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 */
219 #define KSM_FLAG_MASK (SEQNR_MASK|UNSTABLE_FLAG|STABLE_FLAG)
220 /* to mask all the flags */
222 /* The stable and unstable tree heads */
223 static struct rb_root one_stable_tree[1] = { RB_ROOT };
224 static struct rb_root one_unstable_tree[1] = { RB_ROOT };
225 static struct rb_root *root_stable_tree = one_stable_tree;
226 static struct rb_root *root_unstable_tree = one_unstable_tree;
228 /* Recently migrated nodes of stable tree, pending proper placement */
229 static LIST_HEAD(migrate_nodes);
230 #define STABLE_NODE_DUP_HEAD ((struct list_head *)&migrate_nodes.prev)
232 #define MM_SLOTS_HASH_BITS 10
233 static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
235 static struct mm_slot ksm_mm_head = {
236 .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
238 static struct ksm_scan ksm_scan = {
239 .mm_slot = &ksm_mm_head,
242 static struct kmem_cache *rmap_item_cache;
243 static struct kmem_cache *stable_node_cache;
244 static struct kmem_cache *mm_slot_cache;
246 /* The number of nodes in the stable tree */
247 static unsigned long ksm_pages_shared;
249 /* The number of page slots additionally sharing those nodes */
250 static unsigned long ksm_pages_sharing;
252 /* The number of nodes in the unstable tree */
253 static unsigned long ksm_pages_unshared;
255 /* The number of rmap_items in use: to calculate pages_volatile */
256 static unsigned long ksm_rmap_items;
258 /* The number of stable_node chains */
259 static unsigned long ksm_stable_node_chains;
261 /* The number of stable_node dups linked to the stable_node chains */
262 static unsigned long ksm_stable_node_dups;
264 /* Delay in pruning stale stable_node_dups in the stable_node_chains */
265 static int ksm_stable_node_chains_prune_millisecs = 2000;
267 /* Maximum number of page slots sharing a stable node */
268 static int ksm_max_page_sharing = 256;
270 /* Number of pages ksmd should scan in one batch */
271 static unsigned int ksm_thread_pages_to_scan = 100;
273 /* Milliseconds ksmd should sleep between batches */
274 static unsigned int ksm_thread_sleep_millisecs = 20;
276 /* Checksum of an empty (zeroed) page */
277 static unsigned int zero_checksum __read_mostly;
279 /* Whether to merge empty (zeroed) pages with actual zero pages */
280 static bool ksm_use_zero_pages __read_mostly;
283 /* Zeroed when merging across nodes is not allowed */
284 static unsigned int ksm_merge_across_nodes = 1;
285 static int ksm_nr_node_ids = 1;
287 #define ksm_merge_across_nodes 1U
288 #define ksm_nr_node_ids 1
291 #define KSM_RUN_STOP 0
292 #define KSM_RUN_MERGE 1
293 #define KSM_RUN_UNMERGE 2
294 #define KSM_RUN_OFFLINE 4
295 static unsigned long ksm_run = KSM_RUN_STOP;
296 static void wait_while_offlining(void);
298 static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
299 static DEFINE_MUTEX(ksm_thread_mutex);
300 static DEFINE_SPINLOCK(ksm_mmlist_lock);
302 #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
303 sizeof(struct __struct), __alignof__(struct __struct),\
306 static int __init ksm_slab_init(void)
308 rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
309 if (!rmap_item_cache)
312 stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
313 if (!stable_node_cache)
316 mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
323 kmem_cache_destroy(stable_node_cache);
325 kmem_cache_destroy(rmap_item_cache);
330 static void __init ksm_slab_free(void)
332 kmem_cache_destroy(mm_slot_cache);
333 kmem_cache_destroy(stable_node_cache);
334 kmem_cache_destroy(rmap_item_cache);
335 mm_slot_cache = NULL;
338 static __always_inline bool is_stable_node_chain(struct stable_node *chain)
340 return chain->rmap_hlist_len == STABLE_NODE_CHAIN;
343 static __always_inline bool is_stable_node_dup(struct stable_node *dup)
345 return dup->head == STABLE_NODE_DUP_HEAD;
348 static inline void stable_node_chain_add_dup(struct stable_node *dup,
349 struct stable_node *chain)
351 VM_BUG_ON(is_stable_node_dup(dup));
352 dup->head = STABLE_NODE_DUP_HEAD;
353 VM_BUG_ON(!is_stable_node_chain(chain));
354 hlist_add_head(&dup->hlist_dup, &chain->hlist);
355 ksm_stable_node_dups++;
358 static inline void __stable_node_dup_del(struct stable_node *dup)
360 VM_BUG_ON(!is_stable_node_dup(dup));
361 hlist_del(&dup->hlist_dup);
362 ksm_stable_node_dups--;
365 static inline void stable_node_dup_del(struct stable_node *dup)
367 VM_BUG_ON(is_stable_node_chain(dup));
368 if (is_stable_node_dup(dup))
369 __stable_node_dup_del(dup);
371 rb_erase(&dup->node, root_stable_tree + NUMA(dup->nid));
372 #ifdef CONFIG_DEBUG_VM
377 static inline struct rmap_item *alloc_rmap_item(void)
379 struct rmap_item *rmap_item;
381 rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL |
382 __GFP_NORETRY | __GFP_NOWARN);
388 static inline void free_rmap_item(struct rmap_item *rmap_item)
391 rmap_item->mm = NULL; /* debug safety */
392 kmem_cache_free(rmap_item_cache, rmap_item);
395 static inline struct stable_node *alloc_stable_node(void)
398 * The allocation can take too long with GFP_KERNEL when memory is under
399 * pressure, which may lead to hung task warnings. Adding __GFP_HIGH
400 * grants access to memory reserves, helping to avoid this problem.
402 return kmem_cache_alloc(stable_node_cache, GFP_KERNEL | __GFP_HIGH);
405 static inline void free_stable_node(struct stable_node *stable_node)
407 VM_BUG_ON(stable_node->rmap_hlist_len &&
408 !is_stable_node_chain(stable_node));
409 kmem_cache_free(stable_node_cache, stable_node);
412 static inline struct mm_slot *alloc_mm_slot(void)
414 if (!mm_slot_cache) /* initialization failed */
416 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
419 static inline void free_mm_slot(struct mm_slot *mm_slot)
421 kmem_cache_free(mm_slot_cache, mm_slot);
424 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
426 struct mm_slot *slot;
428 hash_for_each_possible(mm_slots_hash, slot, link, (unsigned long)mm)
435 static void insert_to_mm_slots_hash(struct mm_struct *mm,
436 struct mm_slot *mm_slot)
439 hash_add(mm_slots_hash, &mm_slot->link, (unsigned long)mm);
443 * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
444 * page tables after it has passed through ksm_exit() - which, if necessary,
445 * takes mmap_sem briefly to serialize against them. ksm_exit() does not set
446 * a special flag: they can just back out as soon as mm_users goes to zero.
447 * ksm_test_exit() is used throughout to make this test for exit: in some
448 * places for correctness, in some places just to avoid unnecessary work.
450 static inline bool ksm_test_exit(struct mm_struct *mm)
452 return atomic_read(&mm->mm_users) == 0;
456 * We use break_ksm to break COW on a ksm page: it's a stripped down
458 * if (get_user_pages(addr, 1, 1, 1, &page, NULL) == 1)
461 * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
462 * in case the application has unmapped and remapped mm,addr meanwhile.
463 * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP
464 * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
466 * FAULT_FLAG/FOLL_REMOTE are because we do this outside the context
467 * of the process that owns 'vma'. We also do not want to enforce
468 * protection keys here anyway.
470 static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
477 page = follow_page(vma, addr,
478 FOLL_GET | FOLL_MIGRATION | FOLL_REMOTE);
479 if (IS_ERR_OR_NULL(page))
482 ret = handle_mm_fault(vma, addr,
483 FAULT_FLAG_WRITE | FAULT_FLAG_REMOTE);
485 ret = VM_FAULT_WRITE;
487 } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV | VM_FAULT_OOM)));
489 * We must loop because handle_mm_fault() may back out if there's
490 * any difficulty e.g. if pte accessed bit gets updated concurrently.
492 * VM_FAULT_WRITE is what we have been hoping for: it indicates that
493 * COW has been broken, even if the vma does not permit VM_WRITE;
494 * but note that a concurrent fault might break PageKsm for us.
496 * VM_FAULT_SIGBUS could occur if we race with truncation of the
497 * backing file, which also invalidates anonymous pages: that's
498 * okay, that truncation will have unmapped the PageKsm for us.
500 * VM_FAULT_OOM: at the time of writing (late July 2009), setting
501 * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
502 * current task has TIF_MEMDIE set, and will be OOM killed on return
503 * to user; and ksmd, having no mm, would never be chosen for that.
505 * But if the mm is in a limited mem_cgroup, then the fault may fail
506 * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
507 * even ksmd can fail in this way - though it's usually breaking ksm
508 * just to undo a merge it made a moment before, so unlikely to oom.
510 * That's a pity: we might therefore have more kernel pages allocated
511 * than we're counting as nodes in the stable tree; but ksm_do_scan
512 * will retry to break_cow on each pass, so should recover the page
513 * in due course. The important thing is to not let VM_MERGEABLE
514 * be cleared while any such pages might remain in the area.
516 return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
519 static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm,
522 struct vm_area_struct *vma;
523 if (ksm_test_exit(mm))
525 vma = find_vma(mm, addr);
526 if (!vma || vma->vm_start > addr)
528 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
533 static void break_cow(struct rmap_item *rmap_item)
535 struct mm_struct *mm = rmap_item->mm;
536 unsigned long addr = rmap_item->address;
537 struct vm_area_struct *vma;
540 * It is not an accident that whenever we want to break COW
541 * to undo, we also need to drop a reference to the anon_vma.
543 put_anon_vma(rmap_item->anon_vma);
545 down_read(&mm->mmap_sem);
546 vma = find_mergeable_vma(mm, addr);
548 break_ksm(vma, addr);
549 up_read(&mm->mmap_sem);
552 static struct page *get_mergeable_page(struct rmap_item *rmap_item)
554 struct mm_struct *mm = rmap_item->mm;
555 unsigned long addr = rmap_item->address;
556 struct vm_area_struct *vma;
559 down_read(&mm->mmap_sem);
560 vma = find_mergeable_vma(mm, addr);
564 page = follow_page(vma, addr, FOLL_GET);
565 if (IS_ERR_OR_NULL(page))
567 if (PageAnon(page)) {
568 flush_anon_page(vma, page, addr);
569 flush_dcache_page(page);
575 up_read(&mm->mmap_sem);
580 * This helper is used for getting right index into array of tree roots.
