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/sched/mm.h>
23 #include <linux/sched/coredump.h>
24 #include <linux/rwsem.h>
25 #include <linux/pagemap.h>
26 #include <linux/rmap.h>
27 #include <linux/spinlock.h>
28 #include <linux/jhash.h>
29 #include <linux/delay.h>
30 #include <linux/kthread.h>
31 #include <linux/wait.h>
32 #include <linux/slab.h>
33 #include <linux/rbtree.h>
34 #include <linux/memory.h>
35 #include <linux/mmu_notifier.h>
36 #include <linux/swap.h>
37 #include <linux/ksm.h>
38 #include <linux/hashtable.h>
39 #include <linux/freezer.h>
40 #include <linux/oom.h>
41 #include <linux/numa.h>
43 #include <asm/tlbflush.h>
48 #define DO_NUMA(x) do { (x); } while (0)
51 #define DO_NUMA(x) do { } while (0)
55 * A few notes about the KSM scanning process,
56 * to make it easier to understand the data structures below:
58 * In order to reduce excessive scanning, KSM sorts the memory pages by their
59 * contents into a data structure that holds pointers to the pages' locations.
61 * Since the contents of the pages may change at any moment, KSM cannot just
62 * insert the pages into a normal sorted tree and expect it to find anything.
63 * Therefore KSM uses two data structures - the stable and the unstable tree.
65 * The stable tree holds pointers to all the merged pages (ksm pages), sorted
66 * by their contents. Because each such page is write-protected, searching on
67 * this tree is fully assured to be working (except when pages are unmapped),
68 * and therefore this tree is called the stable tree.
70 * In addition to the stable tree, KSM uses a second data structure called the
71 * unstable tree: this tree holds pointers to pages which have been found to
72 * be "unchanged for a period of time". The unstable tree sorts these pages
73 * by their contents, but since they are not write-protected, KSM cannot rely
74 * upon the unstable tree to work correctly - the unstable tree is liable to
75 * be corrupted as its contents are modified, and so it is called unstable.
77 * KSM solves this problem by several techniques:
79 * 1) The unstable tree is flushed every time KSM completes scanning all
80 * memory areas, and then the tree is rebuilt again from the beginning.
81 * 2) KSM will only insert into the unstable tree, pages whose hash value
82 * has not changed since the previous scan of all memory areas.
83 * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
84 * colors of the nodes and not on their contents, assuring that even when
85 * the tree gets "corrupted" it won't get out of balance, so scanning time
86 * remains the same (also, searching and inserting nodes in an rbtree uses
87 * the same algorithm, so we have no overhead when we flush and rebuild).
88 * 4) KSM never flushes the stable tree, which means that even if it were to
89 * take 10 attempts to find a page in the unstable tree, once it is found,
90 * it is secured in the stable tree. (When we scan a new page, we first
91 * compare it against the stable tree, and then against the unstable tree.)
93 * If the merge_across_nodes tunable is unset, then KSM maintains multiple
94 * stable trees and multiple unstable trees: one of each for each NUMA node.
98 * struct mm_slot - ksm information per mm that is being scanned
99 * @link: link to the mm_slots hash list
100 * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
101 * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
102 * @mm: the mm that this information is valid for
105 struct hlist_node link
;
106 struct list_head mm_list
;
107 struct rmap_item
*rmap_list
;
108 struct mm_struct
*mm
;
112 * struct ksm_scan - cursor for scanning
113 * @mm_slot: the current mm_slot we are scanning
114 * @address: the next address inside that to be scanned
115 * @rmap_list: link to the next rmap to be scanned in the rmap_list
116 * @seqnr: count of completed full scans (needed when removing unstable node)
118 * There is only the one ksm_scan instance of this cursor structure.
121 struct mm_slot
*mm_slot
;
122 unsigned long address
;
123 struct rmap_item
**rmap_list
;
128 * struct stable_node - node of the stable rbtree
129 * @node: rb node of this ksm page in the stable tree
130 * @head: (overlaying parent) &migrate_nodes indicates temporarily on that list
131 * @hlist_dup: linked into the stable_node->hlist with a stable_node chain
132 * @list: linked into migrate_nodes, pending placement in the proper node tree
133 * @hlist: hlist head of rmap_items using this ksm page
134 * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid)
135 * @chain_prune_time: time of the last full garbage collection
136 * @rmap_hlist_len: number of rmap_item entries in hlist or STABLE_NODE_CHAIN
137 * @nid: NUMA node id of stable tree in which linked (may not match kpfn)
141 struct rb_node node
; /* when node of stable tree */
142 struct { /* when listed for migration */
143 struct list_head
*head
;
145 struct hlist_node hlist_dup
;
146 struct list_head list
;
150 struct hlist_head hlist
;
153 unsigned long chain_prune_time
;
156 * STABLE_NODE_CHAIN can be any negative number in
157 * rmap_hlist_len negative range, but better not -1 to be able
158 * to reliably detect underflows.
160 #define STABLE_NODE_CHAIN -1024
168 * struct rmap_item - reverse mapping item for virtual addresses
169 * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
170 * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
171 * @nid: NUMA node id of unstable tree in which linked (may not match page)
172 * @mm: the memory structure this rmap_item is pointing into
173 * @address: the virtual address this rmap_item tracks (+ flags in low bits)
174 * @oldchecksum: previous checksum of the page at that virtual address
175 * @node: rb node of this rmap_item in the unstable tree
176 * @head: pointer to stable_node heading this list in the stable tree
177 * @hlist: link into hlist of rmap_items hanging off that stable_node
180 struct rmap_item
*rmap_list
;
182 struct anon_vma
*anon_vma
; /* when stable */
184 int nid
; /* when node of unstable tree */
187 struct mm_struct
*mm
;
188 unsigned long address
; /* + low bits used for flags below */
189 unsigned int oldchecksum
; /* when unstable */
191 struct rb_node node
; /* when node of unstable tree */
192 struct { /* when listed from stable tree */
193 struct stable_node
*head
;
194 struct hlist_node hlist
;
199 #define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */
200 #define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */
201 #define STABLE_FLAG 0x200 /* is listed from the stable tree */
202 #define KSM_FLAG_MASK (SEQNR_MASK|UNSTABLE_FLAG|STABLE_FLAG)
203 /* to mask all the flags */
205 /* The stable and unstable tree heads */
206 static struct rb_root one_stable_tree
[1] = { RB_ROOT
};
207 static struct rb_root one_unstable_tree
[1] = { RB_ROOT
};
208 static struct rb_root
*root_stable_tree
= one_stable_tree
;
209 static struct rb_root
*root_unstable_tree
= one_unstable_tree
;
211 /* Recently migrated nodes of stable tree, pending proper placement */
212 static LIST_HEAD(migrate_nodes
);
213 #define STABLE_NODE_DUP_HEAD ((struct list_head *)&migrate_nodes.prev)
215 #define MM_SLOTS_HASH_BITS 10
216 static DEFINE_HASHTABLE(mm_slots_hash
, MM_SLOTS_HASH_BITS
);
218 static struct mm_slot ksm_mm_head
= {
219 .mm_list
= LIST_HEAD_INIT(ksm_mm_head
.mm_list
),
221 static struct ksm_scan ksm_scan
= {
222 .mm_slot
= &ksm_mm_head
,
225 static struct kmem_cache
*rmap_item_cache
;
226 static struct kmem_cache
*stable_node_cache
;
227 static struct kmem_cache
*mm_slot_cache
;
229 /* The number of nodes in the stable tree */
230 static unsigned long ksm_pages_shared
;
232 /* The number of page slots additionally sharing those nodes */
233 static unsigned long ksm_pages_sharing
;
235 /* The number of nodes in the unstable tree */
236 static unsigned long ksm_pages_unshared
;
238 /* The number of rmap_items in use: to calculate pages_volatile */
239 static unsigned long ksm_rmap_items
;
241 /* The number of stable_node chains */
242 static unsigned long ksm_stable_node_chains
;
244 /* The number of stable_node dups linked to the stable_node chains */
245 static unsigned long ksm_stable_node_dups
;
247 /* Delay in pruning stale stable_node_dups in the stable_node_chains */
248 static int ksm_stable_node_chains_prune_millisecs
= 2000;
250 /* Maximum number of page slots sharing a stable node */
251 static int ksm_max_page_sharing
= 256;
253 /* Number of pages ksmd should scan in one batch */
254 static unsigned int ksm_thread_pages_to_scan
= 100;
256 /* Milliseconds ksmd should sleep between batches */
257 static unsigned int ksm_thread_sleep_millisecs
= 20;
259 /* Checksum of an empty (zeroed) page */
260 static unsigned int zero_checksum __read_mostly
;
262 /* Whether to merge empty (zeroed) pages with actual zero pages */
263 static bool ksm_use_zero_pages __read_mostly
;
266 /* Zeroed when merging across nodes is not allowed */
267 static unsigned int ksm_merge_across_nodes
= 1;
268 static int ksm_nr_node_ids
= 1;
270 #define ksm_merge_across_nodes 1U
271 #define ksm_nr_node_ids 1
274 #define KSM_RUN_STOP 0
275 #define KSM_RUN_MERGE 1
276 #define KSM_RUN_UNMERGE 2
277 #define KSM_RUN_OFFLINE 4
278 static unsigned long ksm_run
= KSM_RUN_STOP
;
279 static void wait_while_offlining(void);
281 static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait
);
282 static DEFINE_MUTEX(ksm_thread_mutex
);
283 static DEFINE_SPINLOCK(ksm_mmlist_lock
);
285 #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
286 sizeof(struct __struct), __alignof__(struct __struct),\
289 static int __init
ksm_slab_init(void)
291 rmap_item_cache
= KSM_KMEM_CACHE(rmap_item
, 0);
292 if (!rmap_item_cache
)
295 stable_node_cache
= KSM_KMEM_CACHE(stable_node
, 0);
296 if (!stable_node_cache
)
299 mm_slot_cache
= KSM_KMEM_CACHE(mm_slot
, 0);
306 kmem_cache_destroy(stable_node_cache
);
308 kmem_cache_destroy(rmap_item_cache
);
313 static void __init
ksm_slab_free(void)
315 kmem_cache_destroy(mm_slot_cache
);
316 kmem_cache_destroy(stable_node_cache
);
317 kmem_cache_destroy(rmap_item_cache
);
318 mm_slot_cache
= NULL
;
321 static __always_inline
bool is_stable_node_chain(struct stable_node
*chain
)
323 return chain
->rmap_hlist_len
== STABLE_NODE_CHAIN
;
326 static __always_inline
bool is_stable_node_dup(struct stable_node
*dup
)
328 return dup
->head
== STABLE_NODE_DUP_HEAD
;
331 static inline void stable_node_chain_add_dup(struct stable_node
*dup
,
332 struct stable_node
*chain
)
334 VM_BUG_ON(is_stable_node_dup(dup
));
335 dup
->head
= STABLE_NODE_DUP_HEAD
;
336 VM_BUG_ON(!is_stable_node_chain(chain
));
337 hlist_add_head(&dup
->hlist_dup
, &chain
->hlist
);
338 ksm_stable_node_dups
++;
341 static inline void __stable_node_dup_del(struct stable_node
*dup
)
343 VM_BUG_ON(!is_stable_node_dup(dup
));
344 hlist_del(&dup
->hlist_dup
);
345 ksm_stable_node_dups
--;
348 static inline void stable_node_dup_del(struct stable_node
*dup
)
350 VM_BUG_ON(is_stable_node_chain(dup
));
351 if (is_stable_node_dup(dup
))
352 __stable_node_dup_del(dup
);
354 rb_erase(&dup
->node
, root_stable_tree
+ NUMA(dup
->nid
));
355 #ifdef CONFIG_DEBUG_VM
360 static inline struct rmap_item
*alloc_rmap_item(void)
362 struct rmap_item
*rmap_item
;
364 rmap_item
= kmem_cache_zalloc(rmap_item_cache
, GFP_KERNEL
|
365 __GFP_NORETRY
| __GFP_NOWARN
);
371 static inline void free_rmap_item(struct rmap_item
*rmap_item
)
374 rmap_item
->mm
= NULL
; /* debug safety */
375 kmem_cache_free(rmap_item_cache
, rmap_item
);
378 static inline struct stable_node
*alloc_stable_node(void)
381 * The allocation can take too long with GFP_KERNEL when memory is under
382 * pressure, which may lead to hung task warnings. Adding __GFP_HIGH
383 * grants access to memory reserves, helping to avoid this problem.
