2 * Memory merging support.
4 * This code enables dynamic sharing of identical pages found in different
5 * memory areas, even if they are not shared by fork()
7 * Copyright (C) 2008-2009 Red Hat, Inc.
14 * This work is licensed under the terms of the GNU GPL, version 2.
17 #include <linux/errno.h>
20 #include <linux/mman.h>
21 #include <linux/sched.h>
22 #include <linux/sched/mm.h>
23 #include <linux/sched/coredump.h>
24 #include <linux/rwsem.h>
25 #include <linux/pagemap.h>
26 #include <linux/rmap.h>
27 #include <linux/spinlock.h>
28 #include <linux/jhash.h>
29 #include <linux/delay.h>
30 #include <linux/kthread.h>
31 #include <linux/wait.h>
32 #include <linux/slab.h>
33 #include <linux/rbtree.h>
34 #include <linux/memory.h>
35 #include <linux/mmu_notifier.h>
36 #include <linux/swap.h>
37 #include <linux/ksm.h>
38 #include <linux/hashtable.h>
39 #include <linux/freezer.h>
40 #include <linux/oom.h>
41 #include <linux/numa.h>
43 #include <asm/tlbflush.h>
48 #define DO_NUMA(x) do { (x); } while (0)
51 #define DO_NUMA(x) do { } while (0)
57 * A few notes about the KSM scanning process,
58 * to make it easier to understand the data structures below:
60 * In order to reduce excessive scanning, KSM sorts the memory pages by their
61 * contents into a data structure that holds pointers to the pages' locations.
63 * Since the contents of the pages may change at any moment, KSM cannot just
64 * insert the pages into a normal sorted tree and expect it to find anything.
65 * Therefore KSM uses two data structures - the stable and the unstable tree.
67 * The stable tree holds pointers to all the merged pages (ksm pages), sorted
68 * by their contents. Because each such page is write-protected, searching on
69 * this tree is fully assured to be working (except when pages are unmapped),
70 * and therefore this tree is called the stable tree.
72 * The stable tree node includes information required for reverse
73 * mapping from a KSM page to virtual addresses that map this page.
75 * In order to avoid large latencies of the rmap walks on KSM pages,
76 * KSM maintains two types of nodes in the stable tree:
78 * * the regular nodes that keep the reverse mapping structures in a
80 * * the "chains" that link nodes ("dups") that represent the same
81 * write protected memory content, but each "dup" corresponds to a
82 * different KSM page copy of that content
84 * Internally, the regular nodes, "dups" and "chains" are represented
85 * using the same :c:type:`struct stable_node` structure.
87 * In addition to the stable tree, KSM uses a second data structure called the
88 * unstable tree: this tree holds pointers to pages which have been found to
89 * be "unchanged for a period of time". The unstable tree sorts these pages
90 * by their contents, but since they are not write-protected, KSM cannot rely
91 * upon the unstable tree to work correctly - the unstable tree is liable to
92 * be corrupted as its contents are modified, and so it is called unstable.
94 * KSM solves this problem by several techniques:
96 * 1) The unstable tree is flushed every time KSM completes scanning all
97 * memory areas, and then the tree is rebuilt again from the beginning.
98 * 2) KSM will only insert into the unstable tree, pages whose hash value
99 * has not changed since the previous scan of all memory areas.
100 * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
101 * colors of the nodes and not on their contents, assuring that even when
102 * the tree gets "corrupted" it won't get out of balance, so scanning time
103 * remains the same (also, searching and inserting nodes in an rbtree uses
104 * the same algorithm, so we have no overhead when we flush and rebuild).
105 * 4) KSM never flushes the stable tree, which means that even if it were to
106 * take 10 attempts to find a page in the unstable tree, once it is found,
107 * it is secured in the stable tree. (When we scan a new page, we first
108 * compare it against the stable tree, and then against the unstable tree.)
110 * If the merge_across_nodes tunable is unset, then KSM maintains multiple
111 * stable trees and multiple unstable trees: one of each for each NUMA node.
115 * struct mm_slot - ksm information per mm that is being scanned
116 * @link: link to the mm_slots hash list
117 * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
118 * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
119 * @mm: the mm that this information is valid for
122 struct hlist_node link;
123 struct list_head mm_list;
124 struct rmap_item *rmap_list;
125 struct mm_struct *mm;
129 * struct ksm_scan - cursor for scanning
130 * @mm_slot: the current mm_slot we are scanning
131 * @address: the next address inside that to be scanned
132 * @rmap_list: link to the next rmap to be scanned in the rmap_list
133 * @seqnr: count of completed full scans (needed when removing unstable node)
135 * There is only the one ksm_scan instance of this cursor structure.
138 struct mm_slot *mm_slot;
139 unsigned long address;
140 struct rmap_item **rmap_list;
145 * struct stable_node - node of the stable rbtree
146 * @node: rb node of this ksm page in the stable tree
147 * @head: (overlaying parent) &migrate_nodes indicates temporarily on that list
148 * @hlist_dup: linked into the stable_node->hlist with a stable_node chain
149 * @list: linked into migrate_nodes, pending placement in the proper node tree
150 * @hlist: hlist head of rmap_items using this ksm page
151 * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid)
152 * @chain_prune_time: time of the last full garbage collection
153 * @rmap_hlist_len: number of rmap_item entries in hlist or STABLE_NODE_CHAIN
154 * @nid: NUMA node id of stable tree in which linked (may not match kpfn)
158 struct rb_node node; /* when node of stable tree */
159 struct { /* when listed for migration */
160 struct list_head *head;
162 struct hlist_node hlist_dup;
163 struct list_head list;
167 struct hlist_head hlist;
170 unsigned long chain_prune_time;
173 * STABLE_NODE_CHAIN can be any negative number in
174 * rmap_hlist_len negative range, but better not -1 to be able
175 * to reliably detect underflows.
177 #define STABLE_NODE_CHAIN -1024
185 * struct rmap_item - reverse mapping item for virtual addresses
186 * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
187 * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
188 * @nid: NUMA node id of unstable tree in which linked (may not match page)
189 * @mm: the memory structure this rmap_item is pointing into
190 * @address: the virtual address this rmap_item tracks (+ flags in low bits)
191 * @oldchecksum: previous checksum of the page at that virtual address
192 * @node: rb node of this rmap_item in the unstable tree
193 * @head: pointer to stable_node heading this list in the stable tree
194 * @hlist: link into hlist of rmap_items hanging off that stable_node
197 struct rmap_item *rmap_list;
199 struct anon_vma *anon_vma; /* when stable */
201 int nid; /* when node of unstable tree */
204 struct mm_struct *mm;
205 unsigned long address; /* + low bits used for flags below */
206 unsigned int oldchecksum; /* when unstable */
208 struct rb_node node; /* when node of unstable tree */
209 struct { /* when listed from stable tree */
210 struct stable_node *head;
211 struct hlist_node hlist;
216 #define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */
217 #define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */
218 #define STABLE_FLAG 0x200 /* is listed from the stable tree */
220 /* The stable and unstable tree heads */
221 static struct rb_root one_stable_tree[1] = { RB_ROOT };
222 static struct rb_root one_unstable_tree[1] = { RB_ROOT };
223 static struct rb_root *root_stable_tree = one_stable_tree;
224 static struct rb_root *root_unstable_tree = one_unstable_tree;
226 /* Recently migrated nodes of stable tree, pending proper placement */
227 static LIST_HEAD(migrate_nodes);
228 #define STABLE_NODE_DUP_HEAD ((struct list_head *)&migrate_nodes.prev)
230 #define MM_SLOTS_HASH_BITS 10
231 static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
233 static struct mm_slot ksm_mm_head = {
234 .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
236 static struct ksm_scan ksm_scan = {
237 .mm_slot = &ksm_mm_head,
240 static struct kmem_cache *rmap_item_cache;
241 static struct kmem_cache *stable_node_cache;
242 static struct kmem_cache *mm_slot_cache;
244 /* The number of nodes in the stable tree */
245 static unsigned long ksm_pages_shared;
247 /* The number of page slots additionally sharing those nodes */
248 static unsigned long ksm_pages_sharing;
250 /* The number of nodes in the unstable tree */
251 static unsigned long ksm_pages_unshared;
253 /* The number of rmap_items in use: to calculate pages_volatile */
254 static unsigned long ksm_rmap_items;
256 /* The number of stable_node chains */
257 static unsigned long ksm_stable_node_chains;
259 /* The number of stable_node dups linked to the stable_node chains */
260 static unsigned long ksm_stable_node_dups;
262 /* Delay in pruning stale stable_node_dups in the stable_node_chains */
263 static int ksm_stable_node_chains_prune_millisecs = 2000;
265 /* Maximum number of page slots sharing a stable node */
266 static int ksm_max_page_sharing = 256;
268 /* Number of pages ksmd should scan in one batch */
269 static unsigned int ksm_thread_pages_to_scan = 100;
271 /* Milliseconds ksmd should sleep between batches */
272 static unsigned int ksm_thread_sleep_millisecs = 20;
274 /* Checksum of an empty (zeroed) page */
275 static unsigned int zero_checksum __read_mostly;
277 /* Whether to merge empty (zeroed) pages with actual zero pages */
278 static bool ksm_use_zero_pages __read_mostly;
281 /* Zeroed when merging across nodes is not allowed */
282 static unsigned int ksm_merge_across_nodes = 1;
283 static int ksm_nr_node_ids = 1;
285 #define ksm_merge_across_nodes 1U
286 #define ksm_nr_node_ids 1
