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). The RCU calls are not for KSM at all, but
518 * to keep the page_count protocol described with page_cache_get_speculative.
520 * Note: it is possible that get_ksm_page() will return NULL one moment,
521 * then page the next, if the page is in between page_freeze_refs() and
522 * page_unfreeze_refs(): this shouldn't be a problem anywhere, the page
523 * is on its way to being freed; but it is an anomaly to bear in mind.
525 static struct page
*get_ksm_page(struct stable_node
*stable_node
)
528 void *expected_mapping
;
530 page
= pfn_to_page(stable_node
->kpfn
);
531 expected_mapping
= (void *)stable_node
+
532 (PAGE_MAPPING_ANON
| PAGE_MAPPING_KSM
);
534 if (page
->mapping
!= expected_mapping
)
536 if (!get_page_unless_zero(page
))
538 if (page
->mapping
!= expected_mapping
) {
546 remove_node_from_stable_tree(stable_node
);
551 * Removing rmap_item from stable or unstable tree.
552 * This function will clean the information from the stable/unstable tree.
554 static void remove_rmap_item_from_tree(struct rmap_item
*rmap_item
)
556 if (rmap_item
->address
& STABLE_FLAG
) {
557 struct stable_node
*stable_node
;
560 stable_node
= rmap_item
->head
;
561 page
= get_ksm_page(stable_node
);
566 hlist_del(&rmap_item
->hlist
);
570 if (stable_node
->hlist
.first
)
575 put_anon_vma(rmap_item
->anon_vma
);
576 rmap_item
->address
&= PAGE_MASK
;
578 } else if (rmap_item
->address
& UNSTABLE_FLAG
) {
581 * Usually ksmd can and must skip the rb_erase, because
582 * root_unstable_tree was already reset to RB_ROOT.
583 * But be careful when an mm is exiting: do the rb_erase
584 * if this rmap_item was inserted by this scan, rather
585 * than left over from before.
587 age
= (unsigned char)(ksm_scan
.seqnr
- rmap_item
->address
);
590 rb_erase(&rmap_item
->node
,
591 &root_unstable_tree
[NUMA(rmap_item
->nid
)]);
592 ksm_pages_unshared
--;
593 rmap_item
->address
&= PAGE_MASK
;
596 cond_resched(); /* we're called from many long loops */
599 static void remove_trailing_rmap_items(struct mm_slot
*mm_slot
,
600 struct rmap_item
**rmap_list
)
603 struct rmap_item
*rmap_item
= *rmap_list
;
604 *rmap_list
= rmap_item
->rmap_list
;
605 remove_rmap_item_from_tree(rmap_item
);
606 free_rmap_item(rmap_item
);
611 * Though it's very tempting to unmerge rmap_items from stable tree rather
612 * than check every pte of a given vma, the locking doesn't quite work for
613 * that - an rmap_item is assigned to the stable tree after inserting ksm
614 * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing
615 * rmap_items from parent to child at fork time (so as not to waste time
616 * if exit comes before the next scan reaches it).
618 * Similarly, although we'd like to remove rmap_items (so updating counts
619 * and freeing memory) when unmerging an area, it's easier to leave that
620 * to the next pass of ksmd - consider, for example, how ksmd might be
621 * in cmp_and_merge_page on one of the rmap_items we would be removing.
623 static int unmerge_ksm_pages(struct vm_area_struct
*vma
,
624 unsigned long start
, unsigned long end
)
629 for (addr
= start
; addr
< end
&& !err
; addr
+= PAGE_SIZE
) {
630 if (ksm_test_exit(vma
->vm_mm
))
632 if (signal_pending(current
))
635 err
= break_ksm(vma
, addr
);
642 * Only called through the sysfs control interface:
644 static int unmerge_and_remove_all_rmap_items(void)
646 struct mm_slot
*mm_slot
;
647 struct mm_struct
*mm
;
648 struct vm_area_struct
*vma
;
651 spin_lock(&ksm_mmlist_lock
);
652 ksm_scan
.mm_slot
= list_entry(ksm_mm_head
.mm_list
.next
,
653 struct mm_slot
, mm_list
);
654 spin_unlock(&ksm_mmlist_lock
);
656 for (mm_slot
= ksm_scan
.mm_slot
;
657 mm_slot
!= &ksm_mm_head
; mm_slot
= ksm_scan
.mm_slot
) {
659 down_read(&mm
->mmap_sem
);
660 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
661 if (ksm_test_exit(mm
))
663 if (!(vma
->vm_flags
& VM_MERGEABLE
) || !vma
->anon_vma
)
665 err
= unmerge_ksm_pages(vma
,
666 vma
->vm_start
, vma
->vm_end
);
671 remove_trailing_rmap_items(mm_slot
, &mm_slot
->rmap_list
);
673 spin_lock(&ksm_mmlist_lock
);
674 ksm_scan
.mm_slot
= list_entry(mm_slot
->mm_list
.next
,
675 struct mm_slot
, mm_list
);
676 if (ksm_test_exit(mm
)) {
677 hash_del(&mm_slot
->link
);
678 list_del(&mm_slot
->mm_list
);
679 spin_unlock(&ksm_mmlist_lock
);
681 free_mm_slot(mm_slot
);
682 clear_bit(MMF_VM_MERGEABLE
, &mm
->flags
);
683 up_read(&mm
->mmap_sem
);
686 spin_unlock(&ksm_mmlist_lock
);
687 up_read(&mm
->mmap_sem
);
695 up_read(&mm
->mmap_sem
);
696 spin_lock(&ksm_mmlist_lock
);
697 ksm_scan
.mm_slot
= &ksm_mm_head
;
698 spin_unlock(&ksm_mmlist_lock
);
701 #endif /* CONFIG_SYSFS */
703 static u32
calc_checksum(struct page
*page
)
706 void *addr
= kmap_atomic(page
);
707 checksum
= jhash2(addr
, PAGE_SIZE
/ 4, 17);
712 static int memcmp_pages(struct page
*page1
, struct page
*page2
)
717 addr1
= kmap_atomic(page1
);
718 addr2
= kmap_atomic(page2
);
719 ret
= memcmp(addr1
, addr2
, PAGE_SIZE
);
720 kunmap_atomic(addr2
);
721 kunmap_atomic(addr1
);
725 static inline int pages_identical(struct page
*page1
, struct page
*page2
)
727 return !memcmp_pages(page1
, page2
);
730 static int write_protect_page(struct vm_area_struct
*vma
, struct page
*page
,
733 struct mm_struct
*mm
= vma
->vm_mm
;
739 unsigned long mmun_start
; /* For mmu_notifiers */
740 unsigned long mmun_end
; /* For mmu_notifiers */
742 addr
= page_address_in_vma(page
, vma
);
746 BUG_ON(PageTransCompound(page
));
749 mmun_end
= addr
+ PAGE_SIZE
;
750 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
752 ptep
= page_check_address(page
, mm
, addr
, &ptl
, 0);
756 if (pte_write(*ptep
) || pte_dirty(*ptep
)) {
759 swapped
= PageSwapCache(page
);
760 flush_cache_page(vma
, addr
, page_to_pfn(page
));
762 * Ok this is tricky, when get_user_pages_fast() run it doesn't
763 * take any lock, therefore the check that we are going to make
764 * with the pagecount against the mapcount is racey and
765 * O_DIRECT can happen right after the check.