581 * When merge_across_nodes knob is set to 1, there are only two rb-trees for
582 * stable and unstable pages from all nodes with roots in index 0. Otherwise,
583 * every node has its own stable and unstable tree.
585 static inline int get_kpfn_nid(unsigned long kpfn)
587 return ksm_merge_across_nodes ? 0 : NUMA(pfn_to_nid(kpfn));
590 static struct stable_node *alloc_stable_node_chain(struct stable_node *dup,
591 struct rb_root *root)
593 struct stable_node *chain = alloc_stable_node();
594 VM_BUG_ON(is_stable_node_chain(dup));
596 INIT_HLIST_HEAD(&chain->hlist);
597 chain->chain_prune_time = jiffies;
598 chain->rmap_hlist_len = STABLE_NODE_CHAIN;
599 #if defined (CONFIG_DEBUG_VM) && defined(CONFIG_NUMA)
600 chain->nid = -1; /* debug */
602 ksm_stable_node_chains++;
605 * Put the stable node chain in the first dimension of
606 * the stable tree and at the same time remove the old
609 rb_replace_node(&dup->node, &chain->node, root);
612 * Move the old stable node to the second dimension
613 * queued in the hlist_dup. The invariant is that all
614 * dup stable_nodes in the chain->hlist point to pages
615 * that are wrprotected and have the exact same
618 stable_node_chain_add_dup(dup, chain);
623 static inline void free_stable_node_chain(struct stable_node *chain,
624 struct rb_root *root)
626 rb_erase(&chain->node, root);
627 free_stable_node(chain);
628 ksm_stable_node_chains--;
631 static void remove_node_from_stable_tree(struct stable_node *stable_node)
633 struct rmap_item *rmap_item;
635 /* check it's not STABLE_NODE_CHAIN or negative */
636 BUG_ON(stable_node->rmap_hlist_len < 0);
638 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
639 if (rmap_item->hlist.next)
643 VM_BUG_ON(stable_node->rmap_hlist_len <= 0);
644 stable_node->rmap_hlist_len--;
645 put_anon_vma(rmap_item->anon_vma);
646 rmap_item->address &= PAGE_MASK;
651 * We need the second aligned pointer of the migrate_nodes
652 * list_head to stay clear from the rb_parent_color union
653 * (aligned and different than any node) and also different
654 * from &migrate_nodes. This will verify that future list.h changes
655 * don't break STABLE_NODE_DUP_HEAD. Only recent gcc can handle it.
657 #if defined(GCC_VERSION) && GCC_VERSION >= 40903
658 BUILD_BUG_ON(STABLE_NODE_DUP_HEAD <= &migrate_nodes);
659 BUILD_BUG_ON(STABLE_NODE_DUP_HEAD >= &migrate_nodes + 1);
662 if (stable_node->head == &migrate_nodes)
663 list_del(&stable_node->list);
665 stable_node_dup_del(stable_node);
666 free_stable_node(stable_node);
670 * get_ksm_page: checks if the page indicated by the stable node
671 * is still its ksm page, despite having held no reference to it.
672 * In which case we can trust the content of the page, and it
673 * returns the gotten page; but if the page has now been zapped,
674 * remove the stale node from the stable tree and return NULL.
675 * But beware, the stable node's page might be being migrated.
677 * You would expect the stable_node to hold a reference to the ksm page.
678 * But if it increments the page's count, swapping out has to wait for
679 * ksmd to come around again before it can free the page, which may take
680 * seconds or even minutes: much too unresponsive. So instead we use a
681 * "keyhole reference": access to the ksm page from the stable node peeps
682 * out through its keyhole to see if that page still holds the right key,
683 * pointing back to this stable node. This relies on freeing a PageAnon
684 * page to reset its page->mapping to NULL, and relies on no other use of
685 * a page to put something that might look like our key in page->mapping.
686 * is on its way to being freed; but it is an anomaly to bear in mind.
688 static struct page *get_ksm_page(struct stable_node *stable_node, bool lock_it)
691 void *expected_mapping;
694 expected_mapping = (void *)((unsigned long)stable_node |
697 kpfn = READ_ONCE(stable_node->kpfn); /* Address dependency. */
698 page = pfn_to_page(kpfn);
699 if (READ_ONCE(page->mapping) != expected_mapping)
703 * We cannot do anything with the page while its refcount is 0.
704 * Usually 0 means free, or tail of a higher-order page: in which
705 * case this node is no longer referenced, and should be freed;
706 * however, it might mean that the page is under page_ref_freeze().
707 * The __remove_mapping() case is easy, again the node is now stale;
708 * but if page is swapcache in migrate_page_move_mapping(), it might
709 * still be our page, in which case it's essential to keep the node.
711 while (!get_page_unless_zero(page)) {
713 * Another check for page->mapping != expected_mapping would
714 * work here too. We have chosen the !PageSwapCache test to
715 * optimize the common case, when the page is or is about to
716 * be freed: PageSwapCache is cleared (under spin_lock_irq)
717 * in the ref_freeze section of __remove_mapping(); but Anon
718 * page->mapping reset to NULL later, in free_pages_prepare().
720 if (!PageSwapCache(page))
725 if (READ_ONCE(page->mapping) != expected_mapping) {
732 if (READ_ONCE(page->mapping) != expected_mapping) {
742 * We come here from above when page->mapping or !PageSwapCache
743 * suggests that the node is stale; but it might be under migration.
744 * We need smp_rmb(), matching the smp_wmb() in ksm_migrate_page(),
745 * before checking whether node->kpfn has been changed.
748 if (READ_ONCE(stable_node->kpfn) != kpfn)
750 remove_node_from_stable_tree(stable_node);
755 * Removing rmap_item from stable or unstable tree.
756 * This function will clean the information from the stable/unstable tree.
758 static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
760 if (rmap_item->address & STABLE_FLAG) {
761 struct stable_node *stable_node;
764 stable_node = rmap_item->head;
765 page = get_ksm_page(stable_node, true);
769 hlist_del(&rmap_item->hlist);
773 if (!hlist_empty(&stable_node->hlist))
777 VM_BUG_ON(stable_node->rmap_hlist_len <= 0);
778 stable_node->rmap_hlist_len--;
780 put_anon_vma(rmap_item->anon_vma);
781 rmap_item->head = NULL;
782 rmap_item->address &= PAGE_MASK;
784 } else if (rmap_item->address & UNSTABLE_FLAG) {
787 * Usually ksmd can and must skip the rb_erase, because
788 * root_unstable_tree was already reset to RB_ROOT.
789 * But be careful when an mm is exiting: do the rb_erase
790 * if this rmap_item was inserted by this scan, rather
791 * than left over from before.
793 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
796 rb_erase(&rmap_item->node,
797 root_unstable_tree + NUMA(rmap_item->nid));
798 ksm_pages_unshared--;
799 rmap_item->address &= PAGE_MASK;
802 cond_resched(); /* we're called from many long loops */
805 static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
806 struct rmap_item **rmap_list)
809 struct rmap_item *rmap_item = *rmap_list;
810 *rmap_list = rmap_item->rmap_list;
811 remove_rmap_item_from_tree(rmap_item);
812 free_rmap_item(rmap_item);
817 * Though it's very tempting to unmerge rmap_items from stable tree rather
818 * than check every pte of a given vma, the locking doesn't quite work for
819 * that - an rmap_item is assigned to the stable tree after inserting ksm
820 * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing
821 * rmap_items from parent to child at fork time (so as not to waste time
822 * if exit comes before the next scan reaches it).
824 * Similarly, although we'd like to remove rmap_items (so updating counts
825 * and freeing memory) when unmerging an area, it's easier to leave that
826 * to the next pass of ksmd - consider, for example, how ksmd might be
827 * in cmp_and_merge_page on one of the rmap_items we would be removing.
829 static int unmerge_ksm_pages(struct vm_area_struct *vma,
830 unsigned long start, unsigned long end)
835 for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
836 if (ksm_test_exit(vma->vm_mm))
838 if (signal_pending(current))
841 err = break_ksm(vma, addr);
846 static inline struct stable_node *page_stable_node(struct page *page)
848 return PageKsm(page) ? page_rmapping(page) : NULL;
851 static inline void set_page_stable_node(struct page *page,
852 struct stable_node *stable_node)
854 page->mapping = (void *)((unsigned long)stable_node | PAGE_MAPPING_KSM);
859 * Only called through the sysfs control interface:
861 static int remove_stable_node(struct stable_node *stable_node)
866 page = get_ksm_page(stable_node, true);
869 * get_ksm_page did remove_node_from_stable_tree itself.
875 * Page could be still mapped if this races with __mmput() running in
876 * between ksm_exit() and exit_mmap(). Just refuse to let
877 * merge_across_nodes/max_page_sharing be switched.
880 if (!page_mapped(page)) {
882 * The stable node did not yet appear stale to get_ksm_page(),
883 * since that allows for an unmapped ksm page to be recognized
884 * right up until it is freed; but the node is safe to remove.
885 * This page might be in a pagevec waiting to be freed,
886 * or it might be PageSwapCache (perhaps under writeback),
887 * or it might have been removed from swapcache a moment ago.
