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/mmu_notifier.h>
33 #include <linux/ksm.h>
35 #include <asm/tlbflush.h>
38 * A few notes about the KSM scanning process,
39 * to make it easier to understand the data structures below:
41 * In order to reduce excessive scanning, KSM sorts the memory pages by their
42 * contents into a data structure that holds pointers to the pages' locations.
44 * Since the contents of the pages may change at any moment, KSM cannot just
45 * insert the pages into a normal sorted tree and expect it to find anything.
46 * Therefore KSM uses two data structures - the stable and the unstable tree.
48 * The stable tree holds pointers to all the merged pages (ksm pages), sorted
49 * by their contents. Because each such page is write-protected, searching on
50 * this tree is fully assured to be working (except when pages are unmapped),
51 * and therefore this tree is called the stable tree.
53 * In addition to the stable tree, KSM uses a second data structure called the
54 * unstable tree: this tree holds pointers to pages which have been found to
55 * be "unchanged for a period of time". The unstable tree sorts these pages
56 * by their contents, but since they are not write-protected, KSM cannot rely
57 * upon the unstable tree to work correctly - the unstable tree is liable to
58 * be corrupted as its contents are modified, and so it is called unstable.
60 * KSM solves this problem by several techniques:
62 * 1) The unstable tree is flushed every time KSM completes scanning all
63 * memory areas, and then the tree is rebuilt again from the beginning.
64 * 2) KSM will only insert into the unstable tree, pages whose hash value
65 * has not changed since the previous scan of all memory areas.
66 * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
67 * colors of the nodes and not on their contents, assuring that even when
68 * the tree gets "corrupted" it won't get out of balance, so scanning time
69 * remains the same (also, searching and inserting nodes in an rbtree uses
70 * the same algorithm, so we have no overhead when we flush and rebuild).
71 * 4) KSM never flushes the stable tree, which means that even if it were to
72 * take 10 attempts to find a page in the unstable tree, once it is found,
73 * it is secured in the stable tree. (When we scan a new page, we first
74 * compare it against the stable tree, and then against the unstable tree.)
78 * struct mm_slot - ksm information per mm that is being scanned
79 * @link: link to the mm_slots hash list
80 * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
81 * @rmap_list: head for this mm_slot's list of rmap_items
82 * @mm: the mm that this information is valid for
85 struct hlist_node link
;
86 struct list_head mm_list
;
87 struct list_head rmap_list
;
92 * struct ksm_scan - cursor for scanning
93 * @mm_slot: the current mm_slot we are scanning
94 * @address: the next address inside that to be scanned
95 * @rmap_item: the current rmap that we are scanning inside the rmap_list
96 * @seqnr: count of completed full scans (needed when removing unstable node)
98 * There is only the one ksm_scan instance of this cursor structure.
101 struct mm_slot
*mm_slot
;
102 unsigned long address
;
103 struct rmap_item
*rmap_item
;
108 * struct rmap_item - reverse mapping item for virtual addresses
109 * @link: link into mm_slot's rmap_list (rmap_list is per mm)
110 * @mm: the memory structure this rmap_item is pointing into
111 * @address: the virtual address this rmap_item tracks (+ flags in low bits)
112 * @oldchecksum: previous checksum of the page at that virtual address
113 * @node: rb_node of this rmap_item in either unstable or stable tree
114 * @next: next rmap_item hanging off the same node of the stable tree
115 * @prev: previous rmap_item hanging off the same node of the stable tree
118 struct list_head link
;
119 struct mm_struct
*mm
;
120 unsigned long address
; /* + low bits used for flags below */
122 unsigned int oldchecksum
; /* when unstable */
123 struct rmap_item
*next
; /* when stable */
126 struct rb_node node
; /* when tree node */
127 struct rmap_item
*prev
; /* in stable list */
131 #define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */
132 #define NODE_FLAG 0x100 /* is a node of unstable or stable tree */
133 #define STABLE_FLAG 0x200 /* is a node or list item of stable tree */
135 /* The stable and unstable tree heads */
136 static struct rb_root root_stable_tree
= RB_ROOT
;
137 static struct rb_root root_unstable_tree
= RB_ROOT
;
139 #define MM_SLOTS_HASH_HEADS 1024
140 static struct hlist_head
*mm_slots_hash
;
142 static struct mm_slot ksm_mm_head
= {
143 .mm_list
= LIST_HEAD_INIT(ksm_mm_head
.mm_list
),
145 static struct ksm_scan ksm_scan
= {
146 .mm_slot
= &ksm_mm_head
,
149 static struct kmem_cache
*rmap_item_cache
;
150 static struct kmem_cache
*mm_slot_cache
;
152 /* The number of nodes in the stable tree */
153 static unsigned long ksm_pages_shared
;
155 /* The number of page slots additionally sharing those nodes */
156 static unsigned long ksm_pages_sharing
;
158 /* The number of nodes in the unstable tree */
159 static unsigned long ksm_pages_unshared
;
161 /* The number of rmap_items in use: to calculate pages_volatile */
162 static unsigned long ksm_rmap_items
;
164 /* Limit on the number of unswappable pages used */
165 static unsigned long ksm_max_kernel_pages
;
