2 * Memory merging support.
4 * This code enables dynamic sharing of identical pages found in different
5 * memory areas, even if they are not shared by fork()
7 * Copyright (C) 2008-2009 Red Hat, Inc.
14 * This work is licensed under the terms of the GNU GPL, version 2.
17 #include <linux/errno.h>
20 #include <linux/mman.h>
21 #include <linux/sched.h>
22 #include <linux/rwsem.h>
23 #include <linux/pagemap.h>
24 #include <linux/rmap.h>
25 #include <linux/spinlock.h>
26 #include <linux/jhash.h>
27 #include <linux/delay.h>
28 #include <linux/kthread.h>
29 #include <linux/wait.h>
30 #include <linux/slab.h>
31 #include <linux/rbtree.h>
32 #include <linux/memory.h>
33 #include <linux/mmu_notifier.h>
34 #include <linux/swap.h>
35 #include <linux/ksm.h>
36 #include <linux/hashtable.h>
37 #include <linux/freezer.h>
38 #include <linux/oom.h>
39 #include <linux/numa.h>
41 #include <asm/tlbflush.h>
46 #define DO_NUMA(x) do { (x); } while (0)
49 #define DO_NUMA(x) do { } while (0)
53 * A few notes about the KSM scanning process,
54 * to make it easier to understand the data structures below:
56 * In order to reduce excessive scanning, KSM sorts the memory pages by their
57 * contents into a data structure that holds pointers to the pages' locations.
59 * Since the contents of the pages may change at any moment, KSM cannot just
60 * insert the pages into a normal sorted tree and expect it to find anything.
61 * Therefore KSM uses two data structures - the stable and the unstable tree.
63 * The stable tree holds pointers to all the merged pages (ksm pages), sorted
64 * by their contents. Because each such page is write-protected, searching on
65 * this tree is fully assured to be working (except when pages are unmapped),
66 * and therefore this tree is called the stable tree.
68 * In addition to the stable tree, KSM uses a second data structure called the
69 * unstable tree: this tree holds pointers to pages which have been found to
70 * be "unchanged for a period of time". The unstable tree sorts these pages
71 * by their contents, but since they are not write-protected, KSM cannot rely
72 * upon the unstable tree to work correctly - the unstable tree is liable to
73 * be corrupted as its contents are modified, and so it is called unstable.
75 * KSM solves this problem by several techniques:
77 * 1) The unstable tree is flushed every time KSM completes scanning all
78 * memory areas, and then the tree is rebuilt again from the beginning.
79 * 2) KSM will only insert into the unstable tree, pages whose hash value
80 * has not changed since the previous scan of all memory areas.
81 * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
82 * colors of the nodes and not on their contents, assuring that even when
83 * the tree gets "corrupted" it won't get out of balance, so scanning time
84 * remains the same (also, searching and inserting nodes in an rbtree uses
85 * the same algorithm, so we have no overhead when we flush and rebuild).
86 * 4) KSM never flushes the stable tree, which means that even if it were to
87 * take 10 attempts to find a page in the unstable tree, once it is found,
88 * it is secured in the stable tree. (When we scan a new page, we first
89 * compare it against the stable tree, and then against the unstable tree.)
91 * If the merge_across_nodes tunable is unset, then KSM maintains multiple
92 * stable trees and multiple unstable trees: one of each for each NUMA node.
96 * struct mm_slot - ksm information per mm that is being scanned
97 * @link: link to the mm_slots hash list
98 * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
99 * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
100 * @mm: the mm that this information is valid for
103 struct hlist_node link
;
104 struct list_head mm_list
;
105 struct rmap_item
*rmap_list
;
106 struct mm_struct
*mm
;
110 * struct ksm_scan - cursor for scanning
111 * @mm_slot: the current mm_slot we are scanning
112 * @address: the next address inside that to be scanned
113 * @rmap_list: link to the next rmap to be scanned in the rmap_list
114 * @seqnr: count of completed full scans (needed when removing unstable node)
116 * There is only the one ksm_scan instance of this cursor structure.
119 struct mm_slot
*mm_slot
;
120 unsigned long address
;
121 struct rmap_item
**rmap_list
;
126 * struct stable_node - node of the stable rbtree
127 * @node: rb node of this ksm page in the stable tree
128 * @head: (overlaying parent) &migrate_nodes indicates temporarily on that list
129 * @hlist_dup: linked into the stable_node->hlist with a stable_node chain
130 * @list: linked into migrate_nodes, pending placement in the proper node tree
131 * @hlist: hlist head of rmap_items using this ksm page
132 * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid)
133 * @chain_prune_time: time of the last full garbage collection
134 * @rmap_hlist_len: number of rmap_item entries in hlist or STABLE_NODE_CHAIN
135 * @nid: NUMA node id of stable tree in which linked (may not match kpfn)
139 struct rb_node node
; /* when node of stable tree */
140 struct { /* when listed for migration */
141 struct list_head
*head
;
143 struct hlist_node hlist_dup
;
144 struct list_head list
;
148 struct hlist_head hlist
;
151 unsigned long chain_prune_time
;
154 * STABLE_NODE_CHAIN can be any negative number in
155 * rmap_hlist_len negative range, but better not -1 to be able
156 * to reliably detect underflows.
158 #define STABLE_NODE_CHAIN -1024
166 * struct rmap_item - reverse mapping item for virtual addresses
167 * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
168 * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
169 * @nid: NUMA node id of unstable tree in which linked (may not match page)
170 * @mm: the memory structure this rmap_item is pointing into
171 * @address: the virtual address this rmap_item tracks (+ flags in low bits)
172 * @oldchecksum: previous checksum of the page at that virtual address
173 * @node: rb node of this rmap_item in the unstable tree
174 * @head: pointer to stable_node heading this list in the stable tree
175 * @hlist: link into hlist of rmap_items hanging off that stable_node
178 struct rmap_item
*rmap_list
;
180 struct anon_vma
*anon_vma
; /* when stable */
182 int nid
; /* when node of unstable tree */
185 struct mm_struct
*mm
;
186 unsigned long address
; /* + low bits used for flags below */
187 unsigned int oldchecksum
; /* when unstable */
189 struct rb_node node
; /* when node of unstable tree */
190 struct { /* when listed from stable tree */
191 struct stable_node
*head
;
192 struct hlist_node hlist
;
197 #define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */
198 #define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */
199 #define STABLE_FLAG 0x200 /* is listed from the stable tree */
201 /* The stable and unstable tree heads */
202 static struct rb_root one_stable_tree
[1] = { RB_ROOT
};
203 static struct rb_root one_unstable_tree
[1] = { RB_ROOT
};
204 static struct rb_root
*root_stable_tree
= one_stable_tree
;
205 static struct rb_root
*root_unstable_tree
= one_unstable_tree
;
207 /* Recently migrated nodes of stable tree, pending proper placement */
208 static LIST_HEAD(migrate_nodes
);
209 #define STABLE_NODE_DUP_HEAD ((struct list_head *)&migrate_nodes.prev)
211 #define MM_SLOTS_HASH_BITS 10
212 static DEFINE_HASHTABLE(mm_slots_hash
, MM_SLOTS_HASH_BITS
);
214 static struct mm_slot ksm_mm_head
= {
215 .mm_list
= LIST_HEAD_INIT(ksm_mm_head
.mm_list
),
217 static struct ksm_scan ksm_scan
= {
218 .mm_slot
= &ksm_mm_head
,
221 static struct kmem_cache
*rmap_item_cache
;
222 static struct kmem_cache
*stable_node_cache
;
223 static struct kmem_cache
*mm_slot_cache
;
225 /* The number of nodes in the stable tree */
226 static unsigned long ksm_pages_shared
;
228 /* The number of page slots additionally sharing those nodes */
229 static unsigned long ksm_pages_sharing
;
231 /* The number of nodes in the unstable tree */
232 static unsigned long ksm_pages_unshared
;
234 /* The number of rmap_items in use: to calculate pages_volatile */
235 static unsigned long ksm_rmap_items
;
237 /* The number of stable_node chains */
238 static unsigned long ksm_stable_node_chains
;
240 /* The number of stable_node dups linked to the stable_node chains */
241 static unsigned long ksm_stable_node_dups
;
243 /* Delay in pruning stale stable_node_dups in the stable_node_chains */
244 static int ksm_stable_node_chains_prune_millisecs
= 2000;
246 /* Maximum number of page slots sharing a stable node */
247 static int ksm_max_page_sharing
= 256;
249 /* Number of pages ksmd should scan in one batch */
250 static unsigned int ksm_thread_pages_to_scan
= 100;
252 /* Milliseconds ksmd should sleep between batches */
253 static unsigned int ksm_thread_sleep_millisecs
= 20;
256 /* Zeroed when merging across nodes is not allowed */
257 static unsigned int ksm_merge_across_nodes
= 1;
258 static int ksm_nr_node_ids
= 1;
260 #define ksm_merge_across_nodes 1U
261 #define ksm_nr_node_ids 1
264 #define KSM_RUN_STOP 0
265 #define KSM_RUN_MERGE 1
266 #define KSM_RUN_UNMERGE 2
267 #define KSM_RUN_OFFLINE 4
268 static unsigned long ksm_run
= KSM_RUN_STOP
;
269 static void wait_while_offlining(void);
271 static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait
);
272 static DEFINE_MUTEX(ksm_thread_mutex
);
273 static DEFINE_SPINLOCK(ksm_mmlist_lock
);
275 #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
276 sizeof(struct __struct), __alignof__(struct __struct),\
279 static int __init
ksm_slab_init(void)
281 rmap_item_cache
= KSM_KMEM_CACHE(rmap_item
, 0);
282 if (!rmap_item_cache
)
285 stable_node_cache
= KSM_KMEM_CACHE(stable_node
, 0);
286 if (!stable_node_cache
)
289 mm_slot_cache
= KSM_KMEM_CACHE(mm_slot
, 0);
296 kmem_cache_destroy(stable_node_cache
);
298 kmem_cache_destroy(rmap_item_cache
);
303 static void __init
ksm_slab_free(void)
305 kmem_cache_destroy(mm_slot_cache
);
306 kmem_cache_destroy(stable_node_cache
);
307 kmem_cache_destroy(rmap_item_cache
);
308 mm_slot_cache
= NULL
;
311 static __always_inline
bool is_stable_node_chain(struct stable_node
*chain
)
313 return chain
->rmap_hlist_len
== STABLE_NODE_CHAIN
;
316 static __always_inline
bool is_stable_node_dup(struct stable_node
*dup
)
318 return dup
->head
== STABLE_NODE_DUP_HEAD
;
321 static inline void stable_node_chain_add_dup(struct stable_node
*dup
,
322 struct stable_node
*chain
)
324 VM_BUG_ON(is_stable_node_dup(dup
));
325 dup
->head
= STABLE_NODE_DUP_HEAD
;
326 VM_BUG_ON(!is_stable_node_chain(chain
));
327 hlist_add_head(&dup
->hlist_dup
, &chain
->hlist
);
328 ksm_stable_node_dups
++;
331 static inline void __stable_node_dup_del(struct stable_node
*dup
)
333 VM_BUG_ON(!is_stable_node_dup(dup
));
334 hlist_del(&dup
->hlist_dup
);
335 ksm_stable_node_dups
--;
338 static inline void stable_node_dup_del(struct stable_node
*dup
)
340 VM_BUG_ON(is_stable_node_chain(dup
));
341 if (is_stable_node_dup(dup
))
342 __stable_node_dup_del(dup
);
344 rb_erase(&dup
->node
, root_stable_tree
+ NUMA(dup
->nid
));
345 #ifdef CONFIG_DEBUG_VM
350 static inline struct rmap_item
*alloc_rmap_item(void)
352 struct rmap_item
*rmap_item
;
354 rmap_item
= kmem_cache_zalloc(rmap_item_cache
, GFP_KERNEL
|
355 __GFP_NORETRY
| __GFP_NOWARN
);
361 static inline void free_rmap_item(struct rmap_item
*rmap_item
)
364 rmap_item
->mm
= NULL
; /* debug safety */
365 kmem_cache_free(rmap_item_cache
, rmap_item
);
368 static inline struct stable_node
*alloc_stable_node(void)
370 return kmem_cache_alloc(stable_node_cache
, GFP_KERNEL
);
373 static inline void free_stable_node(struct stable_node
*stable_node
)
375 VM_BUG_ON(stable_node
->rmap_hlist_len
&&
376 !is_stable_node_chain(stable_node
));
377 kmem_cache_free(stable_node_cache
, stable_node
);
380 static inline struct mm_slot
*alloc_mm_slot(void)
382 if (!mm_slot_cache
) /* initialization failed */
384 return kmem_cache_zalloc(mm_slot_cache
, GFP_KERNEL
);
387 static inline void free_mm_slot(struct mm_slot
*mm_slot
)
389 kmem_cache_free(mm_slot_cache
, mm_slot
);
392 static struct mm_slot
*get_mm_slot(struct mm_struct
*mm
)
394 struct mm_slot
*slot
;
396 hash_for_each_possible(mm_slots_hash
, slot
, link
, (unsigned long)mm
)
403 static void insert_to_mm_slots_hash(struct mm_struct
*mm
,
404 struct mm_slot
*mm_slot
)
407 hash_add(mm_slots_hash
, &mm_slot
->link
, (unsigned long)mm
);
411 * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
412 * page tables after it has passed through ksm_exit() - which, if necessary,
413 * takes mmap_sem briefly to serialize against them. ksm_exit() does not set
414 * a special flag: they can just back out as soon as mm_users goes to zero.
