1 // SPDX-License-Identifier: GPL-2.0-only
3 * Memory merging support.
5 * This code enables dynamic sharing of identical pages found in different
6 * memory areas, even if they are not shared by fork()
8 * Copyright (C) 2008-2009 Red Hat, Inc.
16 #include <linux/errno.h>
19 #include <linux/mman.h>
20 #include <linux/sched.h>
21 #include <linux/sched/mm.h>
22 #include <linux/sched/coredump.h>
23 #include <linux/rwsem.h>
24 #include <linux/pagemap.h>
25 #include <linux/rmap.h>
26 #include <linux/spinlock.h>
27 #include <linux/xxhash.h>
28 #include <linux/delay.h>
29 #include <linux/kthread.h>
30 #include <linux/wait.h>
31 #include <linux/slab.h>
32 #include <linux/rbtree.h>
33 #include <linux/memory.h>
34 #include <linux/mmu_notifier.h>
35 #include <linux/swap.h>
36 #include <linux/ksm.h>
37 #include <linux/hashtable.h>
38 #include <linux/freezer.h>
39 #include <linux/oom.h>
40 #include <linux/numa.h>
42 #include <asm/tlbflush.h>
47 #define DO_NUMA(x) do { (x); } while (0)
50 #define DO_NUMA(x) do { } while (0)
56 * A few notes about the KSM scanning process,
57 * to make it easier to understand the data structures below:
59 * In order to reduce excessive scanning, KSM sorts the memory pages by their
60 * contents into a data structure that holds pointers to the pages' locations.
62 * Since the contents of the pages may change at any moment, KSM cannot just
63 * insert the pages into a normal sorted tree and expect it to find anything.
64 * Therefore KSM uses two data structures - the stable and the unstable tree.
66 * The stable tree holds pointers to all the merged pages (ksm pages), sorted
67 * by their contents. Because each such page is write-protected, searching on
68 * this tree is fully assured to be working (except when pages are unmapped),
69 * and therefore this tree is called the stable tree.
71 * The stable tree node includes information required for reverse
72 * mapping from a KSM page to virtual addresses that map this page.
74 * In order to avoid large latencies of the rmap walks on KSM pages,
75 * KSM maintains two types of nodes in the stable tree:
77 * * the regular nodes that keep the reverse mapping structures in a
79 * * the "chains" that link nodes ("dups") that represent the same
80 * write protected memory content, but each "dup" corresponds to a
81 * different KSM page copy of that content
83 * Internally, the regular nodes, "dups" and "chains" are represented
84 * using the same struct stable_node structure.
86 * In addition to the stable tree, KSM uses a second data structure called the
87 * unstable tree: this tree holds pointers to pages which have been found to
88 * be "unchanged for a period of time". The unstable tree sorts these pages
89 * by their contents, but since they are not write-protected, KSM cannot rely
90 * upon the unstable tree to work correctly - the unstable tree is liable to
91 * be corrupted as its contents are modified, and so it is called unstable.
93 * KSM solves this problem by several techniques:
95 * 1) The unstable tree is flushed every time KSM completes scanning all
96 * memory areas, and then the tree is rebuilt again from the beginning.
97 * 2) KSM will only insert into the unstable tree, pages whose hash value
98 * has not changed since the previous scan of all memory areas.
99 * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
100 * colors of the nodes and not on their contents, assuring that even when
101 * the tree gets "corrupted" it won't get out of balance, so scanning time
102 * remains the same (also, searching and inserting nodes in an rbtree uses
103 * the same algorithm, so we have no overhead when we flush and rebuild).
104 * 4) KSM never flushes the stable tree, which means that even if it were to
105 * take 10 attempts to find a page in the unstable tree, once it is found,
106 * it is secured in the stable tree. (When we scan a new page, we first
107 * compare it against the stable tree, and then against the unstable tree.)
109 * If the merge_across_nodes tunable is unset, then KSM maintains multiple
110 * stable trees and multiple unstable trees: one of each for each NUMA node.
114 * struct mm_slot - ksm information per mm that is being scanned
115 * @link: link to the mm_slots hash list
116 * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
117 * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
118 * @mm: the mm that this information is valid for
121 struct hlist_node link;
122 struct list_head mm_list;
123 struct rmap_item *rmap_list;
124 struct mm_struct *mm;
128 * struct ksm_scan - cursor for scanning
129 * @mm_slot: the current mm_slot we are scanning
130 * @address: the next address inside that to be scanned
131 * @rmap_list: link to the next rmap to be scanned in the rmap_list
132 * @seqnr: count of completed full scans (needed when removing unstable node)
134 * There is only the one ksm_scan instance of this cursor structure.
137 struct mm_slot *mm_slot;
138 unsigned long address;
139 struct rmap_item **rmap_list;
144 * struct stable_node - node of the stable rbtree
145 * @node: rb node of this ksm page in the stable tree
146 * @head: (overlaying parent) &migrate_nodes indicates temporarily on that list
147 * @hlist_dup: linked into the stable_node->hlist with a stable_node chain
148 * @list: linked into migrate_nodes, pending placement in the proper node tree
149 * @hlist: hlist head of rmap_items using this ksm page
150 * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid)
151 * @chain_prune_time: time of the last full garbage collection
152 * @rmap_hlist_len: number of rmap_item entries in hlist or STABLE_NODE_CHAIN
153 * @nid: NUMA node id of stable tree in which linked (may not match kpfn)
157 struct rb_node node; /* when node of stable tree */
158 struct { /* when listed for migration */
159 struct list_head *head;
161 struct hlist_node hlist_dup;
162 struct list_head list;
166 struct hlist_head hlist;
169 unsigned long chain_prune_time;
172 * STABLE_NODE_CHAIN can be any negative number in
173 * rmap_hlist_len negative range, but better not -1 to be able
174 * to reliably detect underflows.
176 #define STABLE_NODE_CHAIN -1024
184 * struct rmap_item - reverse mapping item for virtual addresses
185 * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
186 * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
187 * @nid: NUMA node id of unstable tree in which linked (may not match page)
188 * @mm: the memory structure this rmap_item is pointing into
189 * @address: the virtual address this rmap_item tracks (+ flags in low bits)
190 * @oldchecksum: previous checksum of the page at that virtual address
191 * @node: rb node of this rmap_item in the unstable tree
192 * @head: pointer to stable_node heading this list in the stable tree
193 * @hlist: link into hlist of rmap_items hanging off that stable_node
196 struct rmap_item *rmap_list;
198 struct anon_vma *anon_vma; /* when stable */
200 int nid; /* when node of unstable tree */
203 struct mm_struct *mm;
204 unsigned long address; /* + low bits used for flags below */
205 unsigned int oldchecksum; /* when unstable */
207 struct rb_node node; /* when node of unstable tree */
208 struct { /* when listed from stable tree */
209 struct stable_node *head;
210 struct hlist_node hlist;
215 #define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */
216 #define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */
217 #define STABLE_FLAG 0x200 /* is listed from the stable tree */
218 #define KSM_FLAG_MASK (SEQNR_MASK|UNSTABLE_FLAG|STABLE_FLAG)
219 /* to mask all the flags */
221 /* The stable and unstable tree heads */
222 static struct rb_root one_stable_tree[1] = { RB_ROOT };
223 static struct rb_root one_unstable_tree[1] = { RB_ROOT };
224 static struct rb_root *root_stable_tree = one_stable_tree;
225 static struct rb_root *root_unstable_tree = one_unstable_tree;
227 /* Recently migrated nodes of stable tree, pending proper placement */
228 static LIST_HEAD(migrate_nodes);
229 #define STABLE_NODE_DUP_HEAD ((struct list_head *)&migrate_nodes.prev)
231 #define MM_SLOTS_HASH_BITS 10
232 static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
234 static struct mm_slot ksm_mm_head = {
235 .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
237 static struct ksm_scan ksm_scan = {
238 .mm_slot = &ksm_mm_head,
241 static struct kmem_cache *rmap_item_cache;
242 static struct kmem_cache *stable_node_cache;
243 static struct kmem_cache *mm_slot_cache;
245 /* The number of nodes in the stable tree */
246 static unsigned long ksm_pages_shared;
248 /* The number of page slots additionally sharing those nodes */
249 static unsigned long ksm_pages_sharing;
251 /* The number of nodes in the unstable tree */
252 static unsigned long ksm_pages_unshared;
254 /* The number of rmap_items in use: to calculate pages_volatile */
255 static unsigned long ksm_rmap_items;
257 /* The number of stable_node chains */
258 static unsigned long ksm_stable_node_chains;
260 /* The number of stable_node dups linked to the stable_node chains */
261 static unsigned long ksm_stable_node_dups;
263 /* Delay in pruning stale stable_node_dups in the stable_node_chains */
264 static int ksm_stable_node_chains_prune_millisecs = 2000;
266 /* Maximum number of page slots sharing a stable node */
267 static int ksm_max_page_sharing = 256;
269 /* Number of pages ksmd should scan in one batch */
270 static unsigned int ksm_thread_pages_to_scan = 100;
272 /* Milliseconds ksmd should sleep between batches */
273 static unsigned int ksm_thread_sleep_millisecs = 20;
275 /* Checksum of an empty (zeroed) page */
276 static unsigned int zero_checksum __read_mostly;
278 /* Whether to merge empty (zeroed) pages with actual zero pages */
279 static bool ksm_use_zero_pages __read_mostly;
282 /* Zeroed when merging across nodes is not allowed */
283 static unsigned int ksm_merge_across_nodes = 1;
284 static int ksm_nr_node_ids = 1;
286 #define ksm_merge_across_nodes 1U
287 #define ksm_nr_node_ids 1
290 #define KSM_RUN_STOP 0
291 #define KSM_RUN_MERGE 1
292 #define KSM_RUN_UNMERGE 2
293 #define KSM_RUN_OFFLINE 4
294 static unsigned long ksm_run = KSM_RUN_STOP;
295 static void wait_while_offlining(void);
297 static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
298 static DECLARE_WAIT_QUEUE_HEAD(ksm_iter_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_lock 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, FOLL_WRITE, &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,
486 ret = VM_FAULT_WRITE;
488 } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV | VM_FAULT_OOM)));
490 * We must loop because handle_mm_fault() may back out if there's
491 * any difficulty e.g. if pte accessed bit gets updated concurrently.
493 * VM_FAULT_WRITE is what we have been hoping for: it indicates that
494 * COW has been broken, even if the vma does not permit VM_WRITE;
495 * but note that a concurrent fault might break PageKsm for us.
497 * VM_FAULT_SIGBUS could occur if we race with truncation of the
498 * backing file, which also invalidates anonymous pages: that's
499 * okay, that truncation will have unmapped the PageKsm for us.
