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>
41 #include <asm/tlbflush.h>
46 #define DO_NUMA(x) do { (x); } while (0)
49 #define DO_NUMA(x) do { } while (0)
53 * A few notes about the KSM scanning process,
54 * to make it easier to understand the data structures below:
56 * In order to reduce excessive scanning, KSM sorts the memory pages by their
57 * contents into a data structure that holds pointers to the pages' locations.
59 * Since the contents of the pages may change at any moment, KSM cannot just
60 * insert the pages into a normal sorted tree and expect it to find anything.
61 * Therefore KSM uses two data structures - the stable and the unstable tree.
63 * The stable tree holds pointers to all the merged pages (ksm pages), sorted
64 * by their contents. Because each such page is write-protected, searching on
65 * this tree is fully assured to be working (except when pages are unmapped),
66 * and therefore this tree is called the stable tree.
68 * In addition to the stable tree, KSM uses a second data structure called the
69 * unstable tree: this tree holds pointers to pages which have been found to
70 * be "unchanged for a period of time". The unstable tree sorts these pages
71 * by their contents, but since they are not write-protected, KSM cannot rely
72 * upon the unstable tree to work correctly - the unstable tree is liable to
73 * be corrupted as its contents are modified, and so it is called unstable.
75 * KSM solves this problem by several techniques:
77 * 1) The unstable tree is flushed every time KSM completes scanning all
78 * memory areas, and then the tree is rebuilt again from the beginning.
79 * 2) KSM will only insert into the unstable tree, pages whose hash value
80 * has not changed since the previous scan of all memory areas.
81 * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
82 * colors of the nodes and not on their contents, assuring that even when
83 * the tree gets "corrupted" it won't get out of balance, so scanning time
84 * remains the same (also, searching and inserting nodes in an rbtree uses
85 * the same algorithm, so we have no overhead when we flush and rebuild).
86 * 4) KSM never flushes the stable tree, which means that even if it were to
87 * take 10 attempts to find a page in the unstable tree, once it is found,
88 * it is secured in the stable tree. (When we scan a new page, we first
89 * compare it against the stable tree, and then against the unstable tree.)
91 * If the merge_across_nodes tunable is unset, then KSM maintains multiple
92 * stable trees and multiple unstable trees: one of each for each NUMA node.
96 * struct mm_slot - ksm information per mm that is being scanned
97 * @link: link to the mm_slots hash list
98 * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
99 * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
100 * @mm: the mm that this information is valid for
103 struct hlist_node link;
104 struct list_head mm_list;
105 struct rmap_item *rmap_list;
106 struct mm_struct *mm;
110 * struct ksm_scan - cursor for scanning
111 * @mm_slot: the current mm_slot we are scanning
112 * @address: the next address inside that to be scanned
113 * @rmap_list: link to the next rmap to be scanned in the rmap_list
114 * @seqnr: count of completed full scans (needed when removing unstable node)
116 * There is only the one ksm_scan instance of this cursor structure.
119 struct mm_slot *mm_slot;
120 unsigned long address;
121 struct rmap_item **rmap_list;
126 * struct stable_node - node of the stable rbtree
127 * @node: rb node of this ksm page in the stable tree
128 * @head: (overlaying parent) &migrate_nodes indicates temporarily on that list
129 * @list: linked into migrate_nodes, pending placement in the proper node tree
130 * @hlist: hlist head of rmap_items using this ksm page
131 * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid)
132 * @nid: NUMA node id of stable tree in which linked (may not match kpfn)
136 struct rb_node node; /* when node of stable tree */
137 struct { /* when listed for migration */
138 struct list_head *head;
139 struct list_head list;
142 struct hlist_head hlist;
150 * struct rmap_item - reverse mapping item for virtual addresses
151 * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
152 * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
153 * @nid: NUMA node id of unstable tree in which linked (may not match page)
154 * @mm: the memory structure this rmap_item is pointing into
155 * @address: the virtual address this rmap_item tracks (+ flags in low bits)
156 * @oldchecksum: previous checksum of the page at that virtual address
157 * @node: rb node of this rmap_item in the unstable tree
158 * @head: pointer to stable_node heading this list in the stable tree
159 * @hlist: link into hlist of rmap_items hanging off that stable_node
162 struct rmap_item *rmap_list;
164 struct anon_vma *anon_vma; /* when stable */
166 int nid; /* when node of unstable tree */
169 struct mm_struct *mm;
170 unsigned long address; /* + low bits used for flags below */
171 unsigned int oldchecksum; /* when unstable */
173 struct rb_node node; /* when node of unstable tree */
174 struct { /* when listed from stable tree */
175 struct stable_node *head;
176 struct hlist_node hlist;
181 #define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */
182 #define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */
183 #define STABLE_FLAG 0x200 /* is listed from the stable tree */
185 /* The stable and unstable tree heads */
186 static struct rb_root one_stable_tree[1] = { RB_ROOT };
187 static struct rb_root one_unstable_tree[1] = { RB_ROOT };
188 static struct rb_root *root_stable_tree = one_stable_tree;
189 static struct rb_root *root_unstable_tree = one_unstable_tree;
191 /* Recently migrated nodes of stable tree, pending proper placement */
192 static LIST_HEAD(migrate_nodes);
194 #define MM_SLOTS_HASH_BITS 10
195 static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
197 static struct mm_slot ksm_mm_head = {
198 .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
200 static struct ksm_scan ksm_scan = {
201 .mm_slot = &ksm_mm_head,
204 static struct kmem_cache *rmap_item_cache;
205 static struct kmem_cache *stable_node_cache;
206 static struct kmem_cache *mm_slot_cache;
208 /* The number of nodes in the stable tree */
209 static unsigned long ksm_pages_shared;
211 /* The number of page slots additionally sharing those nodes */
212 static unsigned long ksm_pages_sharing;
214 /* The number of nodes in the unstable tree */
215 static unsigned long ksm_pages_unshared;
217 /* The number of rmap_items in use: to calculate pages_volatile */
218 static unsigned long ksm_rmap_items;
220 /* Number of pages ksmd should scan in one batch */
221 static unsigned int ksm_thread_pages_to_scan = 100;
223 /* Milliseconds ksmd should sleep between batches */
224 static unsigned int ksm_thread_sleep_millisecs = 20;
227 /* Zeroed when merging across nodes is not allowed */
228 static unsigned int ksm_merge_across_nodes = 1;
229 static int ksm_nr_node_ids = 1;
231 #define ksm_merge_across_nodes 1U
232 #define ksm_nr_node_ids 1
235 #define KSM_RUN_STOP 0
236 #define KSM_RUN_MERGE 1
237 #define KSM_RUN_UNMERGE 2
238 #define KSM_RUN_OFFLINE 4
239 static unsigned long ksm_run = KSM_RUN_STOP;
240 static void wait_while_offlining(void);
242 static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
243 static DEFINE_MUTEX(ksm_thread_mutex);
244 static DEFINE_SPINLOCK(ksm_mmlist_lock);
246 #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
247 sizeof(struct __struct), __alignof__(struct __struct),\
250 static int __init ksm_slab_init(void)
252 rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
253 if (!rmap_item_cache)
256 stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
257 if (!stable_node_cache)
260 mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
267 kmem_cache_destroy(stable_node_cache);
269 kmem_cache_destroy(rmap_item_cache);
274 static void __init ksm_slab_free(void)
276 kmem_cache_destroy(mm_slot_cache);
277 kmem_cache_destroy(stable_node_cache);
278 kmem_cache_destroy(rmap_item_cache);
279 mm_slot_cache = NULL;
282 static inline struct rmap_item *alloc_rmap_item(void)
284 struct rmap_item *rmap_item;
286 rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL |
287 __GFP_NORETRY | __GFP_NOWARN);
293 static inline void free_rmap_item(struct rmap_item *rmap_item)
296 rmap_item->mm = NULL; /* debug safety */
297 kmem_cache_free(rmap_item_cache, rmap_item);
300 static inline struct stable_node *alloc_stable_node(void)
303 * The allocation can take too long with GFP_KERNEL when memory is under
304 * pressure, which may lead to hung task warnings. Adding __GFP_HIGH
305 * grants access to memory reserves, helping to avoid this problem.