415 * ksm_test_exit() is used throughout to make this test for exit: in some
416 * places for correctness, in some places just to avoid unnecessary work.
418 static inline bool ksm_test_exit(struct mm_struct
*mm
)
420 return atomic_read(&mm
->mm_users
) == 0;
424 * We use break_ksm to break COW on a ksm page: it's a stripped down
426 * if (get_user_pages(current, mm, addr, 1, 1, 1, &page, NULL) == 1)
429 * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
430 * in case the application has unmapped and remapped mm,addr meanwhile.
431 * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP
432 * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
434 static int break_ksm(struct vm_area_struct
*vma
, unsigned long addr
)
441 page
= follow_page(vma
, addr
, FOLL_GET
| FOLL_MIGRATION
);
442 if (IS_ERR_OR_NULL(page
))
445 ret
= handle_mm_fault(vma
->vm_mm
, vma
, addr
,
448 ret
= VM_FAULT_WRITE
;
450 } while (!(ret
& (VM_FAULT_WRITE
| VM_FAULT_SIGBUS
| VM_FAULT_SIGSEGV
| VM_FAULT_OOM
)));
452 * We must loop because handle_mm_fault() may back out if there's
453 * any difficulty e.g. if pte accessed bit gets updated concurrently.
455 * VM_FAULT_WRITE is what we have been hoping for: it indicates that
456 * COW has been broken, even if the vma does not permit VM_WRITE;
457 * but note that a concurrent fault might break PageKsm for us.
459 * VM_FAULT_SIGBUS could occur if we race with truncation of the
460 * backing file, which also invalidates anonymous pages: that's
461 * okay, that truncation will have unmapped the PageKsm for us.
463 * VM_FAULT_OOM: at the time of writing (late July 2009), setting
464 * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
465 * current task has TIF_MEMDIE set, and will be OOM killed on return
466 * to user; and ksmd, having no mm, would never be chosen for that.
468 * But if the mm is in a limited mem_cgroup, then the fault may fail
469 * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
470 * even ksmd can fail in this way - though it's usually breaking ksm
471 * just to undo a merge it made a moment before, so unlikely to oom.
473 * That's a pity: we might therefore have more kernel pages allocated
474 * than we're counting as nodes in the stable tree; but ksm_do_scan
475 * will retry to break_cow on each pass, so should recover the page
476 * in due course. The important thing is to not let VM_MERGEABLE
477 * be cleared while any such pages might remain in the area.
479 return (ret
& VM_FAULT_OOM
) ? -ENOMEM
: 0;
482 static struct vm_area_struct
*find_mergeable_vma(struct mm_struct
*mm
,
485 struct vm_area_struct
*vma
;
486 if (ksm_test_exit(mm
))
488 vma
= find_vma(mm
, addr
);
489 if (!vma
|| vma
->vm_start
> addr
)
491 if (!(vma
->vm_flags
& VM_MERGEABLE
) || !vma
->anon_vma
)
496 static void break_cow(struct rmap_item
*rmap_item
)
498 struct mm_struct
*mm
= rmap_item
->mm
;
499 unsigned long addr
= rmap_item
->address
;
500 struct vm_area_struct
*vma
;
503 * It is not an accident that whenever we want to break COW
504 * to undo, we also need to drop a reference to the anon_vma.
506 put_anon_vma(rmap_item
->anon_vma
);
508 down_read(&mm
->mmap_sem
);
509 vma
= find_mergeable_vma(mm
, addr
);
511 break_ksm(vma
, addr
);
512 up_read(&mm
->mmap_sem
);
515 static struct page
*page_trans_compound_anon(struct page
*page
)
517 if (PageTransCompound(page
)) {
518 struct page
*head
= compound_head(page
);
520 * head may actually be splitted and freed from under
521 * us but it's ok here.
529 static struct page
*get_mergeable_page(struct rmap_item
*rmap_item
)
531 struct mm_struct
*mm
= rmap_item
->mm
;
532 unsigned long addr
= rmap_item
->address
;
533 struct vm_area_struct
*vma
;
536 down_read(&mm
->mmap_sem
);
537 vma
= find_mergeable_vma(mm
, addr
);
541 page
= follow_page(vma
, addr
, FOLL_GET
);
542 if (IS_ERR_OR_NULL(page
))
544 if (PageAnon(page
) || page_trans_compound_anon(page
)) {
545 flush_anon_page(vma
, page
, addr
);
546 flush_dcache_page(page
);
552 up_read(&mm
->mmap_sem
);
557 * This helper is used for getting right index into array of tree roots.
558 * When merge_across_nodes knob is set to 1, there are only two rb-trees for
559 * stable and unstable pages from all nodes with roots in index 0. Otherwise,
560 * every node has its own stable and unstable tree.
562 static inline int get_kpfn_nid(unsigned long kpfn
)
564 return ksm_merge_across_nodes
? 0 : NUMA(pfn_to_nid(kpfn
));
567 static struct stable_node
*alloc_stable_node_chain(struct stable_node
*dup
,
568 struct rb_root
*root
)
570 struct stable_node
*chain
= alloc_stable_node();
571 VM_BUG_ON(is_stable_node_chain(dup
));
573 INIT_HLIST_HEAD(&chain
->hlist
);
574 chain
->chain_prune_time
= jiffies
;
575 chain
->rmap_hlist_len
= STABLE_NODE_CHAIN
;
576 #if defined (CONFIG_DEBUG_VM) && defined(CONFIG_NUMA)
577 chain
->nid
= -1; /* debug */
579 ksm_stable_node_chains
++;
582 * Put the stable node chain in the first dimension of
583 * the stable tree and at the same time remove the old
586 rb_replace_node(&dup
->node
, &chain
->node
, root
);
589 * Move the old stable node to the second dimension
590 * queued in the hlist_dup. The invariant is that all
591 * dup stable_nodes in the chain->hlist point to pages
592 * that are wrprotected and have the exact same
595 stable_node_chain_add_dup(dup
, chain
);
600 static inline void free_stable_node_chain(struct stable_node
*chain
,
601 struct rb_root
*root
)
603 rb_erase(&chain
->node
, root
);
604 free_stable_node(chain
);
605 ksm_stable_node_chains
--;
608 static void remove_node_from_stable_tree(struct stable_node
*stable_node
)
610 struct rmap_item
*rmap_item
;
612 /* check it's not STABLE_NODE_CHAIN or negative */
613 BUG_ON(stable_node
->rmap_hlist_len
< 0);
615 hlist_for_each_entry(rmap_item
, &stable_node
->hlist
, hlist
) {
616 if (rmap_item
->hlist
.next
)
620 VM_BUG_ON(stable_node
->rmap_hlist_len
<= 0);
621 stable_node
->rmap_hlist_len
--;
622 put_anon_vma(rmap_item
->anon_vma
);
623 rmap_item
->address
&= PAGE_MASK
;
628 * We need the second aligned pointer of the migrate_nodes
629 * list_head to stay clear from the rb_parent_color union
630 * (aligned and different than any node) and also different
631 * from &migrate_nodes. This will verify that future list.h changes
632 * don't break STABLE_NODE_DUP_HEAD.
634 #if GCC_VERSION >= 40903 /* only recent gcc can handle it */
635 BUILD_BUG_ON(STABLE_NODE_DUP_HEAD
<= &migrate_nodes
);
636 BUILD_BUG_ON(STABLE_NODE_DUP_HEAD
>= &migrate_nodes
+ 1);
639 if (stable_node
->head
== &migrate_nodes
)
640 list_del(&stable_node
->list
);
642 stable_node_dup_del(stable_node
);
643 free_stable_node(stable_node
);
647 * get_ksm_page: checks if the page indicated by the stable node
648 * is still its ksm page, despite having held no reference to it.
649 * In which case we can trust the content of the page, and it
650 * returns the gotten page; but if the page has now been zapped,
651 * remove the stale node from the stable tree and return NULL.
652 * But beware, the stable node's page might be being migrated.
654 * You would expect the stable_node to hold a reference to the ksm page.
655 * But if it increments the page's count, swapping out has to wait for
656 * ksmd to come around again before it can free the page, which may take
657 * seconds or even minutes: much too unresponsive. So instead we use a
658 * "keyhole reference": access to the ksm page from the stable node peeps
659 * out through its keyhole to see if that page still holds the right key,
660 * pointing back to this stable node. This relies on freeing a PageAnon
661 * page to reset its page->mapping to NULL, and relies on no other use of
662 * a page to put something that might look like our key in page->mapping.
663 * is on its way to being freed; but it is an anomaly to bear in mind.
665 static struct page
*get_ksm_page(struct stable_node
*stable_node
, bool lock_it
)
668 void *expected_mapping
;
671 expected_mapping
= (void *)stable_node
+
672 (PAGE_MAPPING_ANON
| PAGE_MAPPING_KSM
);
674 kpfn
= READ_ONCE(stable_node
->kpfn
);
675 page
= pfn_to_page(kpfn
);
678 * page is computed from kpfn, so on most architectures reading
679 * page->mapping is naturally ordered after reading node->kpfn,
680 * but on Alpha we need to be more careful.
682 smp_read_barrier_depends();
683 if (READ_ONCE(page
->mapping
) != expected_mapping
)
687 * We cannot do anything with the page while its refcount is 0.
688 * Usually 0 means free, or tail of a higher-order page: in which
689 * case this node is no longer referenced, and should be freed;
690 * however, it might mean that the page is under page_freeze_refs().
691 * The __remove_mapping() case is easy, again the node is now stale;
692 * but if page is swapcache in migrate_page_move_mapping(), it might
693 * still be our page, in which case it's essential to keep the node.
695 while (!get_page_unless_zero(page
)) {
697 * Another check for page->mapping != expected_mapping would
698 * work here too. We have chosen the !PageSwapCache test to
699 * optimize the common case, when the page is or is about to
700 * be freed: PageSwapCache is cleared (under spin_lock_irq)
701 * in the freeze_refs section of __remove_mapping(); but Anon
702 * page->mapping reset to NULL later, in free_pages_prepare().
704 if (!PageSwapCache(page
))
709 if (READ_ONCE(page
->mapping
) != expected_mapping
) {
716 if (READ_ONCE(page
->mapping
) != expected_mapping
) {
726 * We come here from above when page->mapping or !PageSwapCache
727 * suggests that the node is stale; but it might be under migration.