889 set_page_stable_node(page, NULL);
890 remove_node_from_stable_tree(stable_node);
899 static int remove_stable_node_chain(struct stable_node *stable_node,
900 struct rb_root *root)
902 struct stable_node *dup;
903 struct hlist_node *hlist_safe;
905 if (!is_stable_node_chain(stable_node)) {
906 VM_BUG_ON(is_stable_node_dup(stable_node));
907 if (remove_stable_node(stable_node))
913 hlist_for_each_entry_safe(dup, hlist_safe,
914 &stable_node->hlist, hlist_dup) {
915 VM_BUG_ON(!is_stable_node_dup(dup));
916 if (remove_stable_node(dup))
919 BUG_ON(!hlist_empty(&stable_node->hlist));
920 free_stable_node_chain(stable_node, root);
924 static int remove_all_stable_nodes(void)
926 struct stable_node *stable_node, *next;
930 for (nid = 0; nid < ksm_nr_node_ids; nid++) {
931 while (root_stable_tree[nid].rb_node) {
932 stable_node = rb_entry(root_stable_tree[nid].rb_node,
933 struct stable_node, node);
934 if (remove_stable_node_chain(stable_node,
935 root_stable_tree + nid)) {
937 break; /* proceed to next nid */
942 list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
943 if (remove_stable_node(stable_node))
950 static int unmerge_and_remove_all_rmap_items(void)
952 struct mm_slot *mm_slot;
953 struct mm_struct *mm;
954 struct vm_area_struct *vma;
957 spin_lock(&ksm_mmlist_lock);
958 ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
959 struct mm_slot, mm_list);
960 spin_unlock(&ksm_mmlist_lock);
962 for (mm_slot = ksm_scan.mm_slot;
963 mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
965 down_read(&mm->mmap_sem);
966 for (vma = mm->mmap; vma; vma = vma->vm_next) {
967 if (ksm_test_exit(mm))
969 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
971 err = unmerge_ksm_pages(vma,
972 vma->vm_start, vma->vm_end);
977 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
978 up_read(&mm->mmap_sem);
980 spin_lock(&ksm_mmlist_lock);
981 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
982 struct mm_slot, mm_list);
983 if (ksm_test_exit(mm)) {
984 hash_del(&mm_slot->link);
985 list_del(&mm_slot->mm_list);
986 spin_unlock(&ksm_mmlist_lock);
988 free_mm_slot(mm_slot);
989 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
992 spin_unlock(&ksm_mmlist_lock);
995 /* Clean up stable nodes, but don't worry if some are still busy */
996 remove_all_stable_nodes();
1001 up_read(&mm->mmap_sem);
1002 spin_lock(&ksm_mmlist_lock);
1003 ksm_scan.mm_slot = &ksm_mm_head;
1004 spin_unlock(&ksm_mmlist_lock);
1007 #endif /* CONFIG_SYSFS */
1009 static u32 calc_checksum(struct page *page)
1012 void *addr = kmap_atomic(page);
1013 checksum = jhash2(addr, PAGE_SIZE / 4, 17);
1014 kunmap_atomic(addr);
1018 static int memcmp_pages(struct page *page1, struct page *page2)
1020 char *addr1, *addr2;
1023 addr1 = kmap_atomic(page1);
1024 addr2 = kmap_atomic(page2);
1025 ret = memcmp(addr1, addr2, PAGE_SIZE);
1026 kunmap_atomic(addr2);
1027 kunmap_atomic(addr1);
1031 static inline int pages_identical(struct page *page1, struct page *page2)
1033 return !memcmp_pages(page1, page2);
1036 static int write_protect_page(struct vm_area_struct *vma, struct page *page,
1039 struct mm_struct *mm = vma->vm_mm;
1040 struct page_vma_mapped_walk pvmw = {
1046 unsigned long mmun_start; /* For mmu_notifiers */
1047 unsigned long mmun_end; /* For mmu_notifiers */
1049 pvmw.address = page_address_in_vma(page, vma);
1050 if (pvmw.address == -EFAULT)
1053 BUG_ON(PageTransCompound(page));
1055 mmun_start = pvmw.address;
1056 mmun_end = pvmw.address + PAGE_SIZE;
1057 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1059 if (!page_vma_mapped_walk(&pvmw))
1061 if (WARN_ONCE(!pvmw.pte, "Unexpected PMD mapping?"))
1064 if (pte_write(*pvmw.pte) || pte_dirty(*pvmw.pte) ||
1065 (pte_protnone(*pvmw.pte) && pte_savedwrite(*pvmw.pte)) ||
1066 mm_tlb_flush_pending(mm)) {
1069 swapped = PageSwapCache(page);
1070 flush_cache_page(vma, pvmw.address, page_to_pfn(page));
1072 * Ok this is tricky, when get_user_pages_fast() run it doesn't
1073 * take any lock, therefore the check that we are going to make
1074 * with the pagecount against the mapcount is racey and
1075 * O_DIRECT can happen right after the check.
1076 * So we clear the pte and flush the tlb before the check
1077 * this assure us that no O_DIRECT can happen after the check
1078 * or in the middle of the check.
1080 * No need to notify as we are downgrading page table to read
1081 * only not changing it to point to a new page.
1083 * See Documentation/vm/mmu_notifier.rst
1085 entry = ptep_clear_flush(vma, pvmw.address, pvmw.pte);
1087 * Check that no O_DIRECT or similar I/O is in progress on the
1090 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
1091 set_pte_at(mm, pvmw.address, pvmw.pte, entry);
1094 if (pte_dirty(entry))
1095 set_page_dirty(page);
1097 if (pte_protnone(entry))
1098 entry = pte_mkclean(pte_clear_savedwrite(entry));
1100 entry = pte_mkclean(pte_wrprotect(entry));
1101 set_pte_at_notify(mm, pvmw.address, pvmw.pte, entry);
1103 *orig_pte = *pvmw.pte;
1107 page_vma_mapped_walk_done(&pvmw);
1109 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1115 * replace_page - replace page in vma by new ksm page
1116 * @vma: vma that holds the pte pointing to page
1117 * @page: the page we are replacing by kpage
1118 * @kpage: the ksm page we replace page by
1119 * @orig_pte: the original value of the pte
1121 * Returns 0 on success, -EFAULT on failure.
1123 static int replace_page(struct vm_area_struct *vma, struct page *page,
1124 struct page *kpage, pte_t orig_pte)
1126 struct mm_struct *mm = vma->vm_mm;
1133 unsigned long mmun_start; /* For mmu_notifiers */
1134 unsigned long mmun_end; /* For mmu_notifiers */
1136 addr = page_address_in_vma(page, vma);
1137 if (addr == -EFAULT)
1140 pmd = mm_find_pmd(mm, addr);
1145 mmun_end = addr + PAGE_SIZE;
1146 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1148 ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
1149 if (!pte_same(*ptep, orig_pte)) {
1150 pte_unmap_unlock(ptep, ptl);
1155 * No need to check ksm_use_zero_pages here: we can only have a
1156 * zero_page here if ksm_use_zero_pages was enabled alreaady.
1158 if (!is_zero_pfn(page_to_pfn(kpage))) {
1160 page_add_anon_rmap(kpage, vma, addr, false);
1161 newpte = mk_pte(kpage, vma->vm_page_prot);
1163 newpte = pte_mkspecial(pfn_pte(page_to_pfn(kpage),
1164 vma->vm_page_prot));
1166 * We're replacing an anonymous page with a zero page, which is
1167 * not anonymous. We need to do proper accounting otherwise we
1168 * will get wrong values in /proc, and a BUG message in dmesg
1169 * when tearing down the mm.
1171 dec_mm_counter(mm, MM_ANONPAGES);
1174 flush_cache_page(vma, addr, pte_pfn(*ptep));
1176 * No need to notify as we are replacing a read only page with another
1177 * read only page with the same content.
1179 * See Documentation/vm/mmu_notifier.rst
1181 ptep_clear_flush(vma, addr, ptep);
1182 set_pte_at_notify(mm, addr, ptep, newpte);
1184 page_remove_rmap(page, false);
1185 if (!page_mapped(page))
1186 try_to_free_swap(page);
1189 pte_unmap_unlock(ptep, ptl);
1192 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1198 * try_to_merge_one_page - take two pages and merge them into one
1199 * @vma: the vma that holds the pte pointing to page
1200 * @page: the PageAnon page that we want to replace with kpage
1201 * @kpage: the PageKsm page that we want to map instead of page,
1202 * or NULL the first time when we want to use page as kpage.
1204 * This function returns 0 if the pages were merged, -EFAULT otherwise.
1206 static int try_to_merge_one_page(struct vm_area_struct *vma,
1207 struct page *page, struct page *kpage)
1209 pte_t orig_pte = __pte(0);
1212 if (page == kpage) /* ksm page forked */
1215 if (!PageAnon(page))
1219 * We need the page lock to read a stable PageSwapCache in
1220 * write_protect_page(). We use trylock_page() instead of
1221 * lock_page() because we don't want to wait here - we
1222 * prefer to continue scanning and merging different pages,
1223 * then come back to this page when it is unlocked.
1225 if (!trylock_page(page))
1228 if (PageTransCompound(page)) {
1229 if (split_huge_page(page))
1234 * If this anonymous page is mapped only here, its pte may need
1235 * to be write-protected. If it's mapped elsewhere, all of its
1236 * ptes are necessarily already write-protected. But in either
1237 * case, we need to lock and check page_count is not raised.
1239 if (write_protect_page(vma, page, &orig_pte) == 0) {
1242 * While we hold page lock, upgrade page from
1243 * PageAnon+anon_vma to PageKsm+NULL stable_node:
1244 * stable_tree_insert() will update stable_node.
1246 set_page_stable_node(page, NULL);
1247 mark_page_accessed(page);
1249 * Page reclaim just frees a clean page with no dirty
1250 * ptes: make sure that the ksm page would be swapped.
1252 if (!PageDirty(page))
1255 } else if (pages_identical(page, kpage))
1256 err = replace_page(vma, page, kpage, orig_pte);
1259 if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
1260 munlock_vma_page(page);
1261 if (!PageMlocked(kpage)) {
1264 mlock_vma_page(kpage);
1265 page = kpage; /* for final unlock */
1276 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
1277 * but no new kernel page is allocated: kpage must already be a ksm page.
1279 * This function returns 0 if the pages were merged, -EFAULT otherwise.
1281 static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
1282 struct page *page, struct page *kpage)
1284 struct mm_struct *mm = rmap_item->mm;
1285 struct vm_area_struct *vma;
1288 down_read(&mm->mmap_sem);
1289 vma = find_mergeable_vma(mm, rmap_item->address);
1293 err = try_to_merge_one_page(vma, page, kpage);
1297 /* Unstable nid is in union with stable anon_vma: remove first */
1298 remove_rmap_item_from_tree(rmap_item);
1300 /* Must get reference to anon_vma while still holding mmap_sem */
1301 rmap_item->anon_vma = vma->anon_vma;
1302 get_anon_vma(vma->anon_vma);
1304 up_read(&mm->mmap_sem);
1309 * try_to_merge_two_pages - take two identical pages and prepare them
1310 * to be merged into one page.
1312 * This function returns the kpage if we successfully merged two identical
1313 * pages into one ksm page, NULL otherwise.
1315 * Note that this function upgrades page to ksm page: if one of the pages
1316 * is already a ksm page, try_to_merge_with_ksm_page should be used.