385 return kmem_cache_alloc(stable_node_cache
, GFP_KERNEL
| __GFP_HIGH
);
388 static inline void free_stable_node(struct stable_node
*stable_node
)
390 VM_BUG_ON(stable_node
->rmap_hlist_len
&&
391 !is_stable_node_chain(stable_node
));
392 kmem_cache_free(stable_node_cache
, stable_node
);
395 static inline struct mm_slot
*alloc_mm_slot(void)
397 if (!mm_slot_cache
) /* initialization failed */
399 return kmem_cache_zalloc(mm_slot_cache
, GFP_KERNEL
);
402 static inline void free_mm_slot(struct mm_slot
*mm_slot
)
404 kmem_cache_free(mm_slot_cache
, mm_slot
);
407 static struct mm_slot
*get_mm_slot(struct mm_struct
*mm
)
409 struct mm_slot
*slot
;
411 hash_for_each_possible(mm_slots_hash
, slot
, link
, (unsigned long)mm
)
418 static void insert_to_mm_slots_hash(struct mm_struct
*mm
,
419 struct mm_slot
*mm_slot
)
422 hash_add(mm_slots_hash
, &mm_slot
->link
, (unsigned long)mm
);
426 * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
427 * page tables after it has passed through ksm_exit() - which, if necessary,
428 * takes mmap_sem briefly to serialize against them. ksm_exit() does not set
429 * a special flag: they can just back out as soon as mm_users goes to zero.
430 * ksm_test_exit() is used throughout to make this test for exit: in some
431 * places for correctness, in some places just to avoid unnecessary work.
433 static inline bool ksm_test_exit(struct mm_struct
*mm
)
435 return atomic_read(&mm
->mm_users
) == 0;
439 * We use break_ksm to break COW on a ksm page: it's a stripped down
441 * if (get_user_pages(addr, 1, 1, 1, &page, NULL) == 1)
444 * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
445 * in case the application has unmapped and remapped mm,addr meanwhile.
446 * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP
447 * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
449 * FAULT_FLAG/FOLL_REMOTE are because we do this outside the context
450 * of the process that owns 'vma'. We also do not want to enforce
451 * protection keys here anyway.
453 static int break_ksm(struct vm_area_struct
*vma
, unsigned long addr
)
460 page
= follow_page(vma
, addr
,
461 FOLL_GET
| FOLL_MIGRATION
| FOLL_REMOTE
);
462 if (IS_ERR_OR_NULL(page
))
465 ret
= handle_mm_fault(vma
, addr
,
466 FAULT_FLAG_WRITE
| FAULT_FLAG_REMOTE
);
468 ret
= VM_FAULT_WRITE
;
470 } while (!(ret
& (VM_FAULT_WRITE
| VM_FAULT_SIGBUS
| VM_FAULT_SIGSEGV
| VM_FAULT_OOM
)));
472 * We must loop because handle_mm_fault() may back out if there's
473 * any difficulty e.g. if pte accessed bit gets updated concurrently.
475 * VM_FAULT_WRITE is what we have been hoping for: it indicates that
476 * COW has been broken, even if the vma does not permit VM_WRITE;
477 * but note that a concurrent fault might break PageKsm for us.
479 * VM_FAULT_SIGBUS could occur if we race with truncation of the
480 * backing file, which also invalidates anonymous pages: that's
481 * okay, that truncation will have unmapped the PageKsm for us.
483 * VM_FAULT_OOM: at the time of writing (late July 2009), setting
484 * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
485 * current task has TIF_MEMDIE set, and will be OOM killed on return
486 * to user; and ksmd, having no mm, would never be chosen for that.
488 * But if the mm is in a limited mem_cgroup, then the fault may fail
489 * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
490 * even ksmd can fail in this way - though it's usually breaking ksm
491 * just to undo a merge it made a moment before, so unlikely to oom.
493 * That's a pity: we might therefore have more kernel pages allocated
494 * than we're counting as nodes in the stable tree; but ksm_do_scan
495 * will retry to break_cow on each pass, so should recover the page
496 * in due course. The important thing is to not let VM_MERGEABLE
497 * be cleared while any such pages might remain in the area.
499 return (ret
& VM_FAULT_OOM
) ? -ENOMEM
: 0;
502 static struct vm_area_struct
*find_mergeable_vma(struct mm_struct
*mm
,
505 struct vm_area_struct
*vma
;
506 if (ksm_test_exit(mm
))
508 vma
= find_vma(mm
, addr
);
509 if (!vma
|| vma
->vm_start
> addr
)
511 if (!(vma
->vm_flags
& VM_MERGEABLE
) || !vma
->anon_vma
)
516 static void break_cow(struct rmap_item
*rmap_item
)
518 struct mm_struct
*mm
= rmap_item
->mm
;
519 unsigned long addr
= rmap_item
->address
;
520 struct vm_area_struct
*vma
;
523 * It is not an accident that whenever we want to break COW
524 * to undo, we also need to drop a reference to the anon_vma.
526 put_anon_vma(rmap_item
->anon_vma
);
528 down_read(&mm
->mmap_sem
);
529 vma
= find_mergeable_vma(mm
, addr
);
531 break_ksm(vma
, addr
);
532 up_read(&mm
->mmap_sem
);
535 static struct page
*get_mergeable_page(struct rmap_item
*rmap_item
)
537 struct mm_struct
*mm
= rmap_item
->mm
;
538 unsigned long addr
= rmap_item
->address
;
539 struct vm_area_struct
*vma
;
542 down_read(&mm
->mmap_sem
);
543 vma
= find_mergeable_vma(mm
, addr
);
547 page
= follow_page(vma
, addr
, FOLL_GET
);
548 if (IS_ERR_OR_NULL(page
))
550 if (PageAnon(page
)) {
551 flush_anon_page(vma
, page
, addr
);
552 flush_dcache_page(page
);
558 up_read(&mm
->mmap_sem
);
563 * This helper is used for getting right index into array of tree roots.
564 * When merge_across_nodes knob is set to 1, there are only two rb-trees for
565 * stable and unstable pages from all nodes with roots in index 0. Otherwise,
566 * every node has its own stable and unstable tree.
568 static inline int get_kpfn_nid(unsigned long kpfn
)
570 return ksm_merge_across_nodes
? 0 : NUMA(pfn_to_nid(kpfn
));
573 static struct stable_node
*alloc_stable_node_chain(struct stable_node
*dup
,
574 struct rb_root
*root
)
576 struct stable_node
*chain
= alloc_stable_node();
577 VM_BUG_ON(is_stable_node_chain(dup
));
579 INIT_HLIST_HEAD(&chain
->hlist
);
580 chain
->chain_prune_time
= jiffies
;
581 chain
->rmap_hlist_len
= STABLE_NODE_CHAIN
;
582 #if defined (CONFIG_DEBUG_VM) && defined(CONFIG_NUMA)
583 chain
->nid
= -1; /* debug */
585 ksm_stable_node_chains
++;
588 * Put the stable node chain in the first dimension of
589 * the stable tree and at the same time remove the old
592 rb_replace_node(&dup
->node
, &chain
->node
, root
);
595 * Move the old stable node to the second dimension
596 * queued in the hlist_dup. The invariant is that all
597 * dup stable_nodes in the chain->hlist point to pages
598 * that are wrprotected and have the exact same
601 stable_node_chain_add_dup(dup
, chain
);
606 static inline void free_stable_node_chain(struct stable_node
*chain
,
607 struct rb_root
*root
)
609 rb_erase(&chain
->node
, root
);
610 free_stable_node(chain
);
611 ksm_stable_node_chains
--;
614 static void remove_node_from_stable_tree(struct stable_node
*stable_node
)
616 struct rmap_item
*rmap_item
;
618 /* check it's not STABLE_NODE_CHAIN or negative */
619 BUG_ON(stable_node
->rmap_hlist_len
< 0);
621 hlist_for_each_entry(rmap_item
, &stable_node
->hlist
, hlist
) {
622 if (rmap_item
->hlist
.next
)
626 VM_BUG_ON(stable_node
->rmap_hlist_len
<= 0);
627 stable_node
->rmap_hlist_len
--;
628 put_anon_vma(rmap_item
->anon_vma
);
629 rmap_item
->address
&= PAGE_MASK
;
634 * We need the second aligned pointer of the migrate_nodes
635 * list_head to stay clear from the rb_parent_color union
636 * (aligned and different than any node) and also different
637 * from &migrate_nodes. This will verify that future list.h changes
638 * don't break STABLE_NODE_DUP_HEAD.
640 #if GCC_VERSION >= 40903 /* only recent gcc can handle it */
641 BUILD_BUG_ON(STABLE_NODE_DUP_HEAD
<= &migrate_nodes
);
642 BUILD_BUG_ON(STABLE_NODE_DUP_HEAD
>= &migrate_nodes
+ 1);
645 if (stable_node
->head
== &migrate_nodes
)
646 list_del(&stable_node
->list
);
648 stable_node_dup_del(stable_node
);
649 free_stable_node(stable_node
);
653 * get_ksm_page: checks if the page indicated by the stable node
654 * is still its ksm page, despite having held no reference to it.
655 * In which case we can trust the content of the page, and it
656 * returns the gotten page; but if the page has now been zapped,
657 * remove the stale node from the stable tree and return NULL.
658 * But beware, the stable node's page might be being migrated.
660 * You would expect the stable_node to hold a reference to the ksm page.
661 * But if it increments the page's count, swapping out has to wait for
662 * ksmd to come around again before it can free the page, which may take
663 * seconds or even minutes: much too unresponsive. So instead we use a
664 * "keyhole reference": access to the ksm page from the stable node peeps
665 * out through its keyhole to see if that page still holds the right key,
666 * pointing back to this stable node. This relies on freeing a PageAnon
667 * page to reset its page->mapping to NULL, and relies on no other use of
668 * a page to put something that might look like our key in page->mapping.
669 * is on its way to being freed; but it is an anomaly to bear in mind.
671 static struct page
*get_ksm_page(struct stable_node
*stable_node
, bool lock_it
)
674 void *expected_mapping
;
677 expected_mapping
= (void *)((unsigned long)stable_node
|
680 kpfn
= READ_ONCE(stable_node
->kpfn
);
681 page
= pfn_to_page(kpfn
);
684 * page is computed from kpfn, so on most architectures reading
685 * page->mapping is naturally ordered after reading node->kpfn,
686 * but on Alpha we need to be more careful.
688 smp_read_barrier_depends();
689 if (READ_ONCE(page
->mapping
) != expected_mapping
)
693 * We cannot do anything with the page while its refcount is 0.
694 * Usually 0 means free, or tail of a higher-order page: in which
695 * case this node is no longer referenced, and should be freed;
696 * however, it might mean that the page is under page_freeze_refs().
697 * The __remove_mapping() case is easy, again the node is now stale;
698 * but if page is swapcache in migrate_page_move_mapping(), it might
699 * still be our page, in which case it's essential to keep the node.
701 while (!get_page_unless_zero(page
)) {
703 * Another check for page->mapping != expected_mapping would
704 * work here too. We have chosen the !PageSwapCache test to
705 * optimize the common case, when the page is or is about to
706 * be freed: PageSwapCache is cleared (under spin_lock_irq)
707 * in the freeze_refs section of __remove_mapping(); but Anon
708 * page->mapping reset to NULL later, in free_pages_prepare().
710 if (!PageSwapCache(page
))
715 if (READ_ONCE(page
->mapping
) != expected_mapping
) {
722 if (READ_ONCE(page
->mapping
) != expected_mapping
) {
732 * We come here from above when page->mapping or !PageSwapCache
733 * suggests that the node is stale; but it might be under migration.
734 * We need smp_rmb(), matching the smp_wmb() in ksm_migrate_page(),
735 * before checking whether node->kpfn has been changed.
738 if (READ_ONCE(stable_node
->kpfn
) != kpfn
)
740 remove_node_from_stable_tree(stable_node
);
745 * Removing rmap_item from stable or unstable tree.
746 * This function will clean the information from the stable/unstable tree.