728 * We need smp_rmb(), matching the smp_wmb() in ksm_migrate_page(),
729 * before checking whether node->kpfn has been changed.
732 if (READ_ONCE(stable_node
->kpfn
) != kpfn
)
734 remove_node_from_stable_tree(stable_node
);
739 * Removing rmap_item from stable or unstable tree.
740 * This function will clean the information from the stable/unstable tree.
742 static void remove_rmap_item_from_tree(struct rmap_item
*rmap_item
)
744 if (rmap_item
->address
& STABLE_FLAG
) {
745 struct stable_node
*stable_node
;
748 stable_node
= rmap_item
->head
;
749 page
= get_ksm_page(stable_node
, true);
753 hlist_del(&rmap_item
->hlist
);
757 if (!hlist_empty(&stable_node
->hlist
))
761 VM_BUG_ON(stable_node
->rmap_hlist_len
<= 0);
762 stable_node
->rmap_hlist_len
--;
764 put_anon_vma(rmap_item
->anon_vma
);
765 rmap_item
->address
&= PAGE_MASK
;
767 } else if (rmap_item
->address
& UNSTABLE_FLAG
) {
770 * Usually ksmd can and must skip the rb_erase, because
771 * root_unstable_tree was already reset to RB_ROOT.
772 * But be careful when an mm is exiting: do the rb_erase
773 * if this rmap_item was inserted by this scan, rather
774 * than left over from before.
776 age
= (unsigned char)(ksm_scan
.seqnr
- rmap_item
->address
);
779 rb_erase(&rmap_item
->node
,
780 root_unstable_tree
+ NUMA(rmap_item
->nid
));
781 ksm_pages_unshared
--;
782 rmap_item
->address
&= PAGE_MASK
;
785 cond_resched(); /* we're called from many long loops */
788 static void remove_trailing_rmap_items(struct mm_slot
*mm_slot
,
789 struct rmap_item
**rmap_list
)
792 struct rmap_item
*rmap_item
= *rmap_list
;
793 *rmap_list
= rmap_item
->rmap_list
;
794 remove_rmap_item_from_tree(rmap_item
);
795 free_rmap_item(rmap_item
);
800 * Though it's very tempting to unmerge rmap_items from stable tree rather
801 * than check every pte of a given vma, the locking doesn't quite work for
802 * that - an rmap_item is assigned to the stable tree after inserting ksm
803 * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing
804 * rmap_items from parent to child at fork time (so as not to waste time
805 * if exit comes before the next scan reaches it).
807 * Similarly, although we'd like to remove rmap_items (so updating counts
808 * and freeing memory) when unmerging an area, it's easier to leave that
809 * to the next pass of ksmd - consider, for example, how ksmd might be
810 * in cmp_and_merge_page on one of the rmap_items we would be removing.
812 static int unmerge_ksm_pages(struct vm_area_struct
*vma
,
813 unsigned long start
, unsigned long end
)
818 for (addr
= start
; addr
< end
&& !err
; addr
+= PAGE_SIZE
) {
819 if (ksm_test_exit(vma
->vm_mm
))
821 if (signal_pending(current
))
824 err
= break_ksm(vma
, addr
);
831 * Only called through the sysfs control interface:
833 static int remove_stable_node(struct stable_node
*stable_node
)
838 page
= get_ksm_page(stable_node
, true);
841 * get_ksm_page did remove_node_from_stable_tree itself.
846 if (WARN_ON_ONCE(page_mapped(page
))) {
848 * This should not happen: but if it does, just refuse to let
849 * merge_across_nodes be switched - there is no need to panic.
854 * The stable node did not yet appear stale to get_ksm_page(),
855 * since that allows for an unmapped ksm page to be recognized
856 * right up until it is freed; but the node is safe to remove.
857 * This page might be in a pagevec waiting to be freed,
858 * or it might be PageSwapCache (perhaps under writeback),
859 * or it might have been removed from swapcache a moment ago.
861 set_page_stable_node(page
, NULL
);
862 remove_node_from_stable_tree(stable_node
);
871 static int remove_stable_node_chain(struct stable_node
*stable_node
,
872 struct rb_root
*root
)
874 struct stable_node
*dup
;
875 struct hlist_node
*hlist_safe
;
877 if (!is_stable_node_chain(stable_node
)) {
878 VM_BUG_ON(is_stable_node_dup(stable_node
));
879 if (remove_stable_node(stable_node
))
885 hlist_for_each_entry_safe(dup
, hlist_safe
,
886 &stable_node
->hlist
, hlist_dup
) {
887 VM_BUG_ON(!is_stable_node_dup(dup
));
888 if (remove_stable_node(dup
))
891 BUG_ON(!hlist_empty(&stable_node
->hlist
));
892 free_stable_node_chain(stable_node
, root
);
896 static int remove_all_stable_nodes(void)
898 struct stable_node
*stable_node
;
899 struct list_head
*this, *next
;
903 for (nid
= 0; nid
< ksm_nr_node_ids
; nid
++) {
904 while (root_stable_tree
[nid
].rb_node
) {
905 stable_node
= rb_entry(root_stable_tree
[nid
].rb_node
,
906 struct stable_node
, node
);
907 if (remove_stable_node_chain(stable_node
,
908 root_stable_tree
+ nid
)) {
910 break; /* proceed to next nid */
915 list_for_each_safe(this, next
, &migrate_nodes
) {
916 stable_node
= list_entry(this, struct stable_node
, list
);
917 if (remove_stable_node(stable_node
))
924 static int unmerge_and_remove_all_rmap_items(void)
926 struct mm_slot
*mm_slot
;
927 struct mm_struct
*mm
;
928 struct vm_area_struct
*vma
;
931 spin_lock(&ksm_mmlist_lock
);
932 ksm_scan
.mm_slot
= list_entry(ksm_mm_head
.mm_list
.next
,
933 struct mm_slot
, mm_list
);
934 spin_unlock(&ksm_mmlist_lock
);
936 for (mm_slot
= ksm_scan
.mm_slot
;
937 mm_slot
!= &ksm_mm_head
; mm_slot
= ksm_scan
.mm_slot
) {
939 down_read(&mm
->mmap_sem
);
940 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
941 if (ksm_test_exit(mm
))
943 if (!(vma
->vm_flags
& VM_MERGEABLE
) || !vma
->anon_vma
)
945 err
= unmerge_ksm_pages(vma
,
946 vma
->vm_start
, vma
->vm_end
);
951 remove_trailing_rmap_items(mm_slot
, &mm_slot
->rmap_list
);
953 spin_lock(&ksm_mmlist_lock
);
954 ksm_scan
.mm_slot
= list_entry(mm_slot
->mm_list
.next
,
955 struct mm_slot
, mm_list
);
956 if (ksm_test_exit(mm
)) {
957 hash_del(&mm_slot
->link
);
958 list_del(&mm_slot
->mm_list
);
959 spin_unlock(&ksm_mmlist_lock
);
961 free_mm_slot(mm_slot
);
962 clear_bit(MMF_VM_MERGEABLE
, &mm
->flags
);
963 up_read(&mm
->mmap_sem
);
966 spin_unlock(&ksm_mmlist_lock
);
967 up_read(&mm
->mmap_sem
);
971 /* Clean up stable nodes, but don't worry if some are still busy */
972 remove_all_stable_nodes();
977 up_read(&mm
->mmap_sem
);
978 spin_lock(&ksm_mmlist_lock
);
979 ksm_scan
.mm_slot
= &ksm_mm_head
;
980 spin_unlock(&ksm_mmlist_lock
);
983 #endif /* CONFIG_SYSFS */
985 static u32
calc_checksum(struct page
*page
)
988 void *addr
= kmap_atomic(page
);
989 checksum
= jhash2(addr
, PAGE_SIZE
/ 4, 17);
994 static int memcmp_pages(struct page
*page1
, struct page
*page2
)
999 addr1
= kmap_atomic(page1
);
1000 addr2
= kmap_atomic(page2
);
1001 ret
= memcmp(addr1
, addr2
, PAGE_SIZE
);
1002 kunmap_atomic(addr2
);
1003 kunmap_atomic(addr1
);
1007 static inline int pages_identical(struct page
*page1
, struct page
*page2
)
1009 return !memcmp_pages(page1
, page2
);
1012 static int write_protect_page(struct vm_area_struct
*vma
, struct page
*page
,
1015 struct mm_struct
*mm
= vma
->vm_mm
;
1021 unsigned long mmun_start
; /* For mmu_notifiers */
1022 unsigned long mmun_end
; /* For mmu_notifiers */
1024 addr
= page_address_in_vma(page
, vma
);
1025 if (addr
== -EFAULT
)
1028 BUG_ON(PageTransCompound(page
));
1031 mmun_end
= addr
+ PAGE_SIZE
;
1032 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
1034 ptep
= page_check_address(page
, mm
, addr
, &ptl
, 0);
1038 if (pte_write(*ptep
) || pte_dirty(*ptep
)) {
1041 swapped
= PageSwapCache(page
);
1042 flush_cache_page(vma
, addr
, page_to_pfn(page
));
1044 * Ok this is tricky, when get_user_pages_fast() run it doesn't
1045 * take any lock, therefore the check that we are going to make
1046 * with the pagecount against the mapcount is racey and
1047 * O_DIRECT can happen right after the check.
1048 * So we clear the pte and flush the tlb before the check
1049 * this assure us that no O_DIRECT can happen after the check
1050 * or in the middle of the check.
1052 entry
= ptep_clear_flush_notify(vma
, addr
, ptep
);
1054 * Check that no O_DIRECT or similar I/O is in progress on the
1057 if (page_mapcount(page
) + 1 + swapped
!= page_count(page
)) {
1058 set_pte_at(mm
, addr
, ptep
, entry
);
1061 if (pte_dirty(entry
))
1062 set_page_dirty(page
);
1063 entry
= pte_mkclean(pte_wrprotect(entry
));
1064 set_pte_at_notify(mm
, addr
, ptep
, entry
);
1070 pte_unmap_unlock(ptep
, ptl
);
1072 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
1078 * replace_page - replace page in vma by new ksm page
1079 * @vma: vma that holds the pte pointing to page
1080 * @page: the page we are replacing by kpage
1081 * @kpage: the ksm page we replace page by
1082 * @orig_pte: the original value of the pte
1084 * Returns 0 on success, -EFAULT on failure.
1086 static int replace_page(struct vm_area_struct
*vma
, struct page
*page
,
1087 struct page
*kpage
, pte_t orig_pte
)
1089 struct mm_struct
*mm
= vma
->vm_mm
;
1095 unsigned long mmun_start
; /* For mmu_notifiers */
1096 unsigned long mmun_end
; /* For mmu_notifiers */
1098 addr
= page_address_in_vma(page
, vma
);
1099 if (addr
== -EFAULT
)
1102 pmd
= mm_find_pmd(mm
, addr
);
1107 mmun_end
= addr
+ PAGE_SIZE
;
1108 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
1110 ptep
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
1111 if (!pte_same(*ptep
, orig_pte
)) {
1112 pte_unmap_unlock(ptep
, ptl
);
1117 page_add_anon_rmap(kpage
, vma
, addr
);
1119 flush_cache_page(vma
, addr
, pte_pfn(*ptep
));
1120 ptep_clear_flush_notify(vma
, addr
, ptep
);
1121 set_pte_at_notify(mm
, addr
, ptep
, mk_pte(kpage
, vma
->vm_page_prot
));
1123 page_remove_rmap(page
);
1124 if (!page_mapped(page
))
1125 try_to_free_swap(page
);
1128 pte_unmap_unlock(ptep
, ptl
);
1131 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
1136 static int page_trans_compound_anon_split(struct page
*page
)
1139 struct page
*transhuge_head
= page_trans_compound_anon(page
);
1140 if (transhuge_head
) {
1141 /* Get the reference on the head to split it. */
1142 if (get_page_unless_zero(transhuge_head
)) {
1144 * Recheck we got the reference while the head
1145 * was still anonymous.
