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 * @list: linked into migrate_nodes, pending placement in the proper node tree
130 * @hlist: hlist head of rmap_items using this ksm page
131 * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid)
132 * @nid: NUMA node id of stable tree in which linked (may not match kpfn)
136 struct rb_node node
; /* when node of stable tree */
137 struct { /* when listed for migration */
138 struct list_head
*head
;
139 struct list_head list
;
142 struct hlist_head hlist
;
150 * struct rmap_item - reverse mapping item for virtual addresses
151 * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
152 * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
153 * @nid: NUMA node id of unstable tree in which linked (may not match page)
154 * @mm: the memory structure this rmap_item is pointing into
155 * @address: the virtual address this rmap_item tracks (+ flags in low bits)
156 * @oldchecksum: previous checksum of the page at that virtual address
157 * @node: rb node of this rmap_item in the unstable tree
158 * @head: pointer to stable_node heading this list in the stable tree
159 * @hlist: link into hlist of rmap_items hanging off that stable_node
162 struct rmap_item
*rmap_list
;
164 struct anon_vma
*anon_vma
; /* when stable */
166 int nid
; /* when node of unstable tree */
169 struct mm_struct
*mm
;
170 unsigned long address
; /* + low bits used for flags below */
171 unsigned int oldchecksum
; /* when unstable */
173 struct rb_node node
; /* when node of unstable tree */
174 struct { /* when listed from stable tree */
175 struct stable_node
*head
;
176 struct hlist_node hlist
;
181 #define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */
182 #define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */
183 #define STABLE_FLAG 0x200 /* is listed from the stable tree */
185 /* The stable and unstable tree heads */
186 static struct rb_root one_stable_tree
[1] = { RB_ROOT
};
187 static struct rb_root one_unstable_tree
[1] = { RB_ROOT
};
188 static struct rb_root
*root_stable_tree
= one_stable_tree
;
189 static struct rb_root
*root_unstable_tree
= one_unstable_tree
;
191 /* Recently migrated nodes of stable tree, pending proper placement */
192 static LIST_HEAD(migrate_nodes
);
194 #define MM_SLOTS_HASH_BITS 10
195 static DEFINE_HASHTABLE(mm_slots_hash
, MM_SLOTS_HASH_BITS
);
197 static struct mm_slot ksm_mm_head
= {
198 .mm_list
= LIST_HEAD_INIT(ksm_mm_head
.mm_list
),
200 static struct ksm_scan ksm_scan
= {
201 .mm_slot
= &ksm_mm_head
,
204 static struct kmem_cache
*rmap_item_cache
;
205 static struct kmem_cache
*stable_node_cache
;
206 static struct kmem_cache
*mm_slot_cache
;
208 /* The number of nodes in the stable tree */
209 static unsigned long ksm_pages_shared
;
211 /* The number of page slots additionally sharing those nodes */
212 static unsigned long ksm_pages_sharing
;
214 /* The number of nodes in the unstable tree */
215 static unsigned long ksm_pages_unshared
;
217 /* The number of rmap_items in use: to calculate pages_volatile */
218 static unsigned long ksm_rmap_items
;
220 /* Number of pages ksmd should scan in one batch */
221 static unsigned int ksm_thread_pages_to_scan
= 100;
223 /* Milliseconds ksmd should sleep between batches */
224 static unsigned int ksm_thread_sleep_millisecs
= 20;
226 /* Checksum of an empty (zeroed) page */
227 static unsigned int zero_checksum __read_mostly
;
229 /* Whether to merge empty (zeroed) pages with actual zero pages */
230 static bool ksm_use_zero_pages __read_mostly
;
233 /* Zeroed when merging across nodes is not allowed */
234 static unsigned int ksm_merge_across_nodes
= 1;
235 static int ksm_nr_node_ids
= 1;
237 #define ksm_merge_across_nodes 1U
238 #define ksm_nr_node_ids 1
241 #define KSM_RUN_STOP 0
242 #define KSM_RUN_MERGE 1
243 #define KSM_RUN_UNMERGE 2
244 #define KSM_RUN_OFFLINE 4
245 static unsigned long ksm_run
= KSM_RUN_STOP
;
246 static void wait_while_offlining(void);
248 static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait
);
249 static DEFINE_MUTEX(ksm_thread_mutex
);
250 static DEFINE_SPINLOCK(ksm_mmlist_lock
);
252 #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
253 sizeof(struct __struct), __alignof__(struct __struct),\
256 static int __init
ksm_slab_init(void)
258 rmap_item_cache
= KSM_KMEM_CACHE(rmap_item
, 0);
259 if (!rmap_item_cache
)
262 stable_node_cache
= KSM_KMEM_CACHE(stable_node
, 0);
263 if (!stable_node_cache
)
266 mm_slot_cache
= KSM_KMEM_CACHE(mm_slot
, 0);
273 kmem_cache_destroy(stable_node_cache
);
275 kmem_cache_destroy(rmap_item_cache
);
280 static void __init
ksm_slab_free(void)
282 kmem_cache_destroy(mm_slot_cache
);
283 kmem_cache_destroy(stable_node_cache
);
284 kmem_cache_destroy(rmap_item_cache
);
285 mm_slot_cache
= NULL
;
288 static inline struct rmap_item
*alloc_rmap_item(void)
290 struct rmap_item
*rmap_item
;
292 rmap_item
= kmem_cache_zalloc(rmap_item_cache
, GFP_KERNEL
|
293 __GFP_NORETRY
| __GFP_NOWARN
);
299 static inline void free_rmap_item(struct rmap_item
*rmap_item
)
302 rmap_item
->mm
= NULL
; /* debug safety */
303 kmem_cache_free(rmap_item_cache
, rmap_item
);
306 static inline struct stable_node
*alloc_stable_node(void)
309 * The allocation can take too long with GFP_KERNEL when memory is under
310 * pressure, which may lead to hung task warnings. Adding __GFP_HIGH
311 * grants access to memory reserves, helping to avoid this problem.
313 return kmem_cache_alloc(stable_node_cache
, GFP_KERNEL
| __GFP_HIGH
);
316 static inline void free_stable_node(struct stable_node
*stable_node
)
318 kmem_cache_free(stable_node_cache
, stable_node
);
321 static inline struct mm_slot
*alloc_mm_slot(void)
323 if (!mm_slot_cache
) /* initialization failed */
325 return kmem_cache_zalloc(mm_slot_cache
, GFP_KERNEL
);
328 static inline void free_mm_slot(struct mm_slot
*mm_slot
)
330 kmem_cache_free(mm_slot_cache
, mm_slot
);
333 static struct mm_slot
*get_mm_slot(struct mm_struct
*mm
)
335 struct mm_slot
*slot
;
337 hash_for_each_possible(mm_slots_hash
, slot
, link
, (unsigned long)mm
)
344 static void insert_to_mm_slots_hash(struct mm_struct
*mm
,
345 struct mm_slot
*mm_slot
)
348 hash_add(mm_slots_hash
, &mm_slot
->link
, (unsigned long)mm
);
352 * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
353 * page tables after it has passed through ksm_exit() - which, if necessary,
354 * takes mmap_sem briefly to serialize against them. ksm_exit() does not set
355 * a special flag: they can just back out as soon as mm_users goes to zero.
356 * ksm_test_exit() is used throughout to make this test for exit: in some
357 * places for correctness, in some places just to avoid unnecessary work.
359 static inline bool ksm_test_exit(struct mm_struct
*mm
)
361 return atomic_read(&mm
->mm_users
) == 0;
365 * We use break_ksm to break COW on a ksm page: it's a stripped down
367 * if (get_user_pages(addr, 1, 1, 1, &page, NULL) == 1)
370 * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
371 * in case the application has unmapped and remapped mm,addr meanwhile.
372 * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP
373 * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
375 * FAULT_FLAG/FOLL_REMOTE are because we do this outside the context
376 * of the process that owns 'vma'. We also do not want to enforce
377 * protection keys here anyway.
379 static int break_ksm(struct vm_area_struct
*vma
, unsigned long addr
)
386 page
= follow_page(vma
, addr
,
387 FOLL_GET
| FOLL_MIGRATION
| FOLL_REMOTE
);
388 if (IS_ERR_OR_NULL(page
))
391 ret
= handle_mm_fault(vma
, addr
,
392 FAULT_FLAG_WRITE
| FAULT_FLAG_REMOTE
);
394 ret
= VM_FAULT_WRITE
;
396 } while (!(ret
& (VM_FAULT_WRITE
| VM_FAULT_SIGBUS
| VM_FAULT_SIGSEGV
| VM_FAULT_OOM
)));
398 * We must loop because handle_mm_fault() may back out if there's
399 * any difficulty e.g. if pte accessed bit gets updated concurrently.