289 #define KSM_RUN_STOP 0
290 #define KSM_RUN_MERGE 1
291 #define KSM_RUN_UNMERGE 2
292 #define KSM_RUN_OFFLINE 4
293 static unsigned long ksm_run = KSM_RUN_STOP;
294 static void wait_while_offlining(void);
296 static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
297 static DEFINE_MUTEX(ksm_thread_mutex);
298 static DEFINE_SPINLOCK(ksm_mmlist_lock);
300 #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
301 sizeof(struct __struct), __alignof__(struct __struct),\
304 static int __init ksm_slab_init(void)
306 rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
307 if (!rmap_item_cache)
310 stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
311 if (!stable_node_cache)
314 mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
321 kmem_cache_destroy(stable_node_cache);
323 kmem_cache_destroy(rmap_item_cache);
328 static void __init ksm_slab_free(void)
330 kmem_cache_destroy(mm_slot_cache);
331 kmem_cache_destroy(stable_node_cache);
332 kmem_cache_destroy(rmap_item_cache);
333 mm_slot_cache = NULL;
336 static __always_inline bool is_stable_node_chain(struct stable_node *chain)
338 return chain->rmap_hlist_len == STABLE_NODE_CHAIN;
341 static __always_inline bool is_stable_node_dup(struct stable_node *dup)
343 return dup->head == STABLE_NODE_DUP_HEAD;
346 static inline void stable_node_chain_add_dup(struct stable_node *dup,
347 struct stable_node *chain)
349 VM_BUG_ON(is_stable_node_dup(dup));
350 dup->head = STABLE_NODE_DUP_HEAD;
351 VM_BUG_ON(!is_stable_node_chain(chain));
352 hlist_add_head(&dup->hlist_dup, &chain->hlist);
353 ksm_stable_node_dups++;
356 static inline void __stable_node_dup_del(struct stable_node *dup)
358 VM_BUG_ON(!is_stable_node_dup(dup));
359 hlist_del(&dup->hlist_dup);
360 ksm_stable_node_dups--;
363 static inline void stable_node_dup_del(struct stable_node *dup)
365 VM_BUG_ON(is_stable_node_chain(dup));
366 if (is_stable_node_dup(dup))
367 __stable_node_dup_del(dup);
369 rb_erase(&dup->node, root_stable_tree + NUMA(dup->nid));
370 #ifdef CONFIG_DEBUG_VM
375 static inline struct rmap_item *alloc_rmap_item(void)
377 struct rmap_item *rmap_item;
379 rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL |
380 __GFP_NORETRY | __GFP_NOWARN);
386 static inline void free_rmap_item(struct rmap_item *rmap_item)
389 rmap_item->mm = NULL; /* debug safety */
390 kmem_cache_free(rmap_item_cache, rmap_item);
393 static inline struct stable_node *alloc_stable_node(void)
396 * The allocation can take too long with GFP_KERNEL when memory is under
397 * pressure, which may lead to hung task warnings. Adding __GFP_HIGH
398 * grants access to memory reserves, helping to avoid this problem.
400 return kmem_cache_alloc(stable_node_cache, GFP_KERNEL | __GFP_HIGH);
403 static inline void free_stable_node(struct stable_node *stable_node)
405 VM_BUG_ON(stable_node->rmap_hlist_len &&
406 !is_stable_node_chain(stable_node));
407 kmem_cache_free(stable_node_cache, stable_node);
410 static inline struct mm_slot *alloc_mm_slot(void)
412 if (!mm_slot_cache) /* initialization failed */
414 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
417 static inline void free_mm_slot(struct mm_slot *mm_slot)
419 kmem_cache_free(mm_slot_cache, mm_slot);
422 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
424 struct mm_slot *slot;
426 hash_for_each_possible(mm_slots_hash, slot, link, (unsigned long)mm)
433 static void insert_to_mm_slots_hash(struct mm_struct *mm,
434 struct mm_slot *mm_slot)
437 hash_add(mm_slots_hash, &mm_slot->link, (unsigned long)mm);
441 * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
442 * page tables after it has passed through ksm_exit() - which, if necessary,
443 * takes mmap_sem briefly to serialize against them. ksm_exit() does not set
444 * a special flag: they can just back out as soon as mm_users goes to zero.
445 * ksm_test_exit() is used throughout to make this test for exit: in some
446 * places for correctness, in some places just to avoid unnecessary work.
448 static inline bool ksm_test_exit(struct mm_struct *mm)
450 return atomic_read(&mm->mm_users) == 0;
454 * We use break_ksm to break COW on a ksm page: it's a stripped down
456 * if (get_user_pages(addr, 1, 1, 1, &page, NULL) == 1)
459 * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
460 * in case the application has unmapped and remapped mm,addr meanwhile.
461 * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP
462 * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
464 * FAULT_FLAG/FOLL_REMOTE are because we do this outside the context
465 * of the process that owns 'vma'. We also do not want to enforce
466 * protection keys here anyway.
468 static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
475 page = follow_page(vma, addr,
476 FOLL_GET | FOLL_MIGRATION | FOLL_REMOTE);
477 if (IS_ERR_OR_NULL(page))
480 ret = handle_mm_fault(vma, addr,
481 FAULT_FLAG_WRITE | FAULT_FLAG_REMOTE);
483 ret = VM_FAULT_WRITE;
485 } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV | VM_FAULT_OOM)));
487 * We must loop because handle_mm_fault() may back out if there's
488 * any difficulty e.g. if pte accessed bit gets updated concurrently.
490 * VM_FAULT_WRITE is what we have been hoping for: it indicates that
491 * COW has been broken, even if the vma does not permit VM_WRITE;
492 * but note that a concurrent fault might break PageKsm for us.
494 * VM_FAULT_SIGBUS could occur if we race with truncation of the
495 * backing file, which also invalidates anonymous pages: that's
496 * okay, that truncation will have unmapped the PageKsm for us.
498 * VM_FAULT_OOM: at the time of writing (late July 2009), setting
499 * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
500 * current task has TIF_MEMDIE set, and will be OOM killed on return
501 * to user; and ksmd, having no mm, would never be chosen for that.
503 * But if the mm is in a limited mem_cgroup, then the fault may fail
504 * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
505 * even ksmd can fail in this way - though it's usually breaking ksm
506 * just to undo a merge it made a moment before, so unlikely to oom.
508 * That's a pity: we might therefore have more kernel pages allocated
509 * than we're counting as nodes in the stable tree; but ksm_do_scan
510 * will retry to break_cow on each pass, so should recover the page
511 * in due course. The important thing is to not let VM_MERGEABLE
512 * be cleared while any such pages might remain in the area.
514 return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
517 static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm,
520 struct vm_area_struct *vma;
521 if (ksm_test_exit(mm))
523 vma = find_vma(mm, addr);
524 if (!vma || vma->vm_start > addr)
526 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
531 static void break_cow(struct rmap_item *rmap_item)
533 struct mm_struct *mm = rmap_item->mm;
534 unsigned long addr = rmap_item->address;
535 struct vm_area_struct *vma;
538 * It is not an accident that whenever we want to break COW
539 * to undo, we also need to drop a reference to the anon_vma.
541 put_anon_vma(rmap_item->anon_vma);
543 down_read(&mm->mmap_sem);
544 vma = find_mergeable_vma(mm, addr);
546 break_ksm(vma, addr);
547 up_read(&mm->mmap_sem);
550 static struct page *get_mergeable_page(struct rmap_item *rmap_item)
552 struct mm_struct *mm = rmap_item->mm;
553 unsigned long addr = rmap_item->address;
554 struct vm_area_struct *vma;
557 down_read(&mm->mmap_sem);
558 vma = find_mergeable_vma(mm, addr);
562 page = follow_page(vma, addr, FOLL_GET);
563 if (IS_ERR_OR_NULL(page))
565 if (PageAnon(page)) {
566 flush_anon_page(vma, page, addr);
567 flush_dcache_page(page);
573 up_read(&mm->mmap_sem);
578 * This helper is used for getting right index into array of tree roots.
579 * When merge_across_nodes knob is set to 1, there are only two rb-trees for
580 * stable and unstable pages from all nodes with roots in index 0. Otherwise,
581 * every node has its own stable and unstable tree.
583 static inline int get_kpfn_nid(unsigned long kpfn)
585 return ksm_merge_across_nodes ? 0 : NUMA(pfn_to_nid(kpfn));
588 static struct stable_node *alloc_stable_node_chain(struct stable_node *dup,
589 struct rb_root *root)
591 struct stable_node *chain = alloc_stable_node();
592 VM_BUG_ON(is_stable_node_chain(dup));
594 INIT_HLIST_HEAD(&chain->hlist);
595 chain->chain_prune_time = jiffies;
596 chain->rmap_hlist_len = STABLE_NODE_CHAIN;
597 #if defined (CONFIG_DEBUG_VM) && defined(CONFIG_NUMA)
598 chain->nid = -1; /* debug */
600 ksm_stable_node_chains++;
603 * Put the stable node chain in the first dimension of
604 * the stable tree and at the same time remove the old
607 rb_replace_node(&dup->node, &chain->node, root);
610 * Move the old stable node to the second dimension
611 * queued in the hlist_dup. The invariant is that all
612 * dup stable_nodes in the chain->hlist point to pages
613 * that are wrprotected and have the exact same
616 stable_node_chain_add_dup(dup, chain);
621 static inline void free_stable_node_chain(struct stable_node *chain,
622 struct rb_root *root)
624 rb_erase(&chain->node, root);
625 free_stable_node(chain);
626 ksm_stable_node_chains--;
629 static void remove_node_from_stable_tree(struct stable_node *stable_node)
631 struct rmap_item *rmap_item;
633 /* check it's not STABLE_NODE_CHAIN or negative */
634 BUG_ON(stable_node->rmap_hlist_len < 0);
636 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
637 if (rmap_item->hlist.next)
641 VM_BUG_ON(stable_node->rmap_hlist_len <= 0);
642 stable_node->rmap_hlist_len--;
643 put_anon_vma(rmap_item->anon_vma);
644 rmap_item->address &= PAGE_MASK;
649 * We need the second aligned pointer of the migrate_nodes
650 * list_head to stay clear from the rb_parent_color union
651 * (aligned and different than any node) and also different
652 * from &migrate_nodes. This will verify that future list.h changes
653 * don't break STABLE_NODE_DUP_HEAD.
