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.)
93 * struct mm_slot - ksm information per mm that is being scanned
94 * @link: link to the mm_slots hash list
95 * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
96 * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
97 * @mm: the mm that this information is valid for
100 struct hlist_node link
;
101 struct list_head mm_list
;
102 struct rmap_item
*rmap_list
;
103 struct mm_struct
*mm
;
107 * struct ksm_scan - cursor for scanning
108 * @mm_slot: the current mm_slot we are scanning
109 * @address: the next address inside that to be scanned
110 * @rmap_list: link to the next rmap to be scanned in the rmap_list
111 * @seqnr: count of completed full scans (needed when removing unstable node)
113 * There is only the one ksm_scan instance of this cursor structure.
116 struct mm_slot
*mm_slot
;
117 unsigned long address
;
118 struct rmap_item
**rmap_list
;
123 * struct stable_node - node of the stable rbtree
124 * @node: rb node of this ksm page in the stable tree
125 * @hlist: hlist head of rmap_items using this ksm page
126 * @kpfn: page frame number of this ksm page
130 struct hlist_head hlist
;
135 * struct rmap_item - reverse mapping item for virtual addresses
136 * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
137 * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
138 * @mm: the memory structure this rmap_item is pointing into
139 * @address: the virtual address this rmap_item tracks (+ flags in low bits)
140 * @oldchecksum: previous checksum of the page at that virtual address
141 * @nid: NUMA node id of unstable tree in which linked (may not match page)
142 * @node: rb node of this rmap_item in the unstable tree
143 * @head: pointer to stable_node heading this list in the stable tree
144 * @hlist: link into hlist of rmap_items hanging off that stable_node
147 struct rmap_item
*rmap_list
;
148 struct anon_vma
*anon_vma
; /* when stable */
149 struct mm_struct
*mm
;
150 unsigned long address
; /* + low bits used for flags below */
151 unsigned int oldchecksum
; /* when unstable */
156 struct rb_node node
; /* when node of unstable tree */
157 struct { /* when listed from stable tree */
158 struct stable_node
*head
;
159 struct hlist_node hlist
;
164 #define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */
165 #define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */
166 #define STABLE_FLAG 0x200 /* is listed from the stable tree */
168 /* The stable and unstable tree heads */
169 static struct rb_root root_unstable_tree
[MAX_NUMNODES
];
170 static struct rb_root root_stable_tree
[MAX_NUMNODES
];
172 #define MM_SLOTS_HASH_BITS 10
173 static DEFINE_HASHTABLE(mm_slots_hash
, MM_SLOTS_HASH_BITS
);
175 static struct mm_slot ksm_mm_head
= {
176 .mm_list
= LIST_HEAD_INIT(ksm_mm_head
.mm_list
),
178 static struct ksm_scan ksm_scan
= {
179 .mm_slot
= &ksm_mm_head
,
182 static struct kmem_cache
*rmap_item_cache
;
183 static struct kmem_cache
*stable_node_cache
;
184 static struct kmem_cache
*mm_slot_cache
;
186 /* The number of nodes in the stable tree */
187 static unsigned long ksm_pages_shared
;
189 /* The number of page slots additionally sharing those nodes */
190 static unsigned long ksm_pages_sharing
;
192 /* The number of nodes in the unstable tree */
193 static unsigned long ksm_pages_unshared
;
195 /* The number of rmap_items in use: to calculate pages_volatile */
196 static unsigned long ksm_rmap_items
;
198 /* Number of pages ksmd should scan in one batch */
199 static unsigned int ksm_thread_pages_to_scan
= 100;
201 /* Milliseconds ksmd should sleep between batches */
202 static unsigned int ksm_thread_sleep_millisecs
= 20;
205 /* Zeroed when merging across nodes is not allowed */
206 static unsigned int ksm_merge_across_nodes
= 1;
208 #define ksm_merge_across_nodes 1U
211 #define KSM_RUN_STOP 0
212 #define KSM_RUN_MERGE 1
213 #define KSM_RUN_UNMERGE 2
214 static unsigned int ksm_run
= KSM_RUN_STOP
;
216 static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait
);
217 static DEFINE_MUTEX(ksm_thread_mutex
);
218 static DEFINE_SPINLOCK(ksm_mmlist_lock
);
220 #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
221 sizeof(struct __struct), __alignof__(struct __struct),\
224 static int __init
ksm_slab_init(void)
226 rmap_item_cache
= KSM_KMEM_CACHE(rmap_item
, 0);
227 if (!rmap_item_cache
)
230 stable_node_cache
= KSM_KMEM_CACHE(stable_node
, 0);
231 if (!stable_node_cache
)
234 mm_slot_cache
= KSM_KMEM_CACHE(mm_slot
, 0);
241 kmem_cache_destroy(stable_node_cache
);
243 kmem_cache_destroy(rmap_item_cache
);
248 static void __init
ksm_slab_free(void)
250 kmem_cache_destroy(mm_slot_cache
);
251 kmem_cache_destroy(stable_node_cache
);
252 kmem_cache_destroy(rmap_item_cache
);
253 mm_slot_cache
= NULL
;
256 static inline struct rmap_item
*alloc_rmap_item(void)
258 struct rmap_item
*rmap_item
;
260 rmap_item
= kmem_cache_zalloc(rmap_item_cache
, GFP_KERNEL
);
266 static inline void free_rmap_item(struct rmap_item
*rmap_item
)
269 rmap_item
->mm
= NULL
; /* debug safety */
270 kmem_cache_free(rmap_item_cache
, rmap_item
);
273 static inline struct stable_node
*alloc_stable_node(void)
275 return kmem_cache_alloc(stable_node_cache
, GFP_KERNEL
);
278 static inline void free_stable_node(struct stable_node
*stable_node
)
280 kmem_cache_free(stable_node_cache
, stable_node
);
283 static inline struct mm_slot
*alloc_mm_slot(void)
285 if (!mm_slot_cache
) /* initialization failed */
287 return kmem_cache_zalloc(mm_slot_cache
, GFP_KERNEL
);
290 static inline void free_mm_slot(struct mm_slot
*mm_slot
)
292 kmem_cache_free(mm_slot_cache
, mm_slot
);
295 static struct mm_slot
*get_mm_slot(struct mm_struct
*mm
)
297 struct hlist_node
*node
;
298 struct mm_slot
*slot
;
300 hash_for_each_possible(mm_slots_hash
, slot
, node
, link
, (unsigned long)mm
)
307 static void insert_to_mm_slots_hash(struct mm_struct
*mm
,
308 struct mm_slot
*mm_slot
)
311 hash_add(mm_slots_hash
, &mm_slot
->link
, (unsigned long)mm
);
314 static inline int in_stable_tree(struct rmap_item
*rmap_item
)
316 return rmap_item
->address
& STABLE_FLAG
;
320 * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
321 * page tables after it has passed through ksm_exit() - which, if necessary,
322 * takes mmap_sem briefly to serialize against them. ksm_exit() does not set
323 * a special flag: they can just back out as soon as mm_users goes to zero.
324 * ksm_test_exit() is used throughout to make this test for exit: in some
325 * places for correctness, in some places just to avoid unnecessary work.
327 static inline bool ksm_test_exit(struct mm_struct
*mm
)
329 return atomic_read(&mm
->mm_users
) == 0;
333 * We use break_ksm to break COW on a ksm page: it's a stripped down
335 * if (get_user_pages(current, mm, addr, 1, 1, 1, &page, NULL) == 1)
338 * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
339 * in case the application has unmapped and remapped mm,addr meanwhile.
