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>
18 #include <linux/mm_inline.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/xxhash.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>
49 #define DO_NUMA(x) do { (x); } while (0)
52 #define DO_NUMA(x) do { } while (0)
58 * A few notes about the KSM scanning process,
59 * to make it easier to understand the data structures below:
61 * In order to reduce excessive scanning, KSM sorts the memory pages by their
62 * contents into a data structure that holds pointers to the pages' locations.
64 * Since the contents of the pages may change at any moment, KSM cannot just
65 * insert the pages into a normal sorted tree and expect it to find anything.
66 * Therefore KSM uses two data structures - the stable and the unstable tree.
68 * The stable tree holds pointers to all the merged pages (ksm pages), sorted
69 * by their contents. Because each such page is write-protected, searching on
70 * this tree is fully assured to be working (except when pages are unmapped),
71 * and therefore this tree is called the stable tree.
73 * The stable tree node includes information required for reverse
74 * mapping from a KSM page to virtual addresses that map this page.
76 * In order to avoid large latencies of the rmap walks on KSM pages,
77 * KSM maintains two types of nodes in the stable tree:
79 * * the regular nodes that keep the reverse mapping structures in a
81 * * the "chains" that link nodes ("dups") that represent the same
82 * write protected memory content, but each "dup" corresponds to a
83 * different KSM page copy of that content
85 * Internally, the regular nodes, "dups" and "chains" are represented
86 * using the same struct ksm_stable_node structure.
88 * In addition to the stable tree, KSM uses a second data structure called the
89 * unstable tree: this tree holds pointers to pages which have been found to
90 * be "unchanged for a period of time". The unstable tree sorts these pages
91 * by their contents, but since they are not write-protected, KSM cannot rely
92 * upon the unstable tree to work correctly - the unstable tree is liable to
93 * be corrupted as its contents are modified, and so it is called unstable.
95 * KSM solves this problem by several techniques:
97 * 1) The unstable tree is flushed every time KSM completes scanning all
98 * memory areas, and then the tree is rebuilt again from the beginning.
99 * 2) KSM will only insert into the unstable tree, pages whose hash value
100 * has not changed since the previous scan of all memory areas.
101 * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
102 * colors of the nodes and not on their contents, assuring that even when
103 * the tree gets "corrupted" it won't get out of balance, so scanning time
104 * remains the same (also, searching and inserting nodes in an rbtree uses
105 * the same algorithm, so we have no overhead when we flush and rebuild).
106 * 4) KSM never flushes the stable tree, which means that even if it were to
107 * take 10 attempts to find a page in the unstable tree, once it is found,
108 * it is secured in the stable tree. (When we scan a new page, we first
109 * compare it against the stable tree, and then against the unstable tree.)
111 * If the merge_across_nodes tunable is unset, then KSM maintains multiple
112 * stable trees and multiple unstable trees: one of each for each NUMA node.
116 * struct ksm_mm_slot - ksm information per mm that is being scanned
117 * @slot: hash lookup from mm to mm_slot
118 * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
122 struct ksm_rmap_item *rmap_list;
126 * struct ksm_scan - cursor for scanning
127 * @mm_slot: the current mm_slot we are scanning
128 * @address: the next address inside that to be scanned
129 * @rmap_list: link to the next rmap to be scanned in the rmap_list
130 * @seqnr: count of completed full scans (needed when removing unstable node)
132 * There is only the one ksm_scan instance of this cursor structure.
135 struct ksm_mm_slot *mm_slot;
136 unsigned long address;
137 struct ksm_rmap_item **rmap_list;
142 * struct ksm_stable_node - node of the stable rbtree
143 * @node: rb node of this ksm page in the stable tree
144 * @head: (overlaying parent) &migrate_nodes indicates temporarily on that list
145 * @hlist_dup: linked into the stable_node->hlist with a stable_node chain
146 * @list: linked into migrate_nodes, pending placement in the proper node tree
147 * @hlist: hlist head of rmap_items using this ksm page
148 * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid)
149 * @chain_prune_time: time of the last full garbage collection
150 * @rmap_hlist_len: number of rmap_item entries in hlist or STABLE_NODE_CHAIN
151 * @nid: NUMA node id of stable tree in which linked (may not match kpfn)
153 struct ksm_stable_node {
155 struct rb_node node; /* when node of stable tree */
156 struct { /* when listed for migration */
157 struct list_head *head;
159 struct hlist_node hlist_dup;
160 struct list_head list;
164 struct hlist_head hlist;
167 unsigned long chain_prune_time;
170 * STABLE_NODE_CHAIN can be any negative number in
171 * rmap_hlist_len negative range, but better not -1 to be able
172 * to reliably detect underflows.
174 #define STABLE_NODE_CHAIN -1024
182 * struct ksm_rmap_item - reverse mapping item for virtual addresses
183 * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
184 * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
185 * @nid: NUMA node id of unstable tree in which linked (may not match page)
186 * @mm: the memory structure this rmap_item is pointing into
187 * @address: the virtual address this rmap_item tracks (+ flags in low bits)
188 * @oldchecksum: previous checksum of the page at that virtual address
189 * @node: rb node of this rmap_item in the unstable tree
190 * @head: pointer to stable_node heading this list in the stable tree
191 * @hlist: link into hlist of rmap_items hanging off that stable_node
193 struct ksm_rmap_item {
194 struct ksm_rmap_item *rmap_list;
196 struct anon_vma *anon_vma; /* when stable */
198 int nid; /* when node of unstable tree */
201 struct mm_struct *mm;
202 unsigned long address; /* + low bits used for flags below */
203 unsigned int oldchecksum; /* when unstable */
205 struct rb_node node; /* when node of unstable tree */
206 struct { /* when listed from stable tree */
207 struct ksm_stable_node *head;
208 struct hlist_node hlist;
213 #define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */
214 #define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */
215 #define STABLE_FLAG 0x200 /* is listed from the stable tree */
217 /* The stable and unstable tree heads */
218 static struct rb_root one_stable_tree[1] = { RB_ROOT };
219 static struct rb_root one_unstable_tree[1] = { RB_ROOT };
220 static struct rb_root *root_stable_tree = one_stable_tree;
221 static struct rb_root *root_unstable_tree = one_unstable_tree;
223 /* Recently migrated nodes of stable tree, pending proper placement */
224 static LIST_HEAD(migrate_nodes);
225 #define STABLE_NODE_DUP_HEAD ((struct list_head *)&migrate_nodes.prev)
227 #define MM_SLOTS_HASH_BITS 10
228 static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
230 static struct ksm_mm_slot ksm_mm_head = {
231 .slot.mm_node = LIST_HEAD_INIT(ksm_mm_head.slot.mm_node),
233 static struct ksm_scan ksm_scan = {
234 .mm_slot = &ksm_mm_head,
237 static struct kmem_cache *rmap_item_cache;
238 static struct kmem_cache *stable_node_cache;
239 static struct kmem_cache *mm_slot_cache;
241 /* The number of nodes in the stable tree */
242 static unsigned long ksm_pages_shared;
244 /* The number of page slots additionally sharing those nodes */
245 static unsigned long ksm_pages_sharing;
247 /* The number of nodes in the unstable tree */
248 static unsigned long ksm_pages_unshared;
250 /* The number of rmap_items in use: to calculate pages_volatile */
251 static unsigned long ksm_rmap_items;
253 /* The number of stable_node chains */
254 static unsigned long ksm_stable_node_chains;
256 /* The number of stable_node dups linked to the stable_node chains */
257 static unsigned long ksm_stable_node_dups;
259 /* Delay in pruning stale stable_node_dups in the stable_node_chains */
260 static unsigned int ksm_stable_node_chains_prune_millisecs = 2000;
262 /* Maximum number of page slots sharing a stable node */
263 static int ksm_max_page_sharing = 256;
265 /* Number of pages ksmd should scan in one batch */
266 static unsigned int ksm_thread_pages_to_scan = 100;
268 /* Milliseconds ksmd should sleep between batches */
269 static unsigned int ksm_thread_sleep_millisecs = 20;
271 /* Checksum of an empty (zeroed) page */
272 static unsigned int zero_checksum __read_mostly;
274 /* Whether to merge empty (zeroed) pages with actual zero pages */
275 static bool ksm_use_zero_pages __read_mostly;
278 /* Zeroed when merging across nodes is not allowed */
279 static unsigned int ksm_merge_across_nodes = 1;
280 static int ksm_nr_node_ids = 1;
282 #define ksm_merge_across_nodes 1U
283 #define ksm_nr_node_ids 1
286 #define KSM_RUN_STOP 0
287 #define KSM_RUN_MERGE 1
288 #define KSM_RUN_UNMERGE 2
289 #define KSM_RUN_OFFLINE 4
290 static unsigned long ksm_run = KSM_RUN_STOP;
291 static void wait_while_offlining(void);
293 static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
294 static DECLARE_WAIT_QUEUE_HEAD(ksm_iter_wait);
295 static DEFINE_MUTEX(ksm_thread_mutex);
296 static DEFINE_SPINLOCK(ksm_mmlist_lock);
298 #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create(#__struct,\
299 sizeof(struct __struct), __alignof__(struct __struct),\
302 static int __init ksm_slab_init(void)
304 rmap_item_cache = KSM_KMEM_CACHE(ksm_rmap_item, 0);
305 if (!rmap_item_cache)
308 stable_node_cache = KSM_KMEM_CACHE(ksm_stable_node, 0);
309 if (!stable_node_cache)
312 mm_slot_cache = KSM_KMEM_CACHE(ksm_mm_slot, 0);
319 kmem_cache_destroy(stable_node_cache);
321 kmem_cache_destroy(rmap_item_cache);
326 static void __init ksm_slab_free(void)
328 kmem_cache_destroy(mm_slot_cache);
329 kmem_cache_destroy(stable_node_cache);
330 kmem_cache_destroy(rmap_item_cache);
331 mm_slot_cache = NULL;
334 static __always_inline bool is_stable_node_chain(struct ksm_stable_node *chain)
336 return chain->rmap_hlist_len == STABLE_NODE_CHAIN;
339 static __always_inline bool is_stable_node_dup(struct ksm_stable_node *dup)
341 return dup->head == STABLE_NODE_DUP_HEAD;
344 static inline void stable_node_chain_add_dup(struct ksm_stable_node *dup,
345 struct ksm_stable_node *chain)
347 VM_BUG_ON(is_stable_node_dup(dup));
348 dup->head = STABLE_NODE_DUP_HEAD;
349 VM_BUG_ON(!is_stable_node_chain(chain));
350 hlist_add_head(&dup->hlist_dup, &chain->hlist);
351 ksm_stable_node_dups++;
354 static inline void __stable_node_dup_del(struct ksm_stable_node *dup)
356 VM_BUG_ON(!is_stable_node_dup(dup));
357 hlist_del(&dup->hlist_dup);
358 ksm_stable_node_dups--;
361 static inline void stable_node_dup_del(struct ksm_stable_node *dup)
363 VM_BUG_ON(is_stable_node_chain(dup));
364 if (is_stable_node_dup(dup))
365 __stable_node_dup_del(dup);
367 rb_erase(&dup->node, root_stable_tree + NUMA(dup->nid));
368 #ifdef CONFIG_DEBUG_VM
373 static inline struct ksm_rmap_item *alloc_rmap_item(void)
375 struct ksm_rmap_item *rmap_item;
377 rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL |
378 __GFP_NORETRY | __GFP_NOWARN);
384 static inline void free_rmap_item(struct ksm_rmap_item *rmap_item)
387 rmap_item->mm->ksm_rmap_items--;
388 rmap_item->mm = NULL; /* debug safety */
389 kmem_cache_free(rmap_item_cache, rmap_item);
392 static inline struct ksm_stable_node *alloc_stable_node(void)
395 * The allocation can take too long with GFP_KERNEL when memory is under
396 * pressure, which may lead to hung task warnings. Adding __GFP_HIGH
397 * grants access to memory reserves, helping to avoid this problem.