167 /* Number of pages ksmd should scan in one batch */
168 static unsigned int ksm_thread_pages_to_scan
;
170 /* Milliseconds ksmd should sleep between batches */
171 static unsigned int ksm_thread_sleep_millisecs
;
173 #define KSM_RUN_STOP 0
174 #define KSM_RUN_MERGE 1
175 #define KSM_RUN_UNMERGE 2
176 static unsigned int ksm_run
;
178 static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait
);
179 static DEFINE_MUTEX(ksm_thread_mutex
);
180 static DEFINE_SPINLOCK(ksm_mmlist_lock
);
182 #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
183 sizeof(struct __struct), __alignof__(struct __struct),\
186 static int __init
ksm_slab_init(void)
188 rmap_item_cache
= KSM_KMEM_CACHE(rmap_item
, 0);
189 if (!rmap_item_cache
)
192 mm_slot_cache
= KSM_KMEM_CACHE(mm_slot
, 0);
199 kmem_cache_destroy(rmap_item_cache
);
204 static void __init
ksm_slab_free(void)
206 kmem_cache_destroy(mm_slot_cache
);
207 kmem_cache_destroy(rmap_item_cache
);
208 mm_slot_cache
= NULL
;
211 static inline struct rmap_item
*alloc_rmap_item(void)
213 struct rmap_item
*rmap_item
;
215 rmap_item
= kmem_cache_zalloc(rmap_item_cache
, GFP_KERNEL
);
221 static inline void free_rmap_item(struct rmap_item
*rmap_item
)
224 rmap_item
->mm
= NULL
; /* debug safety */
225 kmem_cache_free(rmap_item_cache
, rmap_item
);
228 static inline struct mm_slot
*alloc_mm_slot(void)
230 if (!mm_slot_cache
) /* initialization failed */
232 return kmem_cache_zalloc(mm_slot_cache
, GFP_KERNEL
);
235 static inline void free_mm_slot(struct mm_slot
*mm_slot
)
237 kmem_cache_free(mm_slot_cache
, mm_slot
);
240 static int __init
mm_slots_hash_init(void)
242 mm_slots_hash
= kzalloc(MM_SLOTS_HASH_HEADS
* sizeof(struct hlist_head
),
249 static void __init
mm_slots_hash_free(void)
251 kfree(mm_slots_hash
);
254 static struct mm_slot
*get_mm_slot(struct mm_struct
*mm
)
256 struct mm_slot
*mm_slot
;
257 struct hlist_head
*bucket
;
258 struct hlist_node
*node
;
260 bucket
= &mm_slots_hash
[((unsigned long)mm
/ sizeof(struct mm_struct
))
261 % MM_SLOTS_HASH_HEADS
];
262 hlist_for_each_entry(mm_slot
, node
, bucket
, link
) {
263 if (mm
== mm_slot
->mm
)
269 static void insert_to_mm_slots_hash(struct mm_struct
*mm
,
270 struct mm_slot
*mm_slot
)
272 struct hlist_head
*bucket
;
274 bucket
= &mm_slots_hash
[((unsigned long)mm
/ sizeof(struct mm_struct
))
275 % MM_SLOTS_HASH_HEADS
];
277 INIT_LIST_HEAD(&mm_slot
->rmap_list
);
278 hlist_add_head(&mm_slot
->link
, bucket
);
281 static inline int in_stable_tree(struct rmap_item
*rmap_item
)
283 return rmap_item
->address
& STABLE_FLAG
;
287 * We use break_ksm to break COW on a ksm page: it's a stripped down
289 * if (get_user_pages(current, mm, addr, 1, 1, 1, &page, NULL) == 1)
292 * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
293 * in case the application has unmapped and remapped mm,addr meanwhile.
294 * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP
295 * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
297 static void break_ksm(struct vm_area_struct
*vma
, unsigned long addr
)
304 page
= follow_page(vma
, addr
, FOLL_GET
);
308 ret
= handle_mm_fault(vma
->vm_mm
, vma
, addr
,
311 ret
= VM_FAULT_WRITE
;
313 } while (!(ret
& (VM_FAULT_WRITE
| VM_FAULT_SIGBUS
)));
315 /* Which leaves us looping there if VM_FAULT_OOM: hmmm... */
318 static void break_cow(struct mm_struct
*mm
, unsigned long addr
)
320 struct vm_area_struct
*vma
;
322 down_read(&mm
->mmap_sem
);
323 vma
= find_vma(mm
, addr
);
324 if (!vma
|| vma
->vm_start
> addr
)
326 if (!(vma
->vm_flags
& VM_MERGEABLE
) || !vma
->anon_vma
)
328 break_ksm(vma
, addr
);
330 up_read(&mm
->mmap_sem
);
333 static struct page
*get_mergeable_page(struct rmap_item
*rmap_item
)
335 struct mm_struct
*mm
= rmap_item
->mm
;
336 unsigned long addr
= rmap_item
->address
;
337 struct vm_area_struct
*vma
;
340 down_read(&mm
->mmap_sem
);
341 vma
= find_vma(mm
, addr
);
342 if (!vma
|| vma
->vm_start
> addr
)
344 if (!(vma
->vm_flags
& VM_MERGEABLE
) || !vma
->anon_vma
)
347 page
= follow_page(vma
, addr
, FOLL_GET
);
350 if (PageAnon(page
)) {
351 flush_anon_page(vma
, page
, addr
);
352 flush_dcache_page(page
);
357 up_read(&mm
->mmap_sem
);
362 * get_ksm_page: checks if the page at the virtual address in rmap_item
363 * is still PageKsm, in which case we can trust the content of the page,
364 * and it returns the gotten page; but NULL if the page has been zapped.
366 static struct page
*get_ksm_page(struct rmap_item
*rmap_item
)
370 page
= get_mergeable_page(rmap_item
);
371 if (page
&& !PageKsm(page
)) {
379 * Removing rmap_item from stable or unstable tree.
380 * This function will clean the information from the stable/unstable tree.
382 static void remove_rmap_item_from_tree(struct rmap_item
*rmap_item
)
384 if (in_stable_tree(rmap_item
)) {
385 struct rmap_item
*next_item
= rmap_item
->next
;
387 if (rmap_item
->address
& NODE_FLAG
) {
389 rb_replace_node(&rmap_item
->node
,
392 next_item
->address
|= NODE_FLAG
;
395 rb_erase(&rmap_item
->node
, &root_stable_tree
);
399 struct rmap_item
*prev_item
= rmap_item
->prev
;
401 BUG_ON(prev_item
->next
!= rmap_item
);
402 prev_item
->next
= next_item
;
404 BUG_ON(next_item
->prev
!= rmap_item
);
405 next_item
->prev
= rmap_item
->prev
;
410 rmap_item
->next
= NULL
;
412 } else if (rmap_item
->address
& NODE_FLAG
) {
415 * ksm_thread can and must skip the rb_erase, because
416 * root_unstable_tree was already reset to RB_ROOT.