1318 static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
1320 struct rmap_item *tree_rmap_item,
1321 struct page *tree_page)
1325 err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
1327 err = try_to_merge_with_ksm_page(tree_rmap_item,
1330 * If that fails, we have a ksm page with only one pte
1331 * pointing to it: so break it.
1334 break_cow(rmap_item);
1336 return err ? NULL : page;
1339 static __always_inline
1340 bool __is_page_sharing_candidate(struct stable_node *stable_node, int offset)
1342 VM_BUG_ON(stable_node->rmap_hlist_len < 0);
1344 * Check that at least one mapping still exists, otherwise
1345 * there's no much point to merge and share with this
1346 * stable_node, as the underlying tree_page of the other
1347 * sharer is going to be freed soon.
1349 return stable_node->rmap_hlist_len &&
1350 stable_node->rmap_hlist_len + offset < ksm_max_page_sharing;
1353 static __always_inline
1354 bool is_page_sharing_candidate(struct stable_node *stable_node)
1356 return __is_page_sharing_candidate(stable_node, 0);
1359 static struct page *stable_node_dup(struct stable_node **_stable_node_dup,
1360 struct stable_node **_stable_node,
1361 struct rb_root *root,
1362 bool prune_stale_stable_nodes)
1364 struct stable_node *dup, *found = NULL, *stable_node = *_stable_node;
1365 struct hlist_node *hlist_safe;
1366 struct page *_tree_page, *tree_page = NULL;
1368 int found_rmap_hlist_len;
1370 if (!prune_stale_stable_nodes ||
1371 time_before(jiffies, stable_node->chain_prune_time +
1373 ksm_stable_node_chains_prune_millisecs)))
1374 prune_stale_stable_nodes = false;
1376 stable_node->chain_prune_time = jiffies;
1378 hlist_for_each_entry_safe(dup, hlist_safe,
1379 &stable_node->hlist, hlist_dup) {
1382 * We must walk all stable_node_dup to prune the stale
1383 * stable nodes during lookup.
1385 * get_ksm_page can drop the nodes from the
1386 * stable_node->hlist if they point to freed pages
1387 * (that's why we do a _safe walk). The "dup"
1388 * stable_node parameter itself will be freed from
1389 * under us if it returns NULL.
1391 _tree_page = get_ksm_page(dup, false);
1395 if (is_page_sharing_candidate(dup)) {
1397 dup->rmap_hlist_len > found_rmap_hlist_len) {
1399 put_page(tree_page);
1401 found_rmap_hlist_len = found->rmap_hlist_len;
1402 tree_page = _tree_page;
1404 /* skip put_page for found dup */
1405 if (!prune_stale_stable_nodes)
1410 put_page(_tree_page);
1415 * nr is counting all dups in the chain only if
1416 * prune_stale_stable_nodes is true, otherwise we may
1417 * break the loop at nr == 1 even if there are
1420 if (prune_stale_stable_nodes && nr == 1) {
1422 * If there's not just one entry it would
1423 * corrupt memory, better BUG_ON. In KSM
1424 * context with no lock held it's not even
1427 BUG_ON(stable_node->hlist.first->next);
1430 * There's just one entry and it is below the
1431 * deduplication limit so drop the chain.
1433 rb_replace_node(&stable_node->node, &found->node,
1435 free_stable_node(stable_node);
1436 ksm_stable_node_chains--;
1437 ksm_stable_node_dups--;
1439 * NOTE: the caller depends on the stable_node
1440 * to be equal to stable_node_dup if the chain
1443 *_stable_node = found;
1445 * Just for robustneess as stable_node is
1446 * otherwise left as a stable pointer, the
1447 * compiler shall optimize it away at build
1451 } else if (stable_node->hlist.first != &found->hlist_dup &&
1452 __is_page_sharing_candidate(found, 1)) {
1454 * If the found stable_node dup can accept one
1455 * more future merge (in addition to the one
1456 * that is underway) and is not at the head of
1457 * the chain, put it there so next search will
1458 * be quicker in the !prune_stale_stable_nodes
1461 * NOTE: it would be inaccurate to use nr > 1
1462 * instead of checking the hlist.first pointer
1463 * directly, because in the
1464 * prune_stale_stable_nodes case "nr" isn't
1465 * the position of the found dup in the chain,
1466 * but the total number of dups in the chain.
1468 hlist_del(&found->hlist_dup);
1469 hlist_add_head(&found->hlist_dup,
1470 &stable_node->hlist);
1474 *_stable_node_dup = found;
1478 static struct stable_node *stable_node_dup_any(struct stable_node *stable_node,
1479 struct rb_root *root)
1481 if (!is_stable_node_chain(stable_node))
1483 if (hlist_empty(&stable_node->hlist)) {
1484 free_stable_node_chain(stable_node, root);
1487 return hlist_entry(stable_node->hlist.first,
1488 typeof(*stable_node), hlist_dup);
1492 * Like for get_ksm_page, this function can free the *_stable_node and
1493 * *_stable_node_dup if the returned tree_page is NULL.
1495 * It can also free and overwrite *_stable_node with the found
1496 * stable_node_dup if the chain is collapsed (in which case
1497 * *_stable_node will be equal to *_stable_node_dup like if the chain
1498 * never existed). It's up to the caller to verify tree_page is not
1499 * NULL before dereferencing *_stable_node or *_stable_node_dup.
1501 * *_stable_node_dup is really a second output parameter of this
1502 * function and will be overwritten in all cases, the caller doesn't
1503 * need to initialize it.
1505 static struct page *__stable_node_chain(struct stable_node **_stable_node_dup,
1506 struct stable_node **_stable_node,
1507 struct rb_root *root,
1508 bool prune_stale_stable_nodes)
1510 struct stable_node *stable_node = *_stable_node;
1511 if (!is_stable_node_chain(stable_node)) {
1512 if (is_page_sharing_candidate(stable_node)) {
1513 *_stable_node_dup = stable_node;
1514 return get_ksm_page(stable_node, false);
1517 * _stable_node_dup set to NULL means the stable_node
1518 * reached the ksm_max_page_sharing limit.
1520 *_stable_node_dup = NULL;
1523 return stable_node_dup(_stable_node_dup, _stable_node, root,
1524 prune_stale_stable_nodes);
1527 static __always_inline struct page *chain_prune(struct stable_node **s_n_d,
1528 struct stable_node **s_n,
1529 struct rb_root *root)
1531 return __stable_node_chain(s_n_d, s_n, root, true);
1534 static __always_inline struct page *chain(struct stable_node **s_n_d,
1535 struct stable_node *s_n,
1536 struct rb_root *root)
1538 struct stable_node *old_stable_node = s_n;
1539 struct page *tree_page;
1541 tree_page = __stable_node_chain(s_n_d, &s_n, root, false);
1542 /* not pruning dups so s_n cannot have changed */
1543 VM_BUG_ON(s_n != old_stable_node);
1548 * stable_tree_search - search for page inside the stable tree
1550 * This function checks if there is a page inside the stable tree
1551 * with identical content to the page that we are scanning right now.
1553 * This function returns the stable tree node of identical content if found,
1556 static struct page *stable_tree_search(struct page *page)
1559 struct rb_root *root;
1560 struct rb_node **new;
1561 struct rb_node *parent;
1562 struct stable_node *stable_node, *stable_node_dup, *stable_node_any;
1563 struct stable_node *page_node;
1565 page_node = page_stable_node(page);
1566 if (page_node && page_node->head != &migrate_nodes) {
1567 /* ksm page forked */
1572 nid = get_kpfn_nid(page_to_pfn(page));
1573 root = root_stable_tree + nid;
1575 new = &root->rb_node;
1579 struct page *tree_page;
1583 stable_node = rb_entry(*new, struct stable_node, node);
1584 stable_node_any = NULL;
1585 tree_page = chain_prune(&stable_node_dup, &stable_node, root);
1587 * NOTE: stable_node may have been freed by
1588 * chain_prune() if the returned stable_node_dup is
1589 * not NULL. stable_node_dup may have been inserted in
1590 * the rbtree instead as a regular stable_node (in
1591 * order to collapse the stable_node chain if a single
1592 * stable_node dup was found in it). In such case the
1593 * stable_node is overwritten by the calleee to point
1594 * to the stable_node_dup that was collapsed in the
1595 * stable rbtree and stable_node will be equal to
1596 * stable_node_dup like if the chain never existed.
1598 if (!stable_node_dup) {
1600 * Either all stable_node dups were full in
1601 * this stable_node chain, or this chain was
1602 * empty and should be rb_erased.
1604 stable_node_any = stable_node_dup_any(stable_node,
1606 if (!stable_node_any) {
1607 /* rb_erase just run */
1611 * Take any of the stable_node dups page of
1612 * this stable_node chain to let the tree walk
1613 * continue. All KSM pages belonging to the
1614 * stable_node dups in a stable_node chain
1615 * have the same content and they're
1616 * wrprotected at all times. Any will work
1617 * fine to continue the walk.
1619 tree_page = get_ksm_page(stable_node_any, false);
1621 VM_BUG_ON(!stable_node_dup ^ !!stable_node_any);
1624 * If we walked over a stale stable_node,
1625 * get_ksm_page() will call rb_erase() and it
1626 * may rebalance the tree from under us. So
1627 * restart the search from scratch. Returning
1628 * NULL would be safe too, but we'd generate
1629 * false negative insertions just because some
1630 * stable_node was stale.
1635 ret = memcmp_pages(page, tree_page);
1636 put_page(tree_page);
1640 new = &parent->rb_left;
1642 new = &parent->rb_right;
1645 VM_BUG_ON(page_node->head != &migrate_nodes);
1647 * Test if the migrated page should be merged
1648 * into a stable node dup. If the mapcount is
1649 * 1 we can migrate it with another KSM page
1650 * without adding it to the chain.
1652 if (page_mapcount(page) > 1)
1656 if (!stable_node_dup) {
1658 * If the stable_node is a chain and
1659 * we got a payload match in memcmp
1660 * but we cannot merge the scanned
1661 * page in any of the existing
1662 * stable_node dups because they're
1663 * all full, we need to wait the
1664 * scanned page to find itself a match
1665 * in the unstable tree to create a
1666 * brand new KSM page to add later to
1667 * the dups of this stable_node.
1673 * Lock and unlock the stable_node's page (which
1674 * might already have been migrated) so that page
1675 * migration is sure to notice its raised count.
1676 * It would be more elegant to return stable_node
1677 * than kpage, but that involves more changes.