655 #if GCC_VERSION >= 40903 /* only recent gcc can handle it */
656 BUILD_BUG_ON(STABLE_NODE_DUP_HEAD <= &migrate_nodes);
657 BUILD_BUG_ON(STABLE_NODE_DUP_HEAD >= &migrate_nodes + 1);
660 if (stable_node->head == &migrate_nodes)
661 list_del(&stable_node->list);
663 stable_node_dup_del(stable_node);
664 free_stable_node(stable_node);
668 * get_ksm_page: checks if the page indicated by the stable node
669 * is still its ksm page, despite having held no reference to it.
670 * In which case we can trust the content of the page, and it
671 * returns the gotten page; but if the page has now been zapped,
672 * remove the stale node from the stable tree and return NULL.
673 * But beware, the stable node's page might be being migrated.
675 * You would expect the stable_node to hold a reference to the ksm page.
676 * But if it increments the page's count, swapping out has to wait for
677 * ksmd to come around again before it can free the page, which may take
678 * seconds or even minutes: much too unresponsive. So instead we use a
679 * "keyhole reference": access to the ksm page from the stable node peeps
680 * out through its keyhole to see if that page still holds the right key,
681 * pointing back to this stable node. This relies on freeing a PageAnon
682 * page to reset its page->mapping to NULL, and relies on no other use of
683 * a page to put something that might look like our key in page->mapping.
684 * is on its way to being freed; but it is an anomaly to bear in mind.
686 static struct page *get_ksm_page(struct stable_node *stable_node, bool lock_it)
689 void *expected_mapping;
692 expected_mapping = (void *)((unsigned long)stable_node |
695 kpfn = READ_ONCE(stable_node->kpfn); /* Address dependency. */
696 page = pfn_to_page(kpfn);
697 if (READ_ONCE(page->mapping) != expected_mapping)
701 * We cannot do anything with the page while its refcount is 0.
702 * Usually 0 means free, or tail of a higher-order page: in which
703 * case this node is no longer referenced, and should be freed;
704 * however, it might mean that the page is under page_freeze_refs().
705 * The __remove_mapping() case is easy, again the node is now stale;
706 * but if page is swapcache in migrate_page_move_mapping(), it might
707 * still be our page, in which case it's essential to keep the node.
709 while (!get_page_unless_zero(page)) {
711 * Another check for page->mapping != expected_mapping would
712 * work here too. We have chosen the !PageSwapCache test to
713 * optimize the common case, when the page is or is about to
714 * be freed: PageSwapCache is cleared (under spin_lock_irq)
715 * in the freeze_refs section of __remove_mapping(); but Anon
716 * page->mapping reset to NULL later, in free_pages_prepare().
718 if (!PageSwapCache(page))
723 if (READ_ONCE(page->mapping) != expected_mapping) {
730 if (READ_ONCE(page->mapping) != expected_mapping) {
740 * We come here from above when page->mapping or !PageSwapCache
741 * suggests that the node is stale; but it might be under migration.
742 * We need smp_rmb(), matching the smp_wmb() in ksm_migrate_page(),
743 * before checking whether node->kpfn has been changed.
746 if (READ_ONCE(stable_node->kpfn) != kpfn)
748 remove_node_from_stable_tree(stable_node);
753 * Removing rmap_item from stable or unstable tree.
754 * This function will clean the information from the stable/unstable tree.
756 static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
758 if (rmap_item->address & STABLE_FLAG) {
759 struct stable_node *stable_node;
762 stable_node = rmap_item->head;
763 page = get_ksm_page(stable_node, true);
767 hlist_del(&rmap_item->hlist);
771 if (!hlist_empty(&stable_node->hlist))
775 VM_BUG_ON(stable_node->rmap_hlist_len <= 0);
776 stable_node->rmap_hlist_len--;
778 put_anon_vma(rmap_item->anon_vma);
779 rmap_item->address &= PAGE_MASK;
781 } else if (rmap_item->address & UNSTABLE_FLAG) {
784 * Usually ksmd can and must skip the rb_erase, because
785 * root_unstable_tree was already reset to RB_ROOT.
786 * But be careful when an mm is exiting: do the rb_erase
787 * if this rmap_item was inserted by this scan, rather
788 * than left over from before.
790 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
793 rb_erase(&rmap_item->node,
794 root_unstable_tree + NUMA(rmap_item->nid));
795 ksm_pages_unshared--;
796 rmap_item->address &= PAGE_MASK;
799 cond_resched(); /* we're called from many long loops */
802 static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
803 struct rmap_item **rmap_list)
806 struct rmap_item *rmap_item = *rmap_list;
807 *rmap_list = rmap_item->rmap_list;
808 remove_rmap_item_from_tree(rmap_item);
809 free_rmap_item(rmap_item);
814 * Though it's very tempting to unmerge rmap_items from stable tree rather
815 * than check every pte of a given vma, the locking doesn't quite work for
816 * that - an rmap_item is assigned to the stable tree after inserting ksm
817 * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing
818 * rmap_items from parent to child at fork time (so as not to waste time
819 * if exit comes before the next scan reaches it).
821 * Similarly, although we'd like to remove rmap_items (so updating counts
822 * and freeing memory) when unmerging an area, it's easier to leave that
823 * to the next pass of ksmd - consider, for example, how ksmd might be
824 * in cmp_and_merge_page on one of the rmap_items we would be removing.
826 static int unmerge_ksm_pages(struct vm_area_struct *vma,
827 unsigned long start, unsigned long end)
832 for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
833 if (ksm_test_exit(vma->vm_mm))
835 if (signal_pending(current))
838 err = break_ksm(vma, addr);
843 static inline struct stable_node *page_stable_node(struct page *page)
845 return PageKsm(page) ? page_rmapping(page) : NULL;
848 static inline void set_page_stable_node(struct page *page,
849 struct stable_node *stable_node)
851 page->mapping = (void *)((unsigned long)stable_node | PAGE_MAPPING_KSM);
856 * Only called through the sysfs control interface:
858 static int remove_stable_node(struct stable_node *stable_node)
863 page = get_ksm_page(stable_node, true);
866 * get_ksm_page did remove_node_from_stable_tree itself.
871 if (WARN_ON_ONCE(page_mapped(page))) {
873 * This should not happen: but if it does, just refuse to let
874 * merge_across_nodes be switched - there is no need to panic.
879 * The stable node did not yet appear stale to get_ksm_page(),
880 * since that allows for an unmapped ksm page to be recognized
881 * right up until it is freed; but the node is safe to remove.
882 * This page might be in a pagevec waiting to be freed,
883 * or it might be PageSwapCache (perhaps under writeback),
884 * or it might have been removed from swapcache a moment ago.
886 set_page_stable_node(page, NULL);
887 remove_node_from_stable_tree(stable_node);
896 static int remove_stable_node_chain(struct stable_node *stable_node,
897 struct rb_root *root)
899 struct stable_node *dup;
900 struct hlist_node *hlist_safe;
902 if (!is_stable_node_chain(stable_node)) {
903 VM_BUG_ON(is_stable_node_dup(stable_node));
904 if (remove_stable_node(stable_node))
910 hlist_for_each_entry_safe(dup, hlist_safe,
911 &stable_node->hlist, hlist_dup) {
912 VM_BUG_ON(!is_stable_node_dup(dup));
913 if (remove_stable_node(dup))
916 BUG_ON(!hlist_empty(&stable_node->hlist));
917 free_stable_node_chain(stable_node, root);
921 static int remove_all_stable_nodes(void)
923 struct stable_node *stable_node, *next;
927 for (nid = 0; nid < ksm_nr_node_ids; nid++) {
928 while (root_stable_tree[nid].rb_node) {
929 stable_node = rb_entry(root_stable_tree[nid].rb_node,
930 struct stable_node, node);
931 if (remove_stable_node_chain(stable_node,
932 root_stable_tree + nid)) {
934 break; /* proceed to next nid */
939 list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
940 if (remove_stable_node(stable_node))
947 static int unmerge_and_remove_all_rmap_items(void)
949 struct mm_slot *mm_slot;
950 struct mm_struct *mm;
951 struct vm_area_struct *vma;
954 spin_lock(&ksm_mmlist_lock);
955 ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
956 struct mm_slot, mm_list);
957 spin_unlock(&ksm_mmlist_lock);
959 for (mm_slot = ksm_scan.mm_slot;
960 mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
962 down_read(&mm->mmap_sem);
963 for (vma = mm->mmap; vma; vma = vma->vm_next) {
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, &mm_slot->rmap_list);
975 up_read(&mm->mmap_sem);
977 spin_lock(&ksm_mmlist_lock);
978 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
979 struct mm_slot, mm_list);
980 if (ksm_test_exit(mm)) {
981 hash_del(&mm_slot->link);
982 list_del(&mm_slot->mm_list);
983 spin_unlock(&ksm_mmlist_lock);
985 free_mm_slot(mm_slot);
986 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
989 spin_unlock(&ksm_mmlist_lock);
992 /* Clean up stable nodes, but don't worry if some are still busy */
993 remove_all_stable_nodes();
998 up_read(&mm->mmap_sem);
999 spin_lock(&ksm_mmlist_lock);
1000 ksm_scan.mm_slot = &ksm_mm_head;
1001 spin_unlock(&ksm_mmlist_lock);
1004 #endif /* CONFIG_SYSFS */
1006 static u32 calc_checksum(struct page *page)
1009 void *addr = kmap_atomic(page);
1010 checksum = jhash2(addr, PAGE_SIZE / 4, 17);
1011 kunmap_atomic(addr);
1015 static int memcmp_pages(struct page *page1, struct page *page2)
1017 char *addr1, *addr2;
1020 addr1 = kmap_atomic(page1);
1021 addr2 = kmap_atomic(page2);
1022 ret = memcmp(addr1, addr2, PAGE_SIZE);
1023 kunmap_atomic(addr2);
1024 kunmap_atomic(addr1);
1028 static inline int pages_identical(struct page *page1, struct page *page2)
1030 return !memcmp_pages(page1, page2);
1033 static int write_protect_page(struct vm_area_struct *vma, struct page *page,
1036 struct mm_struct *mm = vma->vm_mm;
1037 struct page_vma_mapped_walk pvmw = {
1043 unsigned long mmun_start; /* For mmu_notifiers */
1044 unsigned long mmun_end; /* For mmu_notifiers */
1046 pvmw.address = page_address_in_vma(page, vma);
1047 if (pvmw.address == -EFAULT)
1050 BUG_ON(PageTransCompound(page));
1052 mmun_start = pvmw.address;
1053 mmun_end = pvmw.address + PAGE_SIZE;
1054 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1056 if (!page_vma_mapped_walk(&pvmw))
1058 if (WARN_ONCE(!pvmw.pte, "Unexpected PMD mapping?"))