501 * VM_FAULT_OOM: at the time of writing (late July 2009), setting
502 * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
503 * current task has TIF_MEMDIE set, and will be OOM killed on return
504 * to user; and ksmd, having no mm, would never be chosen for that.
506 * But if the mm is in a limited mem_cgroup, then the fault may fail
507 * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
508 * even ksmd can fail in this way - though it's usually breaking ksm
509 * just to undo a merge it made a moment before, so unlikely to oom.
511 * That's a pity: we might therefore have more kernel pages allocated
512 * than we're counting as nodes in the stable tree; but ksm_do_scan
513 * will retry to break_cow on each pass, so should recover the page
514 * in due course. The important thing is to not let VM_MERGEABLE
515 * be cleared while any such pages might remain in the area.
517 return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
520 static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm,
523 struct vm_area_struct *vma;
524 if (ksm_test_exit(mm))
526 vma = find_vma(mm, addr);
527 if (!vma || vma->vm_start > addr)
529 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
534 static void break_cow(struct rmap_item *rmap_item)
536 struct mm_struct *mm = rmap_item->mm;
537 unsigned long addr = rmap_item->address;
538 struct vm_area_struct *vma;
541 * It is not an accident that whenever we want to break COW
542 * to undo, we also need to drop a reference to the anon_vma.
544 put_anon_vma(rmap_item->anon_vma);
547 vma = find_mergeable_vma(mm, addr);
549 break_ksm(vma, addr);
550 mmap_read_unlock(mm);
553 static struct page *get_mergeable_page(struct rmap_item *rmap_item)
555 struct mm_struct *mm = rmap_item->mm;
556 unsigned long addr = rmap_item->address;
557 struct vm_area_struct *vma;
561 vma = find_mergeable_vma(mm, addr);
565 page = follow_page(vma, addr, FOLL_GET);
566 if (IS_ERR_OR_NULL(page))
568 if (PageAnon(page)) {
569 flush_anon_page(vma, page, addr);
570 flush_dcache_page(page);
576 mmap_read_unlock(mm);
581 * This helper is used for getting right index into array of tree roots.
582 * When merge_across_nodes knob is set to 1, there are only two rb-trees for
583 * stable and unstable pages from all nodes with roots in index 0. Otherwise,
584 * every node has its own stable and unstable tree.
586 static inline int get_kpfn_nid(unsigned long kpfn)
588 return ksm_merge_across_nodes ? 0 : NUMA(pfn_to_nid(kpfn));
591 static struct stable_node *alloc_stable_node_chain(struct stable_node *dup,
592 struct rb_root *root)
594 struct stable_node *chain = alloc_stable_node();
595 VM_BUG_ON(is_stable_node_chain(dup));
597 INIT_HLIST_HEAD(&chain->hlist);
598 chain->chain_prune_time = jiffies;
599 chain->rmap_hlist_len = STABLE_NODE_CHAIN;
600 #if defined (CONFIG_DEBUG_VM) && defined(CONFIG_NUMA)
601 chain->nid = NUMA_NO_NODE; /* debug */
603 ksm_stable_node_chains++;
606 * Put the stable node chain in the first dimension of
607 * the stable tree and at the same time remove the old
610 rb_replace_node(&dup->node, &chain->node, root);
613 * Move the old stable node to the second dimension
614 * queued in the hlist_dup. The invariant is that all
615 * dup stable_nodes in the chain->hlist point to pages
616 * that are write protected and have the exact same
619 stable_node_chain_add_dup(dup, chain);
624 static inline void free_stable_node_chain(struct stable_node *chain,
625 struct rb_root *root)
627 rb_erase(&chain->node, root);
628 free_stable_node(chain);
629 ksm_stable_node_chains--;
632 static void remove_node_from_stable_tree(struct stable_node *stable_node)
634 struct rmap_item *rmap_item;
636 /* check it's not STABLE_NODE_CHAIN or negative */
637 BUG_ON(stable_node->rmap_hlist_len < 0);
639 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
640 if (rmap_item->hlist.next)
644 VM_BUG_ON(stable_node->rmap_hlist_len <= 0);
645 stable_node->rmap_hlist_len--;
646 put_anon_vma(rmap_item->anon_vma);
647 rmap_item->address &= PAGE_MASK;
652 * We need the second aligned pointer of the migrate_nodes
653 * list_head to stay clear from the rb_parent_color union
654 * (aligned and different than any node) and also different
655 * from &migrate_nodes. This will verify that future list.h changes
656 * don't break STABLE_NODE_DUP_HEAD. Only recent gcc can handle it.
658 #if defined(GCC_VERSION) && GCC_VERSION >= 40903
659 BUILD_BUG_ON(STABLE_NODE_DUP_HEAD <= &migrate_nodes);
660 BUILD_BUG_ON(STABLE_NODE_DUP_HEAD >= &migrate_nodes + 1);
663 if (stable_node->head == &migrate_nodes)
664 list_del(&stable_node->list);
666 stable_node_dup_del(stable_node);
667 free_stable_node(stable_node);
670 enum get_ksm_page_flags {
677 * get_ksm_page: checks if the page indicated by the stable node
678 * is still its ksm page, despite having held no reference to it.
679 * In which case we can trust the content of the page, and it
680 * returns the gotten page; but if the page has now been zapped,
681 * remove the stale node from the stable tree and return NULL.
682 * But beware, the stable node's page might be being migrated.
684 * You would expect the stable_node to hold a reference to the ksm page.
685 * But if it increments the page's count, swapping out has to wait for
686 * ksmd to come around again before it can free the page, which may take
687 * seconds or even minutes: much too unresponsive. So instead we use a
688 * "keyhole reference": access to the ksm page from the stable node peeps
689 * out through its keyhole to see if that page still holds the right key,
690 * pointing back to this stable node. This relies on freeing a PageAnon
691 * page to reset its page->mapping to NULL, and relies on no other use of
692 * a page to put something that might look like our key in page->mapping.
693 * is on its way to being freed; but it is an anomaly to bear in mind.
695 static struct page *get_ksm_page(struct stable_node *stable_node,
696 enum get_ksm_page_flags flags)
699 void *expected_mapping;
702 expected_mapping = (void *)((unsigned long)stable_node |
705 kpfn = READ_ONCE(stable_node->kpfn); /* Address dependency. */
706 page = pfn_to_page(kpfn);
707 if (READ_ONCE(page->mapping) != expected_mapping)
711 * We cannot do anything with the page while its refcount is 0.
712 * Usually 0 means free, or tail of a higher-order page: in which
713 * case this node is no longer referenced, and should be freed;
714 * however, it might mean that the page is under page_ref_freeze().
715 * The __remove_mapping() case is easy, again the node is now stale;
716 * the same is in reuse_ksm_page() case; but if page is swapcache
717 * in migrate_page_move_mapping(), it might still be our page,
718 * in which case it's essential to keep the node.
720 while (!get_page_unless_zero(page)) {
722 * Another check for page->mapping != expected_mapping would
723 * work here too. We have chosen the !PageSwapCache test to
724 * optimize the common case, when the page is or is about to
725 * be freed: PageSwapCache is cleared (under spin_lock_irq)
726 * in the ref_freeze section of __remove_mapping(); but Anon
727 * page->mapping reset to NULL later, in free_pages_prepare().
729 if (!PageSwapCache(page))
734 if (READ_ONCE(page->mapping) != expected_mapping) {
739 if (flags == GET_KSM_PAGE_TRYLOCK) {
740 if (!trylock_page(page)) {
742 return ERR_PTR(-EBUSY);
744 } else if (flags == GET_KSM_PAGE_LOCK)
747 if (flags != GET_KSM_PAGE_NOLOCK) {
748 if (READ_ONCE(page->mapping) != expected_mapping) {
758 * We come here from above when page->mapping or !PageSwapCache
759 * suggests that the node is stale; but it might be under migration.
760 * We need smp_rmb(), matching the smp_wmb() in ksm_migrate_page(),
761 * before checking whether node->kpfn has been changed.
764 if (READ_ONCE(stable_node->kpfn) != kpfn)
766 remove_node_from_stable_tree(stable_node);
771 * Removing rmap_item from stable or unstable tree.
772 * This function will clean the information from the stable/unstable tree.
774 static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
776 if (rmap_item->address & STABLE_FLAG) {
777 struct stable_node *stable_node;
780 stable_node = rmap_item->head;
781 page = get_ksm_page(stable_node, GET_KSM_PAGE_LOCK);
785 hlist_del(&rmap_item->hlist);
789 if (!hlist_empty(&stable_node->hlist))
793 VM_BUG_ON(stable_node->rmap_hlist_len <= 0);
794 stable_node->rmap_hlist_len--;
796 put_anon_vma(rmap_item->anon_vma);
797 rmap_item->head = NULL;
798 rmap_item->address &= PAGE_MASK;
800 } else if (rmap_item->address & UNSTABLE_FLAG) {
803 * Usually ksmd can and must skip the rb_erase, because
804 * root_unstable_tree was already reset to RB_ROOT.
805 * But be careful when an mm is exiting: do the rb_erase
806 * if this rmap_item was inserted by this scan, rather
807 * than left over from before.
809 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
812 rb_erase(&rmap_item->node,
813 root_unstable_tree + NUMA(rmap_item->nid));
814 ksm_pages_unshared--;
815 rmap_item->address &= PAGE_MASK;
818 cond_resched(); /* we're called from many long loops */
821 static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
822 struct rmap_item **rmap_list)
825 struct rmap_item *rmap_item = *rmap_list;
826 *rmap_list = rmap_item->rmap_list;
827 remove_rmap_item_from_tree(rmap_item);
828 free_rmap_item(rmap_item);
833 * Though it's very tempting to unmerge rmap_items from stable tree rather
834 * than check every pte of a given vma, the locking doesn't quite work for
835 * that - an rmap_item is assigned to the stable tree after inserting ksm
836 * page and upping mmap_lock. Nor does it fit with the way we skip dup'ing
837 * rmap_items from parent to child at fork time (so as not to waste time
838 * if exit comes before the next scan reaches it).
840 * Similarly, although we'd like to remove rmap_items (so updating counts
841 * and freeing memory) when unmerging an area, it's easier to leave that
842 * to the next pass of ksmd - consider, for example, how ksmd might be
843 * in cmp_and_merge_page on one of the rmap_items we would be removing.
845 static int unmerge_ksm_pages(struct vm_area_struct *vma,
846 unsigned long start, unsigned long end)
851 for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
852 if (ksm_test_exit(vma->vm_mm))
854 if (signal_pending(current))
857 err = break_ksm(vma, addr);
862 static inline struct stable_node *page_stable_node(struct page *page)
864 return PageKsm(page) ? page_rmapping(page) : NULL;
867 static inline void set_page_stable_node(struct page *page,
868 struct stable_node *stable_node)
870 page->mapping = (void *)((unsigned long)stable_node | PAGE_MAPPING_KSM);
875 * Only called through the sysfs control interface:
877 static int remove_stable_node(struct stable_node *stable_node)
882 page = get_ksm_page(stable_node, GET_KSM_PAGE_LOCK);
885 * get_ksm_page did remove_node_from_stable_tree itself.