340 * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP
341 * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
343 static int break_ksm(struct vm_area_struct
*vma
, unsigned long addr
)
350 page
= follow_page(vma
, addr
, FOLL_GET
);
351 if (IS_ERR_OR_NULL(page
))
354 ret
= handle_mm_fault(vma
->vm_mm
, vma
, addr
,
357 ret
= VM_FAULT_WRITE
;
359 } while (!(ret
& (VM_FAULT_WRITE
| VM_FAULT_SIGBUS
| VM_FAULT_OOM
)));
361 * We must loop because handle_mm_fault() may back out if there's
362 * any difficulty e.g. if pte accessed bit gets updated concurrently.
364 * VM_FAULT_WRITE is what we have been hoping for: it indicates that
365 * COW has been broken, even if the vma does not permit VM_WRITE;
366 * but note that a concurrent fault might break PageKsm for us.
368 * VM_FAULT_SIGBUS could occur if we race with truncation of the
369 * backing file, which also invalidates anonymous pages: that's
370 * okay, that truncation will have unmapped the PageKsm for us.
372 * VM_FAULT_OOM: at the time of writing (late July 2009), setting
373 * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
374 * current task has TIF_MEMDIE set, and will be OOM killed on return
375 * to user; and ksmd, having no mm, would never be chosen for that.
377 * But if the mm is in a limited mem_cgroup, then the fault may fail
378 * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
379 * even ksmd can fail in this way - though it's usually breaking ksm
380 * just to undo a merge it made a moment before, so unlikely to oom.
382 * That's a pity: we might therefore have more kernel pages allocated
383 * than we're counting as nodes in the stable tree; but ksm_do_scan
384 * will retry to break_cow on each pass, so should recover the page
385 * in due course. The important thing is to not let VM_MERGEABLE
386 * be cleared while any such pages might remain in the area.
388 return (ret
& VM_FAULT_OOM
) ? -ENOMEM
: 0;
391 static struct vm_area_struct
*find_mergeable_vma(struct mm_struct
*mm
,
394 struct vm_area_struct
*vma
;
395 if (ksm_test_exit(mm
))
397 vma
= find_vma(mm
, addr
);
398 if (!vma
|| vma
->vm_start
> addr
)
400 if (!(vma
->vm_flags
& VM_MERGEABLE
) || !vma
->anon_vma
)
405 static void break_cow(struct rmap_item
*rmap_item
)
407 struct mm_struct
*mm
= rmap_item
->mm
;
408 unsigned long addr
= rmap_item
->address
;
409 struct vm_area_struct
*vma
;
412 * It is not an accident that whenever we want to break COW
413 * to undo, we also need to drop a reference to the anon_vma.
415 put_anon_vma(rmap_item
->anon_vma
);
417 down_read(&mm
->mmap_sem
);
418 vma
= find_mergeable_vma(mm
, addr
);
420 break_ksm(vma
, addr
);
421 up_read(&mm
->mmap_sem
);
424 static struct page
*page_trans_compound_anon(struct page
*page
)
426 if (PageTransCompound(page
)) {
427 struct page
*head
= compound_trans_head(page
);
429 * head may actually be splitted and freed from under
430 * us but it's ok here.
438 static struct page
*get_mergeable_page(struct rmap_item
*rmap_item
)
440 struct mm_struct
*mm
= rmap_item
->mm
;
441 unsigned long addr
= rmap_item
->address
;
442 struct vm_area_struct
*vma
;
445 down_read(&mm
->mmap_sem
);
446 vma
= find_mergeable_vma(mm
, addr
);
450 page
= follow_page(vma
, addr
, FOLL_GET
);
451 if (IS_ERR_OR_NULL(page
))
453 if (PageAnon(page
) || page_trans_compound_anon(page
)) {
454 flush_anon_page(vma
, page
, addr
);
455 flush_dcache_page(page
);
460 up_read(&mm
->mmap_sem
);
465 * This helper is used for getting right index into array of tree roots.
466 * When merge_across_nodes knob is set to 1, there are only two rb-trees for
467 * stable and unstable pages from all nodes with roots in index 0. Otherwise,
468 * every node has its own stable and unstable tree.
470 static inline int get_kpfn_nid(unsigned long kpfn
)
472 return ksm_merge_across_nodes
? 0 : pfn_to_nid(kpfn
);
475 static void remove_node_from_stable_tree(struct stable_node
*stable_node
)
477 struct rmap_item
*rmap_item
;
478 struct hlist_node
*hlist
;
481 hlist_for_each_entry(rmap_item
, hlist
, &stable_node
->hlist
, hlist
) {
482 if (rmap_item
->hlist
.next
)
486 put_anon_vma(rmap_item
->anon_vma
);
487 rmap_item
->address
&= PAGE_MASK
;
491 nid
= get_kpfn_nid(stable_node
->kpfn
);
492 rb_erase(&stable_node
->node
, &root_stable_tree
[nid
]);
493 free_stable_node(stable_node
);
497 * get_ksm_page: checks if the page indicated by the stable node
498 * is still its ksm page, despite having held no reference to it.
499 * In which case we can trust the content of the page, and it
500 * returns the gotten page; but if the page has now been zapped,
501 * remove the stale node from the stable tree and return NULL.
503 * You would expect the stable_node to hold a reference to the ksm page.
504 * But if it increments the page's count, swapping out has to wait for
505 * ksmd to come around again before it can free the page, which may take
506 * seconds or even minutes: much too unresponsive. So instead we use a
507 * "keyhole reference": access to the ksm page from the stable node peeps
508 * out through its keyhole to see if that page still holds the right key,
509 * pointing back to this stable node. This relies on freeing a PageAnon
510 * page to reset its page->mapping to NULL, and relies on no other use of
511 * a page to put something that might look like our key in page->mapping.
513 * include/linux/pagemap.h page_cache_get_speculative() is a good reference,
514 * but this is different - made simpler by ksm_thread_mutex being held, but
515 * interesting for assuming that no other use of the struct page could ever
516 * put our expected_mapping into page->mapping (or a field of the union which
517 * coincides with page->mapping).
519 * Note: it is possible that get_ksm_page() will return NULL one moment,
520 * then page the next, if the page is in between page_freeze_refs() and
521 * page_unfreeze_refs(): this shouldn't be a problem anywhere, the page
522 * is on its way to being freed; but it is an anomaly to bear in mind.
524 static struct page
*get_ksm_page(struct stable_node
*stable_node
, bool locked
)
527 void *expected_mapping
;
529 page
= pfn_to_page(stable_node
->kpfn
);
530 expected_mapping
= (void *)stable_node
+
531 (PAGE_MAPPING_ANON
| PAGE_MAPPING_KSM
);
532 if (page
->mapping
!= expected_mapping
)
534 if (!get_page_unless_zero(page
))
536 if (page
->mapping
!= expected_mapping
) {
542 if (page
->mapping
!= expected_mapping
) {
550 remove_node_from_stable_tree(stable_node
);
555 * Removing rmap_item from stable or unstable tree.
556 * This function will clean the information from the stable/unstable tree.
558 static void remove_rmap_item_from_tree(struct rmap_item
*rmap_item
)
560 if (rmap_item
->address
& STABLE_FLAG
) {
561 struct stable_node
*stable_node
;
564 stable_node
= rmap_item
->head
;
565 page
= get_ksm_page(stable_node
, true);
569 hlist_del(&rmap_item
->hlist
);
573 if (stable_node
->hlist
.first
)
578 put_anon_vma(rmap_item
->anon_vma
);
579 rmap_item
->address
&= PAGE_MASK
;
581 } else if (rmap_item
->address
& UNSTABLE_FLAG
) {
584 * Usually ksmd can and must skip the rb_erase, because
585 * root_unstable_tree was already reset to RB_ROOT.
586 * But be careful when an mm is exiting: do the rb_erase
587 * if this rmap_item was inserted by this scan, rather
588 * than left over from before.
