xHCI: Set link state support
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / mm / ksm.c
blob942dfc73a2ff89c3c7c96b3f9b1838c99ec16d02
1 /*
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.
8 * Authors:
9 * Izik Eidus
10 * Andrea Arcangeli
11 * Chris Wright
12 * Hugh Dickins
14 * This work is licensed under the terms of the GNU GPL, version 2.
17 #include <linux/errno.h>
18 #include <linux/mm.h>
19 #include <linux/fs.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/hash.h>
37 #include <linux/freezer.h>
39 #include <asm/tlbflush.h>
40 #include "internal.h"
43 * A few notes about the KSM scanning process,
44 * to make it easier to understand the data structures below:
46 * In order to reduce excessive scanning, KSM sorts the memory pages by their
47 * contents into a data structure that holds pointers to the pages' locations.
49 * Since the contents of the pages may change at any moment, KSM cannot just
50 * insert the pages into a normal sorted tree and expect it to find anything.
51 * Therefore KSM uses two data structures - the stable and the unstable tree.
53 * The stable tree holds pointers to all the merged pages (ksm pages), sorted
54 * by their contents. Because each such page is write-protected, searching on
55 * this tree is fully assured to be working (except when pages are unmapped),
56 * and therefore this tree is called the stable tree.
58 * In addition to the stable tree, KSM uses a second data structure called the
59 * unstable tree: this tree holds pointers to pages which have been found to
60 * be "unchanged for a period of time". The unstable tree sorts these pages
61 * by their contents, but since they are not write-protected, KSM cannot rely
62 * upon the unstable tree to work correctly - the unstable tree is liable to
63 * be corrupted as its contents are modified, and so it is called unstable.
65 * KSM solves this problem by several techniques:
67 * 1) The unstable tree is flushed every time KSM completes scanning all
68 * memory areas, and then the tree is rebuilt again from the beginning.
69 * 2) KSM will only insert into the unstable tree, pages whose hash value
70 * has not changed since the previous scan of all memory areas.
71 * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
72 * colors of the nodes and not on their contents, assuring that even when
73 * the tree gets "corrupted" it won't get out of balance, so scanning time
74 * remains the same (also, searching and inserting nodes in an rbtree uses
75 * the same algorithm, so we have no overhead when we flush and rebuild).
76 * 4) KSM never flushes the stable tree, which means that even if it were to
77 * take 10 attempts to find a page in the unstable tree, once it is found,
78 * it is secured in the stable tree. (When we scan a new page, we first
79 * compare it against the stable tree, and then against the unstable tree.)
82 /**
83 * struct mm_slot - ksm information per mm that is being scanned
84 * @link: link to the mm_slots hash list
85 * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
86 * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
87 * @mm: the mm that this information is valid for
89 struct mm_slot {
90 struct hlist_node link;
91 struct list_head mm_list;
92 struct rmap_item *rmap_list;
93 struct mm_struct *mm;
96 /**
97 * struct ksm_scan - cursor for scanning
98 * @mm_slot: the current mm_slot we are scanning
99 * @address: the next address inside that to be scanned
100 * @rmap_list: link to the next rmap to be scanned in the rmap_list
101 * @seqnr: count of completed full scans (needed when removing unstable node)
103 * There is only the one ksm_scan instance of this cursor structure.