1679 tree_page = get_ksm_page(stable_node_dup, true);
1680 if (unlikely(!tree_page))
1682 * The tree may have been rebalanced,
1683 * so re-evaluate parent and new.
1686 unlock_page(tree_page);
1688 if (get_kpfn_nid(stable_node_dup->kpfn) !=
1689 NUMA(stable_node_dup->nid)) {
1690 put_page(tree_page);
1700 list_del(&page_node->list);
1701 DO_NUMA(page_node->nid = nid);
1702 rb_link_node(&page_node->node, parent, new);
1703 rb_insert_color(&page_node->node, root);
1705 if (is_page_sharing_candidate(page_node)) {
1713 * If stable_node was a chain and chain_prune collapsed it,
1714 * stable_node has been updated to be the new regular
1715 * stable_node. A collapse of the chain is indistinguishable
1716 * from the case there was no chain in the stable
1717 * rbtree. Otherwise stable_node is the chain and
1718 * stable_node_dup is the dup to replace.
1720 if (stable_node_dup == stable_node) {
1721 VM_BUG_ON(is_stable_node_chain(stable_node_dup));
1722 VM_BUG_ON(is_stable_node_dup(stable_node_dup));
1723 /* there is no chain */
1725 VM_BUG_ON(page_node->head != &migrate_nodes);
1726 list_del(&page_node->list);
1727 DO_NUMA(page_node->nid = nid);
1728 rb_replace_node(&stable_node_dup->node,
1731 if (is_page_sharing_candidate(page_node))
1736 rb_erase(&stable_node_dup->node, root);
1740 VM_BUG_ON(!is_stable_node_chain(stable_node));
1741 __stable_node_dup_del(stable_node_dup);
1743 VM_BUG_ON(page_node->head != &migrate_nodes);
1744 list_del(&page_node->list);
1745 DO_NUMA(page_node->nid = nid);
1746 stable_node_chain_add_dup(page_node, stable_node);
1747 if (is_page_sharing_candidate(page_node))
1755 stable_node_dup->head = &migrate_nodes;
1756 list_add(&stable_node_dup->list, stable_node_dup->head);
1760 /* stable_node_dup could be null if it reached the limit */
1761 if (!stable_node_dup)
1762 stable_node_dup = stable_node_any;
1764 * If stable_node was a chain and chain_prune collapsed it,
1765 * stable_node has been updated to be the new regular
1766 * stable_node. A collapse of the chain is indistinguishable
1767 * from the case there was no chain in the stable
1768 * rbtree. Otherwise stable_node is the chain and
1769 * stable_node_dup is the dup to replace.
1771 if (stable_node_dup == stable_node) {
1772 VM_BUG_ON(is_stable_node_chain(stable_node_dup));
1773 VM_BUG_ON(is_stable_node_dup(stable_node_dup));
1774 /* chain is missing so create it */
1775 stable_node = alloc_stable_node_chain(stable_node_dup,
1781 * Add this stable_node dup that was
1782 * migrated to the stable_node chain
1783 * of the current nid for this page
1786 VM_BUG_ON(!is_stable_node_chain(stable_node));
1787 VM_BUG_ON(!is_stable_node_dup(stable_node_dup));
1788 VM_BUG_ON(page_node->head != &migrate_nodes);
1789 list_del(&page_node->list);
1790 DO_NUMA(page_node->nid = nid);
1791 stable_node_chain_add_dup(page_node, stable_node);
1796 * stable_tree_insert - insert stable tree node pointing to new ksm page
1797 * into the stable tree.
1799 * This function returns the stable tree node just allocated on success,
1802 static struct stable_node *stable_tree_insert(struct page *kpage)
1806 struct rb_root *root;
1807 struct rb_node **new;
1808 struct rb_node *parent;
1809 struct stable_node *stable_node, *stable_node_dup, *stable_node_any;
1810 bool need_chain = false;
1812 kpfn = page_to_pfn(kpage);
1813 nid = get_kpfn_nid(kpfn);
1814 root = root_stable_tree + nid;
1817 new = &root->rb_node;
1820 struct page *tree_page;
1824 stable_node = rb_entry(*new, struct stable_node, node);
1825 stable_node_any = NULL;
1826 tree_page = chain(&stable_node_dup, stable_node, root);
1827 if (!stable_node_dup) {
1829 * Either all stable_node dups were full in
1830 * this stable_node chain, or this chain was
1831 * empty and should be rb_erased.
1833 stable_node_any = stable_node_dup_any(stable_node,
1835 if (!stable_node_any) {
1836 /* rb_erase just run */
1840 * Take any of the stable_node dups page of
1841 * this stable_node chain to let the tree walk
1842 * continue. All KSM pages belonging to the
1843 * stable_node dups in a stable_node chain
1844 * have the same content and they're
1845 * wrprotected at all times. Any will work
1846 * fine to continue the walk.
1848 tree_page = get_ksm_page(stable_node_any, false);
1850 VM_BUG_ON(!stable_node_dup ^ !!stable_node_any);
1853 * If we walked over a stale stable_node,
1854 * get_ksm_page() will call rb_erase() and it
1855 * may rebalance the tree from under us. So
1856 * restart the search from scratch. Returning
1857 * NULL would be safe too, but we'd generate
1858 * false negative insertions just because some
1859 * stable_node was stale.
1864 ret = memcmp_pages(kpage, tree_page);
1865 put_page(tree_page);
1869 new = &parent->rb_left;
1871 new = &parent->rb_right;
1878 stable_node_dup = alloc_stable_node();
1879 if (!stable_node_dup)
1882 INIT_HLIST_HEAD(&stable_node_dup->hlist);
1883 stable_node_dup->kpfn = kpfn;
1884 set_page_stable_node(kpage, stable_node_dup);
1885 stable_node_dup->rmap_hlist_len = 0;
1886 DO_NUMA(stable_node_dup->nid = nid);
1888 rb_link_node(&stable_node_dup->node, parent, new);
1889 rb_insert_color(&stable_node_dup->node, root);
1891 if (!is_stable_node_chain(stable_node)) {
1892 struct stable_node *orig = stable_node;
1893 /* chain is missing so create it */
1894 stable_node = alloc_stable_node_chain(orig, root);
1896 free_stable_node(stable_node_dup);
1900 stable_node_chain_add_dup(stable_node_dup, stable_node);
1903 return stable_node_dup;
1907 * unstable_tree_search_insert - search for identical page,
1908 * else insert rmap_item into the unstable tree.
1910 * This function searches for a page in the unstable tree identical to the
1911 * page currently being scanned; and if no identical page is found in the
1912 * tree, we insert rmap_item as a new object into the unstable tree.
1914 * This function returns pointer to rmap_item found to be identical
1915 * to the currently scanned page, NULL otherwise.
1917 * This function does both searching and inserting, because they share
1918 * the same walking algorithm in an rbtree.
1921 struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1923 struct page **tree_pagep)
1925 struct rb_node **new;
1926 struct rb_root *root;
1927 struct rb_node *parent = NULL;
1930 nid = get_kpfn_nid(page_to_pfn(page));
1931 root = root_unstable_tree + nid;
1932 new = &root->rb_node;
1935 struct rmap_item *tree_rmap_item;
1936 struct page *tree_page;
1940 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
1941 tree_page = get_mergeable_page(tree_rmap_item);
1946 * Don't substitute a ksm page for a forked page.
1948 if (page == tree_page) {
1949 put_page(tree_page);
1953 ret = memcmp_pages(page, tree_page);
1957 put_page(tree_page);
1958 new = &parent->rb_left;
1959 } else if (ret > 0) {
1960 put_page(tree_page);
1961 new = &parent->rb_right;
1962 } else if (!ksm_merge_across_nodes &&
1963 page_to_nid(tree_page) != nid) {
1965 * If tree_page has been migrated to another NUMA node,
1966 * it will be flushed out and put in the right unstable
1967 * tree next time: only merge with it when across_nodes.
1969 put_page(tree_page);
1972 *tree_pagep = tree_page;
1973 return tree_rmap_item;
1977 rmap_item->address |= UNSTABLE_FLAG;
1978 rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1979 DO_NUMA(rmap_item->nid = nid);
1980 rb_link_node(&rmap_item->node, parent, new);
1981 rb_insert_color(&rmap_item->node, root);
1983 ksm_pages_unshared++;
1988 * stable_tree_append - add another rmap_item to the linked list of
1989 * rmap_items hanging off a given node of the stable tree, all sharing
1990 * the same ksm page.
1992 static void stable_tree_append(struct rmap_item *rmap_item,
1993 struct stable_node *stable_node,
1994 bool max_page_sharing_bypass)
1997 * rmap won't find this mapping if we don't insert the
1998 * rmap_item in the right stable_node
1999 * duplicate. page_migration could break later if rmap breaks,
2000 * so we can as well crash here. We really need to check for
2001 * rmap_hlist_len == STABLE_NODE_CHAIN, but we can as well check
2002 * for other negative values as an undeflow if detected here
2003 * for the first time (and not when decreasing rmap_hlist_len)
2004 * would be sign of memory corruption in the stable_node.
2006 BUG_ON(stable_node->rmap_hlist_len < 0);
2008 stable_node->rmap_hlist_len++;
2009 if (!max_page_sharing_bypass)
2010 /* possibly non fatal but unexpected overflow, only warn */
2011 WARN_ON_ONCE(stable_node->rmap_hlist_len >
2012 ksm_max_page_sharing);
2014 rmap_item->head = stable_node;
2015 rmap_item->address |= STABLE_FLAG;
2016 hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
2018 if (rmap_item->hlist.next)
2019 ksm_pages_sharing++;
2025 * cmp_and_merge_page - first see if page can be merged into the stable tree;
2026 * if not, compare checksum to previous and if it's the same, see if page can
2027 * be inserted into the unstable tree, or merged with a page already there and
2028 * both transferred to the stable tree.
2030 * @page: the page that we are searching identical page to.