307 return kmem_cache_alloc(stable_node_cache, GFP_KERNEL | __GFP_HIGH);
310 static inline void free_stable_node(struct stable_node *stable_node)
312 kmem_cache_free(stable_node_cache, stable_node);
315 static inline struct mm_slot *alloc_mm_slot(void)
317 if (!mm_slot_cache) /* initialization failed */
319 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
322 static inline void free_mm_slot(struct mm_slot *mm_slot)
324 kmem_cache_free(mm_slot_cache, mm_slot);
327 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
329 struct mm_slot *slot;
331 hash_for_each_possible(mm_slots_hash, slot, link, (unsigned long)mm)
338 static void insert_to_mm_slots_hash(struct mm_struct *mm,
339 struct mm_slot *mm_slot)
342 hash_add(mm_slots_hash, &mm_slot->link, (unsigned long)mm);
346 * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
347 * page tables after it has passed through ksm_exit() - which, if necessary,
348 * takes mmap_sem briefly to serialize against them. ksm_exit() does not set
349 * a special flag: they can just back out as soon as mm_users goes to zero.
350 * ksm_test_exit() is used throughout to make this test for exit: in some
351 * places for correctness, in some places just to avoid unnecessary work.
353 static inline bool ksm_test_exit(struct mm_struct *mm)
355 return atomic_read(&mm->mm_users) == 0;
359 * We use break_ksm to break COW on a ksm page: it's a stripped down
361 * if (get_user_pages(addr, 1, 1, 1, &page, NULL) == 1)
364 * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
365 * in case the application has unmapped and remapped mm,addr meanwhile.
366 * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP
367 * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
369 * FAULT_FLAG/FOLL_REMOTE are because we do this outside the context
370 * of the process that owns 'vma'. We also do not want to enforce
371 * protection keys here anyway.
373 static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
380 page = follow_page(vma, addr,
381 FOLL_GET | FOLL_MIGRATION | FOLL_REMOTE);
382 if (IS_ERR_OR_NULL(page))
385 ret = handle_mm_fault(vma, addr,
386 FAULT_FLAG_WRITE | FAULT_FLAG_REMOTE);
388 ret = VM_FAULT_WRITE;
390 } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV | VM_FAULT_OOM)));
392 * We must loop because handle_mm_fault() may back out if there's
393 * any difficulty e.g. if pte accessed bit gets updated concurrently.
395 * VM_FAULT_WRITE is what we have been hoping for: it indicates that
396 * COW has been broken, even if the vma does not permit VM_WRITE;
397 * but note that a concurrent fault might break PageKsm for us.
399 * VM_FAULT_SIGBUS could occur if we race with truncation of the
400 * backing file, which also invalidates anonymous pages: that's
401 * okay, that truncation will have unmapped the PageKsm for us.
403 * VM_FAULT_OOM: at the time of writing (late July 2009), setting
404 * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
405 * current task has TIF_MEMDIE set, and will be OOM killed on return
406 * to user; and ksmd, having no mm, would never be chosen for that.
408 * But if the mm is in a limited mem_cgroup, then the fault may fail
409 * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
410 * even ksmd can fail in this way - though it's usually breaking ksm
411 * just to undo a merge it made a moment before, so unlikely to oom.
413 * That's a pity: we might therefore have more kernel pages allocated
414 * than we're counting as nodes in the stable tree; but ksm_do_scan
415 * will retry to break_cow on each pass, so should recover the page
416 * in due course. The important thing is to not let VM_MERGEABLE
417 * be cleared while any such pages might remain in the area.
419 return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
422 static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm,
425 struct vm_area_struct *vma;
426 if (ksm_test_exit(mm))
428 vma = find_vma(mm, addr);
429 if (!vma || vma->vm_start > addr)
431 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
436 static void break_cow(struct rmap_item *rmap_item)
438 struct mm_struct *mm = rmap_item->mm;
439 unsigned long addr = rmap_item->address;
440 struct vm_area_struct *vma;
443 * It is not an accident that whenever we want to break COW
444 * to undo, we also need to drop a reference to the anon_vma.
446 put_anon_vma(rmap_item->anon_vma);
448 down_read(&mm->mmap_sem);
449 vma = find_mergeable_vma(mm, addr);
451 break_ksm(vma, addr);
452 up_read(&mm->mmap_sem);
455 static struct page *get_mergeable_page(struct rmap_item *rmap_item)
457 struct mm_struct *mm = rmap_item->mm;
458 unsigned long addr = rmap_item->address;
459 struct vm_area_struct *vma;
462 down_read(&mm->mmap_sem);
463 vma = find_mergeable_vma(mm, addr);
467 page = follow_page(vma, addr, FOLL_GET);
468 if (IS_ERR_OR_NULL(page))
470 if (PageAnon(page)) {
471 flush_anon_page(vma, page, addr);
472 flush_dcache_page(page);
478 up_read(&mm->mmap_sem);
483 * This helper is used for getting right index into array of tree roots.
484 * When merge_across_nodes knob is set to 1, there are only two rb-trees for
485 * stable and unstable pages from all nodes with roots in index 0. Otherwise,
486 * every node has its own stable and unstable tree.
488 static inline int get_kpfn_nid(unsigned long kpfn)
490 return ksm_merge_across_nodes ? 0 : NUMA(pfn_to_nid(kpfn));
493 static void remove_node_from_stable_tree(struct stable_node *stable_node)
495 struct rmap_item *rmap_item;
497 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
498 if (rmap_item->hlist.next)
502 put_anon_vma(rmap_item->anon_vma);
503 rmap_item->address &= PAGE_MASK;
507 if (stable_node->head == &migrate_nodes)
508 list_del(&stable_node->list);
510 rb_erase(&stable_node->node,
511 root_stable_tree + NUMA(stable_node->nid));
512 free_stable_node(stable_node);
516 * get_ksm_page: checks if the page indicated by the stable node
517 * is still its ksm page, despite having held no reference to it.
518 * In which case we can trust the content of the page, and it
519 * returns the gotten page; but if the page has now been zapped,
520 * remove the stale node from the stable tree and return NULL.
521 * But beware, the stable node's page might be being migrated.
523 * You would expect the stable_node to hold a reference to the ksm page.
524 * But if it increments the page's count, swapping out has to wait for
525 * ksmd to come around again before it can free the page, which may take
526 * seconds or even minutes: much too unresponsive. So instead we use a
527 * "keyhole reference": access to the ksm page from the stable node peeps
528 * out through its keyhole to see if that page still holds the right key,
529 * pointing back to this stable node. This relies on freeing a PageAnon
530 * page to reset its page->mapping to NULL, and relies on no other use of
531 * a page to put something that might look like our key in page->mapping.
532 * is on its way to being freed; but it is an anomaly to bear in mind.
534 static struct page *get_ksm_page(struct stable_node *stable_node, bool lock_it)
537 void *expected_mapping;
540 expected_mapping = (void *)((unsigned long)stable_node |
543 kpfn = READ_ONCE(stable_node->kpfn);
544 page = pfn_to_page(kpfn);
547 * page is computed from kpfn, so on most architectures reading
548 * page->mapping is naturally ordered after reading node->kpfn,
549 * but on Alpha we need to be more careful.
551 smp_read_barrier_depends();
552 if (READ_ONCE(page->mapping) != expected_mapping)
556 * We cannot do anything with the page while its refcount is 0.
557 * Usually 0 means free, or tail of a higher-order page: in which
558 * case this node is no longer referenced, and should be freed;
559 * however, it might mean that the page is under page_freeze_refs().