590 age
= (unsigned char)(ksm_scan
.seqnr
- rmap_item
->address
);
593 rb_erase(&rmap_item
->node
,
594 &root_unstable_tree
[NUMA(rmap_item
->nid
)]);
595 ksm_pages_unshared
--;
596 rmap_item
->address
&= PAGE_MASK
;
599 cond_resched(); /* we're called from many long loops */
602 static void remove_trailing_rmap_items(struct mm_slot
*mm_slot
,
603 struct rmap_item
**rmap_list
)
606 struct rmap_item
*rmap_item
= *rmap_list
;
607 *rmap_list
= rmap_item
->rmap_list
;
608 remove_rmap_item_from_tree(rmap_item
);
609 free_rmap_item(rmap_item
);
614 * Though it's very tempting to unmerge rmap_items from stable tree rather
615 * than check every pte of a given vma, the locking doesn't quite work for
616 * that - an rmap_item is assigned to the stable tree after inserting ksm
617 * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing
618 * rmap_items from parent to child at fork time (so as not to waste time
619 * if exit comes before the next scan reaches it).
621 * Similarly, although we'd like to remove rmap_items (so updating counts
622 * and freeing memory) when unmerging an area, it's easier to leave that
623 * to the next pass of ksmd - consider, for example, how ksmd might be
624 * in cmp_and_merge_page on one of the rmap_items we would be removing.
626 static int unmerge_ksm_pages(struct vm_area_struct
*vma
,
627 unsigned long start
, unsigned long end
)
632 for (addr
= start
; addr
< end
&& !err
; addr
+= PAGE_SIZE
) {
633 if (ksm_test_exit(vma
->vm_mm
))
635 if (signal_pending(current
))
638 err
= break_ksm(vma
, addr
);
645 * Only called through the sysfs control interface:
647 static int unmerge_and_remove_all_rmap_items(void)
649 struct mm_slot
*mm_slot
;
650 struct mm_struct
*mm
;
651 struct vm_area_struct
*vma
;
654 spin_lock(&ksm_mmlist_lock
);
655 ksm_scan
.mm_slot
= list_entry(ksm_mm_head
.mm_list
.next
,
656 struct mm_slot
, mm_list
);
657 spin_unlock(&ksm_mmlist_lock
);
659 for (mm_slot
= ksm_scan
.mm_slot
;
660 mm_slot
!= &ksm_mm_head
; mm_slot
= ksm_scan
.mm_slot
) {
662 down_read(&mm
->mmap_sem
);
663 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
664 if (ksm_test_exit(mm
))
666 if (!(vma
->vm_flags
& VM_MERGEABLE
) || !vma
->anon_vma
)
668 err
= unmerge_ksm_pages(vma
,
669 vma
->vm_start
, vma
->vm_end
);
674 remove_trailing_rmap_items(mm_slot
, &mm_slot
->rmap_list
);
676 spin_lock(&ksm_mmlist_lock
);
677 ksm_scan
.mm_slot
= list_entry(mm_slot
->mm_list
.next
,
678 struct mm_slot
, mm_list
);
679 if (ksm_test_exit(mm
)) {
680 hash_del(&mm_slot
->link
);
681 list_del(&mm_slot
->mm_list
);
682 spin_unlock(&ksm_mmlist_lock
);
684 free_mm_slot(mm_slot
);
685 clear_bit(MMF_VM_MERGEABLE
, &mm
->flags
);
686 up_read(&mm
->mmap_sem
);
689 spin_unlock(&ksm_mmlist_lock
);
690 up_read(&mm
->mmap_sem
);
698 up_read(&mm
->mmap_sem
);
699 spin_lock(&ksm_mmlist_lock
);
700 ksm_scan
.mm_slot
= &ksm_mm_head
;
701 spin_unlock(&ksm_mmlist_lock
);
704 #endif /* CONFIG_SYSFS */
706 static u32
calc_checksum(struct page
*page
)
709 void *addr
= kmap_atomic(page
);
710 checksum
= jhash2(addr
, PAGE_SIZE
/ 4, 17);
715 static int memcmp_pages(struct page
*page1
, struct page
*page2
)
720 addr1
= kmap_atomic(page1
);
721 addr2
= kmap_atomic(page2
);
722 ret
= memcmp(addr1
, addr2
, PAGE_SIZE
);
723 kunmap_atomic(addr2
);
724 kunmap_atomic(addr1
);
728 static inline int pages_identical(struct page
*page1
, struct page
*page2
)
730 return !memcmp_pages(page1
, page2
);
733 static int write_protect_page(struct vm_area_struct
*vma
, struct page
*page
,
736 struct mm_struct
*mm
= vma
->vm_mm
;
742 unsigned long mmun_start
; /* For mmu_notifiers */
743 unsigned long mmun_end
; /* For mmu_notifiers */
745 addr
= page_address_in_vma(page
, vma
);
749 BUG_ON(PageTransCompound(page
));
752 mmun_end
= addr
+ PAGE_SIZE
;
753 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
755 ptep
= page_check_address(page
, mm
, addr
, &ptl
, 0);
759 if (pte_write(*ptep
) || pte_dirty(*ptep
)) {
762 swapped
= PageSwapCache(page
);
763 flush_cache_page(vma
, addr
, page_to_pfn(page
));
765 * Ok this is tricky, when get_user_pages_fast() run it doesn't
766 * take any lock, therefore the check that we are going to make
767 * with the pagecount against the mapcount is racey and
768 * O_DIRECT can happen right after the check.
769 * So we clear the pte and flush the tlb before the check
770 * this assure us that no O_DIRECT can happen after the check
771 * or in the middle of the check.
773 entry
= ptep_clear_flush(vma
, addr
, ptep
);
775 * Check that no O_DIRECT or similar I/O is in progress on the
778 if (page_mapcount(page
) + 1 + swapped
!= page_count(page
)) {
779 set_pte_at(mm
, addr
, ptep
, entry
);
782 if (pte_dirty(entry
))
783 set_page_dirty(page
);
784 entry
= pte_mkclean(pte_wrprotect(entry
));
785 set_pte_at_notify(mm
, addr
, ptep
, entry
);
791 pte_unmap_unlock(ptep
, ptl
);
793 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
799 * replace_page - replace page in vma by new ksm page
800 * @vma: vma that holds the pte pointing to page
801 * @page: the page we are replacing by kpage
802 * @kpage: the ksm page we replace page by
803 * @orig_pte: the original value of the pte
805 * Returns 0 on success, -EFAULT on failure.
807 static int replace_page(struct vm_area_struct
*vma
, struct page
*page
,
808 struct page
*kpage
, pte_t orig_pte
)
810 struct mm_struct
*mm
= vma
->vm_mm
;
816 unsigned long mmun_start
; /* For mmu_notifiers */
817 unsigned long mmun_end
; /* For mmu_notifiers */
819 addr
= page_address_in_vma(page
, vma
);
823 pmd
= mm_find_pmd(mm
, addr
);
826 BUG_ON(pmd_trans_huge(*pmd
));
829 mmun_end
= addr
+ PAGE_SIZE
;
830 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
832 ptep
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
833 if (!pte_same(*ptep
, orig_pte
)) {
834 pte_unmap_unlock(ptep
, ptl
);
839 page_add_anon_rmap(kpage
, vma
, addr
);
841 flush_cache_page(vma
, addr
, pte_pfn(*ptep
));
842 ptep_clear_flush(vma
, addr
, ptep
);
843 set_pte_at_notify(mm
, addr
, ptep
, mk_pte(kpage
, vma
->vm_page_prot
));
845 page_remove_rmap(page
);
846 if (!page_mapped(page
))
847 try_to_free_swap(page
);
850 pte_unmap_unlock(ptep
, ptl
);
853 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
858 static int page_trans_compound_anon_split(struct page
*page
)
861 struct page
*transhuge_head
= page_trans_compound_anon(page
);
862 if (transhuge_head
) {
863 /* Get the reference on the head to split it. */
864 if (get_page_unless_zero(transhuge_head
)) {
866 * Recheck we got the reference while the head
867 * was still anonymous.
