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)
57 * A few notes about the KSM scanning process,
58 * to make it easier to understand the data structures below:
60 * In order to reduce excessive scanning, KSM sorts the memory pages by their
61 * contents into a data structure that holds pointers to the pages' locations.
63 * Since the contents of the pages may change at any moment, KSM cannot just
64 * insert the pages into a normal sorted tree and expect it to find anything.
65 * Therefore KSM uses two data structures - the stable and the unstable tree.
67 * The stable tree holds pointers to all the merged pages (ksm pages), sorted
68 * by their contents. Because each such page is write-protected, searching on
69 * this tree is fully assured to be working (except when pages are unmapped),
70 * and therefore this tree is called the stable tree.
72 * The stable tree node includes information required for reverse
73 * mapping from a KSM page to virtual addresses that map this page.
75 * In order to avoid large latencies of the rmap walks on KSM pages,
76 * KSM maintains two types of nodes in the stable tree:
78 * * the regular nodes that keep the reverse mapping structures in a
80 * * the "chains" that link nodes ("dups") that represent the same
81 * write protected memory content, but each "dup" corresponds to a
82 * different KSM page copy of that content
84 * Internally, the regular nodes, "dups" and "chains" are represented
85 * using the same :c:type:`struct stable_node` structure.
87 * In addition to the stable tree, KSM uses a second data structure called the
88 * unstable tree: this tree holds pointers to pages which have been found to
89 * be "unchanged for a period of time". The unstable tree sorts these pages
90 * by their contents, but since they are not write-protected, KSM cannot rely
91 * upon the unstable tree to work correctly - the unstable tree is liable to
92 * be corrupted as its contents are modified, and so it is called unstable.
94 * KSM solves this problem by several techniques:
96 * 1) The unstable tree is flushed every time KSM completes scanning all
97 * memory areas, and then the tree is rebuilt again from the beginning.
98 * 2) KSM will only insert into the unstable tree, pages whose hash value
99 * has not changed since the previous scan of all memory areas.
100 * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
101 * colors of the nodes and not on their contents, assuring that even when
102 * the tree gets "corrupted" it won't get out of balance, so scanning time
103 * remains the same (also, searching and inserting nodes in an rbtree uses
104 * the same algorithm, so we have no overhead when we flush and rebuild).
105 * 4) KSM never flushes the stable tree, which means that even if it were to
106 * take 10 attempts to find a page in the unstable tree, once it is found,
107 * it is secured in the stable tree. (When we scan a new page, we first
108 * compare it against the stable tree, and then against the unstable tree.)
110 * If the merge_across_nodes tunable is unset, then KSM maintains multiple
111 * stable trees and multiple unstable trees: one of each for each NUMA node.
115 * struct mm_slot - ksm information per mm that is being scanned
116 * @link: link to the mm_slots hash list
117 * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
118 * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
119 * @mm: the mm that this information is valid for
122 struct hlist_node link;
123 struct list_head mm_list;
124 struct rmap_item *rmap_list;
125 struct mm_struct *mm;
129 * struct ksm_scan - cursor for scanning
130 * @mm_slot: the current mm_slot we are scanning
131 * @address: the next address inside that to be scanned
132 * @rmap_list: link to the next rmap to be scanned in the rmap_list
133 * @seqnr: count of completed full scans (needed when removing unstable node)
135 * There is only the one ksm_scan instance of this cursor structure.
138 struct mm_slot *mm_slot;
139 unsigned long address;
140 struct rmap_item **rmap_list;
145 * struct stable_node - node of the stable rbtree
146 * @node: rb node of this ksm page in the stable tree
147 * @head: (overlaying parent) &migrate_nodes indicates temporarily on that list
148 * @hlist_dup: linked into the stable_node->hlist with a stable_node chain
149 * @list: linked into migrate_nodes, pending placement in the proper node tree
150 * @hlist: hlist head of rmap_items using this ksm page
151 * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid)
152 * @chain_prune_time: time of the last full garbage collection
153 * @rmap_hlist_len: number of rmap_item entries in hlist or STABLE_NODE_CHAIN
154 * @nid: NUMA node id of stable tree in which linked (may not match kpfn)
158 struct rb_node node; /* when node of stable tree */
159 struct { /* when listed for migration */
160 struct list_head *head;
162 struct hlist_node hlist_dup;
163 struct list_head list;
167 struct hlist_head hlist;
170 unsigned long chain_prune_time;
173 * STABLE_NODE_CHAIN can be any negative number in
174 * rmap_hlist_len negative range, but better not -1 to be able
175 * to reliably detect underflows.
177 #define STABLE_NODE_CHAIN -1024
185 * struct rmap_item - reverse mapping item for virtual addresses
186 * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
187 * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
188 * @nid: NUMA node id of unstable tree in which linked (may not match page)
189 * @mm: the memory structure this rmap_item is pointing into
190 * @address: the virtual address this rmap_item tracks (+ flags in low bits)
191 * @oldchecksum: previous checksum of the page at that virtual address
192 * @node: rb node of this rmap_item in the unstable tree
193 * @head: pointer to stable_node heading this list in the stable tree
194 * @hlist: link into hlist of rmap_items hanging off that stable_node
197 struct rmap_item *rmap_list;
199 struct anon_vma *anon_vma; /* when stable */
201 int nid; /* when node of unstable tree */
204 struct mm_struct *mm;
205 unsigned long address; /* + low bits used for flags below */
206 unsigned int oldchecksum; /* when unstable */
208 struct rb_node node; /* when node of unstable tree */
209 struct { /* when listed from stable tree */
210 struct stable_node *head;
211 struct hlist_node hlist;
216 #define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */
217 #define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */
218 #define STABLE_FLAG 0x200 /* is listed from the stable tree */
220 /* The stable and unstable tree heads */
221 static struct rb_root one_stable_tree[1] = { RB_ROOT };
222 static struct rb_root one_unstable_tree[1] = { RB_ROOT };
223 static struct rb_root *root_stable_tree = one_stable_tree;
224 static struct rb_root *root_unstable_tree = one_unstable_tree;
226 /* Recently migrated nodes of stable tree, pending proper placement */
227 static LIST_HEAD(migrate_nodes);
228 #define STABLE_NODE_DUP_HEAD ((struct list_head *)&migrate_nodes.prev)
230 #define MM_SLOTS_HASH_BITS 10
231 static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
233 static struct mm_slot ksm_mm_head = {
234 .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
236 static struct ksm_scan ksm_scan = {
237 .mm_slot = &ksm_mm_head,
240 static struct kmem_cache *rmap_item_cache;
241 static struct kmem_cache *stable_node_cache;
242 static struct kmem_cache *mm_slot_cache;
244 /* The number of nodes in the stable tree */
245 static unsigned long ksm_pages_shared;
247 /* The number of page slots additionally sharing those nodes */
248 static unsigned long ksm_pages_sharing;
250 /* The number of nodes in the unstable tree */
251 static unsigned long ksm_pages_unshared;
253 /* The number of rmap_items in use: to calculate pages_volatile */
254 static unsigned long ksm_rmap_items;
256 /* The number of stable_node chains */
257 static unsigned long ksm_stable_node_chains;
259 /* The number of stable_node dups linked to the stable_node chains */
260 static unsigned long ksm_stable_node_dups;
262 /* Delay in pruning stale stable_node_dups in the stable_node_chains */
263 static int ksm_stable_node_chains_prune_millisecs = 2000;
265 /* Maximum number of page slots sharing a stable node */
266 static int ksm_max_page_sharing = 256;
268 /* Number of pages ksmd should scan in one batch */
269 static unsigned int ksm_thread_pages_to_scan = 100;
271 /* Milliseconds ksmd should sleep between batches */
272 static unsigned int ksm_thread_sleep_millisecs = 20;
274 /* Checksum of an empty (zeroed) page */
275 static unsigned int zero_checksum __read_mostly;
277 /* Whether to merge empty (zeroed) pages with actual zero pages */
278 static bool ksm_use_zero_pages __read_mostly;
281 /* Zeroed when merging across nodes is not allowed */
282 static unsigned int ksm_merge_across_nodes = 1;
283 static int ksm_nr_node_ids = 1;
285 #define ksm_merge_across_nodes 1U
286 #define ksm_nr_node_ids 1
289 #define KSM_RUN_STOP 0
290 #define KSM_RUN_MERGE 1
291 #define KSM_RUN_UNMERGE 2
292 #define KSM_RUN_OFFLINE 4
293 static unsigned long ksm_run = KSM_RUN_STOP;
294 static void wait_while_offlining(void);
296 static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
297 static DEFINE_MUTEX(ksm_thread_mutex);
298 static DEFINE_SPINLOCK(ksm_mmlist_lock);
300 #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
301 sizeof(struct __struct), __alignof__(struct __struct),\
304 static int __init ksm_slab_init(void)
306 rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
307 if (!rmap_item_cache)
310 stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
311 if (!stable_node_cache)
314 mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
321 kmem_cache_destroy(stable_node_cache);
323 kmem_cache_destroy(rmap_item_cache);
328 static void __init ksm_slab_free(void)
330 kmem_cache_destroy(mm_slot_cache);
331 kmem_cache_destroy(stable_node_cache);
332 kmem_cache_destroy(rmap_item_cache);
333 mm_slot_cache = NULL;
336 static __always_inline bool is_stable_node_chain(struct stable_node *chain)
338 return chain->rmap_hlist_len == STABLE_NODE_CHAIN;
341 static __always_inline bool is_stable_node_dup(struct stable_node *dup)
343 return dup->head == STABLE_NODE_DUP_HEAD;
346 static inline void stable_node_chain_add_dup(struct stable_node *dup,
347 struct stable_node *chain)
349 VM_BUG_ON(is_stable_node_dup(dup));
350 dup->head = STABLE_NODE_DUP_HEAD;
351 VM_BUG_ON(!is_stable_node_chain(chain));
352 hlist_add_head(&dup->hlist_dup, &chain->hlist);
353 ksm_stable_node_dups++;
356 static inline void __stable_node_dup_del(struct stable_node *dup)
358 VM_BUG_ON(!is_stable_node_dup(dup));
359 hlist_del(&dup->hlist_dup);
360 ksm_stable_node_dups--;
363 static inline void stable_node_dup_del(struct stable_node *dup)
365 VM_BUG_ON(is_stable_node_chain(dup));
366 if (is_stable_node_dup(dup))
367 __stable_node_dup_del(dup);
369 rb_erase(&dup->node, root_stable_tree + NUMA(dup->nid));
370 #ifdef CONFIG_DEBUG_VM
375 static inline struct rmap_item *alloc_rmap_item(void)
377 struct rmap_item *rmap_item;
379 rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL |
380 __GFP_NORETRY | __GFP_NOWARN);
386 static inline void free_rmap_item(struct rmap_item *rmap_item)
389 rmap_item->mm = NULL; /* debug safety */
390 kmem_cache_free(rmap_item_cache, rmap_item);
393 static inline struct stable_node *alloc_stable_node(void)
396 * The allocation can take too long with GFP_KERNEL when memory is under
397 * pressure, which may lead to hung task warnings. Adding __GFP_HIGH
398 * grants access to memory reserves, helping to avoid this problem.
