tree.
2049 stable_tree_append(rmap_item, page_stable_node(kpage),
2050 max_page_sharing_bypass);
2058 * If the hash value of the page has changed from the last time
2059 * we calculated it, this page is changing frequently: therefore we
2060 * don't want to insert it in the unstable tree, and we don't want
2061 * to waste our time searching for something identical to it there.
2063 checksum = calc_checksum(page);
2064 if (rmap_item->oldchecksum != checksum) {
2065 rmap_item->oldchecksum = checksum;
2070 * Same checksum as an empty page. We attempt to merge it with the
2071 * appropriate zero page if the user enabled this via sysfs.
2073 if (ksm_use_zero_pages && (checksum == zero_checksum)) {
2074 struct vm_area_struct *vma;
2076 down_read(&mm->mmap_sem);
2077 vma = find_mergeable_vma(mm, rmap_item->address);
2079 err = try_to_merge_one_page(vma, page,
2080 ZERO_PAGE(rmap_item->address));
2083 * If the vma is out of date, we do not need to
2088 up_read(&mm->mmap_sem);
2090 * In case of failure, the page was not really empty, so we
2091 * need to continue. Otherwise we're done.
2097 unstable_tree_search_insert(rmap_item, page, &tree_page);
2098 if (tree_rmap_item) {
2101 kpage = try_to_merge_two_pages(rmap_item, page,
2102 tree_rmap_item, tree_page);
2104 * If both pages we tried to merge belong to the same compound
2105 * page, then we actually ended up increasing the reference
2106 * count of the same compound page twice, and split_huge_page
2108 * Here we set a flag if that happened, and we use it later to
2109 * try split_huge_page again. Since we call put_page right
2110 * afterwards, the reference count will be correct and
2111 * split_huge_page should succeed.
2113 split = PageTransCompound(page)
2114 && compound_head(page) == compound_head(tree_page);
2115 put_page(tree_page);
2118 * The pages were successfully merged: insert new
2119 * node in the stable tree and add both rmap_items.
2122 stable_node = stable_tree_insert(kpage);
2124 stable_tree_append(tree_rmap_item, stable_node,
2126 stable_tree_append(rmap_item, stable_node,
2132 * If we fail to insert the page into the stable tree,
2133 * we will have 2 virtual addresses that are pointing
2134 * to a ksm page left outside the stable tree,
2135 * in which case we need to break_cow on both.
2138 break_cow(tree_rmap_item);
2139 break_cow(rmap_item);
2143 * We are here if we tried to merge two pages and
2144 * failed because they both belonged to the same
2145 * compound page. We will split the page now, but no
2146 * merging will take place.
2147 * We do not want to add the cost of a full lock; if
2148 * the page is locked, it is better to skip it and
2149 * perhaps try again later.
2151 if (!trylock_page(page))
2153 split_huge_page(page);
2159 static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
2160 struct rmap_item **rmap_list,
2163 struct rmap_item *rmap_item;
2165 while (*rmap_list) {
2166 rmap_item = *rmap_list;
2167 if ((rmap_item->address & PAGE_MASK) == addr)
2169 if (rmap_item->address > addr)
2171 *rmap_list = rmap_item->rmap_list;
2172 remove_rmap_item_from_tree(rmap_item);
2173 free_rmap_item(rmap_item);
2176 rmap_item = alloc_rmap_item();
2178 /* It has already been zeroed */
2179 rmap_item->mm = mm_slot->mm;
2180 rmap_item->address = addr;
2181 rmap_item->rmap_list = *rmap_list;
2182 *rmap_list = rmap_item;
2187 static struct rmap_item *scan_get_next_rmap_item(struct page **page)
2189 struct mm_struct *mm;
2190 struct mm_slot *slot;
2191 struct vm_area_struct *vma;
2192 struct rmap_item *rmap_item;
2195 if (list_empty(&ksm_mm_head.mm_list))
2198 slot = ksm_scan.mm_slot;
2199 if (slot == &ksm_mm_head) {
2201 * A number of pages can hang around indefinitely on per-cpu
2202 * pagevecs, raised page count preventing write_protect_page
2203 * from merging them. Though it doesn't really matter much,
2204 * it is puzzling to see some stuck in pages_volatile until
2205 * other activity jostles them out, and they also prevented
2206 * LTP's KSM test from succeeding deterministically; so drain
2207 * them here (here rather than on entry to ksm_do_scan(),
2208 * so we don't IPI too often when pages_to_scan is set low).