891 * Page could be still mapped if this races with __mmput() running in
892 * between ksm_exit() and exit_mmap(). Just refuse to let
893 * merge_across_nodes/max_page_sharing be switched.
896 if (!page_mapped(page)) {
898 * The stable node did not yet appear stale to get_ksm_page(),
899 * since that allows for an unmapped ksm page to be recognized
900 * right up until it is freed; but the node is safe to remove.
901 * This page might be in a pagevec waiting to be freed,
902 * or it might be PageSwapCache (perhaps under writeback),
903 * or it might have been removed from swapcache a moment ago.
905 set_page_stable_node(page, NULL);
906 remove_node_from_stable_tree(stable_node);
915 static int remove_stable_node_chain(struct stable_node *stable_node,
916 struct rb_root *root)
918 struct stable_node *dup;
919 struct hlist_node *hlist_safe;
921 if (!is_stable_node_chain(stable_node)) {
922 VM_BUG_ON(is_stable_node_dup(stable_node));
923 if (remove_stable_node(stable_node))
929 hlist_for_each_entry_safe(dup, hlist_safe,
930 &stable_node->hlist, hlist_dup) {
931 VM_BUG_ON(!is_stable_node_dup(dup));
932 if (remove_stable_node(dup))
935 BUG_ON(!hlist_empty(&stable_node->hlist));
936 free_stable_node_chain(stable_node, root);
940 static int remove_all_stable_nodes(void)
942 struct stable_node *stable_node, *next;
946 for (nid = 0; nid < ksm_nr_node_ids; nid++) {
947 while (root_stable_tree[nid].rb_node) {
948 stable_node = rb_entry(root_stable_tree[nid].rb_node,
949 struct stable_node, node);
950 if (remove_stable_node_chain(stable_node,
951 root_stable_tree + nid)) {
953 break; /* proceed to next nid */
958 list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
959 if (remove_stable_node(stable_node))
966 static int unmerge_and_remove_all_rmap_items(void)
968 struct mm_slot *mm_slot;
969 struct mm_struct *mm;
970 struct vm_area_struct *vma;
973 spin_lock(&ksm_mmlist_lock);
974 ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
975 struct mm_slot, mm_list);
976 spin_unlock(&ksm_mmlist_lock);
978 for (mm_slot = ksm_scan.mm_slot;
979 mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
982 for (vma = mm->mmap; vma; vma = vma->vm_next) {
983 if (ksm_test_exit(mm))
985 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
987 err = unmerge_ksm_pages(vma,
988 vma->vm_start, vma->vm_end);
993 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
994 mmap_read_unlock(mm);
996 spin_lock(&ksm_mmlist_lock);
997 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
998 struct mm_slot, mm_list);
999 if (ksm_test_exit(mm)) {
1000 hash_del(&mm_slot->link);
1001 list_del(&mm_slot->mm_list);
1002 spin_unlock(&ksm_mmlist_lock);
1004 free_mm_slot(mm_slot);
1005 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1008 spin_unlock(&ksm_mmlist_lock);
1011 /* Clean up stable nodes, but don't worry if some are still busy */
1012 remove_all_stable_nodes();
1017 mmap_read_unlock(mm);
1018 spin_lock(&ksm_mmlist_lock);
1019 ksm_scan.mm_slot = &ksm_mm_head;
1020 spin_unlock(&ksm_mmlist_lock);
1023 #endif /* CONFIG_SYSFS */
1025 static u32 calc_checksum(struct page *page)
1028 void *addr = kmap_atomic(page);
1029 checksum = xxhash(addr, PAGE_SIZE, 0);
1030 kunmap_atomic(addr);
1034 static int write_protect_page(struct vm_area_struct *vma, struct page *page,
1037 struct mm_struct *mm = vma->vm_mm;
1038 struct page_vma_mapped_walk pvmw = {
1044 struct mmu_notifier_range range;
1046 pvmw.address = page_address_in_vma(page, vma);
1047 if (pvmw.address == -EFAULT)
1050 BUG_ON(PageTransCompound(page));
1052 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
1054 pvmw.address + PAGE_SIZE);
1055 mmu_notifier_invalidate_range_start(&range);
1057 if (!page_vma_mapped_walk(&pvmw))
1059 if (WARN_ONCE(!pvmw.pte, "Unexpected PMD mapping?"))
1062 if (pte_write(*pvmw.pte) || pte_dirty(*pvmw.pte) ||
1063 (pte_protnone(*pvmw.pte) && pte_savedwrite(*pvmw.pte)) ||
1064 mm_tlb_flush_pending(mm)) {
1067 swapped = PageSwapCache(page);
1068 flush_cache_page(vma, pvmw.address, page_to_pfn(page));
1070 * Ok this is tricky, when get_user_pages_fast() run it doesn't
1071 * take any lock, therefore the check that we are going to make
1072 * with the pagecount against the mapcount is racey and
1073 * O_DIRECT can happen right after the check.
1074 * So we clear the pte and flush the tlb before the check
1075 * this assure us that no O_DIRECT can happen after the check
1076 * or in the middle of the check.
1078 * No need to notify as we are downgrading page table to read
1079 * only not changing it to point to a new page.
1081 * See Documentation/vm/mmu_notifier.rst
1083 entry = ptep_clear_flush(vma, pvmw.address, pvmw.pte);
1085 * Check that no O_DIRECT or similar I/O is in progress on the
1088 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
1089 set_pte_at(mm, pvmw.address, pvmw.pte, entry);
1092 if (pte_dirty(entry))
1093 set_page_dirty(page);
1095 if (pte_protnone(entry))
1096 entry = pte_mkclean(pte_clear_savedwrite(entry));
1098 entry = pte_mkclean(pte_wrprotect(entry));
1099 set_pte_at_notify(mm, pvmw.address, pvmw.pte, entry);
1101 *orig_pte = *pvmw.pte;
1105 page_vma_mapped_walk_done(&pvmw);
1107 mmu_notifier_invalidate_range_end(&range);
1113 * replace_page - replace page in vma by new ksm page
1114 * @vma: vma that holds the pte pointing to page
1115 * @page: the page we are replacing by kpage
1116 * @kpage: the ksm page we replace page by
1117 * @orig_pte: the original value of the pte
1119 * Returns 0 on success, -EFAULT on failure.
1121 static int replace_page(struct vm_area_struct *vma, struct page *page,
1122 struct page *kpage, pte_t orig_pte)
1124 struct mm_struct *mm = vma->vm_mm;
1131 struct mmu_notifier_range range;
1133 addr = page_address_in_vma(page, vma);
1134 if (addr == -EFAULT)
1137 pmd = mm_find_pmd(mm, addr);
1141 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, addr,
1143 mmu_notifier_invalidate_range_start(&range);
1145 ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
1146 if (!pte_same(*ptep, orig_pte)) {
1147 pte_unmap_unlock(ptep, ptl);
1152 * No need to check ksm_use_zero_pages here: we can only have a
1153 * zero_page here if ksm_use_zero_pages was enabled already.
1155 if (!is_zero_pfn(page_to_pfn(kpage))) {
1157 page_add_anon_rmap(kpage, vma, addr, false);
1158 newpte = mk_pte(kpage, vma->vm_page_prot);
1160 newpte = pte_mkspecial(pfn_pte(page_to_pfn(kpage),
1161 vma->vm_page_prot));
1163 * We're replacing an anonymous page with a zero page, which is
1164 * not anonymous. We need to do proper accounting otherwise we
1165 * will get wrong values in /proc, and a BUG message in dmesg
1166 * when tearing down the mm.
1168 dec_mm_counter(mm, MM_ANONPAGES);
1171 flush_cache_page(vma, addr, pte_pfn(*ptep));
1173 * No need to notify as we are replacing a read only page with another
1174 * read only page with the same content.
1176 * See Documentation/vm/mmu_notifier.rst
1178 ptep_clear_flush(vma, addr, ptep);
1179 set_pte_at_notify(mm, addr, ptep, newpte);
1181 page_remove_rmap(page, false);
1182 if (!page_mapped(page))
1183 try_to_free_swap(page);
1186 pte_unmap_unlock(ptep, ptl);
1189 mmu_notifier_invalidate_range_end(&range);
1195 * try_to_merge_one_page - take two pages and merge them into one
1196 * @vma: the vma that holds the pte pointing to page
1197 * @page: the PageAnon page that we want to replace with kpage
1198 * @kpage: the PageKsm page that we want to map instead of page,
1199 * or NULL the first time when we want to use page as kpage.
1201 * This function returns 0 if the pages were merged, -EFAULT otherwise.
1203 static int try_to_merge_one_page(struct vm_area_struct *vma,
1204 struct page *page, struct page *kpage)
1206 pte_t orig_pte = __pte(0);
1209 if (page == kpage) /* ksm page forked */
1212 if (!PageAnon(page))
1216 * We need the page lock to read a stable PageSwapCache in
1217 * write_protect_page(). We use trylock_page() instead of
1218 * lock_page() because we don't want to wait here - we
1219 * prefer to continue scanning and merging different pages,
1220 * then come back to this page when it is unlocked.
1222 if (!trylock_page(page))
1225 if (PageTransCompound(page)) {
1226 if (split_huge_page(page))
1231 * If this anonymous page is mapped only here, its pte may need
1232 * to be write-protected. If it's mapped elsewhere, all of its
1233 * ptes are necessarily already write-protected. But in either
1234 * case, we need to lock and check page_count is not raised.
1236 if (write_protect_page(vma, page, &orig_pte) == 0) {
1239 * While we hold page lock, upgrade page from
1240 * PageAnon+anon_vma to PageKsm+NULL stable_node:
1241 * stable_tree_insert() will update stable_node.
1243 set_page_stable_node(page, NULL);
1244 mark_page_accessed(page);
1246 * Page reclaim just frees a clean page with no dirty
1247 * ptes: make sure that the ksm page would be swapped.
1249 if (!PageDirty(page))
1252 } else if (pages_identical(page, kpage))
1253 err = replace_page(vma, page, kpage, orig_pte);
1256 if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
1257 munlock_vma_page(page);
1258 if (!PageMlocked(kpage)) {
1261 mlock_vma_page(kpage);
1262 page = kpage; /* for final unlock */
1273 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
1274 * but no new kernel page is allocated: kpage must already be a ksm page.
1276 * This function returns 0 if the pages were merged, -EFAULT otherwise.