400 return kmem_cache_alloc(stable_node_cache, GFP_KERNEL | __GFP_HIGH);
403 static inline void free_stable_node(struct stable_node *stable_node)
405 VM_BUG_ON(stable_node->rmap_hlist_len &&
406 !is_stable_node_chain(stable_node));
407 kmem_cache_free(stable_node_cache, stable_node);
410 static inline struct mm_slot *alloc_mm_slot(void)
412 if (!mm_slot_cache) /* initialization failed */
414 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
417 static inline void free_mm_slot(struct mm_slot *mm_slot)
419 kmem_cache_free(mm_slot_cache, mm_slot);
422 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
424 struct mm_slot *slot;
426 hash_for_each_possible(mm_slots_hash, slot, link, (unsigned long)mm)
433 static void insert_to_mm_slots_hash(struct mm_struct *mm,
434 struct mm_slot *mm_slot)
437 hash_add(mm_slots_hash, &mm_slot->link, (unsigned long)mm);
441 * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
442 * page tables after it has passed through ksm_exit() - which, if necessary,
443 * takes mmap_sem briefly to serialize against them. ksm_exit() does not set
444 * a special flag: they can just back out as soon as mm_users goes to zero.
445 * ksm_test_exit() is used throughout to make this test for exit: in some
446 * places for correctness, in some places just to avoid unnecessary work.
448 static inline bool ksm_test_exit(struct mm_struct *mm)
450 return atomic_read(&mm->mm_users) == 0;
454 * We use break_ksm to break COW on a ksm page: it's a stripped down
456 * if (get_user_pages(addr, 1, 1, 1, &page, NULL) == 1)
459 * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
460 * in case the application has unmapped and remapped mm,addr meanwhile.
461 * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP
462 * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
464 * FAULT_FLAG/FOLL_REMOTE are because we do this outside the context
465 * of the process that owns 'vma'. We also do not want to enforce
466 * protection keys here anyway.
468 static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
475 page = follow_page(vma, addr,
476 FOLL_GET | FOLL_MIGRATION | FOLL_REMOTE);
477 if (IS_ERR_OR_NULL(page))
480 ret = handle_mm_fault(vma, addr,
481 FAULT_FLAG_WRITE | FAULT_FLAG_REMOTE);
483 ret = VM_FAULT_WRITE;
485 } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV | VM_FAULT_OOM)));
487 * We must loop because handle_mm_fault() may back out if there's
488 * any difficulty e.g. if pte accessed bit gets updated concurrently.
490 * VM_FAULT_WRITE is what we have been hoping for: it indicates that
491 * COW has been broken, even if the vma does not permit VM_WRITE;
492 * but note that a concurrent fault might break PageKsm for us.
494 * VM_FAULT_SIGBUS could occur if we race with truncation of the
495 * backing file, which also invalidates anonymous pages: that's
496 * okay, that truncation will have unmapped the PageKsm for us.
498 * VM_FAULT_OOM: at the time of writing (late July 2009), setting
499 * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
500 * current task has TIF_MEMDIE set, and will be OOM killed on return
501 * to user; and ksmd, having no mm, would never be chosen for that.
503 * But if the mm is in a limited mem_cgroup, then the fault may fail
504 * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
505 * even ksmd can fail in this way - though it's usually breaking ksm
506 * just to undo a merge it made a moment before, so unlikely to oom.
508 * That's a pity: we might therefore have more kernel pages allocated
509 * than we're counting as nodes in the stable tree; but ksm_do_scan
510 * will retry to break_cow on each pass, so should recover the page
511 * in due course. The important thing is to not let VM_MERGEABLE
512 * be cleared while any such pages might remain in the area.
514 return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
517 static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm,
520 struct vm_area_struct *vma;
521 if (ksm_test_exit(mm))
523 vma = find_vma(mm, addr);
524 if (!vma || vma->vm_start > addr)
526 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
531 static void break_cow(struct rmap_item *rmap_item)
533 struct mm_struct *mm = rmap_item->mm;
534 unsigned long addr = rmap_item->address;
535 struct vm_area_struct *vma;
538 * It is not an accident that whenever we want to break COW
539 * to undo, we also need to drop a reference to the anon_vma.
541 put_anon_vma(rmap_item->anon_vma);
543 down_read(&mm->mmap_sem);
544 vma = find_mergeable_vma(mm, addr);
546 break_ksm(vma, addr);
547 up_read(&mm->mmap_sem);
550 static struct page *get_mergeable_page(struct rmap_item *rmap_item)
552 struct mm_struct *mm = rmap_item->mm;
553 unsigned long addr = rmap_item->address;
554 struct vm_area_struct *vma;
557 down_read(&mm->mmap_sem);
558 vma = find_mergeable_vma(mm, addr);
562 page = follow_page(vma, addr, FOLL_GET);
563 if (IS_ERR_OR_NULL(page))
565 if (PageAnon(page)) {
566 flush_anon_page(vma, page, addr);
567 flush_dcache_page(page);
573 up_read(&mm->mmap_sem);
578 * This helper is used for getting right index into array of tree roots.
579 * When merge_across_nodes knob is set to 1, there are only two rb-trees for
580 * stable and unstable pages from all nodes with roots in index 0. Otherwise,
581 * every node has its own stable and unstable tree.
583 static inline int get_kpfn_nid(unsigned long kpfn)
585 return ksm_merge_across_nodes ? 0 : NUMA(pfn_to_nid(kpfn));
588 static struct stable_node *alloc_stable_node_chain(struct stable_node *dup,
589 struct rb_root *root)
591 struct stable_node *chain = alloc_stable_node();
592 VM_BUG_ON(is_stable_node_chain(dup));
594 INIT_HLIST_HEAD(&chain->hlist);
595 chain->chain_prune_time = jiffies;
596 chain->rmap_hlist_len = STABLE_NODE_CHAIN;
597 #if defined (CONFIG_DEBUG_VM) && defined(CONFIG_NUMA)
598 chain->nid = -1; /* debug */
600 ksm_stable_node_chains++;
603 * Put the stable node chain in the first dimension of
604 * the stable tree and at the same time remove the old
607 rb_replace_node(&dup->node, &chain->node, root);
610 * Move the old stable node to the second dimension
611 * queued in the hlist_dup. The invariant is that all
612 * dup stable_nodes in the chain->hlist point to pages
613 * that are wrprotected and have the exact same
616 stable_node_chain_add_dup(dup, chain);
621 static inline void free_stable_node_chain(struct stable_node *chain,
622 struct rb_root *root)
624 rb_erase(&chain->node, root);
625 free_stable_node(chain);
626 ksm_stable_node_chains--;
629 static void remove_node_from_stable_tree(struct stable_node *stable_node)
631 struct rmap_item *rmap_item;
633 /* check it's not STABLE_NODE_CHAIN or negative */
634 BUG_ON(stable_node->rmap_hlist_len < 0);
636 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
637 if (rmap_item->hlist.next)
641 VM_BUG_ON(stable_node->rmap_hlist_len <= 0);
642 stable_node->rmap_hlist_len--;
643 put_anon_vma(rmap_item->anon_vma);
644 rmap_item->address &= PAGE_MASK;
649 * We need the second aligned pointer of the migrate_nodes
650 * list_head to stay clear from the rb_parent_color union
651 * (aligned and different than any node) and also different
652 * from &migrate_nodes. This will verify that future list.h changes
653 * don't break STABLE_NODE_DUP_HEAD.
655 #if GCC_VERSION >= 40903 /* only recent gcc can handle it */
656 BUILD_BUG_ON(STABLE_NODE_DUP_HEAD <= &migrate_nodes);
657 BUILD_BUG_ON(STABLE_NODE_DUP_HEAD >= &migrate_nodes + 1);
660 if (stable_node->head == &migrate_nodes)
661 list_del(&stable_node->list);
663 stable_node_dup_del(stable_node);
664 free_stable_node(stable_node);
668 * get_ksm_page: checks if the page indicated by the stable node
669 * is still its ksm page, despite having held no reference to it.
670 * In which case we can trust the content of the page, and it
671 * returns the gotten page; but if the page has now been zapped,
672 * remove the stale node from the stable tree and return NULL.
673 * But beware, the stable node's page might be being migrated.
675 * You would expect the stable_node to hold a reference to the ksm page.
676 * But if it increments the page's count, swapping out has to wait for
677 * ksmd to come around again before it can free the page, which may take
678 * seconds or even minutes: much too unresponsive. So instead we use a
679 * "keyhole reference": access to the ksm page from the stable node peeps
680 * out through its keyhole to see if that page still holds the right key,
681 * pointing back to this stable node. This relies on freeing a PageAnon
682 * page to reset its page->mapping to NULL, and relies on no other use of
683 * a page to put something that might look like our key in page->mapping.
684 * is on its way to being freed; but it is an anomaly to bear in mind.
686 static struct page *get_ksm_page(struct stable_node *stable_node, bool lock_it)
689 void *expected_mapping;
692 expected_mapping = (void *)((unsigned long)stable_node |
695 kpfn = READ_ONCE(stable_node->kpfn); /* Address dependency. */
696 page = pfn_to_page(kpfn);
697 if (READ_ONCE(page->mapping) != expected_mapping)
701 * We cannot do anything with the page while its refcount is 0.
702 * Usually 0 means free, or tail of a higher-order page: in which
703 * case this node is no longer referenced, and should be freed;
704 * however, it might mean that the page is under page_freeze_refs().
705 * The __remove_mapping() case is easy, again the node is now stale;
706 * but if page is swapcache in migrate_page_move_mapping(), it might
707 * still be our page, in which case it's essential to keep the node.
709 while (!get_page_unless_zero(page)) {
711 * Another check for page->mapping != expected_mapping would
712 * work here too. We have chosen the !PageSwapCache test to
713 * optimize the common case, when the page is or is about to
714 * be freed: PageSwapCache is cleared (under spin_lock_irq)
715 * in the freeze_refs section of __remove_mapping(); but Anon
716 * page->mapping reset to NULL later, in free_pages_prepare().
718 if (!PageSwapCache(page))
723 if (READ_ONCE(page->mapping) != expected_mapping) {
730 if (READ_ONCE(page->mapping) != expected_mapping) {
740 * We come here from above when page->mapping or !PageSwapCache
741 * suggests that the node is stale; but it might be under migration.
742 * We need smp_rmb(), matching the smp_wmb() in ksm_migrate_page(),
743 * before checking whether node->kpfn has been changed.
746 if (READ_ONCE(stable_node->kpfn) != kpfn)
748 remove_node_from_stable_tree(stable_node);
753 * Removing rmap_item from stable or unstable tree.
754 * This function will clean the information from the stable/unstable tree.