399 return kmem_cache_alloc(stable_node_cache, GFP_KERNEL | __GFP_HIGH);
402 static inline void free_stable_node(struct ksm_stable_node *stable_node)
404 VM_BUG_ON(stable_node->rmap_hlist_len &&
405 !is_stable_node_chain(stable_node));
406 kmem_cache_free(stable_node_cache, stable_node);
410 * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
411 * page tables after it has passed through ksm_exit() - which, if necessary,
412 * takes mmap_lock briefly to serialize against them. ksm_exit() does not set
413 * a special flag: they can just back out as soon as mm_users goes to zero.
414 * ksm_test_exit() is used throughout to make this test for exit: in some
415 * places for correctness, in some places just to avoid unnecessary work.
417 static inline bool ksm_test_exit(struct mm_struct *mm)
419 return atomic_read(&mm->mm_users) == 0;
423 * We use break_ksm to break COW on a ksm page: it's a stripped down
425 * if (get_user_pages(addr, 1, FOLL_WRITE, &page, NULL) == 1)
428 * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
429 * in case the application has unmapped and remapped mm,addr meanwhile.
430 * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP
431 * mmap of /dev/mem, where we would not want to touch it.
433 * FAULT_FLAG/FOLL_REMOTE are because we do this outside the context
434 * of the process that owns 'vma'. We also do not want to enforce
435 * protection keys here anyway.
437 static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
444 page = follow_page(vma, addr,
445 FOLL_GET | FOLL_MIGRATION | FOLL_REMOTE);
446 if (IS_ERR_OR_NULL(page))
449 ret = handle_mm_fault(vma, addr,
450 FAULT_FLAG_WRITE | FAULT_FLAG_REMOTE,
453 ret = VM_FAULT_WRITE;
455 } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV | VM_FAULT_OOM)));
457 * We must loop because handle_mm_fault() may back out if there's
458 * any difficulty e.g. if pte accessed bit gets updated concurrently.
460 * VM_FAULT_WRITE is what we have been hoping for: it indicates that
461 * COW has been broken, even if the vma does not permit VM_WRITE;
462 * but note that a concurrent fault might break PageKsm for us.
464 * VM_FAULT_SIGBUS could occur if we race with truncation of the
465 * backing file, which also invalidates anonymous pages: that's
466 * okay, that truncation will have unmapped the PageKsm for us.
468 * VM_FAULT_OOM: at the time of writing (late July 2009), setting
469 * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
470 * current task has TIF_MEMDIE set, and will be OOM killed on return
471 * to user; and ksmd, having no mm, would never be chosen for that.
473 * But if the mm is in a limited mem_cgroup, then the fault may fail
474 * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
475 * even ksmd can fail in this way - though it's usually breaking ksm
476 * just to undo a merge it made a moment before, so unlikely to oom.
478 * That's a pity: we might therefore have more kernel pages allocated
479 * than we're counting as nodes in the stable tree; but ksm_do_scan
480 * will retry to break_cow on each pass, so should recover the page
481 * in due course. The important thing is to not let VM_MERGEABLE
482 * be cleared while any such pages might remain in the area.
484 return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
487 static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm,
490 struct vm_area_struct *vma;
491 if (ksm_test_exit(mm))
493 vma = vma_lookup(mm, addr);
494 if (!vma || !(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
499 static void break_cow(struct ksm_rmap_item *rmap_item)
501 struct mm_struct *mm = rmap_item->mm;
502 unsigned long addr = rmap_item->address;
503 struct vm_area_struct *vma;
506 * It is not an accident that whenever we want to break COW
507 * to undo, we also need to drop a reference to the anon_vma.
509 put_anon_vma(rmap_item->anon_vma);
512 vma = find_mergeable_vma(mm, addr);
514 break_ksm(vma, addr);
515 mmap_read_unlock(mm);
518 static struct page *get_mergeable_page(struct ksm_rmap_item *rmap_item)
520 struct mm_struct *mm = rmap_item->mm;
521 unsigned long addr = rmap_item->address;
522 struct vm_area_struct *vma;
526 vma = find_mergeable_vma(mm, addr);
530 page = follow_page(vma, addr, FOLL_GET);
531 if (IS_ERR_OR_NULL(page))
533 if (is_zone_device_page(page))
535 if (PageAnon(page)) {
536 flush_anon_page(vma, page, addr);
537 flush_dcache_page(page);
544 mmap_read_unlock(mm);
549 * This helper is used for getting right index into array of tree roots.
550 * When merge_across_nodes knob is set to 1, there are only two rb-trees for
551 * stable and unstable pages from all nodes with roots in index 0. Otherwise,
552 * every node has its own stable and unstable tree.
554 static inline int get_kpfn_nid(unsigned long kpfn)
556 return ksm_merge_across_nodes ? 0 : NUMA(pfn_to_nid(kpfn));
559 static struct ksm_stable_node *alloc_stable_node_chain(struct ksm_stable_node *dup,
560 struct rb_root *root)
562 struct ksm_stable_node *chain = alloc_stable_node();
563 VM_BUG_ON(is_stable_node_chain(dup));
565 INIT_HLIST_HEAD(&chain->hlist);
566 chain->chain_prune_time = jiffies;
567 chain->rmap_hlist_len = STABLE_NODE_CHAIN;
568 #if defined (CONFIG_DEBUG_VM) && defined(CONFIG_NUMA)
569 chain->nid = NUMA_NO_NODE; /* debug */
571 ksm_stable_node_chains++;
574 * Put the stable node chain in the first dimension of
575 * the stable tree and at the same time remove the old
578 rb_replace_node(&dup->node, &chain->node, root);
581 * Move the old stable node to the second dimension
582 * queued in the hlist_dup. The invariant is that all
583 * dup stable_nodes in the chain->hlist point to pages
584 * that are write protected and have the exact same
587 stable_node_chain_add_dup(dup, chain);
592 static inline void free_stable_node_chain(struct ksm_stable_node *chain,
593 struct rb_root *root)
595 rb_erase(&chain->node, root);
596 free_stable_node(chain);
597 ksm_stable_node_chains--;
600 static void remove_node_from_stable_tree(struct ksm_stable_node *stable_node)
602 struct ksm_rmap_item *rmap_item;
604 /* check it's not STABLE_NODE_CHAIN or negative */
605 BUG_ON(stable_node->rmap_hlist_len < 0);
607 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
608 if (rmap_item->hlist.next)
613 rmap_item->mm->ksm_merging_pages--;
615 VM_BUG_ON(stable_node->rmap_hlist_len <= 0);
616 stable_node->rmap_hlist_len--;
617 put_anon_vma(rmap_item->anon_vma);
618 rmap_item->address &= PAGE_MASK;
623 * We need the second aligned pointer of the migrate_nodes
624 * list_head to stay clear from the rb_parent_color union
625 * (aligned and different than any node) and also different
626 * from &migrate_nodes. This will verify that future list.h changes
627 * don't break STABLE_NODE_DUP_HEAD. Only recent gcc can handle it.
629 BUILD_BUG_ON(STABLE_NODE_DUP_HEAD <= &migrate_nodes);
630 BUILD_BUG_ON(STABLE_NODE_DUP_HEAD >= &migrate_nodes + 1);
632 if (stable_node->head == &migrate_nodes)
633 list_del(&stable_node->list);
635 stable_node_dup_del(stable_node);
636 free_stable_node(stable_node);
639 enum get_ksm_page_flags {
646 * get_ksm_page: checks if the page indicated by the stable node
647 * is still its ksm page, despite having held no reference to it.
648 * In which case we can trust the content of the page, and it
649 * returns the gotten page; but if the page has now been zapped,
650 * remove the stale node from the stable tree and return NULL.
651 * But beware, the stable node's page might be being migrated.
653 * You would expect the stable_node to hold a reference to the ksm page.
654 * But if it increments the page's count, swapping out has to wait for
655 * ksmd to come around again before it can free the page, which may take
656 * seconds or even minutes: much too unresponsive. So instead we use a
657 * "keyhole reference": access to the ksm page from the stable node peeps
658 * out through its keyhole to see if that page still holds the right key,
659 * pointing back to this stable node. This relies on freeing a PageAnon
660 * page to reset its page->mapping to NULL, and relies on no other use of
661 * a page to put something that might look like our key in page->mapping.
662 * is on its way to being freed; but it is an anomaly to bear in mind.
664 static struct page *get_ksm_page(struct ksm_stable_node *stable_node,
665 enum get_ksm_page_flags flags)
668 void *expected_mapping;
671 expected_mapping = (void *)((unsigned long)stable_node |
674 kpfn = READ_ONCE(stable_node->kpfn); /* Address dependency. */
675 page = pfn_to_page(kpfn);
676 if (READ_ONCE(page->mapping) != expected_mapping)
680 * We cannot do anything with the page while its refcount is 0.
681 * Usually 0 means free, or tail of a higher-order page: in which
682 * case this node is no longer referenced, and should be freed;
683 * however, it might mean that the page is under page_ref_freeze().
684 * The __remove_mapping() case is easy, again the node is now stale;
685 * the same is in reuse_ksm_page() case; but if page is swapcache
686 * in folio_migrate_mapping(), it might still be our page,
687 * in which case it's essential to keep the node.
689 while (!get_page_unless_zero(page)) {
691 * Another check for page->mapping != expected_mapping would
692 * work here too. We have chosen the !PageSwapCache test to
693 * optimize the common case, when the page is or is about to
694 * be freed: PageSwapCache is cleared (under spin_lock_irq)
695 * in the ref_freeze section of __remove_mapping(); but Anon
696 * page->mapping reset to NULL later, in free_pages_prepare().
698 if (!PageSwapCache(page))
703 if (READ_ONCE(page->mapping) != expected_mapping) {
708 if (flags == GET_KSM_PAGE_TRYLOCK) {
709 if (!trylock_page(page)) {
711 return ERR_PTR(-EBUSY);
713 } else if (flags == GET_KSM_PAGE_LOCK)
716 if (flags != GET_KSM_PAGE_NOLOCK) {
717 if (READ_ONCE(page->mapping) != expected_mapping) {
727 * We come here from above when page->mapping or !PageSwapCache
728 * suggests that the node is stale; but it might be under migration.
729 * We need smp_rmb(), matching the smp_wmb() in folio_migrate_ksm(),
730 * before checking whether node->kpfn has been changed.
733 if (READ_ONCE(stable_node->kpfn) != kpfn)
735 remove_node_from_stable_tree(stable_node);
740 * Removing rmap_item from stable or unstable tree.
741 * This function will clean the information from the stable/unstable tree.
743 static void remove_rmap_item_from_tree(struct ksm_rmap_item *rmap_item)
745 if (rmap_item->address & STABLE_FLAG) {
746 struct ksm_stable_node *stable_node;
749 stable_node = rmap_item->head;
750 page = get_ksm_page(stable_node, GET_KSM_PAGE_LOCK);
754 hlist_del(&rmap_item->hlist);
758 if (!hlist_empty(&stable_node->hlist))
763 rmap_item->mm->ksm_merging_pages--;
765 VM_BUG_ON(stable_node->rmap_hlist_len <= 0);
766 stable_node->rmap_hlist_len--;
768 put_anon_vma(rmap_item->anon_vma);
769 rmap_item->head = NULL;
770 rmap_item->address &= PAGE_MASK;
772 } else if (rmap_item->address & UNSTABLE_FLAG) {
775 * Usually ksmd can and must skip the rb_erase, because
776 * root_unstable_tree was already reset to RB_ROOT.
