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/rwsem.h>
23 #include <linux/pagemap.h>
24 #include <linux/rmap.h>
25 #include <linux/spinlock.h>
26 #include <linux/jhash.h>
27 #include <linux/delay.h>
28 #include <linux/kthread.h>
29 #include <linux/wait.h>
30 #include <linux/slab.h>
31 #include <linux/rbtree.h>
32 #include <linux/memory.h>
33 #include <linux/mmu_notifier.h>
34 #include <linux/swap.h>
35 #include <linux/ksm.h>
36 #include <linux/hashtable.h>
37 #include <linux/freezer.h>
38 #include <linux/oom.h>
39 #include <linux/numa.h>
40 #include <linux/show_mem_notifier.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 /* Boolean to indicate whether to use deferred timer or not */
228 static bool use_deferred_timer;
231 /* Zeroed when merging across nodes is not allowed */
232 static unsigned int ksm_merge_across_nodes = 1;
233 static int ksm_nr_node_ids = 1;
235 #define ksm_merge_across_nodes 1U
236 #define ksm_nr_node_ids 1
239 #define KSM_RUN_STOP 0
240 #define KSM_RUN_MERGE 1
241 #define KSM_RUN_UNMERGE 2
242 #define KSM_RUN_OFFLINE 4
243 static unsigned long ksm_run = KSM_RUN_MERGE;
244 static void wait_while_offlining(void);
246 static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
247 static DEFINE_MUTEX(ksm_thread_mutex);
248 static DEFINE_SPINLOCK(ksm_mmlist_lock);
250 #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
251 sizeof(struct __struct), __alignof__(struct __struct),\
254 static int ksm_show_mem_notifier(struct notifier_block *nb,
255 unsigned long action,
258 pr_info("ksm_pages_sharing: %lu\n", ksm_pages_sharing);
259 pr_info("ksm_pages_shared: %lu\n", ksm_pages_shared);
264 static struct notifier_block ksm_show_mem_notifier_block = {
265 .notifier_call = ksm_show_mem_notifier,
268 static int __init ksm_slab_init(void)
270 rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
271 if (!rmap_item_cache)
274 stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
275 if (!stable_node_cache)
278 mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
285 kmem_cache_destroy(stable_node_cache);
287 kmem_cache_destroy(rmap_item_cache);
292 static void __init ksm_slab_free(void)
294 kmem_cache_destroy(mm_slot_cache);
295 kmem_cache_destroy(stable_node_cache);
296 kmem_cache_destroy(rmap_item_cache);
297 mm_slot_cache = NULL;
300 static inline struct rmap_item *alloc_rmap_item(void)
302 struct rmap_item *rmap_item;
304 rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL |
305 __GFP_NORETRY | __GFP_NOWARN);
311 static inline void free_rmap_item(struct rmap_item *rmap_item)
314 rmap_item->mm = NULL; /* debug safety */
315 kmem_cache_free(rmap_item_cache, rmap_item);
318 static inline struct stable_node *alloc_stable_node(void)
320 return kmem_cache_alloc(stable_node_cache, GFP_KERNEL);
323 static inline void free_stable_node(struct stable_node *stable_node)
325 kmem_cache_free(stable_node_cache, stable_node);
328 static inline struct mm_slot *alloc_mm_slot(void)
330 if (!mm_slot_cache) /* initialization failed */
332 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
335 static inline void free_mm_slot(struct mm_slot *mm_slot)
337 kmem_cache_free(mm_slot_cache, mm_slot);
340 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
342 struct mm_slot *slot;
344 hash_for_each_possible(mm_slots_hash, slot, link, (unsigned long)mm)
351 static void insert_to_mm_slots_hash(struct mm_struct *mm,
352 struct mm_slot *mm_slot)
355 hash_add(mm_slots_hash, &mm_slot->link, (unsigned long)mm);
359 * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
360 * page tables after it has passed through ksm_exit() - which, if necessary,
361 * takes mmap_sem briefly to serialize against them. ksm_exit() does not set
362 * a special flag: they can just back out as soon as mm_users goes to zero.
363 * ksm_test_exit() is used throughout to make this test for exit: in some
364 * places for correctness, in some places just to avoid unnecessary work.
366 static inline bool ksm_test_exit(struct mm_struct *mm)
368 return atomic_read(&mm->mm_users) == 0;
372 * We use break_ksm to break COW on a ksm page: it's a stripped down
374 * if (get_user_pages(current, mm, addr, 1, 1, 1, &page, NULL) == 1)
377 * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
378 * in case the application has unmapped and remapped mm,addr meanwhile.
379 * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP
380 * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
382 static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
389 page = follow_page(vma, addr, FOLL_GET | FOLL_MIGRATION);
390 if (IS_ERR_OR_NULL(page))
393 ret = handle_mm_fault(vma->vm_mm, vma, addr,
396 ret = VM_FAULT_WRITE;
398 } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV | VM_FAULT_OOM)));
400 * We must loop because handle_mm_fault() may back out if there's
401 * any difficulty e.g. if pte accessed bit gets updated concurrently.
403 * VM_FAULT_WRITE is what we have been hoping for: it indicates that
404 * COW has been broken, even if the vma does not permit VM_WRITE;
405 * but note that a concurrent fault might break PageKsm for us.
407 * VM_FAULT_SIGBUS could occur if we race with truncation of the
408 * backing file, which also invalidates anonymous pages: that's
409 * okay, that truncation will have unmapped the PageKsm for us.
411 * VM_FAULT_OOM: at the time of writing (late July 2009), setting
412 * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
413 * current task has TIF_MEMDIE set, and will be OOM killed on return
414 * to user; and ksmd, having no mm, would never be chosen for that.
416 * But if the mm is in a limited mem_cgroup, then the fault may fail
417 * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
418 * even ksmd can fail in this way - though it's usually breaking ksm
419 * just to undo a merge it made a moment before, so unlikely to oom.
421 * That's a pity: we might therefore have more kernel pages allocated
422 * than we're counting as nodes in the stable tree; but ksm_do_scan
423 * will retry to break_cow on each pass, so should recover the page
424 * in due course. The important thing is to not let VM_MERGEABLE
425 * be cleared while any such pages might remain in the area.
427 return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
430 static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm,
433 struct vm_area_struct *vma;
434 if (ksm_test_exit(mm))
436 vma = find_vma(mm, addr);
437 if (!vma || vma->vm_start > addr)
439 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
444 static void break_cow(struct rmap_item *rmap_item)
446 struct mm_struct *mm = rmap_item->mm;
447 unsigned long addr = rmap_item->address;
448 struct vm_area_struct *vma;
451 * It is not an accident that whenever we want to break COW
452 * to undo, we also need to drop a reference to the anon_vma.
454 put_anon_vma(rmap_item->anon_vma);
456 down_read(&mm->mmap_sem);
457 vma = find_mergeable_vma(mm, addr);
459 break_ksm(vma, addr);
460 up_read(&mm->mmap_sem);
463 static struct page *page_trans_compound_anon(struct page *page)
465 if (PageTransCompound(page)) {
466 struct page *head = compound_head(page);
468 * head may actually be splitted and freed from under
469 * us but it's ok here.