756 static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
758 if (rmap_item->address & STABLE_FLAG) {
759 struct stable_node *stable_node;
762 stable_node = rmap_item->head;
763 page = get_ksm_page(stable_node, true);
767 hlist_del(&rmap_item->hlist);
771 if (!hlist_empty(&stable_node->hlist))
775 VM_BUG_ON(stable_node->rmap_hlist_len <= 0);
776 stable_node->rmap_hlist_len--;
778 put_anon_vma(rmap_item->anon_vma);
779 rmap_item->address &= PAGE_MASK;
781 } else if (rmap_item->address & UNSTABLE_FLAG) {
784 * Usually ksmd can and must skip the rb_erase, because
785 * root_unstable_tree was already reset to RB_ROOT.
786 * But be careful when an mm is exiting: do the rb_erase
787 * if this rmap_item was inserted by this scan, rather
788 * than left over from before.
790 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
793 rb_erase(&rmap_item->node,
794 root_unstable_tree + NUMA(rmap_item->nid));
795 ksm_pages_unshared--;
796 rmap_item->address &= PAGE_MASK;
799 cond_resched(); /* we're called from many long loops */
802 static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
803 struct rmap_item **rmap_list)
806 struct rmap_item *rmap_item = *rmap_list;
807 *rmap_list = rmap_item->rmap_list;
808 remove_rmap_item_from_tree(rmap_item);
809 free_rmap_item(rmap_item);
814 * Though it's very tempting to unmerge rmap_items from stable tree rather
815 * than check every pte of a given vma, the locking doesn't quite work for
816 * that - an rmap_item is assigned to the stable tree after inserting ksm
817 * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing
818 * rmap_items from parent to child at fork time (so as not to waste time
819 * if exit comes before the next scan reaches it).
821 * Similarly, although we'd like to remove rmap_items (so updating counts
822 * and freeing memory) when unmerging an area, it's easier to leave that
823 * to the next pass of ksmd - consider, for example, how ksmd might be
824 * in cmp_and_merge_page on one of the rmap_items we would be removing.
826 static int unmerge_ksm_pages(struct vm_area_struct *vma,
827 unsigned long start, unsigned long end)
832 for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
833 if (ksm_test_exit(vma->vm_mm))
835 if (signal_pending(current))
838 err = break_ksm(vma, addr);
845 * Only called through the sysfs control interface:
847 static int remove_stable_node(struct stable_node *stable_node)
852 page = get_ksm_page(stable_node, true);
855 * get_ksm_page did remove_node_from_stable_tree itself.
860 if (WARN_ON_ONCE(page_mapped(page))) {
862 * This should not happen: but if it does, just refuse to let
863 * merge_across_nodes be switched - there is no need to panic.
868 * The stable node did not yet appear stale to get_ksm_page(),
869 * since that allows for an unmapped ksm page to be recognized
870 * right up until it is freed; but the node is safe to remove.
871 * This page might be in a pagevec waiting to be freed,
872 * or it might be PageSwapCache (perhaps under writeback),
873 * or it might have been removed from swapcache a moment ago.
875 set_page_stable_node(page, NULL);
876 remove_node_from_stable_tree(stable_node);
885 static int remove_stable_node_chain(struct stable_node *stable_node,
886 struct rb_root *root)
888 struct stable_node *dup;
889 struct hlist_node *hlist_safe;
891 if (!is_stable_node_chain(stable_node)) {
892 VM_BUG_ON(is_stable_node_dup(stable_node));
893 if (remove_stable_node(stable_node))
899 hlist_for_each_entry_safe(dup, hlist_safe,
900 &stable_node->hlist, hlist_dup) {
901 VM_BUG_ON(!is_stable_node_dup(dup));
902 if (remove_stable_node(dup))
905 BUG_ON(!hlist_empty(&stable_node->hlist));
906 free_stable_node_chain(stable_node, root);
910 static int remove_all_stable_nodes(void)
912 struct stable_node *stable_node, *next;
916 for (nid = 0; nid < ksm_nr_node_ids; nid++) {
917 while (root_stable_tree[nid].rb_node) {
918 stable_node = rb_entry(root_stable_tree[nid].rb_node,
919 struct stable_node, node);
920 if (remove_stable_node_chain(stable_node,
921 root_stable_tree + nid)) {
923 break; /* proceed to next nid */
928 list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
929 if (remove_stable_node(stable_node))
936 static int unmerge_and_remove_all_rmap_items(void)
938 struct mm_slot *mm_slot;
939 struct mm_struct *mm;
940 struct vm_area_struct *vma;
943 spin_lock(&ksm_mmlist_lock);
944 ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
945 struct mm_slot, mm_list);
946 spin_unlock(&ksm_mmlist_lock);
948 for (mm_slot = ksm_scan.mm_slot;
949 mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
951 down_read(&mm->mmap_sem);
952 for (vma = mm->mmap; vma; vma = vma->vm_next) {
953 if (ksm_test_exit(mm))
955 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
957 err = unmerge_ksm_pages(vma,
958 vma->vm_start, vma->vm_end);
963 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
964 up_read(&mm->mmap_sem);
966 spin_lock(&ksm_mmlist_lock);
967 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
968 struct mm_slot, mm_list);
969 if (ksm_test_exit(mm)) {
970 hash_del(&mm_slot->link);
971 list_del(&mm_slot->mm_list);
972 spin_unlock(&ksm_mmlist_lock);
974 free_mm_slot(mm_slot);
975 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
978 spin_unlock(&ksm_mmlist_lock);
981 /* Clean up stable nodes, but don't worry if some are still busy */
982 remove_all_stable_nodes();
987 up_read(&mm->mmap_sem);
988 spin_lock(&ksm_mmlist_lock);
989 ksm_scan.mm_slot = &ksm_mm_head;
990 spin_unlock(&ksm_mmlist_lock);
993 #endif /* CONFIG_SYSFS */
995 static u32 calc_checksum(struct page *page)
998 void *addr = kmap_atomic(page);
999 checksum = jhash2(addr, PAGE_SIZE / 4, 17);
1000 kunmap_atomic(addr);
1004 static int memcmp_pages(struct page *page1, struct page *page2)
1006 char *addr1, *addr2;
1009 addr1 = kmap_atomic(page1);
1010 addr2 = kmap_atomic(page2);
1011 ret = memcmp(addr1, addr2, PAGE_SIZE);
1012 kunmap_atomic(addr2);
1013 kunmap_atomic(addr1);
1017 static inline int pages_identical(struct page *page1, struct page *page2)
1019 return !memcmp_pages(page1, page2);
1022 static int write_protect_page(struct vm_area_struct *vma, struct page *page,
1025 struct mm_struct *mm = vma->vm_mm;
1026 struct page_vma_mapped_walk pvmw = {
1032 unsigned long mmun_start; /* For mmu_notifiers */
1033 unsigned long mmun_end; /* For mmu_notifiers */
1035 pvmw.address = page_address_in_vma(page, vma);
1036 if (pvmw.address == -EFAULT)
1039 BUG_ON(PageTransCompound(page));
1041 mmun_start = pvmw.address;
1042 mmun_end = pvmw.address + PAGE_SIZE;
1043 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1045 if (!page_vma_mapped_walk(&pvmw))
1047 if (WARN_ONCE(!pvmw.pte, "Unexpected PMD mapping?"))
1050 if (pte_write(*pvmw.pte) || pte_dirty(*pvmw.pte) ||
1051 (pte_protnone(*pvmw.pte) && pte_savedwrite(*pvmw.pte)) ||
1052 mm_tlb_flush_pending(mm)) {
1055 swapped = PageSwapCache(page);
1056 flush_cache_page(vma, pvmw.address, page_to_pfn(page));
1058 * Ok this is tricky, when get_user_pages_fast() run it doesn't
1059 * take any lock, therefore the check that we are going to make
1060 * with the pagecount against the mapcount is racey and
1061 * O_DIRECT can happen right after the check.
1062 * So we clear the pte and flush the tlb before the check
1063 * this assure us that no O_DIRECT can happen after the check
1064 * or in the middle of the check.
1066 * No need to notify as we are downgrading page table to read
1067 * only not changing it to point to a new page.
1069 * See Documentation/vm/mmu_notifier.rst
1071 entry = ptep_clear_flush(vma, pvmw.address, pvmw.pte);
1073 * Check that no O_DIRECT or similar I/O is in progress on the
1076 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
1077 set_pte_at(mm, pvmw.address, pvmw.pte, entry);
1080 if (pte_dirty(entry))
1081 set_page_dirty(page);
1083 if (pte_protnone(entry))
1084 entry = pte_mkclean(pte_clear_savedwrite(entry));
1086 entry = pte_mkclean(pte_wrprotect(entry));
1087 set_pte_at_notify(mm, pvmw.address, pvmw.pte, entry);
1089 *orig_pte = *pvmw.pte;
1093 page_vma_mapped_walk_done(&pvmw);
1095 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1101 * replace_page - replace page in vma by new ksm page
1102 * @vma: vma that holds the pte pointing to page
1103 * @page: the page we are replacing by kpage
1104 * @kpage: the ksm page we replace page by
1105 * @orig_pte: the original value of the pte
1107 * Returns 0 on success, -EFAULT on failure.
1109 static int replace_page(struct vm_area_struct *vma, struct page *page,
1110 struct page *kpage, pte_t orig_pte)
1112 struct mm_struct *mm = vma->vm_mm;
1119 unsigned long mmun_start; /* For mmu_notifiers */
1120 unsigned long mmun_end; /* For mmu_notifiers */
1122 addr = page_address_in_vma(page, vma);
1123 if (addr == -EFAULT)
1126 pmd = mm_find_pmd(mm, addr);
1131 mmun_end = addr + PAGE_SIZE;
1132 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1134 ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
1135 if (!pte_same(*ptep, orig_pte)) {
1136 pte_unmap_unlock(ptep, ptl);
1141 * No need to check ksm_use_zero_pages here: we can only have a
1142 * zero_page here if ksm_use_zero_pages was enabled alreaady.
1144 if (!is_zero_pfn(page_to_pfn(kpage))) {
1146 page_add_anon_rmap(kpage, vma, addr, false);
1147 newpte = mk_pte(kpage, vma->vm_page_prot);
1149 newpte = pte_mkspecial(pfn_pte(page_to_pfn(kpage),
1150 vma->vm_page_prot));
1152 * We're replacing an anonymous page with a zero page, which is
1153 * not anonymous. We need to do proper accounting otherwise we
1154 * will get wrong values in /proc, and a BUG message in dmesg
1155 * when tearing down the mm.
1157 dec_mm_counter(mm, MM_ANONPAGES);
1160 flush_cache_page(vma, addr, pte_pfn(*ptep));
1162 * No need to notify as we are replacing a read only page with another
1163 * read only page with the same content.
1165 * See Documentation/vm/mmu_notifier.rst
1167 ptep_clear_flush(vma, addr, ptep);
1168 set_pte_at_notify(mm, addr, ptep, newpte);
1170 page_remove_rmap(page, false);
1171 if (!page_mapped(page))
1172 try_to_free_swap(page);
1175 pte_unmap_unlock(ptep, ptl);
1178 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1184 * try_to_merge_one_page - take two pages and merge them into one
1185 * @vma: the vma that holds the pte pointing to page
1186 * @page: the PageAnon page that we want to replace with kpage
1187 * @kpage: the PageKsm page that we want to map instead of page,
1188 * or NULL the first time when we want to use page as kpage.