1147 if (PageAnon(transhuge_head
))
1148 ret
= split_huge_page(transhuge_head
);
1151 * Retry later if split_huge_page run
1155 put_page(transhuge_head
);
1157 /* Retry later if split_huge_page run from under us. */
1164 * try_to_merge_one_page - take two pages and merge them into one
1165 * @vma: the vma that holds the pte pointing to page
1166 * @page: the PageAnon page that we want to replace with kpage
1167 * @kpage: the PageKsm page that we want to map instead of page,
1168 * or NULL the first time when we want to use page as kpage.
1170 * This function returns 0 if the pages were merged, -EFAULT otherwise.
1172 static int try_to_merge_one_page(struct vm_area_struct
*vma
,
1173 struct page
*page
, struct page
*kpage
)
1175 pte_t orig_pte
= __pte(0);
1178 if (page
== kpage
) /* ksm page forked */
1181 if (PageTransCompound(page
) && page_trans_compound_anon_split(page
))
1183 BUG_ON(PageTransCompound(page
));
1184 if (!PageAnon(page
))
1188 * We need the page lock to read a stable PageSwapCache in
1189 * write_protect_page(). We use trylock_page() instead of
1190 * lock_page() because we don't want to wait here - we
1191 * prefer to continue scanning and merging different pages,
1192 * then come back to this page when it is unlocked.
1194 if (!trylock_page(page
))
1197 * If this anonymous page is mapped only here, its pte may need
1198 * to be write-protected. If it's mapped elsewhere, all of its
1199 * ptes are necessarily already write-protected. But in either
1200 * case, we need to lock and check page_count is not raised.
1202 if (write_protect_page(vma
, page
, &orig_pte
) == 0) {
1205 * While we hold page lock, upgrade page from
1206 * PageAnon+anon_vma to PageKsm+NULL stable_node:
1207 * stable_tree_insert() will update stable_node.
1209 set_page_stable_node(page
, NULL
);
1210 mark_page_accessed(page
);
1212 } else if (pages_identical(page
, kpage
))
1213 err
= replace_page(vma
, page
, kpage
, orig_pte
);
1216 if ((vma
->vm_flags
& VM_LOCKED
) && kpage
&& !err
) {
1217 munlock_vma_page(page
);
1218 if (!PageMlocked(kpage
)) {
1221 mlock_vma_page(kpage
);
1222 page
= kpage
; /* for final unlock */
1232 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
1233 * but no new kernel page is allocated: kpage must already be a ksm page.
1235 * This function returns 0 if the pages were merged, -EFAULT otherwise.
1237 static int try_to_merge_with_ksm_page(struct rmap_item
*rmap_item
,
1238 struct page
*page
, struct page
*kpage
)
1240 struct mm_struct
*mm
= rmap_item
->mm
;
1241 struct vm_area_struct
*vma
;
1244 down_read(&mm
->mmap_sem
);
1245 vma
= find_mergeable_vma(mm
, rmap_item
->address
);
1249 err
= try_to_merge_one_page(vma
, page
, kpage
);
1253 /* Unstable nid is in union with stable anon_vma: remove first */
1254 remove_rmap_item_from_tree(rmap_item
);
1256 /* Must get reference to anon_vma while still holding mmap_sem */
1257 rmap_item
->anon_vma
= vma
->anon_vma
;
1258 get_anon_vma(vma
->anon_vma
);
1260 up_read(&mm
->mmap_sem
);
1265 * try_to_merge_two_pages - take two identical pages and prepare them
1266 * to be merged into one page.
1268 * This function returns the kpage if we successfully merged two identical
1269 * pages into one ksm page, NULL otherwise.
1271 * Note that this function upgrades page to ksm page: if one of the pages
1272 * is already a ksm page, try_to_merge_with_ksm_page should be used.
1274 static struct page
*try_to_merge_two_pages(struct rmap_item
*rmap_item
,
1276 struct rmap_item
*tree_rmap_item
,
1277 struct page
*tree_page
)
1281 err
= try_to_merge_with_ksm_page(rmap_item
, page
, NULL
);
1283 err
= try_to_merge_with_ksm_page(tree_rmap_item
,
1286 * If that fails, we have a ksm page with only one pte
1287 * pointing to it: so break it.
1290 break_cow(rmap_item
);
1292 return err
? NULL
: page
;
1295 static __always_inline
1296 bool __is_page_sharing_candidate(struct stable_node
*stable_node
, int offset
)
1298 VM_BUG_ON(stable_node
->rmap_hlist_len
< 0);
1300 * Check that at least one mapping still exists, otherwise
1301 * there's no much point to merge and share with this
1302 * stable_node, as the underlying tree_page of the other
1303 * sharer is going to be freed soon.
1305 return stable_node
->rmap_hlist_len
&&
1306 stable_node
->rmap_hlist_len
+ offset
< ksm_max_page_sharing
;
1309 static __always_inline
1310 bool is_page_sharing_candidate(struct stable_node
*stable_node
)
1312 return __is_page_sharing_candidate(stable_node
, 0);
1315 struct page
*stable_node_dup(struct stable_node
**_stable_node_dup
,
1316 struct stable_node
**_stable_node
,
1317 struct rb_root
*root
,
1318 bool prune_stale_stable_nodes
)
1320 struct stable_node
*dup
, *found
= NULL
, *stable_node
= *_stable_node
;
1321 struct hlist_node
*hlist_safe
;
1322 struct page
*_tree_page
, *tree_page
= NULL
;
1324 int found_rmap_hlist_len
;
1326 if (!prune_stale_stable_nodes
||
1327 time_before(jiffies
, stable_node
->chain_prune_time
+
1329 ksm_stable_node_chains_prune_millisecs
)))
1330 prune_stale_stable_nodes
= false;
1332 stable_node
->chain_prune_time
= jiffies
;
1334 hlist_for_each_entry_safe(dup
, hlist_safe
,
1335 &stable_node
->hlist
, hlist_dup
) {
1338 * We must walk all stable_node_dup to prune the stale
1339 * stable nodes during lookup.
1341 * get_ksm_page can drop the nodes from the
1342 * stable_node->hlist if they point to freed pages
1343 * (that's why we do a _safe walk). The "dup"
1344 * stable_node parameter itself will be freed from
1345 * under us if it returns NULL.
1347 _tree_page
= get_ksm_page(dup
, false);
1351 if (is_page_sharing_candidate(dup
)) {
1353 dup
->rmap_hlist_len
> found_rmap_hlist_len
) {
1355 put_page(tree_page
);
1357 found_rmap_hlist_len
= found
->rmap_hlist_len
;
1358 tree_page
= _tree_page
;
1360 /* skip put_page for found dup */
1361 if (!prune_stale_stable_nodes
)
1366 put_page(_tree_page
);
1371 * nr is counting all dups in the chain only if
1372 * prune_stale_stable_nodes is true, otherwise we may
1373 * break the loop at nr == 1 even if there are
1376 if (prune_stale_stable_nodes
&& nr
== 1) {
1378 * If there's not just one entry it would
1379 * corrupt memory, better BUG_ON. In KSM
1380 * context with no lock held it's not even
1383 BUG_ON(stable_node
->hlist
.first
->next
);
1386 * There's just one entry and it is below the
1387 * deduplication limit so drop the chain.
1389 rb_replace_node(&stable_node
->node
, &found
->node
,
1391 free_stable_node(stable_node
);
1392 ksm_stable_node_chains
--;
1393 ksm_stable_node_dups
--;
1395 * NOTE: the caller depends on the stable_node
1396 * to be equal to stable_node_dup if the chain
1399 *_stable_node
= found
;
1401 * Just for robustneess as stable_node is
1402 * otherwise left as a stable pointer, the
1403 * compiler shall optimize it away at build
1407 } else if (stable_node
->hlist
.first
!= &found
->hlist_dup
&&
1408 __is_page_sharing_candidate(found
, 1)) {
1410 * If the found stable_node dup can accept one
1411 * more future merge (in addition to the one
1412 * that is underway) and is not at the head of
1413 * the chain, put it there so next search will
1414 * be quicker in the !prune_stale_stable_nodes
1417 * NOTE: it would be inaccurate to use nr > 1
1418 * instead of checking the hlist.first pointer
1419 * directly, because in the
1420 * prune_stale_stable_nodes case "nr" isn't
1421 * the position of the found dup in the chain,
1422 * but the total number of dups in the chain.
1424 hlist_del(&found
->hlist_dup
);
1425 hlist_add_head(&found
->hlist_dup
,
1426 &stable_node
->hlist
);
1430 *_stable_node_dup
= found
;
1434 static struct stable_node
*stable_node_dup_any(struct stable_node
*stable_node
,
1435 struct rb_root
*root
)
1437 if (!is_stable_node_chain(stable_node
))
1439 if (hlist_empty(&stable_node
->hlist
)) {
1440 free_stable_node_chain(stable_node
, root
);
1443 return hlist_entry(stable_node
->hlist
.first
,
1444 typeof(*stable_node
), hlist_dup
);
1448 * Like for get_ksm_page, this function can free the *_stable_node and
1449 * *_stable_node_dup if the returned tree_page is NULL.
1451 * It can also free and overwrite *_stable_node with the found
1452 * stable_node_dup if the chain is collapsed (in which case
1453 * *_stable_node will be equal to *_stable_node_dup like if the chain
1454 * never existed). It's up to the caller to verify tree_page is not
1455 * NULL before dereferencing *_stable_node or *_stable_node_dup.
1457 * *_stable_node_dup is really a second output parameter of this
1458 * function and will be overwritten in all cases, the caller doesn't
1459 * need to initialize it.
1461 static struct page
*__stable_node_chain(struct stable_node
**_stable_node_dup
,
1462 struct stable_node
**_stable_node
,
1463 struct rb_root
*root
,
1464 bool prune_stale_stable_nodes
)
1466 struct stable_node
*stable_node
= *_stable_node
;
1467 if (!is_stable_node_chain(stable_node
)) {
1468 if (is_page_sharing_candidate(stable_node
)) {
1469 *_stable_node_dup
= stable_node
;
1470 return get_ksm_page(stable_node
, false);
1473 * _stable_node_dup set to NULL means the stable_node
1474 * reached the ksm_max_page_sharing limit.
1476 *_stable_node_dup
= NULL
;
1479 return stable_node_dup(_stable_node_dup
, _stable_node
, root
,
1480 prune_stale_stable_nodes
);
1483 static __always_inline
struct page
*chain_prune(struct stable_node
**s_n_d
,
1484 struct stable_node
**s_n
,
1485 struct rb_root
*root
)
1487 return __stable_node_chain(s_n_d
, s_n
, root
, true);
1490 static __always_inline
struct page
*chain(struct stable_node
**s_n_d
,
1491 struct stable_node
*s_n
,
1492 struct rb_root
*root
)
1494 struct stable_node
*old_stable_node
= s_n
;
1495 struct page
*tree_page
;
1497 tree_page
= __stable_node_chain(s_n_d
, &s_n
, root
, false);
1498 /* not pruning dups so s_n cannot have changed */
1499 VM_BUG_ON(s_n
!= old_stable_node
);
1504 * stable_tree_search - search for page inside the stable tree
1506 * This function checks if there is a page inside the stable tree
1507 * with identical content to the page that we are scanning right now.