869 if (PageAnon(transhuge_head
))
870 ret
= split_huge_page(transhuge_head
);
873 * Retry later if split_huge_page run
877 put_page(transhuge_head
);
879 /* Retry later if split_huge_page run from under us. */
886 * try_to_merge_one_page - take two pages and merge them into one
887 * @vma: the vma that holds the pte pointing to page
888 * @page: the PageAnon page that we want to replace with kpage
889 * @kpage: the PageKsm page that we want to map instead of page,
890 * or NULL the first time when we want to use page as kpage.
892 * This function returns 0 if the pages were merged, -EFAULT otherwise.
894 static int try_to_merge_one_page(struct vm_area_struct
*vma
,
895 struct page
*page
, struct page
*kpage
)
897 pte_t orig_pte
= __pte(0);
900 if (page
== kpage
) /* ksm page forked */
903 if (!(vma
->vm_flags
& VM_MERGEABLE
))
905 if (PageTransCompound(page
) && page_trans_compound_anon_split(page
))
907 BUG_ON(PageTransCompound(page
));
912 * We need the page lock to read a stable PageSwapCache in
913 * write_protect_page(). We use trylock_page() instead of
914 * lock_page() because we don't want to wait here - we
915 * prefer to continue scanning and merging different pages,
916 * then come back to this page when it is unlocked.
918 if (!trylock_page(page
))
921 * If this anonymous page is mapped only here, its pte may need
922 * to be write-protected. If it's mapped elsewhere, all of its
923 * ptes are necessarily already write-protected. But in either
924 * case, we need to lock and check page_count is not raised.
926 if (write_protect_page(vma
, page
, &orig_pte
) == 0) {
929 * While we hold page lock, upgrade page from
930 * PageAnon+anon_vma to PageKsm+NULL stable_node:
931 * stable_tree_insert() will update stable_node.
933 set_page_stable_node(page
, NULL
);
934 mark_page_accessed(page
);
936 } else if (pages_identical(page
, kpage
))
937 err
= replace_page(vma
, page
, kpage
, orig_pte
);
940 if ((vma
->vm_flags
& VM_LOCKED
) && kpage
&& !err
) {
941 munlock_vma_page(page
);
942 if (!PageMlocked(kpage
)) {
945 mlock_vma_page(kpage
);
946 page
= kpage
; /* for final unlock */
956 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
957 * but no new kernel page is allocated: kpage must already be a ksm page.
959 * This function returns 0 if the pages were merged, -EFAULT otherwise.
961 static int try_to_merge_with_ksm_page(struct rmap_item
*rmap_item
,
962 struct page
*page
, struct page
*kpage
)
964 struct mm_struct
*mm
= rmap_item
->mm
;
965 struct vm_area_struct
*vma
;
968 down_read(&mm
->mmap_sem
);
969 if (ksm_test_exit(mm
))
971 vma
= find_vma(mm
, rmap_item
->address
);
972 if (!vma
|| vma
->vm_start
> rmap_item
->address
)
975 err
= try_to_merge_one_page(vma
, page
, kpage
);
979 /* Must get reference to anon_vma while still holding mmap_sem */
980 rmap_item
->anon_vma
= vma
->anon_vma
;
981 get_anon_vma(vma
->anon_vma
);
983 up_read(&mm
->mmap_sem
);
988 * try_to_merge_two_pages - take two identical pages and prepare them
989 * to be merged into one page.
991 * This function returns the kpage if we successfully merged two identical
992 * pages into one ksm page, NULL otherwise.
994 * Note that this function upgrades page to ksm page: if one of the pages
995 * is already a ksm page, try_to_merge_with_ksm_page should be used.
997 static struct page
*try_to_merge_two_pages(struct rmap_item
*rmap_item
,
999 struct rmap_item
*tree_rmap_item
,
1000 struct page
*tree_page
)
1004 err
= try_to_merge_with_ksm_page(rmap_item
, page
, NULL
);
1006 err
= try_to_merge_with_ksm_page(tree_rmap_item
,
1009 * If that fails, we have a ksm page with only one pte
1010 * pointing to it: so break it.
1013 break_cow(rmap_item
);
1015 return err
? NULL
: page
;
1019 * stable_tree_search - search for page inside the stable tree
1021 * This function checks if there is a page inside the stable tree
1022 * with identical content to the page that we are scanning right now.
1024 * This function returns the stable tree node of identical content if found,
1027 static struct page
*stable_tree_search(struct page
*page
)
1029 struct rb_node
*node
;
1030 struct stable_node
*stable_node
;
1033 stable_node
= page_stable_node(page
);
1034 if (stable_node
) { /* ksm page forked */
1039 nid
= get_kpfn_nid(page_to_pfn(page
));
1040 node
= root_stable_tree
[nid
].rb_node
;
1043 struct page
*tree_page
;
1047 stable_node
= rb_entry(node
, struct stable_node
, node
);
1048 tree_page
= get_ksm_page(stable_node
, false);
1052 ret
= memcmp_pages(page
, tree_page
);
1055 put_page(tree_page
);
1056 node
= node
->rb_left
;
1057 } else if (ret
> 0) {
1058 put_page(tree_page
);
1059 node
= node
->rb_right
;
1068 * stable_tree_insert - insert stable tree node pointing to new ksm page
1069 * into the stable tree.
1071 * This function returns the stable tree node just allocated on success,
1074 static struct stable_node
*stable_tree_insert(struct page
*kpage
)
1078 struct rb_node
**new;
1079 struct rb_node
*parent
= NULL
;
1080 struct stable_node
*stable_node
;
1082 kpfn
= page_to_pfn(kpage
);
1083 nid
= get_kpfn_nid(kpfn
);
1084 new = &root_stable_tree
[nid
].rb_node
;
1087 struct page
*tree_page
;
1091 stable_node
= rb_entry(*new, struct stable_node
, node
);
1092 tree_page
= get_ksm_page(stable_node
, false);
1096 ret
= memcmp_pages(kpage
, tree_page
);
1097 put_page(tree_page
);
1101 new = &parent
->rb_left
;
1103 new = &parent
->rb_right
;
1106 * It is not a bug that stable_tree_search() didn't
1107 * find this node: because at that time our page was
1108 * not yet write-protected, so may have changed since.
1114 stable_node
= alloc_stable_node();
1118 INIT_HLIST_HEAD(&stable_node
->hlist
);
1119 stable_node
->kpfn
= kpfn
;
1120 set_page_stable_node(kpage
, stable_node
);
1121 rb_link_node(&stable_node
->node
, parent
, new);
1122 rb_insert_color(&stable_node
->node
, &root_stable_tree
[nid
]);
1128 * unstable_tree_search_insert - search for identical page,
1129 * else insert rmap_item into the unstable tree.
1131 * This function searches for a page in the unstable tree identical to the
1132 * page currently being scanned; and if no identical page is found in the
1133 * tree, we insert rmap_item as a new object into the unstable tree.
1135 * This function returns pointer to rmap_item found to be identical
1136 * to the currently scanned page, NULL otherwise.
1138 * This function does both searching and inserting, because they share
1139 * the same walking algorithm in an rbtree.
1142 struct rmap_item
*unstable_tree_search_insert(struct rmap_item
*rmap_item
,
1144 struct page
**tree_pagep
)
1146 struct rb_node
**new;
1147 struct rb_root
*root
;
1148 struct rb_node
*parent
= NULL
;
1151 nid
= get_kpfn_nid(page_to_pfn(page
));
1152 root
= &root_unstable_tree
[nid
];
1153 new = &root
->rb_node
;
1156 struct rmap_item
*tree_rmap_item
;
1157 struct page
*tree_page
;
1161 tree_rmap_item
= rb_entry(*new, struct rmap_item
, node
);
1162 tree_page
= get_mergeable_page(tree_rmap_item
);
1163 if (IS_ERR_OR_NULL(tree_page
))
1167 * Don't substitute a ksm page for a forked page.