1190 * This function returns 0 if the pages were merged, -EFAULT otherwise.
1192 static int try_to_merge_one_page(struct vm_area_struct *vma,
1193 struct page *page, struct page *kpage)
1195 pte_t orig_pte = __pte(0);
1198 if (page == kpage) /* ksm page forked */
1201 if (!PageAnon(page))
1205 * We need the page lock to read a stable PageSwapCache in
1206 * write_protect_page(). We use trylock_page() instead of
1207 * lock_page() because we don't want to wait here - we
1208 * prefer to continue scanning and merging different pages,
1209 * then come back to this page when it is unlocked.
1211 if (!trylock_page(page))
1214 if (PageTransCompound(page)) {
1215 if (split_huge_page(page))
1220 * If this anonymous page is mapped only here, its pte may need
1221 * to be write-protected. If it's mapped elsewhere, all of its
1222 * ptes are necessarily already write-protected. But in either
1223 * case, we need to lock and check page_count is not raised.
1225 if (write_protect_page(vma, page, &orig_pte) == 0) {
1228 * While we hold page lock, upgrade page from
1229 * PageAnon+anon_vma to PageKsm+NULL stable_node:
1230 * stable_tree_insert() will update stable_node.
1232 set_page_stable_node(page, NULL);
1233 mark_page_accessed(page);
1235 * Page reclaim just frees a clean page with no dirty
1236 * ptes: make sure that the ksm page would be swapped.
1238 if (!PageDirty(page))
1241 } else if (pages_identical(page, kpage))
1242 err = replace_page(vma, page, kpage, orig_pte);
1245 if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
1246 munlock_vma_page(page);
1247 if (!PageMlocked(kpage)) {
1250 mlock_vma_page(kpage);
1251 page = kpage; /* for final unlock */
1262 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
1263 * but no new kernel page is allocated: kpage must already be a ksm page.
1265 * This function returns 0 if the pages were merged, -EFAULT otherwise.
1267 static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
1268 struct page *page, struct page *kpage)
1270 struct mm_struct *mm = rmap_item->mm;
1271 struct vm_area_struct *vma;
1274 down_read(&mm->mmap_sem);
1275 vma = find_mergeable_vma(mm, rmap_item->address);
1279 err = try_to_merge_one_page(vma, page, kpage);
1283 /* Unstable nid is in union with stable anon_vma: remove first */
1284 remove_rmap_item_from_tree(rmap_item);
1286 /* Must get reference to anon_vma while still holding mmap_sem */
1287 rmap_item->anon_vma = vma->anon_vma;
1288 get_anon_vma(vma->anon_vma);
1290 up_read(&mm->mmap_sem);
1295 * try_to_merge_two_pages - take two identical pages and prepare them
1296 * to be merged into one page.
1298 * This function returns the kpage if we successfully merged two identical
1299 * pages into one ksm page, NULL otherwise.
1301 * Note that this function upgrades page to ksm page: if one of the pages
1302 * is already a ksm page, try_to_merge_with_ksm_page should be used.
1304 static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
1306 struct rmap_item *tree_rmap_item,
1307 struct page *tree_page)
1311 err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
1313 err = try_to_merge_with_ksm_page(tree_rmap_item,
1316 * If that fails, we have a ksm page with only one pte
1317 * pointing to it: so break it.
1320 break_cow(rmap_item);
1322 return err ? NULL : page;
1325 static __always_inline
1326 bool __is_page_sharing_candidate(struct stable_node *stable_node, int offset)
1328 VM_BUG_ON(stable_node->rmap_hlist_len < 0);
1330 * Check that at least one mapping still exists, otherwise
1331 * there's no much point to merge and share with this
1332 * stable_node, as the underlying tree_page of the other
1333 * sharer is going to be freed soon.
1335 return stable_node->rmap_hlist_len &&
1336 stable_node->rmap_hlist_len + offset < ksm_max_page_sharing;
1339 static __always_inline
1340 bool is_page_sharing_candidate(struct stable_node *stable_node)
1342 return __is_page_sharing_candidate(stable_node, 0);
1345 static struct page *stable_node_dup(struct stable_node **_stable_node_dup,
1346 struct stable_node **_stable_node,
1347 struct rb_root *root,
1348 bool prune_stale_stable_nodes)
1350 struct stable_node *dup, *found = NULL, *stable_node = *_stable_node;
1351 struct hlist_node *hlist_safe;
1352 struct page *_tree_page, *tree_page = NULL;
1354 int found_rmap_hlist_len;
1356 if (!prune_stale_stable_nodes ||
1357 time_before(jiffies, stable_node->chain_prune_time +
1359 ksm_stable_node_chains_prune_millisecs)))
1360 prune_stale_stable_nodes = false;
1362 stable_node->chain_prune_time = jiffies;
1364 hlist_for_each_entry_safe(dup, hlist_safe,
1365 &stable_node->hlist, hlist_dup) {
1368 * We must walk all stable_node_dup to prune the stale
1369 * stable nodes during lookup.
1371 * get_ksm_page can drop the nodes from the
1372 * stable_node->hlist if they point to freed pages
1373 * (that's why we do a _safe walk). The "dup"
1374 * stable_node parameter itself will be freed from
1375 * under us if it returns NULL.
1377 _tree_page = get_ksm_page(dup, false);
1381 if (is_page_sharing_candidate(dup)) {
1383 dup->rmap_hlist_len > found_rmap_hlist_len) {
1385 put_page(tree_page);
1387 found_rmap_hlist_len = found->rmap_hlist_len;
1388 tree_page = _tree_page;
1390 /* skip put_page for found dup */
1391 if (!prune_stale_stable_nodes)
1396 put_page(_tree_page);
1401 * nr is counting all dups in the chain only if
1402 * prune_stale_stable_nodes is true, otherwise we may
1403 * break the loop at nr == 1 even if there are
1406 if (prune_stale_stable_nodes && nr == 1) {
1408 * If there's not just one entry it would
1409 * corrupt memory, better BUG_ON. In KSM
1410 * context with no lock held it's not even
1413 BUG_ON(stable_node->hlist.first->next);
1416 * There's just one entry and it is below the
1417 * deduplication limit so drop the chain.
1419 rb_replace_node(&stable_node->node, &found->node,
1421 free_stable_node(stable_node);
1422 ksm_stable_node_chains--;
1423 ksm_stable_node_dups--;
1425 * NOTE: the caller depends on the stable_node
1426 * to be equal to stable_node_dup if the chain
1429 *_stable_node = found;
1431 * Just for robustneess as stable_node is
1432 * otherwise left as a stable pointer, the
1433 * compiler shall optimize it away at build
1437 } else if (stable_node->hlist.first != &found->hlist_dup &&
1438 __is_page_sharing_candidate(found, 1)) {
1440 * If the found stable_node dup can accept one
1441 * more future merge (in addition to the one
1442 * that is underway) and is not at the head of
1443 * the chain, put it there so next search will
1444 * be quicker in the !prune_stale_stable_nodes
1447 * NOTE: it would be inaccurate to use nr > 1
1448 * instead of checking the hlist.first pointer
1449 * directly, because in the
1450 * prune_stale_stable_nodes case "nr" isn't
1451 * the position of the found dup in the chain,
1452 * but the total number of dups in the chain.
1454 hlist_del(&found->hlist_dup);
1455 hlist_add_head(&found->hlist_dup,
1456 &stable_node->hlist);
1460 *_stable_node_dup = found;
1464 static struct stable_node *stable_node_dup_any(struct stable_node *stable_node,
1465 struct rb_root *root)
1467 if (!is_stable_node_chain(stable_node))
1469 if (hlist_empty(&stable_node->hlist)) {
1470 free_stable_node_chain(stable_node, root);
1473 return hlist_entry(stable_node->hlist.first,
1474 typeof(*stable_node), hlist_dup);
1478 * Like for get_ksm_page, this function can free the *_stable_node and
1479 * *_stable_node_dup if the returned tree_page is NULL.
1481 * It can also free and overwrite *_stable_node with the found
1482 * stable_node_dup if the chain is collapsed (in which case
1483 * *_stable_node will be equal to *_stable_node_dup like if the chain
1484 * never existed). It's up to the caller to verify tree_page is not
1485 * NULL before dereferencing *_stable_node or *_stable_node_dup.
1487 * *_stable_node_dup is really a second output parameter of this
1488 * function and will be overwritten in all cases, the caller doesn't
1489 * need to initialize it.
1491 static struct page *__stable_node_chain(struct stable_node **_stable_node_dup,
1492 struct stable_node **_stable_node,
1493 struct rb_root *root,
1494 bool prune_stale_stable_nodes)
1496 struct stable_node *stable_node = *_stable_node;
1497 if (!is_stable_node_chain(stable_node)) {
1498 if (is_page_sharing_candidate(stable_node)) {
1499 *_stable_node_dup = stable_node;
1500 return get_ksm_page(stable_node, false);
1503 * _stable_node_dup set to NULL means the stable_node
1504 * reached the ksm_max_page_sharing limit.
1506 *_stable_node_dup = NULL;
1509 return stable_node_dup(_stable_node_dup, _stable_node, root,
1510 prune_stale_stable_nodes);
1513 static __always_inline struct page *chain_prune(struct stable_node **s_n_d,
1514 struct stable_node **s_n,
1515 struct rb_root *root)
1517 return __stable_node_chain(s_n_d, s_n, root, true);
1520 static __always_inline struct page *chain(struct stable_node **s_n_d,
1521 struct stable_node *s_n,
1522 struct rb_root *root)
1524 struct stable_node *old_stable_node = s_n;
1525 struct page *tree_page;
1527 tree_page = __stable_node_chain(s_n_d, &s_n, root, false);
1528 /* not pruning dups so s_n cannot have changed */
1529 VM_BUG_ON(s_n != old_stable_node);
1534 * stable_tree_search - search for page inside the stable tree
1536 * This function checks if there is a page inside the stable tree
1537 * with identical content to the page that we are scanning right now.
1539 * This function returns the stable tree node of identical content if found,
1542 static struct page *stable_tree_search(struct page *page)
1545 struct rb_root *root;
1546 struct rb_node **new;
1547 struct rb_node *parent;
1548 struct stable_node *stable_node, *stable_node_dup, *stable_node_any;
1549 struct stable_node *page_node;
1551 page_node = page_stable_node(page);
1552 if (page_node && page_node->head != &migrate_nodes) {
1553 /* ksm page forked */
1558 nid = get_kpfn_nid(page_to_pfn(page));
1559 root = root_stable_tree + nid;
1561 new = &root->rb_node;
1565 struct page *tree_page;
1569 stable_node = rb_entry(*new, struct stable_node, node);
1570 stable_node_any = NULL;
1571 tree_page = chain_prune(&stable_node_dup, &stable_node, root);
1573 * NOTE: stable_node may have been freed by
1574 * chain_prune() if the returned stable_node_dup is
1575 * not NULL. stable_node_dup may have been inserted in
1576 * the rbtree instead as a regular stable_node (in
1577 * order to collapse the stable_node chain if a single
1578 * stable_node dup was found in it). In such case the
1579 * stable_node is overwritten by the calleee to point
1580 * to the stable_node_dup that was collapsed in the
1581 * stable rbtree and stable_node will be equal to
1582 * stable_node_dup like if the chain never existed.
