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/rwsem.h>
24 #include <linux/pagemap.h>
25 #include <linux/rmap.h>
26 #include <linux/spinlock.h>
27 #include <linux/jhash.h>
28 #include <linux/delay.h>
29 #include <linux/kthread.h>
30 #include <linux/wait.h>
31 #include <linux/slab.h>
32 #include <linux/rbtree.h>
33 #include <linux/memory.h>
34 #include <linux/mmu_notifier.h>
35 #include <linux/swap.h>
36 #include <linux/ksm.h>
37 #include <linux/hashtable.h>
38 #include <linux/freezer.h>
39 #include <linux/oom.h>
40 #include <linux/numa.h>
42 #include <asm/tlbflush.h>
47 #define DO_NUMA(x) do { (x); } while (0)
50 #define DO_NUMA(x) do { } while (0)
54 * A few notes about the KSM scanning process,
55 * to make it easier to understand the data structures below:
57 * In order to reduce excessive scanning, KSM sorts the memory pages by their
58 * contents into a data structure that holds pointers to the pages' locations.
60 * Since the contents of the pages may change at any moment, KSM cannot just
61 * insert the pages into a normal sorted tree and expect it to find anything.
62 * Therefore KSM uses two data structures - the stable and the unstable tree.
64 * The stable tree holds pointers to all the merged pages (ksm pages), sorted
65 * by their contents. Because each such page is write-protected, searching on
66 * this tree is fully assured to be working (except when pages are unmapped),
67 * and therefore this tree is called the stable tree.
69 * In addition to the stable tree, KSM uses a second data structure called the
70 * unstable tree: this tree holds pointers to pages which have been found to
71 * be "unchanged for a period of time". The unstable tree sorts these pages
72 * by their contents, but since they are not write-protected, KSM cannot rely
73 * upon the unstable tree to work correctly - the unstable tree is liable to
74 * be corrupted as its contents are modified, and so it is called unstable.
76 * KSM solves this problem by several techniques:
78 * 1) The unstable tree is flushed every time KSM completes scanning all
79 * memory areas, and then the tree is rebuilt again from the beginning.
80 * 2) KSM will only insert into the unstable tree, pages whose hash value
81 * has not changed since the previous scan of all memory areas.
82 * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
83 * colors of the nodes and not on their contents, assuring that even when
84 * the tree gets "corrupted" it won't get out of balance, so scanning time
85 * remains the same (also, searching and inserting nodes in an rbtree uses
86 * the same algorithm, so we have no overhead when we flush and rebuild).
87 * 4) KSM never flushes the stable tree, which means that even if it were to
88 * take 10 attempts to find a page in the unstable tree, once it is found,
89 * it is secured in the stable tree. (When we scan a new page, we first
90 * compare it against the stable tree, and then against the unstable tree.)
92 * If the merge_across_nodes tunable is unset, then KSM maintains multiple
93 * stable trees and multiple unstable trees: one of each for each NUMA node.
97 * struct mm_slot - ksm information per mm that is being scanned
98 * @link: link to the mm_slots hash list
99 * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
100 * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
101 * @mm: the mm that this information is valid for
104 struct hlist_node link;
105 struct list_head mm_list;
106 struct rmap_item *rmap_list;
107 struct mm_struct *mm;
111 * struct ksm_scan - cursor for scanning
112 * @mm_slot: the current mm_slot we are scanning
113 * @address: the next address inside that to be scanned
114 * @rmap_list: link to the next rmap to be scanned in the rmap_list
115 * @seqnr: count of completed full scans (needed when removing unstable node)
117 * There is only the one ksm_scan instance of this cursor structure.
120 struct mm_slot *mm_slot;
121 unsigned long address;
122 struct rmap_item **rmap_list;
127 * struct stable_node - node of the stable rbtree
128 * @node: rb node of this ksm page in the stable tree
129 * @head: (overlaying parent) &migrate_nodes indicates temporarily on that list
130 * @list: linked into migrate_nodes, pending placement in the proper node tree
131 * @hlist: hlist head of rmap_items using this ksm page
132 * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid)
133 * @nid: NUMA node id of stable tree in which linked (may not match kpfn)
137 struct rb_node node; /* when node of stable tree */
138 struct { /* when listed for migration */
139 struct list_head *head;
140 struct list_head list;
143 struct hlist_head hlist;
151 * struct rmap_item - reverse mapping item for virtual addresses
152 * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
153 * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
154 * @nid: NUMA node id of unstable tree in which linked (may not match page)
155 * @mm: the memory structure this rmap_item is pointing into
156 * @address: the virtual address this rmap_item tracks (+ flags in low bits)
157 * @oldchecksum: previous checksum of the page at that virtual address
158 * @node: rb node of this rmap_item in the unstable tree
159 * @head: pointer to stable_node heading this list in the stable tree
160 * @hlist: link into hlist of rmap_items hanging off that stable_node
163 struct rmap_item *rmap_list;
165 struct anon_vma *anon_vma; /* when stable */
167 int nid; /* when node of unstable tree */
170 struct mm_struct *mm;
171 unsigned long address; /* + low bits used for flags below */
172 unsigned int oldchecksum; /* when unstable */
174 struct rb_node node; /* when node of unstable tree */
175 struct { /* when listed from stable tree */
176 struct stable_node *head;
177 struct hlist_node hlist;
182 #define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */
183 #define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */
184 #define STABLE_FLAG 0x200 /* is listed from the stable tree */
186 /* The stable and unstable tree heads */
187 static struct rb_root one_stable_tree[1] = { RB_ROOT };
188 static struct rb_root one_unstable_tree[1] = { RB_ROOT };
189 static struct rb_root *root_stable_tree = one_stable_tree;
190 static struct rb_root *root_unstable_tree = one_unstable_tree;
192 /* Recently migrated nodes of stable tree, pending proper placement */
193 static LIST_HEAD(migrate_nodes);
195 #define MM_SLOTS_HASH_BITS 10
196 static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
198 static struct mm_slot ksm_mm_head = {
199 .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
201 static struct ksm_scan ksm_scan = {
202 .mm_slot = &ksm_mm_head,
205 static struct kmem_cache *rmap_item_cache;
206 static struct kmem_cache *stable_node_cache;
207 static struct kmem_cache *mm_slot_cache;
209 /* The number of nodes in the stable tree */
210 static unsigned long ksm_pages_shared;
212 /* The number of page slots additionally sharing those nodes */
213 static unsigned long ksm_pages_sharing;
215 /* The number of nodes in the unstable tree */
216 static unsigned long ksm_pages_unshared;
218 /* The number of rmap_items in use: to calculate pages_volatile */
219 static unsigned long ksm_rmap_items;
221 /* Number of pages ksmd should scan in one batch */
222 static unsigned int ksm_thread_pages_to_scan = 100;
224 /* Milliseconds ksmd should sleep between batches */
225 static unsigned int ksm_thread_sleep_millisecs = 20;
227 /* Checksum of an empty (zeroed) page */
228 static unsigned int zero_checksum __read_mostly;
230 /* Whether to merge empty (zeroed) pages with actual zero pages */
231 static bool ksm_use_zero_pages __read_mostly;
234 /* Zeroed when merging across nodes is not allowed */
235 static unsigned int ksm_merge_across_nodes = 1;
236 static int ksm_nr_node_ids = 1;
238 #define ksm_merge_across_nodes 1U
239 #define ksm_nr_node_ids 1
242 #define KSM_RUN_STOP 0
243 #define KSM_RUN_MERGE 1
244 #define KSM_RUN_UNMERGE 2
245 #define KSM_RUN_OFFLINE 4
246 static unsigned long ksm_run = KSM_RUN_STOP;
247 static void wait_while_offlining(void);
249 static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
250 static DEFINE_MUTEX(ksm_thread_mutex);
251 static DEFINE_SPINLOCK(ksm_mmlist_lock);
253 #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
254 sizeof(struct __struct), __alignof__(struct __struct),\
257 static int __init ksm_slab_init(void)
259 rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
260 if (!rmap_item_cache)
263 stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
264 if (!stable_node_cache)
267 mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
274 kmem_cache_destroy(stable_node_cache);
276 kmem_cache_destroy(rmap_item_cache);
281 static void __init ksm_slab_free(void)
283 kmem_cache_destroy(mm_slot_cache);
284 kmem_cache_destroy(stable_node_cache);
285 kmem_cache_destroy(rmap_item_cache);
286 mm_slot_cache = NULL;
289 static inline struct rmap_item *alloc_rmap_item(void)
291 struct rmap_item *rmap_item;
293 rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL |
294 __GFP_NORETRY | __GFP_NOWARN);
300 static inline void free_rmap_item(struct rmap_item *rmap_item)
303 rmap_item->mm = NULL; /* debug safety */
304 kmem_cache_free(rmap_item_cache, rmap_item);
307 static inline struct stable_node *alloc_stable_node(void)
310 * The allocation can take too long with GFP_KERNEL when memory is under
311 * pressure, which may lead to hung task warnings. Adding __GFP_HIGH
312 * grants access to memory reserves, helping to avoid this problem.