1278 static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
1279 struct page *page, struct page *kpage)
1281 struct mm_struct *mm = rmap_item->mm;
1282 struct vm_area_struct *vma;
1286 vma = find_mergeable_vma(mm, rmap_item->address);
1290 err = try_to_merge_one_page(vma, page, kpage);
1294 /* Unstable nid is in union with stable anon_vma: remove first */
1295 remove_rmap_item_from_tree(rmap_item);
1297 /* Must get reference to anon_vma while still holding mmap_lock */
1298 rmap_item->anon_vma = vma->anon_vma;
1299 get_anon_vma(vma->anon_vma);
1301 mmap_read_unlock(mm);
1306 * try_to_merge_two_pages - take two identical pages and prepare them
1307 * to be merged into one page.
1309 * This function returns the kpage if we successfully merged two identical
1310 * pages into one ksm page, NULL otherwise.
1312 * Note that this function upgrades page to ksm page: if one of the pages
1313 * is already a ksm page, try_to_merge_with_ksm_page should be used.
1315 static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
1317 struct rmap_item *tree_rmap_item,
1318 struct page *tree_page)
1322 err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
1324 err = try_to_merge_with_ksm_page(tree_rmap_item,
1327 * If that fails, we have a ksm page with only one pte
1328 * pointing to it: so break it.
1331 break_cow(rmap_item);
1333 return err ? NULL : page;
1336 static __always_inline
1337 bool __is_page_sharing_candidate(struct stable_node *stable_node, int offset)
1339 VM_BUG_ON(stable_node->rmap_hlist_len < 0);
1341 * Check that at least one mapping still exists, otherwise
1342 * there's no much point to merge and share with this
1343 * stable_node, as the underlying tree_page of the other
1344 * sharer is going to be freed soon.
1346 return stable_node->rmap_hlist_len &&
1347 stable_node->rmap_hlist_len + offset < ksm_max_page_sharing;
1350 static __always_inline
1351 bool is_page_sharing_candidate(struct stable_node *stable_node)
1353 return __is_page_sharing_candidate(stable_node, 0);
1356 static struct page *stable_node_dup(struct stable_node **_stable_node_dup,
1357 struct stable_node **_stable_node,
1358 struct rb_root *root,
1359 bool prune_stale_stable_nodes)
1361 struct stable_node *dup, *found = NULL, *stable_node = *_stable_node;
1362 struct hlist_node *hlist_safe;
1363 struct page *_tree_page, *tree_page = NULL;
1365 int found_rmap_hlist_len;
1367 if (!prune_stale_stable_nodes ||
1368 time_before(jiffies, stable_node->chain_prune_time +
1370 ksm_stable_node_chains_prune_millisecs)))
1371 prune_stale_stable_nodes = false;
1373 stable_node->chain_prune_time = jiffies;
1375 hlist_for_each_entry_safe(dup, hlist_safe,
1376 &stable_node->hlist, hlist_dup) {
1379 * We must walk all stable_node_dup to prune the stale
1380 * stable nodes during lookup.
1382 * get_ksm_page can drop the nodes from the
1383 * stable_node->hlist if they point to freed pages
1384 * (that's why we do a _safe walk). The "dup"
1385 * stable_node parameter itself will be freed from
1386 * under us if it returns NULL.
1388 _tree_page = get_ksm_page(dup, GET_KSM_PAGE_NOLOCK);
1392 if (is_page_sharing_candidate(dup)) {
1394 dup->rmap_hlist_len > found_rmap_hlist_len) {
1396 put_page(tree_page);
1398 found_rmap_hlist_len = found->rmap_hlist_len;
1399 tree_page = _tree_page;
1401 /* skip put_page for found dup */
1402 if (!prune_stale_stable_nodes)
1407 put_page(_tree_page);
1412 * nr is counting all dups in the chain only if
1413 * prune_stale_stable_nodes is true, otherwise we may
1414 * break the loop at nr == 1 even if there are
1417 if (prune_stale_stable_nodes && nr == 1) {
1419 * If there's not just one entry it would
1420 * corrupt memory, better BUG_ON. In KSM
1421 * context with no lock held it's not even
1424 BUG_ON(stable_node->hlist.first->next);
1427 * There's just one entry and it is below the
1428 * deduplication limit so drop the chain.
1430 rb_replace_node(&stable_node->node, &found->node,
1432 free_stable_node(stable_node);
1433 ksm_stable_node_chains--;
1434 ksm_stable_node_dups--;
1436 * NOTE: the caller depends on the stable_node
1437 * to be equal to stable_node_dup if the chain
1440 *_stable_node = found;
1442 * Just for robustneess as stable_node is
1443 * otherwise left as a stable pointer, the
1444 * compiler shall optimize it away at build
1448 } else if (stable_node->hlist.first != &found->hlist_dup &&
1449 __is_page_sharing_candidate(found, 1)) {
1451 * If the found stable_node dup can accept one
1452 * more future merge (in addition to the one
1453 * that is underway) and is not at the head of
1454 * the chain, put it there so next search will
1455 * be quicker in the !prune_stale_stable_nodes
1458 * NOTE: it would be inaccurate to use nr > 1
1459 * instead of checking the hlist.first pointer
1460 * directly, because in the
1461 * prune_stale_stable_nodes case "nr" isn't
1462 * the position of the found dup in the chain,
1463 * but the total number of dups in the chain.
1465 hlist_del(&found->hlist_dup);
1466 hlist_add_head(&found->hlist_dup,
1467 &stable_node->hlist);
1471 *_stable_node_dup = found;
1475 static struct stable_node *stable_node_dup_any(struct stable_node *stable_node,
1476 struct rb_root *root)
1478 if (!is_stable_node_chain(stable_node))
1480 if (hlist_empty(&stable_node->hlist)) {
1481 free_stable_node_chain(stable_node, root);
1484 return hlist_entry(stable_node->hlist.first,
1485 typeof(*stable_node), hlist_dup);
1489 * Like for get_ksm_page, this function can free the *_stable_node and
1490 * *_stable_node_dup if the returned tree_page is NULL.
1492 * It can also free and overwrite *_stable_node with the found
1493 * stable_node_dup if the chain is collapsed (in which case
1494 * *_stable_node will be equal to *_stable_node_dup like if the chain
1495 * never existed). It's up to the caller to verify tree_page is not
1496 * NULL before dereferencing *_stable_node or *_stable_node_dup.
1498 * *_stable_node_dup is really a second output parameter of this
1499 * function and will be overwritten in all cases, the caller doesn't
1500 * need to initialize it.
1502 static struct page *__stable_node_chain(struct stable_node **_stable_node_dup,
1503 struct stable_node **_stable_node,
1504 struct rb_root *root,
1505 bool prune_stale_stable_nodes)
1507 struct stable_node *stable_node = *_stable_node;
1508 if (!is_stable_node_chain(stable_node)) {
1509 if (is_page_sharing_candidate(stable_node)) {
1510 *_stable_node_dup = stable_node;
1511 return get_ksm_page(stable_node, GET_KSM_PAGE_NOLOCK);
1514 * _stable_node_dup set to NULL means the stable_node
1515 * reached the ksm_max_page_sharing limit.
1517 *_stable_node_dup = NULL;
1520 return stable_node_dup(_stable_node_dup, _stable_node, root,
1521 prune_stale_stable_nodes);
1524 static __always_inline struct page *chain_prune(struct stable_node **s_n_d,
1525 struct stable_node **s_n,
1526 struct rb_root *root)
1528 return __stable_node_chain(s_n_d, s_n, root, true);
1531 static __always_inline struct page *chain(struct stable_node **s_n_d,
1532 struct stable_node *s_n,
1533 struct rb_root *root)
1535 struct stable_node *old_stable_node = s_n;
1536 struct page *tree_page;
1538 tree_page = __stable_node_chain(s_n_d, &s_n, root, false);
1539 /* not pruning dups so s_n cannot have changed */
1540 VM_BUG_ON(s_n != old_stable_node);
1545 * stable_tree_search - search for page inside the stable tree
1547 * This function checks if there is a page inside the stable tree
1548 * with identical content to the page that we are scanning right now.
1550 * This function returns the stable tree node of identical content if found,
1553 static struct page *stable_tree_search(struct page *page)
1556 struct rb_root *root;
1557 struct rb_node **new;
1558 struct rb_node *parent;
1559 struct stable_node *stable_node, *stable_node_dup, *stable_node_any;
1560 struct stable_node *page_node;
1562 page_node = page_stable_node(page);
1563 if (page_node && page_node->head != &migrate_nodes) {
1564 /* ksm page forked */
1569 nid = get_kpfn_nid(page_to_pfn(page));
1570 root = root_stable_tree + nid;
1572 new = &root->rb_node;
1576 struct page *tree_page;
1580 stable_node = rb_entry(*new, struct stable_node, node);
1581 stable_node_any = NULL;
1582 tree_page = chain_prune(&stable_node_dup, &stable_node, root);
1584 * NOTE: stable_node may have been freed by
1585 * chain_prune() if the returned stable_node_dup is
1586 * not NULL. stable_node_dup may have been inserted in
1587 * the rbtree instead as a regular stable_node (in
1588 * order to collapse the stable_node chain if a single
1589 * stable_node dup was found in it). In such case the
1590 * stable_node is overwritten by the calleee to point
1591 * to the stable_node_dup that was collapsed in the
1592 * stable rbtree and stable_node will be equal to
1593 * stable_node_dup like if the chain never existed.
1595 if (!stable_node_dup) {
1597 * Either all stable_node dups were full in
1598 * this stable_node chain, or this chain was
1599 * empty and should be rb_erased.
1601 stable_node_any = stable_node_dup_any(stable_node,
1603 if (!stable_node_any) {
1604 /* rb_erase just run */
1608 * Take any of the stable_node dups page of
1609 * this stable_node chain to let the tree walk
1610 * continue. All KSM pages belonging to the
1611 * stable_node dups in a stable_node chain
1612 * have the same content and they're
1613 * write protected at all times. Any will work
1614 * fine to continue the walk.
1616 tree_page = get_ksm_page(stable_node_any,
1617 GET_KSM_PAGE_NOLOCK);
1619 VM_BUG_ON(!stable_node_dup ^ !!stable_node_any);
1622 * If we walked over a stale stable_node,
1623 * get_ksm_page() will call rb_erase() and it
1624 * may rebalance the tree from under us. So
1625 * restart the search from scratch. Returning
1626 * NULL would be safe too, but we'd generate
1627 * false negative insertions just because some
1628 * stable_node was stale.
1633 ret = memcmp_pages(page, tree_page);
1634 put_page(tree_page);
1638 new = &parent->rb_left;
1640 new = &parent->rb_right;
1643 VM_BUG_ON(page_node->head != &migrate_nodes);
1645 * Test if the migrated page should be merged
1646 * into a stable node dup. If the mapcount is
1647 * 1 we can migrate it with another KSM page
1648 * without adding it to the chain.