1584 if (!stable_node_dup) {
1586 * Either all stable_node dups were full in
1587 * this stable_node chain, or this chain was
1588 * empty and should be rb_erased.
1590 stable_node_any = stable_node_dup_any(stable_node,
1592 if (!stable_node_any) {
1593 /* rb_erase just run */
1597 * Take any of the stable_node dups page of
1598 * this stable_node chain to let the tree walk
1599 * continue. All KSM pages belonging to the
1600 * stable_node dups in a stable_node chain
1601 * have the same content and they're
1602 * wrprotected at all times. Any will work
1603 * fine to continue the walk.
1605 tree_page = get_ksm_page(stable_node_any, false);
1607 VM_BUG_ON(!stable_node_dup ^ !!stable_node_any);
1610 * If we walked over a stale stable_node,
1611 * get_ksm_page() will call rb_erase() and it
1612 * may rebalance the tree from under us. So
1613 * restart the search from scratch. Returning
1614 * NULL would be safe too, but we'd generate
1615 * false negative insertions just because some
1616 * stable_node was stale.
1621 ret = memcmp_pages(page, tree_page);
1622 put_page(tree_page);
1626 new = &parent->rb_left;
1628 new = &parent->rb_right;
1631 VM_BUG_ON(page_node->head != &migrate_nodes);
1633 * Test if the migrated page should be merged
1634 * into a stable node dup. If the mapcount is
1635 * 1 we can migrate it with another KSM page
1636 * without adding it to the chain.
1638 if (page_mapcount(page) > 1)
1642 if (!stable_node_dup) {
1644 * If the stable_node is a chain and
1645 * we got a payload match in memcmp
1646 * but we cannot merge the scanned
1647 * page in any of the existing
1648 * stable_node dups because they're
1649 * all full, we need to wait the
1650 * scanned page to find itself a match
1651 * in the unstable tree to create a
1652 * brand new KSM page to add later to
1653 * the dups of this stable_node.
1659 * Lock and unlock the stable_node's page (which
1660 * might already have been migrated) so that page
1661 * migration is sure to notice its raised count.
1662 * It would be more elegant to return stable_node
1663 * than kpage, but that involves more changes.
1665 tree_page = get_ksm_page(stable_node_dup, true);
1666 if (unlikely(!tree_page))
1668 * The tree may have been rebalanced,
1669 * so re-evaluate parent and new.
1672 unlock_page(tree_page);
1674 if (get_kpfn_nid(stable_node_dup->kpfn) !=
1675 NUMA(stable_node_dup->nid)) {
1676 put_page(tree_page);
1686 list_del(&page_node->list);
1687 DO_NUMA(page_node->nid = nid);
1688 rb_link_node(&page_node->node, parent, new);
1689 rb_insert_color(&page_node->node, root);
1691 if (is_page_sharing_candidate(page_node)) {
1699 * If stable_node was a chain and chain_prune collapsed it,
1700 * stable_node has been updated to be the new regular
1701 * stable_node. A collapse of the chain is indistinguishable
1702 * from the case there was no chain in the stable
1703 * rbtree. Otherwise stable_node is the chain and
1704 * stable_node_dup is the dup to replace.
1706 if (stable_node_dup == stable_node) {
1707 VM_BUG_ON(is_stable_node_chain(stable_node_dup));
1708 VM_BUG_ON(is_stable_node_dup(stable_node_dup));
1709 /* there is no chain */
1711 VM_BUG_ON(page_node->head != &migrate_nodes);
1712 list_del(&page_node->list);
1713 DO_NUMA(page_node->nid = nid);
1714 rb_replace_node(&stable_node_dup->node,
1717 if (is_page_sharing_candidate(page_node))
1722 rb_erase(&stable_node_dup->node, root);
1726 VM_BUG_ON(!is_stable_node_chain(stable_node));
1727 __stable_node_dup_del(stable_node_dup);
1729 VM_BUG_ON(page_node->head != &migrate_nodes);
1730 list_del(&page_node->list);
1731 DO_NUMA(page_node->nid = nid);
1732 stable_node_chain_add_dup(page_node, stable_node);
1733 if (is_page_sharing_candidate(page_node))
1741 stable_node_dup->head = &migrate_nodes;
1742 list_add(&stable_node_dup->list, stable_node_dup->head);
1746 /* stable_node_dup could be null if it reached the limit */
1747 if (!stable_node_dup)
1748 stable_node_dup = stable_node_any;
1750 * If stable_node was a chain and chain_prune collapsed it,
1751 * stable_node has been updated to be the new regular
1752 * stable_node. A collapse of the chain is indistinguishable
1753 * from the case there was no chain in the stable
1754 * rbtree. Otherwise stable_node is the chain and
1755 * stable_node_dup is the dup to replace.
1757 if (stable_node_dup == stable_node) {
1758 VM_BUG_ON(is_stable_node_chain(stable_node_dup));
1759 VM_BUG_ON(is_stable_node_dup(stable_node_dup));
1760 /* chain is missing so create it */
1761 stable_node = alloc_stable_node_chain(stable_node_dup,
1767 * Add this stable_node dup that was
1768 * migrated to the stable_node chain
1769 * of the current nid for this page
1772 VM_BUG_ON(!is_stable_node_chain(stable_node));
1773 VM_BUG_ON(!is_stable_node_dup(stable_node_dup));
1774 VM_BUG_ON(page_node->head != &migrate_nodes);
1775 list_del(&page_node->list);
1776 DO_NUMA(page_node->nid = nid);
1777 stable_node_chain_add_dup(page_node, stable_node);
1782 * stable_tree_insert - insert stable tree node pointing to new ksm page
1783 * into the stable tree.
1785 * This function returns the stable tree node just allocated on success,
1788 static struct stable_node *stable_tree_insert(struct page *kpage)
1792 struct rb_root *root;
1793 struct rb_node **new;
1794 struct rb_node *parent;
1795 struct stable_node *stable_node, *stable_node_dup, *stable_node_any;
1796 bool need_chain = false;
1798 kpfn = page_to_pfn(kpage);
1799 nid = get_kpfn_nid(kpfn);
1800 root = root_stable_tree + nid;
1803 new = &root->rb_node;
1806 struct page *tree_page;
1810 stable_node = rb_entry(*new, struct stable_node, node);
1811 stable_node_any = NULL;
1812 tree_page = chain(&stable_node_dup, stable_node, root);
1813 if (!stable_node_dup) {
1815 * Either all stable_node dups were full in
1816 * this stable_node chain, or this chain was
1817 * empty and should be rb_erased.
1819 stable_node_any = stable_node_dup_any(stable_node,
1821 if (!stable_node_any) {
1822 /* rb_erase just run */
1826 * Take any of the stable_node dups page of
1827 * this stable_node chain to let the tree walk
1828 * continue. All KSM pages belonging to the
1829 * stable_node dups in a stable_node chain
1830 * have the same content and they're
1831 * wrprotected at all times. Any will work
1832 * fine to continue the walk.
1834 tree_page = get_ksm_page(stable_node_any, false);
1836 VM_BUG_ON(!stable_node_dup ^ !!stable_node_any);
1839 * If we walked over a stale stable_node,
1840 * get_ksm_page() will call rb_erase() and it
1841 * may rebalance the tree from under us. So
1842 * restart the search from scratch. Returning
1843 * NULL would be safe too, but we'd generate
1844 * false negative insertions just because some
1845 * stable_node was stale.
1850 ret = memcmp_pages(kpage, tree_page);
1851 put_page(tree_page);
1855 new = &parent->rb_left;
1857 new = &parent->rb_right;
1864 stable_node_dup = alloc_stable_node();
1865 if (!stable_node_dup)
1868 INIT_HLIST_HEAD(&stable_node_dup->hlist);
1869 stable_node_dup->kpfn = kpfn;
1870 set_page_stable_node(kpage, stable_node_dup);
1871 stable_node_dup->rmap_hlist_len = 0;
1872 DO_NUMA(stable_node_dup->nid = nid);
1874 rb_link_node(&stable_node_dup->node, parent, new);
1875 rb_insert_color(&stable_node_dup->node, root);
1877 if (!is_stable_node_chain(stable_node)) {
1878 struct stable_node *orig = stable_node;
1879 /* chain is missing so create it */
1880 stable_node = alloc_stable_node_chain(orig, root);
1882 free_stable_node(stable_node_dup);
1886 stable_node_chain_add_dup(stable_node_dup, stable_node);
1889 return stable_node_dup;
1893 * unstable_tree_search_insert - search for identical page,
1894 * else insert rmap_item into the unstable tree.
1896 * This function searches for a page in the unstable tree identical to the
1897 * page currently being scanned; and if no identical page is found in the
1898 * tree, we insert rmap_item as a new object into the unstable tree.
1900 * This function returns pointer to rmap_item found to be identical
1901 * to the currently scanned page, NULL otherwise.
1903 * This function does both searching and inserting, because they share
1904 * the same walking algorithm in an rbtree.
1907 struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1909 struct page **tree_pagep)
1911 struct rb_node **new;
1912 struct rb_root *root;
1913 struct rb_node *parent = NULL;
1916 nid = get_kpfn_nid(page_to_pfn(page));
1917 root = root_unstable_tree + nid;
1918 new = &root->rb_node;
1921 struct rmap_item *tree_rmap_item;
1922 struct page *tree_page;
1926 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
1927 tree_page = get_mergeable_page(tree_rmap_item);
1932 * Don't substitute a ksm page for a forked page.
1934 if (page == tree_page) {
1935 put_page(tree_page);
1939 ret = memcmp_pages(page, tree_page);
1943 put_page(tree_page);
1944 new = &parent->rb_left;
1945 } else if (ret > 0) {
1946 put_page(tree_page);
1947 new = &parent->rb_right;
1948 } else if (!ksm_merge_across_nodes &&
1949 page_to_nid(tree_page) != nid) {
1951 * If tree_page has been migrated to another NUMA node,
1952 * it will be flushed out and put in the right unstable
1953 * tree next time: only merge with it when across_nodes.
1955 put_page(tree_page);
1958 *tree_pagep = tree_page;
1959 return tree_rmap_item;
1963 rmap_item->address |= UNSTABLE_FLAG;
1964 rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1965 DO_NUMA(rmap_item->nid = nid);
1966 rb_link_node(&rmap_item->node, parent, new);
1967 rb_insert_color(&rmap_item->node, root);
1969 ksm_pages_unshared++;
1974 * stable_tree_append - add another rmap_item to the linked list of
1975 * rmap_items hanging off a given node of the stable tree, all sharing
1976 * the same ksm page.