477 static struct page *get_mergeable_page(struct rmap_item *rmap_item)
479 struct mm_struct *mm = rmap_item->mm;
480 unsigned long addr = rmap_item->address;
481 struct vm_area_struct *vma;
484 down_read(&mm->mmap_sem);
485 vma = find_mergeable_vma(mm, addr);
489 page = follow_page(vma, addr, FOLL_GET);
490 if (IS_ERR_OR_NULL(page))
492 if (PageAnon(page) || page_trans_compound_anon(page)) {
493 flush_anon_page(vma, page, addr);
494 flush_dcache_page(page);
500 up_read(&mm->mmap_sem);
505 * This helper is used for getting right index into array of tree roots.
506 * When merge_across_nodes knob is set to 1, there are only two rb-trees for
507 * stable and unstable pages from all nodes with roots in index 0. Otherwise,
508 * every node has its own stable and unstable tree.
510 static inline int get_kpfn_nid(unsigned long kpfn)
512 return ksm_merge_across_nodes ? 0 : NUMA(pfn_to_nid(kpfn));
515 static void remove_node_from_stable_tree(struct stable_node *stable_node)
517 struct rmap_item *rmap_item;
519 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
520 if (rmap_item->hlist.next)
524 put_anon_vma(rmap_item->anon_vma);
525 rmap_item->address &= PAGE_MASK;
529 if (stable_node->head == &migrate_nodes)
530 list_del(&stable_node->list);
532 rb_erase(&stable_node->node,
533 root_stable_tree + NUMA(stable_node->nid));
534 free_stable_node(stable_node);
538 * get_ksm_page: checks if the page indicated by the stable node
539 * is still its ksm page, despite having held no reference to it.
540 * In which case we can trust the content of the page, and it
541 * returns the gotten page; but if the page has now been zapped,
542 * remove the stale node from the stable tree and return NULL.
543 * But beware, the stable node's page might be being migrated.
545 * You would expect the stable_node to hold a reference to the ksm page.
546 * But if it increments the page's count, swapping out has to wait for
547 * ksmd to come around again before it can free the page, which may take
548 * seconds or even minutes: much too unresponsive. So instead we use a
549 * "keyhole reference": access to the ksm page from the stable node peeps
550 * out through its keyhole to see if that page still holds the right key,
551 * pointing back to this stable node. This relies on freeing a PageAnon
552 * page to reset its page->mapping to NULL, and relies on no other use of
553 * a page to put something that might look like our key in page->mapping.
554 * is on its way to being freed; but it is an anomaly to bear in mind.
556 static struct page *get_ksm_page(struct stable_node *stable_node, bool lock_it)
559 void *expected_mapping;
562 expected_mapping = (void *)((unsigned long)stable_node |
565 kpfn = READ_ONCE(stable_node->kpfn);
566 page = pfn_to_page(kpfn);
569 * page is computed from kpfn, so on most architectures reading
570 * page->mapping is naturally ordered after reading node->kpfn,
571 * but on Alpha we need to be more careful.
573 smp_read_barrier_depends();
574 if (READ_ONCE(page->mapping) != expected_mapping)
578 * We cannot do anything with the page while its refcount is 0.
579 * Usually 0 means free, or tail of a higher-order page: in which
580 * case this node is no longer referenced, and should be freed;
581 * however, it might mean that the page is under page_freeze_refs().
582 * The __remove_mapping() case is easy, again the node is now stale;
583 * but if page is swapcache in migrate_page_move_mapping(), it might
584 * still be our page, in which case it's essential to keep the node.
586 while (!get_page_unless_zero(page)) {
588 * Another check for page->mapping != expected_mapping would
589 * work here too. We have chosen the !PageSwapCache test to
590 * optimize the common case, when the page is or is about to
591 * be freed: PageSwapCache is cleared (under spin_lock_irq)
592 * in the freeze_refs section of __remove_mapping(); but Anon
593 * page->mapping reset to NULL later, in free_pages_prepare().
595 if (!PageSwapCache(page))
600 if (READ_ONCE(page->mapping) != expected_mapping) {
607 if (READ_ONCE(page->mapping) != expected_mapping) {
617 * We come here from above when page->mapping or !PageSwapCache
618 * suggests that the node is stale; but it might be under migration.
619 * We need smp_rmb(), matching the smp_wmb() in ksm_migrate_page(),
620 * before checking whether node->kpfn has been changed.
623 if (READ_ONCE(stable_node->kpfn) != kpfn)
625 remove_node_from_stable_tree(stable_node);
630 * Removing rmap_item from stable or unstable tree.
631 * This function will clean the information from the stable/unstable tree.
633 static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
635 if (rmap_item->address & STABLE_FLAG) {
636 struct stable_node *stable_node;
639 stable_node = rmap_item->head;
640 page = get_ksm_page(stable_node, true);
644 hlist_del(&rmap_item->hlist);
648 if (!hlist_empty(&stable_node->hlist))
653 put_anon_vma(rmap_item->anon_vma);
654 rmap_item->address &= PAGE_MASK;
656 } else if (rmap_item->address & UNSTABLE_FLAG) {
659 * Usually ksmd can and must skip the rb_erase, because
660 * root_unstable_tree was already reset to RB_ROOT.
661 * But be careful when an mm is exiting: do the rb_erase
662 * if this rmap_item was inserted by this scan, rather
663 * than left over from before.
665 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
668 rb_erase(&rmap_item->node,
669 root_unstable_tree + NUMA(rmap_item->nid));
670 ksm_pages_unshared--;
671 rmap_item->address &= PAGE_MASK;
674 cond_resched(); /* we're called from many long loops */
677 static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
678 struct rmap_item **rmap_list)
681 struct rmap_item *rmap_item = *rmap_list;
682 *rmap_list = rmap_item->rmap_list;
683 remove_rmap_item_from_tree(rmap_item);
684 free_rmap_item(rmap_item);
689 * Though it's very tempting to unmerge rmap_items from stable tree rather
690 * than check every pte of a given vma, the locking doesn't quite work for
691 * that - an rmap_item is assigned to the stable tree after inserting ksm
692 * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing
693 * rmap_items from parent to child at fork time (so as not to waste time
694 * if exit comes before the next scan reaches it).
696 * Similarly, although we'd like to remove rmap_items (so updating counts
697 * and freeing memory) when unmerging an area, it's easier to leave that
698 * to the next pass of ksmd - consider, for example, how ksmd might be
699 * in cmp_and_merge_page on one of the rmap_items we would be removing.
701 static int unmerge_ksm_pages(struct vm_area_struct *vma,
702 unsigned long start, unsigned long end)
707 for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
708 if (ksm_test_exit(vma->vm_mm))
710 if (signal_pending(current))
713 err = break_ksm(vma, addr);
720 * Only called through the sysfs control interface:
722 static int remove_stable_node(struct stable_node *stable_node)
727 page = get_ksm_page(stable_node, true);
730 * get_ksm_page did remove_node_from_stable_tree itself.