105 struct ksm_scan {
106 struct mm_slot *mm_slot;
107 unsigned long address;
108 struct rmap_item **rmap_list;
109 unsigned long seqnr;
113 * struct stable_node - node of the stable rbtree
114 * @node: rb node of this ksm page in the stable tree
115 * @hlist: hlist head of rmap_items using this ksm page
116 * @kpfn: page frame number of this ksm page
118 struct stable_node {
119 struct rb_node node;
120 struct hlist_head hlist;
121 unsigned long kpfn;
125 * struct rmap_item - reverse mapping item for virtual addresses
126 * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
127 * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
128 * @mm: the memory structure this rmap_item is pointing into
129 * @address: the virtual address this rmap_item tracks (+ flags in low bits)
130 * @oldchecksum: previous checksum of the page at that virtual address
131 * @node: rb node of this rmap_item in the unstable tree
132 * @head: pointer to stable_node heading this list in the stable tree
133 * @hlist: link into hlist of rmap_items hanging off that stable_node
135 struct rmap_item {
136 struct rmap_item *rmap_list;
137 struct anon_vma *anon_vma; /* when stable */
138 struct mm_struct *mm;
139 unsigned long address; /* + low bits used for flags below */
140 unsigned int oldchecksum; /* when unstable */
141 union {
142 struct rb_node node; /* when node of unstable tree */
143 struct { /* when listed from stable tree */
144 struct stable_node *head;
145 struct hlist_node hlist;
150 #define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */
151 #define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */
152 #define STABLE_FLAG 0x200 /* is listed from the stable tree */
154 /* The stable and unstable tree heads */
155 static struct rb_root root_stable_tree = RB_ROOT;
156 static struct rb_root root_unstable_tree = RB_ROOT;
158 #define MM_SLOTS_HASH_SHIFT 10
159 #define MM_SLOTS_HASH_HEADS (1 << MM_SLOTS_HASH_SHIFT)
160 static struct hlist_head mm_slots_hash[MM_SLOTS_HASH_HEADS];
162 static struct mm_slot ksm_mm_head = {
163 .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
165 static struct ksm_scan ksm_scan = {
166 .mm_slot = &ksm_mm_head,
169 static struct kmem_cache *rmap_item_cache;
170 static struct kmem_cache *stable_node_cache;
171 static struct kmem_cache *mm_slot_cache;
173 /* The number of nodes in the stable tree */
174 static unsigned long ksm_pages_shared;
176 /* The number of page slots additionally sharing those nodes */
177 static unsigned long ksm_pages_sharing;
179 /* The number of nodes in the unstable tree */
180 static unsigned long ksm_pages_unshared;
182 /* The number of rmap_items in use: to calculate pages_volatile */
183 static unsigned long ksm_rmap_items;
185 /* Number of pages ksmd should scan in one batch */
186 static unsigned int ksm_thread_pages_to_scan = 100;
188 /* Milliseconds ksmd should sleep between batches */
189 static unsigned int ksm_thread_sleep_millisecs = 20;
191 #define KSM_RUN_STOP 0
192 #define KSM_RUN_MERGE 1
193 #define KSM_RUN_UNMERGE 2
194 static unsigned int ksm_run = KSM_RUN_STOP;
196 static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
197 static DEFINE_MUTEX(ksm_thread_mutex);
198 static DEFINE_SPINLOCK(ksm_mmlist_lock);
200 #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
201 sizeof(struct __struct), __alignof__(struct __struct),\
202 (__flags), NULL)
204 static int __init ksm_slab_init(void)
206 rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
207 if (!rmap_item_cache)
208 goto out;
210 stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
211 if (!stable_node_cache)
212 goto out_free1;
214 mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
215 if (!mm_slot_cache)
216 goto out_free2;
218 return 0;
220 out_free2:
221 kmem_cache_destroy(stable_node_cache);
222 out_free1:
223 kmem_cache_destroy(rmap_item_cache);
224 out:
225 return -ENOMEM;
228 static void __init ksm_slab_free(void)
230 kmem_cache_destroy(mm_slot_cache);
231 kmem_cache_destroy(stable_node_cache);
232 kmem_cache_destroy(rmap_item_cache);
233 mm_slot_cache = NULL;
236 static inline struct rmap_item *alloc_rmap_item(void)
238 struct rmap_item *rmap_item;
240 rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL);
241 if (rmap_item)
242 ksm_rmap_items++;
243 return rmap_item;
246 static inline void free_rmap_item(struct rmap_item *rmap_item)
248 ksm_rmap_items--;
249 rmap_item->mm = NULL; /* debug safety */
250 kmem_cache_free(rmap_item_cache, rmap_item);
253 static inline struct stable_node *alloc_stable_node(void)
255 return kmem_cache_alloc(stable_node_cache, GFP_KERNEL);
258 static inline void free_stable_node(struct stable_node *stable_node)
260 kmem_cache_free(stable_node_cache, stable_node);
263 static inline struct mm_slot *alloc_mm_slot(void)
265 if (!mm_slot_cache) /* initialization failed */
266 return NULL;
267 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
270 static inline void free_mm_slot(struct mm_slot *mm_slot)
272 kmem_cache_free(mm_slot_cache, mm_slot);
275 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
277 struct mm_slot *mm_slot;
278 struct hlist_head *bucket;
279 struct hlist_node *node;
281 bucket = &mm_slots_hash[hash_ptr(mm, MM_SLOTS_HASH_SHIFT)];
282 hlist_for_each_entry(mm_slot, node, bucket, link) {
283 if (mm == mm_slot->mm)
284 return mm_slot;
286 return NULL;
289 static void insert_to_mm_slots_hash(struct mm_struct *mm,
290 struct mm_slot *mm_slot)
292 struct hlist_head *bucket;
294 bucket = &mm_slots_hash[hash_ptr(mm, MM_SLOTS_HASH_SHIFT)];
295 mm_slot->mm = mm;
296 hlist_add_head(&mm_slot->link, bucket);
299 static inline int in_stable_tree(struct rmap_item *rmap_item)
301 return rmap_item->address & STABLE_FLAG;
305 * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
306 * page tables after it has passed through ksm_exit() - which, if necessary,
307 * takes mmap_sem briefly to serialize against them. ksm_exit() does not set
308 * a special flag: they can just back out as soon as mm_users goes to zero.