1650 if (page_mapcount(page) > 1)
1654 if (!stable_node_dup) {
1656 * If the stable_node is a chain and
1657 * we got a payload match in memcmp
1658 * but we cannot merge the scanned
1659 * page in any of the existing
1660 * stable_node dups because they're
1661 * all full, we need to wait the
1662 * scanned page to find itself a match
1663 * in the unstable tree to create a
1664 * brand new KSM page to add later to
1665 * the dups of this stable_node.
1671 * Lock and unlock the stable_node's page (which
1672 * might already have been migrated) so that page
1673 * migration is sure to notice its raised count.
1674 * It would be more elegant to return stable_node
1675 * than kpage, but that involves more changes.
1677 tree_page = get_ksm_page(stable_node_dup,
1678 GET_KSM_PAGE_TRYLOCK);
1680 if (PTR_ERR(tree_page) == -EBUSY)
1681 return ERR_PTR(-EBUSY);
1683 if (unlikely(!tree_page))
1685 * The tree may have been rebalanced,
1686 * so re-evaluate parent and new.
1689 unlock_page(tree_page);
1691 if (get_kpfn_nid(stable_node_dup->kpfn) !=
1692 NUMA(stable_node_dup->nid)) {
1693 put_page(tree_page);
1703 list_del(&page_node->list);
1704 DO_NUMA(page_node->nid = nid);
1705 rb_link_node(&page_node->node, parent, new);
1706 rb_insert_color(&page_node->node, root);
1708 if (is_page_sharing_candidate(page_node)) {
1716 * If stable_node was a chain and chain_prune collapsed it,
1717 * stable_node has been updated to be the new regular
1718 * stable_node. A collapse of the chain is indistinguishable
1719 * from the case there was no chain in the stable
1720 * rbtree. Otherwise stable_node is the chain and
1721 * stable_node_dup is the dup to replace.
1723 if (stable_node_dup == stable_node) {
1724 VM_BUG_ON(is_stable_node_chain(stable_node_dup));
1725 VM_BUG_ON(is_stable_node_dup(stable_node_dup));
1726 /* there is no chain */
1728 VM_BUG_ON(page_node->head != &migrate_nodes);
1729 list_del(&page_node->list);
1730 DO_NUMA(page_node->nid = nid);
1731 rb_replace_node(&stable_node_dup->node,
1734 if (is_page_sharing_candidate(page_node))
1739 rb_erase(&stable_node_dup->node, root);
1743 VM_BUG_ON(!is_stable_node_chain(stable_node));
1744 __stable_node_dup_del(stable_node_dup);
1746 VM_BUG_ON(page_node->head != &migrate_nodes);
1747 list_del(&page_node->list);
1748 DO_NUMA(page_node->nid = nid);
1749 stable_node_chain_add_dup(page_node, stable_node);
1750 if (is_page_sharing_candidate(page_node))
1758 stable_node_dup->head = &migrate_nodes;
1759 list_add(&stable_node_dup->list, stable_node_dup->head);
1763 /* stable_node_dup could be null if it reached the limit */
1764 if (!stable_node_dup)
1765 stable_node_dup = stable_node_any;
1767 * If stable_node was a chain and chain_prune collapsed it,
1768 * stable_node has been updated to be the new regular
1769 * stable_node. A collapse of the chain is indistinguishable
1770 * from the case there was no chain in the stable
1771 * rbtree. Otherwise stable_node is the chain and
1772 * stable_node_dup is the dup to replace.
1774 if (stable_node_dup == stable_node) {
1775 VM_BUG_ON(is_stable_node_chain(stable_node_dup));
1776 VM_BUG_ON(is_stable_node_dup(stable_node_dup));
1777 /* chain is missing so create it */
1778 stable_node = alloc_stable_node_chain(stable_node_dup,
1784 * Add this stable_node dup that was
1785 * migrated to the stable_node chain
1786 * of the current nid for this page
1789 VM_BUG_ON(!is_stable_node_chain(stable_node));
1790 VM_BUG_ON(!is_stable_node_dup(stable_node_dup));
1791 VM_BUG_ON(page_node->head != &migrate_nodes);
1792 list_del(&page_node->list);
1793 DO_NUMA(page_node->nid = nid);
1794 stable_node_chain_add_dup(page_node, stable_node);
1799 * stable_tree_insert - insert stable tree node pointing to new ksm page
1800 * into the stable tree.
1802 * This function returns the stable tree node just allocated on success,
1805 static struct stable_node *stable_tree_insert(struct page *kpage)
1809 struct rb_root *root;
1810 struct rb_node **new;
1811 struct rb_node *parent;
1812 struct stable_node *stable_node, *stable_node_dup, *stable_node_any;
1813 bool need_chain = false;
1815 kpfn = page_to_pfn(kpage);
1816 nid = get_kpfn_nid(kpfn);
1817 root = root_stable_tree + nid;
1820 new = &root->rb_node;
1823 struct page *tree_page;
1827 stable_node = rb_entry(*new, struct stable_node, node);
1828 stable_node_any = NULL;
1829 tree_page = chain(&stable_node_dup, stable_node, root);
1830 if (!stable_node_dup) {
1832 * Either all stable_node dups were full in
1833 * this stable_node chain, or this chain was
1834 * empty and should be rb_erased.
1836 stable_node_any = stable_node_dup_any(stable_node,
1838 if (!stable_node_any) {
1839 /* rb_erase just run */
1843 * Take any of the stable_node dups page of
1844 * this stable_node chain to let the tree walk
1845 * continue. All KSM pages belonging to the
1846 * stable_node dups in a stable_node chain
1847 * have the same content and they're
1848 * write protected at all times. Any will work
1849 * fine to continue the walk.
1851 tree_page = get_ksm_page(stable_node_any,
1852 GET_KSM_PAGE_NOLOCK);
1854 VM_BUG_ON(!stable_node_dup ^ !!stable_node_any);
1857 * If we walked over a stale stable_node,
1858 * get_ksm_page() will call rb_erase() and it
1859 * may rebalance the tree from under us. So
1860 * restart the search from scratch. Returning
1861 * NULL would be safe too, but we'd generate
1862 * false negative insertions just because some
1863 * stable_node was stale.
1868 ret = memcmp_pages(kpage, tree_page);
1869 put_page(tree_page);
1873 new = &parent->rb_left;
1875 new = &parent->rb_right;
1882 stable_node_dup = alloc_stable_node();
1883 if (!stable_node_dup)
1886 INIT_HLIST_HEAD(&stable_node_dup->hlist);
1887 stable_node_dup->kpfn = kpfn;
1888 set_page_stable_node(kpage, stable_node_dup);
1889 stable_node_dup->rmap_hlist_len = 0;
1890 DO_NUMA(stable_node_dup->nid = nid);
1892 rb_link_node(&stable_node_dup->node, parent, new);
1893 rb_insert_color(&stable_node_dup->node, root);
1895 if (!is_stable_node_chain(stable_node)) {
1896 struct stable_node *orig = stable_node;
1897 /* chain is missing so create it */
1898 stable_node = alloc_stable_node_chain(orig, root);
1900 free_stable_node(stable_node_dup);
1904 stable_node_chain_add_dup(stable_node_dup, stable_node);
1907 return stable_node_dup;
1911 * unstable_tree_search_insert - search for identical page,
1912 * else insert rmap_item into the unstable tree.
1914 * This function searches for a page in the unstable tree identical to the
1915 * page currently being scanned; and if no identical page is found in the
1916 * tree, we insert rmap_item as a new object into the unstable tree.
1918 * This function returns pointer to rmap_item found to be identical
1919 * to the currently scanned page, NULL otherwise.
1921 * This function does both searching and inserting, because they share
1922 * the same walking algorithm in an rbtree.
1925 struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1927 struct page **tree_pagep)
1929 struct rb_node **new;
1930 struct rb_root *root;
1931 struct rb_node *parent = NULL;
1934 nid = get_kpfn_nid(page_to_pfn(page));
1935 root = root_unstable_tree + nid;
1936 new = &root->rb_node;
1939 struct rmap_item *tree_rmap_item;
1940 struct page *tree_page;
1944 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
1945 tree_page = get_mergeable_page(tree_rmap_item);
1950 * Don't substitute a ksm page for a forked page.
1952 if (page == tree_page) {
1953 put_page(tree_page);
1957 ret = memcmp_pages(page, tree_page);
1961 put_page(tree_page);
1962 new = &parent->rb_left;
1963 } else if (ret > 0) {
1964 put_page(tree_page);
1965 new = &parent->rb_right;
1966 } else if (!ksm_merge_across_nodes &&
1967 page_to_nid(tree_page) != nid) {
1969 * If tree_page has been migrated to another NUMA node,
1970 * it will be flushed out and put in the right unstable
1971 * tree next time: only merge with it when across_nodes.
1973 put_page(tree_page);
1976 *tree_pagep = tree_page;
1977 return tree_rmap_item;
1981 rmap_item->address |= UNSTABLE_FLAG;
1982 rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1983 DO_NUMA(rmap_item->nid = nid);
1984 rb_link_node(&rmap_item->node, parent, new);
1985 rb_insert_color(&rmap_item->node, root);
1987 ksm_pages_unshared++;
1992 * stable_tree_append - add another rmap_item to the linked list of
1993 * rmap_items hanging off a given node of the stable tree, all sharing
1994 * the same ksm page.
1996 static void stable_tree_append(struct rmap_item *rmap_item,
1997 struct stable_node *stable_node,
1998 bool max_page_sharing_bypass)
2001 * rmap won't find this mapping if we don't insert the
2002 * rmap_item in the right stable_node
2003 * duplicate. page_migration could break later if rmap breaks,
2004 * so we can as well crash here. We really need to check for
2005 * rmap_hlist_len == STABLE_NODE_CHAIN, but we can as well check
2006 * for other negative values as an underflow if detected here
2007 * for the first time (and not when decreasing rmap_hlist_len)
2008 * would be sign of memory corruption in the stable_node.
2010 BUG_ON(stable_node->rmap_hlist_len < 0);
2012 stable_node->rmap_hlist_len++;
2013 if (!max_page_sharing_bypass)
2014 /* possibly non fatal but unexpected overflow, only warn */
2015 WARN_ON_ONCE(stable_node->rmap_hlist_len >
2016 ksm_max_page_sharing);
2018 rmap_item->head = stable_node;
2019 rmap_item->address |= STABLE_FLAG;
2020 hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
2022 if (rmap_item->hlist.next)
2023 ksm_pages_sharing++;
2029 * cmp_and_merge_page - first see if page can be merged into the stable tree;
2030 * if not, compare checksum to previous and if it's the same, see if page can
2031 * be inserted into the unstable tree, or merged with a page already there and
2032 * both transferred to the stable tree.