766 * So we clear the pte and flush the tlb before the check
767 * this assure us that no O_DIRECT can happen after the check
768 * or in the middle of the check.
770 entry
= ptep_clear_flush(vma
, addr
, ptep
);
772 * Check that no O_DIRECT or similar I/O is in progress on the
775 if (page_mapcount(page
) + 1 + swapped
!= page_count(page
)) {
776 set_pte_at(mm
, addr
, ptep
, entry
);
779 if (pte_dirty(entry
))
780 set_page_dirty(page
);
781 entry
= pte_mkclean(pte_wrprotect(entry
));
782 set_pte_at_notify(mm
, addr
, ptep
, entry
);
788 pte_unmap_unlock(ptep
, ptl
);
790 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
796 * replace_page - replace page in vma by new ksm page
797 * @vma: vma that holds the pte pointing to page
798 * @page: the page we are replacing by kpage
799 * @kpage: the ksm page we replace page by
800 * @orig_pte: the original value of the pte
802 * Returns 0 on success, -EFAULT on failure.
804 static int replace_page(struct vm_area_struct
*vma
, struct page
*page
,
805 struct page
*kpage
, pte_t orig_pte
)
807 struct mm_struct
*mm
= vma
->vm_mm
;
813 unsigned long mmun_start
; /* For mmu_notifiers */
814 unsigned long mmun_end
; /* For mmu_notifiers */
816 addr
= page_address_in_vma(page
, vma
);
820 pmd
= mm_find_pmd(mm
, addr
);
823 BUG_ON(pmd_trans_huge(*pmd
));
826 mmun_end
= addr
+ PAGE_SIZE
;
827 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
829 ptep
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
830 if (!pte_same(*ptep
, orig_pte
)) {
831 pte_unmap_unlock(ptep
, ptl
);
836 page_add_anon_rmap(kpage
, vma
, addr
);
838 flush_cache_page(vma
, addr
, pte_pfn(*ptep
));
839 ptep_clear_flush(vma
, addr
, ptep
);
840 set_pte_at_notify(mm
, addr
, ptep
, mk_pte(kpage
, vma
->vm_page_prot
));
842 page_remove_rmap(page
);
843 if (!page_mapped(page
))
844 try_to_free_swap(page
);
847 pte_unmap_unlock(ptep
, ptl
);
850 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
855 static int page_trans_compound_anon_split(struct page
*page
)
858 struct page
*transhuge_head
= page_trans_compound_anon(page
);
859 if (transhuge_head
) {
860 /* Get the reference on the head to split it. */
861 if (get_page_unless_zero(transhuge_head
)) {
863 * Recheck we got the reference while the head
864 * was still anonymous.
866 if (PageAnon(transhuge_head
))
867 ret
= split_huge_page(transhuge_head
);
870 * Retry later if split_huge_page run
874 put_page(transhuge_head
);
876 /* Retry later if split_huge_page run from under us. */
883 * try_to_merge_one_page - take two pages and merge them into one
884 * @vma: the vma that holds the pte pointing to page
885 * @page: the PageAnon page that we want to replace with kpage
886 * @kpage: the PageKsm page that we want to map instead of page,
887 * or NULL the first time when we want to use page as kpage.
889 * This function returns 0 if the pages were merged, -EFAULT otherwise.
891 static int try_to_merge_one_page(struct vm_area_struct
*vma
,
892 struct page
*page
, struct page
*kpage
)
894 pte_t orig_pte
= __pte(0);
897 if (page
== kpage
) /* ksm page forked */
900 if (!(vma
->vm_flags
& VM_MERGEABLE
))
902 if (PageTransCompound(page
) && page_trans_compound_anon_split(page
))
904 BUG_ON(PageTransCompound(page
));
909 * We need the page lock to read a stable PageSwapCache in
910 * write_protect_page(). We use trylock_page() instead of
911 * lock_page() because we don't want to wait here - we
912 * prefer to continue scanning and merging different pages,
913 * then come back to this page when it is unlocked.
915 if (!trylock_page(page
))
918 * If this anonymous page is mapped only here, its pte may need
919 * to be write-protected. If it's mapped elsewhere, all of its
920 * ptes are necessarily already write-protected. But in either
921 * case, we need to lock and check page_count is not raised.
923 if (write_protect_page(vma
, page
, &orig_pte
) == 0) {
926 * While we hold page lock, upgrade page from
927 * PageAnon+anon_vma to PageKsm+NULL stable_node:
928 * stable_tree_insert() will update stable_node.
930 set_page_stable_node(page
, NULL
);
931 mark_page_accessed(page
);
933 } else if (pages_identical(page
, kpage
))
934 err
= replace_page(vma
, page
, kpage
, orig_pte
);
937 if ((vma
->vm_flags
& VM_LOCKED
) && kpage
&& !err
) {
938 munlock_vma_page(page
);
939 if (!PageMlocked(kpage
)) {
942 mlock_vma_page(kpage
);
943 page
= kpage
; /* for final unlock */
953 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
954 * but no new kernel page is allocated: kpage must already be a ksm page.