748 static void remove_rmap_item_from_tree(struct rmap_item
*rmap_item
)
750 if (rmap_item
->address
& STABLE_FLAG
) {
751 struct stable_node
*stable_node
;
754 stable_node
= rmap_item
->head
;
755 page
= get_ksm_page(stable_node
, true);
759 hlist_del(&rmap_item
->hlist
);
763 if (!hlist_empty(&stable_node
->hlist
))
767 VM_BUG_ON(stable_node
->rmap_hlist_len
<= 0);
768 stable_node
->rmap_hlist_len
--;
770 put_anon_vma(rmap_item
->anon_vma
);
771 rmap_item
->address
&= PAGE_MASK
;
773 } else if (rmap_item
->address
& UNSTABLE_FLAG
) {
776 * Usually ksmd can and must skip the rb_erase, because
777 * root_unstable_tree was already reset to RB_ROOT.
778 * But be careful when an mm is exiting: do the rb_erase
779 * if this rmap_item was inserted by this scan, rather
780 * than left over from before.
782 age
= (unsigned char)(ksm_scan
.seqnr
- rmap_item
->address
);
785 rb_erase(&rmap_item
->node
,
786 root_unstable_tree
+ NUMA(rmap_item
->nid
));
787 ksm_pages_unshared
--;
788 rmap_item
->address
&= PAGE_MASK
;
791 cond_resched(); /* we're called from many long loops */
794 static void remove_trailing_rmap_items(struct mm_slot
*mm_slot
,
795 struct rmap_item
**rmap_list
)
798 struct rmap_item
*rmap_item
= *rmap_list
;
799 *rmap_list
= rmap_item
->rmap_list
;
800 remove_rmap_item_from_tree(rmap_item
);
801 free_rmap_item(rmap_item
);
806 * Though it's very tempting to unmerge rmap_items from stable tree rather
807 * than check every pte of a given vma, the locking doesn't quite work for
808 * that - an rmap_item is assigned to the stable tree after inserting ksm
809 * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing
810 * rmap_items from parent to child at fork time (so as not to waste time
811 * if exit comes before the next scan reaches it).
813 * Similarly, although we'd like to remove rmap_items (so updating counts
814 * and freeing memory) when unmerging an area, it's easier to leave that
815 * to the next pass of ksmd - consider, for example, how ksmd might be
816 * in cmp_and_merge_page on one of the rmap_items we would be removing.
818 static int unmerge_ksm_pages(struct vm_area_struct
*vma
,
819 unsigned long start
, unsigned long end
)
824 for (addr
= start
; addr
< end
&& !err
; addr
+= PAGE_SIZE
) {
825 if (ksm_test_exit(vma
->vm_mm
))
827 if (signal_pending(current
))
830 err
= break_ksm(vma
, addr
);
837 * Only called through the sysfs control interface:
839 static int remove_stable_node(struct stable_node
*stable_node
)
844 page
= get_ksm_page(stable_node
, true);
847 * get_ksm_page did remove_node_from_stable_tree itself.
853 * Page could be still mapped if this races with __mmput() running in
854 * between ksm_exit() and exit_mmap(). Just refuse to let
855 * merge_across_nodes/max_page_sharing be switched.
858 if (!page_mapped(page
)) {
860 * The stable node did not yet appear stale to get_ksm_page(),
861 * since that allows for an unmapped ksm page to be recognized
862 * right up until it is freed; but the node is safe to remove.
863 * This page might be in a pagevec waiting to be freed,
864 * or it might be PageSwapCache (perhaps under writeback),
865 * or it might have been removed from swapcache a moment ago.
867 set_page_stable_node(page
, NULL
);
868 remove_node_from_stable_tree(stable_node
);
877 static int remove_stable_node_chain(struct stable_node
*stable_node
,
878 struct rb_root
*root
)
880 struct stable_node
*dup
;
881 struct hlist_node
*hlist_safe
;
883 if (!is_stable_node_chain(stable_node
)) {
884 VM_BUG_ON(is_stable_node_dup(stable_node
));
885 if (remove_stable_node(stable_node
))
891 hlist_for_each_entry_safe(dup
, hlist_safe
,
892 &stable_node
->hlist
, hlist_dup
) {
893 VM_BUG_ON(!is_stable_node_dup(dup
));
894 if (remove_stable_node(dup
))
897 BUG_ON(!hlist_empty(&stable_node
->hlist
));
898 free_stable_node_chain(stable_node
, root
);
902 static int remove_all_stable_nodes(void)
904 struct stable_node
*stable_node
, *next
;
908 for (nid
= 0; nid
< ksm_nr_node_ids
; nid
++) {
909 while (root_stable_tree
[nid
].rb_node
) {
910 stable_node
= rb_entry(root_stable_tree
[nid
].rb_node
,
911 struct stable_node
, node
);
912 if (remove_stable_node_chain(stable_node
,
913 root_stable_tree
+ nid
)) {
915 break; /* proceed to next nid */
920 list_for_each_entry_safe(stable_node
, next
, &migrate_nodes
, list
) {
921 if (remove_stable_node(stable_node
))
928 static int unmerge_and_remove_all_rmap_items(void)
930 struct mm_slot
*mm_slot
;
931 struct mm_struct
*mm
;
932 struct vm_area_struct
*vma
;
935 spin_lock(&ksm_mmlist_lock
);
936 ksm_scan
.mm_slot
= list_entry(ksm_mm_head
.mm_list
.next
,
937 struct mm_slot
, mm_list
);
938 spin_unlock(&ksm_mmlist_lock
);
940 for (mm_slot
= ksm_scan
.mm_slot
;
941 mm_slot
!= &ksm_mm_head
; mm_slot
= ksm_scan
.mm_slot
) {
943 down_read(&mm
->mmap_sem
);
944 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
945 if (ksm_test_exit(mm
))
947 if (!(vma
->vm_flags
& VM_MERGEABLE
) || !vma
->anon_vma
)
949 err
= unmerge_ksm_pages(vma
,
950 vma
->vm_start
, vma
->vm_end
);
955 remove_trailing_rmap_items(mm_slot
, &mm_slot
->rmap_list
);
956 up_read(&mm
->mmap_sem
);
958 spin_lock(&ksm_mmlist_lock
);
959 ksm_scan
.mm_slot
= list_entry(mm_slot
->mm_list
.next
,
960 struct mm_slot
, mm_list
);
961 if (ksm_test_exit(mm
)) {
962 hash_del(&mm_slot
->link
);
963 list_del(&mm_slot
->mm_list
);
964 spin_unlock(&ksm_mmlist_lock
);
966 free_mm_slot(mm_slot
);
967 clear_bit(MMF_VM_MERGEABLE
, &mm
->flags
);
970 spin_unlock(&ksm_mmlist_lock
);
973 /* Clean up stable nodes, but don't worry if some are still busy */
974 remove_all_stable_nodes();
979 up_read(&mm
->mmap_sem
);
980 spin_lock(&ksm_mmlist_lock
);
981 ksm_scan
.mm_slot
= &ksm_mm_head
;
982 spin_unlock(&ksm_mmlist_lock
);
985 #endif /* CONFIG_SYSFS */
987 static u32
calc_checksum(struct page
*page
)
990 void *addr
= kmap_atomic(page
);
991 checksum
= jhash2(addr
, PAGE_SIZE
/ 4, 17);
996 static int memcmp_pages(struct page
*page1
, struct page
*page2
)
1001 addr1
= kmap_atomic(page1
);
1002 addr2
= kmap_atomic(page2
);
1003 ret
= memcmp(addr1
, addr2
, PAGE_SIZE
);
1004 kunmap_atomic(addr2
);
1005 kunmap_atomic(addr1
);
1009 static inline int pages_identical(struct page
*page1
, struct page
*page2
)
1011 return !memcmp_pages(page1
, page2
);
1014 static int write_protect_page(struct vm_area_struct
*vma
, struct page
*page
,
1017 struct mm_struct
*mm
= vma
->vm_mm
;
1018 struct page_vma_mapped_walk pvmw
= {
1024 unsigned long mmun_start
; /* For mmu_notifiers */
1025 unsigned long mmun_end
; /* For mmu_notifiers */
1027 pvmw
.address
= page_address_in_vma(page
, vma
);
1028 if (pvmw
.address
== -EFAULT
)
1031 BUG_ON(PageTransCompound(page
));
1033 mmun_start
= pvmw
.address
;
1034 mmun_end
= pvmw
.address
+ PAGE_SIZE
;
1035 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
1037 if (!page_vma_mapped_walk(&pvmw
))
1039 if (WARN_ONCE(!pvmw
.pte
, "Unexpected PMD mapping?"))
1042 if (pte_write(*pvmw
.pte
) || pte_dirty(*pvmw
.pte
) ||
1043 (pte_protnone(*pvmw
.pte
) && pte_savedwrite(*pvmw
.pte
)) ||
1044 mm_tlb_flush_pending(mm
)) {
1047 swapped
= PageSwapCache(page
);
1048 flush_cache_page(vma
, pvmw
.address
, page_to_pfn(page
));
1050 * Ok this is tricky, when get_user_pages_fast() run it doesn't
1051 * take any lock, therefore the check that we are going to make
1052 * with the pagecount against the mapcount is racey and
1053 * O_DIRECT can happen right after the check.
1054 * So we clear the pte and flush the tlb before the check
1055 * this assure us that no O_DIRECT can happen after the check
1056 * or in the middle of the check.
1058 * No need to notify as we are downgrading page table to read
1059 * only not changing it to point to a new page.
1061 * See Documentation/vm/mmu_notifier.txt
1063 entry
= ptep_clear_flush(vma
, pvmw
.address
, pvmw
.pte
);
1065 * Check that no O_DIRECT or similar I/O is in progress on the
1068 if (page_mapcount(page
) + 1 + swapped
!= page_count(page
)) {
1069 set_pte_at(mm
, pvmw
.address
, pvmw
.pte
, entry
);
1072 if (pte_dirty(entry
))
1073 set_page_dirty(page
);
1075 if (pte_protnone(entry
))
1076 entry
= pte_mkclean(pte_clear_savedwrite(entry
));
1078 entry
= pte_mkclean(pte_wrprotect(entry
));
1079 set_pte_at_notify(mm
, pvmw
.address
, pvmw
.pte
, entry
);
1081 *orig_pte
= *pvmw
.pte
;
1085 page_vma_mapped_walk_done(&pvmw
);
1087 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
1093 * replace_page - replace page in vma by new ksm page
1094 * @vma: vma that holds the pte pointing to page
1095 * @page: the page we are replacing by kpage
1096 * @kpage: the ksm page we replace page by
1097 * @orig_pte: the original value of the pte
1099 * Returns 0 on success, -EFAULT on failure.
1101 static int replace_page(struct vm_area_struct
*vma
, struct page
*page
,
1102 struct page
*kpage
, pte_t orig_pte
)
1104 struct mm_struct
*mm
= vma
->vm_mm
;
1111 unsigned long mmun_start
; /* For mmu_notifiers */
1112 unsigned long mmun_end
; /* For mmu_notifiers */
1114 addr
= page_address_in_vma(page
, vma
);
1115 if (addr
== -EFAULT
)
1118 pmd
= mm_find_pmd(mm
, addr
);
1123 mmun_end
= addr
+ PAGE_SIZE
;
1124 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
1126 ptep
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
1127 if (!pte_same(*ptep
, orig_pte
)) {
1128 pte_unmap_unlock(ptep
, ptl
);
1133 * No need to check ksm_use_zero_pages here: we can only have a
1134 * zero_page here if ksm_use_zero_pages was enabled alreaady.
1136 if (!is_zero_pfn(page_to_pfn(kpage
))) {
1138 page_add_anon_rmap(kpage
, vma
, addr
, false);
1139 newpte
= mk_pte(kpage
, vma
->vm_page_prot
);
1141 newpte
= pte_mkspecial(pfn_pte(page_to_pfn(kpage
),
1142 vma
->vm_page_prot
));
1144 * We're replacing an anonymous page with a zero page, which is
1145 * not anonymous. We need to do proper accounting otherwise we
1146 * will get wrong values in /proc, and a BUG message in dmesg
1147 * when tearing down the mm.
1149 dec_mm_counter(mm
, MM_ANONPAGES
);
1152 flush_cache_page(vma
, addr
, pte_pfn(*ptep
));
1154 * No need to notify as we are replacing a read only page with another
1155 * read only page with the same content.