2034 * @page: the page that we are searching identical page to.
2035 * @rmap_item: the reverse mapping into the virtual address of this page
2037 static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
2039 struct mm_struct *mm = rmap_item->mm;
2040 struct rmap_item *tree_rmap_item;
2041 struct page *tree_page = NULL;
2042 struct stable_node *stable_node;
2044 unsigned int checksum;
2046 bool max_page_sharing_bypass = false;
2048 stable_node = page_stable_node(page);
2050 if (stable_node->head != &migrate_nodes &&
2051 get_kpfn_nid(READ_ONCE(stable_node->kpfn)) !=
2052 NUMA(stable_node->nid)) {
2053 stable_node_dup_del(stable_node);
2054 stable_node->head = &migrate_nodes;
2055 list_add(&stable_node->list, stable_node->head);
2057 if (stable_node->head != &migrate_nodes &&
2058 rmap_item->head == stable_node)
2061 * If it's a KSM fork, allow it to go over the sharing limit
2064 if (!is_page_sharing_candidate(stable_node))
2065 max_page_sharing_bypass = true;
2068 /* We first start with searching the page inside the stable tree */
2069 kpage = stable_tree_search(page);
2070 if (kpage == page && rmap_item->head == stable_node) {
2075 remove_rmap_item_from_tree(rmap_item);
2078 if (PTR_ERR(kpage) == -EBUSY)
2081 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
2084 * The page was successfully merged:
2085 * add its rmap_item to the stable tree.
2088 stable_tree_append(rmap_item, page_stable_node(kpage),
2089 max_page_sharing_bypass);
2097 * If the hash value of the page has changed from the last time
2098 * we calculated it, this page is changing frequently: therefore we
2099 * don't want to insert it in the unstable tree, and we don't want
2100 * to waste our time searching for something identical to it there.
2102 checksum = calc_checksum(page);
2103 if (rmap_item->oldchecksum != checksum) {
2104 rmap_item->oldchecksum = checksum;
2109 * Same checksum as an empty page. We attempt to merge it with the
2110 * appropriate zero page if the user enabled this via sysfs.
2112 if (ksm_use_zero_pages && (checksum == zero_checksum)) {
2113 struct vm_area_struct *vma;
2116 vma = find_mergeable_vma(mm, rmap_item->address);
2118 err = try_to_merge_one_page(vma, page,
2119 ZERO_PAGE(rmap_item->address));
2122 * If the vma is out of date, we do not need to
2127 mmap_read_unlock(mm);
2129 * In case of failure, the page was not really empty, so we
2130 * need to continue. Otherwise we're done.
2136 unstable_tree_search_insert(rmap_item, page, &tree_page);
2137 if (tree_rmap_item) {
2140 kpage = try_to_merge_two_pages(rmap_item, page,
2141 tree_rmap_item, tree_page);
2143 * If both pages we tried to merge belong to the same compound
2144 * page, then we actually ended up increasing the reference
2145 * count of the same compound page twice, and split_huge_page
2147 * Here we set a flag if that happened, and we use it later to
2148 * try split_huge_page again. Since we call put_page right
2149 * afterwards, the reference count will be correct and
2150 * split_huge_page should succeed.
2152 split = PageTransCompound(page)
2153 && compound_head(page) == compound_head(tree_page);
2154 put_page(tree_page);
2157 * The pages were successfully merged: insert new
2158 * node in the stable tree and add both rmap_items.
2161 stable_node = stable_tree_insert(kpage);
2163 stable_tree_append(tree_rmap_item, stable_node,
2165 stable_tree_append(rmap_item, stable_node,
2171 * If we fail to insert the page into the stable tree,
2172 * we will have 2 virtual addresses that are pointing
2173 * to a ksm page left outside the stable tree,
2174 * in which case we need to break_cow on both.
2177 break_cow(tree_rmap_item);
2178 break_cow(rmap_item);
2182 * We are here if we tried to merge two pages and
2183 * failed because they both belonged to the same
2184 * compound page. We will split the page now, but no
2185 * merging will take place.
2186 * We do not want to add the cost of a full lock; if
2187 * the page is locked, it is better to skip it and
2188 * perhaps try again later.
2190 if (!trylock_page(page))
2192 split_huge_page(page);
2198 static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
2199 struct rmap_item **rmap_list,
2202 struct rmap_item *rmap_item;
2204 while (*rmap_list) {
2205 rmap_item = *rmap_list;
2206 if ((rmap_item->address & PAGE_MASK) == addr)
2208 if (rmap_item->address > addr)
2210 *rmap_list = rmap_item->rmap_list;
2211 remove_rmap_item_from_tree(rmap_item);
2212 free_rmap_item(rmap_item);
2215 rmap_item = alloc_rmap_item();
2217 /* It has already been zeroed */
2218 rmap_item->mm = mm_slot->mm;
2219 rmap_item->address = addr;
2220 rmap_item->rmap_list = *rmap_list;
2221 *rmap_list = rmap_item;
2226 static struct rmap_item *scan_get_next_rmap_item(struct page **page)
2228 struct mm_struct *mm;
2229 struct mm_slot *slot;
2230 struct vm_area_struct *vma;
2231 struct rmap_item *rmap_item;
2234 if (list_empty(&ksm_mm_head.mm_list))
2237 slot = ksm_scan.mm_slot;
2238 if (slot == &ksm_mm_head) {
2240 * A number of pages can hang around indefinitely on per-cpu
2241 * pagevecs, raised page count preventing write_protect_page
2242 * from merging them. Though it doesn't really matter much,
2243 * it is puzzling to see some stuck in pages_volatile until
2244 * other activity jostles them out, and they also prevented
2245 * LTP's KSM test from succeeding deterministically; so drain
2246 * them here (here rather than on entry to ksm_do_scan(),
2247 * so we don't IPI too often when pages_to_scan is set low).
2249 lru_add_drain_all();
2252 * Whereas stale stable_nodes on the stable_tree itself
2253 * get pruned in the regular course of stable_tree_search(),
2254 * those moved out to the migrate_nodes list can accumulate:
2255 * so prune them once before each full scan.
2257 if (!ksm_merge_across_nodes) {
2258 struct stable_node *stable_node, *next;
2261 list_for_each_entry_safe(stable_node, next,
2262 &migrate_nodes, list) {
2263 page = get_ksm_page(stable_node,
2264 GET_KSM_PAGE_NOLOCK);
2271 for (nid = 0; nid < ksm_nr_node_ids; nid++)
2272 root_unstable_tree[nid] = RB_ROOT;
2274 spin_lock(&ksm_mmlist_lock);
2275 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
2276 ksm_scan.mm_slot = slot;
2277 spin_unlock(&ksm_mmlist_lock);
2279 * Although we tested list_empty() above, a racing __ksm_exit
2280 * of the last mm on the list may have removed it since then.
2282 if (slot == &ksm_mm_head)
2285 ksm_scan.address = 0;
2286 ksm_scan.rmap_list = &slot->rmap_list;
2291 if (ksm_test_exit(mm))
2294 vma = find_vma(mm, ksm_scan.address);
2296 for (; vma; vma = vma->vm_next) {
2297 if (!(vma->vm_flags & VM_MERGEABLE))
2299 if (ksm_scan.address < vma->vm_start)
2300 ksm_scan.address = vma->vm_start;
2302 ksm_scan.address = vma->vm_end;
2304 while (ksm_scan.address < vma->vm_end) {
2305 if (ksm_test_exit(mm))
2307 *page = follow_page(vma, ksm_scan.address, FOLL_GET);
2308 if (IS_ERR_OR_NULL(*page)) {
2309 ksm_scan.address += PAGE_SIZE;
2313 if (PageAnon(*page)) {
2314 flush_anon_page(vma, *page, ksm_scan.address);
2315 flush_dcache_page(*page);
2316 rmap_item = get_next_rmap_item(slot,
2317 ksm_scan.rmap_list, ksm_scan.address);
2319 ksm_scan.rmap_list =
2320 &rmap_item->rmap_list;
2321 ksm_scan.address += PAGE_SIZE;
2324 mmap_read_unlock(mm);
2328 ksm_scan.address += PAGE_SIZE;
2333 if (ksm_test_exit(mm)) {
2334 ksm_scan.address = 0;
2335 ksm_scan.rmap_list = &slot->rmap_list;
2338 * Nuke all the rmap_items that are above this current rmap:
2339 * because there were no VM_MERGEABLE vmas with such addresses.
2341 remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
2343 spin_lock(&ksm_mmlist_lock);
2344 ksm_scan.mm_slot = list_entry(slot->mm_list.next,
2345 struct mm_slot, mm_list);
2346 if (ksm_scan.address == 0) {
2348 * We've completed a full scan of all vmas, holding mmap_lock
2349 * throughout, and found no VM_MERGEABLE: so do the same as
2350 * __ksm_exit does to remove this mm from all our lists now.
2351 * This applies either when cleaning up after __ksm_exit
2352 * (but beware: we can reach here even before __ksm_exit),
2353 * or when all VM_MERGEABLE areas have been unmapped (and
2354 * mmap_lock then protects against race with MADV_MERGEABLE).
2356 hash_del(&slot->link);
2357 list_del(&slot->mm_list);
2358 spin_unlock(&ksm_mmlist_lock);
2361 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
2362 mmap_read_unlock(mm);
2365 mmap_read_unlock(mm);
2367 * mmap_read_unlock(mm) first because after
2368 * spin_unlock(&ksm_mmlist_lock) run, the "mm" may
2369 * already have been freed under us by __ksm_exit()
2370 * because the "mm_slot" is still hashed and
2371 * ksm_scan.mm_slot doesn't point to it anymore.
2373 spin_unlock(&ksm_mmlist_lock);
2376 /* Repeat until we've completed scanning the whole list */
2377 slot = ksm_scan.mm_slot;
2378 if (slot != &ksm_mm_head)
2386 * ksm_do_scan - the ksm scanner main worker function.
2387 * @scan_npages: number of pages we want to scan before we return.
2389 static void ksm_do_scan(unsigned int scan_npages)
2391 struct rmap_item *rmap_item;
2394 while (scan_npages-- && likely(!freezing(current))) {
2396 rmap_item = scan_get_next_rmap_item(&page);
2399 cmp_and_merge_page(page, rmap_item);
2404 static int ksmd_should_run(void)
2406 return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
2409 static int ksm_scan_thread(void *nothing)
2411 unsigned int sleep_ms;
2414 set_user_nice(current, 5);
2416 while (!kthread_should_stop()) {
2417 mutex_lock(&ksm_thread_mutex);
2418 wait_while_offlining();
2419 if (ksmd_should_run())
2420 ksm_do_scan(ksm_thread_pages_to_scan);
2421 mutex_unlock(&ksm_thread_mutex);
2425 if (ksmd_should_run()) {
2426 sleep_ms = READ_ONCE(ksm_thread_sleep_millisecs);
2427 wait_event_interruptible_timeout(ksm_iter_wait,
2428 sleep_ms != READ_ONCE(ksm_thread_sleep_millisecs),
2429 msecs_to_jiffies(sleep_ms));
2431 wait_event_freezable(ksm_thread_wait,
2432 ksmd_should_run() || kthread_should_stop());
2438 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
2439 unsigned long end, int advice, unsigned long *vm_flags)
2441 struct mm_struct *mm = vma->vm_mm;
2445 case MADV_MERGEABLE:
2447 * Be somewhat over-protective for now!