735 if (WARN_ON_ONCE(page_mapped(page))) {
737 * This should not happen: but if it does, just refuse to let
738 * merge_across_nodes be switched - there is no need to panic.
743 * The stable node did not yet appear stale to get_ksm_page(),
744 * since that allows for an unmapped ksm page to be recognized
745 * right up until it is freed; but the node is safe to remove.
746 * This page might be in a pagevec waiting to be freed,
747 * or it might be PageSwapCache (perhaps under writeback),
748 * or it might have been removed from swapcache a moment ago.
750 set_page_stable_node(page, NULL);
751 remove_node_from_stable_tree(stable_node);
760 static int remove_all_stable_nodes(void)
762 struct stable_node *stable_node;
763 struct list_head *this, *next;
767 for (nid = 0; nid < ksm_nr_node_ids; nid++) {
768 while (root_stable_tree[nid].rb_node) {
769 stable_node = rb_entry(root_stable_tree[nid].rb_node,
770 struct stable_node, node);
771 if (remove_stable_node(stable_node)) {
773 break; /* proceed to next nid */
778 list_for_each_safe(this, next, &migrate_nodes) {
779 stable_node = list_entry(this, struct stable_node, list);
780 if (remove_stable_node(stable_node))
787 static int unmerge_and_remove_all_rmap_items(void)
789 struct mm_slot *mm_slot;
790 struct mm_struct *mm;
791 struct vm_area_struct *vma;
794 spin_lock(&ksm_mmlist_lock);
795 ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
796 struct mm_slot, mm_list);
797 spin_unlock(&ksm_mmlist_lock);
799 for (mm_slot = ksm_scan.mm_slot;
800 mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
802 down_read(&mm->mmap_sem);
803 for (vma = mm->mmap; vma; vma = vma->vm_next) {
804 if (ksm_test_exit(mm))
806 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
808 err = unmerge_ksm_pages(vma,
809 vma->vm_start, vma->vm_end);
814 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
816 spin_lock(&ksm_mmlist_lock);
817 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
818 struct mm_slot, mm_list);
819 if (ksm_test_exit(mm)) {
820 hash_del(&mm_slot->link);
821 list_del(&mm_slot->mm_list);
822 spin_unlock(&ksm_mmlist_lock);
824 free_mm_slot(mm_slot);
825 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
826 up_read(&mm->mmap_sem);
829 spin_unlock(&ksm_mmlist_lock);
830 up_read(&mm->mmap_sem);
834 /* Clean up stable nodes, but don't worry if some are still busy */
835 remove_all_stable_nodes();
840 up_read(&mm->mmap_sem);
841 spin_lock(&ksm_mmlist_lock);
842 ksm_scan.mm_slot = &ksm_mm_head;
843 spin_unlock(&ksm_mmlist_lock);
846 #endif /* CONFIG_SYSFS */
848 static u32 calc_checksum(struct page *page)
851 void *addr = kmap_atomic(page);
852 checksum = jhash2(addr, PAGE_SIZE / 4, 17);
857 static int memcmp_pages(struct page *page1, struct page *page2)
862 addr1 = kmap_atomic(page1);
863 addr2 = kmap_atomic(page2);
864 ret = memcmp(addr1, addr2, PAGE_SIZE);
865 kunmap_atomic(addr2);
866 kunmap_atomic(addr1);
870 static inline int pages_identical(struct page *page1, struct page *page2)
872 return !memcmp_pages(page1, page2);
875 static int write_protect_page(struct vm_area_struct *vma, struct page *page,
878 struct mm_struct *mm = vma->vm_mm;
884 unsigned long mmun_start; /* For mmu_notifiers */
885 unsigned long mmun_end; /* For mmu_notifiers */
887 addr = page_address_in_vma(page, vma);
891 BUG_ON(PageTransCompound(page));
894 mmun_end = addr + PAGE_SIZE;
895 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
897 ptep = page_check_address(page, mm, addr, &ptl, 0);
901 if (pte_write(*ptep) || pte_dirty(*ptep)) {
904 swapped = PageSwapCache(page);
905 flush_cache_page(vma, addr, page_to_pfn(page));
907 * Ok this is tricky, when get_user_pages_fast() run it doesn't
908 * take any lock, therefore the check that we are going to make
909 * with the pagecount against the mapcount is racey and
910 * O_DIRECT can happen right after the check.
911 * So we clear the pte and flush the tlb before the check
912 * this assure us that no O_DIRECT can happen after the check
913 * or in the middle of the check.
915 entry = ptep_clear_flush_notify(vma, addr, ptep);
917 * Check that no O_DIRECT or similar I/O is in progress on the
920 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
921 set_pte_at(mm, addr, ptep, entry);
924 if (pte_dirty(entry))
925 set_page_dirty(page);
926 entry = pte_mkclean(pte_wrprotect(entry));
927 set_pte_at_notify(mm, addr, ptep, entry);
933 pte_unmap_unlock(ptep, ptl);
935 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
941 * replace_page - replace page in vma by new ksm page
942 * @vma: vma that holds the pte pointing to page
943 * @page: the page we are replacing by kpage
944 * @kpage: the ksm page we replace page by
945 * @orig_pte: the original value of the pte
947 * Returns 0 on success, -EFAULT on failure.
949 static int replace_page(struct vm_area_struct *vma, struct page *page,
950 struct page *kpage, pte_t orig_pte)
952 struct mm_struct *mm = vma->vm_mm;
958 unsigned long mmun_start; /* For mmu_notifiers */
959 unsigned long mmun_end; /* For mmu_notifiers */
961 addr = page_address_in_vma(page, vma);
965 pmd = mm_find_pmd(mm, addr);
970 mmun_end = addr + PAGE_SIZE;
971 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
973 ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
974 if (!pte_same(*ptep, orig_pte)) {
975 pte_unmap_unlock(ptep, ptl);
980 page_add_anon_rmap(kpage, vma, addr);
982 flush_cache_page(vma, addr, pte_pfn(*ptep));
983 ptep_clear_flush_notify(vma, addr, ptep);
984 set_pte_at_notify(mm, addr, ptep, mk_pte(kpage, vma->vm_page_prot));
986 page_remove_rmap(page);
987 if (!page_mapped(page))
988 try_to_free_swap(page);
991 pte_unmap_unlock(ptep, ptl);
994 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
999 static int page_trans_compound_anon_split(struct page *page)
1002 struct page *transhuge_head = page_trans_compound_anon(page);
1003 if (transhuge_head) {
1004 /* Get the reference on the head to split it. */
1005 if (get_page_unless_zero(transhuge_head)) {
1007 * Recheck we got the reference while the head
1008 * was still anonymous.
1010 if (PageAnon(transhuge_head))
1011 ret = split_huge_page(transhuge_head);
1014 * Retry later if split_huge_page run
1018 put_page(transhuge_head);
1020 /* Retry later if split_huge_page run from under us. */
1027 * try_to_merge_one_page - take two pages and merge them into one
1028 * @vma: the vma that holds the pte pointing to page
1029 * @page: the PageAnon page that we want to replace with kpage
1030 * @kpage: the PageKsm page that we want to map instead of page,
1031 * or NULL the first time when we want to use page as kpage.