401 * VM_FAULT_WRITE is what we have been hoping for: it indicates that
402 * COW has been broken, even if the vma does not permit VM_WRITE;
403 * but note that a concurrent fault might break PageKsm for us.
405 * VM_FAULT_SIGBUS could occur if we race with truncation of the
406 * backing file, which also invalidates anonymous pages: that's
407 * okay, that truncation will have unmapped the PageKsm for us.
409 * VM_FAULT_OOM: at the time of writing (late July 2009), setting
410 * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
411 * current task has TIF_MEMDIE set, and will be OOM killed on return
412 * to user; and ksmd, having no mm, would never be chosen for that.
414 * But if the mm is in a limited mem_cgroup, then the fault may fail
415 * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
416 * even ksmd can fail in this way - though it's usually breaking ksm
417 * just to undo a merge it made a moment before, so unlikely to oom.
419 * That's a pity: we might therefore have more kernel pages allocated
420 * than we're counting as nodes in the stable tree; but ksm_do_scan
421 * will retry to break_cow on each pass, so should recover the page
422 * in due course. The important thing is to not let VM_MERGEABLE
423 * be cleared while any such pages might remain in the area.
425 return (ret
& VM_FAULT_OOM
) ? -ENOMEM
: 0;
428 static struct vm_area_struct
*find_mergeable_vma(struct mm_struct
*mm
,
431 struct vm_area_struct
*vma
;
432 if (ksm_test_exit(mm
))
434 vma
= find_vma(mm
, addr
);
435 if (!vma
|| vma
->vm_start
> addr
)
437 if (!(vma
->vm_flags
& VM_MERGEABLE
) || !vma
->anon_vma
)
442 static void break_cow(struct rmap_item
*rmap_item
)
444 struct mm_struct
*mm
= rmap_item
->mm
;
445 unsigned long addr
= rmap_item
->address
;
446 struct vm_area_struct
*vma
;
449 * It is not an accident that whenever we want to break COW
450 * to undo, we also need to drop a reference to the anon_vma.
452 put_anon_vma(rmap_item
->anon_vma
);
454 down_read(&mm
->mmap_sem
);
455 vma
= find_mergeable_vma(mm
, addr
);
457 break_ksm(vma
, addr
);
458 up_read(&mm
->mmap_sem
);
461 static struct page
*get_mergeable_page(struct rmap_item
*rmap_item
)
463 struct mm_struct
*mm
= rmap_item
->mm
;
464 unsigned long addr
= rmap_item
->address
;
465 struct vm_area_struct
*vma
;
468 down_read(&mm
->mmap_sem
);
469 vma
= find_mergeable_vma(mm
, addr
);
473 page
= follow_page(vma
, addr
, FOLL_GET
);
474 if (IS_ERR_OR_NULL(page
))
476 if (PageAnon(page
)) {
477 flush_anon_page(vma
, page
, addr
);
478 flush_dcache_page(page
);
484 up_read(&mm
->mmap_sem
);
489 * This helper is used for getting right index into array of tree roots.
490 * When merge_across_nodes knob is set to 1, there are only two rb-trees for
491 * stable and unstable pages from all nodes with roots in index 0. Otherwise,
492 * every node has its own stable and unstable tree.
494 static inline int get_kpfn_nid(unsigned long kpfn
)
496 return ksm_merge_across_nodes
? 0 : NUMA(pfn_to_nid(kpfn
));
499 static void remove_node_from_stable_tree(struct stable_node
*stable_node
)
501 struct rmap_item
*rmap_item
;
503 hlist_for_each_entry(rmap_item
, &stable_node
->hlist
, hlist
) {
504 if (rmap_item
->hlist
.next
)
508 put_anon_vma(rmap_item
->anon_vma
);
509 rmap_item
->address
&= PAGE_MASK
;
513 if (stable_node
->head
== &migrate_nodes
)
514 list_del(&stable_node
->list
);
516 rb_erase(&stable_node
->node
,
517 root_stable_tree
+ NUMA(stable_node
->nid
));
518 free_stable_node(stable_node
);
522 * get_ksm_page: checks if the page indicated by the stable node
523 * is still its ksm page, despite having held no reference to it.
524 * In which case we can trust the content of the page, and it
525 * returns the gotten page; but if the page has now been zapped,
526 * remove the stale node from the stable tree and return NULL.
527 * But beware, the stable node's page might be being migrated.
529 * You would expect the stable_node to hold a reference to the ksm page.
530 * But if it increments the page's count, swapping out has to wait for
531 * ksmd to come around again before it can free the page, which may take
532 * seconds or even minutes: much too unresponsive. So instead we use a
533 * "keyhole reference": access to the ksm page from the stable node peeps
534 * out through its keyhole to see if that page still holds the right key,
535 * pointing back to this stable node. This relies on freeing a PageAnon
536 * page to reset its page->mapping to NULL, and relies on no other use of
537 * a page to put something that might look like our key in page->mapping.
538 * is on its way to being freed; but it is an anomaly to bear in mind.
540 static struct page
*get_ksm_page(struct stable_node
*stable_node
, bool lock_it
)
543 void *expected_mapping
;
546 expected_mapping
= (void *)((unsigned long)stable_node
|
549 kpfn
= READ_ONCE(stable_node
->kpfn
);
550 page
= pfn_to_page(kpfn
);
553 * page is computed from kpfn, so on most architectures reading
554 * page->mapping is naturally ordered after reading node->kpfn,
555 * but on Alpha we need to be more careful.
557 smp_read_barrier_depends();
558 if (READ_ONCE(page
->mapping
) != expected_mapping
)
562 * We cannot do anything with the page while its refcount is 0.
563 * Usually 0 means free, or tail of a higher-order page: in which
564 * case this node is no longer referenced, and should be freed;
565 * however, it might mean that the page is under page_freeze_refs().
566 * The __remove_mapping() case is easy, again the node is now stale;
567 * but if page is swapcache in migrate_page_move_mapping(), it might
568 * still be our page, in which case it's essential to keep the node.
570 while (!get_page_unless_zero(page
)) {
572 * Another check for page->mapping != expected_mapping would
573 * work here too. We have chosen the !PageSwapCache test to
574 * optimize the common case, when the page is or is about to
575 * be freed: PageSwapCache is cleared (under spin_lock_irq)
576 * in the freeze_refs section of __remove_mapping(); but Anon
577 * page->mapping reset to NULL later, in free_pages_prepare().
579 if (!PageSwapCache(page
))
584 if (READ_ONCE(page
->mapping
) != expected_mapping
) {
591 if (READ_ONCE(page
->mapping
) != expected_mapping
) {
601 * We come here from above when page->mapping or !PageSwapCache
602 * suggests that the node is stale; but it might be under migration.
603 * We need smp_rmb(), matching the smp_wmb() in ksm_migrate_page(),
604 * before checking whether node->kpfn has been changed.
607 if (READ_ONCE(stable_node
->kpfn
) != kpfn
)
609 remove_node_from_stable_tree(stable_node
);
614 * Removing rmap_item from stable or unstable tree.
615 * This function will clean the information from the stable/unstable tree.
617 static void remove_rmap_item_from_tree(struct rmap_item
*rmap_item
)
619 if (rmap_item
->address
& STABLE_FLAG
) {
620 struct stable_node
*stable_node
;
623 stable_node
= rmap_item
->head
;
624 page
= get_ksm_page(stable_node
, true);
628 hlist_del(&rmap_item
->hlist
);
632 if (!hlist_empty(&stable_node
->hlist
))
637 put_anon_vma(rmap_item
->anon_vma
);
638 rmap_item
->address
&= PAGE_MASK
;
640 } else if (rmap_item
->address
& UNSTABLE_FLAG
) {
643 * Usually ksmd can and must skip the rb_erase, because
644 * root_unstable_tree was already reset to RB_ROOT.
645 * But be careful when an mm is exiting: do the rb_erase
646 * if this rmap_item was inserted by this scan, rather
647 * than left over from before.
649 age
= (unsigned char)(ksm_scan
.seqnr
- rmap_item
->address
);
652 rb_erase(&rmap_item
->node
,
653 root_unstable_tree
+ NUMA(rmap_item
->nid
));
654 ksm_pages_unshared
--;
655 rmap_item
->address
&= PAGE_MASK
;
658 cond_resched(); /* we're called from many long loops */
661 static void remove_trailing_rmap_items(struct mm_slot
*mm_slot
,
662 struct rmap_item
**rmap_list
)
665 struct rmap_item
*rmap_item
= *rmap_list
;
666 *rmap_list
= rmap_item
->rmap_list
;
667 remove_rmap_item_from_tree(rmap_item
);
668 free_rmap_item(rmap_item
);
673 * Though it's very tempting to unmerge rmap_items from stable tree rather
674 * than check every pte of a given vma, the locking doesn't quite work for
675 * that - an rmap_item is assigned to the stable tree after inserting ksm
676 * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing
677 * rmap_items from parent to child at fork time (so as not to waste time
678 * if exit comes before the next scan reaches it).