2449 if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE |
2450 VM_PFNMAP | VM_IO | VM_DONTEXPAND |
2451 VM_HUGETLB | VM_MIXEDMAP))
2452 return 0; /* just ignore the advice */
2454 if (vma_is_dax(vma))
2458 if (*vm_flags & VM_SAO)
2462 if (*vm_flags & VM_SPARC_ADI)
2466 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
2467 err = __ksm_enter(mm);
2472 *vm_flags |= VM_MERGEABLE;
2475 case MADV_UNMERGEABLE:
2476 if (!(*vm_flags & VM_MERGEABLE))
2477 return 0; /* just ignore the advice */
2479 if (vma->anon_vma) {
2480 err = unmerge_ksm_pages(vma, start, end);
2485 *vm_flags &= ~VM_MERGEABLE;
2491 EXPORT_SYMBOL_GPL(ksm_madvise);
2493 int __ksm_enter(struct mm_struct *mm)
2495 struct mm_slot *mm_slot;
2498 mm_slot = alloc_mm_slot();
2502 /* Check ksm_run too? Would need tighter locking */
2503 needs_wakeup = list_empty(&ksm_mm_head.mm_list);
2505 spin_lock(&ksm_mmlist_lock);
2506 insert_to_mm_slots_hash(mm, mm_slot);
2508 * When KSM_RUN_MERGE (or KSM_RUN_STOP),
2509 * insert just behind the scanning cursor, to let the area settle
2510 * down a little; when fork is followed by immediate exec, we don't
2511 * want ksmd to waste time setting up and tearing down an rmap_list.
2513 * But when KSM_RUN_UNMERGE, it's important to insert ahead of its
2514 * scanning cursor, otherwise KSM pages in newly forked mms will be
2515 * missed: then we might as well insert at the end of the list.
2517 if (ksm_run & KSM_RUN_UNMERGE)
2518 list_add_tail(&mm_slot->mm_list, &ksm_mm_head.mm_list);
2520 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
2521 spin_unlock(&ksm_mmlist_lock);
2523 set_bit(MMF_VM_MERGEABLE, &mm->flags);
2527 wake_up_interruptible(&ksm_thread_wait);
2532 void __ksm_exit(struct mm_struct *mm)
2534 struct mm_slot *mm_slot;
2535 int easy_to_free = 0;
2538 * This process is exiting: if it's straightforward (as is the
2539 * case when ksmd was never running), free mm_slot immediately.
2540 * But if it's at the cursor or has rmap_items linked to it, use
2541 * mmap_lock to synchronize with any break_cows before pagetables
2542 * are freed, and leave the mm_slot on the list for ksmd to free.
2543 * Beware: ksm may already have noticed it exiting and freed the slot.
2546 spin_lock(&ksm_mmlist_lock);
2547 mm_slot = get_mm_slot(mm);
2548 if (mm_slot && ksm_scan.mm_slot != mm_slot) {
2549 if (!mm_slot->rmap_list) {
2550 hash_del(&mm_slot->link);
2551 list_del(&mm_slot->mm_list);
2554 list_move(&mm_slot->mm_list,
2555 &ksm_scan.mm_slot->mm_list);
2558 spin_unlock(&ksm_mmlist_lock);
2561 free_mm_slot(mm_slot);
2562 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
2564 } else if (mm_slot) {
2565 mmap_write_lock(mm);
2566 mmap_write_unlock(mm);
2570 struct page *ksm_might_need_to_copy(struct page *page,
2571 struct vm_area_struct *vma, unsigned long address)
2573 struct anon_vma *anon_vma = page_anon_vma(page);
2574 struct page *new_page;
2576 if (PageKsm(page)) {
2577 if (page_stable_node(page) &&
2578 !(ksm_run & KSM_RUN_UNMERGE))
2579 return page; /* no need to copy it */
2580 } else if (!anon_vma) {
2581 return page; /* no need to copy it */
2582 } else if (anon_vma->root == vma->anon_vma->root &&
2583 page->index == linear_page_index(vma, address)) {
2584 return page; /* still no need to copy it */
2586 if (!PageUptodate(page))
2587 return page; /* let do_swap_page report the error */
2589 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2590 if (new_page && mem_cgroup_charge(new_page, vma->vm_mm, GFP_KERNEL)) {
2595 copy_user_highpage(new_page, page, address, vma);
2597 SetPageDirty(new_page);
2598 __SetPageUptodate(new_page);
2599 __SetPageLocked(new_page);
2605 void rmap_walk_ksm(struct page *page, struct rmap_walk_control *rwc)
2607 struct stable_node *stable_node;
2608 struct rmap_item *rmap_item;
2609 int search_new_forks = 0;
2611 VM_BUG_ON_PAGE(!PageKsm(page), page);
2614 * Rely on the page lock to protect against concurrent modifications
2615 * to that page's node of the stable tree.
2617 VM_BUG_ON_PAGE(!PageLocked(page), page);
2619 stable_node = page_stable_node(page);
2623 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
2624 struct anon_vma *anon_vma = rmap_item->anon_vma;
2625 struct anon_vma_chain *vmac;
2626 struct vm_area_struct *vma;
2629 anon_vma_lock_read(anon_vma);
2630 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
2637 /* Ignore the stable/unstable/sqnr flags */
2638 addr = rmap_item->address & ~KSM_FLAG_MASK;
2640 if (addr < vma->vm_start || addr >= vma->vm_end)
2643 * Initially we examine only the vma which covers this
2644 * rmap_item; but later, if there is still work to do,
2645 * we examine covering vmas in other mms: in case they
2646 * were forked from the original since ksmd passed.
2648 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
2651 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2654 if (!rwc->rmap_one(page, vma, addr, rwc->arg)) {
2655 anon_vma_unlock_read(anon_vma);
2658 if (rwc->done && rwc->done(page)) {
2659 anon_vma_unlock_read(anon_vma);
2663 anon_vma_unlock_read(anon_vma);
2665 if (!search_new_forks++)
2669 #ifdef CONFIG_MIGRATION
2670 void ksm_migrate_page(struct page *newpage, struct page *oldpage)
2672 struct stable_node *stable_node;
2674 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
2675 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
2676 VM_BUG_ON_PAGE(newpage->mapping != oldpage->mapping, newpage);
2678 stable_node = page_stable_node(newpage);
2680 VM_BUG_ON_PAGE(stable_node->kpfn != page_to_pfn(oldpage), oldpage);
2681 stable_node->kpfn = page_to_pfn(newpage);
2683 * newpage->mapping was set in advance; now we need smp_wmb()
2684 * to make sure that the new stable_node->kpfn is visible
2685 * to get_ksm_page() before it can see that oldpage->mapping
2686 * has gone stale (or that PageSwapCache has been cleared).
2689 set_page_stable_node(oldpage, NULL);
2692 #endif /* CONFIG_MIGRATION */
2694 #ifdef CONFIG_MEMORY_HOTREMOVE
2695 static void wait_while_offlining(void)
2697 while (ksm_run & KSM_RUN_OFFLINE) {
2698 mutex_unlock(&ksm_thread_mutex);
2699 wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE),
2700 TASK_UNINTERRUPTIBLE);
2701 mutex_lock(&ksm_thread_mutex);
2705 static bool stable_node_dup_remove_range(struct stable_node *stable_node,
2706 unsigned long start_pfn,
2707 unsigned long end_pfn)
2709 if (stable_node->kpfn >= start_pfn &&
2710 stable_node->kpfn < end_pfn) {
2712 * Don't get_ksm_page, page has already gone:
2713 * which is why we keep kpfn instead of page*
2715 remove_node_from_stable_tree(stable_node);
2721 static bool stable_node_chain_remove_range(struct stable_node *stable_node,
2722 unsigned long start_pfn,
2723 unsigned long end_pfn,
2724 struct rb_root *root)
2726 struct stable_node *dup;
2727 struct hlist_node *hlist_safe;
2729 if (!is_stable_node_chain(stable_node)) {
2730 VM_BUG_ON(is_stable_node_dup(stable_node));
2731 return stable_node_dup_remove_range(stable_node, start_pfn,
2735 hlist_for_each_entry_safe(dup, hlist_safe,
2736 &stable_node->hlist, hlist_dup) {
2737 VM_BUG_ON(!is_stable_node_dup(dup));
2738 stable_node_dup_remove_range(dup, start_pfn, end_pfn);
2740 if (hlist_empty(&stable_node->hlist)) {
2741 free_stable_node_chain(stable_node, root);
2742 return true; /* notify caller that tree was rebalanced */
2747 static void ksm_check_stable_tree(unsigned long start_pfn,
2748 unsigned long end_pfn)
2750 struct stable_node *stable_node, *next;
2751 struct rb_node *node;
2754 for (nid = 0; nid < ksm_nr_node_ids; nid++) {
2755 node = rb_first(root_stable_tree + nid);
2757 stable_node = rb_entry(node, struct stable_node, node);
2758 if (stable_node_chain_remove_range(stable_node,
2762 node = rb_first(root_stable_tree + nid);
2764 node = rb_next(node);
2768 list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
2769 if (stable_node->kpfn >= start_pfn &&
2770 stable_node->kpfn < end_pfn)
2771 remove_node_from_stable_tree(stable_node);
2776 static int ksm_memory_callback(struct notifier_block *self,
2777 unsigned long action, void *arg)
2779 struct memory_notify *mn = arg;
2782 case MEM_GOING_OFFLINE:
2784 * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items()
2785 * and remove_all_stable_nodes() while memory is going offline:
2786 * it is unsafe for them to touch the stable tree at this time.
2787 * But unmerge_ksm_pages(), rmap lookups and other entry points
2788 * which do not need the ksm_thread_mutex are all safe.
2790 mutex_lock(&ksm_thread_mutex);
2791 ksm_run |= KSM_RUN_OFFLINE;
2792 mutex_unlock(&ksm_thread_mutex);
2797 * Most of the work is done by page migration; but there might
2798 * be a few stable_nodes left over, still pointing to struct
2799 * pages which have been offlined: prune those from the tree,
2800 * otherwise get_ksm_page() might later try to access a
2801 * non-existent struct page.