777 * But be careful when an mm is exiting: do the rb_erase
778 * if this rmap_item was inserted by this scan, rather
779 * than left over from before.
781 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
784 rb_erase(&rmap_item->node,
785 root_unstable_tree + NUMA(rmap_item->nid));
786 ksm_pages_unshared--;
787 rmap_item->address &= PAGE_MASK;
790 cond_resched(); /* we're called from many long loops */
793 static void remove_trailing_rmap_items(struct ksm_rmap_item **rmap_list)
796 struct ksm_rmap_item *rmap_item = *rmap_list;
797 *rmap_list = rmap_item->rmap_list;
798 remove_rmap_item_from_tree(rmap_item);
799 free_rmap_item(rmap_item);
804 * Though it's very tempting to unmerge rmap_items from stable tree rather
805 * than check every pte of a given vma, the locking doesn't quite work for
806 * that - an rmap_item is assigned to the stable tree after inserting ksm
807 * page and upping mmap_lock. Nor does it fit with the way we skip dup'ing
808 * rmap_items from parent to child at fork time (so as not to waste time
809 * if exit comes before the next scan reaches it).
811 * Similarly, although we'd like to remove rmap_items (so updating counts
812 * and freeing memory) when unmerging an area, it's easier to leave that
813 * to the next pass of ksmd - consider, for example, how ksmd might be
814 * in cmp_and_merge_page on one of the rmap_items we would be removing.
816 static int unmerge_ksm_pages(struct vm_area_struct *vma,
817 unsigned long start, unsigned long end)
822 for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
823 if (ksm_test_exit(vma->vm_mm))
825 if (signal_pending(current))
828 err = break_ksm(vma, addr);
833 static inline struct ksm_stable_node *folio_stable_node(struct folio *folio)
835 return folio_test_ksm(folio) ? folio_raw_mapping(folio) : NULL;
838 static inline struct ksm_stable_node *page_stable_node(struct page *page)
840 return folio_stable_node(page_folio(page));
843 static inline void set_page_stable_node(struct page *page,
844 struct ksm_stable_node *stable_node)
846 VM_BUG_ON_PAGE(PageAnon(page) && PageAnonExclusive(page), page);
847 page->mapping = (void *)((unsigned long)stable_node | PAGE_MAPPING_KSM);
852 * Only called through the sysfs control interface:
854 static int remove_stable_node(struct ksm_stable_node *stable_node)
859 page = get_ksm_page(stable_node, GET_KSM_PAGE_LOCK);
862 * get_ksm_page did remove_node_from_stable_tree itself.
868 * Page could be still mapped if this races with __mmput() running in
869 * between ksm_exit() and exit_mmap(). Just refuse to let
870 * merge_across_nodes/max_page_sharing be switched.
873 if (!page_mapped(page)) {
875 * The stable node did not yet appear stale to get_ksm_page(),
876 * since that allows for an unmapped ksm page to be recognized
877 * right up until it is freed; but the node is safe to remove.
878 * This page might be in a pagevec waiting to be freed,
879 * or it might be PageSwapCache (perhaps under writeback),
880 * or it might have been removed from swapcache a moment ago.
882 set_page_stable_node(page, NULL);
883 remove_node_from_stable_tree(stable_node);
892 static int remove_stable_node_chain(struct ksm_stable_node *stable_node,
893 struct rb_root *root)
895 struct ksm_stable_node *dup;
896 struct hlist_node *hlist_safe;
898 if (!is_stable_node_chain(stable_node)) {
899 VM_BUG_ON(is_stable_node_dup(stable_node));
900 if (remove_stable_node(stable_node))
906 hlist_for_each_entry_safe(dup, hlist_safe,
907 &stable_node->hlist, hlist_dup) {
908 VM_BUG_ON(!is_stable_node_dup(dup));
909 if (remove_stable_node(dup))
912 BUG_ON(!hlist_empty(&stable_node->hlist));
913 free_stable_node_chain(stable_node, root);
917 static int remove_all_stable_nodes(void)
919 struct ksm_stable_node *stable_node, *next;
923 for (nid = 0; nid < ksm_nr_node_ids; nid++) {
924 while (root_stable_tree[nid].rb_node) {
925 stable_node = rb_entry(root_stable_tree[nid].rb_node,
926 struct ksm_stable_node, node);
927 if (remove_stable_node_chain(stable_node,
928 root_stable_tree + nid)) {
930 break; /* proceed to next nid */
935 list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
936 if (remove_stable_node(stable_node))
943 static int unmerge_and_remove_all_rmap_items(void)
945 struct ksm_mm_slot *mm_slot;
946 struct mm_slot *slot;
947 struct mm_struct *mm;
948 struct vm_area_struct *vma;
951 spin_lock(&ksm_mmlist_lock);
952 slot = list_entry(ksm_mm_head.slot.mm_node.next,
953 struct mm_slot, mm_node);
954 ksm_scan.mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot);
955 spin_unlock(&ksm_mmlist_lock);
957 for (mm_slot = ksm_scan.mm_slot; mm_slot != &ksm_mm_head;
958 mm_slot = ksm_scan.mm_slot) {
959 VMA_ITERATOR(vmi, mm_slot->slot.mm, 0);
961 mm = mm_slot->slot.mm;
963 for_each_vma(vmi, vma) {
964 if (ksm_test_exit(mm))
966 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
968 err = unmerge_ksm_pages(vma,
969 vma->vm_start, vma->vm_end);
974 remove_trailing_rmap_items(&mm_slot->rmap_list);
975 mmap_read_unlock(mm);
977 spin_lock(&ksm_mmlist_lock);
978 slot = list_entry(mm_slot->slot.mm_node.next,
979 struct mm_slot, mm_node);
980 ksm_scan.mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot);
981 if (ksm_test_exit(mm)) {
982 hash_del(&mm_slot->slot.hash);
983 list_del(&mm_slot->slot.mm_node);
984 spin_unlock(&ksm_mmlist_lock);
986 mm_slot_free(mm_slot_cache, mm_slot);
987 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
990 spin_unlock(&ksm_mmlist_lock);
993 /* Clean up stable nodes, but don't worry if some are still busy */
994 remove_all_stable_nodes();
999 mmap_read_unlock(mm);
1000 spin_lock(&ksm_mmlist_lock);
1001 ksm_scan.mm_slot = &ksm_mm_head;
1002 spin_unlock(&ksm_mmlist_lock);
1005 #endif /* CONFIG_SYSFS */
1007 static u32 calc_checksum(struct page *page)
1010 void *addr = kmap_atomic(page);
1011 checksum = xxhash(addr, PAGE_SIZE, 0);
1012 kunmap_atomic(addr);
1016 static int write_protect_page(struct vm_area_struct *vma, struct page *page,
1019 struct mm_struct *mm = vma->vm_mm;
1020 DEFINE_PAGE_VMA_WALK(pvmw, page, vma, 0, 0);
1023 struct mmu_notifier_range range;
1024 bool anon_exclusive;
1026 pvmw.address = page_address_in_vma(page, vma);
1027 if (pvmw.address == -EFAULT)
1030 BUG_ON(PageTransCompound(page));
1032 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
1034 pvmw.address + PAGE_SIZE);
1035 mmu_notifier_invalidate_range_start(&range);
1037 if (!page_vma_mapped_walk(&pvmw))
1039 if (WARN_ONCE(!pvmw.pte, "Unexpected PMD mapping?"))
1042 anon_exclusive = PageAnonExclusive(page);
1043 if (pte_write(*pvmw.pte) || pte_dirty(*pvmw.pte) ||
1044 (pte_protnone(*pvmw.pte) && pte_savedwrite(*pvmw.pte)) ||
1045 anon_exclusive || mm_tlb_flush_pending(mm)) {
1048 swapped = PageSwapCache(page);
1049 flush_cache_page(vma, pvmw.address, page_to_pfn(page));
1051 * Ok this is tricky, when get_user_pages_fast() run it doesn't
1052 * take any lock, therefore the check that we are going to make
1053 * with the pagecount against the mapcount is racy and
1054 * O_DIRECT can happen right after the check.
1055 * So we clear the pte and flush the tlb before the check
1056 * this assure us that no O_DIRECT can happen after the check
1057 * or in the middle of the check.
1059 * No need to notify as we are downgrading page table to read
1060 * only not changing it to point to a new page.
1062 * See Documentation/mm/mmu_notifier.rst
1064 entry = ptep_clear_flush(vma, pvmw.address, pvmw.pte);
1066 * Check that no O_DIRECT or similar I/O is in progress on the
1069 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
1070 set_pte_at(mm, pvmw.address, pvmw.pte, entry);
1074 /* See page_try_share_anon_rmap(): clear PTE first. */
1075 if (anon_exclusive && page_try_share_anon_rmap(page)) {
1076 set_pte_at(mm, pvmw.address, pvmw.pte, entry);
1080 if (pte_dirty(entry))
1081 set_page_dirty(page);
1083 if (pte_protnone(entry))
1084 entry = pte_mkclean(pte_clear_savedwrite(entry));
1086 entry = pte_mkclean(pte_wrprotect(entry));
1087 set_pte_at_notify(mm, pvmw.address, pvmw.pte, entry);
1089 *orig_pte = *pvmw.pte;
1093 page_vma_mapped_walk_done(&pvmw);
1095 mmu_notifier_invalidate_range_end(&range);
1101 * replace_page - replace page in vma by new ksm page
1102 * @vma: vma that holds the pte pointing to page
1103 * @page: the page we are replacing by kpage
1104 * @kpage: the ksm page we replace page by
1105 * @orig_pte: the original value of the pte
1107 * Returns 0 on success, -EFAULT on failure.
1109 static int replace_page(struct vm_area_struct *vma, struct page *page,
1110 struct page *kpage, pte_t orig_pte)
1112 struct mm_struct *mm = vma->vm_mm;
1113 struct folio *folio;
1121 struct mmu_notifier_range range;
1123 addr = page_address_in_vma(page, vma);
1124 if (addr == -EFAULT)
1127 pmd = mm_find_pmd(mm, addr);
1131 * Some THP functions use the sequence pmdp_huge_clear_flush(), set_pmd_at()
1132 * without holding anon_vma lock for write. So when looking for a
1133 * genuine pmde (in which to find pte), test present and !THP together.
1137 if (!pmd_present(pmde) || pmd_trans_huge(pmde))
1140 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, addr,
1142 mmu_notifier_invalidate_range_start(&range);
1144 ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
1145 if (!pte_same(*ptep, orig_pte)) {
1146 pte_unmap_unlock(ptep, ptl);
1149 VM_BUG_ON_PAGE(PageAnonExclusive(page), page);
1150 VM_BUG_ON_PAGE(PageAnon(kpage) && PageAnonExclusive(kpage), kpage);
1153 * No need to check ksm_use_zero_pages here: we can only have a
1154 * zero_page here if ksm_use_zero_pages was enabled already.
1156 if (!is_zero_pfn(page_to_pfn(kpage))) {
1158 page_add_anon_rmap(kpage, vma, addr, RMAP_NONE);
1159 newpte = mk_pte(kpage, vma->vm_page_prot);
1161 newpte = pte_mkspecial(pfn_pte(page_to_pfn(kpage),
1162 vma->vm_page_prot));
1164 * We're replacing an anonymous page with a zero page, which is
1165 * not anonymous. We need to do proper accounting otherwise we
1166 * will get wrong values in /proc, and a BUG message in dmesg
1167 * when tearing down the mm.