956 * This function returns 0 if the pages were merged, -EFAULT otherwise.
958 static int try_to_merge_with_ksm_page(struct rmap_item
*rmap_item
,
959 struct page
*page
, struct page
*kpage
)
961 struct mm_struct
*mm
= rmap_item
->mm
;
962 struct vm_area_struct
*vma
;
965 down_read(&mm
->mmap_sem
);
966 if (ksm_test_exit(mm
))
968 vma
= find_vma(mm
, rmap_item
->address
);
969 if (!vma
|| vma
->vm_start
> rmap_item
->address
)
972 err
= try_to_merge_one_page(vma
, page
, kpage
);
976 /* Must get reference to anon_vma while still holding mmap_sem */
977 rmap_item
->anon_vma
= vma
->anon_vma
;
978 get_anon_vma(vma
->anon_vma
);
980 up_read(&mm
->mmap_sem
);
985 * try_to_merge_two_pages - take two identical pages and prepare them
986 * to be merged into one page.
988 * This function returns the kpage if we successfully merged two identical
989 * pages into one ksm page, NULL otherwise.
991 * Note that this function upgrades page to ksm page: if one of the pages
992 * is already a ksm page, try_to_merge_with_ksm_page should be used.
994 static struct page
*try_to_merge_two_pages(struct rmap_item
*rmap_item
,
996 struct rmap_item
*tree_rmap_item
,
997 struct page
*tree_page
)
1001 err
= try_to_merge_with_ksm_page(rmap_item
, page
, NULL
);
1003 err
= try_to_merge_with_ksm_page(tree_rmap_item
,
1006 * If that fails, we have a ksm page with only one pte
1007 * pointing to it: so break it.
1010 break_cow(rmap_item
);
1012 return err
? NULL
: page
;
1016 * stable_tree_search - search for page inside the stable tree
1018 * This function checks if there is a page inside the stable tree
1019 * with identical content to the page that we are scanning right now.
1021 * This function returns the stable tree node of identical content if found,
1024 static struct page
*stable_tree_search(struct page
*page
)
1026 struct rb_node
*node
;
1027 struct stable_node
*stable_node
;
1030 stable_node
= page_stable_node(page
);
1031 if (stable_node
) { /* ksm page forked */
1036 nid
= get_kpfn_nid(page_to_pfn(page
));
1037 node
= root_stable_tree
[nid
].rb_node
;
1040 struct page
*tree_page
;
1044 stable_node
= rb_entry(node
, struct stable_node
, node
);
1045 tree_page
= get_ksm_page(stable_node
);
1049 ret
= memcmp_pages(page
, tree_page
);
1052 put_page(tree_page
);
1053 node
= node
->rb_left
;
1054 } else if (ret
> 0) {
1055 put_page(tree_page
);
1056 node
= node
->rb_right
;
1065 * stable_tree_insert - insert stable tree node pointing to new ksm page
1066 * into the stable tree.
1068 * This function returns the stable tree node just allocated on success,
1071 static struct stable_node
*stable_tree_insert(struct page
*kpage
)
1075 struct rb_node
**new;
1076 struct rb_node
*parent
= NULL
;
1077 struct stable_node
*stable_node
;
1079 kpfn
= page_to_pfn(kpage
);
1080 nid
= get_kpfn_nid(kpfn
);
1081 new = &root_stable_tree
[nid
].rb_node
;
1084 struct page
*tree_page
;
1088 stable_node
= rb_entry(*new, struct stable_node
, node
);
1089 tree_page
= get_ksm_page(stable_node
);
1093 ret
= memcmp_pages(kpage
, tree_page
);
1094 put_page(tree_page
);
1098 new = &parent
->rb_left
;
1100 new = &parent
->rb_right
;
1103 * It is not a bug that stable_tree_search() didn't
1104 * find this node: because at that time our page was
1105 * not yet write-protected, so may have changed since.
1111 stable_node
= alloc_stable_node();
1115 INIT_HLIST_HEAD(&stable_node
->hlist
);
1116 stable_node
->kpfn
= kpfn
;
1117 set_page_stable_node(kpage
, stable_node
);
1118 rb_link_node(&stable_node
->node
, parent
, new);
1119 rb_insert_color(&stable_node
->node
, &root_stable_tree
[nid
]);
1125 * unstable_tree_search_insert - search for identical page,
1126 * else insert rmap_item into the unstable tree.
1128 * This function searches for a page in the unstable tree identical to the
1129 * page currently being scanned; and if no identical page is found in the
1130 * tree, we insert rmap_item as a new object into the unstable tree.
1132 * This function returns pointer to rmap_item found to be identical
1133 * to the currently scanned page, NULL otherwise.
1135 * This function does both searching and inserting, because they share
1136 * the same walking algorithm in an rbtree.
1139 struct rmap_item
*unstable_tree_search_insert(struct rmap_item
*rmap_item
,
1141 struct page
**tree_pagep
)
1143 struct rb_node
**new;
1144 struct rb_root
*root
;
1145 struct rb_node
*parent
= NULL
;
1148 nid
= get_kpfn_nid(page_to_pfn(page
));
1149 root
= &root_unstable_tree
[nid
];
1150 new = &root
->rb_node
;
1153 struct rmap_item
*tree_rmap_item
;
1154 struct page
*tree_page
;
1158 tree_rmap_item
= rb_entry(*new, struct rmap_item
, node
);
1159 tree_page
= get_mergeable_page(tree_rmap_item
);
1160 if (IS_ERR_OR_NULL(tree_page
))
1164 * Don't substitute a ksm page for a forked page.
1166 if (page
== tree_page
) {
1167 put_page(tree_page
);
1172 * If tree_page has been migrated to another NUMA node, it
1173 * will be flushed out and put into the right unstable tree
1174 * next time: only merge with it if merge_across_nodes.