2210 lru_add_drain_all();
2213 * Whereas stale stable_nodes on the stable_tree itself
2214 * get pruned in the regular course of stable_tree_search(),
2215 * those moved out to the migrate_nodes list can accumulate:
2216 * so prune them once before each full scan.
2218 if (!ksm_merge_across_nodes) {
2219 struct stable_node *stable_node, *next;
2222 list_for_each_entry_safe(stable_node, next,
2223 &migrate_nodes, list) {
2224 page = get_ksm_page(stable_node, false);
2231 for (nid = 0; nid < ksm_nr_node_ids; nid++)
2232 root_unstable_tree[nid] = RB_ROOT;
2234 spin_lock(&ksm_mmlist_lock);
2235 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
2236 ksm_scan.mm_slot = slot;
2237 spin_unlock(&ksm_mmlist_lock);
2239 * Although we tested list_empty() above, a racing __ksm_exit
2240 * of the last mm on the list may have removed it since then.
2242 if (slot == &ksm_mm_head)
2245 ksm_scan.address = 0;
2246 ksm_scan.rmap_list = &slot->rmap_list;
2250 down_read(&mm->mmap_sem);
2251 if (ksm_test_exit(mm))
2254 vma = find_vma(mm, ksm_scan.address);
2256 for (; vma; vma = vma->vm_next) {
2257 if (!(vma->vm_flags & VM_MERGEABLE))
2259 if (ksm_scan.address < vma->vm_start)
2260 ksm_scan.address = vma->vm_start;
2262 ksm_scan.address = vma->vm_end;
2264 while (ksm_scan.address < vma->vm_end) {
2265 if (ksm_test_exit(mm))
2267 *page = follow_page(vma, ksm_scan.address, FOLL_GET);
2268 if (IS_ERR_OR_NULL(*page)) {
2269 ksm_scan.address += PAGE_SIZE;
2273 if (PageAnon(*page)) {
2274 flush_anon_page(vma, *page, ksm_scan.address);
2275 flush_dcache_page(*page);
2276 rmap_item = get_next_rmap_item(slot,
2277 ksm_scan.rmap_list, ksm_scan.address);
2279 ksm_scan.rmap_list =
2280 &rmap_item->rmap_list;
2281 ksm_scan.address += PAGE_SIZE;
2284 up_read(&mm->mmap_sem);
2288 ksm_scan.address += PAGE_SIZE;
2293 if (ksm_test_exit(mm)) {
2294 ksm_scan.address = 0;
2295 ksm_scan.rmap_list = &slot->rmap_list;
2298 * Nuke all the rmap_items that are above this current rmap:
2299 * because there were no VM_MERGEABLE vmas with such addresses.
2301 remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
2303 spin_lock(&ksm_mmlist_lock);
2304 ksm_scan.mm_slot = list_entry(slot->mm_list.next,
2305 struct mm_slot, mm_list);
2306 if (ksm_scan.address == 0) {
2308 * We've completed a full scan of all vmas, holding mmap_sem
2309 * throughout, and found no VM_MERGEABLE: so do the same as
2310 * __ksm_exit does to remove this mm from all our lists now.
2311 * This applies either when cleaning up after __ksm_exit
2312 * (but beware: we can reach here even before __ksm_exit),
2313 * or when all VM_MERGEABLE areas have been unmapped (and
2314 * mmap_sem then protects against race with MADV_MERGEABLE).