1033 * This function returns 0 if the pages were merged, -EFAULT otherwise.
1035 static int try_to_merge_one_page(struct vm_area_struct *vma,
1036 struct page *page, struct page *kpage)
1038 pte_t orig_pte = __pte(0);
1041 if (page == kpage) /* ksm page forked */
1044 if (PageTransCompound(page) && page_trans_compound_anon_split(page))
1046 BUG_ON(PageTransCompound(page));
1047 if (!PageAnon(page))
1051 * We need the page lock to read a stable PageSwapCache in
1052 * write_protect_page(). We use trylock_page() instead of
1053 * lock_page() because we don't want to wait here - we
1054 * prefer to continue scanning and merging different pages,
1055 * then come back to this page when it is unlocked.
1057 if (!trylock_page(page))
1060 * If this anonymous page is mapped only here, its pte may need
1061 * to be write-protected. If it's mapped elsewhere, all of its
1062 * ptes are necessarily already write-protected. But in either
1063 * case, we need to lock and check page_count is not raised.
1065 if (write_protect_page(vma, page, &orig_pte) == 0) {
1068 * While we hold page lock, upgrade page from
1069 * PageAnon+anon_vma to PageKsm+NULL stable_node:
1070 * stable_tree_insert() will update stable_node.
1072 set_page_stable_node(page, NULL);
1073 mark_page_accessed(page);
1075 } else if (pages_identical(page, kpage))
1076 err = replace_page(vma, page, kpage, orig_pte);
1079 if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
1080 munlock_vma_page(page);
1081 if (!PageMlocked(kpage)) {
1084 mlock_vma_page(kpage);
1085 page = kpage; /* for final unlock */
1095 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
1096 * but no new kernel page is allocated: kpage must already be a ksm page.
1098 * This function returns 0 if the pages were merged, -EFAULT otherwise.
1100 static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
1101 struct page *page, struct page *kpage)
1103 struct mm_struct *mm = rmap_item->mm;
1104 struct vm_area_struct *vma;
1107 down_read(&mm->mmap_sem);
1108 vma = find_mergeable_vma(mm, rmap_item->address);
1112 err = try_to_merge_one_page(vma, page, kpage);
1116 /* Unstable nid is in union with stable anon_vma: remove first */
1117 remove_rmap_item_from_tree(rmap_item);
1119 /* Must get reference to anon_vma while still holding mmap_sem */
1120 rmap_item->anon_vma = vma->anon_vma;
1121 get_anon_vma(vma->anon_vma);
1123 up_read(&mm->mmap_sem);
1128 * try_to_merge_two_pages - take two identical pages and prepare them
1129 * to be merged into one page.
1131 * This function returns the kpage if we successfully merged two identical
1132 * pages into one ksm page, NULL otherwise.
1134 * Note that this function upgrades page to ksm page: if one of the pages
1135 * is already a ksm page, try_to_merge_with_ksm_page should be used.
1137 static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
1139 struct rmap_item *tree_rmap_item,
1140 struct page *tree_page)
1144 err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
1146 err = try_to_merge_with_ksm_page(tree_rmap_item,
1149 * If that fails, we have a ksm page with only one pte
1150 * pointing to it: so break it.
1153 break_cow(rmap_item);
1155 return err ? NULL : page;
1159 * stable_tree_search - search for page inside the stable tree
1161 * This function checks if there is a page inside the stable tree
1162 * with identical content to the page that we are scanning right now.
1164 * This function returns the stable tree node of identical content if found,
1167 static struct page *stable_tree_search(struct page *page)
1170 struct rb_root *root;
1171 struct rb_node **new;
1172 struct rb_node *parent;
1173 struct stable_node *stable_node;
1174 struct stable_node *page_node;
1176 page_node = page_stable_node(page);
1177 if (page_node && page_node->head != &migrate_nodes) {
1178 /* ksm page forked */
1183 nid = get_kpfn_nid(page_to_pfn(page));
1184 root = root_stable_tree + nid;
1186 new = &root->rb_node;
1190 struct page *tree_page;
1194 stable_node = rb_entry(*new, struct stable_node, node);
1195 tree_page = get_ksm_page(stable_node, false);
1198 * If we walked over a stale stable_node,
1199 * get_ksm_page() will call rb_erase() and it
1200 * may rebalance the tree from under us. So
1201 * restart the search from scratch. Returning
1202 * NULL would be safe too, but we'd generate
1203 * false negative insertions just because some
1204 * stable_node was stale.
1209 ret = memcmp_pages(page, tree_page);
1210 put_page(tree_page);
1214 new = &parent->rb_left;
1216 new = &parent->rb_right;
1219 * Lock and unlock the stable_node's page (which
1220 * might already have been migrated) so that page
1221 * migration is sure to notice its raised count.
1222 * It would be more elegant to return stable_node
1223 * than kpage, but that involves more changes.
1225 tree_page = get_ksm_page(stable_node, true);
1227 unlock_page(tree_page);
1228 if (get_kpfn_nid(stable_node->kpfn) !=
1229 NUMA(stable_node->nid)) {
1230 put_page(tree_page);
1236 * There is now a place for page_node, but the tree may
1237 * have been rebalanced, so re-evaluate parent and new.
1248 list_del(&page_node->list);
1249 DO_NUMA(page_node->nid = nid);
1250 rb_link_node(&page_node->node, parent, new);
1251 rb_insert_color(&page_node->node, root);
1257 list_del(&page_node->list);
1258 DO_NUMA(page_node->nid = nid);
1259 rb_replace_node(&stable_node->node, &page_node->node, root);
1262 rb_erase(&stable_node->node, root);
1265 stable_node->head = &migrate_nodes;
1266 list_add(&stable_node->list, stable_node->head);
1271 * stable_tree_insert - insert stable tree node pointing to new ksm page
1272 * into the stable tree.
1274 * This function returns the stable tree node just allocated on success,
1277 static struct stable_node *stable_tree_insert(struct page *kpage)
1281 struct rb_root *root;
1282 struct rb_node **new;
1283 struct rb_node *parent;
1284 struct stable_node *stable_node;
1286 kpfn = page_to_pfn(kpage);
1287 nid = get_kpfn_nid(kpfn);
1288 root = root_stable_tree + nid;
1291 new = &root->rb_node;
1294 struct page *tree_page;
1298 stable_node = rb_entry(*new, struct stable_node, node);
1299 tree_page = get_ksm_page(stable_node, false);
1302 * If we walked over a stale stable_node,
1303 * get_ksm_page() will call rb_erase() and it
1304 * may rebalance the tree from under us. So
1305 * restart the search from scratch. Returning
1306 * NULL would be safe too, but we'd generate
1307 * false negative insertions just because some
1308 * stable_node was stale.
1313 ret = memcmp_pages(kpage, tree_page);
1314 put_page(tree_page);
1318 new = &parent->rb_left;
1320 new = &parent->rb_right;
1323 * It is not a bug that stable_tree_search() didn't
1324 * find this node: because at that time our page was
1325 * not yet write-protected, so may have changed since.