560 * The __remove_mapping() case is easy, again the node is now stale;
561 * but if page is swapcache in migrate_page_move_mapping(), it might
562 * still be our page, in which case it's essential to keep the node.
564 while (!get_page_unless_zero(page)) {
566 * Another check for page->mapping != expected_mapping would
567 * work here too. We have chosen the !PageSwapCache test to
568 * optimize the common case, when the page is or is about to
569 * be freed: PageSwapCache is cleared (under spin_lock_irq)
570 * in the freeze_refs section of __remove_mapping(); but Anon
571 * page->mapping reset to NULL later, in free_pages_prepare().
573 if (!PageSwapCache(page))
578 if (READ_ONCE(page->mapping) != expected_mapping) {
585 if (READ_ONCE(page->mapping) != expected_mapping) {
595 * We come here from above when page->mapping or !PageSwapCache
596 * suggests that the node is stale; but it might be under migration.
597 * We need smp_rmb(), matching the smp_wmb() in ksm_migrate_page(),
598 * before checking whether node->kpfn has been changed.
601 if (READ_ONCE(stable_node->kpfn) != kpfn)
603 remove_node_from_stable_tree(stable_node);
608 * Removing rmap_item from stable or unstable tree.
609 * This function will clean the information from the stable/unstable tree.
611 static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
613 if (rmap_item->address & STABLE_FLAG) {
614 struct stable_node *stable_node;
617 stable_node = rmap_item->head;
618 page = get_ksm_page(stable_node, true);
622 hlist_del(&rmap_item->hlist);
626 if (!hlist_empty(&stable_node->hlist))
631 put_anon_vma(rmap_item->anon_vma);
632 rmap_item->head = NULL;
633 rmap_item->address &= PAGE_MASK;
635 } else if (rmap_item->address & UNSTABLE_FLAG) {
638 * Usually ksmd can and must skip the rb_erase, because
639 * root_unstable_tree was already reset to RB_ROOT.
640 * But be careful when an mm is exiting: do the rb_erase
641 * if this rmap_item was inserted by this scan, rather
642 * than left over from before.
644 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
647 rb_erase(&rmap_item->node,
648 root_unstable_tree + NUMA(rmap_item->nid));
649 ksm_pages_unshared--;
650 rmap_item->address &= PAGE_MASK;
653 cond_resched(); /* we're called from many long loops */
656 static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
657 struct rmap_item **rmap_list)
660 struct rmap_item *rmap_item = *rmap_list;
661 *rmap_list = rmap_item->rmap_list;
662 remove_rmap_item_from_tree(rmap_item);
663 free_rmap_item(rmap_item);
668 * Though it's very tempting to unmerge rmap_items from stable tree rather
669 * than check every pte of a given vma, the locking doesn't quite work for
670 * that - an rmap_item is assigned to the stable tree after inserting ksm
671 * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing
672 * rmap_items from parent to child at fork time (so as not to waste time
673 * if exit comes before the next scan reaches it).
675 * Similarly, although we'd like to remove rmap_items (so updating counts
676 * and freeing memory) when unmerging an area, it's easier to leave that
677 * to the next pass of ksmd - consider, for example, how ksmd might be
678 * in cmp_and_merge_page on one of the rmap_items we would be removing.
680 static int unmerge_ksm_pages(struct vm_area_struct *vma,
681 unsigned long start, unsigned long end)
686 for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
687 if (ksm_test_exit(vma->vm_mm))
689 if (signal_pending(current))
692 err = break_ksm(vma, addr);
699 * Only called through the sysfs control interface:
701 static int remove_stable_node(struct stable_node *stable_node)
706 page = get_ksm_page(stable_node, true);
709 * get_ksm_page did remove_node_from_stable_tree itself.
715 * Page could be still mapped if this races with __mmput() running in
716 * between ksm_exit() and exit_mmap(). Just refuse to let
717 * merge_across_nodes/max_page_sharing be switched.
720 if (!page_mapped(page)) {
722 * The stable node did not yet appear stale to get_ksm_page(),
723 * since that allows for an unmapped ksm page to be recognized
724 * right up until it is freed; but the node is safe to remove.
725 * This page might be in a pagevec waiting to be freed,
726 * or it might be PageSwapCache (perhaps under writeback),
727 * or it might have been removed from swapcache a moment ago.
729 set_page_stable_node(page, NULL);
730 remove_node_from_stable_tree(stable_node);
739 static int remove_all_stable_nodes(void)
741 struct stable_node *stable_node, *next;
745 for (nid = 0; nid < ksm_nr_node_ids; nid++) {
746 while (root_stable_tree[nid].rb_node) {
747 stable_node = rb_entry(root_stable_tree[nid].rb_node,
748 struct stable_node, node);
749 if (remove_stable_node(stable_node)) {
751 break; /* proceed to next nid */
756 list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
757 if (remove_stable_node(stable_node))
764 static int unmerge_and_remove_all_rmap_items(void)
766 struct mm_slot *mm_slot;
767 struct mm_struct *mm;
768 struct vm_area_struct *vma;
771 spin_lock(&ksm_mmlist_lock);
772 ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
773 struct mm_slot, mm_list);
774 spin_unlock(&ksm_mmlist_lock);
776 for (mm_slot = ksm_scan.mm_slot;
777 mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
779 down_read(&mm->mmap_sem);
780 for (vma = mm->mmap; vma; vma = vma->vm_next) {
781 if (ksm_test_exit(mm))
783 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
785 err = unmerge_ksm_pages(vma,
786 vma->vm_start, vma->vm_end);
791 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
792 up_read(&mm->mmap_sem);
794 spin_lock(&ksm_mmlist_lock);
795 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
796 struct mm_slot, mm_list);
797 if (ksm_test_exit(mm)) {
798 hash_del(&mm_slot->link);
799 list_del(&mm_slot->mm_list);
800 spin_unlock(&ksm_mmlist_lock);
802 free_mm_slot(mm_slot);
803 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
806 spin_unlock(&ksm_mmlist_lock);
809 /* Clean up stable nodes, but don't worry if some are still busy */
810 remove_all_stable_nodes();
815 up_read(&mm->mmap_sem);
816 spin_lock(&ksm_mmlist_lock);
817 ksm_scan.mm_slot = &ksm_mm_head;
818 spin_unlock(&ksm_mmlist_lock);
821 #endif /* CONFIG_SYSFS */
823 static u32 calc_checksum(struct page *page)
826 void *addr = kmap_atomic(page);
827 checksum = jhash2(addr, PAGE_SIZE / 4, 17);
832 static int memcmp_pages(struct page *page1, struct page *page2)
837 addr1 = kmap_atomic(page1);
838 addr2 = kmap_atomic(page2);
839 ret = memcmp(addr1, addr2, PAGE_SIZE);
840 kunmap_atomic(addr2);
841 kunmap_atomic(addr1);
845 static inline int pages_identical(struct page *page1, struct page *page2)
847 return !memcmp_pages(page1, page2);
850 static int write_protect_page(struct vm_area_struct *vma, struct page *page,
853 struct mm_struct *mm = vma->vm_mm;
859 unsigned long mmun_start; /* For mmu_notifiers */
860 unsigned long mmun_end; /* For mmu_notifiers */
862 addr = page_address_in_vma(page, vma);
866 BUG_ON(PageTransCompound(page));
869 mmun_end = addr + PAGE_SIZE;
870 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
872 ptep = page_check_address(page, mm, addr, &ptl, 0);
876 if (pte_write(*ptep) || pte_dirty(*ptep)) {
879 swapped = PageSwapCache(page);
880 flush_cache_page(vma, addr, page_to_pfn(page));
882 * Ok this is tricky, when get_user_pages_fast() run it doesn't
883 * take any lock, therefore the check that we are going to make
884 * with the pagecount against the mapcount is racey and
885 * O_DIRECT can happen right after the check.