1978 static void stable_tree_append(struct rmap_item *rmap_item,
1979 struct stable_node *stable_node,
1980 bool max_page_sharing_bypass)
1983 * rmap won't find this mapping if we don't insert the
1984 * rmap_item in the right stable_node
1985 * duplicate. page_migration could break later if rmap breaks,
1986 * so we can as well crash here. We really need to check for
1987 * rmap_hlist_len == STABLE_NODE_CHAIN, but we can as well check
1988 * for other negative values as an undeflow if detected here
1989 * for the first time (and not when decreasing rmap_hlist_len)
1990 * would be sign of memory corruption in the stable_node.
1992 BUG_ON(stable_node->rmap_hlist_len < 0);
1994 stable_node->rmap_hlist_len++;
1995 if (!max_page_sharing_bypass)
1996 /* possibly non fatal but unexpected overflow, only warn */
1997 WARN_ON_ONCE(stable_node->rmap_hlist_len >
1998 ksm_max_page_sharing);
2000 rmap_item->head = stable_node;
2001 rmap_item->address |= STABLE_FLAG;
2002 hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
2004 if (rmap_item->hlist.next)
2005 ksm_pages_sharing++;
2011 * cmp_and_merge_page - first see if page can be merged into the stable tree;
2012 * if not, compare checksum to previous and if it's the same, see if page can
2013 * be inserted into the unstable tree, or merged with a page already there and
2014 * both transferred to the stable tree.
2016 * @page: the page that we are searching identical page to.
2017 * @rmap_item: the reverse mapping into the virtual address of this page
2019 static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
2021 struct mm_struct *mm = rmap_item->mm;
2022 struct rmap_item *tree_rmap_item;
2023 struct page *tree_page = NULL;
2024 struct stable_node *stable_node;
2026 unsigned int checksum;
2028 bool max_page_sharing_bypass = false;
2030 stable_node = page_stable_node(page);
2032 if (stable_node->head != &migrate_nodes &&
2033 get_kpfn_nid(READ_ONCE(stable_node->kpfn)) !=
2034 NUMA(stable_node->nid)) {
2035 stable_node_dup_del(stable_node);
2036 stable_node->head = &migrate_nodes;
2037 list_add(&stable_node->list, stable_node->head);
2039 if (stable_node->head != &migrate_nodes &&
2040 rmap_item->head == stable_node)
2043 * If it's a KSM fork, allow it to go over the sharing limit
2046 if (!is_page_sharing_candidate(stable_node))
2047 max_page_sharing_bypass = true;
2050 /* We first start with searching the page inside the stable tree */
2051 kpage = stable_tree_search(page);
2052 if (kpage == page && rmap_item->head == stable_node) {
2057 remove_rmap_item_from_tree(rmap_item);
2060 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
2063 * The page was successfully merged:
2064 * add its rmap_item to the stable tree.
2067 stable_tree_append(rmap_item, page_stable_node(kpage),
2068 max_page_sharing_bypass);
2076 * If the hash value of the page has changed from the last time
2077 * we calculated it, this page is changing frequently: therefore we
2078 * don't want to insert it in the unstable tree, and we don't want
2079 * to waste our time searching for something identical to it there.
2081 checksum = calc_checksum(page);
2082 if (rmap_item->oldchecksum != checksum) {
2083 rmap_item->oldchecksum = checksum;
2088 * Same checksum as an empty page. We attempt to merge it with the
2089 * appropriate zero page if the user enabled this via sysfs.
2091 if (ksm_use_zero_pages && (checksum == zero_checksum)) {
2092 struct vm_area_struct *vma;
2094 down_read(&mm->mmap_sem);
2095 vma = find_mergeable_vma(mm, rmap_item->address);
2096 err = try_to_merge_one_page(vma, page,
2097 ZERO_PAGE(rmap_item->address));
2098 up_read(&mm->mmap_sem);
2100 * In case of failure, the page was not really empty, so we
2101 * need to continue. Otherwise we're done.
2107 unstable_tree_search_insert(rmap_item, page, &tree_page);
2108 if (tree_rmap_item) {
2111 kpage = try_to_merge_two_pages(rmap_item, page,
2112 tree_rmap_item, tree_page);
2114 * If both pages we tried to merge belong to the same compound
2115 * page, then we actually ended up increasing the reference
2116 * count of the same compound page twice, and split_huge_page
2118 * Here we set a flag if that happened, and we use it later to
2119 * try split_huge_page again. Since we call put_page right
2120 * afterwards, the reference count will be correct and
2121 * split_huge_page should succeed.
2123 split = PageTransCompound(page)
2124 && compound_head(page) == compound_head(tree_page);
2125 put_page(tree_page);
2128 * The pages were successfully merged: insert new
2129 * node in the stable tree and add both rmap_items.
2132 stable_node = stable_tree_insert(kpage);
2134 stable_tree_append(tree_rmap_item, stable_node,
2136 stable_tree_append(rmap_item, stable_node,
2142 * If we fail to insert the page into the stable tree,
2143 * we will have 2 virtual addresses that are pointing
2144 * to a ksm page left outside the stable tree,
2145 * in which case we need to break_cow on both.
2148 break_cow(tree_rmap_item);
2149 break_cow(rmap_item);
2153 * We are here if we tried to merge two pages and
2154 * failed because they both belonged to the same
2155 * compound page. We will split the page now, but no
2156 * merging will take place.
2157 * We do not want to add the cost of a full lock; if
2158 * the page is locked, it is better to skip it and
2159 * perhaps try again later.
2161 if (!trylock_page(page))
2163 split_huge_page(page);
2169 static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
2170 struct rmap_item **rmap_list,
2173 struct rmap_item *rmap_item;
2175 while (*rmap_list) {
2176 rmap_item = *rmap_list;
2177 if ((rmap_item->address & PAGE_MASK) == addr)
2179 if (rmap_item->address > addr)
2181 *rmap_list = rmap_item->rmap_list;
2182 remove_rmap_item_from_tree(rmap_item);
2183 free_rmap_item(rmap_item);
2186 rmap_item = alloc_rmap_item();
2188 /* It has already been zeroed */
2189 rmap_item->mm = mm_slot->mm;
2190 rmap_item->address = addr;
2191 rmap_item->rmap_list = *rmap_list;
2192 *rmap_list = rmap_item;
2197 static struct rmap_item *scan_get_next_rmap_item(struct page **page)
2199 struct mm_struct *mm;
2200 struct mm_slot *slot;
2201 struct vm_area_struct *vma;
2202 struct rmap_item *rmap_item;
2205 if (list_empty(&ksm_mm_head.mm_list))
2208 slot = ksm_scan.mm_slot;
2209 if (slot == &ksm_mm_head) {
2211 * A number of pages can hang around indefinitely on per-cpu
2212 * pagevecs, raised page count preventing write_protect_page
2213 * from merging them. Though it doesn't really matter much,
2214 * it is puzzling to see some stuck in pages_volatile until
2215 * other activity jostles them out, and they also prevented
2216 * LTP's KSM test from succeeding deterministically; so drain
2217 * them here (here rather than on entry to ksm_do_scan(),
2218 * so we don't IPI too often when pages_to_scan is set low).
2220 lru_add_drain_all();
2223 * Whereas stale stable_nodes on the stable_tree itself
2224 * get pruned in the regular course of stable_tree_search(),
2225 * those moved out to the migrate_nodes list can accumulate:
2226 * so prune them once before each full scan.
2228 if (!ksm_merge_across_nodes) {
2229 struct stable_node *stable_node, *next;
2232 list_for_each_entry_safe(stable_node, next,
2233 &migrate_nodes, list) {
2234 page = get_ksm_page(stable_node, false);
2241 for (nid = 0; nid < ksm_nr_node_ids; nid++)
2242 root_unstable_tree[nid] = RB_ROOT;
2244 spin_lock(&ksm_mmlist_lock);
2245 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
2246 ksm_scan.mm_slot = slot;
2247 spin_unlock(&ksm_mmlist_lock);
2249 * Although we tested list_empty() above, a racing __ksm_exit
2250 * of the last mm on the list may have removed it since then.
2252 if (slot == &ksm_mm_head)
2255 ksm_scan.address = 0;
2256 ksm_scan.rmap_list = &slot->rmap_list;
2260 down_read(&mm->mmap_sem);
2261 if (ksm_test_exit(mm))
2264 vma = find_vma(mm, ksm_scan.address);
2266 for (; vma; vma = vma->vm_next) {
2267 if (!(vma->vm_flags & VM_MERGEABLE))
2269 if (ksm_scan.address < vma->vm_start)
2270 ksm_scan.address = vma->vm_start;
2272 ksm_scan.address = vma->vm_end;
2274 while (ksm_scan.address < vma->vm_end) {
2275 if (ksm_test_exit(mm))
2277 *page = follow_page(vma, ksm_scan.address, FOLL_GET);
2278 if (IS_ERR_OR_NULL(*page)) {
2279 ksm_scan.address += PAGE_SIZE;
2283 if (PageAnon(*page)) {
2284 flush_anon_page(vma, *page, ksm_scan.address);
2285 flush_dcache_page(*page);
2286 rmap_item = get_next_rmap_item(slot,
2287 ksm_scan.rmap_list, ksm_scan.address);
2289 ksm_scan.rmap_list =
2290 &rmap_item->rmap_list;
2291 ksm_scan.address += PAGE_SIZE;
2294 up_read(&mm->mmap_sem);
2298 ksm_scan.address += PAGE_SIZE;
2303 if (ksm_test_exit(mm)) {
2304 ksm_scan.address = 0;
2305 ksm_scan.rmap_list = &slot->rmap_list;
2308 * Nuke all the rmap_items that are above this current rmap:
2309 * because there were no VM_MERGEABLE vmas with such addresses.
2311 remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
2313 spin_lock(&ksm_mmlist_lock);
2314 ksm_scan.mm_slot = list_entry(slot->mm_list.next,
2315 struct mm_slot, mm_list);
2316 if (ksm_scan.address == 0) {
2318 * We've completed a full scan of all vmas, holding mmap_sem
2319 * throughout, and found no VM_MERGEABLE: so do the same as
2320 * __ksm_exit does to remove this mm from all our lists now.
2321 * This applies either when cleaning up after __ksm_exit
2322 * (but beware: we can reach here even before __ksm_exit),
2323 * or when all VM_MERGEABLE areas have been unmapped (and
2324 * mmap_sem then protects against race with MADV_MERGEABLE).
2326 hash_del(&slot->link);
2327 list_del(&slot->mm_list);
2328 spin_unlock(&ksm_mmlist_lock);
2331 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
2332 up_read(&mm->mmap_sem);
2335 up_read(&mm->mmap_sem);
2337 * up_read(&mm->mmap_sem) first because after
2338 * spin_unlock(&ksm_mmlist_lock) run, the "mm" may
2339 * already have been freed under us by __ksm_exit()
2340 * because the "mm_slot" is still hashed and
2341 * ksm_scan.mm_slot doesn't point to it anymore.