1331 stable_node = alloc_stable_node();
1335 INIT_HLIST_HEAD(&stable_node->hlist);
1336 stable_node->kpfn = kpfn;
1337 set_page_stable_node(kpage, stable_node);
1338 DO_NUMA(stable_node->nid = nid);
1339 rb_link_node(&stable_node->node, parent, new);
1340 rb_insert_color(&stable_node->node, root);
1346 * unstable_tree_search_insert - search for identical page,
1347 * else insert rmap_item into the unstable tree.
1349 * This function searches for a page in the unstable tree identical to the
1350 * page currently being scanned; and if no identical page is found in the
1351 * tree, we insert rmap_item as a new object into the unstable tree.
1353 * This function returns pointer to rmap_item found to be identical
1354 * to the currently scanned page, NULL otherwise.
1356 * This function does both searching and inserting, because they share
1357 * the same walking algorithm in an rbtree.
1360 struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1362 struct page **tree_pagep)
1364 struct rb_node **new;
1365 struct rb_root *root;
1366 struct rb_node *parent = NULL;
1369 nid = get_kpfn_nid(page_to_pfn(page));
1370 root = root_unstable_tree + nid;
1371 new = &root->rb_node;
1374 struct rmap_item *tree_rmap_item;
1375 struct page *tree_page;
1379 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
1380 tree_page = get_mergeable_page(tree_rmap_item);
1385 * Don't substitute a ksm page for a forked page.
1387 if (page == tree_page) {
1388 put_page(tree_page);
1392 ret = memcmp_pages(page, tree_page);
1396 put_page(tree_page);
1397 new = &parent->rb_left;
1398 } else if (ret > 0) {
1399 put_page(tree_page);
1400 new = &parent->rb_right;
1401 } else if (!ksm_merge_across_nodes &&
1402 page_to_nid(tree_page) != nid) {
1404 * If tree_page has been migrated to another NUMA node,
1405 * it will be flushed out and put in the right unstable
1406 * tree next time: only merge with it when across_nodes.
1408 put_page(tree_page);
1411 *tree_pagep = tree_page;
1412 return tree_rmap_item;
1416 rmap_item->address |= UNSTABLE_FLAG;
1417 rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1418 DO_NUMA(rmap_item->nid = nid);
1419 rb_link_node(&rmap_item->node, parent, new);
1420 rb_insert_color(&rmap_item->node, root);
1422 ksm_pages_unshared++;
1427 * stable_tree_append - add another rmap_item to the linked list of
1428 * rmap_items hanging off a given node of the stable tree, all sharing
1429 * the same ksm page.
1431 static void stable_tree_append(struct rmap_item *rmap_item,
1432 struct stable_node *stable_node)
1434 rmap_item->head = stable_node;
1435 rmap_item->address |= STABLE_FLAG;
1436 hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
1438 if (rmap_item->hlist.next)
1439 ksm_pages_sharing++;
1445 * cmp_and_merge_page - first see if page can be merged into the stable tree;
1446 * if not, compare checksum to previous and if it's the same, see if page can
1447 * be inserted into the unstable tree, or merged with a page already there and
1448 * both transferred to the stable tree.
1450 * @page: the page that we are searching identical page to.
1451 * @rmap_item: the reverse mapping into the virtual address of this page
1453 static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1455 struct rmap_item *tree_rmap_item;
1456 struct page *tree_page = NULL;
1457 struct stable_node *stable_node;
1459 unsigned int checksum;
1462 stable_node = page_stable_node(page);
1464 if (stable_node->head != &migrate_nodes &&
1465 get_kpfn_nid(stable_node->kpfn) != NUMA(stable_node->nid)) {
1466 rb_erase(&stable_node->node,
1467 root_stable_tree + NUMA(stable_node->nid));
1468 stable_node->head = &migrate_nodes;
1469 list_add(&stable_node->list, stable_node->head);
1471 if (stable_node->head != &migrate_nodes &&
1472 rmap_item->head == stable_node)
1476 /* We first start with searching the page inside the stable tree */
1477 kpage = stable_tree_search(page);
1478 if (kpage == page && rmap_item->head == stable_node) {
1483 remove_rmap_item_from_tree(rmap_item);
1486 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
1489 * The page was successfully merged:
1490 * add its rmap_item to the stable tree.
1493 stable_tree_append(rmap_item, page_stable_node(kpage));
1501 * If the hash value of the page has changed from the last time
1502 * we calculated it, this page is changing frequently: therefore we
1503 * don't want to insert it in the unstable tree, and we don't want
1504 * to waste our time searching for something identical to it there.
1506 checksum = calc_checksum(page);
1507 if (rmap_item->oldchecksum != checksum) {
1508 rmap_item->oldchecksum = checksum;
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;
1606 struct list_head *this, *next;
1609 list_for_each_safe(this, next, &migrate_nodes) {
1610 stable_node = list_entry(this,
1611 struct stable_node, list);
1612 page = get_ksm_page(stable_node, false);
1619 for (nid = 0; nid < ksm_nr_node_ids; nid++)
1620 root_unstable_tree[nid] = RB_ROOT;
1622 spin_lock(&ksm_mmlist_lock);
1623 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1624 ksm_scan.mm_slot = slot;
1625 spin_unlock(&ksm_mmlist_lock);
1627 * Although we tested list_empty() above, a racing __ksm_exit
1628 * of the last mm on the list may have removed it since then.
1630 if (slot == &ksm_mm_head)
1633 ksm_scan.address = 0;
1634 ksm_scan.rmap_list = &slot->rmap_list;
1638 down_read(&mm->mmap_sem);
1639 if (ksm_test_exit(mm))
1642 vma = find_vma(mm, ksm_scan.address);
1644 for (; vma; vma = vma->vm_next) {
1645 if (!(vma->vm_flags & VM_MERGEABLE))
1647 if (ksm_scan.address < vma->vm_start)
1648 ksm_scan.address = vma->vm_start;
1650 ksm_scan.address = vma->vm_end;
1652 while (ksm_scan.address < vma->vm_end) {
1653 if (ksm_test_exit(mm))
1655 *page = follow_page(vma, ksm_scan.address, FOLL_GET);
1656 if (IS_ERR_OR_NULL(*page)) {
1657 ksm_scan.address += PAGE_SIZE;
1661 if (PageAnon(*page) ||
1662 page_trans_compound_anon(*page)) {
1663 flush_anon_page(vma, *page, ksm_scan.address);
1664 flush_dcache_page(*page);
1665 rmap_item = get_next_rmap_item(slot,
1666 ksm_scan.rmap_list, ksm_scan.address);
1668 ksm_scan.rmap_list =
1669 &rmap_item->rmap_list;
1670 ksm_scan.address += PAGE_SIZE;
1673 up_read(&mm->mmap_sem);
1677 ksm_scan.address += PAGE_SIZE;
1682 if (ksm_test_exit(mm)) {
1683 ksm_scan.address = 0;
1684 ksm_scan.rmap_list = &slot->rmap_list;
1687 * Nuke all the rmap_items that are above this current rmap:
1688 * because there were no VM_MERGEABLE vmas with such addresses.