886 * So we clear the pte and flush the tlb before the check
887 * this assure us that no O_DIRECT can happen after the check
888 * or in the middle of the check.
890 entry = ptep_clear_flush_notify(vma, addr, ptep);
892 * Check that no O_DIRECT or similar I/O is in progress on the
895 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
896 set_pte_at(mm, addr, ptep, entry);
899 if (pte_dirty(entry))
900 set_page_dirty(page);
901 entry = pte_mkclean(pte_wrprotect(entry));
902 set_pte_at_notify(mm, addr, ptep, entry);
908 pte_unmap_unlock(ptep, ptl);
910 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
916 * replace_page - replace page in vma by new ksm page
917 * @vma: vma that holds the pte pointing to page
918 * @page: the page we are replacing by kpage
919 * @kpage: the ksm page we replace page by
920 * @orig_pte: the original value of the pte
922 * Returns 0 on success, -EFAULT on failure.
924 static int replace_page(struct vm_area_struct *vma, struct page *page,
925 struct page *kpage, pte_t orig_pte)
927 struct mm_struct *mm = vma->vm_mm;
933 unsigned long mmun_start; /* For mmu_notifiers */
934 unsigned long mmun_end; /* For mmu_notifiers */
936 addr = page_address_in_vma(page, vma);
940 pmd = mm_find_pmd(mm, addr);
945 mmun_end = addr + PAGE_SIZE;
946 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
948 ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
949 if (!pte_same(*ptep, orig_pte)) {
950 pte_unmap_unlock(ptep, ptl);
955 page_add_anon_rmap(kpage, vma, addr, false);
957 flush_cache_page(vma, addr, pte_pfn(*ptep));
958 ptep_clear_flush_notify(vma, addr, ptep);
959 set_pte_at_notify(mm, addr, ptep, mk_pte(kpage, vma->vm_page_prot));
961 page_remove_rmap(page, false);
962 if (!page_mapped(page))
963 try_to_free_swap(page);
966 pte_unmap_unlock(ptep, ptl);
969 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
975 * try_to_merge_one_page - take two pages and merge them into one
976 * @vma: the vma that holds the pte pointing to page
977 * @page: the PageAnon page that we want to replace with kpage
978 * @kpage: the PageKsm page that we want to map instead of page,
979 * or NULL the first time when we want to use page as kpage.
981 * This function returns 0 if the pages were merged, -EFAULT otherwise.
983 static int try_to_merge_one_page(struct vm_area_struct *vma,
984 struct page *page, struct page *kpage)
986 pte_t orig_pte = __pte(0);
989 if (page == kpage) /* ksm page forked */
996 * We need the page lock to read a stable PageSwapCache in
997 * write_protect_page(). We use trylock_page() instead of
998 * lock_page() because we don't want to wait here - we
999 * prefer to continue scanning and merging different pages,
1000 * then come back to this page when it is unlocked.
1002 if (!trylock_page(page))
1005 if (PageTransCompound(page)) {
1006 if (split_huge_page(page))
1011 * If this anonymous page is mapped only here, its pte may need
1012 * to be write-protected. If it's mapped elsewhere, all of its
1013 * ptes are necessarily already write-protected. But in either
1014 * case, we need to lock and check page_count is not raised.
1016 if (write_protect_page(vma, page, &orig_pte) == 0) {
1019 * While we hold page lock, upgrade page from
1020 * PageAnon+anon_vma to PageKsm+NULL stable_node:
1021 * stable_tree_insert() will update stable_node.
1023 set_page_stable_node(page, NULL);
1024 mark_page_accessed(page);
1026 * Page reclaim just frees a clean page with no dirty
1027 * ptes: make sure that the ksm page would be swapped.
1029 if (!PageDirty(page))
1032 } else if (pages_identical(page, kpage))
1033 err = replace_page(vma, page, kpage, orig_pte);
1036 if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
1037 munlock_vma_page(page);
1038 if (!PageMlocked(kpage)) {
1041 mlock_vma_page(kpage);
1042 page = kpage; /* for final unlock */
1053 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
1054 * but no new kernel page is allocated: kpage must already be a ksm page.
1056 * This function returns 0 if the pages were merged, -EFAULT otherwise.
1058 static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
1059 struct page *page, struct page *kpage)
1061 struct mm_struct *mm = rmap_item->mm;
1062 struct vm_area_struct *vma;
1065 down_read(&mm->mmap_sem);
1066 vma = find_mergeable_vma(mm, rmap_item->address);
1070 err = try_to_merge_one_page(vma, page, kpage);
1074 /* Unstable nid is in union with stable anon_vma: remove first */
1075 remove_rmap_item_from_tree(rmap_item);
1077 /* Must get reference to anon_vma while still holding mmap_sem */
1078 rmap_item->anon_vma = vma->anon_vma;
1079 get_anon_vma(vma->anon_vma);
1081 up_read(&mm->mmap_sem);
1086 * try_to_merge_two_pages - take two identical pages and prepare them
1087 * to be merged into one page.
1089 * This function returns the kpage if we successfully merged two identical
1090 * pages into one ksm page, NULL otherwise.
1092 * Note that this function upgrades page to ksm page: if one of the pages
1093 * is already a ksm page, try_to_merge_with_ksm_page should be used.
1095 static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
1097 struct rmap_item *tree_rmap_item,
1098 struct page *tree_page)
1102 err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
1104 err = try_to_merge_with_ksm_page(tree_rmap_item,
1107 * If that fails, we have a ksm page with only one pte
1108 * pointing to it: so break it.
1111 break_cow(rmap_item);
1113 return err ? NULL : page;
1117 * stable_tree_search - search for page inside the stable tree
1119 * This function checks if there is a page inside the stable tree
1120 * with identical content to the page that we are scanning right now.
1122 * This function returns the stable tree node of identical content if found,
1125 static struct page *stable_tree_search(struct page *page)
1128 struct rb_root *root;
1129 struct rb_node **new;
1130 struct rb_node *parent;
1131 struct stable_node *stable_node;
1132 struct stable_node *page_node;
1134 page_node = page_stable_node(page);
1135 if (page_node && page_node->head != &migrate_nodes) {
1136 /* ksm page forked */
1141 nid = get_kpfn_nid(page_to_pfn(page));
1142 root = root_stable_tree + nid;
1144 new = &root->rb_node;
1148 struct page *tree_page;
1152 stable_node = rb_entry(*new, struct stable_node, node);
1153 tree_page = get_ksm_page(stable_node, false);
1156 * If we walked over a stale stable_node,
1157 * get_ksm_page() will call rb_erase() and it
1158 * may rebalance the tree from under us. So
1159 * restart the search from scratch. Returning
1160 * NULL would be safe too, but we'd generate
1161 * false negative insertions just because some
1162 * stable_node was stale.
1167 ret = memcmp_pages(page, tree_page);
1168 put_page(tree_page);
1172 new = &parent->rb_left;
1174 new = &parent->rb_right;
1177 * Lock and unlock the stable_node's page (which
1178 * might already have been migrated) so that page
1179 * migration is sure to notice its raised count.
1180 * It would be more elegant to return stable_node
1181 * than kpage, but that involves more changes.
1183 tree_page = get_ksm_page(stable_node, true);
1185 unlock_page(tree_page);
1186 if (get_kpfn_nid(stable_node->kpfn) !=
1187 NUMA(stable_node->nid)) {
1188 put_page(tree_page);
1194 * There is now a place for page_node, but the tree may
1195 * have been rebalanced, so re-evaluate parent and new.
1206 list_del(&page_node->list);
1207 DO_NUMA(page_node->nid = nid);
1208 rb_link_node(&page_node->node, parent, new);
1209 rb_insert_color(&page_node->node, root);
1215 list_del(&page_node->list);
1216 DO_NUMA(page_node->nid = nid);
1217 rb_replace_node(&stable_node->node, &page_node->node, root);
1220 rb_erase(&stable_node->node, root);
1223 stable_node->head = &migrate_nodes;
1224 list_add(&stable_node->list, stable_node->head);
1229 * stable_tree_insert - insert stable tree node pointing to new ksm page
1230 * into the stable tree.