1176 if (!ksm_merge_across_nodes
&& page_to_nid(tree_page
) != nid
) {
1177 put_page(tree_page
);
1181 ret
= memcmp_pages(page
, tree_page
);
1185 put_page(tree_page
);
1186 new = &parent
->rb_left
;
1187 } else if (ret
> 0) {
1188 put_page(tree_page
);
1189 new = &parent
->rb_right
;
1191 *tree_pagep
= tree_page
;
1192 return tree_rmap_item
;
1196 rmap_item
->address
|= UNSTABLE_FLAG
;
1197 rmap_item
->address
|= (ksm_scan
.seqnr
& SEQNR_MASK
);
1198 DO_NUMA(rmap_item
->nid
= nid
);
1199 rb_link_node(&rmap_item
->node
, parent
, new);
1200 rb_insert_color(&rmap_item
->node
, root
);
1202 ksm_pages_unshared
++;
1207 * stable_tree_append - add another rmap_item to the linked list of
1208 * rmap_items hanging off a given node of the stable tree, all sharing
1209 * the same ksm page.
1211 static void stable_tree_append(struct rmap_item
*rmap_item
,
1212 struct stable_node
*stable_node
)
1215 * Usually rmap_item->nid is already set correctly,
1216 * but it may be wrong after switching merge_across_nodes.
1218 DO_NUMA(rmap_item
->nid
= get_kpfn_nid(stable_node
->kpfn
));
1219 rmap_item
->head
= stable_node
;
1220 rmap_item
->address
|= STABLE_FLAG
;
1221 hlist_add_head(&rmap_item
->hlist
, &stable_node
->hlist
);
1223 if (rmap_item
->hlist
.next
)
1224 ksm_pages_sharing
++;
1230 * cmp_and_merge_page - first see if page can be merged into the stable tree;
1231 * if not, compare checksum to previous and if it's the same, see if page can
1232 * be inserted into the unstable tree, or merged with a page already there and
1233 * both transferred to the stable tree.
1235 * @page: the page that we are searching identical page to.
1236 * @rmap_item: the reverse mapping into the virtual address of this page
1238 static void cmp_and_merge_page(struct page
*page
, struct rmap_item
*rmap_item
)
1240 struct rmap_item
*tree_rmap_item
;
1241 struct page
*tree_page
= NULL
;
1242 struct stable_node
*stable_node
;
1244 unsigned int checksum
;
1247 remove_rmap_item_from_tree(rmap_item
);
1249 /* We first start with searching the page inside the stable tree */
1250 kpage
= stable_tree_search(page
);
1252 err
= try_to_merge_with_ksm_page(rmap_item
, page
, kpage
);
1255 * The page was successfully merged:
1256 * add its rmap_item to the stable tree.
1259 stable_tree_append(rmap_item
, page_stable_node(kpage
));
1267 * If the hash value of the page has changed from the last time
1268 * we calculated it, this page is changing frequently: therefore we
1269 * don't want to insert it in the unstable tree, and we don't want
1270 * to waste our time searching for something identical to it there.
1272 checksum
= calc_checksum(page
);
1273 if (rmap_item
->oldchecksum
!= checksum
) {
1274 rmap_item
->oldchecksum
= checksum
;
1279 unstable_tree_search_insert(rmap_item
, page
, &tree_page
);
1280 if (tree_rmap_item
) {
1281 kpage
= try_to_merge_two_pages(rmap_item
, page
,
1282 tree_rmap_item
, tree_page
);
1283 put_page(tree_page
);
1285 * As soon as we merge this page, we want to remove the
1286 * rmap_item of the page we have merged with from the unstable
1287 * tree, and insert it instead as new node in the stable tree.
1290 remove_rmap_item_from_tree(tree_rmap_item
);
1293 stable_node
= stable_tree_insert(kpage
);
1295 stable_tree_append(tree_rmap_item
, stable_node
);
1296 stable_tree_append(rmap_item
, stable_node
);
1301 * If we fail to insert the page into the stable tree,
1302 * we will have 2 virtual addresses that are pointing
1303 * to a ksm page left outside the stable tree,
1304 * in which case we need to break_cow on both.
1307 break_cow(tree_rmap_item
);
1308 break_cow(rmap_item
);
1314 static struct rmap_item
*get_next_rmap_item(struct mm_slot
*mm_slot
,
1315 struct rmap_item
**rmap_list
,
1318 struct rmap_item
*rmap_item
;
1320 while (*rmap_list
) {
1321 rmap_item
= *rmap_list
;
1322 if ((rmap_item
->address
& PAGE_MASK
) == addr
)
1324 if (rmap_item
->address
> addr
)
1326 *rmap_list
= rmap_item
->rmap_list
;
1327 remove_rmap_item_from_tree(rmap_item
);
1328 free_rmap_item(rmap_item
);
1331 rmap_item
= alloc_rmap_item();
1333 /* It has already been zeroed */
1334 rmap_item
->mm
= mm_slot
->mm
;
1335 rmap_item
->address
= addr
;
1336 rmap_item
->rmap_list
= *rmap_list
;
1337 *rmap_list
= rmap_item
;
1342 static struct rmap_item
*scan_get_next_rmap_item(struct page
**page
)
1344 struct mm_struct
*mm
;
1345 struct mm_slot
*slot
;
1346 struct vm_area_struct
*vma
;
1347 struct rmap_item
*rmap_item
;
1350 if (list_empty(&ksm_mm_head
.mm_list
))
1353 slot
= ksm_scan
.mm_slot
;
1354 if (slot
== &ksm_mm_head
) {
1356 * A number of pages can hang around indefinitely on per-cpu
1357 * pagevecs, raised page count preventing write_protect_page
1358 * from merging them. Though it doesn't really matter much,
1359 * it is puzzling to see some stuck in pages_volatile until
1360 * other activity jostles them out, and they also prevented
1361 * LTP's KSM test from succeeding deterministically; so drain
1362 * them here (here rather than on entry to ksm_do_scan(),
1363 * so we don't IPI too often when pages_to_scan is set low).
1365 lru_add_drain_all();
1367 for (nid
= 0; nid
< nr_node_ids
; nid
++)
1368 root_unstable_tree
[nid
] = RB_ROOT
;
1370 spin_lock(&ksm_mmlist_lock
);
1371 slot
= list_entry(slot
->mm_list
.next
, struct mm_slot
, mm_list
);
1372 ksm_scan
.mm_slot
= slot
;
1373 spin_unlock(&ksm_mmlist_lock
);
1375 * Although we tested list_empty() above, a racing __ksm_exit
1376 * of the last mm on the list may have removed it since then.