1169 dec_mm_counter(mm, MM_ANONPAGES);
1172 flush_cache_page(vma, addr, pte_pfn(*ptep));
1174 * No need to notify as we are replacing a read only page with another
1175 * read only page with the same content.
1177 * See Documentation/mm/mmu_notifier.rst
1179 ptep_clear_flush(vma, addr, ptep);
1180 set_pte_at_notify(mm, addr, ptep, newpte);
1182 folio = page_folio(page);
1183 page_remove_rmap(page, vma, false);
1184 if (!folio_mapped(folio))
1185 folio_free_swap(folio);
1188 pte_unmap_unlock(ptep, ptl);
1191 mmu_notifier_invalidate_range_end(&range);
1197 * try_to_merge_one_page - take two pages and merge them into one
1198 * @vma: the vma that holds the pte pointing to page
1199 * @page: the PageAnon page that we want to replace with kpage
1200 * @kpage: the PageKsm page that we want to map instead of page,
1201 * or NULL the first time when we want to use page as kpage.
1203 * This function returns 0 if the pages were merged, -EFAULT otherwise.
1205 static int try_to_merge_one_page(struct vm_area_struct *vma,
1206 struct page *page, struct page *kpage)
1208 pte_t orig_pte = __pte(0);
1211 if (page == kpage) /* ksm page forked */
1214 if (!PageAnon(page))
1218 * We need the page lock to read a stable PageSwapCache in
1219 * write_protect_page(). We use trylock_page() instead of
1220 * lock_page() because we don't want to wait here - we
1221 * prefer to continue scanning and merging different pages,
1222 * then come back to this page when it is unlocked.
1224 if (!trylock_page(page))
1227 if (PageTransCompound(page)) {
1228 if (split_huge_page(page))
1233 * If this anonymous page is mapped only here, its pte may need
1234 * to be write-protected. If it's mapped elsewhere, all of its
1235 * ptes are necessarily already write-protected. But in either
1236 * case, we need to lock and check page_count is not raised.
1238 if (write_protect_page(vma, page, &orig_pte) == 0) {
1241 * While we hold page lock, upgrade page from
1242 * PageAnon+anon_vma to PageKsm+NULL stable_node:
1243 * stable_tree_insert() will update stable_node.
1245 set_page_stable_node(page, NULL);
1246 mark_page_accessed(page);
1248 * Page reclaim just frees a clean page with no dirty
1249 * ptes: make sure that the ksm page would be swapped.
1251 if (!PageDirty(page))
1254 } else if (pages_identical(page, kpage))
1255 err = replace_page(vma, page, kpage, orig_pte);
1265 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
1266 * but no new kernel page is allocated: kpage must already be a ksm page.
1268 * This function returns 0 if the pages were merged, -EFAULT otherwise.
1270 static int try_to_merge_with_ksm_page(struct ksm_rmap_item *rmap_item,
1271 struct page *page, struct page *kpage)
1273 struct mm_struct *mm = rmap_item->mm;
1274 struct vm_area_struct *vma;
1278 vma = find_mergeable_vma(mm, rmap_item->address);
1282 err = try_to_merge_one_page(vma, page, kpage);
1286 /* Unstable nid is in union with stable anon_vma: remove first */
1287 remove_rmap_item_from_tree(rmap_item);
1289 /* Must get reference to anon_vma while still holding mmap_lock */
1290 rmap_item->anon_vma = vma->anon_vma;
1291 get_anon_vma(vma->anon_vma);
1293 mmap_read_unlock(mm);
1298 * try_to_merge_two_pages - take two identical pages and prepare them
1299 * to be merged into one page.
1301 * This function returns the kpage if we successfully merged two identical
1302 * pages into one ksm page, NULL otherwise.
1304 * Note that this function upgrades page to ksm page: if one of the pages
1305 * is already a ksm page, try_to_merge_with_ksm_page should be used.
1307 static struct page *try_to_merge_two_pages(struct ksm_rmap_item *rmap_item,
1309 struct ksm_rmap_item *tree_rmap_item,
1310 struct page *tree_page)
1314 err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
1316 err = try_to_merge_with_ksm_page(tree_rmap_item,
1319 * If that fails, we have a ksm page with only one pte
1320 * pointing to it: so break it.
1323 break_cow(rmap_item);
1325 return err ? NULL : page;
1328 static __always_inline
1329 bool __is_page_sharing_candidate(struct ksm_stable_node *stable_node, int offset)
1331 VM_BUG_ON(stable_node->rmap_hlist_len < 0);
1333 * Check that at least one mapping still exists, otherwise
1334 * there's no much point to merge and share with this
1335 * stable_node, as the underlying tree_page of the other
1336 * sharer is going to be freed soon.
1338 return stable_node->rmap_hlist_len &&
1339 stable_node->rmap_hlist_len + offset < ksm_max_page_sharing;
1342 static __always_inline
1343 bool is_page_sharing_candidate(struct ksm_stable_node *stable_node)
1345 return __is_page_sharing_candidate(stable_node, 0);
1348 static struct page *stable_node_dup(struct ksm_stable_node **_stable_node_dup,
1349 struct ksm_stable_node **_stable_node,
1350 struct rb_root *root,
1351 bool prune_stale_stable_nodes)
1353 struct ksm_stable_node *dup, *found = NULL, *stable_node = *_stable_node;
1354 struct hlist_node *hlist_safe;
1355 struct page *_tree_page, *tree_page = NULL;
1357 int found_rmap_hlist_len;
1359 if (!prune_stale_stable_nodes ||
1360 time_before(jiffies, stable_node->chain_prune_time +
1362 ksm_stable_node_chains_prune_millisecs)))
1363 prune_stale_stable_nodes = false;
1365 stable_node->chain_prune_time = jiffies;
1367 hlist_for_each_entry_safe(dup, hlist_safe,
1368 &stable_node->hlist, hlist_dup) {
1371 * We must walk all stable_node_dup to prune the stale
1372 * stable nodes during lookup.
1374 * get_ksm_page can drop the nodes from the
1375 * stable_node->hlist if they point to freed pages
1376 * (that's why we do a _safe walk). The "dup"
1377 * stable_node parameter itself will be freed from
1378 * under us if it returns NULL.
1380 _tree_page = get_ksm_page(dup, GET_KSM_PAGE_NOLOCK);
1384 if (is_page_sharing_candidate(dup)) {
1386 dup->rmap_hlist_len > found_rmap_hlist_len) {
1388 put_page(tree_page);
1390 found_rmap_hlist_len = found->rmap_hlist_len;
1391 tree_page = _tree_page;
1393 /* skip put_page for found dup */
1394 if (!prune_stale_stable_nodes)
1399 put_page(_tree_page);
1404 * nr is counting all dups in the chain only if
1405 * prune_stale_stable_nodes is true, otherwise we may
1406 * break the loop at nr == 1 even if there are
1409 if (prune_stale_stable_nodes && nr == 1) {
1411 * If there's not just one entry it would
1412 * corrupt memory, better BUG_ON. In KSM
1413 * context with no lock held it's not even
1416 BUG_ON(stable_node->hlist.first->next);
1419 * There's just one entry and it is below the
1420 * deduplication limit so drop the chain.
1422 rb_replace_node(&stable_node->node, &found->node,
1424 free_stable_node(stable_node);
1425 ksm_stable_node_chains--;
1426 ksm_stable_node_dups--;
1428 * NOTE: the caller depends on the stable_node
1429 * to be equal to stable_node_dup if the chain
1432 *_stable_node = found;
1434 * Just for robustness, as stable_node is
1435 * otherwise left as a stable pointer, the
1436 * compiler shall optimize it away at build
1440 } else if (stable_node->hlist.first != &found->hlist_dup &&
1441 __is_page_sharing_candidate(found, 1)) {
1443 * If the found stable_node dup can accept one
1444 * more future merge (in addition to the one
1445 * that is underway) and is not at the head of
1446 * the chain, put it there so next search will
1447 * be quicker in the !prune_stale_stable_nodes
1450 * NOTE: it would be inaccurate to use nr > 1
1451 * instead of checking the hlist.first pointer
1452 * directly, because in the
1453 * prune_stale_stable_nodes case "nr" isn't
1454 * the position of the found dup in the chain,
1455 * but the total number of dups in the chain.
1457 hlist_del(&found->hlist_dup);
1458 hlist_add_head(&found->hlist_dup,
1459 &stable_node->hlist);
1463 *_stable_node_dup = found;
1467 static struct ksm_stable_node *stable_node_dup_any(struct ksm_stable_node *stable_node,
1468 struct rb_root *root)
1470 if (!is_stable_node_chain(stable_node))
1472 if (hlist_empty(&stable_node->hlist)) {
1473 free_stable_node_chain(stable_node, root);
1476 return hlist_entry(stable_node->hlist.first,
1477 typeof(*stable_node), hlist_dup);
1481 * Like for get_ksm_page, this function can free the *_stable_node and
1482 * *_stable_node_dup if the returned tree_page is NULL.
1484 * It can also free and overwrite *_stable_node with the found
1485 * stable_node_dup if the chain is collapsed (in which case
1486 * *_stable_node will be equal to *_stable_node_dup like if the chain
1487 * never existed). It's up to the caller to verify tree_page is not
1488 * NULL before dereferencing *_stable_node or *_stable_node_dup.
1490 * *_stable_node_dup is really a second output parameter of this
1491 * function and will be overwritten in all cases, the caller doesn't
1492 * need to initialize it.
1494 static struct page *__stable_node_chain(struct ksm_stable_node **_stable_node_dup,
1495 struct ksm_stable_node **_stable_node,
1496 struct rb_root *root,
1497 bool prune_stale_stable_nodes)
1499 struct ksm_stable_node *stable_node = *_stable_node;
1500 if (!is_stable_node_chain(stable_node)) {
1501 if (is_page_sharing_candidate(stable_node)) {
1502 *_stable_node_dup = stable_node;
1503 return get_ksm_page(stable_node, GET_KSM_PAGE_NOLOCK);
1506 * _stable_node_dup set to NULL means the stable_node
1507 * reached the ksm_max_page_sharing limit.
1509 *_stable_node_dup = NULL;
1512 return stable_node_dup(_stable_node_dup, _stable_node, root,
1513 prune_stale_stable_nodes);
1516 static __always_inline struct page *chain_prune(struct ksm_stable_node **s_n_d,
1517 struct ksm_stable_node **s_n,
1518 struct rb_root *root)
1520 return __stable_node_chain(s_n_d, s_n, root, true);
1523 static __always_inline struct page *chain(struct ksm_stable_node **s_n_d,
1524 struct ksm_stable_node *s_n,
1525 struct rb_root *root)
1527 struct ksm_stable_node *old_stable_node = s_n;
1528 struct page *tree_page;
1530 tree_page = __stable_node_chain(s_n_d, &s_n, root, false);
1531 /* not pruning dups so s_n cannot have changed */
1532 VM_BUG_ON(s_n != old_stable_node);
1537 * stable_tree_search - search for page inside the stable tree
1539 * This function checks if there is a page inside the stable tree
1540 * with identical content to the page that we are scanning right now.