1232 * This function returns the stable tree node just allocated on success,
1235 static struct stable_node *stable_tree_insert(struct page *kpage)
1239 struct rb_root *root;
1240 struct rb_node **new;
1241 struct rb_node *parent;
1242 struct stable_node *stable_node;
1244 kpfn = page_to_pfn(kpage);
1245 nid = get_kpfn_nid(kpfn);
1246 root = root_stable_tree + nid;
1249 new = &root->rb_node;
1252 struct page *tree_page;
1256 stable_node = rb_entry(*new, struct stable_node, node);
1257 tree_page = get_ksm_page(stable_node, false);
1260 * If we walked over a stale stable_node,
1261 * get_ksm_page() will call rb_erase() and it
1262 * may rebalance the tree from under us. So
1263 * restart the search from scratch. Returning
1264 * NULL would be safe too, but we'd generate
1265 * false negative insertions just because some
1266 * stable_node was stale.
1271 ret = memcmp_pages(kpage, tree_page);
1272 put_page(tree_page);
1276 new = &parent->rb_left;
1278 new = &parent->rb_right;
1281 * It is not a bug that stable_tree_search() didn't
1282 * find this node: because at that time our page was
1283 * not yet write-protected, so may have changed since.
1289 stable_node = alloc_stable_node();
1293 INIT_HLIST_HEAD(&stable_node->hlist);
1294 stable_node->kpfn = kpfn;
1295 set_page_stable_node(kpage, stable_node);
1296 DO_NUMA(stable_node->nid = nid);
1297 rb_link_node(&stable_node->node, parent, new);
1298 rb_insert_color(&stable_node->node, root);
1304 * unstable_tree_search_insert - search for identical page,
1305 * else insert rmap_item into the unstable tree.
1307 * This function searches for a page in the unstable tree identical to the
1308 * page currently being scanned; and if no identical page is found in the
1309 * tree, we insert rmap_item as a new object into the unstable tree.
1311 * This function returns pointer to rmap_item found to be identical
1312 * to the currently scanned page, NULL otherwise.
1314 * This function does both searching and inserting, because they share
1315 * the same walking algorithm in an rbtree.
1318 struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1320 struct page **tree_pagep)
1322 struct rb_node **new;
1323 struct rb_root *root;
1324 struct rb_node *parent = NULL;
1327 nid = get_kpfn_nid(page_to_pfn(page));
1328 root = root_unstable_tree + nid;
1329 new = &root->rb_node;
1332 struct rmap_item *tree_rmap_item;
1333 struct page *tree_page;
1337 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
1338 tree_page = get_mergeable_page(tree_rmap_item);
1343 * Don't substitute a ksm page for a forked page.
1345 if (page == tree_page) {
1346 put_page(tree_page);
1350 ret = memcmp_pages(page, tree_page);
1354 put_page(tree_page);
1355 new = &parent->rb_left;
1356 } else if (ret > 0) {
1357 put_page(tree_page);
1358 new = &parent->rb_right;
1359 } else if (!ksm_merge_across_nodes &&
1360 page_to_nid(tree_page) != nid) {
1362 * If tree_page has been migrated to another NUMA node,
1363 * it will be flushed out and put in the right unstable
1364 * tree next time: only merge with it when across_nodes.
1366 put_page(tree_page);
1369 *tree_pagep = tree_page;
1370 return tree_rmap_item;
1374 rmap_item->address |= UNSTABLE_FLAG;
1375 rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1376 DO_NUMA(rmap_item->nid = nid);
1377 rb_link_node(&rmap_item->node, parent, new);
1378 rb_insert_color(&rmap_item->node, root);
1380 ksm_pages_unshared++;
1385 * stable_tree_append - add another rmap_item to the linked list of
1386 * rmap_items hanging off a given node of the stable tree, all sharing
1387 * the same ksm page.
1389 static void stable_tree_append(struct rmap_item *rmap_item,
1390 struct stable_node *stable_node)
1392 rmap_item->head = stable_node;
1393 rmap_item->address |= STABLE_FLAG;
1394 hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
1396 if (rmap_item->hlist.next)
1397 ksm_pages_sharing++;
1403 * cmp_and_merge_page - first see if page can be merged into the stable tree;
1404 * if not, compare checksum to previous and if it's the same, see if page can
1405 * be inserted into the unstable tree, or merged with a page already there and
1406 * both transferred to the stable tree.
1408 * @page: the page that we are searching identical page to.
1409 * @rmap_item: the reverse mapping into the virtual address of this page
1411 static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1413 struct rmap_item *tree_rmap_item;
1414 struct page *tree_page = NULL;
1415 struct stable_node *stable_node;
1417 unsigned int checksum;
1420 stable_node = page_stable_node(page);
1422 if (stable_node->head != &migrate_nodes &&
1423 get_kpfn_nid(stable_node->kpfn) != NUMA(stable_node->nid)) {
1424 rb_erase(&stable_node->node,
1425 root_stable_tree + NUMA(stable_node->nid));
1426 stable_node->head = &migrate_nodes;
1427 list_add(&stable_node->list, stable_node->head);
1429 if (stable_node->head != &migrate_nodes &&
1430 rmap_item->head == stable_node)
1434 /* We first start with searching the page inside the stable tree */
1435 kpage = stable_tree_search(page);
1436 if (kpage == page && rmap_item->head == stable_node) {
1441 remove_rmap_item_from_tree(rmap_item);
1444 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
1447 * The page was successfully merged:
1448 * add its rmap_item to the stable tree.
1451 stable_tree_append(rmap_item, page_stable_node(kpage));
1459 * If the hash value of the page has changed from the last time
1460 * we calculated it, this page is changing frequently: therefore we
1461 * don't want to insert it in the unstable tree, and we don't want
1462 * to waste our time searching for something identical to it there.
1464 checksum = calc_checksum(page);
1465 if (rmap_item->oldchecksum != checksum) {
1466 rmap_item->oldchecksum = checksum;
1471 unstable_tree_search_insert(rmap_item, page, &tree_page);
1472 if (tree_rmap_item) {
1475 kpage = try_to_merge_two_pages(rmap_item, page,
1476 tree_rmap_item, tree_page);
1478 * If both pages we tried to merge belong to the same compound
1479 * page, then we actually ended up increasing the reference
1480 * count of the same compound page twice, and split_huge_page
1482 * Here we set a flag if that happened, and we use it later to
1483 * try split_huge_page again. Since we call put_page right
1484 * afterwards, the reference count will be correct and
1485 * split_huge_page should succeed.
1487 split = PageTransCompound(page)
1488 && compound_head(page) == compound_head(tree_page);
1489 put_page(tree_page);
1492 * The pages were successfully merged: insert new
1493 * node in the stable tree and add both rmap_items.
1496 stable_node = stable_tree_insert(kpage);
1498 stable_tree_append(tree_rmap_item, stable_node);
1499 stable_tree_append(rmap_item, stable_node);
1504 * If we fail to insert the page into the stable tree,
1505 * we will have 2 virtual addresses that are pointing
1506 * to a ksm page left outside the stable tree,
1507 * in which case we need to break_cow on both.
1510 break_cow(tree_rmap_item);
1511 break_cow(rmap_item);
1515 * We are here if we tried to merge two pages and
1516 * failed because they both belonged to the same
1517 * compound page. We will split the page now, but no
1518 * merging will take place.
1519 * We do not want to add the cost of a full lock; if
1520 * the page is locked, it is better to skip it and
1521 * perhaps try again later.