314 return kmem_cache_alloc(stable_node_cache, GFP_KERNEL | __GFP_HIGH);
317 static inline void free_stable_node(struct stable_node *stable_node)
319 kmem_cache_free(stable_node_cache, stable_node);
322 static inline struct mm_slot *alloc_mm_slot(void)
324 if (!mm_slot_cache) /* initialization failed */
326 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
329 static inline void free_mm_slot(struct mm_slot *mm_slot)
331 kmem_cache_free(mm_slot_cache, mm_slot);
334 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
336 struct mm_slot *slot;
338 hash_for_each_possible(mm_slots_hash, slot, link, (unsigned long)mm)
345 static void insert_to_mm_slots_hash(struct mm_struct *mm,
346 struct mm_slot *mm_slot)
349 hash_add(mm_slots_hash, &mm_slot->link, (unsigned long)mm);
353 * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
354 * page tables after it has passed through ksm_exit() - which, if necessary,
355 * takes mmap_sem briefly to serialize against them. ksm_exit() does not set
356 * a special flag: they can just back out as soon as mm_users goes to zero.
357 * ksm_test_exit() is used throughout to make this test for exit: in some
358 * places for correctness, in some places just to avoid unnecessary work.
360 static inline bool ksm_test_exit(struct mm_struct *mm)
362 return atomic_read(&mm->mm_users) == 0;
366 * We use break_ksm to break COW on a ksm page: it's a stripped down
368 * if (get_user_pages(addr, 1, 1, 1, &page, NULL) == 1)
371 * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
372 * in case the application has unmapped and remapped mm,addr meanwhile.
373 * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP
374 * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
376 * FAULT_FLAG/FOLL_REMOTE are because we do this outside the context
377 * of the process that owns 'vma'. We also do not want to enforce
378 * protection keys here anyway.
380 static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
387 page = follow_page(vma, addr,
388 FOLL_GET | FOLL_MIGRATION | FOLL_REMOTE);
389 if (IS_ERR_OR_NULL(page))
392 ret = handle_mm_fault(vma, addr,
393 FAULT_FLAG_WRITE | FAULT_FLAG_REMOTE);
395 ret = VM_FAULT_WRITE;
397 } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV | VM_FAULT_OOM)));
399 * We must loop because handle_mm_fault() may back out if there's
400 * any difficulty e.g. if pte accessed bit gets updated concurrently.
402 * VM_FAULT_WRITE is what we have been hoping for: it indicates that
403 * COW has been broken, even if the vma does not permit VM_WRITE;
404 * but note that a concurrent fault might break PageKsm for us.
406 * VM_FAULT_SIGBUS could occur if we race with truncation of the
407 * backing file, which also invalidates anonymous pages: that's
408 * okay, that truncation will have unmapped the PageKsm for us.
410 * VM_FAULT_OOM: at the time of writing (late July 2009), setting
411 * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
412 * current task has TIF_MEMDIE set, and will be OOM killed on return
413 * to user; and ksmd, having no mm, would never be chosen for that.
415 * But if the mm is in a limited mem_cgroup, then the fault may fail
416 * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
417 * even ksmd can fail in this way - though it's usually breaking ksm
418 * just to undo a merge it made a moment before, so unlikely to oom.
420 * That's a pity: we might therefore have more kernel pages allocated
421 * than we're counting as nodes in the stable tree; but ksm_do_scan
422 * will retry to break_cow on each pass, so should recover the page
423 * in due course. The important thing is to not let VM_MERGEABLE
424 * be cleared while any such pages might remain in the area.
426 return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
429 static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm,
432 struct vm_area_struct *vma;
433 if (ksm_test_exit(mm))
435 vma = find_vma(mm, addr);
436 if (!vma || vma->vm_start > addr)
438 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
443 static void break_cow(struct rmap_item *rmap_item)
445 struct mm_struct *mm = rmap_item->mm;
446 unsigned long addr = rmap_item->address;
447 struct vm_area_struct *vma;
450 * It is not an accident that whenever we want to break COW
451 * to undo, we also need to drop a reference to the anon_vma.
453 put_anon_vma(rmap_item->anon_vma);
455 down_read(&mm->mmap_sem);
456 vma = find_mergeable_vma(mm, addr);
458 break_ksm(vma, addr);
459 up_read(&mm->mmap_sem);
462 static struct page *get_mergeable_page(struct rmap_item *rmap_item)
464 struct mm_struct *mm = rmap_item->mm;
465 unsigned long addr = rmap_item->address;
466 struct vm_area_struct *vma;
469 down_read(&mm->mmap_sem);
470 vma = find_mergeable_vma(mm, addr);
474 page = follow_page(vma, addr, FOLL_GET);
475 if (IS_ERR_OR_NULL(page))
477 if (PageAnon(page)) {
478 flush_anon_page(vma, page, addr);
479 flush_dcache_page(page);
485 up_read(&mm->mmap_sem);
490 * This helper is used for getting right index into array of tree roots.
491 * When merge_across_nodes knob is set to 1, there are only two rb-trees for
492 * stable and unstable pages from all nodes with roots in index 0. Otherwise,
493 * every node has its own stable and unstable tree.
495 static inline int get_kpfn_nid(unsigned long kpfn)
497 return ksm_merge_across_nodes ? 0 : NUMA(pfn_to_nid(kpfn));
500 static void remove_node_from_stable_tree(struct stable_node *stable_node)
502 struct rmap_item *rmap_item;
504 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
505 if (rmap_item->hlist.next)
509 put_anon_vma(rmap_item->anon_vma);
510 rmap_item->address &= PAGE_MASK;
514 if (stable_node->head == &migrate_nodes)
515 list_del(&stable_node->list);
517 rb_erase(&stable_node->node,
518 root_stable_tree + NUMA(stable_node->nid));
519 free_stable_node(stable_node);
523 * get_ksm_page: checks if the page indicated by the stable node
524 * is still its ksm page, despite having held no reference to it.
525 * In which case we can trust the content of the page, and it
526 * returns the gotten page; but if the page has now been zapped,
527 * remove the stale node from the stable tree and return NULL.
528 * But beware, the stable node's page might be being migrated.
530 * You would expect the stable_node to hold a reference to the ksm page.
531 * But if it increments the page's count, swapping out has to wait for
532 * ksmd to come around again before it can free the page, which may take
533 * seconds or even minutes: much too unresponsive. So instead we use a
534 * "keyhole reference": access to the ksm page from the stable node peeps
535 * out through its keyhole to see if that page still holds the right key,
536 * pointing back to this stable node. This relies on freeing a PageAnon
537 * page to reset its page->mapping to NULL, and relies on no other use of
538 * a page to put something that might look like our key in page->mapping.
539 * is on its way to being freed; but it is an anomaly to bear in mind.
541 static struct page *get_ksm_page(struct stable_node *stable_node, bool lock_it)
544 void *expected_mapping;
547 expected_mapping = (void *)((unsigned long)stable_node |
550 kpfn = READ_ONCE(stable_node->kpfn);
551 page = pfn_to_page(kpfn);
554 * page is computed from kpfn, so on most architectures reading
555 * page->mapping is naturally ordered after reading node->kpfn,
556 * but on Alpha we need to be more careful.