1542 * This function returns the stable tree node of identical content if found,
1545 static struct page *stable_tree_search(struct page *page)
1548 struct rb_root *root;
1549 struct rb_node **new;
1550 struct rb_node *parent;
1551 struct ksm_stable_node *stable_node, *stable_node_dup, *stable_node_any;
1552 struct ksm_stable_node *page_node;
1554 page_node = page_stable_node(page);
1555 if (page_node && page_node->head != &migrate_nodes) {
1556 /* ksm page forked */
1561 nid = get_kpfn_nid(page_to_pfn(page));
1562 root = root_stable_tree + nid;
1564 new = &root->rb_node;
1568 struct page *tree_page;
1572 stable_node = rb_entry(*new, struct ksm_stable_node, node);
1573 stable_node_any = NULL;
1574 tree_page = chain_prune(&stable_node_dup, &stable_node, root);
1576 * NOTE: stable_node may have been freed by
1577 * chain_prune() if the returned stable_node_dup is
1578 * not NULL. stable_node_dup may have been inserted in
1579 * the rbtree instead as a regular stable_node (in
1580 * order to collapse the stable_node chain if a single
1581 * stable_node dup was found in it). In such case the
1582 * stable_node is overwritten by the callee to point
1583 * to the stable_node_dup that was collapsed in the
1584 * stable rbtree and stable_node will be equal to
1585 * stable_node_dup like if the chain never existed.
1587 if (!stable_node_dup) {
1589 * Either all stable_node dups were full in
1590 * this stable_node chain, or this chain was
1591 * empty and should be rb_erased.
1593 stable_node_any = stable_node_dup_any(stable_node,
1595 if (!stable_node_any) {
1596 /* rb_erase just run */
1600 * Take any of the stable_node dups page of
1601 * this stable_node chain to let the tree walk
1602 * continue. All KSM pages belonging to the
1603 * stable_node dups in a stable_node chain
1604 * have the same content and they're
1605 * write protected at all times. Any will work
1606 * fine to continue the walk.
1608 tree_page = get_ksm_page(stable_node_any,
1609 GET_KSM_PAGE_NOLOCK);
1611 VM_BUG_ON(!stable_node_dup ^ !!stable_node_any);
1614 * If we walked over a stale stable_node,
1615 * get_ksm_page() will call rb_erase() and it
1616 * may rebalance the tree from under us. So
1617 * restart the search from scratch. Returning
1618 * NULL would be safe too, but we'd generate
1619 * false negative insertions just because some
1620 * stable_node was stale.
1625 ret = memcmp_pages(page, tree_page);
1626 put_page(tree_page);
1630 new = &parent->rb_left;
1632 new = &parent->rb_right;
1635 VM_BUG_ON(page_node->head != &migrate_nodes);
1637 * Test if the migrated page should be merged
1638 * into a stable node dup. If the mapcount is
1639 * 1 we can migrate it with another KSM page
1640 * without adding it to the chain.
1642 if (page_mapcount(page) > 1)
1646 if (!stable_node_dup) {
1648 * If the stable_node is a chain and
1649 * we got a payload match in memcmp
1650 * but we cannot merge the scanned
1651 * page in any of the existing
1652 * stable_node dups because they're
1653 * all full, we need to wait the
1654 * scanned page to find itself a match
1655 * in the unstable tree to create a
1656 * brand new KSM page to add later to
1657 * the dups of this stable_node.
1663 * Lock and unlock the stable_node's page (which
1664 * might already have been migrated) so that page
1665 * migration is sure to notice its raised count.
1666 * It would be more elegant to return stable_node
1667 * than kpage, but that involves more changes.
1669 tree_page = get_ksm_page(stable_node_dup,
1670 GET_KSM_PAGE_TRYLOCK);
1672 if (PTR_ERR(tree_page) == -EBUSY)
1673 return ERR_PTR(-EBUSY);
1675 if (unlikely(!tree_page))
1677 * The tree may have been rebalanced,
1678 * so re-evaluate parent and new.
1681 unlock_page(tree_page);
1683 if (get_kpfn_nid(stable_node_dup->kpfn) !=
1684 NUMA(stable_node_dup->nid)) {
1685 put_page(tree_page);
1695 list_del(&page_node->list);
1696 DO_NUMA(page_node->nid = nid);
1697 rb_link_node(&page_node->node, parent, new);
1698 rb_insert_color(&page_node->node, root);
1700 if (is_page_sharing_candidate(page_node)) {
1708 * If stable_node was a chain and chain_prune collapsed it,
1709 * stable_node has been updated to be the new regular
1710 * stable_node. A collapse of the chain is indistinguishable
1711 * from the case there was no chain in the stable
1712 * rbtree. Otherwise stable_node is the chain and
1713 * stable_node_dup is the dup to replace.
1715 if (stable_node_dup == stable_node) {
1716 VM_BUG_ON(is_stable_node_chain(stable_node_dup));
1717 VM_BUG_ON(is_stable_node_dup(stable_node_dup));
1718 /* there is no chain */
1720 VM_BUG_ON(page_node->head != &migrate_nodes);
1721 list_del(&page_node->list);
1722 DO_NUMA(page_node->nid = nid);
1723 rb_replace_node(&stable_node_dup->node,
1726 if (is_page_sharing_candidate(page_node))
1731 rb_erase(&stable_node_dup->node, root);
1735 VM_BUG_ON(!is_stable_node_chain(stable_node));
1736 __stable_node_dup_del(stable_node_dup);
1738 VM_BUG_ON(page_node->head != &migrate_nodes);
1739 list_del(&page_node->list);
1740 DO_NUMA(page_node->nid = nid);
1741 stable_node_chain_add_dup(page_node, stable_node);
1742 if (is_page_sharing_candidate(page_node))
1750 stable_node_dup->head = &migrate_nodes;
1751 list_add(&stable_node_dup->list, stable_node_dup->head);
1755 /* stable_node_dup could be null if it reached the limit */
1756 if (!stable_node_dup)
1757 stable_node_dup = stable_node_any;
1759 * If stable_node was a chain and chain_prune collapsed it,
1760 * stable_node has been updated to be the new regular
1761 * stable_node. A collapse of the chain is indistinguishable
1762 * from the case there was no chain in the stable
1763 * rbtree. Otherwise stable_node is the chain and
1764 * stable_node_dup is the dup to replace.
1766 if (stable_node_dup == stable_node) {
1767 VM_BUG_ON(is_stable_node_dup(stable_node_dup));
1768 /* chain is missing so create it */
1769 stable_node = alloc_stable_node_chain(stable_node_dup,
1775 * Add this stable_node dup that was
1776 * migrated to the stable_node chain
1777 * of the current nid for this page
1780 VM_BUG_ON(!is_stable_node_dup(stable_node_dup));
1781 VM_BUG_ON(page_node->head != &migrate_nodes);
1782 list_del(&page_node->list);
1783 DO_NUMA(page_node->nid = nid);
1784 stable_node_chain_add_dup(page_node, stable_node);
1789 * stable_tree_insert - insert stable tree node pointing to new ksm page
1790 * into the stable tree.
1792 * This function returns the stable tree node just allocated on success,
1795 static struct ksm_stable_node *stable_tree_insert(struct page *kpage)
1799 struct rb_root *root;
1800 struct rb_node **new;
1801 struct rb_node *parent;
1802 struct ksm_stable_node *stable_node, *stable_node_dup, *stable_node_any;
1803 bool need_chain = false;
1805 kpfn = page_to_pfn(kpage);
1806 nid = get_kpfn_nid(kpfn);
1807 root = root_stable_tree + nid;
1810 new = &root->rb_node;
1813 struct page *tree_page;
1817 stable_node = rb_entry(*new, struct ksm_stable_node, node);
1818 stable_node_any = NULL;
1819 tree_page = chain(&stable_node_dup, stable_node, root);
1820 if (!stable_node_dup) {
1822 * Either all stable_node dups were full in
1823 * this stable_node chain, or this chain was
1824 * empty and should be rb_erased.
1826 stable_node_any = stable_node_dup_any(stable_node,
1828 if (!stable_node_any) {
1829 /* rb_erase just run */
1833 * Take any of the stable_node dups page of
1834 * this stable_node chain to let the tree walk
1835 * continue. All KSM pages belonging to the
1836 * stable_node dups in a stable_node chain
1837 * have the same content and they're
1838 * write protected at all times. Any will work
1839 * fine to continue the walk.
1841 tree_page = get_ksm_page(stable_node_any,
1842 GET_KSM_PAGE_NOLOCK);
1844 VM_BUG_ON(!stable_node_dup ^ !!stable_node_any);
1847 * If we walked over a stale stable_node,
1848 * get_ksm_page() will call rb_erase() and it
1849 * may rebalance the tree from under us. So
1850 * restart the search from scratch. Returning
1851 * NULL would be safe too, but we'd generate
1852 * false negative insertions just because some
1853 * stable_node was stale.
1858 ret = memcmp_pages(kpage, tree_page);
1859 put_page(tree_page);
1863 new = &parent->rb_left;
1865 new = &parent->rb_right;
1872 stable_node_dup = alloc_stable_node();
1873 if (!stable_node_dup)
1876 INIT_HLIST_HEAD(&stable_node_dup->hlist);
1877 stable_node_dup->kpfn = kpfn;
1878 set_page_stable_node(kpage, stable_node_dup);
1879 stable_node_dup->rmap_hlist_len = 0;
1880 DO_NUMA(stable_node_dup->nid = nid);
1882 rb_link_node(&stable_node_dup->node, parent, new);
1883 rb_insert_color(&stable_node_dup->node, root);
1885 if (!is_stable_node_chain(stable_node)) {
1886 struct ksm_stable_node *orig = stable_node;
1887 /* chain is missing so create it */
1888 stable_node = alloc_stable_node_chain(orig, root);
1890 free_stable_node(stable_node_dup);
1894 stable_node_chain_add_dup(stable_node_dup, stable_node);
1897 return stable_node_dup;
1901 * unstable_tree_search_insert - search for identical page,
1902 * else insert rmap_item into the unstable tree.
1904 * This function searches for a page in the unstable tree identical to the
1905 * page currently being scanned; and if no identical page is found in the
1906 * tree, we insert rmap_item as a new object into the unstable tree.
1908 * This function returns pointer to rmap_item found to be identical
1909 * to the currently scanned page, NULL otherwise.
1911 * This function does both searching and inserting, because they share
1912 * the same walking algorithm in an rbtree.
1915 struct ksm_rmap_item *unstable_tree_search_insert(struct ksm_rmap_item *rmap_item,
1917 struct page **tree_pagep)
1919 struct rb_node **new;
1920 struct rb_root *root;
1921 struct rb_node *parent = NULL;
1924 nid = get_kpfn_nid(page_to_pfn(page));
1925 root = root_unstable_tree + nid;
1926 new = &root->rb_node;
1929 struct ksm_rmap_item *tree_rmap_item;
1930 struct page *tree_page;
1934 tree_rmap_item = rb_entry(*new, struct ksm_rmap_item, node);
1935 tree_page = get_mergeable_page(tree_rmap_item);
1940 * Don't substitute a ksm page for a forked page.
1942 if (page == tree_page) {
1943 put_page(tree_page);
1947 ret = memcmp_pages(page, tree_page);
1951 put_page(tree_page);
1952 new = &parent->rb_left;
1953 } else if (ret > 0) {
1954 put_page(tree_page);
1955 new = &parent->rb_right;
1956 } else if (!ksm_merge_across_nodes &&
1957 page_to_nid(tree_page) != nid) {
1959 * If tree_page has been migrated to another NUMA node,
1960 * it will be flushed out and put in the right unstable
1961 * tree next time: only merge with it when across_nodes.
1963 put_page(tree_page);
1966 *tree_pagep = tree_page;
1967 return tree_rmap_item;
1971 rmap_item->address |= UNSTABLE_FLAG;
1972 rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1973 DO_NUMA(rmap_item->nid = nid);
1974 rb_link_node(&rmap_item->node, parent, new);
1975 rb_insert_color(&rmap_item->node, root);
1977 ksm_pages_unshared++;
1982 * stable_tree_append - add another rmap_item to the linked list of
1983 * rmap_items hanging off a given node of the stable tree, all sharing
1984 * the same ksm page.