680 * Similarly, although we'd like to remove rmap_items (so updating counts
681 * and freeing memory) when unmerging an area, it's easier to leave that
682 * to the next pass of ksmd - consider, for example, how ksmd might be
683 * in cmp_and_merge_page on one of the rmap_items we would be removing.
685 static int unmerge_ksm_pages(struct vm_area_struct
*vma
,
686 unsigned long start
, unsigned long end
)
691 for (addr
= start
; addr
< end
&& !err
; addr
+= PAGE_SIZE
) {
692 if (ksm_test_exit(vma
->vm_mm
))
694 if (signal_pending(current
))
697 err
= break_ksm(vma
, addr
);
704 * Only called through the sysfs control interface:
706 static int remove_stable_node(struct stable_node
*stable_node
)
711 page
= get_ksm_page(stable_node
, true);
714 * get_ksm_page did remove_node_from_stable_tree itself.
719 if (WARN_ON_ONCE(page_mapped(page
))) {
721 * This should not happen: but if it does, just refuse to let
722 * merge_across_nodes be switched - there is no need to panic.
727 * The stable node did not yet appear stale to get_ksm_page(),
728 * since that allows for an unmapped ksm page to be recognized
729 * right up until it is freed; but the node is safe to remove.
730 * This page might be in a pagevec waiting to be freed,
731 * or it might be PageSwapCache (perhaps under writeback),
732 * or it might have been removed from swapcache a moment ago.
734 set_page_stable_node(page
, NULL
);
735 remove_node_from_stable_tree(stable_node
);
744 static int remove_all_stable_nodes(void)
746 struct stable_node
*stable_node
, *next
;
750 for (nid
= 0; nid
< ksm_nr_node_ids
; nid
++) {
751 while (root_stable_tree
[nid
].rb_node
) {
752 stable_node
= rb_entry(root_stable_tree
[nid
].rb_node
,
753 struct stable_node
, node
);
754 if (remove_stable_node(stable_node
)) {
756 break; /* proceed to next nid */
761 list_for_each_entry_safe(stable_node
, next
, &migrate_nodes
, list
) {
762 if (remove_stable_node(stable_node
))
769 static int unmerge_and_remove_all_rmap_items(void)
771 struct mm_slot
*mm_slot
;
772 struct mm_struct
*mm
;
773 struct vm_area_struct
*vma
;
776 spin_lock(&ksm_mmlist_lock
);
777 ksm_scan
.mm_slot
= list_entry(ksm_mm_head
.mm_list
.next
,
778 struct mm_slot
, mm_list
);
779 spin_unlock(&ksm_mmlist_lock
);
781 for (mm_slot
= ksm_scan
.mm_slot
;
782 mm_slot
!= &ksm_mm_head
; mm_slot
= ksm_scan
.mm_slot
) {
784 down_read(&mm
->mmap_sem
);
785 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
786 if (ksm_test_exit(mm
))
788 if (!(vma
->vm_flags
& VM_MERGEABLE
) || !vma
->anon_vma
)
790 err
= unmerge_ksm_pages(vma
,
791 vma
->vm_start
, vma
->vm_end
);
796 remove_trailing_rmap_items(mm_slot
, &mm_slot
->rmap_list
);
797 up_read(&mm
->mmap_sem
);
799 spin_lock(&ksm_mmlist_lock
);
800 ksm_scan
.mm_slot
= list_entry(mm_slot
->mm_list
.next
,
801 struct mm_slot
, mm_list
);
802 if (ksm_test_exit(mm
)) {
803 hash_del(&mm_slot
->link
);
804 list_del(&mm_slot
->mm_list
);
805 spin_unlock(&ksm_mmlist_lock
);
807 free_mm_slot(mm_slot
);
808 clear_bit(MMF_VM_MERGEABLE
, &mm
->flags
);
811 spin_unlock(&ksm_mmlist_lock
);
814 /* Clean up stable nodes, but don't worry if some are still busy */
815 remove_all_stable_nodes();
820 up_read(&mm
->mmap_sem
);
821 spin_lock(&ksm_mmlist_lock
);
822 ksm_scan
.mm_slot
= &ksm_mm_head
;
823 spin_unlock(&ksm_mmlist_lock
);
826 #endif /* CONFIG_SYSFS */
828 static u32
calc_checksum(struct page
*page
)
831 void *addr
= kmap_atomic(page
);
832 checksum
= jhash2(addr
, PAGE_SIZE
/ 4, 17);
837 static int memcmp_pages(struct page
*page1
, struct page
*page2
)
842 addr1
= kmap_atomic(page1
);
843 addr2
= kmap_atomic(page2
);
844 ret
= memcmp(addr1
, addr2
, PAGE_SIZE
);
845 kunmap_atomic(addr2
);
846 kunmap_atomic(addr1
);
850 static inline int pages_identical(struct page
*page1
, struct page
*page2
)
852 return !memcmp_pages(page1
, page2
);
855 static int write_protect_page(struct vm_area_struct
*vma
, struct page
*page
,
858 struct mm_struct
*mm
= vma
->vm_mm
;
859 struct page_vma_mapped_walk pvmw
= {
865 unsigned long mmun_start
; /* For mmu_notifiers */
866 unsigned long mmun_end
; /* For mmu_notifiers */
868 pvmw
.address
= page_address_in_vma(page
, vma
);
869 if (pvmw
.address
== -EFAULT
)
872 BUG_ON(PageTransCompound(page
));
874 mmun_start
= pvmw
.address
;
875 mmun_end
= pvmw
.address
+ PAGE_SIZE
;
876 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
878 if (!page_vma_mapped_walk(&pvmw
))
880 if (WARN_ONCE(!pvmw
.pte
, "Unexpected PMD mapping?"))
883 if (pte_write(*pvmw
.pte
) || pte_dirty(*pvmw
.pte
)) {
886 swapped
= PageSwapCache(page
);
887 flush_cache_page(vma
, pvmw
.address
, page_to_pfn(page
));
889 * Ok this is tricky, when get_user_pages_fast() run it doesn't
890 * take any lock, therefore the check that we are going to make
891 * with the pagecount against the mapcount is racey and
892 * O_DIRECT can happen right after the check.
893 * So we clear the pte and flush the tlb before the check
894 * this assure us that no O_DIRECT can happen after the check
895 * or in the middle of the check.
897 entry
= ptep_clear_flush_notify(vma
, pvmw
.address
, pvmw
.pte
);
899 * Check that no O_DIRECT or similar I/O is in progress on the
902 if (page_mapcount(page
) + 1 + swapped
!= page_count(page
)) {
903 set_pte_at(mm
, pvmw
.address
, pvmw
.pte
, entry
);
906 if (pte_dirty(entry
))
907 set_page_dirty(page
);
908 entry
= pte_mkclean(pte_wrprotect(entry
));
909 set_pte_at_notify(mm
, pvmw
.address
, pvmw
.pte
, entry
);
911 *orig_pte
= *pvmw
.pte
;
915 page_vma_mapped_walk_done(&pvmw
);
917 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
923 * replace_page - replace page in vma by new ksm page
924 * @vma: vma that holds the pte pointing to page
925 * @page: the page we are replacing by kpage
926 * @kpage: the ksm page we replace page by
927 * @orig_pte: the original value of the pte
929 * Returns 0 on success, -EFAULT on failure.
931 static int replace_page(struct vm_area_struct
*vma
, struct page
*page
,
932 struct page
*kpage
, pte_t orig_pte
)
934 struct mm_struct
*mm
= vma
->vm_mm
;
941 unsigned long mmun_start
; /* For mmu_notifiers */
942 unsigned long mmun_end
; /* For mmu_notifiers */
944 addr
= page_address_in_vma(page
, vma
);
948 pmd
= mm_find_pmd(mm
, addr
);
953 mmun_end
= addr
+ PAGE_SIZE
;
954 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
956 ptep
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
957 if (!pte_same(*ptep
, orig_pte
)) {
958 pte_unmap_unlock(ptep
, ptl
);
963 * No need to check ksm_use_zero_pages here: we can only have a
964 * zero_page here if ksm_use_zero_pages was enabled alreaady.