558 smp_read_barrier_depends();
559 if (READ_ONCE(page->mapping) != expected_mapping)
563 * We cannot do anything with the page while its refcount is 0.
564 * Usually 0 means free, or tail of a higher-order page: in which
565 * case this node is no longer referenced, and should be freed;
566 * however, it might mean that the page is under page_freeze_refs().
567 * The __remove_mapping() case is easy, again the node is now stale;
568 * but if page is swapcache in migrate_page_move_mapping(), it might
569 * still be our page, in which case it's essential to keep the node.
571 while (!get_page_unless_zero(page)) {
573 * Another check for page->mapping != expected_mapping would
574 * work here too. We have chosen the !PageSwapCache test to
575 * optimize the common case, when the page is or is about to
576 * be freed: PageSwapCache is cleared (under spin_lock_irq)
577 * in the freeze_refs section of __remove_mapping(); but Anon
578 * page->mapping reset to NULL later, in free_pages_prepare().
580 if (!PageSwapCache(page))
585 if (READ_ONCE(page->mapping) != expected_mapping) {
592 if (READ_ONCE(page->mapping) != expected_mapping) {
602 * We come here from above when page->mapping or !PageSwapCache
603 * suggests that the node is stale; but it might be under migration.
604 * We need smp_rmb(), matching the smp_wmb() in ksm_migrate_page(),
605 * before checking whether node->kpfn has been changed.
608 if (READ_ONCE(stable_node->kpfn) != kpfn)
610 remove_node_from_stable_tree(stable_node);
615 * Removing rmap_item from stable or unstable tree.
616 * This function will clean the information from the stable/unstable tree.
618 static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
620 if (rmap_item->address & STABLE_FLAG) {
621 struct stable_node *stable_node;
624 stable_node = rmap_item->head;
625 page = get_ksm_page(stable_node, true);
629 hlist_del(&rmap_item->hlist);
633 if (!hlist_empty(&stable_node->hlist))
638 put_anon_vma(rmap_item->anon_vma);
639 rmap_item->address &= PAGE_MASK;
641 } else if (rmap_item->address & UNSTABLE_FLAG) {
644 * Usually ksmd can and must skip the rb_erase, because
645 * root_unstable_tree was already reset to RB_ROOT.
646 * But be careful when an mm is exiting: do the rb_erase
647 * if this rmap_item was inserted by this scan, rather
648 * than left over from before.
650 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
653 rb_erase(&rmap_item->node,
654 root_unstable_tree + NUMA(rmap_item->nid));
655 ksm_pages_unshared--;
656 rmap_item->address &= PAGE_MASK;
659 cond_resched(); /* we're called from many long loops */
662 static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
663 struct rmap_item **rmap_list)
666 struct rmap_item *rmap_item = *rmap_list;
667 *rmap_list = rmap_item->rmap_list;
668 remove_rmap_item_from_tree(rmap_item);
669 free_rmap_item(rmap_item);
674 * Though it's very tempting to unmerge rmap_items from stable tree rather
675 * than check every pte of a given vma, the locking doesn't quite work for
676 * that - an rmap_item is assigned to the stable tree after inserting ksm
677 * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing
678 * rmap_items from parent to child at fork time (so as not to waste time
679 * if exit comes before the next scan reaches it).
681 * Similarly, although we'd like to remove rmap_items (so updating counts
682 * and freeing memory) when unmerging an area, it's easier to leave that
683 * to the next pass of ksmd - consider, for example, how ksmd might be
684 * in cmp_and_merge_page on one of the rmap_items we would be removing.
686 static int unmerge_ksm_pages(struct vm_area_struct *vma,
687 unsigned long start, unsigned long end)
692 for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
693 if (ksm_test_exit(vma->vm_mm))
695 if (signal_pending(current))
698 err = break_ksm(vma, addr);
705 * Only called through the sysfs control interface:
707 static int remove_stable_node(struct stable_node *stable_node)
712 page = get_ksm_page(stable_node, true);
715 * get_ksm_page did remove_node_from_stable_tree itself.
720 if (WARN_ON_ONCE(page_mapped(page))) {
722 * This should not happen: but if it does, just refuse to let
723 * merge_across_nodes be switched - there is no need to panic.
728 * The stable node did not yet appear stale to get_ksm_page(),
729 * since that allows for an unmapped ksm page to be recognized
730 * right up until it is freed; but the node is safe to remove.
731 * This page might be in a pagevec waiting to be freed,
732 * or it might be PageSwapCache (perhaps under writeback),
733 * or it might have been removed from swapcache a moment ago.
735 set_page_stable_node(page, NULL);
736 remove_node_from_stable_tree(stable_node);
745 static int remove_all_stable_nodes(void)
747 struct stable_node *stable_node, *next;
751 for (nid = 0; nid < ksm_nr_node_ids; nid++) {
752 while (root_stable_tree[nid].rb_node) {
753 stable_node = rb_entry(root_stable_tree[nid].rb_node,
754 struct stable_node, node);
755 if (remove_stable_node(stable_node)) {
757 break; /* proceed to next nid */
762 list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
763 if (remove_stable_node(stable_node))
770 static int unmerge_and_remove_all_rmap_items(void)
772 struct mm_slot *mm_slot;
773 struct mm_struct *mm;
774 struct vm_area_struct *vma;
777 spin_lock(&ksm_mmlist_lock);
778 ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
779 struct mm_slot, mm_list);
780 spin_unlock(&ksm_mmlist_lock);
782 for (mm_slot = ksm_scan.mm_slot;
783 mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
785 down_read(&mm->mmap_sem);
786 for (vma = mm->mmap; vma; vma = vma->vm_next) {
787 if (ksm_test_exit(mm))
789 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
791 err = unmerge_ksm_pages(vma,
792 vma->vm_start, vma->vm_end);
797 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
798 up_read(&mm->mmap_sem);
800 spin_lock(&ksm_mmlist_lock);
801 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
802 struct mm_slot, mm_list);
803 if (ksm_test_exit(mm)) {
804 hash_del(&mm_slot->link);
805 list_del(&mm_slot->mm_list);
806 spin_unlock(&ksm_mmlist_lock);
808 free_mm_slot(mm_slot);
809 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
812 spin_unlock(&ksm_mmlist_lock);
815 /* Clean up stable nodes, but don't worry if some are still busy */
816 remove_all_stable_nodes();
821 up_read(&mm->mmap_sem);
822 spin_lock(&ksm_mmlist_lock);
823 ksm_scan.mm_slot = &ksm_mm_head;
824 spin_unlock(&ksm_mmlist_lock);
827 #endif /* CONFIG_SYSFS */
829 static u32 calc_checksum(struct page *page)
832 void *addr = kmap_atomic(page);
833 checksum = jhash2(addr, PAGE_SIZE / 4, 17);
838 static int memcmp_pages(struct page *page1, struct page *page2)
843 addr1 = kmap_atomic(page1);
844 addr2 = kmap_atomic(page2);
845 ret = memcmp(addr1, addr2, PAGE_SIZE);
846 kunmap_atomic(addr2);
847 kunmap_atomic(addr1);
851 static inline int pages_identical(struct page *page1, struct page *page2)
853 return !memcmp_pages(page1, page2);
856 static int write_protect_page(struct vm_area_struct *vma, struct page *page,
859 struct mm_struct *mm = vma->vm_mm;
860 struct page_vma_mapped_walk pvmw = {
866 unsigned long mmun_start; /* For mmu_notifiers */
867 unsigned long mmun_end; /* For mmu_notifiers */
869 pvmw.address = page_address_in_vma(page, vma);
870 if (pvmw.address == -EFAULT)
873 BUG_ON(PageTransCompound(page));
875 mmun_start = pvmw.address;
876 mmun_end = pvmw.address + PAGE_SIZE;
877 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
879 if (!page_vma_mapped_walk(&pvmw))
881 if (WARN_ONCE(!pvmw.pte, "Unexpected PMD mapping?"))