1523 if (!trylock_page(page))
1525 split_huge_page(page);
1531 static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
1532 struct rmap_item **rmap_list,
1535 struct rmap_item *rmap_item;
1537 while (*rmap_list) {
1538 rmap_item = *rmap_list;
1539 if ((rmap_item->address & PAGE_MASK) == addr)
1541 if (rmap_item->address > addr)
1543 *rmap_list = rmap_item->rmap_list;
1544 remove_rmap_item_from_tree(rmap_item);
1545 free_rmap_item(rmap_item);
1548 rmap_item = alloc_rmap_item();
1550 /* It has already been zeroed */
1551 rmap_item->mm = mm_slot->mm;
1552 rmap_item->address = addr;
1553 rmap_item->rmap_list = *rmap_list;
1554 *rmap_list = rmap_item;
1559 static struct rmap_item *scan_get_next_rmap_item(struct page **page)
1561 struct mm_struct *mm;
1562 struct mm_slot *slot;
1563 struct vm_area_struct *vma;
1564 struct rmap_item *rmap_item;
1567 if (list_empty(&ksm_mm_head.mm_list))
1570 slot = ksm_scan.mm_slot;
1571 if (slot == &ksm_mm_head) {
1573 * A number of pages can hang around indefinitely on per-cpu
1574 * pagevecs, raised page count preventing write_protect_page
1575 * from merging them. Though it doesn't really matter much,
1576 * it is puzzling to see some stuck in pages_volatile until
1577 * other activity jostles them out, and they also prevented
1578 * LTP's KSM test from succeeding deterministically; so drain
1579 * them here (here rather than on entry to ksm_do_scan(),
1580 * so we don't IPI too often when pages_to_scan is set low).
1582 lru_add_drain_all();
1585 * Whereas stale stable_nodes on the stable_tree itself
1586 * get pruned in the regular course of stable_tree_search(),
1587 * those moved out to the migrate_nodes list can accumulate:
1588 * so prune them once before each full scan.
1590 if (!ksm_merge_across_nodes) {
1591 struct stable_node *stable_node, *next;
1594 list_for_each_entry_safe(stable_node, next,
1595 &migrate_nodes, list) {
1596 page = get_ksm_page(stable_node, false);
1603 for (nid = 0; nid < ksm_nr_node_ids; nid++)
1604 root_unstable_tree[nid] = RB_ROOT;
1606 spin_lock(&ksm_mmlist_lock);
1607 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1608 ksm_scan.mm_slot = slot;
1609 spin_unlock(&ksm_mmlist_lock);
1611 * Although we tested list_empty() above, a racing __ksm_exit
1612 * of the last mm on the list may have removed it since then.
1614 if (slot == &ksm_mm_head)
1617 ksm_scan.address = 0;
1618 ksm_scan.rmap_list = &slot->rmap_list;
1622 down_read(&mm->mmap_sem);
1623 if (ksm_test_exit(mm))
1626 vma = find_vma(mm, ksm_scan.address);
1628 for (; vma; vma = vma->vm_next) {
1629 if (!(vma->vm_flags & VM_MERGEABLE))
1631 if (ksm_scan.address < vma->vm_start)
1632 ksm_scan.address = vma->vm_start;
1634 ksm_scan.address = vma->vm_end;
1636 while (ksm_scan.address < vma->vm_end) {
1637 if (ksm_test_exit(mm))
1639 *page = follow_page(vma, ksm_scan.address, FOLL_GET);
1640 if (IS_ERR_OR_NULL(*page)) {
1641 ksm_scan.address += PAGE_SIZE;
1645 if (PageAnon(*page)) {
1646 flush_anon_page(vma, *page, ksm_scan.address);
1647 flush_dcache_page(*page);
1648 rmap_item = get_next_rmap_item(slot,
1649 ksm_scan.rmap_list, ksm_scan.address);
1651 ksm_scan.rmap_list =
1652 &rmap_item->rmap_list;
1653 ksm_scan.address += PAGE_SIZE;
1656 up_read(&mm->mmap_sem);
1660 ksm_scan.address += PAGE_SIZE;
1665 if (ksm_test_exit(mm)) {
1666 ksm_scan.address = 0;
1667 ksm_scan.rmap_list = &slot->rmap_list;
1670 * Nuke all the rmap_items that are above this current rmap:
1671 * because there were no VM_MERGEABLE vmas with such addresses.
1673 remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
1675 spin_lock(&ksm_mmlist_lock);
1676 ksm_scan.mm_slot = list_entry(slot->mm_list.next,
1677 struct mm_slot, mm_list);
1678 if (ksm_scan.address == 0) {
1680 * We've completed a full scan of all vmas, holding mmap_sem
1681 * throughout, and found no VM_MERGEABLE: so do the same as
1682 * __ksm_exit does to remove this mm from all our lists now.
1683 * This applies either when cleaning up after __ksm_exit
1684 * (but beware: we can reach here even before __ksm_exit),
1685 * or when all VM_MERGEABLE areas have been unmapped (and
1686 * mmap_sem then protects against race with MADV_MERGEABLE).
1688 hash_del(&slot->link);
1689 list_del(&slot->mm_list);
1690 spin_unlock(&ksm_mmlist_lock);
1693 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1694 up_read(&mm->mmap_sem);
1697 up_read(&mm->mmap_sem);
1699 * up_read(&mm->mmap_sem) first because after
1700 * spin_unlock(&ksm_mmlist_lock) run, the "mm" may
1701 * already have been freed under us by __ksm_exit()
1702 * because the "mm_slot" is still hashed and
1703 * ksm_scan.mm_slot doesn't point to it anymore.
1705 spin_unlock(&ksm_mmlist_lock);
1708 /* Repeat until we've completed scanning the whole list */
1709 slot = ksm_scan.mm_slot;
1710 if (slot != &ksm_mm_head)
1718 * ksm_do_scan - the ksm scanner main worker function.
1719 * @scan_npages - number of pages we want to scan before we return.
1721 static void ksm_do_scan(unsigned int scan_npages)
1723 struct rmap_item *rmap_item;
1724 struct page *uninitialized_var(page);
1726 while (scan_npages-- && likely(!freezing(current))) {
1728 rmap_item = scan_get_next_rmap_item(&page);
1731 cmp_and_merge_page(page, rmap_item);
1736 static int ksmd_should_run(void)
1738 return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
1741 static int ksm_scan_thread(void *nothing)
1744 set_user_nice(current, 5);
1746 while (!kthread_should_stop()) {
1747 mutex_lock(&ksm_thread_mutex);
1748 wait_while_offlining();
1749 if (ksmd_should_run())
1750 ksm_do_scan(ksm_thread_pages_to_scan);
1751 mutex_unlock(&ksm_thread_mutex);
1755 if (ksmd_should_run()) {
1756 schedule_timeout_interruptible(
1757 msecs_to_jiffies(ksm_thread_sleep_millisecs));
1759 wait_event_freezable(ksm_thread_wait,
1760 ksmd_should_run() || kthread_should_stop());
1766 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1767 unsigned long end, int advice, unsigned long *vm_flags)
1769 struct mm_struct *mm = vma->vm_mm;
1773 case MADV_MERGEABLE:
1775 * Be somewhat over-protective for now!
1777 if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE |
1778 VM_PFNMAP | VM_IO | VM_DONTEXPAND |
1779 VM_HUGETLB | VM_MIXEDMAP))
1780 return 0; /* just ignore the advice */
1783 if (*vm_flags & VM_SAO)
1787 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
1788 err = __ksm_enter(mm);
1793 *vm_flags |= VM_MERGEABLE;
1796 case MADV_UNMERGEABLE:
1797 if (!(*vm_flags & VM_MERGEABLE))
1798 return 0; /* just ignore the advice */
1800 if (vma->anon_vma) {
1801 err = unmerge_ksm_pages(vma, start, end);
1806 *vm_flags &= ~VM_MERGEABLE;
1813 int __ksm_enter(struct mm_struct *mm)
1815 struct mm_slot *mm_slot;
1818 mm_slot = alloc_mm_slot();
1822 /* Check ksm_run too? Would need tighter locking */
1823 needs_wakeup = list_empty(&ksm_mm_head.mm_list);
1825 spin_lock(&ksm_mmlist_lock);
1826 insert_to_mm_slots_hash(mm, mm_slot);
1828 * When KSM_RUN_MERGE (or KSM_RUN_STOP),
1829 * insert just behind the scanning cursor, to let the area settle
1830 * down a little; when fork is followed by immediate exec, we don't
1831 * want ksmd to waste time setting up and tearing down an rmap_list.