2316 hash_del(&slot->link);
2317 list_del(&slot->mm_list);
2318 spin_unlock(&ksm_mmlist_lock);
2321 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
2322 up_read(&mm->mmap_sem);
2325 up_read(&mm->mmap_sem);
2327 * up_read(&mm->mmap_sem) first because after
2328 * spin_unlock(&ksm_mmlist_lock) run, the "mm" may
2329 * already have been freed under us by __ksm_exit()
2330 * because the "mm_slot" is still hashed and
2331 * ksm_scan.mm_slot doesn't point to it anymore.
2333 spin_unlock(&ksm_mmlist_lock);
2336 /* Repeat until we've completed scanning the whole list */
2337 slot = ksm_scan.mm_slot;
2338 if (slot != &ksm_mm_head)
2346 * ksm_do_scan - the ksm scanner main worker function.
2347 * @scan_npages - number of pages we want to scan before we return.
2349 static void ksm_do_scan(unsigned int scan_npages)
2351 struct rmap_item *rmap_item;
2352 struct page *uninitialized_var(page);
2354 while (scan_npages-- && likely(!freezing(current))) {
2356 rmap_item = scan_get_next_rmap_item(&page);
2359 cmp_and_merge_page(page, rmap_item);
2364 static int ksmd_should_run(void)
2366 return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
2369 static int ksm_scan_thread(void *nothing)
2372 set_user_nice(current, 5);
2374 while (!kthread_should_stop()) {
2375 mutex_lock(&ksm_thread_mutex);
2376 wait_while_offlining();
2377 if (ksmd_should_run())
2378 ksm_do_scan(ksm_thread_pages_to_scan);
2379 mutex_unlock(&ksm_thread_mutex);
2383 if (ksmd_should_run()) {
2384 schedule_timeout_interruptible(
2385 msecs_to_jiffies(ksm_thread_sleep_millisecs));
2387 wait_event_freezable(ksm_thread_wait,
2388 ksmd_should_run() || kthread_should_stop());
2394 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
2395 unsigned long end, int advice, unsigned long *vm_flags)
2397 struct mm_struct *mm = vma->vm_mm;
2401 case MADV_MERGEABLE:
2403 * Be somewhat over-protective for now!
2405 if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE |
2406 VM_PFNMAP | VM_IO | VM_DONTEXPAND |
2407 VM_HUGETLB | VM_MIXEDMAP))
2408 return 0; /* just ignore the advice */
2411 if (*vm_flags & VM_SAO)
2415 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
2416 err = __ksm_enter(mm);
2421 *vm_flags |= VM_MERGEABLE;
2424 case MADV_UNMERGEABLE:
2425 if (!(*vm_flags & VM_MERGEABLE))
2426 return 0; /* just ignore the advice */
2428 if (vma->anon_vma) {
2429 err = unmerge_ksm_pages(vma, start, end);
2434 *vm_flags &= ~VM_MERGEABLE;
2441 int __ksm_enter(struct mm_struct *mm)
2443 struct mm_slot *mm_slot;
2446 mm_slot = alloc_mm_slot();
2450 /* Check ksm_run too? Would need tighter locking */
2451 needs_wakeup = list_empty(&ksm_mm_head.mm_list);
2453 spin_lock(&ksm_mmlist_lock);
2454 insert_to_mm_slots_hash(mm, mm_slot);
2456 * When KSM_RUN_MERGE (or KSM_RUN_STOP),
2457 * insert just behind the scanning cursor, to let the area settle
2458 * down a little; when fork is followed by immediate exec, we don't
2459 * want ksmd to waste time setting up and tearing down an rmap_list.
2461 * But when KSM_RUN_UNMERGE, it's important to insert ahead of its
2462 * scanning cursor, otherwise KSM pages in newly forked mms will be
2463 * missed: then we might as well insert at the end of the list.