1509 * This function returns the stable tree node of identical content if found,
1512 static struct page
*stable_tree_search(struct page
*page
)
1515 struct rb_root
*root
;
1516 struct rb_node
**new;
1517 struct rb_node
*parent
;
1518 struct stable_node
*stable_node
, *stable_node_dup
, *stable_node_any
;
1519 struct stable_node
*page_node
;
1521 page_node
= page_stable_node(page
);
1522 if (page_node
&& page_node
->head
!= &migrate_nodes
) {
1523 /* ksm page forked */
1528 nid
= get_kpfn_nid(page_to_pfn(page
));
1529 root
= root_stable_tree
+ nid
;
1531 new = &root
->rb_node
;
1535 struct page
*tree_page
;
1539 stable_node
= rb_entry(*new, struct stable_node
, node
);
1540 stable_node_any
= NULL
;
1541 tree_page
= chain_prune(&stable_node_dup
, &stable_node
, root
);
1543 * NOTE: stable_node may have been freed by
1544 * chain_prune() if the returned stable_node_dup is
1545 * not NULL. stable_node_dup may have been inserted in
1546 * the rbtree instead as a regular stable_node (in
1547 * order to collapse the stable_node chain if a single
1548 * stable_node dup was found in it). In such case the
1549 * stable_node is overwritten by the calleee to point
1550 * to the stable_node_dup that was collapsed in the
1551 * stable rbtree and stable_node will be equal to
1552 * stable_node_dup like if the chain never existed.
1554 if (!stable_node_dup
) {
1556 * Either all stable_node dups were full in
1557 * this stable_node chain, or this chain was
1558 * empty and should be rb_erased.
1560 stable_node_any
= stable_node_dup_any(stable_node
,
1562 if (!stable_node_any
) {
1563 /* rb_erase just run */
1567 * Take any of the stable_node dups page of
1568 * this stable_node chain to let the tree walk
1569 * continue. All KSM pages belonging to the
1570 * stable_node dups in a stable_node chain
1571 * have the same content and they're
1572 * wrprotected at all times. Any will work
1573 * fine to continue the walk.
1575 tree_page
= get_ksm_page(stable_node_any
, false);
1577 VM_BUG_ON(!stable_node_dup
^ !!stable_node_any
);
1580 * If we walked over a stale stable_node,
1581 * get_ksm_page() will call rb_erase() and it
1582 * may rebalance the tree from under us. So
1583 * restart the search from scratch. Returning
1584 * NULL would be safe too, but we'd generate
1585 * false negative insertions just because some
1586 * stable_node was stale.
1591 ret
= memcmp_pages(page
, tree_page
);
1592 put_page(tree_page
);
1596 new = &parent
->rb_left
;
1598 new = &parent
->rb_right
;
1601 VM_BUG_ON(page_node
->head
!= &migrate_nodes
);
1603 * Test if the migrated page should be merged
1604 * into a stable node dup. If the mapcount is
1605 * 1 we can migrate it with another KSM page
1606 * without adding it to the chain.
1608 if (page_mapcount(page
) > 1)
1612 if (!stable_node_dup
) {
1614 * If the stable_node is a chain and
1615 * we got a payload match in memcmp
1616 * but we cannot merge the scanned
1617 * page in any of the existing
1618 * stable_node dups because they're
1619 * all full, we need to wait the
1620 * scanned page to find itself a match
1621 * in the unstable tree to create a
1622 * brand new KSM page to add later to
1623 * the dups of this stable_node.
1629 * Lock and unlock the stable_node's page (which
1630 * might already have been migrated) so that page
1631 * migration is sure to notice its raised count.
1632 * It would be more elegant to return stable_node
1633 * than kpage, but that involves more changes.
1635 tree_page
= get_ksm_page(stable_node_dup
, true);
1636 if (unlikely(!tree_page
))
1638 * The tree may have been rebalanced,
1639 * so re-evaluate parent and new.
1642 unlock_page(tree_page
);
1644 if (get_kpfn_nid(stable_node_dup
->kpfn
) !=
1645 NUMA(stable_node_dup
->nid
)) {
1646 put_page(tree_page
);
1656 list_del(&page_node
->list
);
1657 DO_NUMA(page_node
->nid
= nid
);
1658 rb_link_node(&page_node
->node
, parent
, new);
1659 rb_insert_color(&page_node
->node
, root
);
1661 if (is_page_sharing_candidate(page_node
)) {
1669 * If stable_node was a chain and chain_prune collapsed it,
1670 * stable_node has been updated to be the new regular
1671 * stable_node. A collapse of the chain is indistinguishable
1672 * from the case there was no chain in the stable
1673 * rbtree. Otherwise stable_node is the chain and
1674 * stable_node_dup is the dup to replace.
1676 if (stable_node_dup
== stable_node
) {
1677 VM_BUG_ON(is_stable_node_chain(stable_node_dup
));
1678 VM_BUG_ON(is_stable_node_dup(stable_node_dup
));
1679 /* there is no chain */
1681 VM_BUG_ON(page_node
->head
!= &migrate_nodes
);
1682 list_del(&page_node
->list
);
1683 DO_NUMA(page_node
->nid
= nid
);
1684 rb_replace_node(&stable_node_dup
->node
,
1687 if (is_page_sharing_candidate(page_node
))
1692 rb_erase(&stable_node_dup
->node
, root
);
1696 VM_BUG_ON(!is_stable_node_chain(stable_node
));
1697 __stable_node_dup_del(stable_node_dup
);
1699 VM_BUG_ON(page_node
->head
!= &migrate_nodes
);
1700 list_del(&page_node
->list
);
1701 DO_NUMA(page_node
->nid
= nid
);
1702 stable_node_chain_add_dup(page_node
, stable_node
);
1703 if (is_page_sharing_candidate(page_node
))
1711 stable_node_dup
->head
= &migrate_nodes
;
1712 list_add(&stable_node_dup
->list
, stable_node_dup
->head
);
1716 /* stable_node_dup could be null if it reached the limit */
1717 if (!stable_node_dup
)
1718 stable_node_dup
= stable_node_any
;
1720 * If stable_node was a chain and chain_prune collapsed it,
1721 * stable_node has been updated to be the new regular
1722 * stable_node. A collapse of the chain is indistinguishable
1723 * from the case there was no chain in the stable
1724 * rbtree. Otherwise stable_node is the chain and
1725 * stable_node_dup is the dup to replace.
1727 if (stable_node_dup
== stable_node
) {
1728 VM_BUG_ON(is_stable_node_chain(stable_node_dup
));
1729 VM_BUG_ON(is_stable_node_dup(stable_node_dup
));
1730 /* chain is missing so create it */
1731 stable_node
= alloc_stable_node_chain(stable_node_dup
,
1737 * Add this stable_node dup that was
1738 * migrated to the stable_node chain
1739 * of the current nid for this page
1742 VM_BUG_ON(!is_stable_node_chain(stable_node
));
1743 VM_BUG_ON(!is_stable_node_dup(stable_node_dup
));
1744 VM_BUG_ON(page_node
->head
!= &migrate_nodes
);
1745 list_del(&page_node
->list
);
1746 DO_NUMA(page_node
->nid
= nid
);
1747 stable_node_chain_add_dup(page_node
, stable_node
);
1752 * stable_tree_insert - insert stable tree node pointing to new ksm page
1753 * into the stable tree.
1755 * This function returns the stable tree node just allocated on success,
1758 static struct stable_node
*stable_tree_insert(struct page
*kpage
)
1762 struct rb_root
*root
;
1763 struct rb_node
**new;
1764 struct rb_node
*parent
;
1765 struct stable_node
*stable_node
, *stable_node_dup
, *stable_node_any
;
1766 bool need_chain
= false;
1768 kpfn
= page_to_pfn(kpage
);
1769 nid
= get_kpfn_nid(kpfn
);
1770 root
= root_stable_tree
+ nid
;
1773 new = &root
->rb_node
;
1776 struct page
*tree_page
;
1780 stable_node
= rb_entry(*new, struct stable_node
, node
);
1781 stable_node_any
= NULL
;
1782 tree_page
= chain(&stable_node_dup
, stable_node
, root
);
1783 if (!stable_node_dup
) {
1785 * Either all stable_node dups were full in
1786 * this stable_node chain, or this chain was
1787 * empty and should be rb_erased.
1789 stable_node_any
= stable_node_dup_any(stable_node
,
1791 if (!stable_node_any
) {
1792 /* rb_erase just run */
1796 * Take any of the stable_node dups page of
1797 * this stable_node chain to let the tree walk
1798 * continue. All KSM pages belonging to the
1799 * stable_node dups in a stable_node chain
1800 * have the same content and they're
1801 * wrprotected at all times. Any will work
1802 * fine to continue the walk.
1804 tree_page
= get_ksm_page(stable_node_any
, false);
1806 VM_BUG_ON(!stable_node_dup
^ !!stable_node_any
);
1809 * If we walked over a stale stable_node,
1810 * get_ksm_page() will call rb_erase() and it
1811 * may rebalance the tree from under us. So
1812 * restart the search from scratch. Returning
1813 * NULL would be safe too, but we'd generate
1814 * false negative insertions just because some
1815 * stable_node was stale.
1820 ret
= memcmp_pages(kpage
, tree_page
);
1821 put_page(tree_page
);
1825 new = &parent
->rb_left
;
1827 new = &parent
->rb_right
;
1834 stable_node_dup
= alloc_stable_node();
1835 if (!stable_node_dup
)
1838 INIT_HLIST_HEAD(&stable_node_dup
->hlist
);
1839 stable_node_dup
->kpfn
= kpfn
;
1840 set_page_stable_node(kpage
, stable_node_dup
);
1841 stable_node_dup
->rmap_hlist_len
= 0;
1842 DO_NUMA(stable_node_dup
->nid
= nid
);
1844 rb_link_node(&stable_node_dup
->node
, parent
, new);
1845 rb_insert_color(&stable_node_dup
->node
, root
);
1847 if (!is_stable_node_chain(stable_node
)) {
1848 struct stable_node
*orig
= stable_node
;
1849 /* chain is missing so create it */
1850 stable_node
= alloc_stable_node_chain(orig
, root
);
1852 free_stable_node(stable_node_dup
);
1856 stable_node_chain_add_dup(stable_node_dup
, stable_node
);
1859 return stable_node_dup
;
1863 * unstable_tree_search_insert - search for identical page,
1864 * else insert rmap_item into the unstable tree.
1866 * This function searches for a page in the unstable tree identical to the
1867 * page currently being scanned; and if no identical page is found in the
1868 * tree, we insert rmap_item as a new object into the unstable tree.
1870 * This function returns pointer to rmap_item found to be identical
1871 * to the currently scanned page, NULL otherwise.
1873 * This function does both searching and inserting, because they share
1874 * the same walking algorithm in an rbtree.
1877 struct rmap_item
*unstable_tree_search_insert(struct rmap_item
*rmap_item
,
1879 struct page
**tree_pagep
)
1881 struct rb_node
**new;
1882 struct rb_root
*root
;
1883 struct rb_node
*parent
= NULL
;
1886 nid
= get_kpfn_nid(page_to_pfn(page
));
1887 root
= root_unstable_tree
+ nid
;
1888 new = &root
->rb_node
;
1891 struct rmap_item
*tree_rmap_item
;
1892 struct page
*tree_page
;
1896 tree_rmap_item
= rb_entry(*new, struct rmap_item
, node
);
1897 tree_page
= get_mergeable_page(tree_rmap_item
);
1902 * Don't substitute a ksm page for a forked page.
1904 if (page
== tree_page
) {
1905 put_page(tree_page
);
1909 ret
= memcmp_pages(page
, tree_page
);
1913 put_page(tree_page
);
1914 new = &parent
->rb_left
;
1915 } else if (ret
> 0) {
1916 put_page(tree_page
);
1917 new = &parent
->rb_right
;
1918 } else if (!ksm_merge_across_nodes
&&
1919 page_to_nid(tree_page
) != nid
) {
1921 * If tree_page has been migrated to another NUMA node,
1922 * it will be flushed out and put in the right unstable
1923 * tree next time: only merge with it when across_nodes.