2803 ksm_check_stable_tree(mn->start_pfn,
2804 mn->start_pfn + mn->nr_pages);
2806 case MEM_CANCEL_OFFLINE:
2807 mutex_lock(&ksm_thread_mutex);
2808 ksm_run &= ~KSM_RUN_OFFLINE;
2809 mutex_unlock(&ksm_thread_mutex);
2811 smp_mb(); /* wake_up_bit advises this */
2812 wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE));
2818 static void wait_while_offlining(void)
2821 #endif /* CONFIG_MEMORY_HOTREMOVE */
2825 * This all compiles without CONFIG_SYSFS, but is a waste of space.
2828 #define KSM_ATTR_RO(_name) \
2829 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2830 #define KSM_ATTR(_name) \
2831 static struct kobj_attribute _name##_attr = \
2832 __ATTR(_name, 0644, _name##_show, _name##_store)
2834 static ssize_t sleep_millisecs_show(struct kobject *kobj,
2835 struct kobj_attribute *attr, char *buf)
2837 return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
2840 static ssize_t sleep_millisecs_store(struct kobject *kobj,
2841 struct kobj_attribute *attr,
2842 const char *buf, size_t count)
2844 unsigned long msecs;
2847 err = kstrtoul(buf, 10, &msecs);
2848 if (err || msecs > UINT_MAX)
2851 ksm_thread_sleep_millisecs = msecs;
2852 wake_up_interruptible(&ksm_iter_wait);
2856 KSM_ATTR(sleep_millisecs);
2858 static ssize_t pages_to_scan_show(struct kobject *kobj,
2859 struct kobj_attribute *attr, char *buf)
2861 return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
2864 static ssize_t pages_to_scan_store(struct kobject *kobj,
2865 struct kobj_attribute *attr,
2866 const char *buf, size_t count)
2869 unsigned long nr_pages;
2871 err = kstrtoul(buf, 10, &nr_pages);
2872 if (err || nr_pages > UINT_MAX)
2875 ksm_thread_pages_to_scan = nr_pages;
2879 KSM_ATTR(pages_to_scan);
2881 static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
2884 return sprintf(buf, "%lu\n", ksm_run);
2887 static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
2888 const char *buf, size_t count)
2891 unsigned long flags;
2893 err = kstrtoul(buf, 10, &flags);
2894 if (err || flags > UINT_MAX)
2896 if (flags > KSM_RUN_UNMERGE)
2900 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
2901 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
2902 * breaking COW to free the pages_shared (but leaves mm_slots
2903 * on the list for when ksmd may be set running again).
2906 mutex_lock(&ksm_thread_mutex);
2907 wait_while_offlining();
2908 if (ksm_run != flags) {
2910 if (flags & KSM_RUN_UNMERGE) {
2911 set_current_oom_origin();
2912 err = unmerge_and_remove_all_rmap_items();
2913 clear_current_oom_origin();
2915 ksm_run = KSM_RUN_STOP;
2920 mutex_unlock(&ksm_thread_mutex);
2922 if (flags & KSM_RUN_MERGE)
2923 wake_up_interruptible(&ksm_thread_wait);
2930 static ssize_t merge_across_nodes_show(struct kobject *kobj,
2931 struct kobj_attribute *attr, char *buf)
2933 return sprintf(buf, "%u\n", ksm_merge_across_nodes);
2936 static ssize_t merge_across_nodes_store(struct kobject *kobj,
2937 struct kobj_attribute *attr,
2938 const char *buf, size_t count)
2943 err = kstrtoul(buf, 10, &knob);
2949 mutex_lock(&ksm_thread_mutex);
2950 wait_while_offlining();
2951 if (ksm_merge_across_nodes != knob) {
2952 if (ksm_pages_shared || remove_all_stable_nodes())
2954 else if (root_stable_tree == one_stable_tree) {
2955 struct rb_root *buf;
2957 * This is the first time that we switch away from the
2958 * default of merging across nodes: must now allocate
2959 * a buffer to hold as many roots as may be needed.
2960 * Allocate stable and unstable together:
2961 * MAXSMP NODES_SHIFT 10 will use 16kB.
2963 buf = kcalloc(nr_node_ids + nr_node_ids, sizeof(*buf),
2965 /* Let us assume that RB_ROOT is NULL is zero */
2969 root_stable_tree = buf;
2970 root_unstable_tree = buf + nr_node_ids;
2971 /* Stable tree is empty but not the unstable */
2972 root_unstable_tree[0] = one_unstable_tree[0];
2976 ksm_merge_across_nodes = knob;
2977 ksm_nr_node_ids = knob ? 1 : nr_node_ids;
2980 mutex_unlock(&ksm_thread_mutex);
2982 return err ? err : count;
2984 KSM_ATTR(merge_across_nodes);
2987 static ssize_t use_zero_pages_show(struct kobject *kobj,
2988 struct kobj_attribute *attr, char *buf)
2990 return sprintf(buf, "%u\n", ksm_use_zero_pages);
2992 static ssize_t use_zero_pages_store(struct kobject *kobj,
2993 struct kobj_attribute *attr,
2994 const char *buf, size_t count)
2999 err = kstrtobool(buf, &value);
3003 ksm_use_zero_pages = value;
3007 KSM_ATTR(use_zero_pages);
3009 static ssize_t max_page_sharing_show(struct kobject *kobj,
3010 struct kobj_attribute *attr, char *buf)
3012 return sprintf(buf, "%u\n", ksm_max_page_sharing);
3015 static ssize_t max_page_sharing_store(struct kobject *kobj,
3016 struct kobj_attribute *attr,
3017 const char *buf, size_t count)
3022 err = kstrtoint(buf, 10, &knob);
3026 * When a KSM page is created it is shared by 2 mappings. This
3027 * being a signed comparison, it implicitly verifies it's not
3033 if (READ_ONCE(ksm_max_page_sharing) == knob)
3036 mutex_lock(&ksm_thread_mutex);
3037 wait_while_offlining();
3038 if (ksm_max_page_sharing != knob) {
3039 if (ksm_pages_shared || remove_all_stable_nodes())
3042 ksm_max_page_sharing = knob;
3044 mutex_unlock(&ksm_thread_mutex);
3046 return err ? err : count;
3048 KSM_ATTR(max_page_sharing);
3050 static ssize_t pages_shared_show(struct kobject *kobj,
3051 struct kobj_attribute *attr, char *buf)
3053 return sprintf(buf, "%lu\n", ksm_pages_shared);
3055 KSM_ATTR_RO(pages_shared);
3057 static ssize_t pages_sharing_show(struct kobject *kobj,
3058 struct kobj_attribute *attr, char *buf)
3060 return sprintf(buf, "%lu\n", ksm_pages_sharing);
3062 KSM_ATTR_RO(pages_sharing);
3064 static ssize_t pages_unshared_show(struct kobject *kobj,
3065 struct kobj_attribute *attr, char *buf)
3067 return sprintf(buf, "%lu\n", ksm_pages_unshared);
3069 KSM_ATTR_RO(pages_unshared);
3071 static ssize_t pages_volatile_show(struct kobject *kobj,
3072 struct kobj_attribute *attr, char *buf)
3074 long ksm_pages_volatile;
3076 ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
3077 - ksm_pages_sharing - ksm_pages_unshared;
3079 * It was not worth any locking to calculate that statistic,
3080 * but it might therefore sometimes be negative: conceal that.
3082 if (ksm_pages_volatile < 0)
3083 ksm_pages_volatile = 0;
3084 return sprintf(buf, "%ld\n", ksm_pages_volatile);
3086 KSM_ATTR_RO(pages_volatile);
3088 static ssize_t stable_node_dups_show(struct kobject *kobj,
3089 struct kobj_attribute *attr, char *buf)
3091 return sprintf(buf, "%lu\n", ksm_stable_node_dups);
3093 KSM_ATTR_RO(stable_node_dups);
3095 static ssize_t stable_node_chains_show(struct kobject *kobj,
3096 struct kobj_attribute *attr, char *buf)
3098 return sprintf(buf, "%lu\n", ksm_stable_node_chains);
3100 KSM_ATTR_RO(stable_node_chains);
3103 stable_node_chains_prune_millisecs_show(struct kobject *kobj,
3104 struct kobj_attribute *attr,
3107 return sprintf(buf, "%u\n", ksm_stable_node_chains_prune_millisecs);
3111 stable_node_chains_prune_millisecs_store(struct kobject *kobj,
3112 struct kobj_attribute *attr,
3113 const char *buf, size_t count)
3115 unsigned long msecs;
3118 err = kstrtoul(buf, 10, &msecs);
3119 if (err || msecs > UINT_MAX)
3122 ksm_stable_node_chains_prune_millisecs = msecs;
3126 KSM_ATTR(stable_node_chains_prune_millisecs);
3128 static ssize_t full_scans_show(struct kobject *kobj,
3129 struct kobj_attribute *attr, char *buf)
3131 return sprintf(buf, "%lu\n", ksm_scan.seqnr);
3133 KSM_ATTR_RO(full_scans);
3135 static struct attribute *ksm_attrs[] = {
3136 &sleep_millisecs_attr.attr,
3137 &pages_to_scan_attr.attr,
3139 &pages_shared_attr.attr,
3140 &pages_sharing_attr.attr,
3141 &pages_unshared_attr.attr,
3142 &pages_volatile_attr.attr,
3143 &full_scans_attr.attr,
3145 &merge_across_nodes_attr.attr,
3147 &max_page_sharing_attr.attr,
3148 &stable_node_chains_attr.attr,
3149 &stable_node_dups_attr.attr,
3150 &stable_node_chains_prune_millisecs_attr.attr,
3151 &use_zero_pages_attr.attr,
3155 static const struct attribute_group ksm_attr_group = {
3159 #endif /* CONFIG_SYSFS */
3161 static int __init ksm_init(void)
3163 struct task_struct *ksm_thread;
3166 /* The correct value depends on page size and endianness */
3167 zero_checksum = calc_checksum(ZERO_PAGE(0));
3168 /* Default to false for backwards compatibility */
3169 ksm_use_zero_pages = false;
3171 err = ksm_slab_init();
3175 ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
3176 if (IS_ERR(ksm_thread)) {
3177 pr_err("ksm: creating kthread failed\n");
3178 err = PTR_ERR(ksm_thread);
3183 err = sysfs_create_group(mm_kobj, &ksm_attr_group);
3185 pr_err("ksm: register sysfs failed\n");
3186 kthread_stop(ksm_thread);
3190 ksm_run = KSM_RUN_MERGE; /* no way for user to start it */
3192 #endif /* CONFIG_SYSFS */
3194 #ifdef CONFIG_MEMORY_HOTREMOVE
3195 /* There is no significance to this priority 100 */
3196 hotplug_memory_notifier(ksm_memory_callback, 100);
3205 subsys_initcall(ksm_init);