1378 if (slot
== &ksm_mm_head
)
1381 ksm_scan
.address
= 0;
1382 ksm_scan
.rmap_list
= &slot
->rmap_list
;
1386 down_read(&mm
->mmap_sem
);
1387 if (ksm_test_exit(mm
))
1390 vma
= find_vma(mm
, ksm_scan
.address
);
1392 for (; vma
; vma
= vma
->vm_next
) {
1393 if (!(vma
->vm_flags
& VM_MERGEABLE
))
1395 if (ksm_scan
.address
< vma
->vm_start
)
1396 ksm_scan
.address
= vma
->vm_start
;
1398 ksm_scan
.address
= vma
->vm_end
;
1400 while (ksm_scan
.address
< vma
->vm_end
) {
1401 if (ksm_test_exit(mm
))
1403 *page
= follow_page(vma
, ksm_scan
.address
, FOLL_GET
);
1404 if (IS_ERR_OR_NULL(*page
)) {
1405 ksm_scan
.address
+= PAGE_SIZE
;
1409 if (PageAnon(*page
) ||
1410 page_trans_compound_anon(*page
)) {
1411 flush_anon_page(vma
, *page
, ksm_scan
.address
);
1412 flush_dcache_page(*page
);
1413 rmap_item
= get_next_rmap_item(slot
,
1414 ksm_scan
.rmap_list
, ksm_scan
.address
);
1416 ksm_scan
.rmap_list
=
1417 &rmap_item
->rmap_list
;
1418 ksm_scan
.address
+= PAGE_SIZE
;
1421 up_read(&mm
->mmap_sem
);
1425 ksm_scan
.address
+= PAGE_SIZE
;
1430 if (ksm_test_exit(mm
)) {
1431 ksm_scan
.address
= 0;
1432 ksm_scan
.rmap_list
= &slot
->rmap_list
;
1435 * Nuke all the rmap_items that are above this current rmap:
1436 * because there were no VM_MERGEABLE vmas with such addresses.
1438 remove_trailing_rmap_items(slot
, ksm_scan
.rmap_list
);
1440 spin_lock(&ksm_mmlist_lock
);
1441 ksm_scan
.mm_slot
= list_entry(slot
->mm_list
.next
,
1442 struct mm_slot
, mm_list
);
1443 if (ksm_scan
.address
== 0) {
1445 * We've completed a full scan of all vmas, holding mmap_sem
1446 * throughout, and found no VM_MERGEABLE: so do the same as
1447 * __ksm_exit does to remove this mm from all our lists now.
1448 * This applies either when cleaning up after __ksm_exit
1449 * (but beware: we can reach here even before __ksm_exit),
1450 * or when all VM_MERGEABLE areas have been unmapped (and
1451 * mmap_sem then protects against race with MADV_MERGEABLE).
1453 hash_del(&slot
->link
);
1454 list_del(&slot
->mm_list
);
1455 spin_unlock(&ksm_mmlist_lock
);
1458 clear_bit(MMF_VM_MERGEABLE
, &mm
->flags
);
1459 up_read(&mm
->mmap_sem
);
1462 spin_unlock(&ksm_mmlist_lock
);
1463 up_read(&mm
->mmap_sem
);
1466 /* Repeat until we've completed scanning the whole list */
1467 slot
= ksm_scan
.mm_slot
;
1468 if (slot
!= &ksm_mm_head
)
1476 * ksm_do_scan - the ksm scanner main worker function.
1477 * @scan_npages - number of pages we want to scan before we return.
1479 static void ksm_do_scan(unsigned int scan_npages
)
1481 struct rmap_item
*rmap_item
;
1482 struct page
*uninitialized_var(page
);
1484 while (scan_npages
-- && likely(!freezing(current
))) {
1486 rmap_item
= scan_get_next_rmap_item(&page
);
1489 if (!PageKsm(page
) || !in_stable_tree(rmap_item
))
1490 cmp_and_merge_page(page
, rmap_item
);
1495 static int ksmd_should_run(void)
1497 return (ksm_run
& KSM_RUN_MERGE
) && !list_empty(&ksm_mm_head
.mm_list
);
1500 static int ksm_scan_thread(void *nothing
)
1503 set_user_nice(current
, 5);
1505 while (!kthread_should_stop()) {
1506 mutex_lock(&ksm_thread_mutex
);
1507 if (ksmd_should_run())
1508 ksm_do_scan(ksm_thread_pages_to_scan
);
1509 mutex_unlock(&ksm_thread_mutex
);
1513 if (ksmd_should_run()) {
1514 schedule_timeout_interruptible(
1515 msecs_to_jiffies(ksm_thread_sleep_millisecs
));
1517 wait_event_freezable(ksm_thread_wait
,
1518 ksmd_should_run() || kthread_should_stop());
1524 int ksm_madvise(struct vm_area_struct
*vma
, unsigned long start
,
1525 unsigned long end
, int advice
, unsigned long *vm_flags
)
1527 struct mm_struct
*mm
= vma
->vm_mm
;
1531 case MADV_MERGEABLE
:
1533 * Be somewhat over-protective for now!
1535 if (*vm_flags
& (VM_MERGEABLE
| VM_SHARED
| VM_MAYSHARE
|
1536 VM_PFNMAP
| VM_IO
| VM_DONTEXPAND
|
1537 VM_HUGETLB
| VM_NONLINEAR
| VM_MIXEDMAP
))
1538 return 0; /* just ignore the advice */
1541 if (*vm_flags
& VM_SAO
)
1545 if (!test_bit(MMF_VM_MERGEABLE
, &mm
->flags
)) {
1546 err
= __ksm_enter(mm
);
1551 *vm_flags
|= VM_MERGEABLE
;
1554 case MADV_UNMERGEABLE
:
1555 if (!(*vm_flags
& VM_MERGEABLE
))
1556 return 0; /* just ignore the advice */
1558 if (vma
->anon_vma
) {
1559 err
= unmerge_ksm_pages(vma
, start
, end
);
1564 *vm_flags
&= ~VM_MERGEABLE
;
1571 int __ksm_enter(struct mm_struct
*mm
)
1573 struct mm_slot
*mm_slot
;
1576 mm_slot
= alloc_mm_slot();
1580 /* Check ksm_run too? Would need tighter locking */
1581 needs_wakeup
= list_empty(&ksm_mm_head
.mm_list
);
1583 spin_lock(&ksm_mmlist_lock
);
1584 insert_to_mm_slots_hash(mm
, mm_slot
);
1586 * Insert just behind the scanning cursor, to let the area settle
1587 * down a little; when fork is followed by immediate exec, we don't
1588 * want ksmd to waste time setting up and tearing down an rmap_list.