966 if (!is_zero_pfn(page_to_pfn(kpage
))) {
968 page_add_anon_rmap(kpage
, vma
, addr
, false);
969 newpte
= mk_pte(kpage
, vma
->vm_page_prot
);
971 newpte
= pte_mkspecial(pfn_pte(page_to_pfn(kpage
),
975 flush_cache_page(vma
, addr
, pte_pfn(*ptep
));
976 ptep_clear_flush_notify(vma
, addr
, ptep
);
977 set_pte_at_notify(mm
, addr
, ptep
, newpte
);
979 page_remove_rmap(page
, false);
980 if (!page_mapped(page
))
981 try_to_free_swap(page
);
984 pte_unmap_unlock(ptep
, ptl
);
987 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
993 * try_to_merge_one_page - take two pages and merge them into one
994 * @vma: the vma that holds the pte pointing to page
995 * @page: the PageAnon page that we want to replace with kpage
996 * @kpage: the PageKsm page that we want to map instead of page,
997 * or NULL the first time when we want to use page as kpage.
999 * This function returns 0 if the pages were merged, -EFAULT otherwise.
1001 static int try_to_merge_one_page(struct vm_area_struct
*vma
,
1002 struct page
*page
, struct page
*kpage
)
1004 pte_t orig_pte
= __pte(0);
1007 if (page
== kpage
) /* ksm page forked */
1010 if (!PageAnon(page
))
1014 * We need the page lock to read a stable PageSwapCache in
1015 * write_protect_page(). We use trylock_page() instead of
1016 * lock_page() because we don't want to wait here - we
1017 * prefer to continue scanning and merging different pages,
1018 * then come back to this page when it is unlocked.
1020 if (!trylock_page(page
))
1023 if (PageTransCompound(page
)) {
1024 err
= split_huge_page(page
);
1030 * If this anonymous page is mapped only here, its pte may need
1031 * to be write-protected. If it's mapped elsewhere, all of its
1032 * ptes are necessarily already write-protected. But in either
1033 * case, we need to lock and check page_count is not raised.
1035 if (write_protect_page(vma
, page
, &orig_pte
) == 0) {
1038 * While we hold page lock, upgrade page from
1039 * PageAnon+anon_vma to PageKsm+NULL stable_node:
1040 * stable_tree_insert() will update stable_node.
1042 set_page_stable_node(page
, NULL
);
1043 mark_page_accessed(page
);
1045 * Page reclaim just frees a clean page with no dirty
1046 * ptes: make sure that the ksm page would be swapped.
1048 if (!PageDirty(page
))
1051 } else if (pages_identical(page
, kpage
))
1052 err
= replace_page(vma
, page
, kpage
, orig_pte
);
1055 if ((vma
->vm_flags
& VM_LOCKED
) && kpage
&& !err
) {
1056 munlock_vma_page(page
);
1057 if (!PageMlocked(kpage
)) {
1060 mlock_vma_page(kpage
);
1061 page
= kpage
; /* for final unlock */
1072 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
1073 * but no new kernel page is allocated: kpage must already be a ksm page.
1075 * This function returns 0 if the pages were merged, -EFAULT otherwise.
1077 static int try_to_merge_with_ksm_page(struct rmap_item
*rmap_item
,
1078 struct page
*page
, struct page
*kpage
)
1080 struct mm_struct
*mm
= rmap_item
->mm
;
1081 struct vm_area_struct
*vma
;
1084 down_read(&mm
->mmap_sem
);
1085 vma
= find_mergeable_vma(mm
, rmap_item
->address
);
1089 err
= try_to_merge_one_page(vma
, page
, kpage
);
1093 /* Unstable nid is in union with stable anon_vma: remove first */
1094 remove_rmap_item_from_tree(rmap_item
);
1096 /* Must get reference to anon_vma while still holding mmap_sem */
1097 rmap_item
->anon_vma
= vma
->anon_vma
;
1098 get_anon_vma(vma
->anon_vma
);
1100 up_read(&mm
->mmap_sem
);
1105 * try_to_merge_two_pages - take two identical pages and prepare them
1106 * to be merged into one page.
1108 * This function returns the kpage if we successfully merged two identical
1109 * pages into one ksm page, NULL otherwise.
1111 * Note that this function upgrades page to ksm page: if one of the pages
1112 * is already a ksm page, try_to_merge_with_ksm_page should be used.
1114 static struct page
*try_to_merge_two_pages(struct rmap_item
*rmap_item
,
1116 struct rmap_item
*tree_rmap_item
,
1117 struct page
*tree_page
)
1121 err
= try_to_merge_with_ksm_page(rmap_item
, page
, NULL
);
1123 err
= try_to_merge_with_ksm_page(tree_rmap_item
,
1126 * If that fails, we have a ksm page with only one pte
1127 * pointing to it: so break it.
1130 break_cow(rmap_item
);
1132 return err
? NULL
: page
;
1136 * stable_tree_search - search for page inside the stable tree
1138 * This function checks if there is a page inside the stable tree
1139 * with identical content to the page that we are scanning right now.
1141 * This function returns the stable tree node of identical content if found,
1144 static struct page
*stable_tree_search(struct page
*page
)
1147 struct rb_root
*root
;
1148 struct rb_node
**new;
1149 struct rb_node
*parent
;
1150 struct stable_node
*stable_node
;
1151 struct stable_node
*page_node
;
1153 page_node
= page_stable_node(page
);
1154 if (page_node
&& page_node
->head
!= &migrate_nodes
) {
1155 /* ksm page forked */
1160 nid
= get_kpfn_nid(page_to_pfn(page
));
1161 root
= root_stable_tree
+ nid
;
1163 new = &root
->rb_node
;
1167 struct page
*tree_page
;
1171 stable_node
= rb_entry(*new, struct stable_node
, node
);
1172 tree_page
= get_ksm_page(stable_node
, false);
1175 * If we walked over a stale stable_node,
1176 * get_ksm_page() will call rb_erase() and it
1177 * may rebalance the tree from under us. So
1178 * restart the search from scratch. Returning
1179 * NULL would be safe too, but we'd generate
1180 * false negative insertions just because some
1181 * stable_node was stale.
1186 ret
= memcmp_pages(page
, tree_page
);
1187 put_page(tree_page
);
1191 new = &parent
->rb_left
;
1193 new = &parent
->rb_right
;
1196 * Lock and unlock the stable_node's page (which
1197 * might already have been migrated) so that page
1198 * migration is sure to notice its raised count.
1199 * It would be more elegant to return stable_node
1200 * than kpage, but that involves more changes.
1202 tree_page
= get_ksm_page(stable_node
, true);
1204 unlock_page(tree_page
);
1205 if (get_kpfn_nid(stable_node
->kpfn
) !=
1206 NUMA(stable_node
->nid
)) {
1207 put_page(tree_page
);
1213 * There is now a place for page_node, but the tree may
1214 * have been rebalanced, so re-evaluate parent and new.
1225 list_del(&page_node
->list
);
1226 DO_NUMA(page_node
->nid
= nid
);
1227 rb_link_node(&page_node
->node
, parent
, new);
1228 rb_insert_color(&page_node
->node
, root
);
1234 list_del(&page_node
->list
);
1235 DO_NUMA(page_node
->nid
= nid
);
1236 rb_replace_node(&stable_node
->node
, &page_node
->node
, root
);
1239 rb_erase(&stable_node
->node
, root
);
1242 stable_node
->head
= &migrate_nodes
;
1243 list_add(&stable_node
->list
, stable_node
->head
);
1248 * stable_tree_insert - insert stable tree node pointing to new ksm page
1249 * into the stable tree.
1251 * This function returns the stable tree node just allocated on success,
1254 static struct stable_node
*stable_tree_insert(struct page
*kpage
)
1258 struct rb_root
*root
;
1259 struct rb_node
**new;
1260 struct rb_node
*parent
;
1261 struct stable_node
*stable_node
;
1263 kpfn
= page_to_pfn(kpage
);
1264 nid
= get_kpfn_nid(kpfn
);
1265 root
= root_stable_tree
+ nid
;
1268 new = &root
->rb_node
;
1271 struct page
*tree_page
;
1275 stable_node
= rb_entry(*new, struct stable_node
, node
);
1276 tree_page
= get_ksm_page(stable_node
, false);
1279 * If we walked over a stale stable_node,
1280 * get_ksm_page() will call rb_erase() and it
1281 * may rebalance the tree from under us. So
1282 * restart the search from scratch. Returning
1283 * NULL would be safe too, but we'd generate
1284 * false negative insertions just because some
1285 * stable_node was stale.
1290 ret
= memcmp_pages(kpage
, tree_page
);
1291 put_page(tree_page
);
1295 new = &parent
->rb_left
;
1297 new = &parent
->rb_right
;
1300 * It is not a bug that stable_tree_search() didn't
1301 * find this node: because at that time our page was
1302 * not yet write-protected, so may have changed since.