2343 spin_unlock(&ksm_mmlist_lock);
2346 /* Repeat until we've completed scanning the whole list */
2347 slot = ksm_scan.mm_slot;
2348 if (slot != &ksm_mm_head)
2356 * ksm_do_scan - the ksm scanner main worker function.
2357 * @scan_npages: number of pages we want to scan before we return.
2359 static void ksm_do_scan(unsigned int scan_npages)
2361 struct rmap_item *rmap_item;
2362 struct page *uninitialized_var(page);
2364 while (scan_npages-- && likely(!freezing(current))) {
2366 rmap_item = scan_get_next_rmap_item(&page);
2369 cmp_and_merge_page(page, rmap_item);
2374 static int ksmd_should_run(void)
2376 return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
2379 static int ksm_scan_thread(void *nothing)
2382 set_user_nice(current, 5);
2384 while (!kthread_should_stop()) {
2385 mutex_lock(&ksm_thread_mutex);
2386 wait_while_offlining();
2387 if (ksmd_should_run())
2388 ksm_do_scan(ksm_thread_pages_to_scan);
2389 mutex_unlock(&ksm_thread_mutex);
2393 if (ksmd_should_run()) {
2394 schedule_timeout_interruptible(
2395 msecs_to_jiffies(ksm_thread_sleep_millisecs));
2397 wait_event_freezable(ksm_thread_wait,
2398 ksmd_should_run() || kthread_should_stop());
2404 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
2405 unsigned long end, int advice, unsigned long *vm_flags)
2407 struct mm_struct *mm = vma->vm_mm;
2411 case MADV_MERGEABLE:
2413 * Be somewhat over-protective for now!
2415 if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE |
2416 VM_PFNMAP | VM_IO | VM_DONTEXPAND |
2417 VM_HUGETLB | VM_MIXEDMAP))
2418 return 0; /* just ignore the advice */
2421 if (*vm_flags & VM_SAO)
2425 if (*vm_flags & VM_SPARC_ADI)
2429 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
2430 err = __ksm_enter(mm);
2435 *vm_flags |= VM_MERGEABLE;
2438 case MADV_UNMERGEABLE:
2439 if (!(*vm_flags & VM_MERGEABLE))
2440 return 0; /* just ignore the advice */
2442 if (vma->anon_vma) {
2443 err = unmerge_ksm_pages(vma, start, end);
2448 *vm_flags &= ~VM_MERGEABLE;
2455 int __ksm_enter(struct mm_struct *mm)
2457 struct mm_slot *mm_slot;
2460 mm_slot = alloc_mm_slot();
2464 /* Check ksm_run too? Would need tighter locking */
2465 needs_wakeup = list_empty(&ksm_mm_head.mm_list);
2467 spin_lock(&ksm_mmlist_lock);
2468 insert_to_mm_slots_hash(mm, mm_slot);
2470 * When KSM_RUN_MERGE (or KSM_RUN_STOP),
2471 * insert just behind the scanning cursor, to let the area settle
2472 * down a little; when fork is followed by immediate exec, we don't
2473 * want ksmd to waste time setting up and tearing down an rmap_list.
2475 * But when KSM_RUN_UNMERGE, it's important to insert ahead of its
2476 * scanning cursor, otherwise KSM pages in newly forked mms will be
2477 * missed: then we might as well insert at the end of the list.
2479 if (ksm_run & KSM_RUN_UNMERGE)
2480 list_add_tail(&mm_slot->mm_list, &ksm_mm_head.mm_list);
2482 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
2483 spin_unlock(&ksm_mmlist_lock);
2485 set_bit(MMF_VM_MERGEABLE, &mm->flags);
2489 wake_up_interruptible(&ksm_thread_wait);
2494 void __ksm_exit(struct mm_struct *mm)
2496 struct mm_slot *mm_slot;
2497 int easy_to_free = 0;
2500 * This process is exiting: if it's straightforward (as is the
2501 * case when ksmd was never running), free mm_slot immediately.
2502 * But if it's at the cursor or has rmap_items linked to it, use
2503 * mmap_sem to synchronize with any break_cows before pagetables
2504 * are freed, and leave the mm_slot on the list for ksmd to free.
2505 * Beware: ksm may already have noticed it exiting and freed the slot.
2508 spin_lock(&ksm_mmlist_lock);
2509 mm_slot = get_mm_slot(mm);
2510 if (mm_slot && ksm_scan.mm_slot != mm_slot) {
2511 if (!mm_slot->rmap_list) {
2512 hash_del(&mm_slot->link);
2513 list_del(&mm_slot->mm_list);
2516 list_move(&mm_slot->mm_list,
2517 &ksm_scan.mm_slot->mm_list);
2520 spin_unlock(&ksm_mmlist_lock);
2523 free_mm_slot(mm_slot);
2524 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
2526 } else if (mm_slot) {
2527 down_write(&mm->mmap_sem);
2528 up_write(&mm->mmap_sem);
2532 struct page *ksm_might_need_to_copy(struct page *page,
2533 struct vm_area_struct *vma, unsigned long address)
2535 struct anon_vma *anon_vma = page_anon_vma(page);
2536 struct page *new_page;
2538 if (PageKsm(page)) {
2539 if (page_stable_node(page) &&
2540 !(ksm_run & KSM_RUN_UNMERGE))
2541 return page; /* no need to copy it */
2542 } else if (!anon_vma) {
2543 return page; /* no need to copy it */
2544 } else if (anon_vma->root == vma->anon_vma->root &&
2545 page->index == linear_page_index(vma, address)) {
2546 return page; /* still no need to copy it */
2548 if (!PageUptodate(page))
2549 return page; /* let do_swap_page report the error */
2551 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2553 copy_user_highpage(new_page, page, address, vma);
2555 SetPageDirty(new_page);
2556 __SetPageUptodate(new_page);
2557 __SetPageLocked(new_page);
2563 void rmap_walk_ksm(struct page *page, struct rmap_walk_control *rwc)
2565 struct stable_node *stable_node;
2566 struct rmap_item *rmap_item;
2567 int search_new_forks = 0;
2569 VM_BUG_ON_PAGE(!PageKsm(page), page);
2572 * Rely on the page lock to protect against concurrent modifications
2573 * to that page's node of the stable tree.
2575 VM_BUG_ON_PAGE(!PageLocked(page), page);
2577 stable_node = page_stable_node(page);
2581 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
2582 struct anon_vma *anon_vma = rmap_item->anon_vma;
2583 struct anon_vma_chain *vmac;
2584 struct vm_area_struct *vma;
2587 anon_vma_lock_read(anon_vma);
2588 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
2592 if (rmap_item->address < vma->vm_start ||
2593 rmap_item->address >= vma->vm_end)
2596 * Initially we examine only the vma which covers this
2597 * rmap_item; but later, if there is still work to do,
2598 * we examine covering vmas in other mms: in case they
2599 * were forked from the original since ksmd passed.
2601 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
2604 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2607 if (!rwc->rmap_one(page, vma,
2608 rmap_item->address, rwc->arg)) {
2609 anon_vma_unlock_read(anon_vma);
2612 if (rwc->done && rwc->done(page)) {
2613 anon_vma_unlock_read(anon_vma);
2617 anon_vma_unlock_read(anon_vma);
2619 if (!search_new_forks++)
2623 #ifdef CONFIG_MIGRATION
2624 void ksm_migrate_page(struct page *newpage, struct page *oldpage)
2626 struct stable_node *stable_node;
2628 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
2629 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
2630 VM_BUG_ON_PAGE(newpage->mapping != oldpage->mapping, newpage);
2632 stable_node = page_stable_node(newpage);
2634 VM_BUG_ON_PAGE(stable_node->kpfn != page_to_pfn(oldpage), oldpage);
2635 stable_node->kpfn = page_to_pfn(newpage);
2637 * newpage->mapping was set in advance; now we need smp_wmb()
2638 * to make sure that the new stable_node->kpfn is visible
2639 * to get_ksm_page() before it can see that oldpage->mapping
2640 * has gone stale (or that PageSwapCache has been cleared).
2643 set_page_stable_node(oldpage, NULL);
2646 #endif /* CONFIG_MIGRATION */
2648 #ifdef CONFIG_MEMORY_HOTREMOVE
2649 static void wait_while_offlining(void)
2651 while (ksm_run & KSM_RUN_OFFLINE) {
2652 mutex_unlock(&ksm_thread_mutex);
2653 wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE),
2654 TASK_UNINTERRUPTIBLE);
2655 mutex_lock(&ksm_thread_mutex);
2659 static bool stable_node_dup_remove_range(struct stable_node *stable_node,
2660 unsigned long start_pfn,
2661 unsigned long end_pfn)
2663 if (stable_node->kpfn >= start_pfn &&
2664 stable_node->kpfn < end_pfn) {
2666 * Don't get_ksm_page, page has already gone:
2667 * which is why we keep kpfn instead of page*
2669 remove_node_from_stable_tree(stable_node);
2675 static bool stable_node_chain_remove_range(struct stable_node *stable_node,
2676 unsigned long start_pfn,
2677 unsigned long end_pfn,
2678 struct rb_root *root)
2680 struct stable_node *dup;
2681 struct hlist_node *hlist_safe;
2683 if (!is_stable_node_chain(stable_node)) {
2684 VM_BUG_ON(is_stable_node_dup(stable_node));
2685 return stable_node_dup_remove_range(stable_node, start_pfn,
2689 hlist_for_each_entry_safe(dup, hlist_safe,
2690 &stable_node->hlist, hlist_dup) {
2691 VM_BUG_ON(!is_stable_node_dup(dup));
2692 stable_node_dup_remove_range(dup, start_pfn, end_pfn);
2694 if (hlist_empty(&stable_node->hlist)) {
2695 free_stable_node_chain(stable_node, root);
2696 return true; /* notify caller that tree was rebalanced */
2701 static void ksm_check_stable_tree(unsigned long start_pfn,
2702 unsigned long end_pfn)
2704 struct stable_node *stable_node, *next;
2705 struct rb_node *node;
2708 for (nid = 0; nid < ksm_nr_node_ids; nid++) {
2709 node = rb_first(root_stable_tree + nid);
2711 stable_node = rb_entry(node, struct stable_node, node);
2712 if (stable_node_chain_remove_range(stable_node,
2716 node = rb_first(root_stable_tree + nid);
2718 node = rb_next(node);
2722 list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
2723 if (stable_node->kpfn >= start_pfn &&
2724 stable_node->kpfn < end_pfn)
2725 remove_node_from_stable_tree(stable_node);
2730 static int ksm_memory_callback(struct notifier_block *self,
2731 unsigned long action, void *arg)
2733 struct memory_notify *mn = arg;
2736 case MEM_GOING_OFFLINE:
2738 * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items()
2739 * and remove_all_stable_nodes() while memory is going offline:
2740 * it is unsafe for them to touch the stable tree at this time.
2741 * But unmerge_ksm_pages(), rmap lookups and other entry points
2742 * which do not need the ksm_thread_mutex are all safe.