309 * ksm_test_exit() is used throughout to make this test for exit: in some
310 * places for correctness, in some places just to avoid unnecessary work.
312 static inline bool ksm_test_exit(struct mm_struct *mm)
314 return atomic_read(&mm->mm_users) == 0;
318 * We use break_ksm to break COW on a ksm page: it's a stripped down
320 * if (get_user_pages(current, mm, addr, 1, 1, 1, &page, NULL) == 1)
321 * put_page(page);
323 * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
324 * in case the application has unmapped and remapped mm,addr meanwhile.
325 * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP
326 * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
328 static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
330 struct page *page;
331 int ret = 0;
333 do {
334 cond_resched();
335 page = follow_page(vma, addr, FOLL_GET);
336 if (IS_ERR_OR_NULL(page))
337 break;
338 if (PageKsm(page))
339 ret = handle_mm_fault(vma->vm_mm, vma, addr,
340 FAULT_FLAG_WRITE);
341 else
342 ret = VM_FAULT_WRITE;
343 put_page(page);
344 } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_OOM)));
346 * We must loop because handle_mm_fault() may back out if there's
347 * any difficulty e.g. if pte accessed bit gets updated concurrently.
349 * VM_FAULT_WRITE is what we have been hoping for: it indicates that
350 * COW has been broken, even if the vma does not permit VM_WRITE;
351 * but note that a concurrent fault might break PageKsm for us.
353 * VM_FAULT_SIGBUS could occur if we race with truncation of the
354 * backing file, which also invalidates anonymous pages: that's
355 * okay, that truncation will have unmapped the PageKsm for us.
357 * VM_FAULT_OOM: at the time of writing (late July 2009), setting
358 * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
359 * current task has TIF_MEMDIE set, and will be OOM killed on return
360 * to user; and ksmd, having no mm, would never be chosen for that.
362 * But if the mm is in a limited mem_cgroup, then the fault may fail
363 * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
364 * even ksmd can fail in this way - though it's usually breaking ksm
365 * just to undo a merge it made a moment before, so unlikely to oom.
367 * That's a pity: we might therefore have more kernel pages allocated
368 * than we're counting as nodes in the stable tree; but ksm_do_scan
369 * will retry to break_cow on each pass, so should recover the page
370 * in due course. The important thing is to not let VM_MERGEABLE
371 * be cleared while any such pages might remain in the area.
373 return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
376 static void break_cow(struct rmap_item *rmap_item)
378 struct mm_struct *mm = rmap_item->mm;
379 unsigned long addr = rmap_item->address;
380 struct vm_area_struct *vma;
383 * It is not an accident that whenever we want to break COW
384 * to undo, we also need to drop a reference to the anon_vma.
386 put_anon_vma(rmap_item->anon_vma);
388 down_read(&mm->mmap_sem);
389 if (ksm_test_exit(mm))
390 goto out;
391 vma = find_vma(mm, addr);
392 if (!vma || vma->vm_start > addr)
393 goto out;
394 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
395 goto out;
396 break_ksm(vma, addr);
397 out:
398 up_read(&mm->mmap_sem);
401 static struct page *page_trans_compound_anon(struct page *page)
403 if (PageTransCompound(page)) {
404 struct page *head = compound_trans_head(page);
406 * head may actually be splitted and freed from under
407 * us but it's ok here.