417 * But __ksm_exit has to be careful: do the rb_erase
418 * if it's interrupting a scan, and this rmap_item was
419 * inserted by this scan rather than left from before.
421 * Because of the case in which remove_mm_from_lists
422 * increments seqnr before removing rmaps, unstable_nr
423 * may even be 2 behind seqnr, but should never be
424 * further behind. Yes, I did have trouble with this!
426 age
= (unsigned char)(ksm_scan
.seqnr
- rmap_item
->address
);
429 rb_erase(&rmap_item
->node
, &root_unstable_tree
);
430 ksm_pages_unshared
--;
433 rmap_item
->address
&= PAGE_MASK
;
435 cond_resched(); /* we're called from many long loops */
438 static void remove_trailing_rmap_items(struct mm_slot
*mm_slot
,
439 struct list_head
*cur
)
441 struct rmap_item
*rmap_item
;
443 while (cur
!= &mm_slot
->rmap_list
) {
444 rmap_item
= list_entry(cur
, struct rmap_item
, link
);
446 remove_rmap_item_from_tree(rmap_item
);
447 list_del(&rmap_item
->link
);
448 free_rmap_item(rmap_item
);
453 * Though it's very tempting to unmerge in_stable_tree(rmap_item)s rather
454 * than check every pte of a given vma, the locking doesn't quite work for
455 * that - an rmap_item is assigned to the stable tree after inserting ksm
456 * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing
457 * rmap_items from parent to child at fork time (so as not to waste time
458 * if exit comes before the next scan reaches it).
460 * Similarly, although we'd like to remove rmap_items (so updating counts
461 * and freeing memory) when unmerging an area, it's easier to leave that
462 * to the next pass of ksmd - consider, for example, how ksmd might be
463 * in cmp_and_merge_page on one of the rmap_items we would be removing.
465 static void unmerge_ksm_pages(struct vm_area_struct
*vma
,
466 unsigned long start
, unsigned long end
)
470 for (addr
= start
; addr
< end
; addr
+= PAGE_SIZE
)
471 break_ksm(vma
, addr
);
474 static void unmerge_and_remove_all_rmap_items(void)
476 struct mm_slot
*mm_slot
;
477 struct mm_struct
*mm
;
478 struct vm_area_struct
*vma
;
480 list_for_each_entry(mm_slot
, &ksm_mm_head
.mm_list
, mm_list
) {
482 down_read(&mm
->mmap_sem
);
483 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
484 if (!(vma
->vm_flags
& VM_MERGEABLE
) || !vma
->anon_vma
)
486 unmerge_ksm_pages(vma
, vma
->vm_start
, vma
->vm_end
);
488 remove_trailing_rmap_items(mm_slot
, mm_slot
->rmap_list
.next
);
489 up_read(&mm
->mmap_sem
);
492 spin_lock(&ksm_mmlist_lock
);
493 if (ksm_scan
.mm_slot
!= &ksm_mm_head
) {
494 ksm_scan
.mm_slot
= &ksm_mm_head
;
497 spin_unlock(&ksm_mmlist_lock
);
500 static void remove_mm_from_lists(struct mm_struct
*mm
)
502 struct mm_slot
*mm_slot
;
504 spin_lock(&ksm_mmlist_lock
);
505 mm_slot
= get_mm_slot(mm
);
508 * This mm_slot is always at the scanning cursor when we're
509 * called from scan_get_next_rmap_item; but it's a special
510 * case when we're called from __ksm_exit.
512 if (ksm_scan
.mm_slot
== mm_slot
) {
513 ksm_scan
.mm_slot
= list_entry(
514 mm_slot
->mm_list
.next
, struct mm_slot
, mm_list
);
515 ksm_scan
.address
= 0;
516 ksm_scan
.rmap_item
= list_entry(
517 &ksm_scan
.mm_slot
->rmap_list
, struct rmap_item
, link
);
518 if (ksm_scan
.mm_slot
== &ksm_mm_head
)
522 hlist_del(&mm_slot
->link
);
523 list_del(&mm_slot
->mm_list
);
524 spin_unlock(&ksm_mmlist_lock
);
526 remove_trailing_rmap_items(mm_slot
, mm_slot
->rmap_list
.next
);
527 free_mm_slot(mm_slot
);
528 clear_bit(MMF_VM_MERGEABLE
, &mm
->flags
);
531 static u32
calc_checksum(struct page
*page
)
534 void *addr
= kmap_atomic(page
, KM_USER0
);
535 checksum
= jhash2(addr
, PAGE_SIZE
/ 4, 17);
536 kunmap_atomic(addr
, KM_USER0
);
540 static int memcmp_pages(struct page
*page1
, struct page
*page2
)
545 addr1
= kmap_atomic(page1
, KM_USER0
);
546 addr2
= kmap_atomic(page2
, KM_USER1
);
547 ret
= memcmp(addr1
, addr2
, PAGE_SIZE
);
548 kunmap_atomic(addr2
, KM_USER1
);
549 kunmap_atomic(addr1
, KM_USER0
);
553 static inline int pages_identical(struct page
*page1
, struct page
*page2
)
555 return !memcmp_pages(page1
, page2
);
558 static int write_protect_page(struct vm_area_struct
*vma
, struct page
*page
,
561 struct mm_struct
*mm
= vma
->vm_mm
;
568 addr
= page_address_in_vma(page
, vma
);
572 ptep
= page_check_address(page
, mm
, addr
, &ptl
, 0);
576 if (pte_write(*ptep
)) {
579 swapped
= PageSwapCache(page
);
580 flush_cache_page(vma
, addr
, page_to_pfn(page
));
582 * Ok this is tricky, when get_user_pages_fast() run it doesnt
583 * take any lock, therefore the check that we are going to make
584 * with the pagecount against the mapcount is racey and
585 * O_DIRECT can happen right after the check.