884 if (pte_write(*pvmw.pte) || pte_dirty(*pvmw.pte) ||
885 (pte_protnone(*pvmw.pte) && pte_savedwrite(*pvmw.pte))) {
888 swapped = PageSwapCache(page);
889 flush_cache_page(vma, pvmw.address, page_to_pfn(page));
891 * Ok this is tricky, when get_user_pages_fast() run it doesn't
892 * take any lock, therefore the check that we are going to make
893 * with the pagecount against the mapcount is racey and
894 * O_DIRECT can happen right after the check.
895 * So we clear the pte and flush the tlb before the check
896 * this assure us that no O_DIRECT can happen after the check
897 * or in the middle of the check.
899 entry = ptep_clear_flush_notify(vma, pvmw.address, pvmw.pte);
901 * Check that no O_DIRECT or similar I/O is in progress on the
904 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
905 set_pte_at(mm, pvmw.address, pvmw.pte, entry);
908 if (pte_dirty(entry))
909 set_page_dirty(page);
911 if (pte_protnone(entry))
912 entry = pte_mkclean(pte_clear_savedwrite(entry));
914 entry = pte_mkclean(pte_wrprotect(entry));
915 set_pte_at_notify(mm, pvmw.address, pvmw.pte, entry);
917 *orig_pte = *pvmw.pte;
921 page_vma_mapped_walk_done(&pvmw);
923 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
929 * replace_page - replace page in vma by new ksm page
930 * @vma: vma that holds the pte pointing to page
931 * @page: the page we are replacing by kpage
932 * @kpage: the ksm page we replace page by
933 * @orig_pte: the original value of the pte
935 * Returns 0 on success, -EFAULT on failure.
937 static int replace_page(struct vm_area_struct *vma, struct page *page,
938 struct page *kpage, pte_t orig_pte)
940 struct mm_struct *mm = vma->vm_mm;
947 unsigned long mmun_start; /* For mmu_notifiers */
948 unsigned long mmun_end; /* For mmu_notifiers */
950 addr = page_address_in_vma(page, vma);
954 pmd = mm_find_pmd(mm, addr);
959 mmun_end = addr + PAGE_SIZE;
960 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
962 ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
963 if (!pte_same(*ptep, orig_pte)) {
964 pte_unmap_unlock(ptep, ptl);
969 * No need to check ksm_use_zero_pages here: we can only have a
970 * zero_page here if ksm_use_zero_pages was enabled alreaady.
972 if (!is_zero_pfn(page_to_pfn(kpage))) {
974 page_add_anon_rmap(kpage, vma, addr, false);
975 newpte = mk_pte(kpage, vma->vm_page_prot);
977 newpte = pte_mkspecial(pfn_pte(page_to_pfn(kpage),
981 flush_cache_page(vma, addr, pte_pfn(*ptep));
982 ptep_clear_flush_notify(vma, addr, ptep);
983 set_pte_at_notify(mm, addr, ptep, newpte);
985 page_remove_rmap(page, false);
986 if (!page_mapped(page))
987 try_to_free_swap(page);
990 pte_unmap_unlock(ptep, ptl);
993 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
999 * try_to_merge_one_page - take two pages and merge them into one
1000 * @vma: the vma that holds the pte pointing to page
1001 * @page: the PageAnon page that we want to replace with kpage
1002 * @kpage: the PageKsm page that we want to map instead of page,
1003 * or NULL the first time when we want to use page as kpage.
1005 * This function returns 0 if the pages were merged, -EFAULT otherwise.
1007 static int try_to_merge_one_page(struct vm_area_struct *vma,
1008 struct page *page, struct page *kpage)
1010 pte_t orig_pte = __pte(0);
1013 if (page == kpage) /* ksm page forked */
1016 if (!PageAnon(page))
1020 * We need the page lock to read a stable PageSwapCache in
1021 * write_protect_page(). We use trylock_page() instead of
1022 * lock_page() because we don't want to wait here - we
1023 * prefer to continue scanning and merging different pages,
1024 * then come back to this page when it is unlocked.
1026 if (!trylock_page(page))
1029 if (PageTransCompound(page)) {
1030 err = split_huge_page(page);
1036 * If this anonymous page is mapped only here, its pte may need
1037 * to be write-protected. If it's mapped elsewhere, all of its
1038 * ptes are necessarily already write-protected. But in either
1039 * case, we need to lock and check page_count is not raised.
1041 if (write_protect_page(vma, page, &orig_pte) == 0) {
1044 * While we hold page lock, upgrade page from
1045 * PageAnon+anon_vma to PageKsm+NULL stable_node:
1046 * stable_tree_insert() will update stable_node.
1048 set_page_stable_node(page, NULL);
1049 mark_page_accessed(page);
1051 * Page reclaim just frees a clean page with no dirty
1052 * ptes: make sure that the ksm page would be swapped.
1054 if (!PageDirty(page))
1057 } else if (pages_identical(page, kpage))
1058 err = replace_page(vma, page, kpage, orig_pte);
1061 if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
1062 munlock_vma_page(page);
1063 if (!PageMlocked(kpage)) {
1066 mlock_vma_page(kpage);
1067 page = kpage; /* for final unlock */
1078 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
1079 * but no new kernel page is allocated: kpage must already be a ksm page.
1081 * This function returns 0 if the pages were merged, -EFAULT otherwise.
1083 static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
1084 struct page *page, struct page *kpage)
1086 struct mm_struct *mm = rmap_item->mm;
1087 struct vm_area_struct *vma;
1090 down_read(&mm->mmap_sem);
1091 vma = find_mergeable_vma(mm, rmap_item->address);
1095 err = try_to_merge_one_page(vma, page, kpage);
1099 /* Unstable nid is in union with stable anon_vma: remove first */
1100 remove_rmap_item_from_tree(rmap_item);
1102 /* Must get reference to anon_vma while still holding mmap_sem */
1103 rmap_item->anon_vma = vma->anon_vma;
1104 get_anon_vma(vma->anon_vma);
1106 up_read(&mm->mmap_sem);
1111 * try_to_merge_two_pages - take two identical pages and prepare them
1112 * to be merged into one page.
1114 * This function returns the kpage if we successfully merged two identical
1115 * pages into one ksm page, NULL otherwise.
1117 * Note that this function upgrades page to ksm page: if one of the pages
1118 * is already a ksm page, try_to_merge_with_ksm_page should be used.
1120 static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
1122 struct rmap_item *tree_rmap_item,
1123 struct page *tree_page)
1127 err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
1129 err = try_to_merge_with_ksm_page(tree_rmap_item,
1132 * If that fails, we have a ksm page with only one pte
1133 * pointing to it: so break it.
1136 break_cow(rmap_item);
1138 return err ? NULL : page;
1142 * stable_tree_search - search for page inside the stable tree
1144 * This function checks if there is a page inside the stable tree
1145 * with identical content to the page that we are scanning right now.
1147 * This function returns the stable tree node of identical content if found,
1150 static struct page *stable_tree_search(struct page *page)
1153 struct rb_root *root;
1154 struct rb_node **new;
1155 struct rb_node *parent;
1156 struct stable_node *stable_node;
1157 struct stable_node *page_node;
1159 page_node = page_stable_node(page);
1160 if (page_node && page_node->head != &migrate_nodes) {
1161 /* ksm page forked */
1166 nid = get_kpfn_nid(page_to_pfn(page));
1167 root = root_stable_tree + nid;
1169 new = &root->rb_node;
1173 struct page *tree_page;
1177 stable_node = rb_entry(*new, struct stable_node, node);
1178 tree_page = get_ksm_page(stable_node, false);
1181 * If we walked over a stale stable_node,
1182 * get_ksm_page() will call rb_erase() and it
1183 * may rebalance the tree from under us. So
1184 * restart the search from scratch. Returning
1185 * NULL would be safe too, but we'd generate
1186 * false negative insertions just because some
1187 * stable_node was stale.
1192 ret = memcmp_pages(page, tree_page);
1193 put_page(tree_page);
1197 new = &parent->rb_left;
1199 new = &parent->rb_right;
1202 * Lock and unlock the stable_node's page (which
1203 * might already have been migrated) so that page
1204 * migration is sure to notice its raised count.
1205 * It would be more elegant to return stable_node
1206 * than kpage, but that involves more changes.