1169 if (page
== tree_page
) {
1170 put_page(tree_page
);
1175 * If tree_page has been migrated to another NUMA node, it
1176 * will be flushed out and put into the right unstable tree
1177 * next time: only merge with it if merge_across_nodes.
1179 if (!ksm_merge_across_nodes
&& page_to_nid(tree_page
) != nid
) {
1180 put_page(tree_page
);
1184 ret
= memcmp_pages(page
, tree_page
);
1188 put_page(tree_page
);
1189 new = &parent
->rb_left
;
1190 } else if (ret
> 0) {
1191 put_page(tree_page
);
1192 new = &parent
->rb_right
;
1194 *tree_pagep
= tree_page
;
1195 return tree_rmap_item
;
1199 rmap_item
->address
|= UNSTABLE_FLAG
;
1200 rmap_item
->address
|= (ksm_scan
.seqnr
& SEQNR_MASK
);
1201 DO_NUMA(rmap_item
->nid
= nid
);
1202 rb_link_node(&rmap_item
->node
, parent
, new);
1203 rb_insert_color(&rmap_item
->node
, root
);
1205 ksm_pages_unshared
++;
1210 * stable_tree_append - add another rmap_item to the linked list of
1211 * rmap_items hanging off a given node of the stable tree, all sharing
1212 * the same ksm page.
1214 static void stable_tree_append(struct rmap_item
*rmap_item
,
1215 struct stable_node
*stable_node
)
1218 * Usually rmap_item->nid is already set correctly,
1219 * but it may be wrong after switching merge_across_nodes.
1221 DO_NUMA(rmap_item
->nid
= get_kpfn_nid(stable_node
->kpfn
));
1222 rmap_item
->head
= stable_node
;
1223 rmap_item
->address
|= STABLE_FLAG
;
1224 hlist_add_head(&rmap_item
->hlist
, &stable_node
->hlist
);
1226 if (rmap_item
->hlist
.next
)
1227 ksm_pages_sharing
++;
1233 * cmp_and_merge_page - first see if page can be merged into the stable tree;
1234 * if not, compare checksum to previous and if it's the same, see if page can
1235 * be inserted into the unstable tree, or merged with a page already there and
1236 * both transferred to the stable tree.
1238 * @page: the page that we are searching identical page to.
1239 * @rmap_item: the reverse mapping into the virtual address of this page
1241 static void cmp_and_merge_page(struct page
*page
, struct rmap_item
*rmap_item
)
1243 struct rmap_item
*tree_rmap_item
;
1244 struct page
*tree_page
= NULL
;
1245 struct stable_node
*stable_node
;
1247 unsigned int checksum
;
1250 remove_rmap_item_from_tree(rmap_item
);
1252 /* We first start with searching the page inside the stable tree */
1253 kpage
= stable_tree_search(page
);
1255 err
= try_to_merge_with_ksm_page(rmap_item
, page
, kpage
);
1258 * The page was successfully merged:
1259 * add its rmap_item to the stable tree.
1262 stable_tree_append(rmap_item
, page_stable_node(kpage
));
1270 * If the hash value of the page has changed from the last time
1271 * we calculated it, this page is changing frequently: therefore we
1272 * don't want to insert it in the unstable tree, and we don't want
1273 * to waste our time searching for something identical to it there.
1275 checksum
= calc_checksum(page
);
1276 if (rmap_item
->oldchecksum
!= checksum
) {
1277 rmap_item
->oldchecksum
= checksum
;
1282 unstable_tree_search_insert(rmap_item
, page
, &tree_page
);
1283 if (tree_rmap_item
) {
1284 kpage
= try_to_merge_two_pages(rmap_item
, page
,
1285 tree_rmap_item
, tree_page
);
1286 put_page(tree_page
);
1288 * As soon as we merge this page, we want to remove the
1289 * rmap_item of the page we have merged with from the unstable
1290 * tree, and insert it instead as new node in the stable tree.
1293 remove_rmap_item_from_tree(tree_rmap_item
);
1296 stable_node
= stable_tree_insert(kpage
);
1298 stable_tree_append(tree_rmap_item
, stable_node
);
1299 stable_tree_append(rmap_item
, stable_node
);
1304 * If we fail to insert the page into the stable tree,
1305 * we will have 2 virtual addresses that are pointing
1306 * to a ksm page left outside the stable tree,
1307 * in which case we need to break_cow on both.
1310 break_cow(tree_rmap_item
);
1311 break_cow(rmap_item
);
1317 static struct rmap_item
*get_next_rmap_item(struct mm_slot
*mm_slot
,
1318 struct rmap_item
**rmap_list
,
1321 struct rmap_item
*rmap_item
;
1323 while (*rmap_list
) {
1324 rmap_item
= *rmap_list
;
1325 if ((rmap_item
->address
& PAGE_MASK
) == addr
)
1327 if (rmap_item
->address
> addr
)
1329 *rmap_list
= rmap_item
->rmap_list
;
1330 remove_rmap_item_from_tree(rmap_item
);
1331 free_rmap_item(rmap_item
);
1334 rmap_item
= alloc_rmap_item();
1336 /* It has already been zeroed */
1337 rmap_item
->mm
= mm_slot
->mm
;
1338 rmap_item
->address
= addr
;
1339 rmap_item
->rmap_list
= *rmap_list
;
1340 *rmap_list
= rmap_item
;
1345 static struct rmap_item
*scan_get_next_rmap_item(struct page
**page
)
1347 struct mm_struct
*mm
;
1348 struct mm_slot
*slot
;
1349 struct vm_area_struct
*vma
;
1350 struct rmap_item
*rmap_item
;
1353 if (list_empty(&ksm_mm_head
.mm_list
))
1356 slot
= ksm_scan
.mm_slot
;
1357 if (slot
== &ksm_mm_head
) {
1359 * A number of pages can hang around indefinitely on per-cpu
1360 * pagevecs, raised page count preventing write_protect_page
1361 * from merging them. Though it doesn't really matter much,
1362 * it is puzzling to see some stuck in pages_volatile until
1363 * other activity jostles them out, and they also prevented
1364 * LTP's KSM test from succeeding deterministically; so drain
1365 * them here (here rather than on entry to ksm_do_scan(),
1366 * so we don't IPI too often when pages_to_scan is set low).
1368 lru_add_drain_all();
1370 for (nid
= 0; nid
< nr_node_ids
; nid
++)
1371 root_unstable_tree
[nid
] = RB_ROOT
;
1373 spin_lock(&ksm_mmlist_lock
);
1374 slot
= list_entry(slot
->mm_list
.next
, struct mm_slot
, mm_list
);
1375 ksm_scan
.mm_slot
= slot
;
1376 spin_unlock(&ksm_mmlist_lock
);
1378 * Although we tested list_empty() above, a racing __ksm_exit
1379 * of the last mm on the list may have removed it since then.