1590 list_add_tail(&mm_slot
->mm_list
, &ksm_scan
.mm_slot
->mm_list
);
1591 spin_unlock(&ksm_mmlist_lock
);
1593 set_bit(MMF_VM_MERGEABLE
, &mm
->flags
);
1594 atomic_inc(&mm
->mm_count
);
1597 wake_up_interruptible(&ksm_thread_wait
);
1602 void __ksm_exit(struct mm_struct
*mm
)
1604 struct mm_slot
*mm_slot
;
1605 int easy_to_free
= 0;
1608 * This process is exiting: if it's straightforward (as is the
1609 * case when ksmd was never running), free mm_slot immediately.
1610 * But if it's at the cursor or has rmap_items linked to it, use
1611 * mmap_sem to synchronize with any break_cows before pagetables
1612 * are freed, and leave the mm_slot on the list for ksmd to free.
1613 * Beware: ksm may already have noticed it exiting and freed the slot.
1616 spin_lock(&ksm_mmlist_lock
);
1617 mm_slot
= get_mm_slot(mm
);
1618 if (mm_slot
&& ksm_scan
.mm_slot
!= mm_slot
) {
1619 if (!mm_slot
->rmap_list
) {
1620 hash_del(&mm_slot
->link
);
1621 list_del(&mm_slot
->mm_list
);
1624 list_move(&mm_slot
->mm_list
,
1625 &ksm_scan
.mm_slot
->mm_list
);
1628 spin_unlock(&ksm_mmlist_lock
);
1631 free_mm_slot(mm_slot
);
1632 clear_bit(MMF_VM_MERGEABLE
, &mm
->flags
);
1634 } else if (mm_slot
) {
1635 down_write(&mm
->mmap_sem
);
1636 up_write(&mm
->mmap_sem
);
1640 struct page
*ksm_does_need_to_copy(struct page
*page
,
1641 struct vm_area_struct
*vma
, unsigned long address
)
1643 struct page
*new_page
;
1645 new_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
1647 copy_user_highpage(new_page
, page
, address
, vma
);
1649 SetPageDirty(new_page
);
1650 __SetPageUptodate(new_page
);
1651 __set_page_locked(new_page
);
1657 int page_referenced_ksm(struct page
*page
, struct mem_cgroup
*memcg
,
1658 unsigned long *vm_flags
)
1660 struct stable_node
*stable_node
;
1661 struct rmap_item
*rmap_item
;
1662 struct hlist_node
*hlist
;
1663 unsigned int mapcount
= page_mapcount(page
);
1665 int search_new_forks
= 0;
1667 VM_BUG_ON(!PageKsm(page
));
1668 VM_BUG_ON(!PageLocked(page
));
1670 stable_node
= page_stable_node(page
);
1674 hlist_for_each_entry(rmap_item
, hlist
, &stable_node
->hlist
, hlist
) {
1675 struct anon_vma
*anon_vma
= rmap_item
->anon_vma
;
1676 struct anon_vma_chain
*vmac
;
1677 struct vm_area_struct
*vma
;
1679 anon_vma_lock_read(anon_vma
);
1680 anon_vma_interval_tree_foreach(vmac
, &anon_vma
->rb_root
,
1683 if (rmap_item
->address
< vma
->vm_start
||
1684 rmap_item
->address
>= vma
->vm_end
)
1687 * Initially we examine only the vma which covers this
1688 * rmap_item; but later, if there is still work to do,
1689 * we examine covering vmas in other mms: in case they
1690 * were forked from the original since ksmd passed.
1692 if ((rmap_item
->mm
== vma
->vm_mm
) == search_new_forks
)
1695 if (memcg
&& !mm_match_cgroup(vma
->vm_mm
, memcg
))
1698 referenced
+= page_referenced_one(page
, vma
,
1699 rmap_item
->address
, &mapcount
, vm_flags
);
1700 if (!search_new_forks
|| !mapcount
)
1703 anon_vma_unlock_read(anon_vma
);
1707 if (!search_new_forks
++)
1713 int try_to_unmap_ksm(struct page
*page
, enum ttu_flags flags
)
1715 struct stable_node
*stable_node
;
1716 struct hlist_node
*hlist
;
1717 struct rmap_item
*rmap_item
;
1718 int ret
= SWAP_AGAIN
;
1719 int search_new_forks
= 0;
1721 VM_BUG_ON(!PageKsm(page
));
1722 VM_BUG_ON(!PageLocked(page
));
1724 stable_node
= page_stable_node(page
);
1728 hlist_for_each_entry(rmap_item
, hlist
, &stable_node
->hlist
, hlist
) {
1729 struct anon_vma
*anon_vma
= rmap_item
->anon_vma
;
1730 struct anon_vma_chain
*vmac
;
1731 struct vm_area_struct
*vma
;
1733 anon_vma_lock_read(anon_vma
);
1734 anon_vma_interval_tree_foreach(vmac
, &anon_vma
->rb_root
,
1737 if (rmap_item
->address
< vma
->vm_start
||
1738 rmap_item
->address
>= vma
->vm_end
)
1741 * Initially we examine only the vma which covers this
1742 * rmap_item; but later, if there is still work to do,
1743 * we examine covering vmas in other mms: in case they
1744 * were forked from the original since ksmd passed.