1061 if (pte_write(*pvmw.pte) || pte_dirty(*pvmw.pte) ||
1062 (pte_protnone(*pvmw.pte) && pte_savedwrite(*pvmw.pte)) ||
1063 mm_tlb_flush_pending(mm)) {
1066 swapped = PageSwapCache(page);
1067 flush_cache_page(vma, pvmw.address, page_to_pfn(page));
1069 * Ok this is tricky, when get_user_pages_fast() run it doesn't
1070 * take any lock, therefore the check that we are going to make
1071 * with the pagecount against the mapcount is racey and
1072 * O_DIRECT can happen right after the check.
1073 * So we clear the pte and flush the tlb before the check
1074 * this assure us that no O_DIRECT can happen after the check
1075 * or in the middle of the check.
1077 * No need to notify as we are downgrading page table to read
1078 * only not changing it to point to a new page.
1080 * See Documentation/vm/mmu_notifier.rst
1082 entry = ptep_clear_flush(vma, pvmw.address, pvmw.pte);
1084 * Check that no O_DIRECT or similar I/O is in progress on the
1087 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
1088 set_pte_at(mm, pvmw.address, pvmw.pte, entry);
1091 if (pte_dirty(entry))
1092 set_page_dirty(page);
1094 if (pte_protnone(entry))
1095 entry = pte_mkclean(pte_clear_savedwrite(entry));
1097 entry = pte_mkclean(pte_wrprotect(entry));
1098 set_pte_at_notify(mm, pvmw.address, pvmw.pte, entry);
1100 *orig_pte = *pvmw.pte;
1104 page_vma_mapped_walk_done(&pvmw);
1106 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1112 * replace_page - replace page in vma by new ksm page
1113 * @vma: vma that holds the pte pointing to page
1114 * @page: the page we are replacing by kpage
1115 * @kpage: the ksm page we replace page by
1116 * @orig_pte: the original value of the pte
1118 * Returns 0 on success, -EFAULT on failure.
1120 static int replace_page(struct vm_area_struct *vma, struct page *page,
1121 struct page *kpage, pte_t orig_pte)
1123 struct mm_struct *mm = vma->vm_mm;
1130 unsigned long mmun_start; /* For mmu_notifiers */
1131 unsigned long mmun_end; /* For mmu_notifiers */
1133 addr = page_address_in_vma(page, vma);
1134 if (addr == -EFAULT)
1137 pmd = mm_find_pmd(mm, addr);
1142 mmun_end = addr + PAGE_SIZE;
1143 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
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 alreaady.
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(mm, mmun_start, mmun_end);
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;
1285 down_read(&mm->mmap_sem);
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_sem */
1298 rmap_item->anon_vma = vma->anon_vma;
1299 get_anon_vma(vma->anon_vma);
1301 up_read(&mm->mmap_sem);
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, false);
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, false);
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 * wrprotected at all times. Any will work
1614 * fine to continue the walk.
1616 tree_page = get_ksm_page(stable_node_any, false);
1618 VM_BUG_ON(!stable_node_dup ^ !!stable_node_any);
1621 * If we walked over a stale stable_node,
1622 * get_ksm_page() will call rb_erase() and it
1623 * may rebalance the tree from under us. So
1624 * restart the search from scratch. Returning
1625 * NULL would be safe too, but we'd generate
1626 * false negative insertions just because some
1627 * stable_node was stale.
1632 ret = memcmp_pages(page, tree_page);
1633 put_page(tree_page);
1637 new = &parent->rb_left;
1639 new = &parent->rb_right;
1642 VM_BUG_ON(page_node->head != &migrate_nodes);
1644 * Test if the migrated page should be merged
1645 * into a stable node dup. If the mapcount is
1646 * 1 we can migrate it with another KSM page
1647 * without adding it to the chain.
1649 if (page_mapcount(page) > 1)
1653 if (!stable_node_dup) {
1655 * If the stable_node is a chain and
1656 * we got a payload match in memcmp
1657 * but we cannot merge the scanned
1658 * page in any of the existing
1659 * stable_node dups because they're
1660 * all full, we need to wait the
1661 * scanned page to find itself a match
1662 * in the unstable tree to create a
1663 * brand new KSM page to add later to
1664 * the dups of this stable_node.
1670 * Lock and unlock the stable_node's page (which
1671 * might already have been migrated) so that page
1672 * migration is sure to notice its raised count.
1673 * It would be more elegant to return stable_node
1674 * than kpage, but that involves more changes.
1676 tree_page = get_ksm_page(stable_node_dup, true);
1677 if (unlikely(!tree_page))
1679 * The tree may have been rebalanced,
1680 * so re-evaluate parent and new.
1683 unlock_page(tree_page);
1685 if (get_kpfn_nid(stable_node_dup->kpfn) !=
1686 NUMA(stable_node_dup->nid)) {
1687 put_page(tree_page);
1697 list_del(&page_node->list);
1698 DO_NUMA(page_node->nid = nid);
1699 rb_link_node(&page_node->node, parent, new);
1700 rb_insert_color(&page_node->node, root);
1702 if (is_page_sharing_candidate(page_node)) {
1710 * If stable_node was a chain and chain_prune collapsed it,
1711 * stable_node has been updated to be the new regular
1712 * stable_node. A collapse of the chain is indistinguishable
1713 * from the case there was no chain in the stable
1714 * rbtree. Otherwise stable_node is the chain and
1715 * stable_node_dup is the dup to replace.
1717 if (stable_node_dup == stable_node) {
1718 VM_BUG_ON(is_stable_node_chain(stable_node_dup));
1719 VM_BUG_ON(is_stable_node_dup(stable_node_dup));
1720 /* there is no chain */
1722 VM_BUG_ON(page_node->head != &migrate_nodes);
1723 list_del(&page_node->list);
1724 DO_NUMA(page_node->nid = nid);
1725 rb_replace_node(&stable_node_dup->node,
1728 if (is_page_sharing_candidate(page_node))
1733 rb_erase(&stable_node_dup->node, root);
1737 VM_BUG_ON(!is_stable_node_chain(stable_node));
1738 __stable_node_dup_del(stable_node_dup);
1740 VM_BUG_ON(page_node->head != &migrate_nodes);
1741 list_del(&page_node->list);
1742 DO_NUMA(page_node->nid = nid);
1743 stable_node_chain_add_dup(page_node, stable_node);
1744 if (is_page_sharing_candidate(page_node))
1752 stable_node_dup->head = &migrate_nodes;
1753 list_add(&stable_node_dup->list, stable_node_dup->head);
1757 /* stable_node_dup could be null if it reached the limit */
1758 if (!stable_node_dup)
1759 stable_node_dup = stable_node_any;
1761 * If stable_node was a chain and chain_prune collapsed it,
1762 * stable_node has been updated to be the new regular
1763 * stable_node. A collapse of the chain is indistinguishable
1764 * from the case there was no chain in the stable
1765 * rbtree. Otherwise stable_node is the chain and
1766 * stable_node_dup is the dup to replace.
1768 if (stable_node_dup == stable_node) {
1769 VM_BUG_ON(is_stable_node_chain(stable_node_dup));
1770 VM_BUG_ON(is_stable_node_dup(stable_node_dup));
1771 /* chain is missing so create it */
1772 stable_node = alloc_stable_node_chain(stable_node_dup,
1778 * Add this stable_node dup that was
1779 * migrated to the stable_node chain
1780 * of the current nid for this page
1783 VM_BUG_ON(!is_stable_node_chain(stable_node));
1784 VM_BUG_ON(!is_stable_node_dup(stable_node_dup));
1785 VM_BUG_ON(page_node->head != &migrate_nodes);
1786 list_del(&page_node->list);
1787 DO_NUMA(page_node->nid = nid);
1788 stable_node_chain_add_dup(page_node, stable_node);
1793 * stable_tree_insert - insert stable tree node pointing to new ksm page
1794 * into the stable tree.