586 * So we clear the pte and flush the tlb before the check
587 * this assure us that no O_DIRECT can happen after the check
588 * or in the middle of the check.
590 entry
= ptep_clear_flush(vma
, addr
, ptep
);
592 * Check that no O_DIRECT or similar I/O is in progress on the
595 if ((page_mapcount(page
) + 2 + swapped
) != page_count(page
)) {
596 set_pte_at_notify(mm
, addr
, ptep
, entry
);
599 entry
= pte_wrprotect(entry
);
600 set_pte_at_notify(mm
, addr
, ptep
, entry
);
606 pte_unmap_unlock(ptep
, ptl
);
612 * replace_page - replace page in vma by new ksm page
613 * @vma: vma that holds the pte pointing to oldpage
614 * @oldpage: the page we are replacing by newpage
615 * @newpage: the ksm page we replace oldpage by
616 * @orig_pte: the original value of the pte
618 * Returns 0 on success, -EFAULT on failure.
620 static int replace_page(struct vm_area_struct
*vma
, struct page
*oldpage
,
621 struct page
*newpage
, pte_t orig_pte
)
623 struct mm_struct
*mm
= vma
->vm_mm
;
633 prot
= vm_get_page_prot(vma
->vm_flags
& ~VM_WRITE
);
635 addr
= page_address_in_vma(oldpage
, vma
);
639 pgd
= pgd_offset(mm
, addr
);
640 if (!pgd_present(*pgd
))
643 pud
= pud_offset(pgd
, addr
);
644 if (!pud_present(*pud
))
647 pmd
= pmd_offset(pud
, addr
);
648 if (!pmd_present(*pmd
))
651 ptep
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
652 if (!pte_same(*ptep
, orig_pte
)) {
653 pte_unmap_unlock(ptep
, ptl
);
658 page_add_ksm_rmap(newpage
);
660 flush_cache_page(vma
, addr
, pte_pfn(*ptep
));
661 ptep_clear_flush(vma
, addr
, ptep
);
662 set_pte_at_notify(mm
, addr
, ptep
, mk_pte(newpage
, prot
));
664 page_remove_rmap(oldpage
);
667 pte_unmap_unlock(ptep
, ptl
);
674 * try_to_merge_one_page - take two pages and merge them into one
675 * @vma: the vma that hold the pte pointing into oldpage
676 * @oldpage: the page that we want to replace with newpage
677 * @newpage: the page that we want to map instead of oldpage
680 * oldpage should be a PageAnon page, while newpage should be a PageKsm page,
681 * or a newly allocated kernel page which page_add_ksm_rmap will make PageKsm.
683 * This function returns 0 if the pages were merged, -EFAULT otherwise.
685 static int try_to_merge_one_page(struct vm_area_struct
*vma
,
686 struct page
*oldpage
,
687 struct page
*newpage
)
689 pte_t orig_pte
= __pte(0);
692 if (!(vma
->vm_flags
& VM_MERGEABLE
))
695 if (!PageAnon(oldpage
))
702 * We need the page lock to read a stable PageSwapCache in
703 * write_protect_page(). We use trylock_page() instead of
704 * lock_page() because we don't want to wait here - we
705 * prefer to continue scanning and merging different pages,
706 * then come back to this page when it is unlocked.
708 if (!trylock_page(oldpage
))
711 * If this anonymous page is mapped only here, its pte may need
712 * to be write-protected. If it's mapped elsewhere, all of its
713 * ptes are necessarily already write-protected. But in either
714 * case, we need to lock and check page_count is not raised.
716 if (write_protect_page(vma
, oldpage
, &orig_pte
)) {
717 unlock_page(oldpage
);
720 unlock_page(oldpage
);
722 if (pages_identical(oldpage
, newpage
))
723 err
= replace_page(vma
, oldpage
, newpage
, orig_pte
);
733 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
734 * but no new kernel page is allocated: kpage must already be a ksm page.
736 static int try_to_merge_with_ksm_page(struct mm_struct
*mm1
,
741 struct vm_area_struct
*vma
;
744 down_read(&mm1
->mmap_sem
);
745 vma
= find_vma(mm1
, addr1
);
746 if (!vma
|| vma
->vm_start
> addr1
)
749 err
= try_to_merge_one_page(vma
, page1
, kpage
);
751 up_read(&mm1
->mmap_sem
);
756 * try_to_merge_two_pages - take two identical pages and prepare them
757 * to be merged into one page.
759 * This function returns 0 if we successfully mapped two identical pages
760 * into one page, -EFAULT otherwise.
762 * Note that this function allocates a new kernel page: if one of the pages
763 * is already a ksm page, try_to_merge_with_ksm_page should be used.
765 static int try_to_merge_two_pages(struct mm_struct
*mm1
, unsigned long addr1
,
766 struct page
*page1
, struct mm_struct
*mm2
,
767 unsigned long addr2
, struct page
*page2
)
769 struct vm_area_struct
*vma
;
774 * The number of nodes in the stable tree
775 * is the number of kernel pages that we hold.
777 if (ksm_max_kernel_pages
&&
778 ksm_max_kernel_pages
<= ksm_pages_shared
)
781 kpage
= alloc_page(GFP_HIGHUSER
);
785 down_read(&mm1
->mmap_sem
);
786 vma
= find_vma(mm1
, addr1
);
787 if (!vma
|| vma
->vm_start
> addr1
) {
788 up_read(&mm1
->mmap_sem
);
792 copy_user_highpage(kpage
, page1
, addr1
, vma
);
793 err
= try_to_merge_one_page(vma
, page1
, kpage
);
794 up_read(&mm1
->mmap_sem
);
797 err
= try_to_merge_with_ksm_page(mm2
, addr2
, page2
, kpage
);
799 * If that fails, we have a ksm page with only one pte
800 * pointing to it: so break it.