1157 * See Documentation/vm/mmu_notifier.txt
1159 ptep_clear_flush(vma
, addr
, ptep
);
1160 set_pte_at_notify(mm
, addr
, ptep
, newpte
);
1162 page_remove_rmap(page
, false);
1163 if (!page_mapped(page
))
1164 try_to_free_swap(page
);
1167 pte_unmap_unlock(ptep
, ptl
);
1170 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
1176 * try_to_merge_one_page - take two pages and merge them into one
1177 * @vma: the vma that holds the pte pointing to page
1178 * @page: the PageAnon page that we want to replace with kpage
1179 * @kpage: the PageKsm page that we want to map instead of page,
1180 * or NULL the first time when we want to use page as kpage.
1182 * This function returns 0 if the pages were merged, -EFAULT otherwise.
1184 static int try_to_merge_one_page(struct vm_area_struct
*vma
,
1185 struct page
*page
, struct page
*kpage
)
1187 pte_t orig_pte
= __pte(0);
1190 if (page
== kpage
) /* ksm page forked */
1193 if (!PageAnon(page
))
1197 * We need the page lock to read a stable PageSwapCache in
1198 * write_protect_page(). We use trylock_page() instead of
1199 * lock_page() because we don't want to wait here - we
1200 * prefer to continue scanning and merging different pages,
1201 * then come back to this page when it is unlocked.
1203 if (!trylock_page(page
))
1206 if (PageTransCompound(page
)) {
1207 if (split_huge_page(page
))
1212 * If this anonymous page is mapped only here, its pte may need
1213 * to be write-protected. If it's mapped elsewhere, all of its
1214 * ptes are necessarily already write-protected. But in either
1215 * case, we need to lock and check page_count is not raised.
1217 if (write_protect_page(vma
, page
, &orig_pte
) == 0) {
1220 * While we hold page lock, upgrade page from
1221 * PageAnon+anon_vma to PageKsm+NULL stable_node:
1222 * stable_tree_insert() will update stable_node.
1224 set_page_stable_node(page
, NULL
);
1225 mark_page_accessed(page
);
1227 * Page reclaim just frees a clean page with no dirty
1228 * ptes: make sure that the ksm page would be swapped.
1230 if (!PageDirty(page
))
1233 } else if (pages_identical(page
, kpage
))
1234 err
= replace_page(vma
, page
, kpage
, orig_pte
);
1237 if ((vma
->vm_flags
& VM_LOCKED
) && kpage
&& !err
) {
1238 munlock_vma_page(page
);
1239 if (!PageMlocked(kpage
)) {
1242 mlock_vma_page(kpage
);
1243 page
= kpage
; /* for final unlock */
1254 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
1255 * but no new kernel page is allocated: kpage must already be a ksm page.
1257 * This function returns 0 if the pages were merged, -EFAULT otherwise.
1259 static int try_to_merge_with_ksm_page(struct rmap_item
*rmap_item
,
1260 struct page
*page
, struct page
*kpage
)
1262 struct mm_struct
*mm
= rmap_item
->mm
;
1263 struct vm_area_struct
*vma
;
1266 down_read(&mm
->mmap_sem
);
1267 vma
= find_mergeable_vma(mm
, rmap_item
->address
);
1271 err
= try_to_merge_one_page(vma
, page
, kpage
);
1275 /* Unstable nid is in union with stable anon_vma: remove first */
1276 remove_rmap_item_from_tree(rmap_item
);
1278 /* Must get reference to anon_vma while still holding mmap_sem */
1279 rmap_item
->anon_vma
= vma
->anon_vma
;
1280 get_anon_vma(vma
->anon_vma
);
1282 up_read(&mm
->mmap_sem
);
1287 * try_to_merge_two_pages - take two identical pages and prepare them
1288 * to be merged into one page.
1290 * This function returns the kpage if we successfully merged two identical
1291 * pages into one ksm page, NULL otherwise.
1293 * Note that this function upgrades page to ksm page: if one of the pages
1294 * is already a ksm page, try_to_merge_with_ksm_page should be used.
1296 static struct page
*try_to_merge_two_pages(struct rmap_item
*rmap_item
,
1298 struct rmap_item
*tree_rmap_item
,
1299 struct page
*tree_page
)
1303 err
= try_to_merge_with_ksm_page(rmap_item
, page
, NULL
);
1305 err
= try_to_merge_with_ksm_page(tree_rmap_item
,
1308 * If that fails, we have a ksm page with only one pte
1309 * pointing to it: so break it.
1312 break_cow(rmap_item
);
1314 return err
? NULL
: page
;
1317 static __always_inline
1318 bool __is_page_sharing_candidate(struct stable_node
*stable_node
, int offset
)
1320 VM_BUG_ON(stable_node
->rmap_hlist_len
< 0);
1322 * Check that at least one mapping still exists, otherwise
1323 * there's no much point to merge and share with this
1324 * stable_node, as the underlying tree_page of the other
1325 * sharer is going to be freed soon.
1327 return stable_node
->rmap_hlist_len
&&
1328 stable_node
->rmap_hlist_len
+ offset
< ksm_max_page_sharing
;
1331 static __always_inline
1332 bool is_page_sharing_candidate(struct stable_node
*stable_node
)
1334 return __is_page_sharing_candidate(stable_node
, 0);
1337 struct page
*stable_node_dup(struct stable_node
**_stable_node_dup
,
1338 struct stable_node
**_stable_node
,
1339 struct rb_root
*root
,
1340 bool prune_stale_stable_nodes
)
1342 struct stable_node
*dup
, *found
= NULL
, *stable_node
= *_stable_node
;
1343 struct hlist_node
*hlist_safe
;
1344 struct page
*_tree_page
, *tree_page
= NULL
;
1346 int found_rmap_hlist_len
;
1348 if (!prune_stale_stable_nodes
||
1349 time_before(jiffies
, stable_node
->chain_prune_time
+
1351 ksm_stable_node_chains_prune_millisecs
)))
1352 prune_stale_stable_nodes
= false;
1354 stable_node
->chain_prune_time
= jiffies
;
1356 hlist_for_each_entry_safe(dup
, hlist_safe
,
1357 &stable_node
->hlist
, hlist_dup
) {
1360 * We must walk all stable_node_dup to prune the stale
1361 * stable nodes during lookup.
1363 * get_ksm_page can drop the nodes from the
1364 * stable_node->hlist if they point to freed pages
1365 * (that's why we do a _safe walk). The "dup"
1366 * stable_node parameter itself will be freed from
1367 * under us if it returns NULL.
1369 _tree_page
= get_ksm_page(dup
, false);
1373 if (is_page_sharing_candidate(dup
)) {
1375 dup
->rmap_hlist_len
> found_rmap_hlist_len
) {
1377 put_page(tree_page
);
1379 found_rmap_hlist_len
= found
->rmap_hlist_len
;
1380 tree_page
= _tree_page
;
1382 /* skip put_page for found dup */
1383 if (!prune_stale_stable_nodes
)
1388 put_page(_tree_page
);
1393 * nr is counting all dups in the chain only if
1394 * prune_stale_stable_nodes is true, otherwise we may
1395 * break the loop at nr == 1 even if there are
1398 if (prune_stale_stable_nodes
&& nr
== 1) {
1400 * If there's not just one entry it would
1401 * corrupt memory, better BUG_ON. In KSM
1402 * context with no lock held it's not even
1405 BUG_ON(stable_node
->hlist
.first
->next
);
1408 * There's just one entry and it is below the
1409 * deduplication limit so drop the chain.
1411 rb_replace_node(&stable_node
->node
, &found
->node
,
1413 free_stable_node(stable_node
);
1414 ksm_stable_node_chains
--;
1415 ksm_stable_node_dups
--;
1417 * NOTE: the caller depends on the stable_node
1418 * to be equal to stable_node_dup if the chain
1421 *_stable_node
= found
;
1423 * Just for robustneess as stable_node is
1424 * otherwise left as a stable pointer, the
1425 * compiler shall optimize it away at build
1429 } else if (stable_node
->hlist
.first
!= &found
->hlist_dup
&&
1430 __is_page_sharing_candidate(found
, 1)) {
1432 * If the found stable_node dup can accept one
1433 * more future merge (in addition to the one
1434 * that is underway) and is not at the head of
1435 * the chain, put it there so next search will
1436 * be quicker in the !prune_stale_stable_nodes
1439 * NOTE: it would be inaccurate to use nr > 1
1440 * instead of checking the hlist.first pointer
1441 * directly, because in the
1442 * prune_stale_stable_nodes case "nr" isn't
1443 * the position of the found dup in the chain,
1444 * but the total number of dups in the chain.
1446 hlist_del(&found
->hlist_dup
);
1447 hlist_add_head(&found
->hlist_dup
,
1448 &stable_node
->hlist
);
1452 *_stable_node_dup
= found
;
1456 static struct stable_node
*stable_node_dup_any(struct stable_node
*stable_node
,
1457 struct rb_root
*root
)
1459 if (!is_stable_node_chain(stable_node
))
1461 if (hlist_empty(&stable_node
->hlist
)) {
1462 free_stable_node_chain(stable_node
, root
);
1465 return hlist_entry(stable_node
->hlist
.first
,
1466 typeof(*stable_node
), hlist_dup
);
1470 * Like for get_ksm_page, this function can free the *_stable_node and
1471 * *_stable_node_dup if the returned tree_page is NULL.
1473 * It can also free and overwrite *_stable_node with the found
1474 * stable_node_dup if the chain is collapsed (in which case
1475 * *_stable_node will be equal to *_stable_node_dup like if the chain
1476 * never existed). It's up to the caller to verify tree_page is not
1477 * NULL before dereferencing *_stable_node or *_stable_node_dup.
1479 * *_stable_node_dup is really a second output parameter of this
1480 * function and will be overwritten in all cases, the caller doesn't
1481 * need to initialize it.
1483 static struct page
*__stable_node_chain(struct stable_node
**_stable_node_dup
,
1484 struct stable_node
**_stable_node
,
1485 struct rb_root
*root
,
1486 bool prune_stale_stable_nodes
)
1488 struct stable_node
*stable_node
= *_stable_node
;
1489 if (!is_stable_node_chain(stable_node
)) {
1490 if (is_page_sharing_candidate(stable_node
)) {
1491 *_stable_node_dup
= stable_node
;
1492 return get_ksm_page(stable_node
, false);
1495 * _stable_node_dup set to NULL means the stable_node
1496 * reached the ksm_max_page_sharing limit.
1498 *_stable_node_dup
= NULL
;
1501 return stable_node_dup(_stable_node_dup
, _stable_node
, root
,
1502 prune_stale_stable_nodes
);
1505 static __always_inline
struct page
*chain_prune(struct stable_node
**s_n_d
,
1506 struct stable_node
**s_n
,
1507 struct rb_root
*root
)
1509 return __stable_node_chain(s_n_d
, s_n
, root
, true);
1512 static __always_inline
struct page
*chain(struct stable_node
**s_n_d
,
1513 struct stable_node
*s_n
,
1514 struct rb_root
*root
)
1516 struct stable_node
*old_stable_node
= s_n
;
1517 struct page
*tree_page
;
1519 tree_page
= __stable_node_chain(s_n_d
, &s_n
, root
, false);
1520 /* not pruning dups so s_n cannot have changed */
1521 VM_BUG_ON(s_n
!= old_stable_node
);
1526 * stable_tree_search - search for page inside the stable tree
1528 * This function checks if there is a page inside the stable tree
1529 * with identical content to the page that we are scanning right now.
1531 * This function returns the stable tree node of identical content if found,
1534 static struct page
*stable_tree_search(struct page
*page
)
1537 struct rb_root
*root
;
1538 struct rb_node
**new;
1539 struct rb_node
*parent
;
1540 struct stable_node
*stable_node
, *stable_node_dup
, *stable_node_any
;
1541 struct stable_node
*page_node
;
1543 page_node
= page_stable_node(page
);
1544 if (page_node
&& page_node
->head
!= &migrate_nodes
) {
1545 /* ksm page forked */
1550 nid
= get_kpfn_nid(page_to_pfn(page
));
1551 root
= root_stable_tree
+ nid
;
1553 new = &root
->rb_node
;
1557 struct page
*tree_page
;
1561 stable_node
= rb_entry(*new, struct stable_node
, node
);
1562 stable_node_any
= NULL
;
1563 tree_page
= chain_prune(&stable_node_dup
, &stable_node
, root
);
1565 * NOTE: stable_node may have been freed by
1566 * chain_prune() if the returned stable_node_dup is
1567 * not NULL. stable_node_dup may have been inserted in
1568 * the rbtree instead as a regular stable_node (in
1569 * order to collapse the stable_node chain if a single
1570 * stable_node dup was found in it). In such case the
1571 * stable_node is overwritten by the calleee to point
1572 * to the stable_node_dup that was collapsed in the
1573 * stable rbtree and stable_node will be equal to
1574 * stable_node_dup like if the chain never existed.