2465 if (ksm_run & KSM_RUN_UNMERGE)
2466 list_add_tail(&mm_slot->mm_list, &ksm_mm_head.mm_list);
2468 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
2469 spin_unlock(&ksm_mmlist_lock);
2471 set_bit(MMF_VM_MERGEABLE, &mm->flags);
2475 wake_up_interruptible(&ksm_thread_wait);
2480 void __ksm_exit(struct mm_struct *mm)
2482 struct mm_slot *mm_slot;
2483 int easy_to_free = 0;
2486 * This process is exiting: if it's straightforward (as is the
2487 * case when ksmd was never running), free mm_slot immediately.
2488 * But if it's at the cursor or has rmap_items linked to it, use
2489 * mmap_sem to synchronize with any break_cows before pagetables
2490 * are freed, and leave the mm_slot on the list for ksmd to free.
2491 * Beware: ksm may already have noticed it exiting and freed the slot.
2494 spin_lock(&ksm_mmlist_lock);
2495 mm_slot = get_mm_slot(mm);
2496 if (mm_slot && ksm_scan.mm_slot != mm_slot) {
2497 if (!mm_slot->rmap_list) {
2498 hash_del(&mm_slot->link);
2499 list_del(&mm_slot->mm_list);
2502 list_move(&mm_slot->mm_list,
2503 &ksm_scan.mm_slot->mm_list);
2506 spin_unlock(&ksm_mmlist_lock);
2509 free_mm_slot(mm_slot);
2510 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
2512 } else if (mm_slot) {
2513 down_write(&mm->mmap_sem);
2514 up_write(&mm->mmap_sem);
2518 struct page *ksm_might_need_to_copy(struct page *page,
2519 struct vm_area_struct *vma, unsigned long address)
2521 struct anon_vma *anon_vma = page_anon_vma(page);
2522 struct page *new_page;
2524 if (PageKsm(page)) {
2525 if (page_stable_node(page) &&
2526 !(ksm_run & KSM_RUN_UNMERGE))
2527 return page; /* no need to copy it */
2528 } else if (!anon_vma) {
2529 return page; /* no need to copy it */
2530 } else if (anon_vma->root == vma->anon_vma->root &&
2531 page->index == linear_page_index(vma, address)) {
2532 return page; /* still no need to copy it */
2534 if (!PageUptodate(page))
2535 return page; /* let do_swap_page report the error */
2537 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2539 copy_user_highpage(new_page, page, address, vma);
2541 SetPageDirty(new_page);
2542 __SetPageUptodate(new_page);
2543 __SetPageLocked(new_page);
2549 void rmap_walk_ksm(struct page *page, struct rmap_walk_control *rwc)
2551 struct stable_node *stable_node;
2552 struct rmap_item *rmap_item;
2553 int search_new_forks = 0;
2555 VM_BUG_ON_PAGE(!PageKsm(page), page);
2558 * Rely on the page lock to protect against concurrent modifications
2559 * to that page's node of the stable tree.
2561 VM_BUG_ON_PAGE(!PageLocked(page), page);
2563 stable_node = page_stable_node(page);
2567 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
2568 struct anon_vma *anon_vma = rmap_item->anon_vma;
2569 struct anon_vma_chain *vmac;
2570 struct vm_area_struct *vma;
2573 anon_vma_lock_read(anon_vma);
2574 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
2581 /* Ignore the stable/unstable/sqnr flags */
2582 addr = rmap_item->address & ~KSM_FLAG_MASK;
2584 if (addr < vma->vm_start || addr >= vma->vm_end)
2587 * Initially we examine only the vma which covers this
2588 * rmap_item; but later, if there is still work to do,
2589 * we examine covering vmas in other mms: in case they
2590 * were forked from the original since ksmd passed.