1925 put_page(tree_page
);
1928 *tree_pagep
= tree_page
;
1929 return tree_rmap_item
;
1933 rmap_item
->address
|= UNSTABLE_FLAG
;
1934 rmap_item
->address
|= (ksm_scan
.seqnr
& SEQNR_MASK
);
1935 DO_NUMA(rmap_item
->nid
= nid
);
1936 rb_link_node(&rmap_item
->node
, parent
, new);
1937 rb_insert_color(&rmap_item
->node
, root
);
1939 ksm_pages_unshared
++;
1944 * stable_tree_append - add another rmap_item to the linked list of
1945 * rmap_items hanging off a given node of the stable tree, all sharing
1946 * the same ksm page.
1948 static void stable_tree_append(struct rmap_item
*rmap_item
,
1949 struct stable_node
*stable_node
,
1950 bool max_page_sharing_bypass
)
1953 * rmap won't find this mapping if we don't insert the
1954 * rmap_item in the right stable_node
1955 * duplicate. page_migration could break later if rmap breaks,
1956 * so we can as well crash here. We really need to check for
1957 * rmap_hlist_len == STABLE_NODE_CHAIN, but we can as well check
1958 * for other negative values as an undeflow if detected here
1959 * for the first time (and not when decreasing rmap_hlist_len)
1960 * would be sign of memory corruption in the stable_node.
1962 BUG_ON(stable_node
->rmap_hlist_len
< 0);
1964 stable_node
->rmap_hlist_len
++;
1965 if (!max_page_sharing_bypass
)
1966 /* possibly non fatal but unexpected overflow, only warn */
1967 WARN_ON_ONCE(stable_node
->rmap_hlist_len
>
1968 ksm_max_page_sharing
);
1970 rmap_item
->head
= stable_node
;
1971 rmap_item
->address
|= STABLE_FLAG
;
1972 hlist_add_head(&rmap_item
->hlist
, &stable_node
->hlist
);
1974 if (rmap_item
->hlist
.next
)
1975 ksm_pages_sharing
++;
1981 * cmp_and_merge_page - first see if page can be merged into the stable tree;
1982 * if not, compare checksum to previous and if it's the same, see if page can
1983 * be inserted into the unstable tree, or merged with a page already there and
1984 * both transferred to the stable tree.
1986 * @page: the page that we are searching identical page to.
1987 * @rmap_item: the reverse mapping into the virtual address of this page
1989 static void cmp_and_merge_page(struct page
*page
, struct rmap_item
*rmap_item
)
1991 struct rmap_item
*tree_rmap_item
;
1992 struct page
*tree_page
= NULL
;
1993 struct stable_node
*stable_node
;
1995 unsigned int checksum
;
1997 bool max_page_sharing_bypass
= false;
1999 stable_node
= page_stable_node(page
);
2001 if (stable_node
->head
!= &migrate_nodes
&&
2002 get_kpfn_nid(READ_ONCE(stable_node
->kpfn
)) !=
2003 NUMA(stable_node
->nid
)) {
2004 stable_node_dup_del(stable_node
);
2005 stable_node
->head
= &migrate_nodes
;
2006 list_add(&stable_node
->list
, stable_node
->head
);
2008 if (stable_node
->head
!= &migrate_nodes
&&
2009 rmap_item
->head
== stable_node
)
2012 * If it's a KSM fork, allow it to go over the sharing limit
2015 if (!is_page_sharing_candidate(stable_node
))
2016 max_page_sharing_bypass
= true;
2019 /* We first start with searching the page inside the stable tree */
2020 kpage
= stable_tree_search(page
);
2021 if (kpage
== page
&& rmap_item
->head
== stable_node
) {
2026 remove_rmap_item_from_tree(rmap_item
);
2029 err
= try_to_merge_with_ksm_page(rmap_item
, page
, kpage
);
2032 * The page was successfully merged:
2033 * add its rmap_item to the stable tree.
2036 stable_tree_append(rmap_item
, page_stable_node(kpage
),
2037 max_page_sharing_bypass
);
2045 * If the hash value of the page has changed from the last time
2046 * we calculated it, this page is changing frequently: therefore we
2047 * don't want to insert it in the unstable tree, and we don't want
2048 * to waste our time searching for something identical to it there.
2050 checksum
= calc_checksum(page
);
2051 if (rmap_item
->oldchecksum
!= checksum
) {
2052 rmap_item
->oldchecksum
= checksum
;
2057 unstable_tree_search_insert(rmap_item
, page
, &tree_page
);
2058 if (tree_rmap_item
) {
2059 kpage
= try_to_merge_two_pages(rmap_item
, page
,
2060 tree_rmap_item
, tree_page
);
2061 put_page(tree_page
);
2064 * The pages were successfully merged: insert new
2065 * node in the stable tree and add both rmap_items.
2068 stable_node
= stable_tree_insert(kpage
);
2070 stable_tree_append(tree_rmap_item
, stable_node
,
2072 stable_tree_append(rmap_item
, stable_node
,
2078 * If we fail to insert the page into the stable tree,
2079 * we will have 2 virtual addresses that are pointing
2080 * to a ksm page left outside the stable tree,
2081 * in which case we need to break_cow on both.
2084 break_cow(tree_rmap_item
);
2085 break_cow(rmap_item
);
2091 static struct rmap_item
*get_next_rmap_item(struct mm_slot
*mm_slot
,
2092 struct rmap_item
**rmap_list
,
2095 struct rmap_item
*rmap_item
;
2097 while (*rmap_list
) {
2098 rmap_item
= *rmap_list
;
2099 if ((rmap_item
->address
& PAGE_MASK
) == addr
)
2101 if (rmap_item
->address
> addr
)
2103 *rmap_list
= rmap_item
->rmap_list
;
2104 remove_rmap_item_from_tree(rmap_item
);
2105 free_rmap_item(rmap_item
);
2108 rmap_item
= alloc_rmap_item();
2110 /* It has already been zeroed */
2111 rmap_item
->mm
= mm_slot
->mm
;
2112 rmap_item
->address
= addr
;
2113 rmap_item
->rmap_list
= *rmap_list
;
2114 *rmap_list
= rmap_item
;
2119 static struct rmap_item
*scan_get_next_rmap_item(struct page
**page
)
2121 struct mm_struct
*mm
;
2122 struct mm_slot
*slot
;
2123 struct vm_area_struct
*vma
;
2124 struct rmap_item
*rmap_item
;
2127 if (list_empty(&ksm_mm_head
.mm_list
))
2130 slot
= ksm_scan
.mm_slot
;
2131 if (slot
== &ksm_mm_head
) {
2133 * A number of pages can hang around indefinitely on per-cpu
2134 * pagevecs, raised page count preventing write_protect_page
2135 * from merging them. Though it doesn't really matter much,
2136 * it is puzzling to see some stuck in pages_volatile until
2137 * other activity jostles them out, and they also prevented
2138 * LTP's KSM test from succeeding deterministically; so drain
2139 * them here (here rather than on entry to ksm_do_scan(),
2140 * so we don't IPI too often when pages_to_scan is set low).
2142 lru_add_drain_all();
2145 * Whereas stale stable_nodes on the stable_tree itself
2146 * get pruned in the regular course of stable_tree_search(),
2147 * those moved out to the migrate_nodes list can accumulate:
2148 * so prune them once before each full scan.
2150 if (!ksm_merge_across_nodes
) {
2151 struct stable_node
*stable_node
;
2152 struct list_head
*this, *next
;
2155 list_for_each_safe(this, next
, &migrate_nodes
) {
2156 stable_node
= list_entry(this,
2157 struct stable_node
, list
);
2158 page
= get_ksm_page(stable_node
, false);
2165 for (nid
= 0; nid
< ksm_nr_node_ids
; nid
++)
2166 root_unstable_tree
[nid
] = RB_ROOT
;
2168 spin_lock(&ksm_mmlist_lock
);
2169 slot
= list_entry(slot
->mm_list
.next
, struct mm_slot
, mm_list
);
2170 ksm_scan
.mm_slot
= slot
;
2171 spin_unlock(&ksm_mmlist_lock
);
2173 * Although we tested list_empty() above, a racing __ksm_exit
2174 * of the last mm on the list may have removed it since then.
2176 if (slot
== &ksm_mm_head
)
2179 ksm_scan
.address
= 0;
2180 ksm_scan
.rmap_list
= &slot
->rmap_list
;
2184 down_read(&mm
->mmap_sem
);
2185 if (ksm_test_exit(mm
))
2188 vma
= find_vma(mm
, ksm_scan
.address
);
2190 for (; vma
; vma
= vma
->vm_next
) {
2191 if (!(vma
->vm_flags
& VM_MERGEABLE
))
2193 if (ksm_scan
.address
< vma
->vm_start
)
2194 ksm_scan
.address
= vma
->vm_start
;
2196 ksm_scan
.address
= vma
->vm_end
;
2198 while (ksm_scan
.address
< vma
->vm_end
) {
2199 if (ksm_test_exit(mm
))
2201 *page
= follow_page(vma
, ksm_scan
.address
, FOLL_GET
);
2202 if (IS_ERR_OR_NULL(*page
)) {
2203 ksm_scan
.address
+= PAGE_SIZE
;
2207 if (PageAnon(*page
) ||
2208 page_trans_compound_anon(*page
)) {
2209 flush_anon_page(vma
, *page
, ksm_scan
.address
);
2210 flush_dcache_page(*page
);
2211 rmap_item
= get_next_rmap_item(slot
,
2212 ksm_scan
.rmap_list
, ksm_scan
.address
);
2214 ksm_scan
.rmap_list
=
2215 &rmap_item
->rmap_list
;
2216 ksm_scan
.address
+= PAGE_SIZE
;
2219 up_read(&mm
->mmap_sem
);
2223 ksm_scan
.address
+= PAGE_SIZE
;
2228 if (ksm_test_exit(mm
)) {
2229 ksm_scan
.address
= 0;
2230 ksm_scan
.rmap_list
= &slot
->rmap_list
;
2233 * Nuke all the rmap_items that are above this current rmap:
2234 * because there were no VM_MERGEABLE vmas with such addresses.
2236 remove_trailing_rmap_items(slot
, ksm_scan
.rmap_list
);
2238 spin_lock(&ksm_mmlist_lock
);
2239 ksm_scan
.mm_slot
= list_entry(slot
->mm_list
.next
,
2240 struct mm_slot
, mm_list
);
2241 if (ksm_scan
.address
== 0) {
2243 * We've completed a full scan of all vmas, holding mmap_sem
2244 * throughout, and found no VM_MERGEABLE: so do the same as
2245 * __ksm_exit does to remove this mm from all our lists now.
2246 * This applies either when cleaning up after __ksm_exit
2247 * (but beware: we can reach here even before __ksm_exit),
2248 * or when all VM_MERGEABLE areas have been unmapped (and
2249 * mmap_sem then protects against race with MADV_MERGEABLE).
2251 hash_del(&slot
->link
);
2252 list_del(&slot
->mm_list
);
2253 spin_unlock(&ksm_mmlist_lock
);
2256 clear_bit(MMF_VM_MERGEABLE
, &mm
->flags
);
2257 up_read(&mm
->mmap_sem
);
2260 spin_unlock(&ksm_mmlist_lock
);
2261 up_read(&mm
->mmap_sem
);
2264 /* Repeat until we've completed scanning the whole list */
2265 slot
= ksm_scan
.mm_slot
;
2266 if (slot
!= &ksm_mm_head
)
2274 * ksm_do_scan - the ksm scanner main worker function.
2275 * @scan_npages - number of pages we want to scan before we return.