1986 static void stable_tree_append(struct ksm_rmap_item *rmap_item,
1987 struct ksm_stable_node *stable_node,
1988 bool max_page_sharing_bypass)
1991 * rmap won't find this mapping if we don't insert the
1992 * rmap_item in the right stable_node
1993 * duplicate. page_migration could break later if rmap breaks,
1994 * so we can as well crash here. We really need to check for
1995 * rmap_hlist_len == STABLE_NODE_CHAIN, but we can as well check
1996 * for other negative values as an underflow if detected here
1997 * for the first time (and not when decreasing rmap_hlist_len)
1998 * would be sign of memory corruption in the stable_node.
2000 BUG_ON(stable_node->rmap_hlist_len < 0);
2002 stable_node->rmap_hlist_len++;
2003 if (!max_page_sharing_bypass)
2004 /* possibly non fatal but unexpected overflow, only warn */
2005 WARN_ON_ONCE(stable_node->rmap_hlist_len >
2006 ksm_max_page_sharing);
2008 rmap_item->head = stable_node;
2009 rmap_item->address |= STABLE_FLAG;
2010 hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
2012 if (rmap_item->hlist.next)
2013 ksm_pages_sharing++;
2017 rmap_item->mm->ksm_merging_pages++;
2021 * cmp_and_merge_page - first see if page can be merged into the stable tree;
2022 * if not, compare checksum to previous and if it's the same, see if page can
2023 * be inserted into the unstable tree, or merged with a page already there and
2024 * both transferred to the stable tree.
2026 * @page: the page that we are searching identical page to.
2027 * @rmap_item: the reverse mapping into the virtual address of this page
2029 static void cmp_and_merge_page(struct page *page, struct ksm_rmap_item *rmap_item)
2031 struct mm_struct *mm = rmap_item->mm;
2032 struct ksm_rmap_item *tree_rmap_item;
2033 struct page *tree_page = NULL;
2034 struct ksm_stable_node *stable_node;
2036 unsigned int checksum;
2038 bool max_page_sharing_bypass = false;
2040 stable_node = page_stable_node(page);
2042 if (stable_node->head != &migrate_nodes &&
2043 get_kpfn_nid(READ_ONCE(stable_node->kpfn)) !=
2044 NUMA(stable_node->nid)) {
2045 stable_node_dup_del(stable_node);
2046 stable_node->head = &migrate_nodes;
2047 list_add(&stable_node->list, stable_node->head);
2049 if (stable_node->head != &migrate_nodes &&
2050 rmap_item->head == stable_node)
2053 * If it's a KSM fork, allow it to go over the sharing limit
2056 if (!is_page_sharing_candidate(stable_node))
2057 max_page_sharing_bypass = true;
2060 /* We first start with searching the page inside the stable tree */
2061 kpage = stable_tree_search(page);
2062 if (kpage == page && rmap_item->head == stable_node) {
2067 remove_rmap_item_from_tree(rmap_item);
2070 if (PTR_ERR(kpage) == -EBUSY)
2073 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
2076 * The page was successfully merged:
2077 * add its rmap_item to the stable tree.
2080 stable_tree_append(rmap_item, page_stable_node(kpage),
2081 max_page_sharing_bypass);
2089 * If the hash value of the page has changed from the last time
2090 * we calculated it, this page is changing frequently: therefore we
2091 * don't want to insert it in the unstable tree, and we don't want
2092 * to waste our time searching for something identical to it there.
2094 checksum = calc_checksum(page);
2095 if (rmap_item->oldchecksum != checksum) {
2096 rmap_item->oldchecksum = checksum;
2101 * Same checksum as an empty page. We attempt to merge it with the
2102 * appropriate zero page if the user enabled this via sysfs.
2104 if (ksm_use_zero_pages && (checksum == zero_checksum)) {
2105 struct vm_area_struct *vma;
2108 vma = find_mergeable_vma(mm, rmap_item->address);
2110 err = try_to_merge_one_page(vma, page,
2111 ZERO_PAGE(rmap_item->address));
2114 * If the vma is out of date, we do not need to
2119 mmap_read_unlock(mm);
2121 * In case of failure, the page was not really empty, so we
2122 * need to continue. Otherwise we're done.
2128 unstable_tree_search_insert(rmap_item, page, &tree_page);
2129 if (tree_rmap_item) {
2132 kpage = try_to_merge_two_pages(rmap_item, page,
2133 tree_rmap_item, tree_page);
2135 * If both pages we tried to merge belong to the same compound
2136 * page, then we actually ended up increasing the reference
2137 * count of the same compound page twice, and split_huge_page
2139 * Here we set a flag if that happened, and we use it later to
2140 * try split_huge_page again. Since we call put_page right
2141 * afterwards, the reference count will be correct and
2142 * split_huge_page should succeed.
2144 split = PageTransCompound(page)
2145 && compound_head(page) == compound_head(tree_page);
2146 put_page(tree_page);
2149 * The pages were successfully merged: insert new
2150 * node in the stable tree and add both rmap_items.
2153 stable_node = stable_tree_insert(kpage);
2155 stable_tree_append(tree_rmap_item, stable_node,
2157 stable_tree_append(rmap_item, stable_node,
2163 * If we fail to insert the page into the stable tree,
2164 * we will have 2 virtual addresses that are pointing
2165 * to a ksm page left outside the stable tree,
2166 * in which case we need to break_cow on both.
2169 break_cow(tree_rmap_item);
2170 break_cow(rmap_item);
2174 * We are here if we tried to merge two pages and
2175 * failed because they both belonged to the same
2176 * compound page. We will split the page now, but no
2177 * merging will take place.
2178 * We do not want to add the cost of a full lock; if
2179 * the page is locked, it is better to skip it and
2180 * perhaps try again later.
2182 if (!trylock_page(page))
2184 split_huge_page(page);
2190 static struct ksm_rmap_item *get_next_rmap_item(struct ksm_mm_slot *mm_slot,
2191 struct ksm_rmap_item **rmap_list,
2194 struct ksm_rmap_item *rmap_item;
2196 while (*rmap_list) {
2197 rmap_item = *rmap_list;
2198 if ((rmap_item->address & PAGE_MASK) == addr)
2200 if (rmap_item->address > addr)
2202 *rmap_list = rmap_item->rmap_list;
2203 remove_rmap_item_from_tree(rmap_item);
2204 free_rmap_item(rmap_item);
2207 rmap_item = alloc_rmap_item();
2209 /* It has already been zeroed */
2210 rmap_item->mm = mm_slot->slot.mm;
2211 rmap_item->mm->ksm_rmap_items++;
2212 rmap_item->address = addr;
2213 rmap_item->rmap_list = *rmap_list;
2214 *rmap_list = rmap_item;
2219 static struct ksm_rmap_item *scan_get_next_rmap_item(struct page **page)
2221 struct mm_struct *mm;
2222 struct ksm_mm_slot *mm_slot;
2223 struct mm_slot *slot;
2224 struct vm_area_struct *vma;
2225 struct ksm_rmap_item *rmap_item;
2226 struct vma_iterator vmi;
2229 if (list_empty(&ksm_mm_head.slot.mm_node))
2232 mm_slot = ksm_scan.mm_slot;
2233 if (mm_slot == &ksm_mm_head) {
2235 * A number of pages can hang around indefinitely on per-cpu
2236 * pagevecs, raised page count preventing write_protect_page
2237 * from merging them. Though it doesn't really matter much,
2238 * it is puzzling to see some stuck in pages_volatile until
2239 * other activity jostles them out, and they also prevented
2240 * LTP's KSM test from succeeding deterministically; so drain
2241 * them here (here rather than on entry to ksm_do_scan(),
2242 * so we don't IPI too often when pages_to_scan is set low).
2244 lru_add_drain_all();
2247 * Whereas stale stable_nodes on the stable_tree itself
2248 * get pruned in the regular course of stable_tree_search(),
2249 * those moved out to the migrate_nodes list can accumulate:
2250 * so prune them once before each full scan.
2252 if (!ksm_merge_across_nodes) {
2253 struct ksm_stable_node *stable_node, *next;
2256 list_for_each_entry_safe(stable_node, next,
2257 &migrate_nodes, list) {
2258 page = get_ksm_page(stable_node,
2259 GET_KSM_PAGE_NOLOCK);
2266 for (nid = 0; nid < ksm_nr_node_ids; nid++)
2267 root_unstable_tree[nid] = RB_ROOT;
2269 spin_lock(&ksm_mmlist_lock);
2270 slot = list_entry(mm_slot->slot.mm_node.next,
2271 struct mm_slot, mm_node);
2272 mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot);
2273 ksm_scan.mm_slot = mm_slot;
2274 spin_unlock(&ksm_mmlist_lock);
2276 * Although we tested list_empty() above, a racing __ksm_exit
2277 * of the last mm on the list may have removed it since then.
2279 if (mm_slot == &ksm_mm_head)
2282 ksm_scan.address = 0;
2283 ksm_scan.rmap_list = &mm_slot->rmap_list;
2286 slot = &mm_slot->slot;
2288 vma_iter_init(&vmi, mm, ksm_scan.address);
2291 if (ksm_test_exit(mm))
2294 for_each_vma(vmi, vma) {
2295 if (!(vma->vm_flags & VM_MERGEABLE))
2297 if (ksm_scan.address < vma->vm_start)
2298 ksm_scan.address = vma->vm_start;
2300 ksm_scan.address = vma->vm_end;
2302 while (ksm_scan.address < vma->vm_end) {
2303 if (ksm_test_exit(mm))
2305 *page = follow_page(vma, ksm_scan.address, FOLL_GET);
2306 if (IS_ERR_OR_NULL(*page)) {
2307 ksm_scan.address += PAGE_SIZE;
2311 if (is_zone_device_page(*page))
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(mm_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);
2329 ksm_scan.address += PAGE_SIZE;
2334 if (ksm_test_exit(mm)) {
2336 ksm_scan.address = 0;
2337 ksm_scan.rmap_list = &mm_slot->rmap_list;
2340 * Nuke all the rmap_items that are above this current rmap:
2341 * because there were no VM_MERGEABLE vmas with such addresses.
2343 remove_trailing_rmap_items(ksm_scan.rmap_list);
2345 spin_lock(&ksm_mmlist_lock);
2346 slot = list_entry(mm_slot->slot.mm_node.next,
2347 struct mm_slot, mm_node);
2348 ksm_scan.mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot);
2349 if (ksm_scan.address == 0) {
2351 * We've completed a full scan of all vmas, holding mmap_lock
2352 * throughout, and found no VM_MERGEABLE: so do the same as
2353 * __ksm_exit does to remove this mm from all our lists now.
2354 * This applies either when cleaning up after __ksm_exit
2355 * (but beware: we can reach here even before __ksm_exit),
2356 * or when all VM_MERGEABLE areas have been unmapped (and
2357 * mmap_lock then protects against race with MADV_MERGEABLE).
2359 hash_del(&mm_slot->slot.hash);
2360 list_del(&mm_slot->slot.mm_node);
2361 spin_unlock(&ksm_mmlist_lock);
2363 mm_slot_free(mm_slot_cache, mm_slot);
2364 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
2365 mmap_read_unlock(mm);
2368 mmap_read_unlock(mm);
2370 * mmap_read_unlock(mm) first because after
2371 * spin_unlock(&ksm_mmlist_lock) run, the "mm" may
2372 * already have been freed under us by __ksm_exit()
2373 * because the "mm_slot" is still hashed and
2374 * ksm_scan.mm_slot doesn't point to it anymore.