2592 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
2595 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2598 if (!rwc->rmap_one(page, vma, addr, rwc->arg)) {
2599 anon_vma_unlock_read(anon_vma);
2602 if (rwc->done && rwc->done(page)) {
2603 anon_vma_unlock_read(anon_vma);
2607 anon_vma_unlock_read(anon_vma);
2609 if (!search_new_forks++)
2613 #ifdef CONFIG_MIGRATION
2614 void ksm_migrate_page(struct page *newpage, struct page *oldpage)
2616 struct stable_node *stable_node;
2618 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
2619 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
2620 VM_BUG_ON_PAGE(newpage->mapping != oldpage->mapping, newpage);
2622 stable_node = page_stable_node(newpage);
2624 VM_BUG_ON_PAGE(stable_node->kpfn != page_to_pfn(oldpage), oldpage);
2625 stable_node->kpfn = page_to_pfn(newpage);
2627 * newpage->mapping was set in advance; now we need smp_wmb()
2628 * to make sure that the new stable_node->kpfn is visible
2629 * to get_ksm_page() before it can see that oldpage->mapping
2630 * has gone stale (or that PageSwapCache has been cleared).
2633 set_page_stable_node(oldpage, NULL);
2636 #endif /* CONFIG_MIGRATION */
2638 #ifdef CONFIG_MEMORY_HOTREMOVE
2639 static void wait_while_offlining(void)
2641 while (ksm_run & KSM_RUN_OFFLINE) {
2642 mutex_unlock(&ksm_thread_mutex);
2643 wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE),
2644 TASK_UNINTERRUPTIBLE);
2645 mutex_lock(&ksm_thread_mutex);
2649 static bool stable_node_dup_remove_range(struct stable_node *stable_node,
2650 unsigned long start_pfn,
2651 unsigned long end_pfn)
2653 if (stable_node->kpfn >= start_pfn &&
2654 stable_node->kpfn < end_pfn) {
2656 * Don't get_ksm_page, page has already gone:
2657 * which is why we keep kpfn instead of page*
2659 remove_node_from_stable_tree(stable_node);
2665 static bool stable_node_chain_remove_range(struct stable_node *stable_node,
2666 unsigned long start_pfn,
2667 unsigned long end_pfn,
2668 struct rb_root *root)
2670 struct stable_node *dup;
2671 struct hlist_node *hlist_safe;
2673 if (!is_stable_node_chain(stable_node)) {
2674 VM_BUG_ON(is_stable_node_dup(stable_node));
2675 return stable_node_dup_remove_range(stable_node, start_pfn,
2679 hlist_for_each_entry_safe(dup, hlist_safe,
2680 &stable_node->hlist, hlist_dup) {
2681 VM_BUG_ON(!is_stable_node_dup(dup));
2682 stable_node_dup_remove_range(dup, start_pfn, end_pfn);
2684 if (hlist_empty(&stable_node->hlist)) {
2685 free_stable_node_chain(stable_node, root);
2686 return true; /* notify caller that tree was rebalanced */
2691 static void ksm_check_stable_tree(unsigned long start_pfn,
2692 unsigned long end_pfn)
2694 struct stable_node *stable_node, *next;
2695 struct rb_node *node;
2698 for (nid = 0; nid < ksm_nr_node_ids; nid++) {
2699 node = rb_first(root_stable_tree + nid);
2701 stable_node = rb_entry(node, struct stable_node, node);
2702 if (stable_node_chain_remove_range(stable_node,
2706 node = rb_first(root_stable_tree + nid);
2708 node = rb_next(node);
2712 list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
2713 if (stable_node->kpfn >= start_pfn &&
2714 stable_node->kpfn < end_pfn)
2715 remove_node_from_stable_tree(stable_node);
2720 static int ksm_memory_callback(struct notifier_block *self,
2721 unsigned long action, void *arg)
2723 struct memory_notify *mn = arg;
2726 case MEM_GOING_OFFLINE:
2728 * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items()
2729 * and remove_all_stable_nodes() while memory is going offline:
2730 * it is unsafe for them to touch the stable tree at this time.