1208 tree_page = get_ksm_page(stable_node, true);
1210 unlock_page(tree_page);
1211 if (get_kpfn_nid(stable_node->kpfn) !=
1212 NUMA(stable_node->nid)) {
1213 put_page(tree_page);
1219 * There is now a place for page_node, but the tree may
1220 * have been rebalanced, so re-evaluate parent and new.
1231 list_del(&page_node->list);
1232 DO_NUMA(page_node->nid = nid);
1233 rb_link_node(&page_node->node, parent, new);
1234 rb_insert_color(&page_node->node, root);
1240 list_del(&page_node->list);
1241 DO_NUMA(page_node->nid = nid);
1242 rb_replace_node(&stable_node->node, &page_node->node, root);
1245 rb_erase(&stable_node->node, root);
1248 stable_node->head = &migrate_nodes;
1249 list_add(&stable_node->list, stable_node->head);
1254 * stable_tree_insert - insert stable tree node pointing to new ksm page
1255 * into the stable tree.
1257 * This function returns the stable tree node just allocated on success,
1260 static struct stable_node *stable_tree_insert(struct page *kpage)
1264 struct rb_root *root;
1265 struct rb_node **new;
1266 struct rb_node *parent;
1267 struct stable_node *stable_node;
1269 kpfn = page_to_pfn(kpage);
1270 nid = get_kpfn_nid(kpfn);
1271 root = root_stable_tree + nid;
1274 new = &root->rb_node;
1277 struct page *tree_page;
1281 stable_node = rb_entry(*new, struct stable_node, node);
1282 tree_page = get_ksm_page(stable_node, false);
1285 * If we walked over a stale stable_node,
1286 * get_ksm_page() will call rb_erase() and it
1287 * may rebalance the tree from under us. So
1288 * restart the search from scratch. Returning
1289 * NULL would be safe too, but we'd generate
1290 * false negative insertions just because some
1291 * stable_node was stale.
1296 ret = memcmp_pages(kpage, tree_page);
1297 put_page(tree_page);
1301 new = &parent->rb_left;
1303 new = &parent->rb_right;
1306 * It is not a bug that stable_tree_search() didn't
1307 * find this node: because at that time our page was
1308 * not yet write-protected, so may have changed since.
1314 stable_node = alloc_stable_node();
1318 INIT_HLIST_HEAD(&stable_node->hlist);
1319 stable_node->kpfn = kpfn;
1320 set_page_stable_node(kpage, stable_node);
1321 DO_NUMA(stable_node->nid = nid);
1322 rb_link_node(&stable_node->node, parent, new);
1323 rb_insert_color(&stable_node->node, root);
1329 * unstable_tree_search_insert - search for identical page,
1330 * else insert rmap_item into the unstable tree.
1332 * This function searches for a page in the unstable tree identical to the
1333 * page currently being scanned; and if no identical page is found in the
1334 * tree, we insert rmap_item as a new object into the unstable tree.
1336 * This function returns pointer to rmap_item found to be identical
1337 * to the currently scanned page, NULL otherwise.
1339 * This function does both searching and inserting, because they share
1340 * the same walking algorithm in an rbtree.
1343 struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1345 struct page **tree_pagep)
1347 struct rb_node **new;
1348 struct rb_root *root;
1349 struct rb_node *parent = NULL;
1352 nid = get_kpfn_nid(page_to_pfn(page));
1353 root = root_unstable_tree + nid;
1354 new = &root->rb_node;
1357 struct rmap_item *tree_rmap_item;
1358 struct page *tree_page;
1362 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
1363 tree_page = get_mergeable_page(tree_rmap_item);
1368 * Don't substitute a ksm page for a forked page.
1370 if (page == tree_page) {
1371 put_page(tree_page);
1375 ret = memcmp_pages(page, tree_page);
1379 put_page(tree_page);
1380 new = &parent->rb_left;
1381 } else if (ret > 0) {
1382 put_page(tree_page);
1383 new = &parent->rb_right;
1384 } else if (!ksm_merge_across_nodes &&
1385 page_to_nid(tree_page) != nid) {
1387 * If tree_page has been migrated to another NUMA node,
1388 * it will be flushed out and put in the right unstable
1389 * tree next time: only merge with it when across_nodes.
1391 put_page(tree_page);
1394 *tree_pagep = tree_page;
1395 return tree_rmap_item;
1399 rmap_item->address |= UNSTABLE_FLAG;
1400 rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1401 DO_NUMA(rmap_item->nid = nid);
1402 rb_link_node(&rmap_item->node, parent, new);
1403 rb_insert_color(&rmap_item->node, root);
1405 ksm_pages_unshared++;
1410 * stable_tree_append - add another rmap_item to the linked list of
1411 * rmap_items hanging off a given node of the stable tree, all sharing
1412 * the same ksm page.
1414 static void stable_tree_append(struct rmap_item *rmap_item,
1415 struct stable_node *stable_node)
1417 rmap_item->head = stable_node;
1418 rmap_item->address |= STABLE_FLAG;
1419 hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
1421 if (rmap_item->hlist.next)
1422 ksm_pages_sharing++;
1428 * cmp_and_merge_page - first see if page can be merged into the stable tree;
1429 * if not, compare checksum to previous and if it's the same, see if page can
1430 * be inserted into the unstable tree, or merged with a page already there and
1431 * both transferred to the stable tree.
1433 * @page: the page that we are searching identical page to.
1434 * @rmap_item: the reverse mapping into the virtual address of this page
1436 static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1438 struct rmap_item *tree_rmap_item;
1439 struct page *tree_page = NULL;
1440 struct stable_node *stable_node;
1442 unsigned int checksum;
1445 stable_node = page_stable_node(page);
1447 if (stable_node->head != &migrate_nodes &&
1448 get_kpfn_nid(stable_node->kpfn) != NUMA(stable_node->nid)) {
1449 rb_erase(&stable_node->node,
1450 root_stable_tree + NUMA(stable_node->nid));
1451 stable_node->head = &migrate_nodes;
1452 list_add(&stable_node->list, stable_node->head);
1454 if (stable_node->head != &migrate_nodes &&
1455 rmap_item->head == stable_node)
1459 /* We first start with searching the page inside the stable tree */
1460 kpage = stable_tree_search(page);
1461 if (kpage == page && rmap_item->head == stable_node) {
1466 remove_rmap_item_from_tree(rmap_item);
1469 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
1472 * The page was successfully merged:
1473 * add its rmap_item to the stable tree.
1476 stable_tree_append(rmap_item, page_stable_node(kpage));
1484 * If the hash value of the page has changed from the last time
1485 * we calculated it, this page is changing frequently: therefore we
1486 * don't want to insert it in the unstable tree, and we don't want
1487 * to waste our time searching for something identical to it there.
1489 checksum = calc_checksum(page);
1490 if (rmap_item->oldchecksum != checksum) {
1491 rmap_item->oldchecksum = checksum;
1496 * Same checksum as an empty page. We attempt to merge it with the
1497 * appropriate zero page if the user enabled this via sysfs.
1499 if (ksm_use_zero_pages && (checksum == zero_checksum)) {
1500 struct vm_area_struct *vma;
1502 vma = find_mergeable_vma(rmap_item->mm, rmap_item->address);
1503 err = try_to_merge_one_page(vma, page,
1504 ZERO_PAGE(rmap_item->address));
1506 * In case of failure, the page was not really empty, so we
1507 * need to continue. Otherwise we're done.
1513 unstable_tree_search_insert(rmap_item, page, &tree_page);
1514 if (tree_rmap_item) {
1515 kpage = try_to_merge_two_pages(rmap_item, page,
1516 tree_rmap_item, tree_page);
1517 put_page(tree_page);
1520 * The pages were successfully merged: insert new
1521 * node in the stable tree and add both rmap_items.
1524 stable_node = stable_tree_insert(kpage);
1526 stable_tree_append(tree_rmap_item, stable_node);
1527 stable_tree_append(rmap_item, stable_node);
1532 * If we fail to insert the page into the stable tree,
1533 * we will have 2 virtual addresses that are pointing
1534 * to a ksm page left outside the stable tree,
1535 * in which case we need to break_cow on both.