803 break_cow(mm1
, addr1
);
811 * stable_tree_search - search page inside the stable tree
812 * @page: the page that we are searching identical pages to.
813 * @page2: pointer into identical page that we are holding inside the stable
814 * tree that we have found.
815 * @rmap_item: the reverse mapping item
817 * This function checks if there is a page inside the stable tree
818 * with identical content to the page that we are scanning right now.
820 * This function return rmap_item pointer to the identical item if found,
823 static struct rmap_item
*stable_tree_search(struct page
*page
,
825 struct rmap_item
*rmap_item
)
827 struct rb_node
*node
= root_stable_tree
.rb_node
;
830 struct rmap_item
*tree_rmap_item
, *next_rmap_item
;
833 tree_rmap_item
= rb_entry(node
, struct rmap_item
, node
);
834 while (tree_rmap_item
) {
835 BUG_ON(!in_stable_tree(tree_rmap_item
));
837 page2
[0] = get_ksm_page(tree_rmap_item
);
840 next_rmap_item
= tree_rmap_item
->next
;
841 remove_rmap_item_from_tree(tree_rmap_item
);
842 tree_rmap_item
= next_rmap_item
;
847 ret
= memcmp_pages(page
, page2
[0]);
851 node
= node
->rb_left
;
852 } else if (ret
> 0) {
854 node
= node
->rb_right
;
856 return tree_rmap_item
;
864 * stable_tree_insert - insert rmap_item pointing to new ksm page
865 * into the stable tree.
867 * @page: the page that we are searching identical page to inside the stable
869 * @rmap_item: pointer to the reverse mapping item.
871 * This function returns rmap_item if success, NULL otherwise.
873 static struct rmap_item
*stable_tree_insert(struct page
*page
,
874 struct rmap_item
*rmap_item
)
876 struct rb_node
**new = &root_stable_tree
.rb_node
;
877 struct rb_node
*parent
= NULL
;
880 struct rmap_item
*tree_rmap_item
, *next_rmap_item
;
881 struct page
*tree_page
;
884 tree_rmap_item
= rb_entry(*new, struct rmap_item
, node
);
885 while (tree_rmap_item
) {
886 BUG_ON(!in_stable_tree(tree_rmap_item
));
888 tree_page
= get_ksm_page(tree_rmap_item
);
891 next_rmap_item
= tree_rmap_item
->next
;
892 remove_rmap_item_from_tree(tree_rmap_item
);
893 tree_rmap_item
= next_rmap_item
;
898 ret
= memcmp_pages(page
, tree_page
);
903 new = &parent
->rb_left
;
905 new = &parent
->rb_right
;
908 * It is not a bug that stable_tree_search() didn't
909 * find this node: because at that time our page was
910 * not yet write-protected, so may have changed since.
916 rmap_item
->address
|= NODE_FLAG
| STABLE_FLAG
;
917 rmap_item
->next
= NULL
;
918 rb_link_node(&rmap_item
->node
, parent
, new);
919 rb_insert_color(&rmap_item
->node
, &root_stable_tree
);
926 * unstable_tree_search_insert - search and insert items into the unstable tree.
928 * @page: the page that we are going to search for identical page or to insert
929 * into the unstable tree
930 * @page2: pointer into identical page that was found inside the unstable tree
931 * @rmap_item: the reverse mapping item of page
933 * This function searches for a page in the unstable tree identical to the
934 * page currently being scanned; and if no identical page is found in the
935 * tree, we insert rmap_item as a new object into the unstable tree.
937 * This function returns pointer to rmap_item found to be identical
938 * to the currently scanned page, NULL otherwise.
940 * This function does both searching and inserting, because they share
941 * the same walking algorithm in an rbtree.
943 static struct rmap_item
*unstable_tree_search_insert(struct page
*page
,
945 struct rmap_item
*rmap_item
)
947 struct rb_node
**new = &root_unstable_tree
.rb_node
;
948 struct rb_node
*parent
= NULL
;
951 struct rmap_item
*tree_rmap_item
;
954 tree_rmap_item
= rb_entry(*new, struct rmap_item
, node
);
955 page2
[0] = get_mergeable_page(tree_rmap_item
);
960 * Don't substitute an unswappable ksm page
961 * just for one good swappable forked page.
963 if (page
== page2
[0]) {
968 ret
= memcmp_pages(page
, page2
[0]);
973 new = &parent
->rb_left
;
974 } else if (ret
> 0) {
976 new = &parent
->rb_right
;
978 return tree_rmap_item
;
982 rmap_item
->address
|= NODE_FLAG
;
983 rmap_item
->address
|= (ksm_scan
.seqnr
& SEQNR_MASK
);
984 rb_link_node(&rmap_item
->node
, parent
, new);
985 rb_insert_color(&rmap_item
->node
, &root_unstable_tree
);
987 ksm_pages_unshared
++;
992 * stable_tree_append - add another rmap_item to the linked list of
993 * rmap_items hanging off a given node of the stable tree, all sharing
996 static void stable_tree_append(struct rmap_item
*rmap_item
,
997 struct rmap_item
*tree_rmap_item
)
999 rmap_item
->next
= tree_rmap_item
->next
;
1000 rmap_item
->prev
= tree_rmap_item
;
1002 if (tree_rmap_item
->next
)
1003 tree_rmap_item
->next
->prev
= rmap_item
;
1005 tree_rmap_item
->next
= rmap_item
;
1006 rmap_item
->address
|= STABLE_FLAG
;
1008 ksm_pages_sharing
++;
1012 * cmp_and_merge_page - first see if page can be merged into the stable tree;
1013 * if not, compare checksum to previous and if it's the same, see if page can
1014 * be inserted into the unstable tree, or merged with a page already there and
1015 * both transferred to the stable tree.
1017 * @page: the page that we are searching identical page to.