409 if (PageAnon(head))
410 return head;
412 return NULL;
415 static struct page *get_mergeable_page(struct rmap_item *rmap_item)
417 struct mm_struct *mm = rmap_item->mm;
418 unsigned long addr = rmap_item->address;
419 struct vm_area_struct *vma;
420 struct page *page;
422 down_read(&mm->mmap_sem);
423 if (ksm_test_exit(mm))
424 goto out;
425 vma = find_vma(mm, addr);
426 if (!vma || vma->vm_start > addr)
427 goto out;
428 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
429 goto out;
431 page = follow_page(vma, addr, FOLL_GET);
432 if (IS_ERR_OR_NULL(page))
433 goto out;
434 if (PageAnon(page) || page_trans_compound_anon(page)) {
435 flush_anon_page(vma, page, addr);
436 flush_dcache_page(page);
437 } else {
438 put_page(page);
439 out: page = NULL;
441 up_read(&mm->mmap_sem);
442 return page;
445 static void remove_node_from_stable_tree(struct stable_node *stable_node)
447 struct rmap_item *rmap_item;
448 struct hlist_node *hlist;
450 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
451 if (rmap_item->hlist.next)
452 ksm_pages_sharing--;
453 else
454 ksm_pages_shared--;
455 put_anon_vma(rmap_item->anon_vma);
456 rmap_item->address &= PAGE_MASK;
457 cond_resched();
460 rb_erase(&stable_node->node, &root_stable_tree);
461 free_stable_node(stable_node);
465 * get_ksm_page: checks if the page indicated by the stable node
466 * is still its ksm page, despite having held no reference to it.
467 * In which case we can trust the content of the page, and it
468 * returns the gotten page; but if the page has now been zapped,
469 * remove the stale node from the stable tree and return NULL.
471 * You would expect the stable_node to hold a reference to the ksm page.
472 * But if it increments the page's count, swapping out has to wait for
473 * ksmd to come around again before it can free the page, which may take
474 * seconds or even minutes: much too unresponsive. So instead we use a
475 * "keyhole reference": access to the ksm page from the stable node peeps
476 * out through its keyhole to see if that page still holds the right key,
477 * pointing back to this stable node. This relies on freeing a PageAnon
478 * page to reset its page->mapping to NULL, and relies on no other use of
479 * a page to put something that might look like our key in page->mapping.
481 * include/linux/pagemap.h page_cache_get_speculative() is a good reference,
482 * but this is different - made simpler by ksm_thread_mutex being held, but
483 * interesting for assuming that no other use of the struct page could ever
484 * put our expected_mapping into page->mapping (or a field of the union which
485 * coincides with page->mapping). The RCU calls are not for KSM at all, but
486 * to keep the page_count protocol described with page_cache_get_speculative.
488 * Note: it is possible that get_ksm_page() will return NULL one moment,
489 * then page the next, if the page is in between page_freeze_refs() and
490 * page_unfreeze_refs(): this shouldn't be a problem anywhere, the page
491 * is on its way to being freed; but it is an anomaly to bear in mind.
493 static struct page *get_ksm_page(struct stable_node *stable_node)
495 struct page *page;
496 void *expected_mapping;
498 page = pfn_to_page(stable_node->kpfn);
499 expected_mapping = (void *)stable_node +
500 (PAGE_MAPPING_ANON | PAGE_MAPPING_KSM);
501 rcu_read_lock();
502 if (page->mapping != expected_mapping)
503 goto stale;
504 if (!get_page_unless_zero(page))
505 goto stale;
506 if (page->mapping != expected_mapping) {
507 put_page(page);
508 goto stale;
510 rcu_read_unlock();
511 return page;
512 stale:
513 rcu_read_unlock();
514 remove_node_from_stable_tree(stable_node);
515 return NULL;
519 * Removing rmap_item from stable or unstable tree.
520 * This function will clean the information from the stable/unstable tree.
522 static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
524 if (rmap_item->address & STABLE_FLAG) {
525 struct stable_node *stable_node;
526 struct page *page;
528 stable_node = rmap_item->head;
529 page = get_ksm_page(stable_node);
530 if (!page)
531 goto out;
533 lock_page(page);
534 hlist_del(&rmap_item->hlist);
535 unlock_page(page);
536 put_page(page);
538 if (stable_node->hlist.first)
539 ksm_pages_sharing--;
540 else
541 ksm_pages_shared--;
543 put_anon_vma(rmap_item->anon_vma);
544 rmap_item->address &= PAGE_MASK;
546 } else if (rmap_item->address & UNSTABLE_FLAG) {
547 unsigned char age;
549 * Usually ksmd can and must skip the rb_erase, because
550 * root_unstable_tree was already reset to RB_ROOT.
551 * But be careful when an mm is exiting: do the rb_erase
552 * if this rmap_item was inserted by this scan, rather
553 * than left over from before.