2031 * @rmap_item: the reverse mapping into the virtual address of this page
2033 static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
2035 struct mm_struct *mm = rmap_item->mm;
2036 struct rmap_item *tree_rmap_item;
2037 struct page *tree_page = NULL;
2038 struct stable_node *stable_node;
2040 unsigned int checksum;
2042 bool max_page_sharing_bypass = false;
2044 stable_node = page_stable_node(page);
2046 if (stable_node->head != &migrate_nodes &&
2047 get_kpfn_nid(READ_ONCE(stable_node->kpfn)) !=
2048 NUMA(stable_node->nid)) {
2049 stable_node_dup_del(stable_node);
2050 stable_node->head = &migrate_nodes;
2051 list_add(&stable_node->list, stable_node->head);
2053 if (stable_node->head != &migrate_nodes &&
2054 rmap_item->head == stable_node)
2057 * If it's a KSM fork, allow it to go over the sharing limit
2060 if (!is_page_sharing_candidate(stable_node))
2061 max_page_sharing_bypass = true;
2064 /* We first start with searching the page inside the stable tree */
2065 kpage = stable_tree_search(page);
2066 if (kpage == page && rmap_item->head == stable_node) {
2071 remove_rmap_item_from_tree(rmap_item);
2074 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
2077 * The page was successfully merged:
2078 * add its rmap_item to the stable tree.
2081 stable_tree_append(rmap_item, page_stable_node(kpage),
2082 max_page_sharing_bypass);
2090 * If the hash value of the page has changed from the last time
2091 * we calculated it, this page is changing frequently: therefore we
2092 * don't want to insert it in the unstable tree, and we don't want
2093 * to waste our time searching for something identical to it there.
2095 checksum = calc_checksum(page);
2096 if (rmap_item->oldchecksum != checksum) {
2097 rmap_item->oldchecksum = checksum;
2102 * Same checksum as an empty page. We attempt to merge it with the
2103 * appropriate zero page if the user enabled this via sysfs.
2105 if (ksm_use_zero_pages && (checksum == zero_checksum)) {
2106 struct vm_area_struct *vma;
2108 down_read(&mm->mmap_sem);
2109 vma = find_mergeable_vma(mm, rmap_item->address);
2111 err = try_to_merge_one_page(vma, page,
2112 ZERO_PAGE(rmap_item->address));
2115 * If the vma is out of date, we do not need to
2120 up_read(&mm->mmap_sem);
2122 * In case of failure, the page was not really empty, so we
2123 * need to continue. Otherwise we're done.
2129 unstable_tree_search_insert(rmap_item, page, &tree_page);
2130 if (tree_rmap_item) {
2133 kpage = try_to_merge_two_pages(rmap_item, page,
2134 tree_rmap_item, tree_page);
2136 * If both pages we tried to merge belong to the same compound
2137 * page, then we actually ended up increasing the reference
2138 * count of the same compound page twice, and split_huge_page
2140 * Here we set a flag if that happened, and we use it later to
2141 * try split_huge_page again. Since we call put_page right
2142 * afterwards, the reference count will be correct and
2143 * split_huge_page should succeed.
2145 split = PageTransCompound(page)
2146 && compound_head(page) == compound_head(tree_page);
2147 put_page(tree_page);
2150 * The pages were successfully merged: insert new
2151 * node in the stable tree and add both rmap_items.
2154 stable_node = stable_tree_insert(kpage);
2156 stable_tree_append(tree_rmap_item, stable_node,
2158 stable_tree_append(rmap_item, stable_node,
2164 * If we fail to insert the page into the stable tree,
2165 * we will have 2 virtual addresses that are pointing
2166 * to a ksm page left outside the stable tree,
2167 * in which case we need to break_cow on both.
2170 break_cow(tree_rmap_item);
2171 break_cow(rmap_item);
2175 * We are here if we tried to merge two pages and
2176 * failed because they both belonged to the same
2177 * compound page. We will split the page now, but no
2178 * merging will take place.
2179 * We do not want to add the cost of a full lock; if
2180 * the page is locked, it is better to skip it and
2181 * perhaps try again later.
2183 if (!trylock_page(page))
2185 split_huge_page(page);
2191 static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
2192 struct rmap_item **rmap_list,
2195 struct rmap_item *rmap_item;
2197 while (*rmap_list) {
2198 rmap_item = *rmap_list;
2199 if ((rmap_item->address & PAGE_MASK) == addr)
2201 if (rmap_item->address > addr)
2203 *rmap_list = rmap_item->rmap_list;
2204 remove_rmap_item_from_tree(rmap_item);
2205 free_rmap_item(rmap_item);
2208 rmap_item = alloc_rmap_item();
2210 /* It has already been zeroed */
2211 rmap_item->mm = mm_slot->mm;
2212 rmap_item->address = addr;
2213 rmap_item->rmap_list = *rmap_list;
2214 *rmap_list = rmap_item;
2219 static struct rmap_item *scan_get_next_rmap_item(struct page **page)
2221 struct mm_struct *mm;
2222 struct mm_slot *slot;
2223 struct vm_area_struct *vma;
2224 struct rmap_item *rmap_item;
2227 if (list_empty(&ksm_mm_head.mm_list))
2230 slot = ksm_scan.mm_slot;
2231 if (slot == &ksm_mm_head) {
2233 * A number of pages can hang around indefinitely on per-cpu
2234 * pagevecs, raised page count preventing write_protect_page
2235 * from merging them. Though it doesn't really matter much,
2236 * it is puzzling to see some stuck in pages_volatile until
2237 * other activity jostles them out, and they also prevented
2238 * LTP's KSM test from succeeding deterministically; so drain
2239 * them here (here rather than on entry to ksm_do_scan(),
2240 * so we don't IPI too often when pages_to_scan is set low).
2242 lru_add_drain_all();
2245 * Whereas stale stable_nodes on the stable_tree itself
2246 * get pruned in the regular course of stable_tree_search(),
2247 * those moved out to the migrate_nodes list can accumulate:
2248 * so prune them once before each full scan.
2250 if (!ksm_merge_across_nodes) {
2251 struct stable_node *stable_node, *next;
2254 list_for_each_entry_safe(stable_node, next,
2255 &migrate_nodes, list) {
2256 page = get_ksm_page(stable_node, false);
2263 for (nid = 0; nid < ksm_nr_node_ids; nid++)
2264 root_unstable_tree[nid] = RB_ROOT;
2266 spin_lock(&ksm_mmlist_lock);
2267 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
2268 ksm_scan.mm_slot = slot;
2269 spin_unlock(&ksm_mmlist_lock);
2271 * Although we tested list_empty() above, a racing __ksm_exit
2272 * of the last mm on the list may have removed it since then.
2274 if (slot == &ksm_mm_head)
2277 ksm_scan.address = 0;
2278 ksm_scan.rmap_list = &slot->rmap_list;
2282 down_read(&mm->mmap_sem);
2283 if (ksm_test_exit(mm))
2286 vma = find_vma(mm, ksm_scan.address);
2288 for (; vma; vma = vma->vm_next) {
2289 if (!(vma->vm_flags & VM_MERGEABLE))
2291 if (ksm_scan.address < vma->vm_start)
2292 ksm_scan.address = vma->vm_start;
2294 ksm_scan.address = vma->vm_end;
2296 while (ksm_scan.address < vma->vm_end) {
2297 if (ksm_test_exit(mm))
2299 *page = follow_page(vma, ksm_scan.address, FOLL_GET);
2300 if (IS_ERR_OR_NULL(*page)) {
2301 ksm_scan.address += PAGE_SIZE;
2305 if (PageAnon(*page)) {
2306 flush_anon_page(vma, *page, ksm_scan.address);
2307 flush_dcache_page(*page);
2308 rmap_item = get_next_rmap_item(slot,
2309 ksm_scan.rmap_list, ksm_scan.address);
2311 ksm_scan.rmap_list =
2312 &rmap_item->rmap_list;
2313 ksm_scan.address += PAGE_SIZE;
2316 up_read(&mm->mmap_sem);
2320 ksm_scan.address += PAGE_SIZE;
2325 if (ksm_test_exit(mm)) {
2326 ksm_scan.address = 0;
2327 ksm_scan.rmap_list = &slot->rmap_list;
2330 * Nuke all the rmap_items that are above this current rmap:
2331 * because there were no VM_MERGEABLE vmas with such addresses.
2333 remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
2335 spin_lock(&ksm_mmlist_lock);
2336 ksm_scan.mm_slot = list_entry(slot->mm_list.next,
2337 struct mm_slot, mm_list);
2338 if (ksm_scan.address == 0) {
2340 * We've completed a full scan of all vmas, holding mmap_sem
2341 * throughout, and found no VM_MERGEABLE: so do the same as
2342 * __ksm_exit does to remove this mm from all our lists now.
2343 * This applies either when cleaning up after __ksm_exit
2344 * (but beware: we can reach here even before __ksm_exit),
2345 * or when all VM_MERGEABLE areas have been unmapped (and
2346 * mmap_sem then protects against race with MADV_MERGEABLE).
2348 hash_del(&slot->link);
2349 list_del(&slot->mm_list);
2350 spin_unlock(&ksm_mmlist_lock);
2353 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
2354 up_read(&mm->mmap_sem);
2357 up_read(&mm->mmap_sem);
2359 * up_read(&mm->mmap_sem) first because after
2360 * spin_unlock(&ksm_mmlist_lock) run, the "mm" may
2361 * already have been freed under us by __ksm_exit()
2362 * because the "mm_slot" is still hashed and
2363 * ksm_scan.mm_slot doesn't point to it anymore.
2365 spin_unlock(&ksm_mmlist_lock);
2368 /* Repeat until we've completed scanning the whole list */
2369 slot = ksm_scan.mm_slot;
2370 if (slot != &ksm_mm_head)
2378 * ksm_do_scan - the ksm scanner main worker function.
2379 * @scan_npages: number of pages we want to scan before we return.
2381 static void ksm_do_scan(unsigned int scan_npages)
2383 struct rmap_item *rmap_item;
2384 struct page *uninitialized_var(page);
2386 while (scan_npages-- && likely(!freezing(current))) {
2388 rmap_item = scan_get_next_rmap_item(&page);
2391 cmp_and_merge_page(page, rmap_item);
2396 static int ksmd_should_run(void)
2398 return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
2401 static int ksm_scan_thread(void *nothing)
2404 set_user_nice(current, 5);
2406 while (!kthread_should_stop()) {
2407 mutex_lock(&ksm_thread_mutex);
2408 wait_while_offlining();
2409 if (ksmd_should_run())
2410 ksm_do_scan(ksm_thread_pages_to_scan);
2411 mutex_unlock(&ksm_thread_mutex);
2415 if (ksmd_should_run()) {
2416 schedule_timeout_interruptible(
2417 msecs_to_jiffies(ksm_thread_sleep_millisecs));
2419 wait_event_freezable(ksm_thread_wait,
2420 ksmd_should_run() || kthread_should_stop());
2426 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
2427 unsigned long end, int advice, unsigned long *vm_flags)
2429 struct mm_struct *mm = vma->vm_mm;
2433 case MADV_MERGEABLE:
2435 * Be somewhat over-protective for now!