1381 if (slot
== &ksm_mm_head
)
1384 ksm_scan
.address
= 0;
1385 ksm_scan
.rmap_list
= &slot
->rmap_list
;
1389 down_read(&mm
->mmap_sem
);
1390 if (ksm_test_exit(mm
))
1393 vma
= find_vma(mm
, ksm_scan
.address
);
1395 for (; vma
; vma
= vma
->vm_next
) {
1396 if (!(vma
->vm_flags
& VM_MERGEABLE
))
1398 if (ksm_scan
.address
< vma
->vm_start
)
1399 ksm_scan
.address
= vma
->vm_start
;
1401 ksm_scan
.address
= vma
->vm_end
;
1403 while (ksm_scan
.address
< vma
->vm_end
) {
1404 if (ksm_test_exit(mm
))
1406 *page
= follow_page(vma
, ksm_scan
.address
, FOLL_GET
);
1407 if (IS_ERR_OR_NULL(*page
)) {
1408 ksm_scan
.address
+= PAGE_SIZE
;
1412 if (PageAnon(*page
) ||
1413 page_trans_compound_anon(*page
)) {
1414 flush_anon_page(vma
, *page
, ksm_scan
.address
);
1415 flush_dcache_page(*page
);
1416 rmap_item
= get_next_rmap_item(slot
,
1417 ksm_scan
.rmap_list
, ksm_scan
.address
);
1419 ksm_scan
.rmap_list
=
1420 &rmap_item
->rmap_list
;
1421 ksm_scan
.address
+= PAGE_SIZE
;
1424 up_read(&mm
->mmap_sem
);
1428 ksm_scan
.address
+= PAGE_SIZE
;
1433 if (ksm_test_exit(mm
)) {
1434 ksm_scan
.address
= 0;
1435 ksm_scan
.rmap_list
= &slot
->rmap_list
;
1438 * Nuke all the rmap_items that are above this current rmap:
1439 * because there were no VM_MERGEABLE vmas with such addresses.
1441 remove_trailing_rmap_items(slot
, ksm_scan
.rmap_list
);
1443 spin_lock(&ksm_mmlist_lock
);
1444 ksm_scan
.mm_slot
= list_entry(slot
->mm_list
.next
,
1445 struct mm_slot
, mm_list
);
1446 if (ksm_scan
.address
== 0) {
1448 * We've completed a full scan of all vmas, holding mmap_sem
1449 * throughout, and found no VM_MERGEABLE: so do the same as
1450 * __ksm_exit does to remove this mm from all our lists now.
1451 * This applies either when cleaning up after __ksm_exit
1452 * (but beware: we can reach here even before __ksm_exit),
1453 * or when all VM_MERGEABLE areas have been unmapped (and
1454 * mmap_sem then protects against race with MADV_MERGEABLE).
1456 hash_del(&slot
->link
);
1457 list_del(&slot
->mm_list
);
1458 spin_unlock(&ksm_mmlist_lock
);
1461 clear_bit(MMF_VM_MERGEABLE
, &mm
->flags
);
1462 up_read(&mm
->mmap_sem
);
1465 spin_unlock(&ksm_mmlist_lock
);
1466 up_read(&mm
->mmap_sem
);
1469 /* Repeat until we've completed scanning the whole list */
1470 slot
= ksm_scan
.mm_slot
;
1471 if (slot
!= &ksm_mm_head
)
1479 * ksm_do_scan - the ksm scanner main worker function.
1480 * @scan_npages - number of pages we want to scan before we return.
1482 static void ksm_do_scan(unsigned int scan_npages
)
1484 struct rmap_item
*rmap_item
;
1485 struct page
*uninitialized_var(page
);
1487 while (scan_npages
-- && likely(!freezing(current
))) {
1489 rmap_item
= scan_get_next_rmap_item(&page
);
1492 if (!PageKsm(page
) || !in_stable_tree(rmap_item
))
1493 cmp_and_merge_page(page
, rmap_item
);
1498 static int ksmd_should_run(void)
1500 return (ksm_run
& KSM_RUN_MERGE
) && !list_empty(&ksm_mm_head
.mm_list
);
1503 static int ksm_scan_thread(void *nothing
)
1506 set_user_nice(current
, 5);
1508 while (!kthread_should_stop()) {
1509 mutex_lock(&ksm_thread_mutex
);
1510 if (ksmd_should_run())
1511 ksm_do_scan(ksm_thread_pages_to_scan
);
1512 mutex_unlock(&ksm_thread_mutex
);
1516 if (ksmd_should_run()) {
1517 schedule_timeout_interruptible(
1518 msecs_to_jiffies(ksm_thread_sleep_millisecs
));
1520 wait_event_freezable(ksm_thread_wait
,
1521 ksmd_should_run() || kthread_should_stop());
1527 int ksm_madvise(struct vm_area_struct
*vma
, unsigned long start
,
1528 unsigned long end
, int advice
, unsigned long *vm_flags
)
1530 struct mm_struct
*mm
= vma
->vm_mm
;
1534 case MADV_MERGEABLE
:
1536 * Be somewhat over-protective for now!
1538 if (*vm_flags
& (VM_MERGEABLE
| VM_SHARED
| VM_MAYSHARE
|
1539 VM_PFNMAP
| VM_IO
| VM_DONTEXPAND
|
1540 VM_HUGETLB
| VM_NONLINEAR
| VM_MIXEDMAP
))
1541 return 0; /* just ignore the advice */
1544 if (*vm_flags
& VM_SAO
)
1548 if (!test_bit(MMF_VM_MERGEABLE
, &mm
->flags
)) {
1549 err
= __ksm_enter(mm
);
1554 *vm_flags
|= VM_MERGEABLE
;
1557 case MADV_UNMERGEABLE
:
1558 if (!(*vm_flags
& VM_MERGEABLE
))
1559 return 0; /* just ignore the advice */
1561 if (vma
->anon_vma
) {
1562 err
= unmerge_ksm_pages(vma
, start
, end
);
1567 *vm_flags
&= ~VM_MERGEABLE
;
1574 int __ksm_enter(struct mm_struct
*mm
)
1576 struct mm_slot
*mm_slot
;
1579 mm_slot
= alloc_mm_slot();
1583 /* Check ksm_run too? Would need tighter locking */
1584 needs_wakeup
= list_empty(&ksm_mm_head
.mm_list
);
1586 spin_lock(&ksm_mmlist_lock
);
1587 insert_to_mm_slots_hash(mm
, mm_slot
);
1589 * Insert just behind the scanning cursor, to let the area settle
1590 * down a little; when fork is followed by immediate exec, we don't
1591 * want ksmd to waste time setting up and tearing down an rmap_list.
1593 list_add_tail(&mm_slot
->mm_list
, &ksm_scan
.mm_slot
->mm_list
);
1594 spin_unlock(&ksm_mmlist_lock
);
1596 set_bit(MMF_VM_MERGEABLE
, &mm
->flags
);
1597 atomic_inc(&mm
->mm_count
);
1600 wake_up_interruptible(&ksm_thread_wait
);
1605 void __ksm_exit(struct mm_struct
*mm
)
1607 struct mm_slot
*mm_slot
;
1608 int easy_to_free
= 0;
1611 * This process is exiting: if it's straightforward (as is the
1612 * case when ksmd was never running), free mm_slot immediately.
1613 * But if it's at the cursor or has rmap_items linked to it, use
1614 * mmap_sem to synchronize with any break_cows before pagetables
1615 * are freed, and leave the mm_slot on the list for ksmd to free.
1616 * Beware: ksm may already have noticed it exiting and freed the slot.