2277 static void ksm_do_scan(unsigned int scan_npages
)
2279 struct rmap_item
*rmap_item
;
2280 struct page
*uninitialized_var(page
);
2282 while (scan_npages
-- && likely(!freezing(current
))) {
2284 rmap_item
= scan_get_next_rmap_item(&page
);
2287 cmp_and_merge_page(page
, rmap_item
);
2292 static int ksmd_should_run(void)
2294 return (ksm_run
& KSM_RUN_MERGE
) && !list_empty(&ksm_mm_head
.mm_list
);
2297 static int ksm_scan_thread(void *nothing
)
2300 set_user_nice(current
, 5);
2302 while (!kthread_should_stop()) {
2303 mutex_lock(&ksm_thread_mutex
);
2304 wait_while_offlining();
2305 if (ksmd_should_run())
2306 ksm_do_scan(ksm_thread_pages_to_scan
);
2307 mutex_unlock(&ksm_thread_mutex
);
2311 if (ksmd_should_run()) {
2312 schedule_timeout_interruptible(
2313 msecs_to_jiffies(ksm_thread_sleep_millisecs
));
2315 wait_event_freezable(ksm_thread_wait
,
2316 ksmd_should_run() || kthread_should_stop());
2322 int ksm_madvise(struct vm_area_struct
*vma
, unsigned long start
,
2323 unsigned long end
, int advice
, unsigned long *vm_flags
)
2325 struct mm_struct
*mm
= vma
->vm_mm
;
2329 case MADV_MERGEABLE
:
2331 * Be somewhat over-protective for now!
2333 if (*vm_flags
& (VM_MERGEABLE
| VM_SHARED
| VM_MAYSHARE
|
2334 VM_PFNMAP
| VM_IO
| VM_DONTEXPAND
|
2335 VM_HUGETLB
| VM_MIXEDMAP
))
2336 return 0; /* just ignore the advice */
2339 if (*vm_flags
& VM_SAO
)
2343 if (!test_bit(MMF_VM_MERGEABLE
, &mm
->flags
)) {
2344 err
= __ksm_enter(mm
);
2349 *vm_flags
|= VM_MERGEABLE
;
2352 case MADV_UNMERGEABLE
:
2353 if (!(*vm_flags
& VM_MERGEABLE
))
2354 return 0; /* just ignore the advice */
2356 if (vma
->anon_vma
) {
2357 err
= unmerge_ksm_pages(vma
, start
, end
);
2362 *vm_flags
&= ~VM_MERGEABLE
;
2369 int __ksm_enter(struct mm_struct
*mm
)
2371 struct mm_slot
*mm_slot
;
2374 mm_slot
= alloc_mm_slot();
2378 /* Check ksm_run too? Would need tighter locking */
2379 needs_wakeup
= list_empty(&ksm_mm_head
.mm_list
);
2381 spin_lock(&ksm_mmlist_lock
);
2382 insert_to_mm_slots_hash(mm
, mm_slot
);
2384 * When KSM_RUN_MERGE (or KSM_RUN_STOP),
2385 * insert just behind the scanning cursor, to let the area settle
2386 * down a little; when fork is followed by immediate exec, we don't
2387 * want ksmd to waste time setting up and tearing down an rmap_list.
2389 * But when KSM_RUN_UNMERGE, it's important to insert ahead of its
2390 * scanning cursor, otherwise KSM pages in newly forked mms will be
2391 * missed: then we might as well insert at the end of the list.
2393 if (ksm_run
& KSM_RUN_UNMERGE
)
2394 list_add_tail(&mm_slot
->mm_list
, &ksm_mm_head
.mm_list
);
2396 list_add_tail(&mm_slot
->mm_list
, &ksm_scan
.mm_slot
->mm_list
);
2397 spin_unlock(&ksm_mmlist_lock
);
2399 set_bit(MMF_VM_MERGEABLE
, &mm
->flags
);
2400 atomic_inc(&mm
->mm_count
);
2403 wake_up_interruptible(&ksm_thread_wait
);
2408 void __ksm_exit(struct mm_struct
*mm
)
2410 struct mm_slot
*mm_slot
;
2411 int easy_to_free
= 0;
2414 * This process is exiting: if it's straightforward (as is the
2415 * case when ksmd was never running), free mm_slot immediately.
2416 * But if it's at the cursor or has rmap_items linked to it, use
2417 * mmap_sem to synchronize with any break_cows before pagetables
2418 * are freed, and leave the mm_slot on the list for ksmd to free.
2419 * Beware: ksm may already have noticed it exiting and freed the slot.
2422 spin_lock(&ksm_mmlist_lock
);
2423 mm_slot
= get_mm_slot(mm
);
2424 if (mm_slot
&& ksm_scan
.mm_slot
!= mm_slot
) {
2425 if (!mm_slot
->rmap_list
) {
2426 hash_del(&mm_slot
->link
);
2427 list_del(&mm_slot
->mm_list
);
2430 list_move(&mm_slot
->mm_list
,
2431 &ksm_scan
.mm_slot
->mm_list
);
2434 spin_unlock(&ksm_mmlist_lock
);
2437 free_mm_slot(mm_slot
);
2438 clear_bit(MMF_VM_MERGEABLE
, &mm
->flags
);
2440 } else if (mm_slot
) {
2441 down_write(&mm
->mmap_sem
);
2442 up_write(&mm
->mmap_sem
);
2446 struct page
*ksm_might_need_to_copy(struct page
*page
,
2447 struct vm_area_struct
*vma
, unsigned long address
)
2449 struct anon_vma
*anon_vma
= page_anon_vma(page
);
2450 struct page
*new_page
;
2452 if (PageKsm(page
)) {
2453 if (page_stable_node(page
) &&
2454 !(ksm_run
& KSM_RUN_UNMERGE
))
2455 return page
; /* no need to copy it */
2456 } else if (!anon_vma
) {
2457 return page
; /* no need to copy it */
2458 } else if (anon_vma
->root
== vma
->anon_vma
->root
&&
2459 page
->index
== linear_page_index(vma
, address
)) {
2460 return page
; /* still no need to copy it */
2462 if (!PageUptodate(page
))
2463 return page
; /* let do_swap_page report the error */
2465 new_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
2467 copy_user_highpage(new_page
, page
, address
, vma
);
2469 SetPageDirty(new_page
);
2470 __SetPageUptodate(new_page
);
2471 __set_page_locked(new_page
);
2477 int rmap_walk_ksm(struct page
*page
, struct rmap_walk_control
*rwc
)
2479 struct stable_node
*stable_node
;
2480 struct rmap_item
*rmap_item
;
2481 int ret
= SWAP_AGAIN
;
2482 int search_new_forks
= 0;
2484 VM_BUG_ON_PAGE(!PageKsm(page
), page
);
2487 * Rely on the page lock to protect against concurrent modifications
2488 * to that page's node of the stable tree.
2490 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
2492 stable_node
= page_stable_node(page
);
2496 hlist_for_each_entry(rmap_item
, &stable_node
->hlist
, hlist
) {
2497 struct anon_vma
*anon_vma
= rmap_item
->anon_vma
;
2498 struct anon_vma_chain
*vmac
;
2499 struct vm_area_struct
*vma
;
2502 anon_vma_lock_read(anon_vma
);
2503 anon_vma_interval_tree_foreach(vmac
, &anon_vma
->rb_root
,
2507 if (rmap_item
->address
< vma
->vm_start
||
2508 rmap_item
->address
>= vma
->vm_end
)
2511 * Initially we examine only the vma which covers this
2512 * rmap_item; but later, if there is still work to do,
2513 * we examine covering vmas in other mms: in case they
2514 * were forked from the original since ksmd passed.
2516 if ((rmap_item
->mm
== vma
->vm_mm
) == search_new_forks
)
2519 if (rwc
->invalid_vma
&& rwc
->invalid_vma(vma
, rwc
->arg
))
2522 ret
= rwc
->rmap_one(page
, vma
,
2523 rmap_item
->address
, rwc
->arg
);
2524 if (ret
!= SWAP_AGAIN
) {
2525 anon_vma_unlock_read(anon_vma
);
2528 if (rwc
->done
&& rwc
->done(page
)) {
2529 anon_vma_unlock_read(anon_vma
);
2533 anon_vma_unlock_read(anon_vma
);
2535 if (!search_new_forks
++)
2541 #ifdef CONFIG_MIGRATION
2542 void ksm_migrate_page(struct page
*newpage
, struct page
*oldpage
)
2544 struct stable_node
*stable_node
;
2546 VM_BUG_ON_PAGE(!PageLocked(oldpage
), oldpage
);
2547 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
2548 VM_BUG_ON_PAGE(newpage
->mapping
!= oldpage
->mapping
, newpage
);
2550 stable_node
= page_stable_node(newpage
);
2552 VM_BUG_ON_PAGE(stable_node
->kpfn
!= page_to_pfn(oldpage
), oldpage
);
2553 stable_node
->kpfn
= page_to_pfn(newpage
);
2555 * newpage->mapping was set in advance; now we need smp_wmb()
2556 * to make sure that the new stable_node->kpfn is visible
2557 * to get_ksm_page() before it can see that oldpage->mapping
2558 * has gone stale (or that PageSwapCache has been cleared).
2561 set_page_stable_node(oldpage
, NULL
);
2564 #endif /* CONFIG_MIGRATION */
2566 #ifdef CONFIG_MEMORY_HOTREMOVE
2567 static void wait_while_offlining(void)
2569 while (ksm_run
& KSM_RUN_OFFLINE
) {
2570 mutex_unlock(&ksm_thread_mutex
);
2571 wait_on_bit(&ksm_run
, ilog2(KSM_RUN_OFFLINE
),
2572 TASK_UNINTERRUPTIBLE
);
2573 mutex_lock(&ksm_thread_mutex
);
2577 static bool stable_node_dup_remove_range(struct stable_node
*stable_node
,
2578 unsigned long start_pfn
,
2579 unsigned long end_pfn
)
2581 if (stable_node
->kpfn
>= start_pfn
&&
2582 stable_node
->kpfn
< end_pfn
) {
2584 * Don't get_ksm_page, page has already gone:
2585 * which is why we keep kpfn instead of page*
2587 remove_node_from_stable_tree(stable_node
);
2593 static bool stable_node_chain_remove_range(struct stable_node
*stable_node
,
2594 unsigned long start_pfn
,
2595 unsigned long end_pfn
,
2596 struct rb_root
*root
)
2598 struct stable_node
*dup
;
2599 struct hlist_node
*hlist_safe
;
2601 if (!is_stable_node_chain(stable_node
)) {
2602 VM_BUG_ON(is_stable_node_dup(stable_node
));
2603 return stable_node_dup_remove_range(stable_node
, start_pfn
,
2607 hlist_for_each_entry_safe(dup
, hlist_safe
,
2608 &stable_node
->hlist
, hlist_dup
) {
2609 VM_BUG_ON(!is_stable_node_dup(dup
));
2610 stable_node_dup_remove_range(dup
, start_pfn
, end_pfn
);
2612 if (hlist_empty(&stable_node
->hlist
)) {
2613 free_stable_node_chain(stable_node
, root
);
2614 return true; /* notify caller that tree was rebalanced */
2619 static void ksm_check_stable_tree(unsigned long start_pfn
,
2620 unsigned long end_pfn
)
2622 struct stable_node
*stable_node
;
2623 struct list_head
*this, *next
;
2624 struct rb_node
*node
;
2627 for (nid
= 0; nid
< ksm_nr_node_ids
; nid
++) {
2628 node
= rb_first(root_stable_tree
+ nid
);
2630 stable_node
= rb_entry(node
, struct stable_node
, node
);
2631 if (stable_node_chain_remove_range(stable_node
,
2635 node
= rb_first(root_stable_tree
+ nid
);
2637 node
= rb_next(node
);
2641 list_for_each_safe(this, next
, &migrate_nodes
) {
2642 stable_node
= list_entry(this, struct stable_node
, list
);
2643 if (stable_node
->kpfn
>= start_pfn
&&
2644 stable_node
->kpfn
< end_pfn
)
2645 remove_node_from_stable_tree(stable_node
);
2650 static int ksm_memory_callback(struct notifier_block
*self
,
2651 unsigned long action
, void *arg
)
2653 struct memory_notify
*mn
= arg
;
2656 case MEM_GOING_OFFLINE
:
2658 * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items()
2659 * and remove_all_stable_nodes() while memory is going offline:
2660 * it is unsafe for them to touch the stable tree at this time.