2437 if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE |
2438 VM_PFNMAP | VM_IO | VM_DONTEXPAND |
2439 VM_HUGETLB | VM_MIXEDMAP))
2440 return 0; /* just ignore the advice */
2442 if (vma_is_dax(vma))
2446 if (*vm_flags & VM_SAO)
2450 if (*vm_flags & VM_SPARC_ADI)
2454 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
2455 err = __ksm_enter(mm);
2460 *vm_flags |= VM_MERGEABLE;
2463 case MADV_UNMERGEABLE:
2464 if (!(*vm_flags & VM_MERGEABLE))
2465 return 0; /* just ignore the advice */
2467 if (vma->anon_vma) {
2468 err = unmerge_ksm_pages(vma, start, end);
2473 *vm_flags &= ~VM_MERGEABLE;
2480 int __ksm_enter(struct mm_struct *mm)
2482 struct mm_slot *mm_slot;
2485 mm_slot = alloc_mm_slot();
2489 /* Check ksm_run too? Would need tighter locking */
2490 needs_wakeup = list_empty(&ksm_mm_head.mm_list);
2492 spin_lock(&ksm_mmlist_lock);
2493 insert_to_mm_slots_hash(mm, mm_slot);
2495 * When KSM_RUN_MERGE (or KSM_RUN_STOP),
2496 * insert just behind the scanning cursor, to let the area settle
2497 * down a little; when fork is followed by immediate exec, we don't
2498 * want ksmd to waste time setting up and tearing down an rmap_list.
2500 * But when KSM_RUN_UNMERGE, it's important to insert ahead of its
2501 * scanning cursor, otherwise KSM pages in newly forked mms will be
2502 * missed: then we might as well insert at the end of the list.
2504 if (ksm_run & KSM_RUN_UNMERGE)
2505 list_add_tail(&mm_slot->mm_list, &ksm_mm_head.mm_list);
2507 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
2508 spin_unlock(&ksm_mmlist_lock);
2510 set_bit(MMF_VM_MERGEABLE, &mm->flags);
2514 wake_up_interruptible(&ksm_thread_wait);
2519 void __ksm_exit(struct mm_struct *mm)
2521 struct mm_slot *mm_slot;
2522 int easy_to_free = 0;
2525 * This process is exiting: if it's straightforward (as is the
2526 * case when ksmd was never running), free mm_slot immediately.
2527 * But if it's at the cursor or has rmap_items linked to it, use
2528 * mmap_sem to synchronize with any break_cows before pagetables
2529 * are freed, and leave the mm_slot on the list for ksmd to free.
2530 * Beware: ksm may already have noticed it exiting and freed the slot.
2533 spin_lock(&ksm_mmlist_lock);
2534 mm_slot = get_mm_slot(mm);
2535 if (mm_slot && ksm_scan.mm_slot != mm_slot) {
2536 if (!mm_slot->rmap_list) {
2537 hash_del(&mm_slot->link);
2538 list_del(&mm_slot->mm_list);
2541 list_move(&mm_slot->mm_list,
2542 &ksm_scan.mm_slot->mm_list);
2545 spin_unlock(&ksm_mmlist_lock);
2548 free_mm_slot(mm_slot);
2549 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
2551 } else if (mm_slot) {
2552 down_write(&mm->mmap_sem);
2553 up_write(&mm->mmap_sem);
2557 struct page *ksm_might_need_to_copy(struct page *page,
2558 struct vm_area_struct *vma, unsigned long address)
2560 struct anon_vma *anon_vma = page_anon_vma(page);
2561 struct page *new_page;
2563 if (PageKsm(page)) {
2564 if (page_stable_node(page) &&
2565 !(ksm_run & KSM_RUN_UNMERGE))
2566 return page; /* no need to copy it */
2567 } else if (!anon_vma) {
2568 return page; /* no need to copy it */
2569 } else if (anon_vma->root == vma->anon_vma->root &&
2570 page->index == linear_page_index(vma, address)) {
2571 return page; /* still no need to copy it */
2573 if (!PageUptodate(page))
2574 return page; /* let do_swap_page report the error */
2576 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2578 copy_user_highpage(new_page, page, address, vma);
2580 SetPageDirty(new_page);
2581 __SetPageUptodate(new_page);
2582 __SetPageLocked(new_page);
2588 void rmap_walk_ksm(struct page *page, struct rmap_walk_control *rwc)
2590 struct stable_node *stable_node;
2591 struct rmap_item *rmap_item;
2592 int search_new_forks = 0;
2594 VM_BUG_ON_PAGE(!PageKsm(page), page);
2597 * Rely on the page lock to protect against concurrent modifications
2598 * to that page's node of the stable tree.
2600 VM_BUG_ON_PAGE(!PageLocked(page), page);
2602 stable_node = page_stable_node(page);
2606 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
2607 struct anon_vma *anon_vma = rmap_item->anon_vma;
2608 struct anon_vma_chain *vmac;
2609 struct vm_area_struct *vma;
2612 anon_vma_lock_read(anon_vma);
2613 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
2620 /* Ignore the stable/unstable/sqnr flags */
2621 addr = rmap_item->address & ~KSM_FLAG_MASK;
2623 if (addr < vma->vm_start || addr >= vma->vm_end)
2626 * Initially we examine only the vma which covers this
2627 * rmap_item; but later, if there is still work to do,
2628 * we examine covering vmas in other mms: in case they
2629 * were forked from the original since ksmd passed.
2631 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
2634 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2637 if (!rwc->rmap_one(page, vma, addr, rwc->arg)) {
2638 anon_vma_unlock_read(anon_vma);
2641 if (rwc->done && rwc->done(page)) {
2642 anon_vma_unlock_read(anon_vma);
2646 anon_vma_unlock_read(anon_vma);
2648 if (!search_new_forks++)
2652 #ifdef CONFIG_MIGRATION
2653 void ksm_migrate_page(struct page *newpage, struct page *oldpage)
2655 struct stable_node *stable_node;
2657 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
2658 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
2659 VM_BUG_ON_PAGE(newpage->mapping != oldpage->mapping, newpage);
2661 stable_node = page_stable_node(newpage);
2663 VM_BUG_ON_PAGE(stable_node->kpfn != page_to_pfn(oldpage), oldpage);
2664 stable_node->kpfn = page_to_pfn(newpage);
2666 * newpage->mapping was set in advance; now we need smp_wmb()
2667 * to make sure that the new stable_node->kpfn is visible
2668 * to get_ksm_page() before it can see that oldpage->mapping
2669 * has gone stale (or that PageSwapCache has been cleared).
2672 set_page_stable_node(oldpage, NULL);
2675 #endif /* CONFIG_MIGRATION */
2677 #ifdef CONFIG_MEMORY_HOTREMOVE
2678 static void wait_while_offlining(void)
2680 while (ksm_run & KSM_RUN_OFFLINE) {
2681 mutex_unlock(&ksm_thread_mutex);
2682 wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE),
2683 TASK_UNINTERRUPTIBLE);
2684 mutex_lock(&ksm_thread_mutex);
2688 static bool stable_node_dup_remove_range(struct stable_node *stable_node,
2689 unsigned long start_pfn,
2690 unsigned long end_pfn)
2692 if (stable_node->kpfn >= start_pfn &&
2693 stable_node->kpfn < end_pfn) {
2695 * Don't get_ksm_page, page has already gone:
2696 * which is why we keep kpfn instead of page*
2698 remove_node_from_stable_tree(stable_node);
2704 static bool stable_node_chain_remove_range(struct stable_node *stable_node,
2705 unsigned long start_pfn,
2706 unsigned long end_pfn,
2707 struct rb_root *root)
2709 struct stable_node *dup;
2710 struct hlist_node *hlist_safe;
2712 if (!is_stable_node_chain(stable_node)) {
2713 VM_BUG_ON(is_stable_node_dup(stable_node));
2714 return stable_node_dup_remove_range(stable_node, start_pfn,
2718 hlist_for_each_entry_safe(dup, hlist_safe,
2719 &stable_node->hlist, hlist_dup) {
2720 VM_BUG_ON(!is_stable_node_dup(dup));
2721 stable_node_dup_remove_range(dup, start_pfn, end_pfn);
2723 if (hlist_empty(&stable_node->hlist)) {
2724 free_stable_node_chain(stable_node, root);
2725 return true; /* notify caller that tree was rebalanced */
2730 static void ksm_check_stable_tree(unsigned long start_pfn,
2731 unsigned long end_pfn)
2733 struct stable_node *stable_node, *next;
2734 struct rb_node *node;
2737 for (nid = 0; nid < ksm_nr_node_ids; nid++) {
2738 node = rb_first(root_stable_tree + nid);
2740 stable_node = rb_entry(node, struct stable_node, node);
2741 if (stable_node_chain_remove_range(stable_node,
2745 node = rb_first(root_stable_tree + nid);
2747 node = rb_next(node);
2751 list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
2752 if (stable_node->kpfn >= start_pfn &&
2753 stable_node->kpfn < end_pfn)
2754 remove_node_from_stable_tree(stable_node);
2759 static int ksm_memory_callback(struct notifier_block *self,
2760 unsigned long action, void *arg)
2762 struct memory_notify *mn = arg;
2765 case MEM_GOING_OFFLINE:
2767 * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items()
2768 * and remove_all_stable_nodes() while memory is going offline:
2769 * it is unsafe for them to touch the stable tree at this time.
2770 * But unmerge_ksm_pages(), rmap lookups and other entry points
2771 * which do not need the ksm_thread_mutex are all safe.
2773 mutex_lock(&ksm_thread_mutex);
2774 ksm_run |= KSM_RUN_OFFLINE;
2775 mutex_unlock(&ksm_thread_mutex);
2780 * Most of the work is done by page migration; but there might
2781 * be a few stable_nodes left over, still pointing to struct
2782 * pages which have been offlined: prune those from the tree,
2783 * otherwise get_ksm_page() might later try to access a
2784 * non-existent struct page.
2786 ksm_check_stable_tree(mn->start_pfn,
2787 mn->start_pfn + mn->nr_pages);
2790 case MEM_CANCEL_OFFLINE:
2791 mutex_lock(&ksm_thread_mutex);
2792 ksm_run &= ~KSM_RUN_OFFLINE;
2793 mutex_unlock(&ksm_thread_mutex);
2795 smp_mb(); /* wake_up_bit advises this */
2796 wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE));
2802 static void wait_while_offlining(void)
2805 #endif /* CONFIG_MEMORY_HOTREMOVE */
2809 * This all compiles without CONFIG_SYSFS, but is a waste of space.