1690 remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
1692 spin_lock(&ksm_mmlist_lock);
1693 ksm_scan.mm_slot = list_entry(slot->mm_list.next,
1694 struct mm_slot, mm_list);
1695 if (ksm_scan.address == 0) {
1697 * We've completed a full scan of all vmas, holding mmap_sem
1698 * throughout, and found no VM_MERGEABLE: so do the same as
1699 * __ksm_exit does to remove this mm from all our lists now.
1700 * This applies either when cleaning up after __ksm_exit
1701 * (but beware: we can reach here even before __ksm_exit),
1702 * or when all VM_MERGEABLE areas have been unmapped (and
1703 * mmap_sem then protects against race with MADV_MERGEABLE).
1705 hash_del(&slot->link);
1706 list_del(&slot->mm_list);
1707 spin_unlock(&ksm_mmlist_lock);
1710 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1711 up_read(&mm->mmap_sem);
1714 spin_unlock(&ksm_mmlist_lock);
1715 up_read(&mm->mmap_sem);
1718 /* Repeat until we've completed scanning the whole list */
1719 slot = ksm_scan.mm_slot;
1720 if (slot != &ksm_mm_head)
1728 * ksm_do_scan - the ksm scanner main worker function.
1729 * @scan_npages - number of pages we want to scan before we return.
1731 static void ksm_do_scan(unsigned int scan_npages)
1733 struct rmap_item *rmap_item;
1734 struct page *uninitialized_var(page);
1736 while (scan_npages-- && likely(!freezing(current))) {
1738 rmap_item = scan_get_next_rmap_item(&page);
1741 cmp_and_merge_page(page, rmap_item);
1746 static void process_timeout(unsigned long __data)
1748 wake_up_process((struct task_struct *)__data);
1751 static signed long __sched deferred_schedule_timeout(signed long timeout)
1753 struct timer_list timer;
1754 unsigned long expire;
1756 __set_current_state(TASK_INTERRUPTIBLE);
1758 pr_err("schedule_timeout: wrong timeout value %lx\n",
1760 __set_current_state(TASK_RUNNING);
1764 expire = timeout + jiffies;
1766 setup_deferrable_timer_on_stack(&timer, process_timeout,
1767 (unsigned long)current);
1768 mod_timer(&timer, expire);
1770 del_singleshot_timer_sync(&timer);
1772 /* Remove the timer from the object tracker */
1773 destroy_timer_on_stack(&timer);
1775 timeout = expire - jiffies;
1778 return timeout < 0 ? 0 : timeout;
1781 static int ksmd_should_run(void)
1783 return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
1786 static int ksm_scan_thread(void *nothing)
1789 set_user_nice(current, 5);
1791 while (!kthread_should_stop()) {
1792 mutex_lock(&ksm_thread_mutex);
1793 wait_while_offlining();
1794 if (ksmd_should_run())
1795 ksm_do_scan(ksm_thread_pages_to_scan);
1796 mutex_unlock(&ksm_thread_mutex);
1800 if (ksmd_should_run()) {
1801 if (use_deferred_timer)
1802 deferred_schedule_timeout(
1803 msecs_to_jiffies(ksm_thread_sleep_millisecs));
1805 schedule_timeout_interruptible(
1806 msecs_to_jiffies(ksm_thread_sleep_millisecs));
1808 wait_event_freezable(ksm_thread_wait,
1809 ksmd_should_run() || kthread_should_stop());
1815 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1816 unsigned long end, int advice, unsigned long *vm_flags)
1818 struct mm_struct *mm = vma->vm_mm;
1822 case MADV_MERGEABLE:
1824 * Be somewhat over-protective for now!
1826 if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE |
1827 VM_PFNMAP | VM_IO | VM_DONTEXPAND |
1828 VM_HUGETLB | VM_MIXEDMAP))
1829 return 0; /* just ignore the advice */
1832 if (*vm_flags & VM_SAO)
1836 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
1837 err = __ksm_enter(mm);
1842 *vm_flags |= VM_MERGEABLE;
1845 case MADV_UNMERGEABLE:
1846 if (!(*vm_flags & VM_MERGEABLE))
1847 return 0; /* just ignore the advice */
1849 if (vma->anon_vma) {
1850 err = unmerge_ksm_pages(vma, start, end);
1855 *vm_flags &= ~VM_MERGEABLE;
1862 int __ksm_enter(struct mm_struct *mm)
1864 struct mm_slot *mm_slot;
1867 mm_slot = alloc_mm_slot();
1871 /* Check ksm_run too? Would need tighter locking */
1872 needs_wakeup = list_empty(&ksm_mm_head.mm_list);
1874 spin_lock(&ksm_mmlist_lock);
1875 insert_to_mm_slots_hash(mm, mm_slot);
1877 * When KSM_RUN_MERGE (or KSM_RUN_STOP),
1878 * insert just behind the scanning cursor, to let the area settle
1879 * down a little; when fork is followed by immediate exec, we don't
1880 * want ksmd to waste time setting up and tearing down an rmap_list.
1882 * But when KSM_RUN_UNMERGE, it's important to insert ahead of its
1883 * scanning cursor, otherwise KSM pages in newly forked mms will be
1884 * missed: then we might as well insert at the end of the list.
1886 if (ksm_run & KSM_RUN_UNMERGE)
1887 list_add_tail(&mm_slot->mm_list, &ksm_mm_head.mm_list);
1889 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
1890 spin_unlock(&ksm_mmlist_lock);
1892 set_bit(MMF_VM_MERGEABLE, &mm->flags);
1893 atomic_inc(&mm->mm_count);
1896 wake_up_interruptible(&ksm_thread_wait);
1901 void __ksm_exit(struct mm_struct *mm)
1903 struct mm_slot *mm_slot;
1904 int easy_to_free = 0;
1907 * This process is exiting: if it's straightforward (as is the
1908 * case when ksmd was never running), free mm_slot immediately.
1909 * But if it's at the cursor or has rmap_items linked to it, use
1910 * mmap_sem to synchronize with any break_cows before pagetables
1911 * are freed, and leave the mm_slot on the list for ksmd to free.
1912 * Beware: ksm may already have noticed it exiting and freed the slot.