1619 spin_lock(&ksm_mmlist_lock
);
1620 mm_slot
= get_mm_slot(mm
);
1621 if (mm_slot
&& ksm_scan
.mm_slot
!= mm_slot
) {
1622 if (!mm_slot
->rmap_list
) {
1623 hash_del(&mm_slot
->link
);
1624 list_del(&mm_slot
->mm_list
);
1627 list_move(&mm_slot
->mm_list
,
1628 &ksm_scan
.mm_slot
->mm_list
);
1631 spin_unlock(&ksm_mmlist_lock
);
1634 free_mm_slot(mm_slot
);
1635 clear_bit(MMF_VM_MERGEABLE
, &mm
->flags
);
1637 } else if (mm_slot
) {
1638 down_write(&mm
->mmap_sem
);
1639 up_write(&mm
->mmap_sem
);
1643 struct page
*ksm_does_need_to_copy(struct page
*page
,
1644 struct vm_area_struct
*vma
, unsigned long address
)
1646 struct page
*new_page
;
1648 new_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
1650 copy_user_highpage(new_page
, page
, address
, vma
);
1652 SetPageDirty(new_page
);
1653 __SetPageUptodate(new_page
);
1654 __set_page_locked(new_page
);
1660 int page_referenced_ksm(struct page
*page
, struct mem_cgroup
*memcg
,
1661 unsigned long *vm_flags
)
1663 struct stable_node
*stable_node
;
1664 struct rmap_item
*rmap_item
;
1665 struct hlist_node
*hlist
;
1666 unsigned int mapcount
= page_mapcount(page
);
1668 int search_new_forks
= 0;
1670 VM_BUG_ON(!PageKsm(page
));
1671 VM_BUG_ON(!PageLocked(page
));
1673 stable_node
= page_stable_node(page
);
1677 hlist_for_each_entry(rmap_item
, hlist
, &stable_node
->hlist
, hlist
) {
1678 struct anon_vma
*anon_vma
= rmap_item
->anon_vma
;
1679 struct anon_vma_chain
*vmac
;
1680 struct vm_area_struct
*vma
;
1682 anon_vma_lock_read(anon_vma
);
1683 anon_vma_interval_tree_foreach(vmac
, &anon_vma
->rb_root
,
1686 if (rmap_item
->address
< vma
->vm_start
||
1687 rmap_item
->address
>= vma
->vm_end
)
1690 * Initially we examine only the vma which covers this
1691 * rmap_item; but later, if there is still work to do,
1692 * we examine covering vmas in other mms: in case they
1693 * were forked from the original since ksmd passed.
1695 if ((rmap_item
->mm
== vma
->vm_mm
) == search_new_forks
)
1698 if (memcg
&& !mm_match_cgroup(vma
->vm_mm
, memcg
))
1701 referenced
+= page_referenced_one(page
, vma
,
1702 rmap_item
->address
, &mapcount
, vm_flags
);
1703 if (!search_new_forks
|| !mapcount
)
1706 anon_vma_unlock_read(anon_vma
);
1710 if (!search_new_forks
++)
1716 int try_to_unmap_ksm(struct page
*page
, enum ttu_flags flags
)
1718 struct stable_node
*stable_node
;
1719 struct hlist_node
*hlist
;
1720 struct rmap_item
*rmap_item
;
1721 int ret
= SWAP_AGAIN
;
1722 int search_new_forks
= 0;
1724 VM_BUG_ON(!PageKsm(page
));
1725 VM_BUG_ON(!PageLocked(page
));
1727 stable_node
= page_stable_node(page
);
1731 hlist_for_each_entry(rmap_item
, hlist
, &stable_node
->hlist
, hlist
) {
1732 struct anon_vma
*anon_vma
= rmap_item
->anon_vma
;
1733 struct anon_vma_chain
*vmac
;
1734 struct vm_area_struct
*vma
;
1736 anon_vma_lock_read(anon_vma
);
1737 anon_vma_interval_tree_foreach(vmac
, &anon_vma
->rb_root
,
1740 if (rmap_item
->address
< vma
->vm_start
||
1741 rmap_item
->address
>= vma
->vm_end
)
1744 * Initially we examine only the vma which covers this
1745 * rmap_item; but later, if there is still work to do,
1746 * we examine covering vmas in other mms: in case they
1747 * were forked from the original since ksmd passed.
1749 if ((rmap_item
->mm
== vma
->vm_mm
) == search_new_forks
)
1752 ret
= try_to_unmap_one(page
, vma
,
1753 rmap_item
->address
, flags
);
1754 if (ret
!= SWAP_AGAIN
|| !page_mapped(page
)) {
1755 anon_vma_unlock_read(anon_vma
);
1759 anon_vma_unlock_read(anon_vma
);
1761 if (!search_new_forks
++)
1767 #ifdef CONFIG_MIGRATION
1768 int rmap_walk_ksm(struct page
*page
, int (*rmap_one
)(struct page
*,
1769 struct vm_area_struct
*, unsigned long, void *), void *arg
)
1771 struct stable_node
*stable_node
;
1772 struct hlist_node
*hlist
;
1773 struct rmap_item
*rmap_item
;
1774 int ret
= SWAP_AGAIN
;
1775 int search_new_forks
= 0;
1777 VM_BUG_ON(!PageKsm(page
));
1778 VM_BUG_ON(!PageLocked(page
));
1780 stable_node
= page_stable_node(page
);
1784 hlist_for_each_entry(rmap_item
, hlist
, &stable_node
->hlist
, hlist
) {
1785 struct anon_vma
*anon_vma
= rmap_item
->anon_vma
;
1786 struct anon_vma_chain
*vmac
;
1787 struct vm_area_struct
*vma
;
1789 anon_vma_lock_read(anon_vma
);
1790 anon_vma_interval_tree_foreach(vmac
, &anon_vma
->rb_root
,
1793 if (rmap_item
->address
< vma
->vm_start
||
1794 rmap_item
->address
>= vma
->vm_end
)
1797 * Initially we examine only the vma which covers this
1798 * rmap_item; but later, if there is still work to do,
1799 * we examine covering vmas in other mms: in case they
1800 * were forked from the original since ksmd passed.
1802 if ((rmap_item
->mm
== vma
->vm_mm
) == search_new_forks
)
1805 ret
= rmap_one(page
, vma
, rmap_item
->address
, arg
);
1806 if (ret
!= SWAP_AGAIN
) {
1807 anon_vma_unlock_read(anon_vma
);
1811 anon_vma_unlock_read(anon_vma
);
1813 if (!search_new_forks
++)
1819 void ksm_migrate_page(struct page
*newpage
, struct page
*oldpage
)
1821 struct stable_node
*stable_node
;
1823 VM_BUG_ON(!PageLocked(oldpage
));
1824 VM_BUG_ON(!PageLocked(newpage
));
1825 VM_BUG_ON(newpage
->mapping
!= oldpage
->mapping
);
1827 stable_node
= page_stable_node(newpage
);
1829 VM_BUG_ON(stable_node
->kpfn
!= page_to_pfn(oldpage
));
1830 stable_node
->kpfn
= page_to_pfn(newpage
);
1833 #endif /* CONFIG_MIGRATION */
1835 #ifdef CONFIG_MEMORY_HOTREMOVE
1836 static void ksm_check_stable_tree(unsigned long start_pfn
,
1837 unsigned long end_pfn
)
1839 struct stable_node
*stable_node
;
1840 struct rb_node
*node
;
1843 for (nid
= 0; nid
< nr_node_ids
; nid
++) {
1844 node
= rb_first(&root_stable_tree
[nid
]);
1846 stable_node
= rb_entry(node
, struct stable_node
, node
);
1847 if (stable_node
->kpfn
>= start_pfn
&&
1848 stable_node
->kpfn
< end_pfn
) {
1850 * Don't get_ksm_page, page has already gone:
1851 * which is why we keep kpfn instead of page*
1853 remove_node_from_stable_tree(stable_node
);
1854 node
= rb_first(&root_stable_tree
[nid
]);
1856 node
= rb_next(node
);
1862 static int ksm_memory_callback(struct notifier_block
*self
,
1863 unsigned long action
, void *arg
)
1865 struct memory_notify
*mn
= arg
;
1868 case MEM_GOING_OFFLINE
:
1870 * Keep it very simple for now: just lock out ksmd and
1871 * MADV_UNMERGEABLE while any memory is going offline.
1872 * mutex_lock_nested() is necessary because lockdep was alarmed
1873 * that here we take ksm_thread_mutex inside notifier chain
1874 * mutex, and later take notifier chain mutex inside
1875 * ksm_thread_mutex to unlock it. But that's safe because both
1876 * are inside mem_hotplug_mutex.
1878 mutex_lock_nested(&ksm_thread_mutex
, SINGLE_DEPTH_NESTING
);
1883 * Most of the work is done by page migration; but there might
1884 * be a few stable_nodes left over, still pointing to struct
1885 * pages which have been offlined: prune those from the tree,
1886 * otherwise get_ksm_page() might later try to access a
1887 * non-existent struct page.