1018 * @rmap_item: the reverse mapping into the virtual address of this page
1020 static void cmp_and_merge_page(struct page
*page
, struct rmap_item
*rmap_item
)
1022 struct page
*page2
[1];
1023 struct rmap_item
*tree_rmap_item
;
1024 unsigned int checksum
;
1027 if (in_stable_tree(rmap_item
))
1028 remove_rmap_item_from_tree(rmap_item
);
1030 /* We first start with searching the page inside the stable tree */
1031 tree_rmap_item
= stable_tree_search(page
, page2
, rmap_item
);
1032 if (tree_rmap_item
) {
1033 if (page
== page2
[0]) /* forked */
1036 err
= try_to_merge_with_ksm_page(rmap_item
->mm
,
1043 * The page was successfully merged:
1044 * add its rmap_item to the stable tree.
1046 stable_tree_append(rmap_item
, tree_rmap_item
);
1052 * A ksm page might have got here by fork, but its other
1053 * references have already been removed from the stable tree.
1056 break_cow(rmap_item
->mm
, rmap_item
->address
);
1059 * In case the hash value of the page was changed from the last time we
1060 * have calculated it, this page to be changed frequely, therefore we
1061 * don't want to insert it to the unstable tree, and we don't want to
1062 * waste our time to search if there is something identical to it there.
1064 checksum
= calc_checksum(page
);
1065 if (rmap_item
->oldchecksum
!= checksum
) {
1066 rmap_item
->oldchecksum
= checksum
;
1070 tree_rmap_item
= unstable_tree_search_insert(page
, page2
, rmap_item
);
1071 if (tree_rmap_item
) {
1072 err
= try_to_merge_two_pages(rmap_item
->mm
,
1073 rmap_item
->address
, page
,
1075 tree_rmap_item
->address
, page2
[0]);
1077 * As soon as we merge this page, we want to remove the
1078 * rmap_item of the page we have merged with from the unstable
1079 * tree, and insert it instead as new node in the stable tree.
1082 rb_erase(&tree_rmap_item
->node
, &root_unstable_tree
);
1083 tree_rmap_item
->address
&= ~NODE_FLAG
;
1084 ksm_pages_unshared
--;
1087 * If we fail to insert the page into the stable tree,
1088 * we will have 2 virtual addresses that are pointing
1089 * to a ksm page left outside the stable tree,
1090 * in which case we need to break_cow on both.
1092 if (stable_tree_insert(page2
[0], tree_rmap_item
))
1093 stable_tree_append(rmap_item
, tree_rmap_item
);
1095 break_cow(tree_rmap_item
->mm
,
1096 tree_rmap_item
->address
);
1097 break_cow(rmap_item
->mm
, rmap_item
->address
);
1105 static struct rmap_item
*get_next_rmap_item(struct mm_slot
*mm_slot
,
1106 struct list_head
*cur
,
1109 struct rmap_item
*rmap_item
;
1111 while (cur
!= &mm_slot
->rmap_list
) {
1112 rmap_item
= list_entry(cur
, struct rmap_item
, link
);
1113 if ((rmap_item
->address
& PAGE_MASK
) == addr
) {
1114 if (!in_stable_tree(rmap_item
))
1115 remove_rmap_item_from_tree(rmap_item
);
1118 if (rmap_item
->address
> addr
)
1121 remove_rmap_item_from_tree(rmap_item
);
1122 list_del(&rmap_item
->link
);
1123 free_rmap_item(rmap_item
);
1126 rmap_item
= alloc_rmap_item();
1128 /* It has already been zeroed */
1129 rmap_item
->mm
= mm_slot
->mm
;
1130 rmap_item
->address
= addr
;
1131 list_add_tail(&rmap_item
->link
, cur
);
1136 static struct rmap_item
*scan_get_next_rmap_item(struct page
**page
)
1138 struct mm_struct
*mm
;
1139 struct mm_slot
*slot
;
1140 struct vm_area_struct
*vma
;
1141 struct rmap_item
*rmap_item
;
1143 if (list_empty(&ksm_mm_head
.mm_list
))
1146 slot
= ksm_scan
.mm_slot
;
1147 if (slot
== &ksm_mm_head
) {
1148 root_unstable_tree
= RB_ROOT
;
1150 spin_lock(&ksm_mmlist_lock
);
1151 slot
= list_entry(slot
->mm_list
.next
, struct mm_slot
, mm_list
);
1152 ksm_scan
.mm_slot
= slot
;
1153 spin_unlock(&ksm_mmlist_lock
);
1155 ksm_scan
.address
= 0;
1156 ksm_scan
.rmap_item
= list_entry(&slot
->rmap_list
,
1157 struct rmap_item
, link
);
1161 down_read(&mm
->mmap_sem
);
1162 for (vma
= find_vma(mm
, ksm_scan
.address
); vma
; vma
= vma
->vm_next
) {
1163 if (!(vma
->vm_flags
& VM_MERGEABLE
))
1165 if (ksm_scan
.address
< vma
->vm_start
)
1166 ksm_scan
.address
= vma
->vm_start
;
1168 ksm_scan
.address
= vma
->vm_end
;
1170 while (ksm_scan
.address
< vma
->vm_end
) {
1171 *page
= follow_page(vma
, ksm_scan
.address
, FOLL_GET
);
1172 if (*page
&& PageAnon(*page
)) {
1173 flush_anon_page(vma
, *page
, ksm_scan
.address
);
1174 flush_dcache_page(*page
);
1175 rmap_item
= get_next_rmap_item(slot
,
1176 ksm_scan
.rmap_item
->link
.next
,
1179 ksm_scan
.rmap_item
= rmap_item
;
1180 ksm_scan
.address
+= PAGE_SIZE
;
1183 up_read(&mm
->mmap_sem
);
1188 ksm_scan
.address
+= PAGE_SIZE
;
1193 if (!ksm_scan
.address
) {
1195 * We've completed a full scan of all vmas, holding mmap_sem
1196 * throughout, and found no VM_MERGEABLE: so do the same as
1197 * __ksm_exit does to remove this mm from all our lists now.