1308 stable_node
= alloc_stable_node();
1312 INIT_HLIST_HEAD(&stable_node
->hlist
);
1313 stable_node
->kpfn
= kpfn
;
1314 set_page_stable_node(kpage
, stable_node
);
1315 DO_NUMA(stable_node
->nid
= nid
);
1316 rb_link_node(&stable_node
->node
, parent
, new);
1317 rb_insert_color(&stable_node
->node
, root
);
1323 * unstable_tree_search_insert - search for identical page,
1324 * else insert rmap_item into the unstable tree.
1326 * This function searches for a page in the unstable tree identical to the
1327 * page currently being scanned; and if no identical page is found in the
1328 * tree, we insert rmap_item as a new object into the unstable tree.
1330 * This function returns pointer to rmap_item found to be identical
1331 * to the currently scanned page, NULL otherwise.
1333 * This function does both searching and inserting, because they share
1334 * the same walking algorithm in an rbtree.
1337 struct rmap_item
*unstable_tree_search_insert(struct rmap_item
*rmap_item
,
1339 struct page
**tree_pagep
)
1341 struct rb_node
**new;
1342 struct rb_root
*root
;
1343 struct rb_node
*parent
= NULL
;
1346 nid
= get_kpfn_nid(page_to_pfn(page
));
1347 root
= root_unstable_tree
+ nid
;
1348 new = &root
->rb_node
;
1351 struct rmap_item
*tree_rmap_item
;
1352 struct page
*tree_page
;
1356 tree_rmap_item
= rb_entry(*new, struct rmap_item
, node
);
1357 tree_page
= get_mergeable_page(tree_rmap_item
);
1362 * Don't substitute a ksm page for a forked page.
1364 if (page
== tree_page
) {
1365 put_page(tree_page
);
1369 ret
= memcmp_pages(page
, tree_page
);
1373 put_page(tree_page
);
1374 new = &parent
->rb_left
;
1375 } else if (ret
> 0) {
1376 put_page(tree_page
);
1377 new = &parent
->rb_right
;
1378 } else if (!ksm_merge_across_nodes
&&
1379 page_to_nid(tree_page
) != nid
) {
1381 * If tree_page has been migrated to another NUMA node,
1382 * it will be flushed out and put in the right unstable
1383 * tree next time: only merge with it when across_nodes.
1385 put_page(tree_page
);
1388 *tree_pagep
= tree_page
;
1389 return tree_rmap_item
;
1393 rmap_item
->address
|= UNSTABLE_FLAG
;
1394 rmap_item
->address
|= (ksm_scan
.seqnr
& SEQNR_MASK
);
1395 DO_NUMA(rmap_item
->nid
= nid
);
1396 rb_link_node(&rmap_item
->node
, parent
, new);
1397 rb_insert_color(&rmap_item
->node
, root
);
1399 ksm_pages_unshared
++;
1404 * stable_tree_append - add another rmap_item to the linked list of
1405 * rmap_items hanging off a given node of the stable tree, all sharing
1406 * the same ksm page.
1408 static void stable_tree_append(struct rmap_item
*rmap_item
,
1409 struct stable_node
*stable_node
)
1411 rmap_item
->head
= stable_node
;
1412 rmap_item
->address
|= STABLE_FLAG
;
1413 hlist_add_head(&rmap_item
->hlist
, &stable_node
->hlist
);
1415 if (rmap_item
->hlist
.next
)
1416 ksm_pages_sharing
++;
1422 * cmp_and_merge_page - first see if page can be merged into the stable tree;
1423 * if not, compare checksum to previous and if it's the same, see if page can
1424 * be inserted into the unstable tree, or merged with a page already there and
1425 * both transferred to the stable tree.
1427 * @page: the page that we are searching identical page to.
1428 * @rmap_item: the reverse mapping into the virtual address of this page
1430 static void cmp_and_merge_page(struct page
*page
, struct rmap_item
*rmap_item
)
1432 struct rmap_item
*tree_rmap_item
;
1433 struct page
*tree_page
= NULL
;
1434 struct stable_node
*stable_node
;
1436 unsigned int checksum
;
1439 stable_node
= page_stable_node(page
);
1441 if (stable_node
->head
!= &migrate_nodes
&&
1442 get_kpfn_nid(stable_node
->kpfn
) != NUMA(stable_node
->nid
)) {
1443 rb_erase(&stable_node
->node
,
1444 root_stable_tree
+ NUMA(stable_node
->nid
));
1445 stable_node
->head
= &migrate_nodes
;
1446 list_add(&stable_node
->list
, stable_node
->head
);
1448 if (stable_node
->head
!= &migrate_nodes
&&
1449 rmap_item
->head
== stable_node
)
1453 /* We first start with searching the page inside the stable tree */
1454 kpage
= stable_tree_search(page
);
1455 if (kpage
== page
&& rmap_item
->head
== stable_node
) {
1460 remove_rmap_item_from_tree(rmap_item
);
1463 err
= try_to_merge_with_ksm_page(rmap_item
, page
, kpage
);
1466 * The page was successfully merged:
1467 * add its rmap_item to the stable tree.
1470 stable_tree_append(rmap_item
, page_stable_node(kpage
));
1478 * If the hash value of the page has changed from the last time
1479 * we calculated it, this page is changing frequently: therefore we
1480 * don't want to insert it in the unstable tree, and we don't want
1481 * to waste our time searching for something identical to it there.
1483 checksum
= calc_checksum(page
);
1484 if (rmap_item
->oldchecksum
!= checksum
) {
1485 rmap_item
->oldchecksum
= checksum
;
1490 * Same checksum as an empty page. We attempt to merge it with the
1491 * appropriate zero page if the user enabled this via sysfs.
1493 if (ksm_use_zero_pages
&& (checksum
== zero_checksum
)) {
1494 struct vm_area_struct
*vma
;
1496 vma
= find_mergeable_vma(rmap_item
->mm
, rmap_item
->address
);
1497 err
= try_to_merge_one_page(vma
, page
,
1498 ZERO_PAGE(rmap_item
->address
));
1500 * In case of failure, the page was not really empty, so we
1501 * need to continue. Otherwise we're done.
1507 unstable_tree_search_insert(rmap_item
, page
, &tree_page
);
1508 if (tree_rmap_item
) {
1509 kpage
= try_to_merge_two_pages(rmap_item
, page
,
1510 tree_rmap_item
, tree_page
);
1511 put_page(tree_page
);
1514 * The pages were successfully merged: insert new
1515 * node in the stable tree and add both rmap_items.
1518 stable_node
= stable_tree_insert(kpage
);
1520 stable_tree_append(tree_rmap_item
, stable_node
);
1521 stable_tree_append(rmap_item
, stable_node
);
1526 * If we fail to insert the page into the stable tree,
1527 * we will have 2 virtual addresses that are pointing
1528 * to a ksm page left outside the stable tree,
1529 * in which case we need to break_cow on both.
1532 break_cow(tree_rmap_item
);
1533 break_cow(rmap_item
);
1539 static struct rmap_item
*get_next_rmap_item(struct mm_slot
*mm_slot
,
1540 struct rmap_item
**rmap_list
,
1543 struct rmap_item
*rmap_item
;
1545 while (*rmap_list
) {
1546 rmap_item
= *rmap_list
;
1547 if ((rmap_item
->address
& PAGE_MASK
) == addr
)
1549 if (rmap_item
->address
> addr
)
1551 *rmap_list
= rmap_item
->rmap_list
;
1552 remove_rmap_item_from_tree(rmap_item
);
1553 free_rmap_item(rmap_item
);
1556 rmap_item
= alloc_rmap_item();
1558 /* It has already been zeroed */
1559 rmap_item
->mm
= mm_slot
->mm
;
1560 rmap_item
->address
= addr
;
1561 rmap_item
->rmap_list
= *rmap_list
;
1562 *rmap_list
= rmap_item
;
1567 static struct rmap_item
*scan_get_next_rmap_item(struct page
**page
)
1569 struct mm_struct
*mm
;
1570 struct mm_slot
*slot
;
1571 struct vm_area_struct
*vma
;
1572 struct rmap_item
*rmap_item
;
1575 if (list_empty(&ksm_mm_head
.mm_list
))
1578 slot
= ksm_scan
.mm_slot
;
1579 if (slot
== &ksm_mm_head
) {
1581 * A number of pages can hang around indefinitely on per-cpu
1582 * pagevecs, raised page count preventing write_protect_page
1583 * from merging them. Though it doesn't really matter much,
1584 * it is puzzling to see some stuck in pages_volatile until
1585 * other activity jostles them out, and they also prevented
1586 * LTP's KSM test from succeeding deterministically; so drain
1587 * them here (here rather than on entry to ksm_do_scan(),
1588 * so we don't IPI too often when pages_to_scan is set low).
1590 lru_add_drain_all();
1593 * Whereas stale stable_nodes on the stable_tree itself
1594 * get pruned in the regular course of stable_tree_search(),
1595 * those moved out to the migrate_nodes list can accumulate:
1596 * so prune them once before each full scan.