2731 * But unmerge_ksm_pages(), rmap lookups and other entry points
2732 * which do not need the ksm_thread_mutex are all safe.
2734 mutex_lock(&ksm_thread_mutex);
2735 ksm_run |= KSM_RUN_OFFLINE;
2736 mutex_unlock(&ksm_thread_mutex);
2741 * Most of the work is done by page migration; but there might
2742 * be a few stable_nodes left over, still pointing to struct
2743 * pages which have been offlined: prune those from the tree,
2744 * otherwise get_ksm_page() might later try to access a
2745 * non-existent struct page.
2747 ksm_check_stable_tree(mn->start_pfn,
2748 mn->start_pfn + mn->nr_pages);
2751 case MEM_CANCEL_OFFLINE:
2752 mutex_lock(&ksm_thread_mutex);
2753 ksm_run &= ~KSM_RUN_OFFLINE;
2754 mutex_unlock(&ksm_thread_mutex);
2756 smp_mb(); /* wake_up_bit advises this */
2757 wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE));
2763 static void wait_while_offlining(void)
2766 #endif /* CONFIG_MEMORY_HOTREMOVE */
2770 * This all compiles without CONFIG_SYSFS, but is a waste of space.
2773 #define KSM_ATTR_RO(_name) \
2774 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2775 #define KSM_ATTR(_name) \
2776 static struct kobj_attribute _name##_attr = \
2777 __ATTR(_name, 0644, _name##_show, _name##_store)
2779 static ssize_t sleep_millisecs_show(struct kobject *kobj,
2780 struct kobj_attribute *attr, char *buf)
2782 return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
2785 static ssize_t sleep_millisecs_store(struct kobject *kobj,
2786 struct kobj_attribute *attr,
2787 const char *buf, size_t count)
2789 unsigned long msecs;
2792 err = kstrtoul(buf, 10, &msecs);
2793 if (err || msecs > UINT_MAX)
2796 ksm_thread_sleep_millisecs = msecs;
2800 KSM_ATTR(sleep_millisecs);
2802 static ssize_t pages_to_scan_show(struct kobject *kobj,
2803 struct kobj_attribute *attr, char *buf)
2805 return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
2808 static ssize_t pages_to_scan_store(struct kobject *kobj,
2809 struct kobj_attribute *attr,
2810 const char *buf, size_t count)
2813 unsigned long nr_pages;
2815 err = kstrtoul(buf, 10, &nr_pages);
2816 if (err || nr_pages > UINT_MAX)
2819 ksm_thread_pages_to_scan = nr_pages;
2823 KSM_ATTR(pages_to_scan);
2825 static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
2828 return sprintf(buf, "%lu\n", ksm_run);
2831 static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
2832 const char *buf, size_t count)
2835 unsigned long flags;
2837 err = kstrtoul(buf, 10, &flags);
2838 if (err || flags > UINT_MAX)
2840 if (flags > KSM_RUN_UNMERGE)
2844 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
2845 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
2846 * breaking COW to free the pages_shared (but leaves mm_slots
2847 * on the list for when ksmd may be set running again).