1915 spin_lock(&ksm_mmlist_lock);
1916 mm_slot = get_mm_slot(mm);
1917 if (mm_slot && ksm_scan.mm_slot != mm_slot) {
1918 if (!mm_slot->rmap_list) {
1919 hash_del(&mm_slot->link);
1920 list_del(&mm_slot->mm_list);
1923 list_move(&mm_slot->mm_list,
1924 &ksm_scan.mm_slot->mm_list);
1927 spin_unlock(&ksm_mmlist_lock);
1930 free_mm_slot(mm_slot);
1931 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1933 } else if (mm_slot) {
1934 down_write(&mm->mmap_sem);
1935 up_write(&mm->mmap_sem);
1939 struct page *ksm_might_need_to_copy(struct page *page,
1940 struct vm_area_struct *vma, unsigned long address)
1942 struct anon_vma *anon_vma = page_anon_vma(page);
1943 struct page *new_page;
1945 if (PageKsm(page)) {
1946 if (page_stable_node(page) &&
1947 !(ksm_run & KSM_RUN_UNMERGE))
1948 return page; /* no need to copy it */
1949 } else if (!anon_vma) {
1950 return page; /* no need to copy it */
1951 } else if (anon_vma->root == vma->anon_vma->root &&
1952 page->index == linear_page_index(vma, address)) {
1953 return page; /* still no need to copy it */
1955 if (!PageUptodate(page))
1956 return page; /* let do_swap_page report the error */
1958 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1960 copy_user_highpage(new_page, page, address, vma);
1962 SetPageDirty(new_page);
1963 __SetPageUptodate(new_page);
1964 __set_page_locked(new_page);
1970 int rmap_walk_ksm(struct page *page, struct rmap_walk_control *rwc)
1972 struct stable_node *stable_node;
1973 struct rmap_item *rmap_item;
1974 int ret = SWAP_AGAIN;
1975 int search_new_forks = 0;
1977 VM_BUG_ON_PAGE(!PageKsm(page), page);
1980 * Rely on the page lock to protect against concurrent modifications
1981 * to that page's node of the stable tree.
1983 VM_BUG_ON_PAGE(!PageLocked(page), page);
1985 stable_node = page_stable_node(page);
1990 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
1991 struct anon_vma *anon_vma = rmap_item->anon_vma;
1992 struct anon_vma_chain *vmac;
1993 struct vm_area_struct *vma;
1996 anon_vma_lock_read(anon_vma);
1997 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
2001 if (rmap_item->address < vma->vm_start ||
2002 rmap_item->address >= vma->vm_end)
2005 * Initially we examine only the vma which covers this
2006 * rmap_item; but later, if there is still work to do,
2007 * we examine covering vmas in other mms: in case they
2008 * were forked from the original since ksmd passed.
2010 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
2013 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2016 ret = rwc->rmap_one(page, vma,
2017 rmap_item->address, rwc->arg);
2018 if (ret != SWAP_AGAIN) {
2019 anon_vma_unlock_read(anon_vma);
2022 if (rwc->done && rwc->done(page)) {
2023 anon_vma_unlock_read(anon_vma);
2027 anon_vma_unlock_read(anon_vma);
2029 if (!search_new_forks++)
2035 #ifdef CONFIG_MIGRATION
2036 void ksm_migrate_page(struct page *newpage, struct page *oldpage)
2038 struct stable_node *stable_node;
2040 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
2041 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
2042 VM_BUG_ON_PAGE(newpage->mapping != oldpage->mapping, newpage);
2044 stable_node = page_stable_node(newpage);
2046 VM_BUG_ON_PAGE(stable_node->kpfn != page_to_pfn(oldpage), oldpage);
2047 stable_node->kpfn = page_to_pfn(newpage);
2049 * newpage->mapping was set in advance; now we need smp_wmb()
2050 * to make sure that the new stable_node->kpfn is visible
2051 * to get_ksm_page() before it can see that oldpage->mapping
2052 * has gone stale (or that PageSwapCache has been cleared).
2055 set_page_stable_node(oldpage, NULL);
2058 #endif /* CONFIG_MIGRATION */
2060 #ifdef CONFIG_MEMORY_HOTREMOVE
2061 static void wait_while_offlining(void)
2063 while (ksm_run & KSM_RUN_OFFLINE) {
2064 mutex_unlock(&ksm_thread_mutex);
2065 wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE),
2066 TASK_UNINTERRUPTIBLE);
2067 mutex_lock(&ksm_thread_mutex);
2071 static void ksm_check_stable_tree(unsigned long start_pfn,
2072 unsigned long end_pfn)
2074 struct stable_node *stable_node;
2075 struct list_head *this, *next;
2076 struct rb_node *node;
2079 for (nid = 0; nid < ksm_nr_node_ids; nid++) {
2080 node = rb_first(root_stable_tree + nid);
2082 stable_node = rb_entry(node, struct stable_node, node);
2083 if (stable_node->kpfn >= start_pfn &&
2084 stable_node->kpfn < end_pfn) {
2086 * Don't get_ksm_page, page has already gone:
2087 * which is why we keep kpfn instead of page*
2089 remove_node_from_stable_tree(stable_node);
2090 node = rb_first(root_stable_tree + nid);
2092 node = rb_next(node);
2096 list_for_each_safe(this, next, &migrate_nodes) {
2097 stable_node = list_entry(this, struct stable_node, list);
2098 if (stable_node->kpfn >= start_pfn &&
2099 stable_node->kpfn < end_pfn)
2100 remove_node_from_stable_tree(stable_node);
2105 static int ksm_memory_callback(struct notifier_block *self,
2106 unsigned long action, void *arg)
2108 struct memory_notify *mn = arg;
2111 case MEM_GOING_OFFLINE:
2113 * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items()
2114 * and remove_all_stable_nodes() while memory is going offline:
2115 * it is unsafe for them to touch the stable tree at this time.
2116 * But unmerge_ksm_pages(), rmap lookups and other entry points
2117 * which do not need the ksm_thread_mutex are all safe.
2119 mutex_lock(&ksm_thread_mutex);
2120 ksm_run |= KSM_RUN_OFFLINE;
2121 mutex_unlock(&ksm_thread_mutex);
2126 * Most of the work is done by page migration; but there might
2127 * be a few stable_nodes left over, still pointing to struct
2128 * pages which have been offlined: prune those from the tree,
2129 * otherwise get_ksm_page() might later try to access a
2130 * non-existent struct page.
2132 ksm_check_stable_tree(mn->start_pfn,
2133 mn->start_pfn + mn->nr_pages);
2136 case MEM_CANCEL_OFFLINE:
2137 mutex_lock(&ksm_thread_mutex);
2138 ksm_run &= ~KSM_RUN_OFFLINE;
2139 mutex_unlock(&ksm_thread_mutex);
2141 smp_mb(); /* wake_up_bit advises this */
2142 wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE));
2148 static void wait_while_offlining(void)
2151 #endif /* CONFIG_MEMORY_HOTREMOVE */
2155 * This all compiles without CONFIG_SYSFS, but is a waste of space.