1598 if (!ksm_merge_across_nodes
) {
1599 struct stable_node
*stable_node
, *next
;
1602 list_for_each_entry_safe(stable_node
, next
,
1603 &migrate_nodes
, list
) {
1604 page
= get_ksm_page(stable_node
, false);
1611 for (nid
= 0; nid
< ksm_nr_node_ids
; nid
++)
1612 root_unstable_tree
[nid
] = RB_ROOT
;
1614 spin_lock(&ksm_mmlist_lock
);
1615 slot
= list_entry(slot
->mm_list
.next
, struct mm_slot
, mm_list
);
1616 ksm_scan
.mm_slot
= slot
;
1617 spin_unlock(&ksm_mmlist_lock
);
1619 * Although we tested list_empty() above, a racing __ksm_exit
1620 * of the last mm on the list may have removed it since then.
1622 if (slot
== &ksm_mm_head
)
1625 ksm_scan
.address
= 0;
1626 ksm_scan
.rmap_list
= &slot
->rmap_list
;
1630 down_read(&mm
->mmap_sem
);
1631 if (ksm_test_exit(mm
))
1634 vma
= find_vma(mm
, ksm_scan
.address
);
1636 for (; vma
; vma
= vma
->vm_next
) {
1637 if (!(vma
->vm_flags
& VM_MERGEABLE
))
1639 if (ksm_scan
.address
< vma
->vm_start
)
1640 ksm_scan
.address
= vma
->vm_start
;
1642 ksm_scan
.address
= vma
->vm_end
;
1644 while (ksm_scan
.address
< vma
->vm_end
) {
1645 if (ksm_test_exit(mm
))
1647 *page
= follow_page(vma
, ksm_scan
.address
, FOLL_GET
);
1648 if (IS_ERR_OR_NULL(*page
)) {
1649 ksm_scan
.address
+= PAGE_SIZE
;
1653 if (PageAnon(*page
)) {
1654 flush_anon_page(vma
, *page
, ksm_scan
.address
);
1655 flush_dcache_page(*page
);
1656 rmap_item
= get_next_rmap_item(slot
,
1657 ksm_scan
.rmap_list
, ksm_scan
.address
);
1659 ksm_scan
.rmap_list
=
1660 &rmap_item
->rmap_list
;
1661 ksm_scan
.address
+= PAGE_SIZE
;
1664 up_read(&mm
->mmap_sem
);
1668 ksm_scan
.address
+= PAGE_SIZE
;
1673 if (ksm_test_exit(mm
)) {
1674 ksm_scan
.address
= 0;
1675 ksm_scan
.rmap_list
= &slot
->rmap_list
;
1678 * Nuke all the rmap_items that are above this current rmap:
1679 * because there were no VM_MERGEABLE vmas with such addresses.
1681 remove_trailing_rmap_items(slot
, ksm_scan
.rmap_list
);
1683 spin_lock(&ksm_mmlist_lock
);
1684 ksm_scan
.mm_slot
= list_entry(slot
->mm_list
.next
,
1685 struct mm_slot
, mm_list
);
1686 if (ksm_scan
.address
== 0) {
1688 * We've completed a full scan of all vmas, holding mmap_sem
1689 * throughout, and found no VM_MERGEABLE: so do the same as
1690 * __ksm_exit does to remove this mm from all our lists now.
1691 * This applies either when cleaning up after __ksm_exit
1692 * (but beware: we can reach here even before __ksm_exit),
1693 * or when all VM_MERGEABLE areas have been unmapped (and
1694 * mmap_sem then protects against race with MADV_MERGEABLE).
1696 hash_del(&slot
->link
);
1697 list_del(&slot
->mm_list
);
1698 spin_unlock(&ksm_mmlist_lock
);
1701 clear_bit(MMF_VM_MERGEABLE
, &mm
->flags
);
1702 up_read(&mm
->mmap_sem
);
1705 up_read(&mm
->mmap_sem
);
1707 * up_read(&mm->mmap_sem) first because after
1708 * spin_unlock(&ksm_mmlist_lock) run, the "mm" may
1709 * already have been freed under us by __ksm_exit()
1710 * because the "mm_slot" is still hashed and
1711 * ksm_scan.mm_slot doesn't point to it anymore.
1713 spin_unlock(&ksm_mmlist_lock
);
1716 /* Repeat until we've completed scanning the whole list */
1717 slot
= ksm_scan
.mm_slot
;
1718 if (slot
!= &ksm_mm_head
)
1726 * ksm_do_scan - the ksm scanner main worker function.
1727 * @scan_npages - number of pages we want to scan before we return.
1729 static void ksm_do_scan(unsigned int scan_npages
)
1731 struct rmap_item
*rmap_item
;
1732 struct page
*uninitialized_var(page
);
1734 while (scan_npages
-- && likely(!freezing(current
))) {
1736 rmap_item
= scan_get_next_rmap_item(&page
);
1739 cmp_and_merge_page(page
, rmap_item
);
1744 static int ksmd_should_run(void)
1746 return (ksm_run
& KSM_RUN_MERGE
) && !list_empty(&ksm_mm_head
.mm_list
);
1749 static int ksm_scan_thread(void *nothing
)
1752 set_user_nice(current
, 5);
1754 while (!kthread_should_stop()) {
1755 mutex_lock(&ksm_thread_mutex
);
1756 wait_while_offlining();
1757 if (ksmd_should_run())
1758 ksm_do_scan(ksm_thread_pages_to_scan
);
1759 mutex_unlock(&ksm_thread_mutex
);
1763 if (ksmd_should_run()) {
1764 if (ksm_thread_sleep_millisecs
>= 1000)
1765 schedule_timeout_interruptible(
1766 msecs_to_jiffies(round_jiffies_relative(ksm_thread_sleep_millisecs
)));
1768 schedule_timeout_interruptible(
1769 msecs_to_jiffies(ksm_thread_sleep_millisecs
));
1771 wait_event_freezable(ksm_thread_wait
,
1772 ksmd_should_run() || kthread_should_stop());
1778 int ksm_madvise(struct vm_area_struct
*vma
, unsigned long start
,
1779 unsigned long end
, int advice
, unsigned long *vm_flags
)
1781 struct mm_struct
*mm
= vma
->vm_mm
;
1785 case MADV_MERGEABLE
:
1787 * Be somewhat over-protective for now!
1789 if (*vm_flags
& (VM_MERGEABLE
| VM_SHARED
| VM_MAYSHARE
|
1790 VM_PFNMAP
| VM_IO
| VM_DONTEXPAND
|
1791 VM_HUGETLB
| VM_MIXEDMAP
))
1792 return 0; /* just ignore the advice */
1795 if (*vm_flags
& VM_SAO
)
1799 if (!test_bit(MMF_VM_MERGEABLE
, &mm
->flags
)) {
1800 err
= __ksm_enter(mm
);
1805 *vm_flags
|= VM_MERGEABLE
;
1808 case MADV_UNMERGEABLE
:
1809 if (!(*vm_flags
& VM_MERGEABLE
))
1810 return 0; /* just ignore the advice */
1812 if (vma
->anon_vma
) {
1813 err
= unmerge_ksm_pages(vma
, start
, end
);
1818 *vm_flags
&= ~VM_MERGEABLE
;
1825 int __ksm_enter(struct mm_struct
*mm
)
1827 struct mm_slot
*mm_slot
;
1830 mm_slot
= alloc_mm_slot();
1834 /* Check ksm_run too? Would need tighter locking */
1835 needs_wakeup
= list_empty(&ksm_mm_head
.mm_list
);
1837 spin_lock(&ksm_mmlist_lock
);
1838 insert_to_mm_slots_hash(mm
, mm_slot
);
1840 * When KSM_RUN_MERGE (or KSM_RUN_STOP),
1841 * insert just behind the scanning cursor, to let the area settle
1842 * down a little; when fork is followed by immediate exec, we don't
1843 * want ksmd to waste time setting up and tearing down an rmap_list.
1845 * But when KSM_RUN_UNMERGE, it's important to insert ahead of its
1846 * scanning cursor, otherwise KSM pages in newly forked mms will be
1847 * missed: then we might as well insert at the end of the list.
1849 if (ksm_run
& KSM_RUN_UNMERGE
)
1850 list_add_tail(&mm_slot
->mm_list
, &ksm_mm_head
.mm_list
);
1852 list_add_tail(&mm_slot
->mm_list
, &ksm_scan
.mm_slot
->mm_list
);
1853 spin_unlock(&ksm_mmlist_lock
);
1855 set_bit(MMF_VM_MERGEABLE
, &mm
->flags
);
1856 atomic_inc(&mm
->mm_count
);
1859 wake_up_interruptible(&ksm_thread_wait
);
1864 void __ksm_exit(struct mm_struct
*mm
)
1866 struct mm_slot
*mm_slot
;
1867 int easy_to_free
= 0;
1870 * This process is exiting: if it's straightforward (as is the
1871 * case when ksmd was never running), free mm_slot immediately.