1576 if (!stable_node_dup
) {
1578 * Either all stable_node dups were full in
1579 * this stable_node chain, or this chain was
1580 * empty and should be rb_erased.
1582 stable_node_any
= stable_node_dup_any(stable_node
,
1584 if (!stable_node_any
) {
1585 /* rb_erase just run */
1589 * Take any of the stable_node dups page of
1590 * this stable_node chain to let the tree walk
1591 * continue. All KSM pages belonging to the
1592 * stable_node dups in a stable_node chain
1593 * have the same content and they're
1594 * wrprotected at all times. Any will work
1595 * fine to continue the walk.
1597 tree_page
= get_ksm_page(stable_node_any
, false);
1599 VM_BUG_ON(!stable_node_dup
^ !!stable_node_any
);
1602 * If we walked over a stale stable_node,
1603 * get_ksm_page() will call rb_erase() and it
1604 * may rebalance the tree from under us. So
1605 * restart the search from scratch. Returning
1606 * NULL would be safe too, but we'd generate
1607 * false negative insertions just because some
1608 * stable_node was stale.
1613 ret
= memcmp_pages(page
, tree_page
);
1614 put_page(tree_page
);
1618 new = &parent
->rb_left
;
1620 new = &parent
->rb_right
;
1623 VM_BUG_ON(page_node
->head
!= &migrate_nodes
);
1625 * Test if the migrated page should be merged
1626 * into a stable node dup. If the mapcount is
1627 * 1 we can migrate it with another KSM page
1628 * without adding it to the chain.
1630 if (page_mapcount(page
) > 1)
1634 if (!stable_node_dup
) {
1636 * If the stable_node is a chain and
1637 * we got a payload match in memcmp
1638 * but we cannot merge the scanned
1639 * page in any of the existing
1640 * stable_node dups because they're
1641 * all full, we need to wait the
1642 * scanned page to find itself a match
1643 * in the unstable tree to create a
1644 * brand new KSM page to add later to
1645 * the dups of this stable_node.
1651 * Lock and unlock the stable_node's page (which
1652 * might already have been migrated) so that page
1653 * migration is sure to notice its raised count.
1654 * It would be more elegant to return stable_node
1655 * than kpage, but that involves more changes.
1657 tree_page
= get_ksm_page(stable_node_dup
, true);
1658 if (unlikely(!tree_page
))
1660 * The tree may have been rebalanced,
1661 * so re-evaluate parent and new.
1664 unlock_page(tree_page
);
1666 if (get_kpfn_nid(stable_node_dup
->kpfn
) !=
1667 NUMA(stable_node_dup
->nid
)) {
1668 put_page(tree_page
);
1678 list_del(&page_node
->list
);
1679 DO_NUMA(page_node
->nid
= nid
);
1680 rb_link_node(&page_node
->node
, parent
, new);
1681 rb_insert_color(&page_node
->node
, root
);
1683 if (is_page_sharing_candidate(page_node
)) {
1691 * If stable_node was a chain and chain_prune collapsed it,
1692 * stable_node has been updated to be the new regular
1693 * stable_node. A collapse of the chain is indistinguishable
1694 * from the case there was no chain in the stable
1695 * rbtree. Otherwise stable_node is the chain and
1696 * stable_node_dup is the dup to replace.
1698 if (stable_node_dup
== stable_node
) {
1699 VM_BUG_ON(is_stable_node_chain(stable_node_dup
));
1700 VM_BUG_ON(is_stable_node_dup(stable_node_dup
));
1701 /* there is no chain */
1703 VM_BUG_ON(page_node
->head
!= &migrate_nodes
);
1704 list_del(&page_node
->list
);
1705 DO_NUMA(page_node
->nid
= nid
);
1706 rb_replace_node(&stable_node_dup
->node
,
1709 if (is_page_sharing_candidate(page_node
))
1714 rb_erase(&stable_node_dup
->node
, root
);
1718 VM_BUG_ON(!is_stable_node_chain(stable_node
));
1719 __stable_node_dup_del(stable_node_dup
);
1721 VM_BUG_ON(page_node
->head
!= &migrate_nodes
);
1722 list_del(&page_node
->list
);
1723 DO_NUMA(page_node
->nid
= nid
);
1724 stable_node_chain_add_dup(page_node
, stable_node
);
1725 if (is_page_sharing_candidate(page_node
))
1733 stable_node_dup
->head
= &migrate_nodes
;
1734 list_add(&stable_node_dup
->list
, stable_node_dup
->head
);
1738 /* stable_node_dup could be null if it reached the limit */
1739 if (!stable_node_dup
)
1740 stable_node_dup
= stable_node_any
;
1742 * If stable_node was a chain and chain_prune collapsed it,
1743 * stable_node has been updated to be the new regular
1744 * stable_node. A collapse of the chain is indistinguishable
1745 * from the case there was no chain in the stable
1746 * rbtree. Otherwise stable_node is the chain and
1747 * stable_node_dup is the dup to replace.
1749 if (stable_node_dup
== stable_node
) {
1750 VM_BUG_ON(is_stable_node_chain(stable_node_dup
));
1751 VM_BUG_ON(is_stable_node_dup(stable_node_dup
));
1752 /* chain is missing so create it */
1753 stable_node
= alloc_stable_node_chain(stable_node_dup
,
1759 * Add this stable_node dup that was
1760 * migrated to the stable_node chain
1761 * of the current nid for this page
1764 VM_BUG_ON(!is_stable_node_chain(stable_node
));
1765 VM_BUG_ON(!is_stable_node_dup(stable_node_dup
));
1766 VM_BUG_ON(page_node
->head
!= &migrate_nodes
);
1767 list_del(&page_node
->list
);
1768 DO_NUMA(page_node
->nid
= nid
);
1769 stable_node_chain_add_dup(page_node
, stable_node
);
1774 * stable_tree_insert - insert stable tree node pointing to new ksm page
1775 * into the stable tree.
1777 * This function returns the stable tree node just allocated on success,
1780 static struct stable_node
*stable_tree_insert(struct page
*kpage
)
1784 struct rb_root
*root
;
1785 struct rb_node
**new;
1786 struct rb_node
*parent
;
1787 struct stable_node
*stable_node
, *stable_node_dup
, *stable_node_any
;
1788 bool need_chain
= false;
1790 kpfn
= page_to_pfn(kpage
);
1791 nid
= get_kpfn_nid(kpfn
);
1792 root
= root_stable_tree
+ nid
;
1795 new = &root
->rb_node
;
1798 struct page
*tree_page
;
1802 stable_node
= rb_entry(*new, struct stable_node
, node
);
1803 stable_node_any
= NULL
;
1804 tree_page
= chain(&stable_node_dup
, stable_node
, root
);
1805 if (!stable_node_dup
) {
1807 * Either all stable_node dups were full in
1808 * this stable_node chain, or this chain was
1809 * empty and should be rb_erased.
1811 stable_node_any
= stable_node_dup_any(stable_node
,
1813 if (!stable_node_any
) {
1814 /* rb_erase just run */
1818 * Take any of the stable_node dups page of
1819 * this stable_node chain to let the tree walk
1820 * continue. All KSM pages belonging to the
1821 * stable_node dups in a stable_node chain
1822 * have the same content and they're
1823 * wrprotected at all times. Any will work
1824 * fine to continue the walk.
1826 tree_page
= get_ksm_page(stable_node_any
, false);
1828 VM_BUG_ON(!stable_node_dup
^ !!stable_node_any
);
1831 * If we walked over a stale stable_node,
1832 * get_ksm_page() will call rb_erase() and it
1833 * may rebalance the tree from under us. So
1834 * restart the search from scratch. Returning
1835 * NULL would be safe too, but we'd generate
1836 * false negative insertions just because some
1837 * stable_node was stale.
1842 ret
= memcmp_pages(kpage
, tree_page
);
1843 put_page(tree_page
);
1847 new = &parent
->rb_left
;
1849 new = &parent
->rb_right
;
1856 stable_node_dup
= alloc_stable_node();
1857 if (!stable_node_dup
)
1860 INIT_HLIST_HEAD(&stable_node_dup
->hlist
);
1861 stable_node_dup
->kpfn
= kpfn
;
1862 set_page_stable_node(kpage
, stable_node_dup
);
1863 stable_node_dup
->rmap_hlist_len
= 0;
1864 DO_NUMA(stable_node_dup
->nid
= nid
);
1866 rb_link_node(&stable_node_dup
->node
, parent
, new);
1867 rb_insert_color(&stable_node_dup
->node
, root
);
1869 if (!is_stable_node_chain(stable_node
)) {
1870 struct stable_node
*orig
= stable_node
;
1871 /* chain is missing so create it */
1872 stable_node
= alloc_stable_node_chain(orig
, root
);
1874 free_stable_node(stable_node_dup
);
1878 stable_node_chain_add_dup(stable_node_dup
, stable_node
);
1881 return stable_node_dup
;
1885 * unstable_tree_search_insert - search for identical page,
1886 * else insert rmap_item into the unstable tree.
1888 * This function searches for a page in the unstable tree identical to the
1889 * page currently being scanned; and if no identical page is found in the
1890 * tree, we insert rmap_item as a new object into the unstable tree.
1892 * This function returns pointer to rmap_item found to be identical
1893 * to the currently scanned page, NULL otherwise.
1895 * This function does both searching and inserting, because they share
1896 * the same walking algorithm in an rbtree.
1899 struct rmap_item
*unstable_tree_search_insert(struct rmap_item
*rmap_item
,
1901 struct page
**tree_pagep
)
1903 struct rb_node
**new;
1904 struct rb_root
*root
;
1905 struct rb_node
*parent
= NULL
;
1908 nid
= get_kpfn_nid(page_to_pfn(page
));
1909 root
= root_unstable_tree
+ nid
;
1910 new = &root
->rb_node
;
1913 struct rmap_item
*tree_rmap_item
;
1914 struct page
*tree_page
;
1918 tree_rmap_item
= rb_entry(*new, struct rmap_item
, node
);
1919 tree_page
= get_mergeable_page(tree_rmap_item
);
1924 * Don't substitute a ksm page for a forked page.
1926 if (page
== tree_page
) {
1927 put_page(tree_page
);
1931 ret
= memcmp_pages(page
, tree_page
);
1935 put_page(tree_page
);
1936 new = &parent
->rb_left
;
1937 } else if (ret
> 0) {
1938 put_page(tree_page
);
1939 new = &parent
->rb_right
;
1940 } else if (!ksm_merge_across_nodes
&&
1941 page_to_nid(tree_page
) != nid
) {
1943 * If tree_page has been migrated to another NUMA node,
1944 * it will be flushed out and put in the right unstable
1945 * tree next time: only merge with it when across_nodes.
1947 put_page(tree_page
);
1950 *tree_pagep
= tree_page
;
1951 return tree_rmap_item
;
1955 rmap_item
->address
|= UNSTABLE_FLAG
;
1956 rmap_item
->address
|= (ksm_scan
.seqnr
& SEQNR_MASK
);
1957 DO_NUMA(rmap_item
->nid
= nid
);
1958 rb_link_node(&rmap_item
->node
, parent
, new);
1959 rb_insert_color(&rmap_item
->node
, root
);
1961 ksm_pages_unshared
++;
1966 * stable_tree_append - add another rmap_item to the linked list of
1967 * rmap_items hanging off a given node of the stable tree, all sharing
1968 * the same ksm page.
1970 static void stable_tree_append(struct rmap_item
*rmap_item
,
1971 struct stable_node
*stable_node
,
1972 bool max_page_sharing_bypass
)
1975 * rmap won't find this mapping if we don't insert the
1976 * rmap_item in the right stable_node
1977 * duplicate. page_migration could break later if rmap breaks,
1978 * so we can as well crash here. We really need to check for
1979 * rmap_hlist_len == STABLE_NODE_CHAIN, but we can as well check
1980 * for other negative values as an undeflow if detected here
1981 * for the first time (and not when decreasing rmap_hlist_len)
1982 * would be sign of memory corruption in the stable_node.