2158 #define KSM_ATTR_RO(_name) \
2159 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2160 #define KSM_ATTR(_name) \
2161 static struct kobj_attribute _name##_attr = \
2162 __ATTR(_name, 0644, _name##_show, _name##_store)
2164 static ssize_t sleep_millisecs_show(struct kobject *kobj,
2165 struct kobj_attribute *attr, char *buf)
2167 return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
2170 static ssize_t sleep_millisecs_store(struct kobject *kobj,
2171 struct kobj_attribute *attr,
2172 const char *buf, size_t count)
2174 unsigned long msecs;
2177 err = kstrtoul(buf, 10, &msecs);
2178 if (err || msecs > UINT_MAX)
2181 ksm_thread_sleep_millisecs = msecs;
2185 KSM_ATTR(sleep_millisecs);
2187 static ssize_t pages_to_scan_show(struct kobject *kobj,
2188 struct kobj_attribute *attr, char *buf)
2190 return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
2193 static ssize_t pages_to_scan_store(struct kobject *kobj,
2194 struct kobj_attribute *attr,
2195 const char *buf, size_t count)
2198 unsigned long nr_pages;
2200 err = kstrtoul(buf, 10, &nr_pages);
2201 if (err || nr_pages > UINT_MAX)
2204 ksm_thread_pages_to_scan = nr_pages;
2208 KSM_ATTR(pages_to_scan);
2210 static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
2213 return sprintf(buf, "%lu\n", ksm_run);
2216 static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
2217 const char *buf, size_t count)
2220 unsigned long flags;
2222 err = kstrtoul(buf, 10, &flags);
2223 if (err || flags > UINT_MAX)
2225 if (flags > KSM_RUN_UNMERGE)
2229 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
2230 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
2231 * breaking COW to free the pages_shared (but leaves mm_slots
2232 * on the list for when ksmd may be set running again).
2235 mutex_lock(&ksm_thread_mutex);
2236 wait_while_offlining();
2237 if (ksm_run != flags) {
2239 if (flags & KSM_RUN_UNMERGE) {
2240 set_current_oom_origin();
2241 err = unmerge_and_remove_all_rmap_items();
2242 clear_current_oom_origin();
2244 ksm_run = KSM_RUN_STOP;
2249 mutex_unlock(&ksm_thread_mutex);
2251 if (flags & KSM_RUN_MERGE)
2252 wake_up_interruptible(&ksm_thread_wait);
2258 static ssize_t deferred_timer_show(struct kobject *kobj,
2259 struct kobj_attribute *attr, char *buf)
2261 return snprintf(buf, 8, "%d\n", use_deferred_timer);
2264 static ssize_t deferred_timer_store(struct kobject *kobj,
2265 struct kobj_attribute *attr,
2266 const char *buf, size_t count)
2268 unsigned long enable;
2271 err = kstrtoul(buf, 10, &enable);
2272 use_deferred_timer = enable;
2276 KSM_ATTR(deferred_timer);
2279 static ssize_t merge_across_nodes_show(struct kobject *kobj,
2280 struct kobj_attribute *attr, char *buf)
2282 return sprintf(buf, "%u\n", ksm_merge_across_nodes);
2285 static ssize_t merge_across_nodes_store(struct kobject *kobj,
2286 struct kobj_attribute *attr,
2287 const char *buf, size_t count)
2292 err = kstrtoul(buf, 10, &knob);
2298 mutex_lock(&ksm_thread_mutex);
2299 wait_while_offlining();
2300 if (ksm_merge_across_nodes != knob) {
2301 if (ksm_pages_shared || remove_all_stable_nodes())
2303 else if (root_stable_tree == one_stable_tree) {
2304 struct rb_root *buf;
2306 * This is the first time that we switch away from the
2307 * default of merging across nodes: must now allocate
2308 * a buffer to hold as many roots as may be needed.
2309 * Allocate stable and unstable together:
2310 * MAXSMP NODES_SHIFT 10 will use 16kB.
2312 buf = kcalloc(nr_node_ids + nr_node_ids, sizeof(*buf),
2314 /* Let us assume that RB_ROOT is NULL is zero */
2318 root_stable_tree = buf;
2319 root_unstable_tree = buf + nr_node_ids;
2320 /* Stable tree is empty but not the unstable */
2321 root_unstable_tree[0] = one_unstable_tree[0];
2325 ksm_merge_across_nodes = knob;
2326 ksm_nr_node_ids = knob ? 1 : nr_node_ids;
2329 mutex_unlock(&ksm_thread_mutex);
2331 return err ? err : count;
2333 KSM_ATTR(merge_across_nodes);
2336 static ssize_t pages_shared_show(struct kobject *kobj,
2337 struct kobj_attribute *attr, char *buf)
2339 return sprintf(buf, "%lu\n", ksm_pages_shared);
2341 KSM_ATTR_RO(pages_shared);
2343 static ssize_t pages_sharing_show(struct kobject *kobj,
2344 struct kobj_attribute *attr, char *buf)
2346 return sprintf(buf, "%lu\n", ksm_pages_sharing);
2348 KSM_ATTR_RO(pages_sharing);
2350 static ssize_t pages_unshared_show(struct kobject *kobj,
2351 struct kobj_attribute *attr, char *buf)
2353 return sprintf(buf, "%lu\n", ksm_pages_unshared);
2355 KSM_ATTR_RO(pages_unshared);
2357 static ssize_t pages_volatile_show(struct kobject *kobj,
2358 struct kobj_attribute *attr, char *buf)
2360 long ksm_pages_volatile;
2362 ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
2363 - ksm_pages_sharing - ksm_pages_unshared;
2365 * It was not worth any locking to calculate that statistic,
2366 * but it might therefore sometimes be negative: conceal that.
2368 if (ksm_pages_volatile < 0)
2369 ksm_pages_volatile = 0;
2370 return sprintf(buf, "%ld\n", ksm_pages_volatile);
2372 KSM_ATTR_RO(pages_volatile);
2374 static ssize_t full_scans_show(struct kobject *kobj,
2375 struct kobj_attribute *attr, char *buf)
2377 return sprintf(buf, "%lu\n", ksm_scan.seqnr);
2379 KSM_ATTR_RO(full_scans);
2381 static struct attribute *ksm_attrs[] = {
2382 &sleep_millisecs_attr.attr,
2383 &pages_to_scan_attr.attr,
2385 &pages_shared_attr.attr,
2386 &pages_sharing_attr.attr,
2387 &pages_unshared_attr.attr,
2388 &pages_volatile_attr.attr,
2389 &full_scans_attr.attr,
2390 &deferred_timer_attr.attr,
2392 &merge_across_nodes_attr.attr,
2397 static struct attribute_group ksm_attr_group = {
2401 #endif /* CONFIG_SYSFS */
2403 static int __init ksm_init(void)
2405 struct task_struct *ksm_thread;
2408 err = ksm_slab_init();
2412 ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
2413 if (IS_ERR(ksm_thread)) {
2414 pr_err("ksm: creating kthread failed\n");
2415 err = PTR_ERR(ksm_thread);
2420 err = sysfs_create_group(mm_kobj, &ksm_attr_group);
2422 pr_err("ksm: register sysfs failed\n");
2423 kthread_stop(ksm_thread);
2427 ksm_run = KSM_RUN_MERGE; /* no way for user to start it */
2429 #endif /* CONFIG_SYSFS */
2431 #ifdef CONFIG_MEMORY_HOTREMOVE
2432 /* There is no significance to this priority 100 */
2433 hotplug_memory_notifier(ksm_memory_callback, 100);
2436 show_mem_notifier_register(&ksm_show_mem_notifier_block);
2444 subsys_initcall(ksm_init);