1746 if ((rmap_item
->mm
== vma
->vm_mm
) == search_new_forks
)
1749 ret
= try_to_unmap_one(page
, vma
,
1750 rmap_item
->address
, flags
);
1751 if (ret
!= SWAP_AGAIN
|| !page_mapped(page
)) {
1752 anon_vma_unlock_read(anon_vma
);
1756 anon_vma_unlock_read(anon_vma
);
1758 if (!search_new_forks
++)
1764 #ifdef CONFIG_MIGRATION
1765 int rmap_walk_ksm(struct page
*page
, int (*rmap_one
)(struct page
*,
1766 struct vm_area_struct
*, unsigned long, void *), void *arg
)
1768 struct stable_node
*stable_node
;
1769 struct hlist_node
*hlist
;
1770 struct rmap_item
*rmap_item
;
1771 int ret
= SWAP_AGAIN
;
1772 int search_new_forks
= 0;
1774 VM_BUG_ON(!PageKsm(page
));
1775 VM_BUG_ON(!PageLocked(page
));
1777 stable_node
= page_stable_node(page
);
1781 hlist_for_each_entry(rmap_item
, hlist
, &stable_node
->hlist
, hlist
) {
1782 struct anon_vma
*anon_vma
= rmap_item
->anon_vma
;
1783 struct anon_vma_chain
*vmac
;
1784 struct vm_area_struct
*vma
;
1786 anon_vma_lock_read(anon_vma
);
1787 anon_vma_interval_tree_foreach(vmac
, &anon_vma
->rb_root
,
1790 if (rmap_item
->address
< vma
->vm_start
||
1791 rmap_item
->address
>= vma
->vm_end
)
1794 * Initially we examine only the vma which covers this
1795 * rmap_item; but later, if there is still work to do,
1796 * we examine covering vmas in other mms: in case they
1797 * were forked from the original since ksmd passed.
1799 if ((rmap_item
->mm
== vma
->vm_mm
) == search_new_forks
)
1802 ret
= rmap_one(page
, vma
, rmap_item
->address
, arg
);
1803 if (ret
!= SWAP_AGAIN
) {
1804 anon_vma_unlock_read(anon_vma
);
1808 anon_vma_unlock_read(anon_vma
);
1810 if (!search_new_forks
++)
1816 void ksm_migrate_page(struct page
*newpage
, struct page
*oldpage
)
1818 struct stable_node
*stable_node
;
1820 VM_BUG_ON(!PageLocked(oldpage
));
1821 VM_BUG_ON(!PageLocked(newpage
));
1822 VM_BUG_ON(newpage
->mapping
!= oldpage
->mapping
);
1824 stable_node
= page_stable_node(newpage
);
1826 VM_BUG_ON(stable_node
->kpfn
!= page_to_pfn(oldpage
));
1827 stable_node
->kpfn
= page_to_pfn(newpage
);
1830 #endif /* CONFIG_MIGRATION */
1832 #ifdef CONFIG_MEMORY_HOTREMOVE
1833 static void ksm_check_stable_tree(unsigned long start_pfn
,
1834 unsigned long end_pfn
)
1836 struct stable_node
*stable_node
;
1837 struct rb_node
*node
;
1840 for (nid
= 0; nid
< nr_node_ids
; nid
++) {
1841 node
= rb_first(&root_stable_tree
[nid
]);
1843 stable_node
= rb_entry(node
, struct stable_node
, node
);
1844 if (stable_node
->kpfn
>= start_pfn
&&
1845 stable_node
->kpfn
< end_pfn
) {
1847 * Don't get_ksm_page, page has already gone:
1848 * which is why we keep kpfn instead of page*
1850 remove_node_from_stable_tree(stable_node
);
1851 node
= rb_first(&root_stable_tree
[nid
]);
1853 node
= rb_next(node
);
1859 static int ksm_memory_callback(struct notifier_block
*self
,
1860 unsigned long action
, void *arg
)
1862 struct memory_notify
*mn
= arg
;
1865 case MEM_GOING_OFFLINE
:
1867 * Keep it very simple for now: just lock out ksmd and
1868 * MADV_UNMERGEABLE while any memory is going offline.
1869 * mutex_lock_nested() is necessary because lockdep was alarmed
1870 * that here we take ksm_thread_mutex inside notifier chain
1871 * mutex, and later take notifier chain mutex inside
1872 * ksm_thread_mutex to unlock it. But that's safe because both
1873 * are inside mem_hotplug_mutex.
1875 mutex_lock_nested(&ksm_thread_mutex
, SINGLE_DEPTH_NESTING
);
1880 * Most of the work is done by page migration; but there might
1881 * be a few stable_nodes left over, still pointing to struct
1882 * pages which have been offlined: prune those from the tree,
1883 * otherwise get_ksm_page() might later try to access a
1884 * non-existent struct page.
1886 ksm_check_stable_tree(mn
->start_pfn
,
1887 mn
->start_pfn
+ mn
->nr_pages
);
1890 case MEM_CANCEL_OFFLINE
:
1891 mutex_unlock(&ksm_thread_mutex
);
1896 #endif /* CONFIG_MEMORY_HOTREMOVE */
1900 * This all compiles without CONFIG_SYSFS, but is a waste of space.
1903 #define KSM_ATTR_RO(_name) \
1904 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1905 #define KSM_ATTR(_name) \
1906 static struct kobj_attribute _name##_attr = \
1907 __ATTR(_name, 0644, _name##_show, _name##_store)
1909 static ssize_t
sleep_millisecs_show(struct kobject
*kobj
,
1910 struct kobj_attribute
*attr
, char *buf
)
1912 return sprintf(buf
, "%u\n", ksm_thread_sleep_millisecs
);
1915 static ssize_t
sleep_millisecs_store(struct kobject
*kobj
,
1916 struct kobj_attribute
*attr
,
1917 const char *buf
, size_t count
)
1919 unsigned long msecs
;
1922 err
= strict_strtoul(buf
, 10, &msecs
);
1923 if (err
|| msecs
> UINT_MAX
)
1926 ksm_thread_sleep_millisecs
= msecs
;
1930 KSM_ATTR(sleep_millisecs
);
1932 static ssize_t
pages_to_scan_show(struct kobject
*kobj
,
1933 struct kobj_attribute
*attr
, char *buf
)
1935 return sprintf(buf
, "%u\n", ksm_thread_pages_to_scan
);
1938 static ssize_t
pages_to_scan_store(struct kobject
*kobj
,
1939 struct kobj_attribute
*attr
,
1940 const char *buf
, size_t count
)
1943 unsigned long nr_pages
;
1945 err
= strict_strtoul(buf
, 10, &nr_pages
);
1946 if (err
|| nr_pages
> UINT_MAX
)
1949 ksm_thread_pages_to_scan
= nr_pages
;
1953 KSM_ATTR(pages_to_scan
);
1955 static ssize_t
run_show(struct kobject
*kobj
, struct kobj_attribute
*attr
,
1958 return sprintf(buf
, "%u\n", ksm_run
);
1961 static ssize_t
run_store(struct kobject
*kobj
, struct kobj_attribute
*attr
,
1962 const char *buf
, size_t count
)
1965 unsigned long flags
;
1967 err
= strict_strtoul(buf
, 10, &flags
);
1968 if (err
|| flags
> UINT_MAX
)
1970 if (flags
> KSM_RUN_UNMERGE
)
1974 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
1975 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
1976 * breaking COW to free the pages_shared (but leaves mm_slots
1977 * on the list for when ksmd may be set running again).