2661 * But unmerge_ksm_pages(), rmap lookups and other entry points
2662 * which do not need the ksm_thread_mutex are all safe.
2664 mutex_lock(&ksm_thread_mutex
);
2665 ksm_run
|= KSM_RUN_OFFLINE
;
2666 mutex_unlock(&ksm_thread_mutex
);
2671 * Most of the work is done by page migration; but there might
2672 * be a few stable_nodes left over, still pointing to struct
2673 * pages which have been offlined: prune those from the tree,
2674 * otherwise get_ksm_page() might later try to access a
2675 * non-existent struct page.
2677 ksm_check_stable_tree(mn
->start_pfn
,
2678 mn
->start_pfn
+ mn
->nr_pages
);
2681 case MEM_CANCEL_OFFLINE
:
2682 mutex_lock(&ksm_thread_mutex
);
2683 ksm_run
&= ~KSM_RUN_OFFLINE
;
2684 mutex_unlock(&ksm_thread_mutex
);
2686 smp_mb(); /* wake_up_bit advises this */
2687 wake_up_bit(&ksm_run
, ilog2(KSM_RUN_OFFLINE
));
2693 static void wait_while_offlining(void)
2696 #endif /* CONFIG_MEMORY_HOTREMOVE */
2700 * This all compiles without CONFIG_SYSFS, but is a waste of space.
2703 #define KSM_ATTR_RO(_name) \
2704 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2705 #define KSM_ATTR(_name) \
2706 static struct kobj_attribute _name##_attr = \
2707 __ATTR(_name, 0644, _name##_show, _name##_store)
2709 static ssize_t
sleep_millisecs_show(struct kobject
*kobj
,
2710 struct kobj_attribute
*attr
, char *buf
)
2712 return sprintf(buf
, "%u\n", ksm_thread_sleep_millisecs
);
2715 static ssize_t
sleep_millisecs_store(struct kobject
*kobj
,
2716 struct kobj_attribute
*attr
,
2717 const char *buf
, size_t count
)
2719 unsigned long msecs
;
2722 err
= kstrtoul(buf
, 10, &msecs
);
2723 if (err
|| msecs
> UINT_MAX
)
2726 ksm_thread_sleep_millisecs
= msecs
;
2730 KSM_ATTR(sleep_millisecs
);
2732 static ssize_t
pages_to_scan_show(struct kobject
*kobj
,
2733 struct kobj_attribute
*attr
, char *buf
)
2735 return sprintf(buf
, "%u\n", ksm_thread_pages_to_scan
);
2738 static ssize_t
pages_to_scan_store(struct kobject
*kobj
,
2739 struct kobj_attribute
*attr
,
2740 const char *buf
, size_t count
)
2743 unsigned long nr_pages
;
2745 err
= kstrtoul(buf
, 10, &nr_pages
);
2746 if (err
|| nr_pages
> UINT_MAX
)
2749 ksm_thread_pages_to_scan
= nr_pages
;
2753 KSM_ATTR(pages_to_scan
);
2755 static ssize_t
run_show(struct kobject
*kobj
, struct kobj_attribute
*attr
,
2758 return sprintf(buf
, "%lu\n", ksm_run
);
2761 static ssize_t
run_store(struct kobject
*kobj
, struct kobj_attribute
*attr
,
2762 const char *buf
, size_t count
)
2765 unsigned long flags
;
2767 err
= kstrtoul(buf
, 10, &flags
);
2768 if (err
|| flags
> UINT_MAX
)
2770 if (flags
> KSM_RUN_UNMERGE
)
2774 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
2775 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
2776 * breaking COW to free the pages_shared (but leaves mm_slots
2777 * on the list for when ksmd may be set running again).
2780 mutex_lock(&ksm_thread_mutex
);
2781 wait_while_offlining();
2782 if (ksm_run
!= flags
) {
2784 if (flags
& KSM_RUN_UNMERGE
) {
2785 set_current_oom_origin();
2786 err
= unmerge_and_remove_all_rmap_items();
2787 clear_current_oom_origin();
2789 ksm_run
= KSM_RUN_STOP
;
2794 mutex_unlock(&ksm_thread_mutex
);
2796 if (flags
& KSM_RUN_MERGE
)
2797 wake_up_interruptible(&ksm_thread_wait
);
2804 static ssize_t
merge_across_nodes_show(struct kobject
*kobj
,
2805 struct kobj_attribute
*attr
, char *buf
)
2807 return sprintf(buf
, "%u\n", ksm_merge_across_nodes
);
2810 static ssize_t
merge_across_nodes_store(struct kobject
*kobj
,
2811 struct kobj_attribute
*attr
,
2812 const char *buf
, size_t count
)
2817 err
= kstrtoul(buf
, 10, &knob
);
2823 mutex_lock(&ksm_thread_mutex
);
2824 wait_while_offlining();
2825 if (ksm_merge_across_nodes
!= knob
) {
2826 if (ksm_pages_shared
|| remove_all_stable_nodes())
2828 else if (root_stable_tree
== one_stable_tree
) {
2829 struct rb_root
*buf
;
2831 * This is the first time that we switch away from the
2832 * default of merging across nodes: must now allocate
2833 * a buffer to hold as many roots as may be needed.
2834 * Allocate stable and unstable together:
2835 * MAXSMP NODES_SHIFT 10 will use 16kB.
2837 buf
= kcalloc(nr_node_ids
+ nr_node_ids
, sizeof(*buf
),
2839 /* Let us assume that RB_ROOT is NULL is zero */
2843 root_stable_tree
= buf
;
2844 root_unstable_tree
= buf
+ nr_node_ids
;
2845 /* Stable tree is empty but not the unstable */
2846 root_unstable_tree
[0] = one_unstable_tree
[0];
2850 ksm_merge_across_nodes
= knob
;
2851 ksm_nr_node_ids
= knob
? 1 : nr_node_ids
;
2854 mutex_unlock(&ksm_thread_mutex
);
2856 return err
? err
: count
;
2858 KSM_ATTR(merge_across_nodes
);
2861 static ssize_t
max_page_sharing_show(struct kobject
*kobj
,
2862 struct kobj_attribute
*attr
, char *buf
)
2864 return sprintf(buf
, "%u\n", ksm_max_page_sharing
);
2867 static ssize_t
max_page_sharing_store(struct kobject
*kobj
,
2868 struct kobj_attribute
*attr
,
2869 const char *buf
, size_t count
)
2874 err
= kstrtoint(buf
, 10, &knob
);
2878 * When a KSM page is created it is shared by 2 mappings. This
2879 * being a signed comparison, it implicitly verifies it's not
2885 if (READ_ONCE(ksm_max_page_sharing
) == knob
)
2888 mutex_lock(&ksm_thread_mutex
);
2889 wait_while_offlining();
2890 if (ksm_max_page_sharing
!= knob
) {
2891 if (ksm_pages_shared
|| remove_all_stable_nodes())
2894 ksm_max_page_sharing
= knob
;
2896 mutex_unlock(&ksm_thread_mutex
);
2898 return err
? err
: count
;
2900 KSM_ATTR(max_page_sharing
);
2902 static ssize_t
pages_shared_show(struct kobject
*kobj
,
2903 struct kobj_attribute
*attr
, char *buf
)
2905 return sprintf(buf
, "%lu\n", ksm_pages_shared
);
2907 KSM_ATTR_RO(pages_shared
);
2909 static ssize_t
pages_sharing_show(struct kobject
*kobj
,
2910 struct kobj_attribute
*attr
, char *buf
)
2912 return sprintf(buf
, "%lu\n", ksm_pages_sharing
);
2914 KSM_ATTR_RO(pages_sharing
);
2916 static ssize_t
pages_unshared_show(struct kobject
*kobj
,
2917 struct kobj_attribute
*attr
, char *buf
)
2919 return sprintf(buf
, "%lu\n", ksm_pages_unshared
);
2921 KSM_ATTR_RO(pages_unshared
);
2923 static ssize_t
pages_volatile_show(struct kobject
*kobj
,
2924 struct kobj_attribute
*attr
, char *buf
)
2926 long ksm_pages_volatile
;
2928 ksm_pages_volatile
= ksm_rmap_items
- ksm_pages_shared
2929 - ksm_pages_sharing
- ksm_pages_unshared
;
2931 * It was not worth any locking to calculate that statistic,
2932 * but it might therefore sometimes be negative: conceal that.
2934 if (ksm_pages_volatile
< 0)
2935 ksm_pages_volatile
= 0;
2936 return sprintf(buf
, "%ld\n", ksm_pages_volatile
);
2938 KSM_ATTR_RO(pages_volatile
);
2940 static ssize_t
stable_node_dups_show(struct kobject
*kobj
,
2941 struct kobj_attribute
*attr
, char *buf
)
2943 return sprintf(buf
, "%lu\n", ksm_stable_node_dups
);
2945 KSM_ATTR_RO(stable_node_dups
);
2947 static ssize_t
stable_node_chains_show(struct kobject
*kobj
,
2948 struct kobj_attribute
*attr
, char *buf
)
2950 return sprintf(buf
, "%lu\n", ksm_stable_node_chains
);
2952 KSM_ATTR_RO(stable_node_chains
);
2955 stable_node_chains_prune_millisecs_show(struct kobject
*kobj
,
2956 struct kobj_attribute
*attr
,
2959 return sprintf(buf
, "%u\n", ksm_stable_node_chains_prune_millisecs
);
2963 stable_node_chains_prune_millisecs_store(struct kobject
*kobj
,
2964 struct kobj_attribute
*attr
,
2965 const char *buf
, size_t count
)
2967 unsigned long msecs
;
2970 err
= kstrtoul(buf
, 10, &msecs
);
2971 if (err
|| msecs
> UINT_MAX
)
2974 ksm_stable_node_chains_prune_millisecs
= msecs
;
2978 KSM_ATTR(stable_node_chains_prune_millisecs
);
2980 static ssize_t
full_scans_show(struct kobject
*kobj
,
2981 struct kobj_attribute
*attr
, char *buf
)
2983 return sprintf(buf
, "%lu\n", ksm_scan
.seqnr
);
2985 KSM_ATTR_RO(full_scans
);
2987 static struct attribute
*ksm_attrs
[] = {
2988 &sleep_millisecs_attr
.attr
,
2989 &pages_to_scan_attr
.attr
,
2991 &pages_shared_attr
.attr
,
2992 &pages_sharing_attr
.attr
,
2993 &pages_unshared_attr
.attr
,
2994 &pages_volatile_attr
.attr
,
2995 &full_scans_attr
.attr
,
2997 &merge_across_nodes_attr
.attr
,
2999 &max_page_sharing_attr
.attr
,
3000 &stable_node_chains_attr
.attr
,
3001 &stable_node_dups_attr
.attr
,
3002 &stable_node_chains_prune_millisecs_attr
.attr
,
3006 static struct attribute_group ksm_attr_group
= {
3010 #endif /* CONFIG_SYSFS */
3012 static int __init
ksm_init(void)
3014 struct task_struct
*ksm_thread
;
3017 err
= ksm_slab_init();
3021 ksm_thread
= kthread_run(ksm_scan_thread
, NULL
, "ksmd");
3022 if (IS_ERR(ksm_thread
)) {
3023 pr_err("ksm: creating kthread failed\n");
3024 err
= PTR_ERR(ksm_thread
);
3029 err
= sysfs_create_group(mm_kobj
, &ksm_attr_group
);
3031 pr_err("ksm: register sysfs failed\n");
3032 kthread_stop(ksm_thread
);
3036 ksm_run
= KSM_RUN_MERGE
; /* no way for user to start it */
3038 #endif /* CONFIG_SYSFS */
3040 #ifdef CONFIG_MEMORY_HOTREMOVE
3041 /* There is no significance to this priority 100 */
3042 hotplug_memory_notifier(ksm_memory_callback
, 100);
3051 subsys_initcall(ksm_init
);