1199 remove_mm_from_lists(mm
);
1200 up_read(&mm
->mmap_sem
);
1201 slot
= ksm_scan
.mm_slot
;
1202 if (slot
!= &ksm_mm_head
)
1208 * Nuke all the rmap_items that are above this current rmap:
1209 * because there were no VM_MERGEABLE vmas with such addresses.
1211 remove_trailing_rmap_items(slot
, ksm_scan
.rmap_item
->link
.next
);
1212 up_read(&mm
->mmap_sem
);
1214 spin_lock(&ksm_mmlist_lock
);
1215 slot
= list_entry(slot
->mm_list
.next
, struct mm_slot
, mm_list
);
1216 ksm_scan
.mm_slot
= slot
;
1217 spin_unlock(&ksm_mmlist_lock
);
1219 /* Repeat until we've completed scanning the whole list */
1220 if (slot
!= &ksm_mm_head
)
1224 * Bump seqnr here rather than at top, so that __ksm_exit
1225 * can skip rb_erase on unstable tree until we run again.
1232 * ksm_do_scan - the ksm scanner main worker function.
1233 * @scan_npages - number of pages we want to scan before we return.
1235 static void ksm_do_scan(unsigned int scan_npages
)
1237 struct rmap_item
*rmap_item
;
1240 while (scan_npages
--) {
1242 rmap_item
= scan_get_next_rmap_item(&page
);
1245 if (!PageKsm(page
) || !in_stable_tree(rmap_item
))
1246 cmp_and_merge_page(page
, rmap_item
);
1247 else if (page_mapcount(page
) == 1) {
1249 * Replace now-unshared ksm page by ordinary page.
1251 break_cow(rmap_item
->mm
, rmap_item
->address
);
1252 remove_rmap_item_from_tree(rmap_item
);
1253 rmap_item
->oldchecksum
= calc_checksum(page
);
1259 static int ksmd_should_run(void)
1261 return (ksm_run
& KSM_RUN_MERGE
) && !list_empty(&ksm_mm_head
.mm_list
);
1264 static int ksm_scan_thread(void *nothing
)
1266 set_user_nice(current
, 5);
1268 while (!kthread_should_stop()) {
1269 mutex_lock(&ksm_thread_mutex
);
1270 if (ksmd_should_run())
1271 ksm_do_scan(ksm_thread_pages_to_scan
);
1272 mutex_unlock(&ksm_thread_mutex
);
1274 if (ksmd_should_run()) {
1275 schedule_timeout_interruptible(
1276 msecs_to_jiffies(ksm_thread_sleep_millisecs
));
1278 wait_event_interruptible(ksm_thread_wait
,
1279 ksmd_should_run() || kthread_should_stop());
1285 int ksm_madvise(struct vm_area_struct
*vma
, unsigned long start
,
1286 unsigned long end
, int advice
, unsigned long *vm_flags
)
1288 struct mm_struct
*mm
= vma
->vm_mm
;
1291 case MADV_MERGEABLE
:
1293 * Be somewhat over-protective for now!
1295 if (*vm_flags
& (VM_MERGEABLE
| VM_SHARED
| VM_MAYSHARE
|
1296 VM_PFNMAP
| VM_IO
| VM_DONTEXPAND
|
1297 VM_RESERVED
| VM_HUGETLB
| VM_INSERTPAGE
|
1298 VM_MIXEDMAP
| VM_SAO
))
1299 return 0; /* just ignore the advice */
1301 if (!test_bit(MMF_VM_MERGEABLE
, &mm
->flags
))
1302 if (__ksm_enter(mm
) < 0)
1305 *vm_flags
|= VM_MERGEABLE
;
1308 case MADV_UNMERGEABLE
:
1309 if (!(*vm_flags
& VM_MERGEABLE
))
1310 return 0; /* just ignore the advice */
1313 unmerge_ksm_pages(vma
, start
, end
);
1315 *vm_flags
&= ~VM_MERGEABLE
;
1322 int __ksm_enter(struct mm_struct
*mm
)
1324 struct mm_slot
*mm_slot
;
1327 mm_slot
= alloc_mm_slot();
1331 /* Check ksm_run too? Would need tighter locking */
1332 needs_wakeup
= list_empty(&ksm_mm_head
.mm_list
);
1334 spin_lock(&ksm_mmlist_lock
);
1335 insert_to_mm_slots_hash(mm
, mm_slot
);
1337 * Insert just behind the scanning cursor, to let the area settle
1338 * down a little; when fork is followed by immediate exec, we don't
1339 * want ksmd to waste time setting up and tearing down an rmap_list.
1341 list_add_tail(&mm_slot
->mm_list
, &ksm_scan
.mm_slot
->mm_list
);
1342 spin_unlock(&ksm_mmlist_lock
);
1344 set_bit(MMF_VM_MERGEABLE
, &mm
->flags
);
1347 wake_up_interruptible(&ksm_thread_wait
);
1352 void __ksm_exit(struct mm_struct
*mm
)
1355 * This process is exiting: doesn't hold and doesn't need mmap_sem;
1356 * but we do need to exclude ksmd and other exiters while we modify
1357 * the various lists and trees.