1872 * But if it's at the cursor or has rmap_items linked to it, use
1873 * mmap_sem to synchronize with any break_cows before pagetables
1874 * are freed, and leave the mm_slot on the list for ksmd to free.
1875 * Beware: ksm may already have noticed it exiting and freed the slot.
1878 spin_lock(&ksm_mmlist_lock
);
1879 mm_slot
= get_mm_slot(mm
);
1880 if (mm_slot
&& ksm_scan
.mm_slot
!= mm_slot
) {
1881 if (!mm_slot
->rmap_list
) {
1882 hash_del(&mm_slot
->link
);
1883 list_del(&mm_slot
->mm_list
);
1886 list_move(&mm_slot
->mm_list
,
1887 &ksm_scan
.mm_slot
->mm_list
);
1890 spin_unlock(&ksm_mmlist_lock
);
1893 free_mm_slot(mm_slot
);
1894 clear_bit(MMF_VM_MERGEABLE
, &mm
->flags
);
1896 } else if (mm_slot
) {
1897 down_write(&mm
->mmap_sem
);
1898 up_write(&mm
->mmap_sem
);
1902 struct page
*ksm_might_need_to_copy(struct page
*page
,
1903 struct vm_area_struct
*vma
, unsigned long address
)
1905 struct anon_vma
*anon_vma
= page_anon_vma(page
);
1906 struct page
*new_page
;
1908 if (PageKsm(page
)) {
1909 if (page_stable_node(page
) &&
1910 !(ksm_run
& KSM_RUN_UNMERGE
))
1911 return page
; /* no need to copy it */
1912 } else if (!anon_vma
) {
1913 return page
; /* no need to copy it */
1914 } else if (anon_vma
->root
== vma
->anon_vma
->root
&&
1915 page
->index
== linear_page_index(vma
, address
)) {
1916 return page
; /* still no need to copy it */
1918 if (!PageUptodate(page
))
1919 return page
; /* let do_swap_page report the error */
1921 new_page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
1923 copy_user_highpage(new_page
, page
, address
, vma
);
1925 SetPageDirty(new_page
);
1926 __SetPageUptodate(new_page
);
1927 __SetPageLocked(new_page
);
1933 int rmap_walk_ksm(struct page
*page
, struct rmap_walk_control
*rwc
)
1935 struct stable_node
*stable_node
;
1936 struct rmap_item
*rmap_item
;
1937 int ret
= SWAP_AGAIN
;
1938 int search_new_forks
= 0;
1940 VM_BUG_ON_PAGE(!PageKsm(page
), page
);
1943 * Rely on the page lock to protect against concurrent modifications
1944 * to that page's node of the stable tree.
1946 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
1948 stable_node
= page_stable_node(page
);
1952 hlist_for_each_entry(rmap_item
, &stable_node
->hlist
, hlist
) {
1953 struct anon_vma
*anon_vma
= rmap_item
->anon_vma
;
1954 struct anon_vma_chain
*vmac
;
1955 struct vm_area_struct
*vma
;
1958 anon_vma_lock_read(anon_vma
);
1959 anon_vma_interval_tree_foreach(vmac
, &anon_vma
->rb_root
,
1963 if (rmap_item
->address
< vma
->vm_start
||
1964 rmap_item
->address
>= vma
->vm_end
)
1967 * Initially we examine only the vma which covers this
1968 * rmap_item; but later, if there is still work to do,
1969 * we examine covering vmas in other mms: in case they
1970 * were forked from the original since ksmd passed.
1972 if ((rmap_item
->mm
== vma
->vm_mm
) == search_new_forks
)
1975 if (rwc
->invalid_vma
&& rwc
->invalid_vma(vma
, rwc
->arg
))
1978 ret
= rwc
->rmap_one(page
, vma
,
1979 rmap_item
->address
, rwc
->arg
);
1980 if (ret
!= SWAP_AGAIN
) {
1981 anon_vma_unlock_read(anon_vma
);
1984 if (rwc
->done
&& rwc
->done(page
)) {
1985 anon_vma_unlock_read(anon_vma
);
1989 anon_vma_unlock_read(anon_vma
);
1991 if (!search_new_forks
++)
1997 #ifdef CONFIG_MIGRATION
1998 void ksm_migrate_page(struct page
*newpage
, struct page
*oldpage
)
2000 struct stable_node
*stable_node
;
2002 VM_BUG_ON_PAGE(!PageLocked(oldpage
), oldpage
);
2003 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
2004 VM_BUG_ON_PAGE(newpage
->mapping
!= oldpage
->mapping
, newpage
);
2006 stable_node
= page_stable_node(newpage
);
2008 VM_BUG_ON_PAGE(stable_node
->kpfn
!= page_to_pfn(oldpage
), oldpage
);
2009 stable_node
->kpfn
= page_to_pfn(newpage
);
2011 * newpage->mapping was set in advance; now we need smp_wmb()
2012 * to make sure that the new stable_node->kpfn is visible
2013 * to get_ksm_page() before it can see that oldpage->mapping
2014 * has gone stale (or that PageSwapCache has been cleared).
2017 set_page_stable_node(oldpage
, NULL
);
2020 #endif /* CONFIG_MIGRATION */
2022 #ifdef CONFIG_MEMORY_HOTREMOVE
2023 static void wait_while_offlining(void)
2025 while (ksm_run
& KSM_RUN_OFFLINE
) {
2026 mutex_unlock(&ksm_thread_mutex
);
2027 wait_on_bit(&ksm_run
, ilog2(KSM_RUN_OFFLINE
),
2028 TASK_UNINTERRUPTIBLE
);
2029 mutex_lock(&ksm_thread_mutex
);
2033 static void ksm_check_stable_tree(unsigned long start_pfn
,
2034 unsigned long end_pfn
)
2036 struct stable_node
*stable_node
, *next
;
2037 struct rb_node
*node
;
2040 for (nid
= 0; nid
< ksm_nr_node_ids
; nid
++) {
2041 node
= rb_first(root_stable_tree
+ nid
);
2043 stable_node
= rb_entry(node
, struct stable_node
, node
);
2044 if (stable_node
->kpfn
>= start_pfn
&&
2045 stable_node
->kpfn
< end_pfn
) {
2047 * Don't get_ksm_page, page has already gone:
2048 * which is why we keep kpfn instead of page*
2050 remove_node_from_stable_tree(stable_node
);
2051 node
= rb_first(root_stable_tree
+ nid
);
2053 node
= rb_next(node
);
2057 list_for_each_entry_safe(stable_node
, next
, &migrate_nodes
, list
) {
2058 if (stable_node
->kpfn
>= start_pfn
&&
2059 stable_node
->kpfn
< end_pfn
)
2060 remove_node_from_stable_tree(stable_node
);
2065 static int ksm_memory_callback(struct notifier_block
*self
,
2066 unsigned long action
, void *arg
)
2068 struct memory_notify
*mn
= arg
;
2071 case MEM_GOING_OFFLINE
:
2073 * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items()
2074 * and remove_all_stable_nodes() while memory is going offline:
2075 * it is unsafe for them to touch the stable tree at this time.
2076 * But unmerge_ksm_pages(), rmap lookups and other entry points
2077 * which do not need the ksm_thread_mutex are all safe.
2079 mutex_lock(&ksm_thread_mutex
);
2080 ksm_run
|= KSM_RUN_OFFLINE
;
2081 mutex_unlock(&ksm_thread_mutex
);
2086 * Most of the work is done by page migration; but there might
2087 * be a few stable_nodes left over, still pointing to struct
2088 * pages which have been offlined: prune those from the tree,
2089 * otherwise get_ksm_page() might later try to access a
2090 * non-existent struct page.
2092 ksm_check_stable_tree(mn
->start_pfn
,
2093 mn
->start_pfn
+ mn
->nr_pages
);
2096 case MEM_CANCEL_OFFLINE
:
2097 mutex_lock(&ksm_thread_mutex
);
2098 ksm_run
&= ~KSM_RUN_OFFLINE
;
2099 mutex_unlock(&ksm_thread_mutex
);
2101 smp_mb(); /* wake_up_bit advises this */
2102 wake_up_bit(&ksm_run
, ilog2(KSM_RUN_OFFLINE
));
2108 static void wait_while_offlining(void)
2111 #endif /* CONFIG_MEMORY_HOTREMOVE */
2115 * This all compiles without CONFIG_SYSFS, but is a waste of space.