1980 mutex_lock(&ksm_thread_mutex
);
1981 if (ksm_run
!= flags
) {
1983 if (flags
& KSM_RUN_UNMERGE
) {
1984 set_current_oom_origin();
1985 err
= unmerge_and_remove_all_rmap_items();
1986 clear_current_oom_origin();
1988 ksm_run
= KSM_RUN_STOP
;
1993 mutex_unlock(&ksm_thread_mutex
);
1995 if (flags
& KSM_RUN_MERGE
)
1996 wake_up_interruptible(&ksm_thread_wait
);
2003 static ssize_t
merge_across_nodes_show(struct kobject
*kobj
,
2004 struct kobj_attribute
*attr
, char *buf
)
2006 return sprintf(buf
, "%u\n", ksm_merge_across_nodes
);
2009 static ssize_t
merge_across_nodes_store(struct kobject
*kobj
,
2010 struct kobj_attribute
*attr
,
2011 const char *buf
, size_t count
)
2016 err
= kstrtoul(buf
, 10, &knob
);
2022 mutex_lock(&ksm_thread_mutex
);
2023 if (ksm_merge_across_nodes
!= knob
) {
2024 if (ksm_pages_shared
)
2027 ksm_merge_across_nodes
= knob
;
2029 mutex_unlock(&ksm_thread_mutex
);
2031 return err
? err
: count
;
2033 KSM_ATTR(merge_across_nodes
);
2036 static ssize_t
pages_shared_show(struct kobject
*kobj
,
2037 struct kobj_attribute
*attr
, char *buf
)
2039 return sprintf(buf
, "%lu\n", ksm_pages_shared
);
2041 KSM_ATTR_RO(pages_shared
);
2043 static ssize_t
pages_sharing_show(struct kobject
*kobj
,
2044 struct kobj_attribute
*attr
, char *buf
)
2046 return sprintf(buf
, "%lu\n", ksm_pages_sharing
);
2048 KSM_ATTR_RO(pages_sharing
);
2050 static ssize_t
pages_unshared_show(struct kobject
*kobj
,
2051 struct kobj_attribute
*attr
, char *buf
)
2053 return sprintf(buf
, "%lu\n", ksm_pages_unshared
);
2055 KSM_ATTR_RO(pages_unshared
);
2057 static ssize_t
pages_volatile_show(struct kobject
*kobj
,
2058 struct kobj_attribute
*attr
, char *buf
)
2060 long ksm_pages_volatile
;
2062 ksm_pages_volatile
= ksm_rmap_items
- ksm_pages_shared
2063 - ksm_pages_sharing
- ksm_pages_unshared
;
2065 * It was not worth any locking to calculate that statistic,
2066 * but it might therefore sometimes be negative: conceal that.
2068 if (ksm_pages_volatile
< 0)
2069 ksm_pages_volatile
= 0;
2070 return sprintf(buf
, "%ld\n", ksm_pages_volatile
);
2072 KSM_ATTR_RO(pages_volatile
);
2074 static ssize_t
full_scans_show(struct kobject
*kobj
,
2075 struct kobj_attribute
*attr
, char *buf
)
2077 return sprintf(buf
, "%lu\n", ksm_scan
.seqnr
);
2079 KSM_ATTR_RO(full_scans
);
2081 static struct attribute
*ksm_attrs
[] = {
2082 &sleep_millisecs_attr
.attr
,
2083 &pages_to_scan_attr
.attr
,
2085 &pages_shared_attr
.attr
,
2086 &pages_sharing_attr
.attr
,
2087 &pages_unshared_attr
.attr
,
2088 &pages_volatile_attr
.attr
,
2089 &full_scans_attr
.attr
,
2091 &merge_across_nodes_attr
.attr
,
2096 static struct attribute_group ksm_attr_group
= {
2100 #endif /* CONFIG_SYSFS */
2102 static int __init
ksm_init(void)
2104 struct task_struct
*ksm_thread
;
2108 err
= ksm_slab_init();
2112 for (nid
= 0; nid
< nr_node_ids
; nid
++)
2113 root_stable_tree
[nid
] = RB_ROOT
;
2115 ksm_thread
= kthread_run(ksm_scan_thread
, NULL
, "ksmd");
2116 if (IS_ERR(ksm_thread
)) {
2117 printk(KERN_ERR
"ksm: creating kthread failed\n");
2118 err
= PTR_ERR(ksm_thread
);
2123 err
= sysfs_create_group(mm_kobj
, &ksm_attr_group
);
2125 printk(KERN_ERR
"ksm: register sysfs failed\n");
2126 kthread_stop(ksm_thread
);
2130 ksm_run
= KSM_RUN_MERGE
; /* no way for user to start it */
2132 #endif /* CONFIG_SYSFS */
2134 #ifdef CONFIG_MEMORY_HOTREMOVE
2136 * Choose a high priority since the callback takes ksm_thread_mutex:
2137 * later callbacks could only be taking locks which nest within that.
2139 hotplug_memory_notifier(ksm_memory_callback
, 100);
2148 module_init(ksm_init
)