1538 break_cow(tree_rmap_item);
1539 break_cow(rmap_item);
1545 static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
1546 struct rmap_item **rmap_list,
1549 struct rmap_item *rmap_item;
1551 while (*rmap_list) {
1552 rmap_item = *rmap_list;
1553 if ((rmap_item->address & PAGE_MASK) == addr)
1555 if (rmap_item->address > addr)
1557 *rmap_list = rmap_item->rmap_list;
1558 remove_rmap_item_from_tree(rmap_item);
1559 free_rmap_item(rmap_item);
1562 rmap_item = alloc_rmap_item();
1564 /* It has already been zeroed */
1565 rmap_item->mm = mm_slot->mm;
1566 rmap_item->address = addr;
1567 rmap_item->rmap_list = *rmap_list;
1568 *rmap_list = rmap_item;
1573 static struct rmap_item *scan_get_next_rmap_item(struct page **page)
1575 struct mm_struct *mm;
1576 struct mm_slot *slot;
1577 struct vm_area_struct *vma;
1578 struct rmap_item *rmap_item;
1581 if (list_empty(&ksm_mm_head.mm_list))
1584 slot = ksm_scan.mm_slot;
1585 if (slot == &ksm_mm_head) {
1587 * A number of pages can hang around indefinitely on per-cpu
1588 * pagevecs, raised page count preventing write_protect_page
1589 * from merging them. Though it doesn't really matter much,
1590 * it is puzzling to see some stuck in pages_volatile until
1591 * other activity jostles them out, and they also prevented
1592 * LTP's KSM test from succeeding deterministically; so drain
1593 * them here (here rather than on entry to ksm_do_scan(),
1594 * so we don't IPI too often when pages_to_scan is set low).
1596 lru_add_drain_all();
1599 * Whereas stale stable_nodes on the stable_tree itself
1600 * get pruned in the regular course of stable_tree_search(),
1601 * those moved out to the migrate_nodes list can accumulate:
1602 * so prune them once before each full scan.
1604 if (!ksm_merge_across_nodes) {
1605 struct stable_node *stable_node, *next;
1608 list_for_each_entry_safe(stable_node, next,
1609 &migrate_nodes, list) {
1610 page = get_ksm_page(stable_node, false);
1617 for (nid = 0; nid < ksm_nr_node_ids; nid++)
1618 root_unstable_tree[nid] = RB_ROOT;
1620 spin_lock(&ksm_mmlist_lock);
1621 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1622 ksm_scan.mm_slot = slot;
1623 spin_unlock(&ksm_mmlist_lock);
1625 * Although we tested list_empty() above, a racing __ksm_exit
1626 * of the last mm on the list may have removed it since then.
1628 if (slot == &ksm_mm_head)
1631 ksm_scan.address = 0;
1632 ksm_scan.rmap_list = &slot->rmap_list;
1636 down_read(&mm->mmap_sem);
1637 if (ksm_test_exit(mm))
1640 vma = find_vma(mm, ksm_scan.address);
1642 for (; vma; vma = vma->vm_next) {
1643 if (!(vma->vm_flags & VM_MERGEABLE))
1645 if (ksm_scan.address < vma->vm_start)
1646 ksm_scan.address = vma->vm_start;
1648 ksm_scan.address = vma->vm_end;
1650 while (ksm_scan.address < vma->vm_end) {
1651 if (ksm_test_exit(mm))
1653 *page = follow_page(vma, ksm_scan.address, FOLL_GET);
1654 if (IS_ERR_OR_NULL(*page)) {
1655 ksm_scan.address += PAGE_SIZE;
1659 if (PageAnon(*page)) {
1660 flush_anon_page(vma, *page, ksm_scan.address);
1661 flush_dcache_page(*page);
1662 rmap_item = get_next_rmap_item(slot,
1663 ksm_scan.rmap_list, ksm_scan.address);
1665 ksm_scan.rmap_list =
1666 &rmap_item->rmap_list;
1667 ksm_scan.address += PAGE_SIZE;
1670 up_read(&mm->mmap_sem);
1674 ksm_scan.address += PAGE_SIZE;
1679 if (ksm_test_exit(mm)) {
1680 ksm_scan.address = 0;
1681 ksm_scan.rmap_list = &slot->rmap_list;
1684 * Nuke all the rmap_items that are above this current rmap:
1685 * because there were no VM_MERGEABLE vmas with such addresses.
1687 remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
1689 spin_lock(&ksm_mmlist_lock);
1690 ksm_scan.mm_slot = list_entry(slot->mm_list.next,
1691 struct mm_slot, mm_list);
1692 if (ksm_scan.address == 0) {
1694 * We've completed a full scan of all vmas, holding mmap_sem
1695 * throughout, and found no VM_MERGEABLE: so do the same as
1696 * __ksm_exit does to remove this mm from all our lists now.
1697 * This applies either when cleaning up after __ksm_exit
1698 * (but beware: we can reach here even before __ksm_exit),
1699 * or when all VM_MERGEABLE areas have been unmapped (and
1700 * mmap_sem then protects against race with MADV_MERGEABLE).
1702 hash_del(&slot->link);
1703 list_del(&slot->mm_list);
1704 spin_unlock(&ksm_mmlist_lock);
1707 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1708 up_read(&mm->mmap_sem);
1711 up_read(&mm->mmap_sem);
1713 * up_read(&mm->mmap_sem) first because after
1714 * spin_unlock(&ksm_mmlist_lock) run, the "mm" may
1715 * already have been freed under us by __ksm_exit()
1716 * because the "mm_slot" is still hashed and
1717 * ksm_scan.mm_slot doesn't point to it anymore.
1719 spin_unlock(&ksm_mmlist_lock);
1722 /* Repeat until we've completed scanning the whole list */
1723 slot = ksm_scan.mm_slot;
1724 if (slot != &ksm_mm_head)
1732 * ksm_do_scan - the ksm scanner main worker function.
1733 * @scan_npages - number of pages we want to scan before we return.
1735 static void ksm_do_scan(unsigned int scan_npages)
1737 struct rmap_item *rmap_item;
1738 struct page *uninitialized_var(page);
1740 while (scan_npages-- && likely(!freezing(current))) {
1742 rmap_item = scan_get_next_rmap_item(&page);
1745 cmp_and_merge_page(page, rmap_item);
1750 static int ksmd_should_run(void)
1752 return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
1755 static int ksm_scan_thread(void *nothing)
1758 set_user_nice(current, 5);
1760 while (!kthread_should_stop()) {
1761 mutex_lock(&ksm_thread_mutex);
1762 wait_while_offlining();
1763 if (ksmd_should_run())
1764 ksm_do_scan(ksm_thread_pages_to_scan);
1765 mutex_unlock(&ksm_thread_mutex);
1769 if (ksmd_should_run()) {
1770 schedule_timeout_interruptible(
1771 msecs_to_jiffies(ksm_thread_sleep_millisecs));
1773 wait_event_freezable(ksm_thread_wait,
1774 ksmd_should_run() || kthread_should_stop());
1780 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1781 unsigned long end, int advice, unsigned long *vm_flags)
1783 struct mm_struct *mm = vma->vm_mm;
1787 case MADV_MERGEABLE:
1789 * Be somewhat over-protective for now!
1791 if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE |
1792 VM_PFNMAP | VM_IO | VM_DONTEXPAND |
1793 VM_HUGETLB | VM_MIXEDMAP))
1794 return 0; /* just ignore the advice */
1797 if (*vm_flags & VM_SAO)
1801 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
1802 err = __ksm_enter(mm);
1807 *vm_flags |= VM_MERGEABLE;
1810 case MADV_UNMERGEABLE:
1811 if (!(*vm_flags & VM_MERGEABLE))
1812 return 0; /* just ignore the advice */
1814 if (vma->anon_vma) {
1815 err = unmerge_ksm_pages(vma, start, end);
1820 *vm_flags &= ~VM_MERGEABLE;
1827 int __ksm_enter(struct mm_struct *mm)
1829 struct mm_slot *mm_slot;
1832 mm_slot = alloc_mm_slot();
1836 /* Check ksm_run too? Would need tighter locking */
1837 needs_wakeup = list_empty(&ksm_mm_head.mm_list);
1839 spin_lock(&ksm_mmlist_lock);
1840 insert_to_mm_slots_hash(mm, mm_slot);
1842 * When KSM_RUN_MERGE (or KSM_RUN_STOP),
1843 * insert just behind the scanning cursor, to let the area settle
1844 * down a little; when fork is followed by immediate exec, we don't
1845 * want ksmd to waste time setting up and tearing down an rmap_list.