555 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
556 BUG_ON(age > 1);
557 if (!age)
558 rb_erase(&rmap_item->node, &root_unstable_tree);
560 ksm_pages_unshared--;
561 rmap_item->address &= PAGE_MASK;
563 out:
564 cond_resched(); /* we're called from many long loops */
567 static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
568 struct rmap_item **rmap_list)
570 while (*rmap_list) {
571 struct rmap_item *rmap_item = *rmap_list;
572 *rmap_list = rmap_item->rmap_list;
573 remove_rmap_item_from_tree(rmap_item);
574 free_rmap_item(rmap_item);
579 * Though it's very tempting to unmerge in_stable_tree(rmap_item)s rather
580 * than check every pte of a given vma, the locking doesn't quite work for
581 * that - an rmap_item is assigned to the stable tree after inserting ksm
582 * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing
583 * rmap_items from parent to child at fork time (so as not to waste time
584 * if exit comes before the next scan reaches it).
586 * Similarly, although we'd like to remove rmap_items (so updating counts
587 * and freeing memory) when unmerging an area, it's easier to leave that
588 * to the next pass of ksmd - consider, for example, how ksmd might be
589 * in cmp_and_merge_page on one of the rmap_items we would be removing.
591 static int unmerge_ksm_pages(struct vm_area_struct *vma,
592 unsigned long start, unsigned long end)
594 unsigned long addr;
595 int err = 0;
597 for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
598 if (ksm_test_exit(vma->vm_mm))
599 break;
600 if (signal_pending(current))
601 err = -ERESTARTSYS;
602 else
603 err = break_ksm(vma, addr);
605 return err;
608 #ifdef CONFIG_SYSFS
610 * Only called through the sysfs control interface:
612 static int unmerge_and_remove_all_rmap_items(void)
614 struct mm_slot *mm_slot;
615 struct mm_struct *mm;
616 struct vm_area_struct *vma;
617 int err = 0;
619 spin_lock(&ksm_mmlist_lock);
620 ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
621 struct mm_slot, mm_list);
622 spin_unlock(&ksm_mmlist_lock);
624 for (mm_slot = ksm_scan.mm_slot;
625 mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
626 mm = mm_slot->mm;
627 down_read(&mm->mmap_sem);
628 for (vma = mm->mmap; vma; vma = vma->vm_next) {
629 if (ksm_test_exit(mm))
630 break;
631 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
632 continue;
633 err = unmerge_ksm_pages(vma,
634 vma->vm_start, vma->vm_end);
635 if (err)
636 goto error;
639 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
641 spin_lock(&ksm_mmlist_lock);
642 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
643 struct mm_slot, mm_list);
644 if (ksm_test_exit(mm)) {
645 hlist_del(&mm_slot->link);
646 list_del(&mm_slot->mm_list);
647 spin_unlock(&ksm_mmlist_lock);
649 free_mm_slot(mm_slot);
650 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
651 up_read(&mm->mmap_sem);
652 mmdrop(mm);
653 } else {
654 spin_unlock(&ksm_mmlist_lock);
655 up_read(&mm->mmap_sem);
659 ksm_scan.seqnr = 0;
660 return 0;
662 error:
663 up_read(&mm->mmap_sem);
664 spin_lock(&ksm_mmlist_lock);
665 ksm_scan.mm_slot = &ksm_mm_head;
666 spin_unlock(&ksm_mmlist_lock);
667 return err;
669 #endif /* CONFIG_SYSFS */
671 static u32 calc_checksum(struct page *page)
673 u32 checksum;
674 void *addr = kmap_atomic(page, KM_USER0);
675 checksum = jhash2(addr, PAGE_SIZE / 4, 17);
676 kunmap_atomic(addr, KM_USER0);
677 return checksum;
680 static int memcmp_pages(struct page *page1, struct page *page2)
682 char *addr1, *addr2;
683 int ret;
685 addr1 = kmap_atomic(page1, KM_USER0);
686 addr2 = kmap_atomic(page2, KM_USER1);
687 ret = memcmp(addr1, addr2, PAGE_SIZE);
688 kunmap_atomic(addr2, KM_USER1);
689 kunmap_atomic(addr1, KM_USER0);
690 return ret;
693 static inline int pages_identical(struct page *page1, struct page *page2)
695 return !memcmp_pages(page1, page2);
698 static int write_protect_page(struct vm_area_struct *vma, struct page *page,
699 pte_t *orig_pte)
701 struct mm_struct *mm = vma->vm_mm;
702 unsigned long addr;
703 pte_t *ptep;
704 spinlock_t *ptl;
705 int swapped;
706 int err = -EFAULT;
708 addr = page_address_in_vma(page, vma);
709 if (addr == -EFAULT)
710 goto out;
712 BUG_ON(PageTransCompound(page));
713 ptep = page_check_address(page, mm, addr, &ptl, 0);
714 if (!ptep)
715 goto out;
717 if (pte_write(*ptep) || pte_dirty(*ptep)) {
718 pte_t entry;
720 swapped = PageSwapCache(page);
721 flush_cache_page(vma, addr, page_to_pfn(page));
723 * Ok this is tricky, when get_user_pages_fast() run it doesn't
724 * take any lock, therefore the check that we are going to make
725 * with the pagecount against the mapcount is racey and
726 * O_DIRECT can happen right after the check.