1796 * This function returns the stable tree node just allocated on success,
1799 static struct stable_node *stable_tree_insert(struct page *kpage)
1803 struct rb_root *root;
1804 struct rb_node **new;
1805 struct rb_node *parent;
1806 struct stable_node *stable_node, *stable_node_dup, *stable_node_any;
1807 bool need_chain = false;
1809 kpfn = page_to_pfn(kpage);
1810 nid = get_kpfn_nid(kpfn);
1811 root = root_stable_tree + nid;
1814 new = &root->rb_node;
1817 struct page *tree_page;
1821 stable_node = rb_entry(*new, struct stable_node, node);
1822 stable_node_any = NULL;
1823 tree_page = chain(&stable_node_dup, stable_node, root);
1824 if (!stable_node_dup) {
1826 * Either all stable_node dups were full in
1827 * this stable_node chain, or this chain was
1828 * empty and should be rb_erased.
1830 stable_node_any = stable_node_dup_any(stable_node,
1832 if (!stable_node_any) {
1833 /* rb_erase just run */
1837 * Take any of the stable_node dups page of
1838 * this stable_node chain to let the tree walk
1839 * continue. All KSM pages belonging to the
1840 * stable_node dups in a stable_node chain
1841 * have the same content and they're
1842 * wrprotected at all times. Any will work
1843 * fine to continue the walk.
1845 tree_page = get_ksm_page(stable_node_any, false);
1847 VM_BUG_ON(!stable_node_dup ^ !!stable_node_any);
1850 * If we walked over a stale stable_node,
1851 * get_ksm_page() will call rb_erase() and it
1852 * may rebalance the tree from under us. So
1853 * restart the search from scratch. Returning
1854 * NULL would be safe too, but we'd generate
1855 * false negative insertions just because some
1856 * stable_node was stale.
1861 ret = memcmp_pages(kpage, tree_page);
1862 put_page(tree_page);
1866 new = &parent->rb_left;
1868 new = &parent->rb_right;
1875 stable_node_dup = alloc_stable_node();
1876 if (!stable_node_dup)
1879 INIT_HLIST_HEAD(&stable_node_dup->hlist);
1880 stable_node_dup->kpfn = kpfn;
1881 set_page_stable_node(kpage, stable_node_dup);
1882 stable_node_dup->rmap_hlist_len = 0;
1883 DO_NUMA(stable_node_dup->nid = nid);
1885 rb_link_node(&stable_node_dup->node, parent, new);
1886 rb_insert_color(&stable_node_dup->node, root);
1888 if (!is_stable_node_chain(stable_node)) {
1889 struct stable_node *orig = stable_node;
1890 /* chain is missing so create it */
1891 stable_node = alloc_stable_node_chain(orig, root);
1893 free_stable_node(stable_node_dup);
1897 stable_node_chain_add_dup(stable_node_dup, stable_node);
1900 return stable_node_dup;
1904 * unstable_tree_search_insert - search for identical page,
1905 * else insert rmap_item into the unstable tree.
1907 * This function searches for a page in the unstable tree identical to the
1908 * page currently being scanned; and if no identical page is found in the
1909 * tree, we insert rmap_item as a new object into the unstable tree.
1911 * This function returns pointer to rmap_item found to be identical
1912 * to the currently scanned page, NULL otherwise.
1914 * This function does both searching and inserting, because they share
1915 * the same walking algorithm in an rbtree.
1918 struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1920 struct page **tree_pagep)
1922 struct rb_node **new;
1923 struct rb_root *root;
1924 struct rb_node *parent = NULL;
1927 nid = get_kpfn_nid(page_to_pfn(page));
1928 root = root_unstable_tree + nid;
1929 new = &root->rb_node;
1932 struct rmap_item *tree_rmap_item;
1933 struct page *tree_page;
1937 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
1938 tree_page = get_mergeable_page(tree_rmap_item);
1943 * Don't substitute a ksm page for a forked page.
1945 if (page == tree_page) {
1946 put_page(tree_page);
1950 ret = memcmp_pages(page, tree_page);
1954 put_page(tree_page);
1955 new = &parent->rb_left;
1956 } else if (ret > 0) {
1957 put_page(tree_page);
1958 new = &parent->rb_right;
1959 } else if (!ksm_merge_across_nodes &&
1960 page_to_nid(tree_page) != nid) {
1962 * If tree_page has been migrated to another NUMA node,
1963 * it will be flushed out and put in the right unstable
1964 * tree next time: only merge with it when across_nodes.
1966 put_page(tree_page);
1969 *tree_pagep = tree_page;
1970 return tree_rmap_item;
1974 rmap_item->address |= UNSTABLE_FLAG;
1975 rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1976 DO_NUMA(rmap_item->nid = nid);
1977 rb_link_node(&rmap_item->node, parent, new);
1978 rb_insert_color(&rmap_item->node, root);
1980 ksm_pages_unshared++;
1985 * stable_tree_append - add another rmap_item to the linked list of
1986 * rmap_items hanging off a given node of the stable tree, all sharing
1987 * the same ksm page.
1989 static void stable_tree_append(struct rmap_item *rmap_item,
1990 struct stable_node *stable_node,
1991 bool max_page_sharing_bypass)
1994 * rmap won't find this mapping if we don't insert the
1995 * rmap_item in the right stable_node
1996 * duplicate. page_migration could break later if rmap breaks,
1997 * so we can as well crash here. We really need to check for
1998 * rmap_hlist_len == STABLE_NODE_CHAIN, but we can as well check
1999 * for other negative values as an undeflow if detected here
2000 * for the first time (and not when decreasing rmap_hlist_len)
2001 * would be sign of memory corruption in the stable_node.
2003 BUG_ON(stable_node->rmap_hlist_len < 0);
2005 stable_node->rmap_hlist_len++;
2006 if (!max_page_sharing_bypass)
2007 /* possibly non fatal but unexpected overflow, only warn */
2008 WARN_ON_ONCE(stable_node->rmap_hlist_len >
2009 ksm_max_page_sharing);
2011 rmap_item->head = stable_node;
2012 rmap_item->address |= STABLE_FLAG;
2013 hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
2015 if (rmap_item->hlist.next)
2016 ksm_pages_sharing++;
2022 * cmp_and_merge_page - first see if page can be merged into the stable tree;
2023 * if not, compare checksum to previous and if it's the same, see if page can
2024 * be inserted into the unstable tree, or merged with a page already there and
2025 * both transferred to the stable tree.
2027 * @page: the page that we are searching identical page to.
2028 * @rmap_item: the reverse mapping into the virtual address of this page
2030 static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
2032 struct mm_struct *mm = rmap_item->mm;
2033 struct rmap_item *tree_rmap_item;
2034 struct page *tree_page = NULL;
2035 struct stable_node *stable_node;
2037 unsigned int checksum;
2039 bool max_page_sharing_bypass = false;
2041 stable_node = page_stable_node(page);
2043 if (stable_node->head != &migrate_nodes &&
2044 get_kpfn_nid(READ_ONCE(stable_node->kpfn)) !=
2045 NUMA(stable_node->nid)) {
2046 stable_node_dup_del(stable_node);
2047 stable_node->head = &migrate_nodes;
2048 list_add(&stable_node->list, stable_node->head);
2050 if (stable_node->head != &migrate_nodes &&
2051 rmap_item->head == stable_node)
2054 * If it's a KSM fork, allow it to go over the sharing limit
2057 if (!is_page_sharing_candidate(stable_node))
2058 max_page_sharing_bypass = true;
2061 /* We first start with searching the page inside the stable tree */
2062 kpage = stable_tree_search(page);
2063 if (kpage == page && rmap_item->head == stable_node) {
2068 remove_rmap_item_from_tree(rmap_item);
2071 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
2074 * The page was successfully merged:
2075 * add its rmap_item to the stable tree.
2078 stable_tree_append(rmap_item, page_stable_node(kpage),
2079 max_page_sharing_bypass);
2087 * If the hash value of the page has changed from the last time
2088 * we calculated it, this page is changing frequently: therefore we
2089 * don't want to insert it in the unstable tree, and we don't want
2090 * to waste our time searching for something identical to it there.
2092 checksum = calc_checksum(page);
2093 if (rmap_item->oldchecksum != checksum) {
2094 rmap_item->oldchecksum = checksum;
2099 * Same checksum as an empty page. We attempt to merge it with the
2100 * appropriate zero page if the user enabled this via sysfs.
2102 if (ksm_use_zero_pages && (checksum == zero_checksum)) {
2103 struct vm_area_struct *vma;
2105 down_read(&mm->mmap_sem);
2106 vma = find_mergeable_vma(mm, rmap_item->address);
2107 err = try_to_merge_one_page(vma, page,
2108 ZERO_PAGE(rmap_item->address));
2109 up_read(&mm->mmap_sem);
2111 * In case of failure, the page was not really empty, so we
2112 * need to continue. Otherwise we're done.
2118 unstable_tree_search_insert(rmap_item, page, &tree_page);
2119 if (tree_rmap_item) {
2122 kpage = try_to_merge_two_pages(rmap_item, page,
2123 tree_rmap_item, tree_page);
2125 * If both pages we tried to merge belong to the same compound
2126 * page, then we actually ended up increasing the reference
2127 * count of the same compound page twice, and split_huge_page
2129 * Here we set a flag if that happened, and we use it later to
2130 * try split_huge_page again. Since we call put_page right
2131 * afterwards, the reference count will be correct and
2132 * split_huge_page should succeed.