1833 * But when KSM_RUN_UNMERGE, it's important to insert ahead of its
1834 * scanning cursor, otherwise KSM pages in newly forked mms will be
1835 * missed: then we might as well insert at the end of the list.
1837 if (ksm_run & KSM_RUN_UNMERGE)
1838 list_add_tail(&mm_slot->mm_list, &ksm_mm_head.mm_list);
1840 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
1841 spin_unlock(&ksm_mmlist_lock);
1843 set_bit(MMF_VM_MERGEABLE, &mm->flags);
1844 atomic_inc(&mm->mm_count);
1847 wake_up_interruptible(&ksm_thread_wait);
1852 void __ksm_exit(struct mm_struct *mm)
1854 struct mm_slot *mm_slot;
1855 int easy_to_free = 0;
1858 * This process is exiting: if it's straightforward (as is the
1859 * case when ksmd was never running), free mm_slot immediately.
1860 * But if it's at the cursor or has rmap_items linked to it, use
1861 * mmap_sem to synchronize with any break_cows before pagetables
1862 * are freed, and leave the mm_slot on the list for ksmd to free.
1863 * Beware: ksm may already have noticed it exiting and freed the slot.
1866 spin_lock(&ksm_mmlist_lock);
1867 mm_slot = get_mm_slot(mm);
1868 if (mm_slot && ksm_scan.mm_slot != mm_slot) {
1869 if (!mm_slot->rmap_list) {
1870 hash_del(&mm_slot->link);
1871 list_del(&mm_slot->mm_list);
1874 list_move(&mm_slot->mm_list,
1875 &ksm_scan.mm_slot->mm_list);
1878 spin_unlock(&ksm_mmlist_lock);
1881 free_mm_slot(mm_slot);
1882 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1884 } else if (mm_slot) {
1885 down_write(&mm->mmap_sem);
1886 up_write(&mm->mmap_sem);
1890 struct page *ksm_might_need_to_copy(struct page *page,
1891 struct vm_area_struct *vma, unsigned long address)
1893 struct anon_vma *anon_vma = page_anon_vma(page);
1894 struct page *new_page;
1896 if (PageKsm(page)) {
1897 if (page_stable_node(page) &&
1898 !(ksm_run & KSM_RUN_UNMERGE))
1899 return page; /* no need to copy it */
1900 } else if (!anon_vma) {
1901 return page; /* no need to copy it */
1902 } else if (anon_vma->root == vma->anon_vma->root &&
1903 page->index == linear_page_index(vma, address)) {
1904 return page; /* still no need to copy it */
1906 if (!PageUptodate(page))
1907 return page; /* let do_swap_page report the error */
1909 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1911 copy_user_highpage(new_page, page, address, vma);
1913 SetPageDirty(new_page);
1914 __SetPageUptodate(new_page);
1915 __SetPageLocked(new_page);
1921 int rmap_walk_ksm(struct page *page, struct rmap_walk_control *rwc)
1923 struct stable_node *stable_node;
1924 struct rmap_item *rmap_item;
1925 int ret = SWAP_AGAIN;
1926 int search_new_forks = 0;
1928 VM_BUG_ON_PAGE(!PageKsm(page), page);
1931 * Rely on the page lock to protect against concurrent modifications
1932 * to that page's node of the stable tree.
1934 VM_BUG_ON_PAGE(!PageLocked(page), page);
1936 stable_node = page_stable_node(page);
1940 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
1941 struct anon_vma *anon_vma = rmap_item->anon_vma;
1942 struct anon_vma_chain *vmac;
1943 struct vm_area_struct *vma;
1946 anon_vma_lock_read(anon_vma);
1947 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
1951 if (rmap_item->address < vma->vm_start ||
1952 rmap_item->address >= vma->vm_end)
1955 * Initially we examine only the vma which covers this
1956 * rmap_item; but later, if there is still work to do,
1957 * we examine covering vmas in other mms: in case they
1958 * were forked from the original since ksmd passed.
1960 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1963 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1966 ret = rwc->rmap_one(page, vma,
1967 rmap_item->address, rwc->arg);
1968 if (ret != SWAP_AGAIN) {
1969 anon_vma_unlock_read(anon_vma);
1972 if (rwc->done && rwc->done(page)) {
1973 anon_vma_unlock_read(anon_vma);
1977 anon_vma_unlock_read(anon_vma);
1979 if (!search_new_forks++)
1985 #ifdef CONFIG_MIGRATION
1986 void ksm_migrate_page(struct page *newpage, struct page *oldpage)
1988 struct stable_node *stable_node;
1990 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
1991 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
1992 VM_BUG_ON_PAGE(newpage->mapping != oldpage->mapping, newpage);
1994 stable_node = page_stable_node(newpage);
1996 VM_BUG_ON_PAGE(stable_node->kpfn != page_to_pfn(oldpage), oldpage);
1997 stable_node->kpfn = page_to_pfn(newpage);
1999 * newpage->mapping was set in advance; now we need smp_wmb()
2000 * to make sure that the new stable_node->kpfn is visible
2001 * to get_ksm_page() before it can see that oldpage->mapping
2002 * has gone stale (or that PageSwapCache has been cleared).
2005 set_page_stable_node(oldpage, NULL);
2008 #endif /* CONFIG_MIGRATION */
2010 #ifdef CONFIG_MEMORY_HOTREMOVE
2011 static void wait_while_offlining(void)
2013 while (ksm_run & KSM_RUN_OFFLINE) {
2014 mutex_unlock(&ksm_thread_mutex);
2015 wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE),
2016 TASK_UNINTERRUPTIBLE);
2017 mutex_lock(&ksm_thread_mutex);
2021 static void ksm_check_stable_tree(unsigned long start_pfn,
2022 unsigned long end_pfn)
2024 struct stable_node *stable_node, *next;
2025 struct rb_node *node;
2028 for (nid = 0; nid < ksm_nr_node_ids; nid++) {
2029 node = rb_first(root_stable_tree + nid);
2031 stable_node = rb_entry(node, struct stable_node, node);
2032 if (stable_node->kpfn >= start_pfn &&
2033 stable_node->kpfn < end_pfn) {
2035 * Don't get_ksm_page, page has already gone:
2036 * which is why we keep kpfn instead of page*
2038 remove_node_from_stable_tree(stable_node);
2039 node = rb_first(root_stable_tree + nid);
2041 node = rb_next(node);
2045 list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
2046 if (stable_node->kpfn >= start_pfn &&
2047 stable_node->kpfn < end_pfn)
2048 remove_node_from_stable_tree(stable_node);
2053 static int ksm_memory_callback(struct notifier_block *self,
2054 unsigned long action, void *arg)
2056 struct memory_notify *mn = arg;
2059 case MEM_GOING_OFFLINE:
2061 * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items()
2062 * and remove_all_stable_nodes() while memory is going offline:
2063 * it is unsafe for them to touch the stable tree at this time.
2064 * But unmerge_ksm_pages(), rmap lookups and other entry points
2065 * which do not need the ksm_thread_mutex are all safe.
2067 mutex_lock(&ksm_thread_mutex);
2068 ksm_run |= KSM_RUN_OFFLINE;
2069 mutex_unlock(&ksm_thread_mutex);
2074 * Most of the work is done by page migration; but there might
2075 * be a few stable_nodes left over, still pointing to struct
2076 * pages which have been offlined: prune those from the tree,
2077 * otherwise get_ksm_page() might later try to access a
2078 * non-existent struct page.