1847 * But when KSM_RUN_UNMERGE, it's important to insert ahead of its
1848 * scanning cursor, otherwise KSM pages in newly forked mms will be
1849 * missed: then we might as well insert at the end of the list.
1851 if (ksm_run & KSM_RUN_UNMERGE)
1852 list_add_tail(&mm_slot->mm_list, &ksm_mm_head.mm_list);
1854 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
1855 spin_unlock(&ksm_mmlist_lock);
1857 set_bit(MMF_VM_MERGEABLE, &mm->flags);
1861 wake_up_interruptible(&ksm_thread_wait);
1866 void __ksm_exit(struct mm_struct *mm)
1868 struct mm_slot *mm_slot;
1869 int easy_to_free = 0;
1872 * This process is exiting: if it's straightforward (as is the
1873 * case when ksmd was never running), free mm_slot immediately.
1874 * But if it's at the cursor or has rmap_items linked to it, use
1875 * mmap_sem to synchronize with any break_cows before pagetables
1876 * are freed, and leave the mm_slot on the list for ksmd to free.
1877 * Beware: ksm may already have noticed it exiting and freed the slot.
1880 spin_lock(&ksm_mmlist_lock);
1881 mm_slot = get_mm_slot(mm);
1882 if (mm_slot && ksm_scan.mm_slot != mm_slot) {
1883 if (!mm_slot->rmap_list) {
1884 hash_del(&mm_slot->link);
1885 list_del(&mm_slot->mm_list);
1888 list_move(&mm_slot->mm_list,
1889 &ksm_scan.mm_slot->mm_list);
1892 spin_unlock(&ksm_mmlist_lock);
1895 free_mm_slot(mm_slot);
1896 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1898 } else if (mm_slot) {
1899 down_write(&mm->mmap_sem);
1900 up_write(&mm->mmap_sem);
1904 struct page *ksm_might_need_to_copy(struct page *page,
1905 struct vm_area_struct *vma, unsigned long address)
1907 struct anon_vma *anon_vma = page_anon_vma(page);
1908 struct page *new_page;
1910 if (PageKsm(page)) {
1911 if (page_stable_node(page) &&
1912 !(ksm_run & KSM_RUN_UNMERGE))
1913 return page; /* no need to copy it */
1914 } else if (!anon_vma) {
1915 return page; /* no need to copy it */
1916 } else if (anon_vma->root == vma->anon_vma->root &&
1917 page->index == linear_page_index(vma, address)) {
1918 return page; /* still no need to copy it */
1920 if (!PageUptodate(page))
1921 return page; /* let do_swap_page report the error */
1923 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1925 copy_user_highpage(new_page, page, address, vma);
1927 SetPageDirty(new_page);
1928 __SetPageUptodate(new_page);
1929 __SetPageLocked(new_page);
1935 int rmap_walk_ksm(struct page *page, struct rmap_walk_control *rwc)
1937 struct stable_node *stable_node;
1938 struct rmap_item *rmap_item;
1939 int ret = SWAP_AGAIN;
1940 int search_new_forks = 0;
1942 VM_BUG_ON_PAGE(!PageKsm(page), page);
1945 * Rely on the page lock to protect against concurrent modifications
1946 * to that page's node of the stable tree.
1948 VM_BUG_ON_PAGE(!PageLocked(page), page);
1950 stable_node = page_stable_node(page);
1954 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
1955 struct anon_vma *anon_vma = rmap_item->anon_vma;
1956 struct anon_vma_chain *vmac;
1957 struct vm_area_struct *vma;
1960 anon_vma_lock_read(anon_vma);
1961 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
1965 if (rmap_item->address < vma->vm_start ||
1966 rmap_item->address >= vma->vm_end)
1969 * Initially we examine only the vma which covers this
1970 * rmap_item; but later, if there is still work to do,
1971 * we examine covering vmas in other mms: in case they
1972 * were forked from the original since ksmd passed.
1974 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1977 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1980 ret = rwc->rmap_one(page, vma,
1981 rmap_item->address, rwc->arg);
1982 if (ret != SWAP_AGAIN) {
1983 anon_vma_unlock_read(anon_vma);
1986 if (rwc->done && rwc->done(page)) {
1987 anon_vma_unlock_read(anon_vma);
1991 anon_vma_unlock_read(anon_vma);
1993 if (!search_new_forks++)
1999 #ifdef CONFIG_MIGRATION
2000 void ksm_migrate_page(struct page *newpage, struct page *oldpage)
2002 struct stable_node *stable_node;
2004 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
2005 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
2006 VM_BUG_ON_PAGE(newpage->mapping != oldpage->mapping, newpage);
2008 stable_node = page_stable_node(newpage);
2010 VM_BUG_ON_PAGE(stable_node->kpfn != page_to_pfn(oldpage), oldpage);
2011 stable_node->kpfn = page_to_pfn(newpage);
2013 * newpage->mapping was set in advance; now we need smp_wmb()
2014 * to make sure that the new stable_node->kpfn is visible
2015 * to get_ksm_page() before it can see that oldpage->mapping
2016 * has gone stale (or that PageSwapCache has been cleared).
2019 set_page_stable_node(oldpage, NULL);
2022 #endif /* CONFIG_MIGRATION */
2024 #ifdef CONFIG_MEMORY_HOTREMOVE
2025 static void wait_while_offlining(void)
2027 while (ksm_run & KSM_RUN_OFFLINE) {
2028 mutex_unlock(&ksm_thread_mutex);
2029 wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE),
2030 TASK_UNINTERRUPTIBLE);
2031 mutex_lock(&ksm_thread_mutex);
2035 static void ksm_check_stable_tree(unsigned long start_pfn,
2036 unsigned long end_pfn)
2038 struct stable_node *stable_node, *next;
2039 struct rb_node *node;
2042 for (nid = 0; nid < ksm_nr_node_ids; nid++) {
2043 node = rb_first(root_stable_tree + nid);
2045 stable_node = rb_entry(node, struct stable_node, node);
2046 if (stable_node->kpfn >= start_pfn &&
2047 stable_node->kpfn < end_pfn) {
2049 * Don't get_ksm_page, page has already gone:
2050 * which is why we keep kpfn instead of page*
2052 remove_node_from_stable_tree(stable_node);
2053 node = rb_first(root_stable_tree + nid);
2055 node = rb_next(node);
2059 list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
2060 if (stable_node->kpfn >= start_pfn &&
2061 stable_node->kpfn < end_pfn)
2062 remove_node_from_stable_tree(stable_node);
2067 static int ksm_memory_callback(struct notifier_block *self,
2068 unsigned long action, void *arg)
2070 struct memory_notify *mn = arg;
2073 case MEM_GOING_OFFLINE:
2075 * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items()
2076 * and remove_all_stable_nodes() while memory is going offline:
2077 * it is unsafe for them to touch the stable tree at this time.
2078 * But unmerge_ksm_pages(), rmap lookups and other entry points
2079 * which do not need the ksm_thread_mutex are all safe.
2081 mutex_lock(&ksm_thread_mutex);
2082 ksm_run |= KSM_RUN_OFFLINE;
2083 mutex_unlock(&ksm_thread_mutex);
2088 * Most of the work is done by page migration; but there might
2089 * be a few stable_nodes left over, still pointing to struct
2090 * pages which have been offlined: prune those from the tree,
2091 * otherwise get_ksm_page() might later try to access a
2092 * non-existent struct page.