2134 split = PageTransCompound(page)
2135 && compound_head(page) == compound_head(tree_page);
2136 put_page(tree_page);
2139 * The pages were successfully merged: insert new
2140 * node in the stable tree and add both rmap_items.
2143 stable_node = stable_tree_insert(kpage);
2145 stable_tree_append(tree_rmap_item, stable_node,
2147 stable_tree_append(rmap_item, stable_node,
2153 * If we fail to insert the page into the stable tree,
2154 * we will have 2 virtual addresses that are pointing
2155 * to a ksm page left outside the stable tree,
2156 * in which case we need to break_cow on both.
2159 break_cow(tree_rmap_item);
2160 break_cow(rmap_item);
2164 * We are here if we tried to merge two pages and
2165 * failed because they both belonged to the same
2166 * compound page. We will split the page now, but no
2167 * merging will take place.
2168 * We do not want to add the cost of a full lock; if
2169 * the page is locked, it is better to skip it and
2170 * perhaps try again later.
2172 if (!trylock_page(page))
2174 split_huge_page(page);
2180 static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
2181 struct rmap_item **rmap_list,
2184 struct rmap_item *rmap_item;
2186 while (*rmap_list) {
2187 rmap_item = *rmap_list;
2188 if ((rmap_item->address & PAGE_MASK) == addr)
2190 if (rmap_item->address > addr)
2192 *rmap_list = rmap_item->rmap_list;
2193 remove_rmap_item_from_tree(rmap_item);
2194 free_rmap_item(rmap_item);
2197 rmap_item = alloc_rmap_item();
2199 /* It has already been zeroed */
2200 rmap_item->mm = mm_slot->mm;
2201 rmap_item->address = addr;
2202 rmap_item->rmap_list = *rmap_list;
2203 *rmap_list = rmap_item;
2208 static struct rmap_item *scan_get_next_rmap_item(struct page **page)
2210 struct mm_struct *mm;
2211 struct mm_slot *slot;
2212 struct vm_area_struct *vma;
2213 struct rmap_item *rmap_item;
2216 if (list_empty(&ksm_mm_head.mm_list))
2219 slot = ksm_scan.mm_slot;
2220 if (slot == &ksm_mm_head) {
2222 * A number of pages can hang around indefinitely on per-cpu
2223 * pagevecs, raised page count preventing write_protect_page
2224 * from merging them. Though it doesn't really matter much,
2225 * it is puzzling to see some stuck in pages_volatile until
2226 * other activity jostles them out, and they also prevented
2227 * LTP's KSM test from succeeding deterministically; so drain
2228 * them here (here rather than on entry to ksm_do_scan(),
2229 * so we don't IPI too often when pages_to_scan is set low).
2231 lru_add_drain_all();
2234 * Whereas stale stable_nodes on the stable_tree itself
2235 * get pruned in the regular course of stable_tree_search(),
2236 * those moved out to the migrate_nodes list can accumulate:
2237 * so prune them once before each full scan.
2239 if (!ksm_merge_across_nodes) {
2240 struct stable_node *stable_node, *next;
2243 list_for_each_entry_safe(stable_node, next,
2244 &migrate_nodes, list) {
2245 page = get_ksm_page(stable_node, false);
2252 for (nid = 0; nid < ksm_nr_node_ids; nid++)
2253 root_unstable_tree[nid] = RB_ROOT;
2255 spin_lock(&ksm_mmlist_lock);
2256 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
2257 ksm_scan.mm_slot = slot;
2258 spin_unlock(&ksm_mmlist_lock);
2260 * Although we tested list_empty() above, a racing __ksm_exit
2261 * of the last mm on the list may have removed it since then.
2263 if (slot == &ksm_mm_head)
2266 ksm_scan.address = 0;
2267 ksm_scan.rmap_list = &slot->rmap_list;
2271 down_read(&mm->mmap_sem);
2272 if (ksm_test_exit(mm))
2275 vma = find_vma(mm, ksm_scan.address);
2277 for (; vma; vma = vma->vm_next) {
2278 if (!(vma->vm_flags & VM_MERGEABLE))
2280 if (ksm_scan.address < vma->vm_start)
2281 ksm_scan.address = vma->vm_start;
2283 ksm_scan.address = vma->vm_end;
2285 while (ksm_scan.address < vma->vm_end) {
2286 if (ksm_test_exit(mm))
2288 *page = follow_page(vma, ksm_scan.address, FOLL_GET);
2289 if (IS_ERR_OR_NULL(*page)) {
2290 ksm_scan.address += PAGE_SIZE;
2294 if (PageAnon(*page)) {
2295 flush_anon_page(vma, *page, ksm_scan.address);
2296 flush_dcache_page(*page);
2297 rmap_item = get_next_rmap_item(slot,
2298 ksm_scan.rmap_list, ksm_scan.address);
2300 ksm_scan.rmap_list =
2301 &rmap_item->rmap_list;
2302 ksm_scan.address += PAGE_SIZE;
2305 up_read(&mm->mmap_sem);
2309 ksm_scan.address += PAGE_SIZE;
2314 if (ksm_test_exit(mm)) {
2315 ksm_scan.address = 0;
2316 ksm_scan.rmap_list = &slot->rmap_list;
2319 * Nuke all the rmap_items that are above this current rmap:
2320 * because there were no VM_MERGEABLE vmas with such addresses.
2322 remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
2324 spin_lock(&ksm_mmlist_lock);
2325 ksm_scan.mm_slot = list_entry(slot->mm_list.next,
2326 struct mm_slot, mm_list);
2327 if (ksm_scan.address == 0) {
2329 * We've completed a full scan of all vmas, holding mmap_sem
2330 * throughout, and found no VM_MERGEABLE: so do the same as
2331 * __ksm_exit does to remove this mm from all our lists now.
2332 * This applies either when cleaning up after __ksm_exit
2333 * (but beware: we can reach here even before __ksm_exit),
2334 * or when all VM_MERGEABLE areas have been unmapped (and
2335 * mmap_sem then protects against race with MADV_MERGEABLE).
2337 hash_del(&slot->link);
2338 list_del(&slot->mm_list);
2339 spin_unlock(&ksm_mmlist_lock);
2342 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
2343 up_read(&mm->mmap_sem);
2346 up_read(&mm->mmap_sem);
2348 * up_read(&mm->mmap_sem) first because after
2349 * spin_unlock(&ksm_mmlist_lock) run, the "mm" may
2350 * already have been freed under us by __ksm_exit()
2351 * because the "mm_slot" is still hashed and
2352 * ksm_scan.mm_slot doesn't point to it anymore.
2354 spin_unlock(&ksm_mmlist_lock);
2357 /* Repeat until we've completed scanning the whole list */
2358 slot = ksm_scan.mm_slot;
2359 if (slot != &ksm_mm_head)
2367 * ksm_do_scan - the ksm scanner main worker function.
2368 * @scan_npages: number of pages we want to scan before we return.
2370 static void ksm_do_scan(unsigned int scan_npages)
2372 struct rmap_item *rmap_item;
2373 struct page *uninitialized_var(page);
2375 while (scan_npages-- && likely(!freezing(current))) {
2377 rmap_item = scan_get_next_rmap_item(&page);
2380 cmp_and_merge_page(page, rmap_item);
2385 static int ksmd_should_run(void)
2387 return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
2390 static int ksm_scan_thread(void *nothing)
2393 set_user_nice(current, 5);
2395 while (!kthread_should_stop()) {
2396 mutex_lock(&ksm_thread_mutex);
2397 wait_while_offlining();
2398 if (ksmd_should_run())
2399 ksm_do_scan(ksm_thread_pages_to_scan);
2400 mutex_unlock(&ksm_thread_mutex);
2404 if (ksmd_should_run()) {
2405 schedule_timeout_interruptible(
2406 msecs_to_jiffies(ksm_thread_sleep_millisecs));
2408 wait_event_freezable(ksm_thread_wait,
2409 ksmd_should_run() || kthread_should_stop());
2415 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
2416 unsigned long end, int advice, unsigned long *vm_flags)
2418 struct mm_struct *mm = vma->vm_mm;
2422 case MADV_MERGEABLE:
2424 * Be somewhat over-protective for now!
2426 if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE |
2427 VM_PFNMAP | VM_IO | VM_DONTEXPAND |
2428 VM_HUGETLB | VM_MIXEDMAP))
2429 return 0; /* just ignore the advice */
2432 if (*vm_flags & VM_SAO)
2436 if (*vm_flags & VM_SPARC_ADI)
2440 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
2441 err = __ksm_enter(mm);
2446 *vm_flags |= VM_MERGEABLE;
2449 case MADV_UNMERGEABLE:
2450 if (!(*vm_flags & VM_MERGEABLE))
2451 return 0; /* just ignore the advice */
2453 if (vma->anon_vma) {
2454 err = unmerge_ksm_pages(vma, start, end);
2459 *vm_flags &= ~VM_MERGEABLE;
2466 int __ksm_enter(struct mm_struct *mm)
2468 struct mm_slot *mm_slot;
2471 mm_slot = alloc_mm_slot();
2475 /* Check ksm_run too? Would need tighter locking */
2476 needs_wakeup = list_empty(&ksm_mm_head.mm_list);
2478 spin_lock(&ksm_mmlist_lock);
2479 insert_to_mm_slots_hash(mm, mm_slot);
2481 * When KSM_RUN_MERGE (or KSM_RUN_STOP),
2482 * insert just behind the scanning cursor, to let the area settle
2483 * down a little; when fork is followed by immediate exec, we don't
2484 * want ksmd to waste time setting up and tearing down an rmap_list.