2812 #define KSM_ATTR_RO(_name) \
2813 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2814 #define KSM_ATTR(_name) \
2815 static struct kobj_attribute _name##_attr = \
2816 __ATTR(_name, 0644, _name##_show, _name##_store)
2818 static ssize_t sleep_millisecs_show(struct kobject *kobj,
2819 struct kobj_attribute *attr, char *buf)
2821 return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
2824 static ssize_t sleep_millisecs_store(struct kobject *kobj,
2825 struct kobj_attribute *attr,
2826 const char *buf, size_t count)
2828 unsigned long msecs;
2831 err = kstrtoul(buf, 10, &msecs);
2832 if (err || msecs > UINT_MAX)
2835 ksm_thread_sleep_millisecs = msecs;
2839 KSM_ATTR(sleep_millisecs);
2841 static ssize_t pages_to_scan_show(struct kobject *kobj,
2842 struct kobj_attribute *attr, char *buf)
2844 return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
2847 static ssize_t pages_to_scan_store(struct kobject *kobj,
2848 struct kobj_attribute *attr,
2849 const char *buf, size_t count)
2852 unsigned long nr_pages;
2854 err = kstrtoul(buf, 10, &nr_pages);
2855 if (err || nr_pages > UINT_MAX)
2858 ksm_thread_pages_to_scan = nr_pages;
2862 KSM_ATTR(pages_to_scan);
2864 static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
2867 return sprintf(buf, "%lu\n", ksm_run);
2870 static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
2871 const char *buf, size_t count)
2874 unsigned long flags;
2876 err = kstrtoul(buf, 10, &flags);
2877 if (err || flags > UINT_MAX)
2879 if (flags > KSM_RUN_UNMERGE)
2883 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
2884 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
2885 * breaking COW to free the pages_shared (but leaves mm_slots
2886 * on the list for when ksmd may be set running again).
2889 mutex_lock(&ksm_thread_mutex);
2890 wait_while_offlining();
2891 if (ksm_run != flags) {
2893 if (flags & KSM_RUN_UNMERGE) {
2894 set_current_oom_origin();
2895 err = unmerge_and_remove_all_rmap_items();
2896 clear_current_oom_origin();
2898 ksm_run = KSM_RUN_STOP;
2903 mutex_unlock(&ksm_thread_mutex);
2905 if (flags & KSM_RUN_MERGE)
2906 wake_up_interruptible(&ksm_thread_wait);
2913 static ssize_t merge_across_nodes_show(struct kobject *kobj,
2914 struct kobj_attribute *attr, char *buf)
2916 return sprintf(buf, "%u\n", ksm_merge_across_nodes);
2919 static ssize_t merge_across_nodes_store(struct kobject *kobj,
2920 struct kobj_attribute *attr,
2921 const char *buf, size_t count)
2926 err = kstrtoul(buf, 10, &knob);
2932 mutex_lock(&ksm_thread_mutex);
2933 wait_while_offlining();
2934 if (ksm_merge_across_nodes != knob) {
2935 if (ksm_pages_shared || remove_all_stable_nodes())
2937 else if (root_stable_tree == one_stable_tree) {
2938 struct rb_root *buf;
2940 * This is the first time that we switch away from the
2941 * default of merging across nodes: must now allocate
2942 * a buffer to hold as many roots as may be needed.
2943 * Allocate stable and unstable together:
2944 * MAXSMP NODES_SHIFT 10 will use 16kB.
2946 buf = kcalloc(nr_node_ids + nr_node_ids, sizeof(*buf),
2948 /* Let us assume that RB_ROOT is NULL is zero */
2952 root_stable_tree = buf;
2953 root_unstable_tree = buf + nr_node_ids;
2954 /* Stable tree is empty but not the unstable */
2955 root_unstable_tree[0] = one_unstable_tree[0];
2959 ksm_merge_across_nodes = knob;
2960 ksm_nr_node_ids = knob ? 1 : nr_node_ids;
2963 mutex_unlock(&ksm_thread_mutex);
2965 return err ? err : count;
2967 KSM_ATTR(merge_across_nodes);
2970 static ssize_t use_zero_pages_show(struct kobject *kobj,
2971 struct kobj_attribute *attr, char *buf)
2973 return sprintf(buf, "%u\n", ksm_use_zero_pages);
2975 static ssize_t use_zero_pages_store(struct kobject *kobj,
2976 struct kobj_attribute *attr,
2977 const char *buf, size_t count)
2982 err = kstrtobool(buf, &value);
2986 ksm_use_zero_pages = value;
2990 KSM_ATTR(use_zero_pages);
2992 static ssize_t max_page_sharing_show(struct kobject *kobj,
2993 struct kobj_attribute *attr, char *buf)
2995 return sprintf(buf, "%u\n", ksm_max_page_sharing);
2998 static ssize_t max_page_sharing_store(struct kobject *kobj,
2999 struct kobj_attribute *attr,
3000 const char *buf, size_t count)
3005 err = kstrtoint(buf, 10, &knob);
3009 * When a KSM page is created it is shared by 2 mappings. This
3010 * being a signed comparison, it implicitly verifies it's not
3016 if (READ_ONCE(ksm_max_page_sharing) == knob)
3019 mutex_lock(&ksm_thread_mutex);
3020 wait_while_offlining();
3021 if (ksm_max_page_sharing != knob) {
3022 if (ksm_pages_shared || remove_all_stable_nodes())
3025 ksm_max_page_sharing = knob;
3027 mutex_unlock(&ksm_thread_mutex);
3029 return err ? err : count;
3031 KSM_ATTR(max_page_sharing);
3033 static ssize_t pages_shared_show(struct kobject *kobj,
3034 struct kobj_attribute *attr, char *buf)
3036 return sprintf(buf, "%lu\n", ksm_pages_shared);
3038 KSM_ATTR_RO(pages_shared);
3040 static ssize_t pages_sharing_show(struct kobject *kobj,
3041 struct kobj_attribute *attr, char *buf)
3043 return sprintf(buf, "%lu\n", ksm_pages_sharing);
3045 KSM_ATTR_RO(pages_sharing);
3047 static ssize_t pages_unshared_show(struct kobject *kobj,
3048 struct kobj_attribute *attr, char *buf)
3050 return sprintf(buf, "%lu\n", ksm_pages_unshared);
3052 KSM_ATTR_RO(pages_unshared);
3054 static ssize_t pages_volatile_show(struct kobject *kobj,
3055 struct kobj_attribute *attr, char *buf)
3057 long ksm_pages_volatile;
3059 ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
3060 - ksm_pages_sharing - ksm_pages_unshared;
3062 * It was not worth any locking to calculate that statistic,
3063 * but it might therefore sometimes be negative: conceal that.
3065 if (ksm_pages_volatile < 0)
3066 ksm_pages_volatile = 0;
3067 return sprintf(buf, "%ld\n", ksm_pages_volatile);
3069 KSM_ATTR_RO(pages_volatile);
3071 static ssize_t stable_node_dups_show(struct kobject *kobj,
3072 struct kobj_attribute *attr, char *buf)
3074 return sprintf(buf, "%lu\n", ksm_stable_node_dups);
3076 KSM_ATTR_RO(stable_node_dups);
3078 static ssize_t stable_node_chains_show(struct kobject *kobj,
3079 struct kobj_attribute *attr, char *buf)
3081 return sprintf(buf, "%lu\n", ksm_stable_node_chains);
3083 KSM_ATTR_RO(stable_node_chains);
3086 stable_node_chains_prune_millisecs_show(struct kobject *kobj,
3087 struct kobj_attribute *attr,
3090 return sprintf(buf, "%u\n", ksm_stable_node_chains_prune_millisecs);
3094 stable_node_chains_prune_millisecs_store(struct kobject *kobj,
3095 struct kobj_attribute *attr,
3096 const char *buf, size_t count)
3098 unsigned long msecs;
3101 err = kstrtoul(buf, 10, &msecs);
3102 if (err || msecs > UINT_MAX)
3105 ksm_stable_node_chains_prune_millisecs = msecs;
3109 KSM_ATTR(stable_node_chains_prune_millisecs);
3111 static ssize_t full_scans_show(struct kobject *kobj,
3112 struct kobj_attribute *attr, char *buf)
3114 return sprintf(buf, "%lu\n", ksm_scan.seqnr);
3116 KSM_ATTR_RO(full_scans);
3118 static struct attribute *ksm_attrs[] = {
3119 &sleep_millisecs_attr.attr,
3120 &pages_to_scan_attr.attr,
3122 &pages_shared_attr.attr,
3123 &pages_sharing_attr.attr,
3124 &pages_unshared_attr.attr,
3125 &pages_volatile_attr.attr,
3126 &full_scans_attr.attr,
3128 &merge_across_nodes_attr.attr,
3130 &max_page_sharing_attr.attr,
3131 &stable_node_chains_attr.attr,
3132 &stable_node_dups_attr.attr,
3133 &stable_node_chains_prune_millisecs_attr.attr,
3134 &use_zero_pages_attr.attr,
3138 static const struct attribute_group ksm_attr_group = {
3142 #endif /* CONFIG_SYSFS */
3144 static int __init ksm_init(void)
3146 struct task_struct *ksm_thread;
3149 /* The correct value depends on page size and endianness */
3150 zero_checksum = calc_checksum(ZERO_PAGE(0));
3151 /* Default to false for backwards compatibility */
3152 ksm_use_zero_pages = false;
3154 err = ksm_slab_init();
3158 ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
3159 if (IS_ERR(ksm_thread)) {
3160 pr_err("ksm: creating kthread failed\n");
3161 err = PTR_ERR(ksm_thread);
3166 err = sysfs_create_group(mm_kobj, &ksm_attr_group);
3168 pr_err("ksm: register sysfs failed\n");
3169 kthread_stop(ksm_thread);
3173 ksm_run = KSM_RUN_MERGE; /* no way for user to start it */
3175 #endif /* CONFIG_SYSFS */
3177 #ifdef CONFIG_MEMORY_HOTREMOVE
3178 /* There is no significance to this priority 100 */
3179 hotplug_memory_notifier(ksm_memory_callback, 100);
3188 subsys_initcall(ksm_init);