727 * So we clear the pte and flush the tlb before the check
728 * this assure us that no O_DIRECT can happen after the check
729 * or in the middle of the check.
731 entry = ptep_clear_flush(vma, addr, ptep);
733 * Check that no O_DIRECT or similar I/O is in progress on the
734 * page
736 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
737 set_pte_at(mm, addr, ptep, entry);
738 goto out_unlock;
740 if (pte_dirty(entry))
741 set_page_dirty(page);
742 entry = pte_mkclean(pte_wrprotect(entry));
743 set_pte_at_notify(mm, addr, ptep, entry);
745 *orig_pte = *ptep;
746 err = 0;
748 out_unlock:
749 pte_unmap_unlock(ptep, ptl);
750 out:
751 return err;
755 * replace_page - replace page in vma by new ksm page
756 * @vma: vma that holds the pte pointing to page
757 * @page: the page we are replacing by kpage
758 * @kpage: the ksm page we replace page by
759 * @orig_pte: the original value of the pte
761 * Returns 0 on success, -EFAULT on failure.
763 static int replace_page(struct vm_area_struct *vma, struct page *page,
764 struct page *kpage, pte_t orig_pte)
766 struct mm_struct *mm = vma->vm_mm;
767 pgd_t *pgd;
768 pud_t *pud;
769 pmd_t *pmd;
770 pte_t *ptep;
771 spinlock_t *ptl;
772 unsigned long addr;
773 int err = -EFAULT;
775 addr = page_address_in_vma(page, vma);
776 if (addr == -EFAULT)
777 goto out;
779 pgd = pgd_offset(mm, addr);
780 if (!pgd_present(*pgd))
781 goto out;
783 pud = pud_offset(pgd, addr);
784 if (!pud_present(*pud))
785 goto out;
787 pmd = pmd_offset(pud, addr);
788 BUG_ON(pmd_trans_huge(*pmd));
789 if (!pmd_present(*pmd))
790 goto out;
792 ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
793 if (!pte_same(*ptep, orig_pte)) {
794 pte_unmap_unlock(ptep, ptl);
795 goto out;
798 get_page(kpage);
799 page_add_anon_rmap(kpage, vma, addr);
801 flush_cache_page(vma, addr, pte_pfn(*ptep));
802 ptep_clear_flush(vma, addr, ptep);
803 set_pte_at_notify(mm, addr, ptep, mk_pte(kpage, vma->vm_page_prot));
805 page_remove_rmap(page);
806 if (!page_mapped(page))
807 try_to_free_swap(page);
808 put_page(page);
810 pte_unmap_unlock(ptep, ptl);
811 err = 0;
812 out:
813 return err;
816 static int page_trans_compound_anon_split(struct page *page)
818 int ret = 0;
819 struct page *transhuge_head = page_trans_compound_anon(page);
820 if (transhuge_head) {
821 /* Get the reference on the head to split it. */
822 if (get_page_unless_zero(transhuge_head)) {
824 * Recheck we got the reference while the head
825 * was still anonymous.
827 if (PageAnon(transhuge_head))
828 ret = split_huge_page(transhuge_head);
829 else
831 * Retry later if split_huge_page run
832 * from under us.
834 ret = 1;
835 put_page(transhuge_head);
836 } else
837 /* Retry later if split_huge_page run from under us. */
838 ret = 1;
840 return ret;
844 * try_to_merge_one_page - take two pages and merge them into one
845 * @vma: the vma that holds the pte pointing to page
846 * @page: the PageAnon page that we want to replace with kpage
847 * @kpage: the PageKsm page that we want to map instead of page,
848 * or NULL the first time when we want to use page as kpage.
850 * This function returns 0 if the pages were merged, -EFAULT otherwise.
852 static int try_to_merge_one_page(struct vm_area_struct *vma,
853 struct page *page, struct page *kpage)
855 pte_t orig_pte = __pte(0);
856 int err = -EFAULT;
858 if (page == kpage) /* ksm page forked */
859 return 0;
861 if (!(vma->vm_flags & VM_MERGEABLE))
862 goto out;
863 if (PageTransCompound(page) && page_trans_compound_anon_split(page))
864 goto out;
865 BUG_ON(PageTransCompound(page));
866 if (!PageAnon(page))
867 goto out;
870 * We need the page lock to read a stable PageSwapCache in
871 * write_protect_page(). We use trylock_page() instead of
872 * lock_page() because we don't want to wait here - we
873 * prefer to continue scanning and merging different pages,
874 * then come back to this page when it is unlocked.