1889 ksm_check_stable_tree(mn
->start_pfn
,
1890 mn
->start_pfn
+ mn
->nr_pages
);
1893 case MEM_CANCEL_OFFLINE
:
1894 mutex_unlock(&ksm_thread_mutex
);
1899 #endif /* CONFIG_MEMORY_HOTREMOVE */
1903 * This all compiles without CONFIG_SYSFS, but is a waste of space.
1906 #define KSM_ATTR_RO(_name) \
1907 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1908 #define KSM_ATTR(_name) \
1909 static struct kobj_attribute _name##_attr = \
1910 __ATTR(_name, 0644, _name##_show, _name##_store)
1912 static ssize_t
sleep_millisecs_show(struct kobject
*kobj
,
1913 struct kobj_attribute
*attr
, char *buf
)
1915 return sprintf(buf
, "%u\n", ksm_thread_sleep_millisecs
);
1918 static ssize_t
sleep_millisecs_store(struct kobject
*kobj
,
1919 struct kobj_attribute
*attr
,
1920 const char *buf
, size_t count
)
1922 unsigned long msecs
;
1925 err
= strict_strtoul(buf
, 10, &msecs
);
1926 if (err
|| msecs
> UINT_MAX
)
1929 ksm_thread_sleep_millisecs
= msecs
;
1933 KSM_ATTR(sleep_millisecs
);
1935 static ssize_t
pages_to_scan_show(struct kobject
*kobj
,
1936 struct kobj_attribute
*attr
, char *buf
)
1938 return sprintf(buf
, "%u\n", ksm_thread_pages_to_scan
);
1941 static ssize_t
pages_to_scan_store(struct kobject
*kobj
,
1942 struct kobj_attribute
*attr
,
1943 const char *buf
, size_t count
)
1946 unsigned long nr_pages
;
1948 err
= strict_strtoul(buf
, 10, &nr_pages
);
1949 if (err
|| nr_pages
> UINT_MAX
)
1952 ksm_thread_pages_to_scan
= nr_pages
;
1956 KSM_ATTR(pages_to_scan
);
1958 static ssize_t
run_show(struct kobject
*kobj
, struct kobj_attribute
*attr
,
1961 return sprintf(buf
, "%u\n", ksm_run
);
1964 static ssize_t
run_store(struct kobject
*kobj
, struct kobj_attribute
*attr
,
1965 const char *buf
, size_t count
)
1968 unsigned long flags
;
1970 err
= strict_strtoul(buf
, 10, &flags
);
1971 if (err
|| flags
> UINT_MAX
)
1973 if (flags
> KSM_RUN_UNMERGE
)
1977 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
1978 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
1979 * breaking COW to free the pages_shared (but leaves mm_slots
1980 * on the list for when ksmd may be set running again).
1983 mutex_lock(&ksm_thread_mutex
);
1984 if (ksm_run
!= flags
) {
1986 if (flags
& KSM_RUN_UNMERGE
) {
1987 set_current_oom_origin();
1988 err
= unmerge_and_remove_all_rmap_items();
1989 clear_current_oom_origin();
1991 ksm_run
= KSM_RUN_STOP
;
1996 mutex_unlock(&ksm_thread_mutex
);
1998 if (flags
& KSM_RUN_MERGE
)
1999 wake_up_interruptible(&ksm_thread_wait
);
2006 static ssize_t
merge_across_nodes_show(struct kobject
*kobj
,
2007 struct kobj_attribute
*attr
, char *buf
)
2009 return sprintf(buf
, "%u\n", ksm_merge_across_nodes
);
2012 static ssize_t
merge_across_nodes_store(struct kobject
*kobj
,
2013 struct kobj_attribute
*attr
,
2014 const char *buf
, size_t count
)
2019 err
= kstrtoul(buf
, 10, &knob
);
2025 mutex_lock(&ksm_thread_mutex
);
2026 if (ksm_merge_across_nodes
!= knob
) {
2027 if (ksm_pages_shared
)
2030 ksm_merge_across_nodes
= knob
;
2032 mutex_unlock(&ksm_thread_mutex
);
2034 return err
? err
: count
;
2036 KSM_ATTR(merge_across_nodes
);
2039 static ssize_t
pages_shared_show(struct kobject
*kobj
,
2040 struct kobj_attribute
*attr
, char *buf
)
2042 return sprintf(buf
, "%lu\n", ksm_pages_shared
);
2044 KSM_ATTR_RO(pages_shared
);
2046 static ssize_t
pages_sharing_show(struct kobject
*kobj
,
2047 struct kobj_attribute
*attr
, char *buf
)
2049 return sprintf(buf
, "%lu\n", ksm_pages_sharing
);
2051 KSM_ATTR_RO(pages_sharing
);
2053 static ssize_t
pages_unshared_show(struct kobject
*kobj
,
2054 struct kobj_attribute
*attr
, char *buf
)
2056 return sprintf(buf
, "%lu\n", ksm_pages_unshared
);
2058 KSM_ATTR_RO(pages_unshared
);
2060 static ssize_t
pages_volatile_show(struct kobject
*kobj
,
2061 struct kobj_attribute
*attr
, char *buf
)
2063 long ksm_pages_volatile
;
2065 ksm_pages_volatile
= ksm_rmap_items
- ksm_pages_shared
2066 - ksm_pages_sharing
- ksm_pages_unshared
;
2068 * It was not worth any locking to calculate that statistic,
2069 * but it might therefore sometimes be negative: conceal that.
2071 if (ksm_pages_volatile
< 0)
2072 ksm_pages_volatile
= 0;
2073 return sprintf(buf
, "%ld\n", ksm_pages_volatile
);
2075 KSM_ATTR_RO(pages_volatile
);
2077 static ssize_t
full_scans_show(struct kobject
*kobj
,
2078 struct kobj_attribute
*attr
, char *buf
)
2080 return sprintf(buf
, "%lu\n", ksm_scan
.seqnr
);
2082 KSM_ATTR_RO(full_scans
);
2084 static struct attribute
*ksm_attrs
[] = {
2085 &sleep_millisecs_attr
.attr
,
2086 &pages_to_scan_attr
.attr
,
2088 &pages_shared_attr
.attr
,
2089 &pages_sharing_attr
.attr
,
2090 &pages_unshared_attr
.attr
,
2091 &pages_volatile_attr
.attr
,
2092 &full_scans_attr
.attr
,
2094 &merge_across_nodes_attr
.attr
,
2099 static struct attribute_group ksm_attr_group
= {
2103 #endif /* CONFIG_SYSFS */
2105 static int __init
ksm_init(void)
2107 struct task_struct
*ksm_thread
;
2111 err
= ksm_slab_init();
2115 for (nid
= 0; nid
< nr_node_ids
; nid
++)
2116 root_stable_tree
[nid
] = RB_ROOT
;
2118 ksm_thread
= kthread_run(ksm_scan_thread
, NULL
, "ksmd");
2119 if (IS_ERR(ksm_thread
)) {
2120 printk(KERN_ERR
"ksm: creating kthread failed\n");
2121 err
= PTR_ERR(ksm_thread
);
2126 err
= sysfs_create_group(mm_kobj
, &ksm_attr_group
);
2128 printk(KERN_ERR
"ksm: register sysfs failed\n");
2129 kthread_stop(ksm_thread
);
2133 ksm_run
= KSM_RUN_MERGE
; /* no way for user to start it */
2135 #endif /* CONFIG_SYSFS */
2137 #ifdef CONFIG_MEMORY_HOTREMOVE
2139 * Choose a high priority since the callback takes ksm_thread_mutex:
2140 * later callbacks could only be taking locks which nest within that.
2142 hotplug_memory_notifier(ksm_memory_callback
, 100);
2151 module_init(ksm_init
)