2850 mutex_lock(&ksm_thread_mutex);
2851 wait_while_offlining();
2852 if (ksm_run != flags) {
2854 if (flags & KSM_RUN_UNMERGE) {
2855 set_current_oom_origin();
2856 err = unmerge_and_remove_all_rmap_items();
2857 clear_current_oom_origin();
2859 ksm_run = KSM_RUN_STOP;
2864 mutex_unlock(&ksm_thread_mutex);
2866 if (flags & KSM_RUN_MERGE)
2867 wake_up_interruptible(&ksm_thread_wait);
2874 static ssize_t merge_across_nodes_show(struct kobject *kobj,
2875 struct kobj_attribute *attr, char *buf)
2877 return sprintf(buf, "%u\n", ksm_merge_across_nodes);
2880 static ssize_t merge_across_nodes_store(struct kobject *kobj,
2881 struct kobj_attribute *attr,
2882 const char *buf, size_t count)
2887 err = kstrtoul(buf, 10, &knob);
2893 mutex_lock(&ksm_thread_mutex);
2894 wait_while_offlining();
2895 if (ksm_merge_across_nodes != knob) {
2896 if (ksm_pages_shared || remove_all_stable_nodes())
2898 else if (root_stable_tree == one_stable_tree) {
2899 struct rb_root *buf;
2901 * This is the first time that we switch away from the
2902 * default of merging across nodes: must now allocate
2903 * a buffer to hold as many roots as may be needed.
2904 * Allocate stable and unstable together:
2905 * MAXSMP NODES_SHIFT 10 will use 16kB.
2907 buf = kcalloc(nr_node_ids + nr_node_ids, sizeof(*buf),
2909 /* Let us assume that RB_ROOT is NULL is zero */
2913 root_stable_tree = buf;
2914 root_unstable_tree = buf + nr_node_ids;
2915 /* Stable tree is empty but not the unstable */
2916 root_unstable_tree[0] = one_unstable_tree[0];
2920 ksm_merge_across_nodes = knob;
2921 ksm_nr_node_ids = knob ? 1 : nr_node_ids;
2924 mutex_unlock(&ksm_thread_mutex);
2926 return err ? err : count;
2928 KSM_ATTR(merge_across_nodes);
2931 static ssize_t use_zero_pages_show(struct kobject *kobj,
2932 struct kobj_attribute *attr, char *buf)
2934 return sprintf(buf, "%u\n", ksm_use_zero_pages);
2936 static ssize_t use_zero_pages_store(struct kobject *kobj,
2937 struct kobj_attribute *attr,
2938 const char *buf, size_t count)
2943 err = kstrtobool(buf, &value);
2947 ksm_use_zero_pages = value;
2951 KSM_ATTR(use_zero_pages);
2953 static ssize_t max_page_sharing_show(struct kobject *kobj,
2954 struct kobj_attribute *attr, char *buf)
2956 return sprintf(buf, "%u\n", ksm_max_page_sharing);
2959 static ssize_t max_page_sharing_store(struct kobject *kobj,
2960 struct kobj_attribute *attr,
2961 const char *buf, size_t count)
2966 err = kstrtoint(buf, 10, &knob);
2970 * When a KSM page is created it is shared by 2 mappings. This
2971 * being a signed comparison, it implicitly verifies it's not
2977 if (READ_ONCE(ksm_max_page_sharing) == knob)
2980 mutex_lock(&ksm_thread_mutex);
2981 wait_while_offlining();
2982 if (ksm_max_page_sharing != knob) {
2983 if (ksm_pages_shared || remove_all_stable_nodes())
2986 ksm_max_page_sharing = knob;
2988 mutex_unlock(&ksm_thread_mutex);
2990 return err ? err : count;
2992 KSM_ATTR(max_page_sharing);
2994 static ssize_t pages_shared_show(struct kobject *kobj,
2995 struct kobj_attribute *attr, char *buf)
2997 return sprintf(buf, "%lu\n", ksm_pages_shared);
2999 KSM_ATTR_RO(pages_shared);
3001 static ssize_t pages_sharing_show(struct kobject *kobj,
3002 struct kobj_attribute *attr, char *buf)
3004 return sprintf(buf, "%lu\n", ksm_pages_sharing);
3006 KSM_ATTR_RO(pages_sharing);
3008 static ssize_t pages_unshared_show(struct kobject *kobj,
3009 struct kobj_attribute *attr, char *buf)
3011 return sprintf(buf, "%lu\n", ksm_pages_unshared);
3013 KSM_ATTR_RO(pages_unshared);
3015 static ssize_t pages_volatile_show(struct kobject *kobj,
3016 struct kobj_attribute *attr, char *buf)
3018 long ksm_pages_volatile;
3020 ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
3021 - ksm_pages_sharing - ksm_pages_unshared;
3023 * It was not worth any locking to calculate that statistic,
3024 * but it might therefore sometimes be negative: conceal that.