2118 #define KSM_ATTR_RO(_name) \
2119 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2120 #define KSM_ATTR(_name) \
2121 static struct kobj_attribute _name##_attr = \
2122 __ATTR(_name, 0644, _name##_show, _name##_store)
2124 static ssize_t
sleep_millisecs_show(struct kobject
*kobj
,
2125 struct kobj_attribute
*attr
, char *buf
)
2127 return sprintf(buf
, "%u\n", ksm_thread_sleep_millisecs
);
2130 static ssize_t
sleep_millisecs_store(struct kobject
*kobj
,
2131 struct kobj_attribute
*attr
,
2132 const char *buf
, size_t count
)
2134 unsigned long msecs
;
2137 err
= kstrtoul(buf
, 10, &msecs
);
2138 if (err
|| msecs
> UINT_MAX
)
2141 ksm_thread_sleep_millisecs
= msecs
;
2145 KSM_ATTR(sleep_millisecs
);
2147 static ssize_t
pages_to_scan_show(struct kobject
*kobj
,
2148 struct kobj_attribute
*attr
, char *buf
)
2150 return sprintf(buf
, "%u\n", ksm_thread_pages_to_scan
);
2153 static ssize_t
pages_to_scan_store(struct kobject
*kobj
,
2154 struct kobj_attribute
*attr
,
2155 const char *buf
, size_t count
)
2158 unsigned long nr_pages
;
2160 err
= kstrtoul(buf
, 10, &nr_pages
);
2161 if (err
|| nr_pages
> UINT_MAX
)
2164 ksm_thread_pages_to_scan
= nr_pages
;
2168 KSM_ATTR(pages_to_scan
);
2170 static ssize_t
run_show(struct kobject
*kobj
, struct kobj_attribute
*attr
,
2173 return sprintf(buf
, "%lu\n", ksm_run
);
2176 static ssize_t
run_store(struct kobject
*kobj
, struct kobj_attribute
*attr
,
2177 const char *buf
, size_t count
)
2180 unsigned long flags
;
2182 err
= kstrtoul(buf
, 10, &flags
);
2183 if (err
|| flags
> UINT_MAX
)
2185 if (flags
> KSM_RUN_UNMERGE
)
2189 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
2190 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
2191 * breaking COW to free the pages_shared (but leaves mm_slots
2192 * on the list for when ksmd may be set running again).
2195 mutex_lock(&ksm_thread_mutex
);
2196 wait_while_offlining();
2197 if (ksm_run
!= flags
) {
2199 if (flags
& KSM_RUN_UNMERGE
) {
2200 set_current_oom_origin();
2201 err
= unmerge_and_remove_all_rmap_items();
2202 clear_current_oom_origin();
2204 ksm_run
= KSM_RUN_STOP
;
2209 mutex_unlock(&ksm_thread_mutex
);
2211 if (flags
& KSM_RUN_MERGE
)
2212 wake_up_interruptible(&ksm_thread_wait
);
2219 static ssize_t
merge_across_nodes_show(struct kobject
*kobj
,
2220 struct kobj_attribute
*attr
, char *buf
)
2222 return sprintf(buf
, "%u\n", ksm_merge_across_nodes
);
2225 static ssize_t
merge_across_nodes_store(struct kobject
*kobj
,
2226 struct kobj_attribute
*attr
,
2227 const char *buf
, size_t count
)
2232 err
= kstrtoul(buf
, 10, &knob
);
2238 mutex_lock(&ksm_thread_mutex
);
2239 wait_while_offlining();
2240 if (ksm_merge_across_nodes
!= knob
) {
2241 if (ksm_pages_shared
|| remove_all_stable_nodes())
2243 else if (root_stable_tree
== one_stable_tree
) {
2244 struct rb_root
*buf
;
2246 * This is the first time that we switch away from the
2247 * default of merging across nodes: must now allocate
2248 * a buffer to hold as many roots as may be needed.
2249 * Allocate stable and unstable together:
2250 * MAXSMP NODES_SHIFT 10 will use 16kB.
2252 buf
= kcalloc(nr_node_ids
+ nr_node_ids
, sizeof(*buf
),
2254 /* Let us assume that RB_ROOT is NULL is zero */
2258 root_stable_tree
= buf
;
2259 root_unstable_tree
= buf
+ nr_node_ids
;
2260 /* Stable tree is empty but not the unstable */
2261 root_unstable_tree
[0] = one_unstable_tree
[0];
2265 ksm_merge_across_nodes
= knob
;
2266 ksm_nr_node_ids
= knob
? 1 : nr_node_ids
;
2269 mutex_unlock(&ksm_thread_mutex
);
2271 return err
? err
: count
;
2273 KSM_ATTR(merge_across_nodes
);
2276 static ssize_t
use_zero_pages_show(struct kobject
*kobj
,
2277 struct kobj_attribute
*attr
, char *buf
)
2279 return sprintf(buf
, "%u\n", ksm_use_zero_pages
);
2281 static ssize_t
use_zero_pages_store(struct kobject
*kobj
,
2282 struct kobj_attribute
*attr
,
2283 const char *buf
, size_t count
)
2288 err
= kstrtobool(buf
, &value
);
2292 ksm_use_zero_pages
= value
;
2296 KSM_ATTR(use_zero_pages
);
2298 static ssize_t
pages_shared_show(struct kobject
*kobj
,
2299 struct kobj_attribute
*attr
, char *buf
)
2301 return sprintf(buf
, "%lu\n", ksm_pages_shared
);
2303 KSM_ATTR_RO(pages_shared
);
2305 static ssize_t
pages_sharing_show(struct kobject
*kobj
,
2306 struct kobj_attribute
*attr
, char *buf
)
2308 return sprintf(buf
, "%lu\n", ksm_pages_sharing
);
2310 KSM_ATTR_RO(pages_sharing
);
2312 static ssize_t
pages_unshared_show(struct kobject
*kobj
,
2313 struct kobj_attribute
*attr
, char *buf
)
2315 return sprintf(buf
, "%lu\n", ksm_pages_unshared
);
2317 KSM_ATTR_RO(pages_unshared
);
2319 static ssize_t
pages_volatile_show(struct kobject
*kobj
,
2320 struct kobj_attribute
*attr
, char *buf
)
2322 long ksm_pages_volatile
;
2324 ksm_pages_volatile
= ksm_rmap_items
- ksm_pages_shared
2325 - ksm_pages_sharing
- ksm_pages_unshared
;
2327 * It was not worth any locking to calculate that statistic,
2328 * but it might therefore sometimes be negative: conceal that.
2330 if (ksm_pages_volatile
< 0)
2331 ksm_pages_volatile
= 0;
2332 return sprintf(buf
, "%ld\n", ksm_pages_volatile
);
2334 KSM_ATTR_RO(pages_volatile
);
2336 static ssize_t
full_scans_show(struct kobject
*kobj
,
2337 struct kobj_attribute
*attr
, char *buf
)
2339 return sprintf(buf
, "%lu\n", ksm_scan
.seqnr
);
2341 KSM_ATTR_RO(full_scans
);
2343 static struct attribute
*ksm_attrs
[] = {
2344 &sleep_millisecs_attr
.attr
,
2345 &pages_to_scan_attr
.attr
,
2347 &pages_shared_attr
.attr
,
2348 &pages_sharing_attr
.attr
,
2349 &pages_unshared_attr
.attr
,
2350 &pages_volatile_attr
.attr
,
2351 &full_scans_attr
.attr
,
2353 &merge_across_nodes_attr
.attr
,
2355 &use_zero_pages_attr
.attr
,
2359 static struct attribute_group ksm_attr_group
= {
2363 #endif /* CONFIG_SYSFS */
2365 static int __init
ksm_init(void)
2367 struct task_struct
*ksm_thread
;
2370 /* The correct value depends on page size and endianness */
2371 zero_checksum
= calc_checksum(ZERO_PAGE(0));
2372 /* Default to false for backwards compatibility */
2373 ksm_use_zero_pages
= false;
2375 err
= ksm_slab_init();
2379 ksm_thread
= kthread_run(ksm_scan_thread
, NULL
, "ksmd");
2380 if (IS_ERR(ksm_thread
)) {
2381 pr_err("ksm: creating kthread failed\n");
2382 err
= PTR_ERR(ksm_thread
);
2387 err
= sysfs_create_group(mm_kobj
, &ksm_attr_group
);
2389 pr_err("ksm: register sysfs failed\n");
2390 kthread_stop(ksm_thread
);
2394 ksm_run
= KSM_RUN_MERGE
; /* no way for user to start it */
2396 #endif /* CONFIG_SYSFS */
2398 #ifdef CONFIG_MEMORY_HOTREMOVE
2399 /* There is no significance to this priority 100 */
2400 hotplug_memory_notifier(ksm_memory_callback
, 100);
2409 subsys_initcall(ksm_init
);