2486 * But when KSM_RUN_UNMERGE, it's important to insert ahead of its
2487 * scanning cursor, otherwise KSM pages in newly forked mms will be
2488 * missed: then we might as well insert at the end of the list.
2490 if (ksm_run & KSM_RUN_UNMERGE)
2491 list_add_tail(&mm_slot->mm_list, &ksm_mm_head.mm_list);
2493 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
2494 spin_unlock(&ksm_mmlist_lock);
2496 set_bit(MMF_VM_MERGEABLE, &mm->flags);
2500 wake_up_interruptible(&ksm_thread_wait);
2505 void __ksm_exit(struct mm_struct *mm)
2507 struct mm_slot *mm_slot;
2508 int easy_to_free = 0;
2511 * This process is exiting: if it's straightforward (as is the
2512 * case when ksmd was never running), free mm_slot immediately.
2513 * But if it's at the cursor or has rmap_items linked to it, use
2514 * mmap_sem to synchronize with any break_cows before pagetables
2515 * are freed, and leave the mm_slot on the list for ksmd to free.
2516 * Beware: ksm may already have noticed it exiting and freed the slot.
2519 spin_lock(&ksm_mmlist_lock);
2520 mm_slot = get_mm_slot(mm);
2521 if (mm_slot && ksm_scan.mm_slot != mm_slot) {
2522 if (!mm_slot->rmap_list) {
2523 hash_del(&mm_slot->link);
2524 list_del(&mm_slot->mm_list);
2527 list_move(&mm_slot->mm_list,
2528 &ksm_scan.mm_slot->mm_list);
2531 spin_unlock(&ksm_mmlist_lock);
2534 free_mm_slot(mm_slot);
2535 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
2537 } else if (mm_slot) {
2538 down_write(&mm->mmap_sem);
2539 up_write(&mm->mmap_sem);
2543 struct page *ksm_might_need_to_copy(struct page *page,
2544 struct vm_area_struct *vma, unsigned long address)
2546 struct anon_vma *anon_vma = page_anon_vma(page);
2547 struct page *new_page;
2549 if (PageKsm(page)) {
2550 if (page_stable_node(page) &&
2551 !(ksm_run & KSM_RUN_UNMERGE))
2552 return page; /* no need to copy it */
2553 } else if (!anon_vma) {
2554 return page; /* no need to copy it */
2555 } else if (anon_vma->root == vma->anon_vma->root &&
2556 page->index == linear_page_index(vma, address)) {
2557 return page; /* still no need to copy it */
2559 if (!PageUptodate(page))
2560 return page; /* let do_swap_page report the error */
2562 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2564 copy_user_highpage(new_page, page, address, vma);
2566 SetPageDirty(new_page);
2567 __SetPageUptodate(new_page);
2568 __SetPageLocked(new_page);
2574 void rmap_walk_ksm(struct page *page, struct rmap_walk_control *rwc)
2576 struct stable_node *stable_node;
2577 struct rmap_item *rmap_item;
2578 int search_new_forks = 0;
2580 VM_BUG_ON_PAGE(!PageKsm(page), page);
2583 * Rely on the page lock to protect against concurrent modifications
2584 * to that page's node of the stable tree.
2586 VM_BUG_ON_PAGE(!PageLocked(page), page);
2588 stable_node = page_stable_node(page);
2592 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
2593 struct anon_vma *anon_vma = rmap_item->anon_vma;
2594 struct anon_vma_chain *vmac;
2595 struct vm_area_struct *vma;
2598 anon_vma_lock_read(anon_vma);
2599 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
2603 if (rmap_item->address < vma->vm_start ||
2604 rmap_item->address >= vma->vm_end)
2607 * Initially we examine only the vma which covers this
2608 * rmap_item; but later, if there is still work to do,
2609 * we examine covering vmas in other mms: in case they
2610 * were forked from the original since ksmd passed.
2612 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
2615 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2618 if (!rwc->rmap_one(page, vma,
2619 rmap_item->address, rwc->arg)) {
2620 anon_vma_unlock_read(anon_vma);
2623 if (rwc->done && rwc->done(page)) {
2624 anon_vma_unlock_read(anon_vma);
2628 anon_vma_unlock_read(anon_vma);
2630 if (!search_new_forks++)
2634 #ifdef CONFIG_MIGRATION
2635 void ksm_migrate_page(struct page *newpage, struct page *oldpage)
2637 struct stable_node *stable_node;
2639 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
2640 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
2641 VM_BUG_ON_PAGE(newpage->mapping != oldpage->mapping, newpage);
2643 stable_node = page_stable_node(newpage);
2645 VM_BUG_ON_PAGE(stable_node->kpfn != page_to_pfn(oldpage), oldpage);
2646 stable_node->kpfn = page_to_pfn(newpage);
2648 * newpage->mapping was set in advance; now we need smp_wmb()
2649 * to make sure that the new stable_node->kpfn is visible
2650 * to get_ksm_page() before it can see that oldpage->mapping
2651 * has gone stale (or that PageSwapCache has been cleared).
2654 set_page_stable_node(oldpage, NULL);
2657 #endif /* CONFIG_MIGRATION */
2659 #ifdef CONFIG_MEMORY_HOTREMOVE
2660 static void wait_while_offlining(void)
2662 while (ksm_run & KSM_RUN_OFFLINE) {
2663 mutex_unlock(&ksm_thread_mutex);
2664 wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE),
2665 TASK_UNINTERRUPTIBLE);
2666 mutex_lock(&ksm_thread_mutex);
2670 static bool stable_node_dup_remove_range(struct stable_node *stable_node,
2671 unsigned long start_pfn,
2672 unsigned long end_pfn)
2674 if (stable_node->kpfn >= start_pfn &&
2675 stable_node->kpfn < end_pfn) {
2677 * Don't get_ksm_page, page has already gone:
2678 * which is why we keep kpfn instead of page*
2680 remove_node_from_stable_tree(stable_node);
2686 static bool stable_node_chain_remove_range(struct stable_node *stable_node,
2687 unsigned long start_pfn,
2688 unsigned long end_pfn,
2689 struct rb_root *root)
2691 struct stable_node *dup;
2692 struct hlist_node *hlist_safe;
2694 if (!is_stable_node_chain(stable_node)) {
2695 VM_BUG_ON(is_stable_node_dup(stable_node));
2696 return stable_node_dup_remove_range(stable_node, start_pfn,
2700 hlist_for_each_entry_safe(dup, hlist_safe,
2701 &stable_node->hlist, hlist_dup) {
2702 VM_BUG_ON(!is_stable_node_dup(dup));
2703 stable_node_dup_remove_range(dup, start_pfn, end_pfn);
2705 if (hlist_empty(&stable_node->hlist)) {
2706 free_stable_node_chain(stable_node, root);
2707 return true; /* notify caller that tree was rebalanced */
2712 static void ksm_check_stable_tree(unsigned long start_pfn,
2713 unsigned long end_pfn)
2715 struct stable_node *stable_node, *next;
2716 struct rb_node *node;
2719 for (nid = 0; nid < ksm_nr_node_ids; nid++) {
2720 node = rb_first(root_stable_tree + nid);
2722 stable_node = rb_entry(node, struct stable_node, node);
2723 if (stable_node_chain_remove_range(stable_node,
2727 node = rb_first(root_stable_tree + nid);
2729 node = rb_next(node);
2733 list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
2734 if (stable_node->kpfn >= start_pfn &&
2735 stable_node->kpfn < end_pfn)
2736 remove_node_from_stable_tree(stable_node);
2741 static int ksm_memory_callback(struct notifier_block *self,
2742 unsigned long action, void *arg)
2744 struct memory_notify *mn = arg;
2747 case MEM_GOING_OFFLINE:
2749 * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items()
2750 * and remove_all_stable_nodes() while memory is going offline:
2751 * it is unsafe for them to touch the stable tree at this time.
2752 * But unmerge_ksm_pages(), rmap lookups and other entry points
2753 * which do not need the ksm_thread_mutex are all safe.
2755 mutex_lock(&ksm_thread_mutex);
2756 ksm_run |= KSM_RUN_OFFLINE;
2757 mutex_unlock(&ksm_thread_mutex);
2762 * Most of the work is done by page migration; but there might
2763 * be a few stable_nodes left over, still pointing to struct
2764 * pages which have been offlined: prune those from the tree,
2765 * otherwise get_ksm_page() might later try to access a
2766 * non-existent struct page.
2768 ksm_check_stable_tree(mn->start_pfn,
2769 mn->start_pfn + mn->nr_pages);
2772 case MEM_CANCEL_OFFLINE:
2773 mutex_lock(&ksm_thread_mutex);
2774 ksm_run &= ~KSM_RUN_OFFLINE;
2775 mutex_unlock(&ksm_thread_mutex);
2777 smp_mb(); /* wake_up_bit advises this */
2778 wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE));
2784 static void wait_while_offlining(void)
2787 #endif /* CONFIG_MEMORY_HOTREMOVE */
2791 * This all compiles without CONFIG_SYSFS, but is a waste of space.