1984 BUG_ON(stable_node
->rmap_hlist_len
< 0);
1986 stable_node
->rmap_hlist_len
++;
1987 if (!max_page_sharing_bypass
)
1988 /* possibly non fatal but unexpected overflow, only warn */
1989 WARN_ON_ONCE(stable_node
->rmap_hlist_len
>
1990 ksm_max_page_sharing
);
1992 rmap_item
->head
= stable_node
;
1993 rmap_item
->address
|= STABLE_FLAG
;
1994 hlist_add_head(&rmap_item
->hlist
, &stable_node
->hlist
);
1996 if (rmap_item
->hlist
.next
)
1997 ksm_pages_sharing
++;
2003 * cmp_and_merge_page - first see if page can be merged into the stable tree;
2004 * if not, compare checksum to previous and if it's the same, see if page can
2005 * be inserted into the unstable tree, or merged with a page already there and
2006 * both transferred to the stable tree.
2008 * @page: the page that we are searching identical page to.
2009 * @rmap_item: the reverse mapping into the virtual address of this page
2011 static void cmp_and_merge_page(struct page
*page
, struct rmap_item
*rmap_item
)
2013 struct mm_struct
*mm
= rmap_item
->mm
;
2014 struct rmap_item
*tree_rmap_item
;
2015 struct page
*tree_page
= NULL
;
2016 struct stable_node
*stable_node
;
2018 unsigned int checksum
;
2020 bool max_page_sharing_bypass
= false;
2022 stable_node
= page_stable_node(page
);
2024 if (stable_node
->head
!= &migrate_nodes
&&
2025 get_kpfn_nid(READ_ONCE(stable_node
->kpfn
)) !=
2026 NUMA(stable_node
->nid
)) {
2027 stable_node_dup_del(stable_node
);
2028 stable_node
->head
= &migrate_nodes
;
2029 list_add(&stable_node
->list
, stable_node
->head
);
2031 if (stable_node
->head
!= &migrate_nodes
&&
2032 rmap_item
->head
== stable_node
)
2035 * If it's a KSM fork, allow it to go over the sharing limit
2038 if (!is_page_sharing_candidate(stable_node
))
2039 max_page_sharing_bypass
= true;
2042 /* We first start with searching the page inside the stable tree */
2043 kpage
= stable_tree_search(page
);
2044 if (kpage
== page
&& rmap_item
->head
== stable_node
) {
2049 remove_rmap_item_from_tree(rmap_item
);
2052 err
= try_to_merge_with_ksm_page(rmap_item
, page
, kpage
);
2055 * The page was successfully merged:
2056 * add its rmap_item to the stable tree.
2059 stable_tree_append(rmap_item
, page_stable_node(kpage
),
2060 max_page_sharing_bypass
);
2068 * If the hash value of the page has changed from the last time
2069 * we calculated it, this page is changing frequently: therefore we
2070 * don't want to insert it in the unstable tree, and we don't want
2071 * to waste our time searching for something identical to it there.
2073 checksum
= calc_checksum(page
);
2074 if (rmap_item
->oldchecksum
!= checksum
) {
2075 rmap_item
->oldchecksum
= checksum
;
2080 * Same checksum as an empty page. We attempt to merge it with the
2081 * appropriate zero page if the user enabled this via sysfs.
2083 if (ksm_use_zero_pages
&& (checksum
== zero_checksum
)) {
2084 struct vm_area_struct
*vma
;
2086 down_read(&mm
->mmap_sem
);
2087 vma
= find_mergeable_vma(mm
, rmap_item
->address
);
2088 err
= try_to_merge_one_page(vma
, page
,
2089 ZERO_PAGE(rmap_item
->address
));
2090 up_read(&mm
->mmap_sem
);
2092 * In case of failure, the page was not really empty, so we
2093 * need to continue. Otherwise we're done.
2099 unstable_tree_search_insert(rmap_item
, page
, &tree_page
);
2100 if (tree_rmap_item
) {
2103 kpage
= try_to_merge_two_pages(rmap_item
, page
,
2104 tree_rmap_item
, tree_page
);
2106 * If both pages we tried to merge belong to the same compound
2107 * page, then we actually ended up increasing the reference
2108 * count of the same compound page twice, and split_huge_page
2110 * Here we set a flag if that happened, and we use it later to
2111 * try split_huge_page again. Since we call put_page right
2112 * afterwards, the reference count will be correct and
2113 * split_huge_page should succeed.
2115 split
= PageTransCompound(page
)
2116 && compound_head(page
) == compound_head(tree_page
);
2117 put_page(tree_page
);
2120 * The pages were successfully merged: insert new
2121 * node in the stable tree and add both rmap_items.
2124 stable_node
= stable_tree_insert(kpage
);
2126 stable_tree_append(tree_rmap_item
, stable_node
,
2128 stable_tree_append(rmap_item
, stable_node
,
2134 * If we fail to insert the page into the stable tree,
2135 * we will have 2 virtual addresses that are pointing
2136 * to a ksm page left outside the stable tree,
2137 * in which case we need to break_cow on both.
2140 break_cow(tree_rmap_item
);
2141 break_cow(rmap_item
);
2145 * We are here if we tried to merge two pages and
2146 * failed because they both belonged to the same
2147 * compound page. We will split the page now, but no
2148 * merging will take place.
2149 * We do not want to add the cost of a full lock; if
2150 * the page is locked, it is better to skip it and
2151 * perhaps try again later.
2153 if (!trylock_page(page
))
2155 split_huge_page(page
);
2161 static struct rmap_item
*get_next_rmap_item(struct mm_slot
*mm_slot
,
2162 struct rmap_item
**rmap_list
,
2165 struct rmap_item
*rmap_item
;
2167 while (*rmap_list
) {
2168 rmap_item
= *rmap_list
;
2169 if ((rmap_item
->address
& PAGE_MASK
) == addr
)
2171 if (rmap_item
->address
> addr
)
2173 *rmap_list
= rmap_item
->rmap_list
;
2174 remove_rmap_item_from_tree(rmap_item
);
2175 free_rmap_item(rmap_item
);
2178 rmap_item
= alloc_rmap_item();
2180 /* It has already been zeroed */
2181 rmap_item
->mm
= mm_slot
->mm
;
2182 rmap_item
->address
= addr
;
2183 rmap_item
->rmap_list
= *rmap_list
;
2184 *rmap_list
= rmap_item
;
2189 static struct rmap_item
*scan_get_next_rmap_item(struct page
**page
)
2191 struct mm_struct
*mm
;
2192 struct mm_slot
*slot
;
2193 struct vm_area_struct
*vma
;
2194 struct rmap_item
*rmap_item
;
2197 if (list_empty(&ksm_mm_head
.mm_list
))
2200 slot
= ksm_scan
.mm_slot
;
2201 if (slot
== &ksm_mm_head
) {
2203 * A number of pages can hang around indefinitely on per-cpu
2204 * pagevecs, raised page count preventing write_protect_page
2205 * from merging them. Though it doesn't really matter much,
2206 * it is puzzling to see some stuck in pages_volatile until
2207 * other activity jostles them out, and they also prevented
2208 * LTP's KSM test from succeeding deterministically; so drain
2209 * them here (here rather than on entry to ksm_do_scan(),
2210 * so we don't IPI too often when pages_to_scan is set low).
2212 lru_add_drain_all();
2215 * Whereas stale stable_nodes on the stable_tree itself
2216 * get pruned in the regular course of stable_tree_search(),
2217 * those moved out to the migrate_nodes list can accumulate:
2218 * so prune them once before each full scan.
2220 if (!ksm_merge_across_nodes
) {
2221 struct stable_node
*stable_node
, *next
;
2224 list_for_each_entry_safe(stable_node
, next
,
2225 &migrate_nodes
, list
) {
2226 page
= get_ksm_page(stable_node
, false);
2233 for (nid
= 0; nid
< ksm_nr_node_ids
; nid
++)
2234 root_unstable_tree
[nid
] = RB_ROOT
;
2236 spin_lock(&ksm_mmlist_lock
);
2237 slot
= list_entry(slot
->mm_list
.next
, struct mm_slot
, mm_list
);
2238 ksm_scan
.mm_slot
= slot
;
2239 spin_unlock(&ksm_mmlist_lock
);
2241 * Although we tested list_empty() above, a racing __ksm_exit
2242 * of the last mm on the list may have removed it since then.
2244 if (slot
== &ksm_mm_head
)
2247 ksm_scan
.address
= 0;
2248 ksm_scan
.rmap_list
= &slot
->rmap_list
;
2252 down_read(&mm
->mmap_sem
);
2253 if (ksm_test_exit(mm
))
2256 vma
= find_vma(mm
, ksm_scan
.address
);
2258 for (; vma
; vma
= vma
->vm_next
) {
2259 if (!(vma
->vm_flags
& VM_MERGEABLE
))
2261 if (ksm_scan
.address
< vma
->vm_start
)
2262 ksm_scan
.address
= vma
->vm_start
;
2264 ksm_scan
.address
= vma
->vm_end
;
2266 while (ksm_scan
.address
< vma
->vm_end
) {
2267 if (ksm_test_exit(mm
))
2269 *page
= follow_page(vma
, ksm_scan
.address
, FOLL_GET
);
2270 if (IS_ERR_OR_NULL(*page
)) {
2271 ksm_scan
.address
+= PAGE_SIZE
;
2275 if (PageAnon(*page
)) {
2276 flush_anon_page(vma
, *page
, ksm_scan
.address
);
2277 flush_dcache_page(*page
);
2278 rmap_item
= get_next_rmap_item(slot
,
2279 ksm_scan
.rmap_list
, ksm_scan
.address
);
2281 ksm_scan
.rmap_list
=
2282 &rmap_item
->rmap_list
;
2283 ksm_scan
.address
+= PAGE_SIZE
;
2286 up_read(&mm
->mmap_sem
);
2290 ksm_scan
.address
+= PAGE_SIZE
;
2295 if (ksm_test_exit(mm
)) {
2296 ksm_scan
.address
= 0;
2297 ksm_scan
.rmap_list
= &slot
->rmap_list
;
2300 * Nuke all the rmap_items that are above this current rmap:
2301 * because there were no VM_MERGEABLE vmas with such addresses.
2303 remove_trailing_rmap_items(slot
, ksm_scan
.rmap_list
);
2305 spin_lock(&ksm_mmlist_lock
);
2306 ksm_scan
.mm_slot
= list_entry(slot
->mm_list
.next
,
2307 struct mm_slot
, mm_list
);
2308 if (ksm_scan
.address
== 0) {
2310 * We've completed a full scan of all vmas, holding mmap_sem
2311 * throughout, and found no VM_MERGEABLE: so do the same as
2312 * __ksm_exit does to remove this mm from all our lists now.
2313 * This applies either when cleaning up after __ksm_exit
2314 * (but beware: we can reach here even before __ksm_exit),
2315 * or when all VM_MERGEABLE areas have been unmapped (and
2316 * mmap_sem then protects against race with MADV_MERGEABLE).
2318 hash_del(&slot
->link
);
2319 list_del(&slot
->mm_list
);
2320 spin_unlock(&ksm_mmlist_lock
);
2323 clear_bit(MMF_VM_MERGEABLE
, &mm
->flags
);
2324 up_read(&mm
->mmap_sem
);
2327 up_read(&mm
->mmap_sem
);
2329 * up_read(&mm->mmap_sem) first because after
2330 * spin_unlock(&ksm_mmlist_lock) run, the "mm" may
2331 * already have been freed under us by __ksm_exit()
2332 * because the "mm_slot" is still hashed and
2333 * ksm_scan.mm_slot doesn't point to it anymore.
2335 spin_unlock(&ksm_mmlist_lock
);
2338 /* Repeat until we've completed scanning the whole list */
2339 slot
= ksm_scan
.mm_slot
;
2340 if (slot
!= &ksm_mm_head
)
2348 * ksm_do_scan - the ksm scanner main worker function.
2349 * @scan_npages - number of pages we want to scan before we return.