2376 spin_unlock(&ksm_mmlist_lock);
2379 /* Repeat until we've completed scanning the whole list */
2380 mm_slot = ksm_scan.mm_slot;
2381 if (mm_slot != &ksm_mm_head)
2389 * ksm_do_scan - the ksm scanner main worker function.
2390 * @scan_npages: number of pages we want to scan before we return.
2392 static void ksm_do_scan(unsigned int scan_npages)
2394 struct ksm_rmap_item *rmap_item;
2397 while (scan_npages-- && likely(!freezing(current))) {
2399 rmap_item = scan_get_next_rmap_item(&page);
2402 cmp_and_merge_page(page, rmap_item);
2407 static int ksmd_should_run(void)
2409 return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.slot.mm_node);
2412 static int ksm_scan_thread(void *nothing)
2414 unsigned int sleep_ms;
2417 set_user_nice(current, 5);
2419 while (!kthread_should_stop()) {
2420 mutex_lock(&ksm_thread_mutex);
2421 wait_while_offlining();
2422 if (ksmd_should_run())
2423 ksm_do_scan(ksm_thread_pages_to_scan);
2424 mutex_unlock(&ksm_thread_mutex);
2428 if (ksmd_should_run()) {
2429 sleep_ms = READ_ONCE(ksm_thread_sleep_millisecs);
2430 wait_event_interruptible_timeout(ksm_iter_wait,
2431 sleep_ms != READ_ONCE(ksm_thread_sleep_millisecs),
2432 msecs_to_jiffies(sleep_ms));
2434 wait_event_freezable(ksm_thread_wait,
2435 ksmd_should_run() || kthread_should_stop());
2441 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
2442 unsigned long end, int advice, unsigned long *vm_flags)
2444 struct mm_struct *mm = vma->vm_mm;
2448 case MADV_MERGEABLE:
2450 * Be somewhat over-protective for now!
2452 if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE |
2453 VM_PFNMAP | VM_IO | VM_DONTEXPAND |
2454 VM_HUGETLB | VM_MIXEDMAP))
2455 return 0; /* just ignore the advice */
2457 if (vma_is_dax(vma))
2461 if (*vm_flags & VM_SAO)
2465 if (*vm_flags & VM_SPARC_ADI)
2469 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
2470 err = __ksm_enter(mm);
2475 *vm_flags |= VM_MERGEABLE;
2478 case MADV_UNMERGEABLE:
2479 if (!(*vm_flags & VM_MERGEABLE))
2480 return 0; /* just ignore the advice */
2482 if (vma->anon_vma) {
2483 err = unmerge_ksm_pages(vma, start, end);
2488 *vm_flags &= ~VM_MERGEABLE;
2494 EXPORT_SYMBOL_GPL(ksm_madvise);
2496 int __ksm_enter(struct mm_struct *mm)
2498 struct ksm_mm_slot *mm_slot;
2499 struct mm_slot *slot;
2502 mm_slot = mm_slot_alloc(mm_slot_cache);
2506 slot = &mm_slot->slot;
2508 /* Check ksm_run too? Would need tighter locking */
2509 needs_wakeup = list_empty(&ksm_mm_head.slot.mm_node);
2511 spin_lock(&ksm_mmlist_lock);
2512 mm_slot_insert(mm_slots_hash, mm, slot);
2514 * When KSM_RUN_MERGE (or KSM_RUN_STOP),
2515 * insert just behind the scanning cursor, to let the area settle
2516 * down a little; when fork is followed by immediate exec, we don't
2517 * want ksmd to waste time setting up and tearing down an rmap_list.
2519 * But when KSM_RUN_UNMERGE, it's important to insert ahead of its
2520 * scanning cursor, otherwise KSM pages in newly forked mms will be
2521 * missed: then we might as well insert at the end of the list.
2523 if (ksm_run & KSM_RUN_UNMERGE)
2524 list_add_tail(&slot->mm_node, &ksm_mm_head.slot.mm_node);
2526 list_add_tail(&slot->mm_node, &ksm_scan.mm_slot->slot.mm_node);
2527 spin_unlock(&ksm_mmlist_lock);
2529 set_bit(MMF_VM_MERGEABLE, &mm->flags);
2533 wake_up_interruptible(&ksm_thread_wait);
2538 void __ksm_exit(struct mm_struct *mm)
2540 struct ksm_mm_slot *mm_slot;
2541 struct mm_slot *slot;
2542 int easy_to_free = 0;
2545 * This process is exiting: if it's straightforward (as is the
2546 * case when ksmd was never running), free mm_slot immediately.
2547 * But if it's at the cursor or has rmap_items linked to it, use
2548 * mmap_lock to synchronize with any break_cows before pagetables
2549 * are freed, and leave the mm_slot on the list for ksmd to free.
2550 * Beware: ksm may already have noticed it exiting and freed the slot.
2553 spin_lock(&ksm_mmlist_lock);
2554 slot = mm_slot_lookup(mm_slots_hash, mm);
2555 mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot);
2556 if (mm_slot && ksm_scan.mm_slot != mm_slot) {
2557 if (!mm_slot->rmap_list) {
2558 hash_del(&slot->hash);
2559 list_del(&slot->mm_node);
2562 list_move(&slot->mm_node,
2563 &ksm_scan.mm_slot->slot.mm_node);
2566 spin_unlock(&ksm_mmlist_lock);
2569 mm_slot_free(mm_slot_cache, mm_slot);
2570 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
2572 } else if (mm_slot) {
2573 mmap_write_lock(mm);
2574 mmap_write_unlock(mm);
2578 struct page *ksm_might_need_to_copy(struct page *page,
2579 struct vm_area_struct *vma, unsigned long address)
2581 struct folio *folio = page_folio(page);
2582 struct anon_vma *anon_vma = folio_anon_vma(folio);
2583 struct page *new_page;
2585 if (PageKsm(page)) {
2586 if (page_stable_node(page) &&
2587 !(ksm_run & KSM_RUN_UNMERGE))
2588 return page; /* no need to copy it */
2589 } else if (!anon_vma) {
2590 return page; /* no need to copy it */
2591 } else if (page->index == linear_page_index(vma, address) &&
2592 anon_vma->root == vma->anon_vma->root) {
2593 return page; /* still no need to copy it */
2595 if (!PageUptodate(page))
2596 return page; /* let do_swap_page report the error */
2598 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2600 mem_cgroup_charge(page_folio(new_page), vma->vm_mm, GFP_KERNEL)) {
2605 copy_user_highpage(new_page, page, address, vma);
2607 SetPageDirty(new_page);
2608 __SetPageUptodate(new_page);
2609 __SetPageLocked(new_page);
2611 count_vm_event(KSM_SWPIN_COPY);
2618 void rmap_walk_ksm(struct folio *folio, struct rmap_walk_control *rwc)
2620 struct ksm_stable_node *stable_node;
2621 struct ksm_rmap_item *rmap_item;
2622 int search_new_forks = 0;
2624 VM_BUG_ON_FOLIO(!folio_test_ksm(folio), folio);
2627 * Rely on the page lock to protect against concurrent modifications
2628 * to that page's node of the stable tree.
2630 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
2632 stable_node = folio_stable_node(folio);
2636 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
2637 struct anon_vma *anon_vma = rmap_item->anon_vma;
2638 struct anon_vma_chain *vmac;
2639 struct vm_area_struct *vma;
2642 if (!anon_vma_trylock_read(anon_vma)) {
2643 if (rwc->try_lock) {
2644 rwc->contended = true;
2647 anon_vma_lock_read(anon_vma);
2649 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
2656 /* Ignore the stable/unstable/sqnr flags */
2657 addr = rmap_item->address & PAGE_MASK;
2659 if (addr < vma->vm_start || addr >= vma->vm_end)
2662 * Initially we examine only the vma which covers this
2663 * rmap_item; but later, if there is still work to do,
2664 * we examine covering vmas in other mms: in case they
2665 * were forked from the original since ksmd passed.
2667 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
2670 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2673 if (!rwc->rmap_one(folio, vma, addr, rwc->arg)) {
2674 anon_vma_unlock_read(anon_vma);
2677 if (rwc->done && rwc->done(folio)) {
2678 anon_vma_unlock_read(anon_vma);
2682 anon_vma_unlock_read(anon_vma);
2684 if (!search_new_forks++)
2688 #ifdef CONFIG_MIGRATION
2689 void folio_migrate_ksm(struct folio *newfolio, struct folio *folio)
2691 struct ksm_stable_node *stable_node;
2693 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
2694 VM_BUG_ON_FOLIO(!folio_test_locked(newfolio), newfolio);
2695 VM_BUG_ON_FOLIO(newfolio->mapping != folio->mapping, newfolio);
2697 stable_node = folio_stable_node(folio);
2699 VM_BUG_ON_FOLIO(stable_node->kpfn != folio_pfn(folio), folio);
2700 stable_node->kpfn = folio_pfn(newfolio);
2702 * newfolio->mapping was set in advance; now we need smp_wmb()
2703 * to make sure that the new stable_node->kpfn is visible
2704 * to get_ksm_page() before it can see that folio->mapping
2705 * has gone stale (or that folio_test_swapcache has been cleared).
2708 set_page_stable_node(&folio->page, NULL);
2711 #endif /* CONFIG_MIGRATION */
2713 #ifdef CONFIG_MEMORY_HOTREMOVE
2714 static void wait_while_offlining(void)
2716 while (ksm_run & KSM_RUN_OFFLINE) {
2717 mutex_unlock(&ksm_thread_mutex);
2718 wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE),
2719 TASK_UNINTERRUPTIBLE);
2720 mutex_lock(&ksm_thread_mutex);
2724 static bool stable_node_dup_remove_range(struct ksm_stable_node *stable_node,
2725 unsigned long start_pfn,
2726 unsigned long end_pfn)
2728 if (stable_node->kpfn >= start_pfn &&
2729 stable_node->kpfn < end_pfn) {
2731 * Don't get_ksm_page, page has already gone:
2732 * which is why we keep kpfn instead of page*
2734 remove_node_from_stable_tree(stable_node);
2740 static bool stable_node_chain_remove_range(struct ksm_stable_node *stable_node,
2741 unsigned long start_pfn,
2742 unsigned long end_pfn,
2743 struct rb_root *root)
2745 struct ksm_stable_node *dup;
2746 struct hlist_node *hlist_safe;
2748 if (!is_stable_node_chain(stable_node)) {
2749 VM_BUG_ON(is_stable_node_dup(stable_node));
2750 return stable_node_dup_remove_range(stable_node, start_pfn,
2754 hlist_for_each_entry_safe(dup, hlist_safe,
2755 &stable_node->hlist, hlist_dup) {
2756 VM_BUG_ON(!is_stable_node_dup(dup));
2757 stable_node_dup_remove_range(dup, start_pfn, end_pfn);
2759 if (hlist_empty(&stable_node->hlist)) {
2760 free_stable_node_chain(stable_node, root);
2761 return true; /* notify caller that tree was rebalanced */
2766 static void ksm_check_stable_tree(unsigned long start_pfn,
2767 unsigned long end_pfn)
2769 struct ksm_stable_node *stable_node, *next;
2770 struct rb_node *node;
2773 for (nid = 0; nid < ksm_nr_node_ids; nid++) {
2774 node = rb_first(root_stable_tree + nid);
2776 stable_node = rb_entry(node, struct ksm_stable_node, node);
2777 if (stable_node_chain_remove_range(stable_node,
2781 node = rb_first(root_stable_tree + nid);
2783 node = rb_next(node);
2787 list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
2788 if (stable_node->kpfn >= start_pfn &&
2789 stable_node->kpfn < end_pfn)
2790 remove_node_from_stable_tree(stable_node);
2795 static int ksm_memory_callback(struct notifier_block *self,
2796 unsigned long action, void *arg)
2798 struct memory_notify *mn = arg;
2801 case MEM_GOING_OFFLINE:
2803 * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items()
2804 * and remove_all_stable_nodes() while memory is going offline:
2805 * it is unsafe for them to touch the stable tree at this time.