876 if (!trylock_page(page))
877 goto out;
879 * If this anonymous page is mapped only here, its pte may need
880 * to be write-protected. If it's mapped elsewhere, all of its
881 * ptes are necessarily already write-protected. But in either
882 * case, we need to lock and check page_count is not raised.
884 if (write_protect_page(vma, page, &orig_pte) == 0) {
885 if (!kpage) {
887 * While we hold page lock, upgrade page from
888 * PageAnon+anon_vma to PageKsm+NULL stable_node:
889 * stable_tree_insert() will update stable_node.
891 set_page_stable_node(page, NULL);
892 mark_page_accessed(page);
893 err = 0;
894 } else if (pages_identical(page, kpage))
895 err = replace_page(vma, page, kpage, orig_pte);
898 if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
899 munlock_vma_page(page);
900 if (!PageMlocked(kpage)) {
901 unlock_page(page);
902 lock_page(kpage);
903 mlock_vma_page(kpage);
904 page = kpage; /* for final unlock */
908 unlock_page(page);
909 out:
910 return err;
914 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
915 * but no new kernel page is allocated: kpage must already be a ksm page.
917 * This function returns 0 if the pages were merged, -EFAULT otherwise.
919 static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
920 struct page *page, struct page *kpage)
922 struct mm_struct *mm = rmap_item->mm;
923 struct vm_area_struct *vma;
924 int err = -EFAULT;
926 down_read(&mm->mmap_sem);
927 if (ksm_test_exit(mm))
928 goto out;
929 vma = find_vma(mm, rmap_item->address);
930 if (!vma || vma->vm_start > rmap_item->address)
931 goto out;
933 err = try_to_merge_one_page(vma, page, kpage);
934 if (err)
935 goto out;
937 /* Must get reference to anon_vma while still holding mmap_sem */
938 rmap_item->anon_vma = vma->anon_vma;
939 get_anon_vma(vma->anon_vma);
940 out:
941 up_read(&mm->mmap_sem);
942 return err;
946 * try_to_merge_two_pages - take two identical pages and prepare them
947 * to be merged into one page.
949 * This function returns the kpage if we successfully merged two identical
950 * pages into one ksm page, NULL otherwise.
952 * Note that this function upgrades page to ksm page: if one of the pages
953 * is already a ksm page, try_to_merge_with_ksm_page should be used.
955 static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
956 struct page *page,
957 struct rmap_item *tree_rmap_item,
958 struct page *tree_page)
960 int err;
962 err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
963 if (!err) {
964 err = try_to_merge_with_ksm_page(tree_rmap_item,
965 tree_page, page);
967 * If that fails, we have a ksm page with only one pte
968 * pointing to it: so break it.
970 if (err)
971 break_cow(rmap_item);
973 return err ? NULL : page;
977 * stable_tree_search - search for page inside the stable tree
979 * This function checks if there is a page inside the stable tree
980 * with identical content to the page that we are scanning right now.
982 * This function returns the stable tree node of identical content if found,
983 * NULL otherwise.
985 static struct page *stable_tree_search(struct page *page)
987 struct rb_node *node = root_stable_tree.rb_node;
988 struct stable_node *stable_node;
990 stable_node = page_stable_node(page);
991 if (stable_node) { /* ksm page forked */
992 get_page(page);
993 return page;
996 while (node) {
997 struct page *tree_page;
998 int ret;
1000 cond_resched();
1001 stable_node = rb_entry(node, struct stable_node, node);
1002 tree_page = get_ksm_page(stable_node);
1003 if (!tree_page)
1004 return NULL;
1006 ret = memcmp_pages(page, tree_page);
1008 if (ret < 0) {
1009 put_page(tree_page);
1010 node = node->rb_left;
1011 } else if (ret > 0) {
1012 put_page(tree_page);
1013 node = node->rb_right;
1014 } else
1015 return tree_page;
1018 return NULL;
1022 * stable_tree_insert - insert rmap_item pointing to new ksm page
1023 * into the stable tree.
1025 * This function returns the stable tree node just allocated on success,
1026 * NULL otherwise.