2744 mutex_lock(&ksm_thread_mutex);
2745 ksm_run |= KSM_RUN_OFFLINE;
2746 mutex_unlock(&ksm_thread_mutex);
2751 * Most of the work is done by page migration; but there might
2752 * be a few stable_nodes left over, still pointing to struct
2753 * pages which have been offlined: prune those from the tree,
2754 * otherwise get_ksm_page() might later try to access a
2755 * non-existent struct page.
2757 ksm_check_stable_tree(mn->start_pfn,
2758 mn->start_pfn + mn->nr_pages);
2761 case MEM_CANCEL_OFFLINE:
2762 mutex_lock(&ksm_thread_mutex);
2763 ksm_run &= ~KSM_RUN_OFFLINE;
2764 mutex_unlock(&ksm_thread_mutex);
2766 smp_mb(); /* wake_up_bit advises this */
2767 wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE));
2773 static void wait_while_offlining(void)
2776 #endif /* CONFIG_MEMORY_HOTREMOVE */
2780 * This all compiles without CONFIG_SYSFS, but is a waste of space.
2783 #define KSM_ATTR_RO(_name) \
2784 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2785 #define KSM_ATTR(_name) \
2786 static struct kobj_attribute _name##_attr = \
2787 __ATTR(_name, 0644, _name##_show, _name##_store)
2789 static ssize_t sleep_millisecs_show(struct kobject *kobj,
2790 struct kobj_attribute *attr, char *buf)
2792 return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
2795 static ssize_t sleep_millisecs_store(struct kobject *kobj,
2796 struct kobj_attribute *attr,
2797 const char *buf, size_t count)
2799 unsigned long msecs;
2802 err = kstrtoul(buf, 10, &msecs);
2803 if (err || msecs > UINT_MAX)
2806 ksm_thread_sleep_millisecs = msecs;
2810 KSM_ATTR(sleep_millisecs);
2812 static ssize_t pages_to_scan_show(struct kobject *kobj,
2813 struct kobj_attribute *attr, char *buf)
2815 return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
2818 static ssize_t pages_to_scan_store(struct kobject *kobj,
2819 struct kobj_attribute *attr,
2820 const char *buf, size_t count)
2823 unsigned long nr_pages;
2825 err = kstrtoul(buf, 10, &nr_pages);
2826 if (err || nr_pages > UINT_MAX)
2829 ksm_thread_pages_to_scan = nr_pages;
2833 KSM_ATTR(pages_to_scan);
2835 static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
2838 return sprintf(buf, "%lu\n", ksm_run);
2841 static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
2842 const char *buf, size_t count)
2845 unsigned long flags;
2847 err = kstrtoul(buf, 10, &flags);
2848 if (err || flags > UINT_MAX)
2850 if (flags > KSM_RUN_UNMERGE)
2854 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
2855 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
2856 * breaking COW to free the pages_shared (but leaves mm_slots
2857 * on the list for when ksmd may be set running again).
2860 mutex_lock(&ksm_thread_mutex);
2861 wait_while_offlining();
2862 if (ksm_run != flags) {
2864 if (flags & KSM_RUN_UNMERGE) {
2865 set_current_oom_origin();
2866 err = unmerge_and_remove_all_rmap_items();
2867 clear_current_oom_origin();
2869 ksm_run = KSM_RUN_STOP;
2874 mutex_unlock(&ksm_thread_mutex);
2876 if (flags & KSM_RUN_MERGE)
2877 wake_up_interruptible(&ksm_thread_wait);
2884 static ssize_t merge_across_nodes_show(struct kobject *kobj,
2885 struct kobj_attribute *attr, char *buf)
2887 return sprintf(buf, "%u\n", ksm_merge_across_nodes);
2890 static ssize_t merge_across_nodes_store(struct kobject *kobj,
2891 struct kobj_attribute *attr,
2892 const char *buf, size_t count)
2897 err = kstrtoul(buf, 10, &knob);
2903 mutex_lock(&ksm_thread_mutex);
2904 wait_while_offlining();
2905 if (ksm_merge_across_nodes != knob) {
2906 if (ksm_pages_shared || remove_all_stable_nodes())
2908 else if (root_stable_tree == one_stable_tree) {
2909 struct rb_root *buf;
2911 * This is the first time that we switch away from the
2912 * default of merging across nodes: must now allocate
2913 * a buffer to hold as many roots as may be needed.
2914 * Allocate stable and unstable together:
2915 * MAXSMP NODES_SHIFT 10 will use 16kB.
2917 buf = kcalloc(nr_node_ids + nr_node_ids, sizeof(*buf),
2919 /* Let us assume that RB_ROOT is NULL is zero */
2923 root_stable_tree = buf;
2924 root_unstable_tree = buf + nr_node_ids;
2925 /* Stable tree is empty but not the unstable */
2926 root_unstable_tree[0] = one_unstable_tree[0];
2930 ksm_merge_across_nodes = knob;
2931 ksm_nr_node_ids = knob ? 1 : nr_node_ids;
2934 mutex_unlock(&ksm_thread_mutex);
2936 return err ? err : count;
2938 KSM_ATTR(merge_across_nodes);
2941 static ssize_t use_zero_pages_show(struct kobject *kobj,
2942 struct kobj_attribute *attr, char *buf)
2944 return sprintf(buf, "%u\n", ksm_use_zero_pages);
2946 static ssize_t use_zero_pages_store(struct kobject *kobj,
2947 struct kobj_attribute *attr,
2948 const char *buf, size_t count)
2953 err = kstrtobool(buf, &value);
2957 ksm_use_zero_pages = value;
2961 KSM_ATTR(use_zero_pages);
2963 static ssize_t max_page_sharing_show(struct kobject *kobj,
2964 struct kobj_attribute *attr, char *buf)
2966 return sprintf(buf, "%u\n", ksm_max_page_sharing);
2969 static ssize_t max_page_sharing_store(struct kobject *kobj,
2970 struct kobj_attribute *attr,
2971 const char *buf, size_t count)
2976 err = kstrtoint(buf, 10, &knob);
2980 * When a KSM page is created it is shared by 2 mappings. This
2981 * being a signed comparison, it implicitly verifies it's not
2987 if (READ_ONCE(ksm_max_page_sharing) == knob)
2990 mutex_lock(&ksm_thread_mutex);
2991 wait_while_offlining();
2992 if (ksm_max_page_sharing != knob) {
2993 if (ksm_pages_shared || remove_all_stable_nodes())
2996 ksm_max_page_sharing = knob;
2998 mutex_unlock(&ksm_thread_mutex);
3000 return err ? err : count;
3002 KSM_ATTR(max_page_sharing);
3004 static ssize_t pages_shared_show(struct kobject *kobj,
3005 struct kobj_attribute *attr, char *buf)
3007 return sprintf(buf, "%lu\n", ksm_pages_shared);
3009 KSM_ATTR_RO(pages_shared);
3011 static ssize_t pages_sharing_show(struct kobject *kobj,
3012 struct kobj_attribute *attr, char *buf)
3014 return sprintf(buf, "%lu\n", ksm_pages_sharing);
3016 KSM_ATTR_RO(pages_sharing);
3018 static ssize_t pages_unshared_show(struct kobject *kobj,
3019 struct kobj_attribute *attr, char *buf)
3021 return sprintf(buf, "%lu\n", ksm_pages_unshared);
3023 KSM_ATTR_RO(pages_unshared);
3025 static ssize_t pages_volatile_show(struct kobject *kobj,
3026 struct kobj_attribute *attr, char *buf)
3028 long ksm_pages_volatile;
3030 ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
3031 - ksm_pages_sharing - ksm_pages_unshared;
3033 * It was not worth any locking to calculate that statistic,
3034 * but it might therefore sometimes be negative: conceal that.
3036 if (ksm_pages_volatile < 0)
3037 ksm_pages_volatile = 0;
3038 return sprintf(buf, "%ld\n", ksm_pages_volatile);
3040 KSM_ATTR_RO(pages_volatile);
3042 static ssize_t stable_node_dups_show(struct kobject *kobj,
3043 struct kobj_attribute *attr, char *buf)
3045 return sprintf(buf, "%lu\n", ksm_stable_node_dups);
3047 KSM_ATTR_RO(stable_node_dups);
3049 static ssize_t stable_node_chains_show(struct kobject *kobj,
3050 struct kobj_attribute *attr, char *buf)
3052 return sprintf(buf, "%lu\n", ksm_stable_node_chains);
3054 KSM_ATTR_RO(stable_node_chains);
3057 stable_node_chains_prune_millisecs_show(struct kobject *kobj,
3058 struct kobj_attribute *attr,
3061 return sprintf(buf, "%u\n", ksm_stable_node_chains_prune_millisecs);
3065 stable_node_chains_prune_millisecs_store(struct kobject *kobj,
3066 struct kobj_attribute *attr,
3067 const char *buf, size_t count)
3069 unsigned long msecs;
3072 err = kstrtoul(buf, 10, &msecs);
3073 if (err || msecs > UINT_MAX)
3076 ksm_stable_node_chains_prune_millisecs = msecs;
3080 KSM_ATTR(stable_node_chains_prune_millisecs);
3082 static ssize_t full_scans_show(struct kobject *kobj,
3083 struct kobj_attribute *attr, char *buf)
3085 return sprintf(buf, "%lu\n", ksm_scan.seqnr);
3087 KSM_ATTR_RO(full_scans);
3089 static struct attribute *ksm_attrs[] = {
3090 &sleep_millisecs_attr.attr,
3091 &pages_to_scan_attr.attr,
3093 &pages_shared_attr.attr,
3094 &pages_sharing_attr.attr,
3095 &pages_unshared_attr.attr,
3096 &pages_volatile_attr.attr,
3097 &full_scans_attr.attr,
3099 &merge_across_nodes_attr.attr,
3101 &max_page_sharing_attr.attr,
3102 &stable_node_chains_attr.attr,
3103 &stable_node_dups_attr.attr,
3104 &stable_node_chains_prune_millisecs_attr.attr,
3105 &use_zero_pages_attr.attr,
3109 static const struct attribute_group ksm_attr_group = {
3113 #endif /* CONFIG_SYSFS */
3115 static int __init ksm_init(void)
3117 struct task_struct *ksm_thread;
3120 /* The correct value depends on page size and endianness */
3121 zero_checksum = calc_checksum(ZERO_PAGE(0));
3122 /* Default to false for backwards compatibility */
3123 ksm_use_zero_pages = false;
3125 err = ksm_slab_init();
3129 ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
3130 if (IS_ERR(ksm_thread)) {
3131 pr_err("ksm: creating kthread failed\n");
3132 err = PTR_ERR(ksm_thread);
3137 err = sysfs_create_group(mm_kobj, &ksm_attr_group);
3139 pr_err("ksm: register sysfs failed\n");
3140 kthread_stop(ksm_thread);
3144 ksm_run = KSM_RUN_MERGE; /* no way for user to start it */
3146 #endif /* CONFIG_SYSFS */
3148 #ifdef CONFIG_MEMORY_HOTREMOVE
3149 /* There is no significance to this priority 100 */
3150 hotplug_memory_notifier(ksm_memory_callback, 100);
3159 subsys_initcall(ksm_init);