2351 static void ksm_do_scan(unsigned int scan_npages
)
2353 struct rmap_item
*rmap_item
;
2354 struct page
*uninitialized_var(page
);
2356 while (scan_npages
-- && likely(!freezing(current
))) {
2358 rmap_item
= scan_get_next_rmap_item(&page
);
2361 cmp_and_merge_page(page
, rmap_item
);
2366 static int ksmd_should_run(void)
2368 return (ksm_run
& KSM_RUN_MERGE
) && !list_empty(&ksm_mm_head
.mm_list
);
2371 static int ksm_scan_thread(void *nothing
)
2374 set_user_nice(current
, 5);
2376 while (!kthread_should_stop()) {
2377 mutex_lock(&ksm_thread_mutex
);
2378 wait_while_offlining();
2379 if (ksmd_should_run())
2380 ksm_do_scan(ksm_thread_pages_to_scan
);
2381 mutex_unlock(&ksm_thread_mutex
);
2385 if (ksmd_should_run()) {
2386 if (ksm_thread_sleep_millisecs
>= 1000)
2387 schedule_timeout_interruptible(
2388 msecs_to_jiffies(round_jiffies_relative(ksm_thread_sleep_millisecs
)));
2390 schedule_timeout_interruptible(
2391 msecs_to_jiffies(ksm_thread_sleep_millisecs
));
2393 wait_event_freezable(ksm_thread_wait
,
2394 ksmd_should_run() || kthread_should_stop());
2400 int ksm_madvise(struct vm_area_struct
*vma
, unsigned long start
,
2401 unsigned long end
, int advice
, unsigned long *vm_flags
)
2403 struct mm_struct
*mm
= vma
->vm_mm
;
2407 case MADV_MERGEABLE
:
2409 * Be somewhat over-protective for now!
2411 if (*vm_flags
& (VM_MERGEABLE
| VM_SHARED
| VM_MAYSHARE
|
2412 VM_PFNMAP
| VM_IO
| VM_DONTEXPAND
|
2413 VM_HUGETLB
| VM_MIXEDMAP
))
2414 return 0; /* just ignore the advice */
2417 if (*vm_flags
& VM_SAO
)
2421 if (!test_bit(MMF_VM_MERGEABLE
, &mm
->flags
)) {
2422 err
= __ksm_enter(mm
);
2427 *vm_flags
|= VM_MERGEABLE
;
2430 case MADV_UNMERGEABLE
:
2431 if (!(*vm_flags
& VM_MERGEABLE
))
2432 return 0; /* just ignore the advice */
2434 if (vma
->anon_vma
) {
2435 err
= unmerge_ksm_pages(vma
, start
, end
);
2440 *vm_flags
&= ~VM_MERGEABLE
;
2447 int __ksm_enter(struct mm_struct
*mm
)
2449 struct mm_slot
*mm_slot
;
2452 mm_slot
= alloc_mm_slot();
2456 /* Check ksm_run too? Would need tighter locking */
2457 needs_wakeup
= list_empty(&ksm_mm_head
.mm_list
);
2459 spin_lock(&ksm_mmlist_lock
);
2460 insert_to_mm_slots_hash(mm
, mm_slot
);
2462 * When KSM_RUN_MERGE (or KSM_RUN_STOP),
2463 * insert just behind the scanning cursor, to let the area settle
2464 * down a little; when fork is followed by immediate exec, we don't
2465 * want ksmd to waste time setting up and tearing down an rmap_list.
2467 * But when KSM_RUN_UNMERGE, it's important to insert ahead of its
2468 * scanning cursor, otherwise KSM pages in newly forked mms will be
2469 * missed: then we might as well insert at the end of the list.
2471 if (ksm_run
& KSM_RUN_UNMERGE
)
2472 list_add_tail(&mm_slot
->mm_list
, &ksm_mm_head
.mm_list
);
2474 list_add_tail(&mm_slot
->mm_list
, &ksm_scan
.mm_slot
->mm_list
);
2475 spin_unlock(&ksm_mmlist_lock
);
2477 set_bit(MMF_VM_MERGEABLE
, &mm
->flags
);
2481 wake_up_interruptible(&ksm_thread_wait
);
2486 void __ksm_exit(struct mm_struct
*mm
)
2488 struct mm_slot
*mm_slot
;
2489 int easy_to_free
= 0;
2492 * This process is exiting: if it's straightforward (as is the
2493 * case when ksmd was never running), free mm_slot immediately.
2494 * But if it's at the cursor or has rmap_items linked to it, use
2495 * mmap_sem to synchronize with any break_cows before pagetables
2496 * are freed, and leave the mm_slot on the list for ksmd to free.
2497 * Beware: ksm may already have noticed it exiting and freed the slot.
2500 spin_lock(&ksm_mmlist_lock
);
2501 mm_slot
= get_mm_slot(mm
);
2502 if (mm_slot
&& ksm_scan
.mm_slot
!= mm_slot
) {
2503 if (!mm_slot
->rmap_list
) {
2504 hash_del(&mm_slot
->link
);
2505 list_del(&mm_slot
->mm_list
);
2508 list_move(&mm_slot
->mm_list
,
2509 &ksm_scan
.mm_slot
->mm_list
);
2512 spin_unlock(&ksm_mmlist_lock
);
2515 free_mm_slot(mm_slot
);
2516 clear_bit(MMF_VM_MERGEABLE
, &mm
->flags
);
2518 } else if (mm_slot
) {
2519 down_write(&mm
->mmap_sem
);
2520 up_write(&mm
->mmap_sem
);
2524 struct page
*ksm_might_need_to_copy(struct page
*page
,
2525 struct vm_area_struct
*vma
, unsigned long address
)
2527 struct anon_vma
*anon_vma
= page_anon_vma(page
);
2528 struct page
*new_page
;
2530 if (PageKsm(page
)) {
2531 if (page_stable_node(page
) &&
2532 !(ksm_run
& KSM_RUN_UNMERGE
))
2533 return page
; /* no need to copy it */
2534 } else if (!anon_vma
) {
2535 return page
; /* no need to copy it */
2536 } else if (anon_vma
->root
== vma
->anon_vma
->root
&&
2537 page
->index
== linear_page_index(vma
, address
)) {
2538 return page
; /* still no need to copy it */
2540 if (!PageUptodate(page
))
2541 return page
; /* let do_swap_page report the error */
2543 new_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
2545 copy_user_highpage(new_page
, page
, address
, vma
);
2547 SetPageDirty(new_page
);
2548 __SetPageUptodate(new_page
);
2549 __SetPageLocked(new_page
);
2555 void rmap_walk_ksm(struct page
*page
, struct rmap_walk_control
*rwc
)
2557 struct stable_node
*stable_node
;
2558 struct rmap_item
*rmap_item
;
2559 int search_new_forks
= 0;
2561 VM_BUG_ON_PAGE(!PageKsm(page
), page
);
2564 * Rely on the page lock to protect against concurrent modifications
2565 * to that page's node of the stable tree.
2567 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
2569 stable_node
= page_stable_node(page
);
2573 hlist_for_each_entry(rmap_item
, &stable_node
->hlist
, hlist
) {
2574 struct anon_vma
*anon_vma
= rmap_item
->anon_vma
;
2575 struct anon_vma_chain
*vmac
;
2576 struct vm_area_struct
*vma
;
2579 anon_vma_lock_read(anon_vma
);
2580 anon_vma_interval_tree_foreach(vmac
, &anon_vma
->rb_root
,
2587 /* Ignore the stable/unstable/sqnr flags */
2588 addr
= rmap_item
->address
& ~KSM_FLAG_MASK
;
2590 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
2593 * Initially we examine only the vma which covers this
2594 * rmap_item; but later, if there is still work to do,
2595 * we examine covering vmas in other mms: in case they
2596 * were forked from the original since ksmd passed.
2598 if ((rmap_item
->mm
== vma
->vm_mm
) == search_new_forks
)
2601 if (rwc
->invalid_vma
&& rwc
->invalid_vma(vma
, rwc
->arg
))
2604 if (!rwc
->rmap_one(page
, vma
, addr
, rwc
->arg
)) {
2605 anon_vma_unlock_read(anon_vma
);
2608 if (rwc
->done
&& rwc
->done(page
)) {
2609 anon_vma_unlock_read(anon_vma
);
2613 anon_vma_unlock_read(anon_vma
);
2615 if (!search_new_forks
++)
2619 #ifdef CONFIG_MIGRATION
2620 void ksm_migrate_page(struct page
*newpage
, struct page
*oldpage
)
2622 struct stable_node
*stable_node
;
2624 VM_BUG_ON_PAGE(!PageLocked(oldpage
), oldpage
);
2625 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
2626 VM_BUG_ON_PAGE(newpage
->mapping
!= oldpage
->mapping
, newpage
);
2628 stable_node
= page_stable_node(newpage
);
2630 VM_BUG_ON_PAGE(stable_node
->kpfn
!= page_to_pfn(oldpage
), oldpage
);
2631 stable_node
->kpfn
= page_to_pfn(newpage
);
2633 * newpage->mapping was set in advance; now we need smp_wmb()
2634 * to make sure that the new stable_node->kpfn is visible
2635 * to get_ksm_page() before it can see that oldpage->mapping
2636 * has gone stale (or that PageSwapCache has been cleared).
2639 set_page_stable_node(oldpage
, NULL
);
2642 #endif /* CONFIG_MIGRATION */
2644 #ifdef CONFIG_MEMORY_HOTREMOVE
2645 static void wait_while_offlining(void)
2647 while (ksm_run
& KSM_RUN_OFFLINE
) {
2648 mutex_unlock(&ksm_thread_mutex
);
2649 wait_on_bit(&ksm_run
, ilog2(KSM_RUN_OFFLINE
),
2650 TASK_UNINTERRUPTIBLE
);
2651 mutex_lock(&ksm_thread_mutex
);
2655 static bool stable_node_dup_remove_range(struct stable_node
*stable_node
,
2656 unsigned long start_pfn
,
2657 unsigned long end_pfn
)
2659 if (stable_node
->kpfn
>= start_pfn
&&
2660 stable_node
->kpfn
< end_pfn
) {
2662 * Don't get_ksm_page, page has already gone:
2663 * which is why we keep kpfn instead of page*
2665 remove_node_from_stable_tree(stable_node
);
2671 static bool stable_node_chain_remove_range(struct stable_node
*stable_node
,
2672 unsigned long start_pfn
,
2673 unsigned long end_pfn
,
2674 struct rb_root
*root
)
2676 struct stable_node
*dup
;
2677 struct hlist_node
*hlist_safe
;
2679 if (!is_stable_node_chain(stable_node
)) {
2680 VM_BUG_ON(is_stable_node_dup(stable_node
));
2681 return stable_node_dup_remove_range(stable_node
, start_pfn
,
2685 hlist_for_each_entry_safe(dup
, hlist_safe
,
2686 &stable_node
->hlist
, hlist_dup
) {
2687 VM_BUG_ON(!is_stable_node_dup(dup
));
2688 stable_node_dup_remove_range(dup
, start_pfn
, end_pfn
);
2690 if (hlist_empty(&stable_node
->hlist
)) {
2691 free_stable_node_chain(stable_node
, root
);
2692 return true; /* notify caller that tree was rebalanced */
2697 static void ksm_check_stable_tree(unsigned long start_pfn
,
2698 unsigned long end_pfn
)
2700 struct stable_node
*stable_node
, *next
;
2701 struct rb_node
*node
;
2704 for (nid
= 0; nid
< ksm_nr_node_ids
; nid
++) {
2705 node
= rb_first(root_stable_tree
+ nid
);
2707 stable_node
= rb_entry(node
, struct stable_node
, node
);
2708 if (stable_node_chain_remove_range(stable_node
,
2712 node
= rb_first(root_stable_tree
+ nid
);
2714 node
= rb_next(node
);
2718 list_for_each_entry_safe(stable_node
, next
, &migrate_nodes
, list
) {
2719 if (stable_node
->kpfn
>= start_pfn
&&
2720 stable_node
->kpfn
< end_pfn
)
2721 remove_node_from_stable_tree(stable_node
);
2726 static int ksm_memory_callback(struct notifier_block
*self
,
2727 unsigned long action
, void *arg
)
2729 struct memory_notify
*mn
= arg
;
2732 case MEM_GOING_OFFLINE
:
2734 * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items()
2735 * and remove_all_stable_nodes() while memory is going offline:
2736 * it is unsafe for them to touch the stable tree at this time.
2737 * But unmerge_ksm_pages(), rmap lookups and other entry points
2738 * which do not need the ksm_thread_mutex are all safe.