1359 mutex_lock(&ksm_thread_mutex
);
1360 remove_mm_from_lists(mm
);
1361 mutex_unlock(&ksm_thread_mutex
);
1364 #define KSM_ATTR_RO(_name) \
1365 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1366 #define KSM_ATTR(_name) \
1367 static struct kobj_attribute _name##_attr = \
1368 __ATTR(_name, 0644, _name##_show, _name##_store)
1370 static ssize_t
sleep_millisecs_show(struct kobject
*kobj
,
1371 struct kobj_attribute
*attr
, char *buf
)
1373 return sprintf(buf
, "%u\n", ksm_thread_sleep_millisecs
);
1376 static ssize_t
sleep_millisecs_store(struct kobject
*kobj
,
1377 struct kobj_attribute
*attr
,
1378 const char *buf
, size_t count
)
1380 unsigned long msecs
;
1383 err
= strict_strtoul(buf
, 10, &msecs
);
1384 if (err
|| msecs
> UINT_MAX
)
1387 ksm_thread_sleep_millisecs
= msecs
;
1391 KSM_ATTR(sleep_millisecs
);
1393 static ssize_t
pages_to_scan_show(struct kobject
*kobj
,
1394 struct kobj_attribute
*attr
, char *buf
)
1396 return sprintf(buf
, "%u\n", ksm_thread_pages_to_scan
);
1399 static ssize_t
pages_to_scan_store(struct kobject
*kobj
,
1400 struct kobj_attribute
*attr
,
1401 const char *buf
, size_t count
)
1404 unsigned long nr_pages
;
1406 err
= strict_strtoul(buf
, 10, &nr_pages
);
1407 if (err
|| nr_pages
> UINT_MAX
)
1410 ksm_thread_pages_to_scan
= nr_pages
;
1414 KSM_ATTR(pages_to_scan
);
1416 static ssize_t
run_show(struct kobject
*kobj
, struct kobj_attribute
*attr
,
1419 return sprintf(buf
, "%u\n", ksm_run
);
1422 static ssize_t
run_store(struct kobject
*kobj
, struct kobj_attribute
*attr
,
1423 const char *buf
, size_t count
)
1426 unsigned long flags
;
1428 err
= strict_strtoul(buf
, 10, &flags
);
1429 if (err
|| flags
> UINT_MAX
)
1431 if (flags
> KSM_RUN_UNMERGE
)
1435 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
1436 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
1437 * breaking COW to free the unswappable pages_shared (but leaves
1438 * mm_slots on the list for when ksmd may be set running again).
1441 mutex_lock(&ksm_thread_mutex
);
1442 if (ksm_run
!= flags
) {
1444 if (flags
& KSM_RUN_UNMERGE
)
1445 unmerge_and_remove_all_rmap_items();
1447 mutex_unlock(&ksm_thread_mutex
);
1449 if (flags
& KSM_RUN_MERGE
)
1450 wake_up_interruptible(&ksm_thread_wait
);
1456 static ssize_t
max_kernel_pages_store(struct kobject
*kobj
,
1457 struct kobj_attribute
*attr
,
1458 const char *buf
, size_t count
)
1461 unsigned long nr_pages
;
1463 err
= strict_strtoul(buf
, 10, &nr_pages
);
1467 ksm_max_kernel_pages
= nr_pages
;
1472 static ssize_t
max_kernel_pages_show(struct kobject
*kobj
,
1473 struct kobj_attribute
*attr
, char *buf
)
1475 return sprintf(buf
, "%lu\n", ksm_max_kernel_pages
);
1477 KSM_ATTR(max_kernel_pages
);
1479 static ssize_t
pages_shared_show(struct kobject
*kobj
,
1480 struct kobj_attribute
*attr
, char *buf
)
1482 return sprintf(buf
, "%lu\n", ksm_pages_shared
);
1484 KSM_ATTR_RO(pages_shared
);
1486 static ssize_t
pages_sharing_show(struct kobject
*kobj
,
1487 struct kobj_attribute
*attr
, char *buf
)
1489 return sprintf(buf
, "%lu\n", ksm_pages_sharing
);
1491 KSM_ATTR_RO(pages_sharing
);
1493 static ssize_t
pages_unshared_show(struct kobject
*kobj
,
1494 struct kobj_attribute
*attr
, char *buf
)
1496 return sprintf(buf
, "%lu\n", ksm_pages_unshared
);
1498 KSM_ATTR_RO(pages_unshared
);
1500 static ssize_t
pages_volatile_show(struct kobject
*kobj
,
1501 struct kobj_attribute
*attr
, char *buf
)
1503 long ksm_pages_volatile
;
1505 ksm_pages_volatile
= ksm_rmap_items
- ksm_pages_shared
1506 - ksm_pages_sharing
- ksm_pages_unshared
;
1508 * It was not worth any locking to calculate that statistic,
1509 * but it might therefore sometimes be negative: conceal that.
1511 if (ksm_pages_volatile
< 0)
1512 ksm_pages_volatile
= 0;
1513 return sprintf(buf
, "%ld\n", ksm_pages_volatile
);
1515 KSM_ATTR_RO(pages_volatile
);
1517 static ssize_t
full_scans_show(struct kobject
*kobj
,
1518 struct kobj_attribute
*attr
, char *buf
)
1520 return sprintf(buf
, "%lu\n", ksm_scan
.seqnr
);
1522 KSM_ATTR_RO(full_scans
);
1524 static struct attribute
*ksm_attrs
[] = {
1525 &sleep_millisecs_attr
.attr
,
1526 &pages_to_scan_attr
.attr
,
1528 &max_kernel_pages_attr
.attr
,
1529 &pages_shared_attr
.attr
,
1530 &pages_sharing_attr
.attr
,
1531 &pages_unshared_attr
.attr
,
1532 &pages_volatile_attr
.attr
,
1533 &full_scans_attr
.attr
,
1537 static struct attribute_group ksm_attr_group
= {
1542 static int __init
ksm_init(void)
1544 struct task_struct
*ksm_thread
;
1547 err
= ksm_slab_init();
1551 err
= mm_slots_hash_init();
1555 ksm_thread
= kthread_run(ksm_scan_thread
, NULL
, "ksmd");
1556 if (IS_ERR(ksm_thread
)) {
1557 printk(KERN_ERR
"ksm: creating kthread failed\n");
1558 err
= PTR_ERR(ksm_thread
);
1562 err
= sysfs_create_group(mm_kobj
, &ksm_attr_group
);
1564 printk(KERN_ERR
"ksm: register sysfs failed\n");
1571 kthread_stop(ksm_thread
);
1573 mm_slots_hash_free();
1579 module_init(ksm_init
)