2080 ksm_check_stable_tree(mn->start_pfn,
2081 mn->start_pfn + mn->nr_pages);
2084 case MEM_CANCEL_OFFLINE:
2085 mutex_lock(&ksm_thread_mutex);
2086 ksm_run &= ~KSM_RUN_OFFLINE;
2087 mutex_unlock(&ksm_thread_mutex);
2089 smp_mb(); /* wake_up_bit advises this */
2090 wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE));
2096 static void wait_while_offlining(void)
2099 #endif /* CONFIG_MEMORY_HOTREMOVE */
2103 * This all compiles without CONFIG_SYSFS, but is a waste of space.
2106 #define KSM_ATTR_RO(_name) \
2107 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2108 #define KSM_ATTR(_name) \
2109 static struct kobj_attribute _name##_attr = \
2110 __ATTR(_name, 0644, _name##_show, _name##_store)
2112 static ssize_t sleep_millisecs_show(struct kobject *kobj,
2113 struct kobj_attribute *attr, char *buf)
2115 return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
2118 static ssize_t sleep_millisecs_store(struct kobject *kobj,
2119 struct kobj_attribute *attr,
2120 const char *buf, size_t count)
2122 unsigned long msecs;
2125 err = kstrtoul(buf, 10, &msecs);
2126 if (err || msecs > UINT_MAX)
2129 ksm_thread_sleep_millisecs = msecs;
2133 KSM_ATTR(sleep_millisecs);
2135 static ssize_t pages_to_scan_show(struct kobject *kobj,
2136 struct kobj_attribute *attr, char *buf)
2138 return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
2141 static ssize_t pages_to_scan_store(struct kobject *kobj,
2142 struct kobj_attribute *attr,
2143 const char *buf, size_t count)
2146 unsigned long nr_pages;
2148 err = kstrtoul(buf, 10, &nr_pages);
2149 if (err || nr_pages > UINT_MAX)
2152 ksm_thread_pages_to_scan = nr_pages;
2156 KSM_ATTR(pages_to_scan);
2158 static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
2161 return sprintf(buf, "%lu\n", ksm_run);
2164 static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
2165 const char *buf, size_t count)
2168 unsigned long flags;
2170 err = kstrtoul(buf, 10, &flags);
2171 if (err || flags > UINT_MAX)
2173 if (flags > KSM_RUN_UNMERGE)
2177 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
2178 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
2179 * breaking COW to free the pages_shared (but leaves mm_slots
2180 * on the list for when ksmd may be set running again).
2183 mutex_lock(&ksm_thread_mutex);
2184 wait_while_offlining();
2185 if (ksm_run != flags) {
2187 if (flags & KSM_RUN_UNMERGE) {
2188 set_current_oom_origin();
2189 err = unmerge_and_remove_all_rmap_items();
2190 clear_current_oom_origin();
2192 ksm_run = KSM_RUN_STOP;
2197 mutex_unlock(&ksm_thread_mutex);
2199 if (flags & KSM_RUN_MERGE)
2200 wake_up_interruptible(&ksm_thread_wait);
2207 static ssize_t merge_across_nodes_show(struct kobject *kobj,
2208 struct kobj_attribute *attr, char *buf)
2210 return sprintf(buf, "%u\n", ksm_merge_across_nodes);
2213 static ssize_t merge_across_nodes_store(struct kobject *kobj,
2214 struct kobj_attribute *attr,
2215 const char *buf, size_t count)
2220 err = kstrtoul(buf, 10, &knob);
2226 mutex_lock(&ksm_thread_mutex);
2227 wait_while_offlining();
2228 if (ksm_merge_across_nodes != knob) {
2229 if (ksm_pages_shared || remove_all_stable_nodes())
2231 else if (root_stable_tree == one_stable_tree) {
2232 struct rb_root *buf;
2234 * This is the first time that we switch away from the
2235 * default of merging across nodes: must now allocate
2236 * a buffer to hold as many roots as may be needed.
2237 * Allocate stable and unstable together:
2238 * MAXSMP NODES_SHIFT 10 will use 16kB.
2240 buf = kcalloc(nr_node_ids + nr_node_ids, sizeof(*buf),
2242 /* Let us assume that RB_ROOT is NULL is zero */
2246 root_stable_tree = buf;
2247 root_unstable_tree = buf + nr_node_ids;
2248 /* Stable tree is empty but not the unstable */
2249 root_unstable_tree[0] = one_unstable_tree[0];
2253 ksm_merge_across_nodes = knob;
2254 ksm_nr_node_ids = knob ? 1 : nr_node_ids;
2257 mutex_unlock(&ksm_thread_mutex);
2259 return err ? err : count;
2261 KSM_ATTR(merge_across_nodes);
2264 static ssize_t pages_shared_show(struct kobject *kobj,
2265 struct kobj_attribute *attr, char *buf)
2267 return sprintf(buf, "%lu\n", ksm_pages_shared);
2269 KSM_ATTR_RO(pages_shared);
2271 static ssize_t pages_sharing_show(struct kobject *kobj,
2272 struct kobj_attribute *attr, char *buf)
2274 return sprintf(buf, "%lu\n", ksm_pages_sharing);
2276 KSM_ATTR_RO(pages_sharing);
2278 static ssize_t pages_unshared_show(struct kobject *kobj,
2279 struct kobj_attribute *attr, char *buf)
2281 return sprintf(buf, "%lu\n", ksm_pages_unshared);
2283 KSM_ATTR_RO(pages_unshared);
2285 static ssize_t pages_volatile_show(struct kobject *kobj,
2286 struct kobj_attribute *attr, char *buf)
2288 long ksm_pages_volatile;
2290 ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
2291 - ksm_pages_sharing - ksm_pages_unshared;
2293 * It was not worth any locking to calculate that statistic,
2294 * but it might therefore sometimes be negative: conceal that.
2296 if (ksm_pages_volatile < 0)
2297 ksm_pages_volatile = 0;
2298 return sprintf(buf, "%ld\n", ksm_pages_volatile);
2300 KSM_ATTR_RO(pages_volatile);
2302 static ssize_t full_scans_show(struct kobject *kobj,
2303 struct kobj_attribute *attr, char *buf)
2305 return sprintf(buf, "%lu\n", ksm_scan.seqnr);
2307 KSM_ATTR_RO(full_scans);
2309 static struct attribute *ksm_attrs[] = {
2310 &sleep_millisecs_attr.attr,
2311 &pages_to_scan_attr.attr,
2313 &pages_shared_attr.attr,
2314 &pages_sharing_attr.attr,
2315 &pages_unshared_attr.attr,
2316 &pages_volatile_attr.attr,
2317 &full_scans_attr.attr,
2319 &merge_across_nodes_attr.attr,
2324 static struct attribute_group ksm_attr_group = {
2328 #endif /* CONFIG_SYSFS */
2330 static int __init ksm_init(void)
2332 struct task_struct *ksm_thread;
2335 err = ksm_slab_init();
2339 ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
2340 if (IS_ERR(ksm_thread)) {
2341 pr_err("ksm: creating kthread failed\n");
2342 err = PTR_ERR(ksm_thread);
2347 err = sysfs_create_group(mm_kobj, &ksm_attr_group);
2349 pr_err("ksm: register sysfs failed\n");
2350 kthread_stop(ksm_thread);
2354 ksm_run = KSM_RUN_MERGE; /* no way for user to start it */
2356 #endif /* CONFIG_SYSFS */
2358 #ifdef CONFIG_MEMORY_HOTREMOVE
2359 /* There is no significance to this priority 100 */
2360 hotplug_memory_notifier(ksm_memory_callback, 100);
2369 subsys_initcall(ksm_init);