2740 mutex_lock(&ksm_thread_mutex
);
2741 ksm_run
|= KSM_RUN_OFFLINE
;
2742 mutex_unlock(&ksm_thread_mutex
);
2747 * Most of the work is done by page migration; but there might
2748 * be a few stable_nodes left over, still pointing to struct
2749 * pages which have been offlined: prune those from the tree,
2750 * otherwise get_ksm_page() might later try to access a
2751 * non-existent struct page.
2753 ksm_check_stable_tree(mn
->start_pfn
,
2754 mn
->start_pfn
+ mn
->nr_pages
);
2757 case MEM_CANCEL_OFFLINE
:
2758 mutex_lock(&ksm_thread_mutex
);
2759 ksm_run
&= ~KSM_RUN_OFFLINE
;
2760 mutex_unlock(&ksm_thread_mutex
);
2762 smp_mb(); /* wake_up_bit advises this */
2763 wake_up_bit(&ksm_run
, ilog2(KSM_RUN_OFFLINE
));
2769 static void wait_while_offlining(void)
2772 #endif /* CONFIG_MEMORY_HOTREMOVE */
2776 * This all compiles without CONFIG_SYSFS, but is a waste of space.
2779 #define KSM_ATTR_RO(_name) \
2780 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2781 #define KSM_ATTR(_name) \
2782 static struct kobj_attribute _name##_attr = \
2783 __ATTR(_name, 0644, _name##_show, _name##_store)
2785 static ssize_t
sleep_millisecs_show(struct kobject
*kobj
,
2786 struct kobj_attribute
*attr
, char *buf
)
2788 return sprintf(buf
, "%u\n", ksm_thread_sleep_millisecs
);
2791 static ssize_t
sleep_millisecs_store(struct kobject
*kobj
,
2792 struct kobj_attribute
*attr
,
2793 const char *buf
, size_t count
)
2795 unsigned long msecs
;
2798 err
= kstrtoul(buf
, 10, &msecs
);
2799 if (err
|| msecs
> UINT_MAX
)
2802 ksm_thread_sleep_millisecs
= msecs
;
2806 KSM_ATTR(sleep_millisecs
);
2808 static ssize_t
pages_to_scan_show(struct kobject
*kobj
,
2809 struct kobj_attribute
*attr
, char *buf
)
2811 return sprintf(buf
, "%u\n", ksm_thread_pages_to_scan
);
2814 static ssize_t
pages_to_scan_store(struct kobject
*kobj
,
2815 struct kobj_attribute
*attr
,
2816 const char *buf
, size_t count
)
2819 unsigned long nr_pages
;
2821 err
= kstrtoul(buf
, 10, &nr_pages
);
2822 if (err
|| nr_pages
> UINT_MAX
)
2825 ksm_thread_pages_to_scan
= nr_pages
;
2829 KSM_ATTR(pages_to_scan
);
2831 static ssize_t
run_show(struct kobject
*kobj
, struct kobj_attribute
*attr
,
2834 return sprintf(buf
, "%lu\n", ksm_run
);
2837 static ssize_t
run_store(struct kobject
*kobj
, struct kobj_attribute
*attr
,
2838 const char *buf
, size_t count
)
2841 unsigned long flags
;
2843 err
= kstrtoul(buf
, 10, &flags
);
2844 if (err
|| flags
> UINT_MAX
)
2846 if (flags
> KSM_RUN_UNMERGE
)
2850 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
2851 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
2852 * breaking COW to free the pages_shared (but leaves mm_slots
2853 * on the list for when ksmd may be set running again).
2856 mutex_lock(&ksm_thread_mutex
);
2857 wait_while_offlining();
2858 if (ksm_run
!= flags
) {
2860 if (flags
& KSM_RUN_UNMERGE
) {
2861 set_current_oom_origin();
2862 err
= unmerge_and_remove_all_rmap_items();
2863 clear_current_oom_origin();
2865 ksm_run
= KSM_RUN_STOP
;
2870 mutex_unlock(&ksm_thread_mutex
);
2872 if (flags
& KSM_RUN_MERGE
)
2873 wake_up_interruptible(&ksm_thread_wait
);
2880 static ssize_t
merge_across_nodes_show(struct kobject
*kobj
,
2881 struct kobj_attribute
*attr
, char *buf
)
2883 return sprintf(buf
, "%u\n", ksm_merge_across_nodes
);
2886 static ssize_t
merge_across_nodes_store(struct kobject
*kobj
,
2887 struct kobj_attribute
*attr
,
2888 const char *buf
, size_t count
)
2893 err
= kstrtoul(buf
, 10, &knob
);
2899 mutex_lock(&ksm_thread_mutex
);
2900 wait_while_offlining();
2901 if (ksm_merge_across_nodes
!= knob
) {
2902 if (ksm_pages_shared
|| remove_all_stable_nodes())
2904 else if (root_stable_tree
== one_stable_tree
) {
2905 struct rb_root
*buf
;
2907 * This is the first time that we switch away from the
2908 * default of merging across nodes: must now allocate
2909 * a buffer to hold as many roots as may be needed.
2910 * Allocate stable and unstable together:
2911 * MAXSMP NODES_SHIFT 10 will use 16kB.
2913 buf
= kcalloc(nr_node_ids
+ nr_node_ids
, sizeof(*buf
),
2915 /* Let us assume that RB_ROOT is NULL is zero */
2919 root_stable_tree
= buf
;
2920 root_unstable_tree
= buf
+ nr_node_ids
;
2921 /* Stable tree is empty but not the unstable */
2922 root_unstable_tree
[0] = one_unstable_tree
[0];
2926 ksm_merge_across_nodes
= knob
;
2927 ksm_nr_node_ids
= knob
? 1 : nr_node_ids
;
2930 mutex_unlock(&ksm_thread_mutex
);
2932 return err
? err
: count
;
2934 KSM_ATTR(merge_across_nodes
);
2937 static ssize_t
use_zero_pages_show(struct kobject
*kobj
,
2938 struct kobj_attribute
*attr
, char *buf
)
2940 return sprintf(buf
, "%u\n", ksm_use_zero_pages
);
2942 static ssize_t
use_zero_pages_store(struct kobject
*kobj
,
2943 struct kobj_attribute
*attr
,
2944 const char *buf
, size_t count
)
2949 err
= kstrtobool(buf
, &value
);
2953 ksm_use_zero_pages
= value
;
2957 KSM_ATTR(use_zero_pages
);
2959 static ssize_t
max_page_sharing_show(struct kobject
*kobj
,
2960 struct kobj_attribute
*attr
, char *buf
)
2962 return sprintf(buf
, "%u\n", ksm_max_page_sharing
);
2965 static ssize_t
max_page_sharing_store(struct kobject
*kobj
,
2966 struct kobj_attribute
*attr
,
2967 const char *buf
, size_t count
)
2972 err
= kstrtoint(buf
, 10, &knob
);
2976 * When a KSM page is created it is shared by 2 mappings. This
2977 * being a signed comparison, it implicitly verifies it's not
2983 if (READ_ONCE(ksm_max_page_sharing
) == knob
)
2986 mutex_lock(&ksm_thread_mutex
);
2987 wait_while_offlining();
2988 if (ksm_max_page_sharing
!= knob
) {
2989 if (ksm_pages_shared
|| remove_all_stable_nodes())
2992 ksm_max_page_sharing
= knob
;
2994 mutex_unlock(&ksm_thread_mutex
);
2996 return err
? err
: count
;
2998 KSM_ATTR(max_page_sharing
);
3000 static ssize_t
pages_shared_show(struct kobject
*kobj
,
3001 struct kobj_attribute
*attr
, char *buf
)
3003 return sprintf(buf
, "%lu\n", ksm_pages_shared
);
3005 KSM_ATTR_RO(pages_shared
);
3007 static ssize_t
pages_sharing_show(struct kobject
*kobj
,
3008 struct kobj_attribute
*attr
, char *buf
)
3010 return sprintf(buf
, "%lu\n", ksm_pages_sharing
);
3012 KSM_ATTR_RO(pages_sharing
);
3014 static ssize_t
pages_unshared_show(struct kobject
*kobj
,
3015 struct kobj_attribute
*attr
, char *buf
)
3017 return sprintf(buf
, "%lu\n", ksm_pages_unshared
);
3019 KSM_ATTR_RO(pages_unshared
);
3021 static ssize_t
pages_volatile_show(struct kobject
*kobj
,
3022 struct kobj_attribute
*attr
, char *buf
)
3024 long ksm_pages_volatile
;
3026 ksm_pages_volatile
= ksm_rmap_items
- ksm_pages_shared
3027 - ksm_pages_sharing
- ksm_pages_unshared
;
3029 * It was not worth any locking to calculate that statistic,
3030 * but it might therefore sometimes be negative: conceal that.
3032 if (ksm_pages_volatile
< 0)
3033 ksm_pages_volatile
= 0;
3034 return sprintf(buf
, "%ld\n", ksm_pages_volatile
);
3036 KSM_ATTR_RO(pages_volatile
);
3038 static ssize_t
stable_node_dups_show(struct kobject
*kobj
,
3039 struct kobj_attribute
*attr
, char *buf
)
3041 return sprintf(buf
, "%lu\n", ksm_stable_node_dups
);
3043 KSM_ATTR_RO(stable_node_dups
);
3045 static ssize_t
stable_node_chains_show(struct kobject
*kobj
,
3046 struct kobj_attribute
*attr
, char *buf
)
3048 return sprintf(buf
, "%lu\n", ksm_stable_node_chains
);
3050 KSM_ATTR_RO(stable_node_chains
);
3053 stable_node_chains_prune_millisecs_show(struct kobject
*kobj
,
3054 struct kobj_attribute
*attr
,
3057 return sprintf(buf
, "%u\n", ksm_stable_node_chains_prune_millisecs
);
3061 stable_node_chains_prune_millisecs_store(struct kobject
*kobj
,
3062 struct kobj_attribute
*attr
,
3063 const char *buf
, size_t count
)
3065 unsigned long msecs
;
3068 err
= kstrtoul(buf
, 10, &msecs
);
3069 if (err
|| msecs
> UINT_MAX
)
3072 ksm_stable_node_chains_prune_millisecs
= msecs
;
3076 KSM_ATTR(stable_node_chains_prune_millisecs
);
3078 static ssize_t
full_scans_show(struct kobject
*kobj
,
3079 struct kobj_attribute
*attr
, char *buf
)
3081 return sprintf(buf
, "%lu\n", ksm_scan
.seqnr
);
3083 KSM_ATTR_RO(full_scans
);
3085 static struct attribute
*ksm_attrs
[] = {
3086 &sleep_millisecs_attr
.attr
,
3087 &pages_to_scan_attr
.attr
,
3089 &pages_shared_attr
.attr
,
3090 &pages_sharing_attr
.attr
,
3091 &pages_unshared_attr
.attr
,
3092 &pages_volatile_attr
.attr
,
3093 &full_scans_attr
.attr
,
3095 &merge_across_nodes_attr
.attr
,
3097 &max_page_sharing_attr
.attr
,
3098 &stable_node_chains_attr
.attr
,
3099 &stable_node_dups_attr
.attr
,
3100 &stable_node_chains_prune_millisecs_attr
.attr
,
3101 &use_zero_pages_attr
.attr
,
3105 static const struct attribute_group ksm_attr_group
= {
3109 #endif /* CONFIG_SYSFS */
3111 static int __init
ksm_init(void)
3113 struct task_struct
*ksm_thread
;
3116 /* The correct value depends on page size and endianness */
3117 zero_checksum
= calc_checksum(ZERO_PAGE(0));
3118 /* Default to false for backwards compatibility */
3119 ksm_use_zero_pages
= false;
3121 err
= ksm_slab_init();
3125 ksm_thread
= kthread_run(ksm_scan_thread
, NULL
, "ksmd");
3126 if (IS_ERR(ksm_thread
)) {
3127 pr_err("ksm: creating kthread failed\n");
3128 err
= PTR_ERR(ksm_thread
);
3133 err
= sysfs_create_group(mm_kobj
, &ksm_attr_group
);
3135 pr_err("ksm: register sysfs failed\n");
3136 kthread_stop(ksm_thread
);
3140 ksm_run
= KSM_RUN_MERGE
; /* no way for user to start it */
3142 #endif /* CONFIG_SYSFS */
3144 #ifdef CONFIG_MEMORY_HOTREMOVE
3145 /* There is no significance to this priority 100 */
3146 hotplug_memory_notifier(ksm_memory_callback
, 100);
3155 subsys_initcall(ksm_init
);