2806 * But unmerge_ksm_pages(), rmap lookups and other entry points
2807 * which do not need the ksm_thread_mutex are all safe.
2809 mutex_lock(&ksm_thread_mutex);
2810 ksm_run |= KSM_RUN_OFFLINE;
2811 mutex_unlock(&ksm_thread_mutex);
2816 * Most of the work is done by page migration; but there might
2817 * be a few stable_nodes left over, still pointing to struct
2818 * pages which have been offlined: prune those from the tree,
2819 * otherwise get_ksm_page() might later try to access a
2820 * non-existent struct page.
2822 ksm_check_stable_tree(mn->start_pfn,
2823 mn->start_pfn + mn->nr_pages);
2825 case MEM_CANCEL_OFFLINE:
2826 mutex_lock(&ksm_thread_mutex);
2827 ksm_run &= ~KSM_RUN_OFFLINE;
2828 mutex_unlock(&ksm_thread_mutex);
2830 smp_mb(); /* wake_up_bit advises this */
2831 wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE));
2837 static void wait_while_offlining(void)
2840 #endif /* CONFIG_MEMORY_HOTREMOVE */
2844 * This all compiles without CONFIG_SYSFS, but is a waste of space.
2847 #define KSM_ATTR_RO(_name) \
2848 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2849 #define KSM_ATTR(_name) \
2850 static struct kobj_attribute _name##_attr = __ATTR_RW(_name)
2852 static ssize_t sleep_millisecs_show(struct kobject *kobj,
2853 struct kobj_attribute *attr, char *buf)
2855 return sysfs_emit(buf, "%u\n", ksm_thread_sleep_millisecs);
2858 static ssize_t sleep_millisecs_store(struct kobject *kobj,
2859 struct kobj_attribute *attr,
2860 const char *buf, size_t count)
2865 err = kstrtouint(buf, 10, &msecs);
2869 ksm_thread_sleep_millisecs = msecs;
2870 wake_up_interruptible(&ksm_iter_wait);
2874 KSM_ATTR(sleep_millisecs);
2876 static ssize_t pages_to_scan_show(struct kobject *kobj,
2877 struct kobj_attribute *attr, char *buf)
2879 return sysfs_emit(buf, "%u\n", ksm_thread_pages_to_scan);
2882 static ssize_t pages_to_scan_store(struct kobject *kobj,
2883 struct kobj_attribute *attr,
2884 const char *buf, size_t count)
2886 unsigned int nr_pages;
2889 err = kstrtouint(buf, 10, &nr_pages);
2893 ksm_thread_pages_to_scan = nr_pages;
2897 KSM_ATTR(pages_to_scan);
2899 static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
2902 return sysfs_emit(buf, "%lu\n", ksm_run);
2905 static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
2906 const char *buf, size_t count)
2911 err = kstrtouint(buf, 10, &flags);
2914 if (flags > KSM_RUN_UNMERGE)
2918 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
2919 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
2920 * breaking COW to free the pages_shared (but leaves mm_slots
2921 * on the list for when ksmd may be set running again).
2924 mutex_lock(&ksm_thread_mutex);
2925 wait_while_offlining();
2926 if (ksm_run != flags) {
2928 if (flags & KSM_RUN_UNMERGE) {
2929 set_current_oom_origin();
2930 err = unmerge_and_remove_all_rmap_items();
2931 clear_current_oom_origin();
2933 ksm_run = KSM_RUN_STOP;
2938 mutex_unlock(&ksm_thread_mutex);
2940 if (flags & KSM_RUN_MERGE)
2941 wake_up_interruptible(&ksm_thread_wait);
2948 static ssize_t merge_across_nodes_show(struct kobject *kobj,
2949 struct kobj_attribute *attr, char *buf)
2951 return sysfs_emit(buf, "%u\n", ksm_merge_across_nodes);
2954 static ssize_t merge_across_nodes_store(struct kobject *kobj,
2955 struct kobj_attribute *attr,
2956 const char *buf, size_t count)
2961 err = kstrtoul(buf, 10, &knob);
2967 mutex_lock(&ksm_thread_mutex);
2968 wait_while_offlining();
2969 if (ksm_merge_across_nodes != knob) {
2970 if (ksm_pages_shared || remove_all_stable_nodes())
2972 else if (root_stable_tree == one_stable_tree) {
2973 struct rb_root *buf;
2975 * This is the first time that we switch away from the
2976 * default of merging across nodes: must now allocate
2977 * a buffer to hold as many roots as may be needed.
2978 * Allocate stable and unstable together:
2979 * MAXSMP NODES_SHIFT 10 will use 16kB.
2981 buf = kcalloc(nr_node_ids + nr_node_ids, sizeof(*buf),
2983 /* Let us assume that RB_ROOT is NULL is zero */
2987 root_stable_tree = buf;
2988 root_unstable_tree = buf + nr_node_ids;
2989 /* Stable tree is empty but not the unstable */
2990 root_unstable_tree[0] = one_unstable_tree[0];
2994 ksm_merge_across_nodes = knob;
2995 ksm_nr_node_ids = knob ? 1 : nr_node_ids;
2998 mutex_unlock(&ksm_thread_mutex);
3000 return err ? err : count;
3002 KSM_ATTR(merge_across_nodes);
3005 static ssize_t use_zero_pages_show(struct kobject *kobj,
3006 struct kobj_attribute *attr, char *buf)
3008 return sysfs_emit(buf, "%u\n", ksm_use_zero_pages);
3010 static ssize_t use_zero_pages_store(struct kobject *kobj,
3011 struct kobj_attribute *attr,
3012 const char *buf, size_t count)
3017 err = kstrtobool(buf, &value);
3021 ksm_use_zero_pages = value;
3025 KSM_ATTR(use_zero_pages);
3027 static ssize_t max_page_sharing_show(struct kobject *kobj,
3028 struct kobj_attribute *attr, char *buf)
3030 return sysfs_emit(buf, "%u\n", ksm_max_page_sharing);
3033 static ssize_t max_page_sharing_store(struct kobject *kobj,
3034 struct kobj_attribute *attr,
3035 const char *buf, size_t count)
3040 err = kstrtoint(buf, 10, &knob);
3044 * When a KSM page is created it is shared by 2 mappings. This
3045 * being a signed comparison, it implicitly verifies it's not
3051 if (READ_ONCE(ksm_max_page_sharing) == knob)
3054 mutex_lock(&ksm_thread_mutex);
3055 wait_while_offlining();
3056 if (ksm_max_page_sharing != knob) {
3057 if (ksm_pages_shared || remove_all_stable_nodes())
3060 ksm_max_page_sharing = knob;
3062 mutex_unlock(&ksm_thread_mutex);
3064 return err ? err : count;
3066 KSM_ATTR(max_page_sharing);
3068 static ssize_t pages_shared_show(struct kobject *kobj,
3069 struct kobj_attribute *attr, char *buf)
3071 return sysfs_emit(buf, "%lu\n", ksm_pages_shared);
3073 KSM_ATTR_RO(pages_shared);
3075 static ssize_t pages_sharing_show(struct kobject *kobj,
3076 struct kobj_attribute *attr, char *buf)
3078 return sysfs_emit(buf, "%lu\n", ksm_pages_sharing);
3080 KSM_ATTR_RO(pages_sharing);
3082 static ssize_t pages_unshared_show(struct kobject *kobj,
3083 struct kobj_attribute *attr, char *buf)
3085 return sysfs_emit(buf, "%lu\n", ksm_pages_unshared);
3087 KSM_ATTR_RO(pages_unshared);
3089 static ssize_t pages_volatile_show(struct kobject *kobj,
3090 struct kobj_attribute *attr, char *buf)
3092 long ksm_pages_volatile;
3094 ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
3095 - ksm_pages_sharing - ksm_pages_unshared;
3097 * It was not worth any locking to calculate that statistic,
3098 * but it might therefore sometimes be negative: conceal that.
3100 if (ksm_pages_volatile < 0)
3101 ksm_pages_volatile = 0;
3102 return sysfs_emit(buf, "%ld\n", ksm_pages_volatile);
3104 KSM_ATTR_RO(pages_volatile);
3106 static ssize_t stable_node_dups_show(struct kobject *kobj,
3107 struct kobj_attribute *attr, char *buf)
3109 return sysfs_emit(buf, "%lu\n", ksm_stable_node_dups);
3111 KSM_ATTR_RO(stable_node_dups);
3113 static ssize_t stable_node_chains_show(struct kobject *kobj,
3114 struct kobj_attribute *attr, char *buf)
3116 return sysfs_emit(buf, "%lu\n", ksm_stable_node_chains);
3118 KSM_ATTR_RO(stable_node_chains);
3121 stable_node_chains_prune_millisecs_show(struct kobject *kobj,
3122 struct kobj_attribute *attr,
3125 return sysfs_emit(buf, "%u\n", ksm_stable_node_chains_prune_millisecs);
3129 stable_node_chains_prune_millisecs_store(struct kobject *kobj,
3130 struct kobj_attribute *attr,
3131 const char *buf, size_t count)
3136 err = kstrtouint(buf, 10, &msecs);
3140 ksm_stable_node_chains_prune_millisecs = msecs;
3144 KSM_ATTR(stable_node_chains_prune_millisecs);
3146 static ssize_t full_scans_show(struct kobject *kobj,
3147 struct kobj_attribute *attr, char *buf)
3149 return sysfs_emit(buf, "%lu\n", ksm_scan.seqnr);
3151 KSM_ATTR_RO(full_scans);
3153 static struct attribute *ksm_attrs[] = {
3154 &sleep_millisecs_attr.attr,
3155 &pages_to_scan_attr.attr,
3157 &pages_shared_attr.attr,
3158 &pages_sharing_attr.attr,
3159 &pages_unshared_attr.attr,
3160 &pages_volatile_attr.attr,
3161 &full_scans_attr.attr,
3163 &merge_across_nodes_attr.attr,
3165 &max_page_sharing_attr.attr,
3166 &stable_node_chains_attr.attr,
3167 &stable_node_dups_attr.attr,
3168 &stable_node_chains_prune_millisecs_attr.attr,
3169 &use_zero_pages_attr.attr,
3173 static const struct attribute_group ksm_attr_group = {
3177 #endif /* CONFIG_SYSFS */
3179 static int __init ksm_init(void)
3181 struct task_struct *ksm_thread;
3184 /* The correct value depends on page size and endianness */
3185 zero_checksum = calc_checksum(ZERO_PAGE(0));
3186 /* Default to false for backwards compatibility */
3187 ksm_use_zero_pages = false;
3189 err = ksm_slab_init();
3193 ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
3194 if (IS_ERR(ksm_thread)) {
3195 pr_err("ksm: creating kthread failed\n");
3196 err = PTR_ERR(ksm_thread);
3201 err = sysfs_create_group(mm_kobj, &ksm_attr_group);
3203 pr_err("ksm: register sysfs failed\n");
3204 kthread_stop(ksm_thread);
3208 ksm_run = KSM_RUN_MERGE; /* no way for user to start it */
3210 #endif /* CONFIG_SYSFS */
3212 #ifdef CONFIG_MEMORY_HOTREMOVE
3213 /* There is no significance to this priority 100 */
3214 hotplug_memory_notifier(ksm_memory_callback, KSM_CALLBACK_PRI);
3223 subsys_initcall(ksm_init);