2094 ksm_check_stable_tree(mn->start_pfn,
2095 mn->start_pfn + mn->nr_pages);
2098 case MEM_CANCEL_OFFLINE:
2099 mutex_lock(&ksm_thread_mutex);
2100 ksm_run &= ~KSM_RUN_OFFLINE;
2101 mutex_unlock(&ksm_thread_mutex);
2103 smp_mb(); /* wake_up_bit advises this */
2104 wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE));
2110 static void wait_while_offlining(void)
2113 #endif /* CONFIG_MEMORY_HOTREMOVE */
2117 * This all compiles without CONFIG_SYSFS, but is a waste of space.
2120 #define KSM_ATTR_RO(_name) \
2121 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2122 #define KSM_ATTR(_name) \
2123 static struct kobj_attribute _name##_attr = \
2124 __ATTR(_name, 0644, _name##_show, _name##_store)
2126 static ssize_t sleep_millisecs_show(struct kobject *kobj,
2127 struct kobj_attribute *attr, char *buf)
2129 return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
2132 static ssize_t sleep_millisecs_store(struct kobject *kobj,
2133 struct kobj_attribute *attr,
2134 const char *buf, size_t count)
2136 unsigned long msecs;
2139 err = kstrtoul(buf, 10, &msecs);
2140 if (err || msecs > UINT_MAX)
2143 ksm_thread_sleep_millisecs = msecs;
2147 KSM_ATTR(sleep_millisecs);
2149 static ssize_t pages_to_scan_show(struct kobject *kobj,
2150 struct kobj_attribute *attr, char *buf)
2152 return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
2155 static ssize_t pages_to_scan_store(struct kobject *kobj,
2156 struct kobj_attribute *attr,
2157 const char *buf, size_t count)
2160 unsigned long nr_pages;
2162 err = kstrtoul(buf, 10, &nr_pages);
2163 if (err || nr_pages > UINT_MAX)
2166 ksm_thread_pages_to_scan = nr_pages;
2170 KSM_ATTR(pages_to_scan);
2172 static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
2175 return sprintf(buf, "%lu\n", ksm_run);
2178 static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
2179 const char *buf, size_t count)
2182 unsigned long flags;
2184 err = kstrtoul(buf, 10, &flags);
2185 if (err || flags > UINT_MAX)
2187 if (flags > KSM_RUN_UNMERGE)
2191 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
2192 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
2193 * breaking COW to free the pages_shared (but leaves mm_slots
2194 * on the list for when ksmd may be set running again).
2197 mutex_lock(&ksm_thread_mutex);
2198 wait_while_offlining();
2199 if (ksm_run != flags) {
2201 if (flags & KSM_RUN_UNMERGE) {
2202 set_current_oom_origin();
2203 err = unmerge_and_remove_all_rmap_items();
2204 clear_current_oom_origin();
2206 ksm_run = KSM_RUN_STOP;
2211 mutex_unlock(&ksm_thread_mutex);
2213 if (flags & KSM_RUN_MERGE)
2214 wake_up_interruptible(&ksm_thread_wait);
2221 static ssize_t merge_across_nodes_show(struct kobject *kobj,
2222 struct kobj_attribute *attr, char *buf)
2224 return sprintf(buf, "%u\n", ksm_merge_across_nodes);
2227 static ssize_t merge_across_nodes_store(struct kobject *kobj,
2228 struct kobj_attribute *attr,
2229 const char *buf, size_t count)
2234 err = kstrtoul(buf, 10, &knob);
2240 mutex_lock(&ksm_thread_mutex);
2241 wait_while_offlining();
2242 if (ksm_merge_across_nodes != knob) {
2243 if (ksm_pages_shared || remove_all_stable_nodes())
2245 else if (root_stable_tree == one_stable_tree) {
2246 struct rb_root *buf;
2248 * This is the first time that we switch away from the
2249 * default of merging across nodes: must now allocate
2250 * a buffer to hold as many roots as may be needed.
2251 * Allocate stable and unstable together:
2252 * MAXSMP NODES_SHIFT 10 will use 16kB.
2254 buf = kcalloc(nr_node_ids + nr_node_ids, sizeof(*buf),
2256 /* Let us assume that RB_ROOT is NULL is zero */
2260 root_stable_tree = buf;
2261 root_unstable_tree = buf + nr_node_ids;
2262 /* Stable tree is empty but not the unstable */
2263 root_unstable_tree[0] = one_unstable_tree[0];
2267 ksm_merge_across_nodes = knob;
2268 ksm_nr_node_ids = knob ? 1 : nr_node_ids;
2271 mutex_unlock(&ksm_thread_mutex);
2273 return err ? err : count;
2275 KSM_ATTR(merge_across_nodes);
2278 static ssize_t use_zero_pages_show(struct kobject *kobj,
2279 struct kobj_attribute *attr, char *buf)
2281 return sprintf(buf, "%u\n", ksm_use_zero_pages);
2283 static ssize_t use_zero_pages_store(struct kobject *kobj,
2284 struct kobj_attribute *attr,
2285 const char *buf, size_t count)
2290 err = kstrtobool(buf, &value);
2294 ksm_use_zero_pages = value;
2298 KSM_ATTR(use_zero_pages);
2300 static ssize_t pages_shared_show(struct kobject *kobj,
2301 struct kobj_attribute *attr, char *buf)
2303 return sprintf(buf, "%lu\n", ksm_pages_shared);
2305 KSM_ATTR_RO(pages_shared);
2307 static ssize_t pages_sharing_show(struct kobject *kobj,
2308 struct kobj_attribute *attr, char *buf)
2310 return sprintf(buf, "%lu\n", ksm_pages_sharing);
2312 KSM_ATTR_RO(pages_sharing);
2314 static ssize_t pages_unshared_show(struct kobject *kobj,
2315 struct kobj_attribute *attr, char *buf)
2317 return sprintf(buf, "%lu\n", ksm_pages_unshared);
2319 KSM_ATTR_RO(pages_unshared);
2321 static ssize_t pages_volatile_show(struct kobject *kobj,
2322 struct kobj_attribute *attr, char *buf)
2324 long ksm_pages_volatile;
2326 ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
2327 - ksm_pages_sharing - ksm_pages_unshared;
2329 * It was not worth any locking to calculate that statistic,
2330 * but it might therefore sometimes be negative: conceal that.
2332 if (ksm_pages_volatile < 0)
2333 ksm_pages_volatile = 0;
2334 return sprintf(buf, "%ld\n", ksm_pages_volatile);
2336 KSM_ATTR_RO(pages_volatile);
2338 static ssize_t full_scans_show(struct kobject *kobj,
2339 struct kobj_attribute *attr, char *buf)
2341 return sprintf(buf, "%lu\n", ksm_scan.seqnr);
2343 KSM_ATTR_RO(full_scans);
2345 static struct attribute *ksm_attrs[] = {
2346 &sleep_millisecs_attr.attr,
2347 &pages_to_scan_attr.attr,
2349 &pages_shared_attr.attr,
2350 &pages_sharing_attr.attr,
2351 &pages_unshared_attr.attr,
2352 &pages_volatile_attr.attr,
2353 &full_scans_attr.attr,
2355 &merge_across_nodes_attr.attr,
2357 &use_zero_pages_attr.attr,
2361 static struct attribute_group ksm_attr_group = {
2365 #endif /* CONFIG_SYSFS */
2367 static int __init ksm_init(void)
2369 struct task_struct *ksm_thread;
2372 /* The correct value depends on page size and endianness */
2373 zero_checksum = calc_checksum(ZERO_PAGE(0));
2374 /* Default to false for backwards compatibility */
2375 ksm_use_zero_pages = false;
2377 err = ksm_slab_init();
2381 ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
2382 if (IS_ERR(ksm_thread)) {
2383 pr_err("ksm: creating kthread failed\n");
2384 err = PTR_ERR(ksm_thread);
2389 err = sysfs_create_group(mm_kobj, &ksm_attr_group);
2391 pr_err("ksm: register sysfs failed\n");
2392 kthread_stop(ksm_thread);
2396 ksm_run = KSM_RUN_MERGE; /* no way for user to start it */
2398 #endif /* CONFIG_SYSFS */
2400 #ifdef CONFIG_MEMORY_HOTREMOVE
2401 /* There is no significance to this priority 100 */
2402 hotplug_memory_notifier(ksm_memory_callback, 100);
2411 subsys_initcall(ksm_init);