2794 #define KSM_ATTR_RO(_name) \
2795 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2796 #define KSM_ATTR(_name) \
2797 static struct kobj_attribute _name##_attr = \
2798 __ATTR(_name, 0644, _name##_show, _name##_store)
2800 static ssize_t sleep_millisecs_show(struct kobject *kobj,
2801 struct kobj_attribute *attr, char *buf)
2803 return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
2806 static ssize_t sleep_millisecs_store(struct kobject *kobj,
2807 struct kobj_attribute *attr,
2808 const char *buf, size_t count)
2810 unsigned long msecs;
2813 err = kstrtoul(buf, 10, &msecs);
2814 if (err || msecs > UINT_MAX)
2817 ksm_thread_sleep_millisecs = msecs;
2821 KSM_ATTR(sleep_millisecs);
2823 static ssize_t pages_to_scan_show(struct kobject *kobj,
2824 struct kobj_attribute *attr, char *buf)
2826 return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
2829 static ssize_t pages_to_scan_store(struct kobject *kobj,
2830 struct kobj_attribute *attr,
2831 const char *buf, size_t count)
2834 unsigned long nr_pages;
2836 err = kstrtoul(buf, 10, &nr_pages);
2837 if (err || nr_pages > UINT_MAX)
2840 ksm_thread_pages_to_scan = nr_pages;
2844 KSM_ATTR(pages_to_scan);
2846 static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
2849 return sprintf(buf, "%lu\n", ksm_run);
2852 static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
2853 const char *buf, size_t count)
2856 unsigned long flags;
2858 err = kstrtoul(buf, 10, &flags);
2859 if (err || flags > UINT_MAX)
2861 if (flags > KSM_RUN_UNMERGE)
2865 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
2866 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
2867 * breaking COW to free the pages_shared (but leaves mm_slots
2868 * on the list for when ksmd may be set running again).
2871 mutex_lock(&ksm_thread_mutex);
2872 wait_while_offlining();
2873 if (ksm_run != flags) {
2875 if (flags & KSM_RUN_UNMERGE) {
2876 set_current_oom_origin();
2877 err = unmerge_and_remove_all_rmap_items();
2878 clear_current_oom_origin();
2880 ksm_run = KSM_RUN_STOP;
2885 mutex_unlock(&ksm_thread_mutex);
2887 if (flags & KSM_RUN_MERGE)
2888 wake_up_interruptible(&ksm_thread_wait);
2895 static ssize_t merge_across_nodes_show(struct kobject *kobj,
2896 struct kobj_attribute *attr, char *buf)
2898 return sprintf(buf, "%u\n", ksm_merge_across_nodes);
2901 static ssize_t merge_across_nodes_store(struct kobject *kobj,
2902 struct kobj_attribute *attr,
2903 const char *buf, size_t count)
2908 err = kstrtoul(buf, 10, &knob);
2914 mutex_lock(&ksm_thread_mutex);
2915 wait_while_offlining();
2916 if (ksm_merge_across_nodes != knob) {
2917 if (ksm_pages_shared || remove_all_stable_nodes())
2919 else if (root_stable_tree == one_stable_tree) {
2920 struct rb_root *buf;
2922 * This is the first time that we switch away from the
2923 * default of merging across nodes: must now allocate
2924 * a buffer to hold as many roots as may be needed.
2925 * Allocate stable and unstable together:
2926 * MAXSMP NODES_SHIFT 10 will use 16kB.
2928 buf = kcalloc(nr_node_ids + nr_node_ids, sizeof(*buf),
2930 /* Let us assume that RB_ROOT is NULL is zero */
2934 root_stable_tree = buf;
2935 root_unstable_tree = buf + nr_node_ids;
2936 /* Stable tree is empty but not the unstable */
2937 root_unstable_tree[0] = one_unstable_tree[0];
2941 ksm_merge_across_nodes = knob;
2942 ksm_nr_node_ids = knob ? 1 : nr_node_ids;
2945 mutex_unlock(&ksm_thread_mutex);
2947 return err ? err : count;
2949 KSM_ATTR(merge_across_nodes);
2952 static ssize_t use_zero_pages_show(struct kobject *kobj,
2953 struct kobj_attribute *attr, char *buf)
2955 return sprintf(buf, "%u\n", ksm_use_zero_pages);
2957 static ssize_t use_zero_pages_store(struct kobject *kobj,
2958 struct kobj_attribute *attr,
2959 const char *buf, size_t count)
2964 err = kstrtobool(buf, &value);
2968 ksm_use_zero_pages = value;
2972 KSM_ATTR(use_zero_pages);
2974 static ssize_t max_page_sharing_show(struct kobject *kobj,
2975 struct kobj_attribute *attr, char *buf)
2977 return sprintf(buf, "%u\n", ksm_max_page_sharing);
2980 static ssize_t max_page_sharing_store(struct kobject *kobj,
2981 struct kobj_attribute *attr,
2982 const char *buf, size_t count)
2987 err = kstrtoint(buf, 10, &knob);
2991 * When a KSM page is created it is shared by 2 mappings. This
2992 * being a signed comparison, it implicitly verifies it's not
2998 if (READ_ONCE(ksm_max_page_sharing) == knob)
3001 mutex_lock(&ksm_thread_mutex);
3002 wait_while_offlining();
3003 if (ksm_max_page_sharing != knob) {
3004 if (ksm_pages_shared || remove_all_stable_nodes())
3007 ksm_max_page_sharing = knob;
3009 mutex_unlock(&ksm_thread_mutex);
3011 return err ? err : count;
3013 KSM_ATTR(max_page_sharing);
3015 static ssize_t pages_shared_show(struct kobject *kobj,
3016 struct kobj_attribute *attr, char *buf)
3018 return sprintf(buf, "%lu\n", ksm_pages_shared);
3020 KSM_ATTR_RO(pages_shared);
3022 static ssize_t pages_sharing_show(struct kobject *kobj,
3023 struct kobj_attribute *attr, char *buf)
3025 return sprintf(buf, "%lu\n", ksm_pages_sharing);
3027 KSM_ATTR_RO(pages_sharing);
3029 static ssize_t pages_unshared_show(struct kobject *kobj,
3030 struct kobj_attribute *attr, char *buf)
3032 return sprintf(buf, "%lu\n", ksm_pages_unshared);
3034 KSM_ATTR_RO(pages_unshared);
3036 static ssize_t pages_volatile_show(struct kobject *kobj,
3037 struct kobj_attribute *attr, char *buf)
3039 long ksm_pages_volatile;
3041 ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
3042 - ksm_pages_sharing - ksm_pages_unshared;
3044 * It was not worth any locking to calculate that statistic,
3045 * but it might therefore sometimes be negative: conceal that.
3047 if (ksm_pages_volatile < 0)
3048 ksm_pages_volatile = 0;
3049 return sprintf(buf, "%ld\n", ksm_pages_volatile);
3051 KSM_ATTR_RO(pages_volatile);
3053 static ssize_t stable_node_dups_show(struct kobject *kobj,
3054 struct kobj_attribute *attr, char *buf)
3056 return sprintf(buf, "%lu\n", ksm_stable_node_dups);
3058 KSM_ATTR_RO(stable_node_dups);
3060 static ssize_t stable_node_chains_show(struct kobject *kobj,
3061 struct kobj_attribute *attr, char *buf)
3063 return sprintf(buf, "%lu\n", ksm_stable_node_chains);
3065 KSM_ATTR_RO(stable_node_chains);
3068 stable_node_chains_prune_millisecs_show(struct kobject *kobj,
3069 struct kobj_attribute *attr,
3072 return sprintf(buf, "%u\n", ksm_stable_node_chains_prune_millisecs);
3076 stable_node_chains_prune_millisecs_store(struct kobject *kobj,
3077 struct kobj_attribute *attr,
3078 const char *buf, size_t count)
3080 unsigned long msecs;
3083 err = kstrtoul(buf, 10, &msecs);
3084 if (err || msecs > UINT_MAX)
3087 ksm_stable_node_chains_prune_millisecs = msecs;
3091 KSM_ATTR(stable_node_chains_prune_millisecs);
3093 static ssize_t full_scans_show(struct kobject *kobj,
3094 struct kobj_attribute *attr, char *buf)
3096 return sprintf(buf, "%lu\n", ksm_scan.seqnr);
3098 KSM_ATTR_RO(full_scans);
3100 static struct attribute *ksm_attrs[] = {
3101 &sleep_millisecs_attr.attr,
3102 &pages_to_scan_attr.attr,
3104 &pages_shared_attr.attr,
3105 &pages_sharing_attr.attr,
3106 &pages_unshared_attr.attr,
3107 &pages_volatile_attr.attr,
3108 &full_scans_attr.attr,
3110 &merge_across_nodes_attr.attr,
3112 &max_page_sharing_attr.attr,
3113 &stable_node_chains_attr.attr,
3114 &stable_node_dups_attr.attr,
3115 &stable_node_chains_prune_millisecs_attr.attr,
3116 &use_zero_pages_attr.attr,
3120 static const struct attribute_group ksm_attr_group = {
3124 #endif /* CONFIG_SYSFS */
3126 static int __init ksm_init(void)
3128 struct task_struct *ksm_thread;
3131 /* The correct value depends on page size and endianness */
3132 zero_checksum = calc_checksum(ZERO_PAGE(0));
3133 /* Default to false for backwards compatibility */
3134 ksm_use_zero_pages = false;
3136 err = ksm_slab_init();
3140 ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
3141 if (IS_ERR(ksm_thread)) {
3142 pr_err("ksm: creating kthread failed\n");
3143 err = PTR_ERR(ksm_thread);
3148 err = sysfs_create_group(mm_kobj, &ksm_attr_group);
3150 pr_err("ksm: register sysfs failed\n");
3151 kthread_stop(ksm_thread);
3155 ksm_run = KSM_RUN_MERGE; /* no way for user to start it */
3157 #endif /* CONFIG_SYSFS */
3159 #ifdef CONFIG_MEMORY_HOTREMOVE
3160 /* There is no significance to this priority 100 */
3161 hotplug_memory_notifier(ksm_memory_callback, 100);
3170 subsys_initcall(ksm_init);