3026 if (ksm_pages_volatile < 0)
3027 ksm_pages_volatile = 0;
3028 return sprintf(buf, "%ld\n", ksm_pages_volatile);
3030 KSM_ATTR_RO(pages_volatile);
3032 static ssize_t stable_node_dups_show(struct kobject *kobj,
3033 struct kobj_attribute *attr, char *buf)
3035 return sprintf(buf, "%lu\n", ksm_stable_node_dups);
3037 KSM_ATTR_RO(stable_node_dups);
3039 static ssize_t stable_node_chains_show(struct kobject *kobj,
3040 struct kobj_attribute *attr, char *buf)
3042 return sprintf(buf, "%lu\n", ksm_stable_node_chains);
3044 KSM_ATTR_RO(stable_node_chains);
3047 stable_node_chains_prune_millisecs_show(struct kobject *kobj,
3048 struct kobj_attribute *attr,
3051 return sprintf(buf, "%u\n", ksm_stable_node_chains_prune_millisecs);
3055 stable_node_chains_prune_millisecs_store(struct kobject *kobj,
3056 struct kobj_attribute *attr,
3057 const char *buf, size_t count)
3059 unsigned long msecs;
3062 err = kstrtoul(buf, 10, &msecs);
3063 if (err || msecs > UINT_MAX)
3066 ksm_stable_node_chains_prune_millisecs = msecs;
3070 KSM_ATTR(stable_node_chains_prune_millisecs);
3072 static ssize_t full_scans_show(struct kobject *kobj,
3073 struct kobj_attribute *attr, char *buf)
3075 return sprintf(buf, "%lu\n", ksm_scan.seqnr);
3077 KSM_ATTR_RO(full_scans);
3079 static struct attribute *ksm_attrs[] = {
3080 &sleep_millisecs_attr.attr,
3081 &pages_to_scan_attr.attr,
3083 &pages_shared_attr.attr,
3084 &pages_sharing_attr.attr,
3085 &pages_unshared_attr.attr,
3086 &pages_volatile_attr.attr,
3087 &full_scans_attr.attr,
3089 &merge_across_nodes_attr.attr,
3091 &max_page_sharing_attr.attr,
3092 &stable_node_chains_attr.attr,
3093 &stable_node_dups_attr.attr,
3094 &stable_node_chains_prune_millisecs_attr.attr,
3095 &use_zero_pages_attr.attr,
3099 static const struct attribute_group ksm_attr_group = {
3103 #endif /* CONFIG_SYSFS */
3105 static int __init ksm_init(void)
3107 struct task_struct *ksm_thread;
3110 /* The correct value depends on page size and endianness */
3111 zero_checksum = calc_checksum(ZERO_PAGE(0));
3112 /* Default to false for backwards compatibility */
3113 ksm_use_zero_pages = false;
3115 err = ksm_slab_init();
3119 ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
3120 if (IS_ERR(ksm_thread)) {
3121 pr_err("ksm: creating kthread failed\n");
3122 err = PTR_ERR(ksm_thread);
3127 err = sysfs_create_group(mm_kobj, &ksm_attr_group);
3129 pr_err("ksm: register sysfs failed\n");
3130 kthread_stop(ksm_thread);
3134 ksm_run = KSM_RUN_MERGE; /* no way for user to start it */
3136 #endif /* CONFIG_SYSFS */
3138 #ifdef CONFIG_MEMORY_HOTREMOVE
3139 /* There is no significance to this priority 100 */
3140 hotplug_memory_notifier(ksm_memory_callback, 100);
3149 subsys_initcall(ksm_init);