1028 static struct stable_node *stable_tree_insert(struct page *kpage)
1030 struct rb_node **new = &root_stable_tree.rb_node;
1031 struct rb_node *parent = NULL;
1032 struct stable_node *stable_node;
1034 while (*new) {
1035 struct page *tree_page;
1036 int ret;
1038 cond_resched();
1039 stable_node = rb_entry(*new, struct stable_node, node);
1040 tree_page = get_ksm_page(stable_node);
1041 if (!tree_page)
1042 return NULL;
1044 ret = memcmp_pages(kpage, tree_page);
1045 put_page(tree_page);
1047 parent = *new;
1048 if (ret < 0)
1049 new = &parent->rb_left;
1050 else if (ret > 0)
1051 new = &parent->rb_right;
1052 else {
1054 * It is not a bug that stable_tree_search() didn't
1055 * find this node: because at that time our page was
1056 * not yet write-protected, so may have changed since.
1058 return NULL;
1062 stable_node = alloc_stable_node();
1063 if (!stable_node)
1064 return NULL;
1066 rb_link_node(&stable_node->node, parent, new);
1067 rb_insert_color(&stable_node->node, &root_stable_tree);
1069 INIT_HLIST_HEAD(&stable_node->hlist);
1071 stable_node->kpfn = page_to_pfn(kpage);
1072 set_page_stable_node(kpage, stable_node);
1074 return stable_node;
1078 * unstable_tree_search_insert - search for identical page,
1079 * else insert rmap_item into the unstable tree.
1081 * This function searches for a page in the unstable tree identical to the
1082 * page currently being scanned; and if no identical page is found in the
1083 * tree, we insert rmap_item as a new object into the unstable tree.
1085 * This function returns pointer to rmap_item found to be identical
1086 * to the currently scanned page, NULL otherwise.
1088 * This function does both searching and inserting, because they share
1089 * the same walking algorithm in an rbtree.
1091 static
1092 struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1093 struct page *page,
1094 struct page **tree_pagep)
1097 struct rb_node **new = &root_unstable_tree.rb_node;
1098 struct rb_node *parent = NULL;
1100 while (*new) {
1101 struct rmap_item *tree_rmap_item;
1102 struct page *tree_page;
1103 int ret;
1105 cond_resched();
1106 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
1107 tree_page = get_mergeable_page(tree_rmap_item);
1108 if (IS_ERR_OR_NULL(tree_page))
1109 return NULL;
1112 * Don't substitute a ksm page for a forked page.
1114 if (page == tree_page) {
1115 put_page(tree_page);
1116 return NULL;
1119 ret = memcmp_pages(page, tree_page);
1121 parent = *new;
1122 if (ret < 0) {
1123 put_page(tree_page);
1124 new = &parent->rb_left;
1125 } else if (ret > 0) {
1126 put_page(tree_page);
1127 new = &parent->rb_right;
1128 } else {
1129 *tree_pagep = tree_page;
1130 return tree_rmap_item;
1134 rmap_item->address |= UNSTABLE_FLAG;
1135 rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1136 rb_link_node(&rmap_item->node, parent, new);
1137 rb_insert_color(&rmap_item->node, &root_unstable_tree);
1139 ksm_pages_unshared++;
1140 return NULL;
1144 * stable_tree_append - add another rmap_item to the linked list of
1145 * rmap_items hanging off a given node of the stable tree, all sharing
1146 * the same ksm page.
1148 static void stable_tree_append(struct rmap_item *rmap_item,
1149 struct stable_node *stable_node)
1151 rmap_item->head = stable_node;
1152 rmap_item->address |= STABLE_FLAG;
1153 hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
1155 if (rmap_item->hlist.next)
1156 ksm_pages_sharing++;
1157 else
1158 ksm_pages_shared++;
1162 * cmp_and_merge_page - first see if page can be merged into the stable tree;
1163 * if not, compare checksum to previous and if it's the same, see if page can
1164 * be inserted into the unstable tree, or merged with a page already there and
1165 * both transferred to the stable tree.
1167 * @page: the page that we are searching identical page to.
1168 * @rmap_item: the reverse mapping into the virtual address of this page
1170 static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1172 struct rmap_item *tree_rmap_item;
1173 struct page *tree_page = NULL;
1174 struct stable_node *stable_node;
1175 struct page *kpage;
1176 unsigned int checksum;
1177 int err;
1179 remove_rmap_item_from_tree(rmap_item);
1181 /* We first start with searching the page inside the stable tree */
1182 kpage = stable_tree_search(page);
1183 if (kpage) {
1184 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
1185 if (!err) {
1187 * The page was successfully merged:
1188 * add its rmap_item to the stable tree.
1190 lock_page(kpage);
1191 stable_tree_append(rmap_item, page_stable_node(kpage));
1192 unlock_page(kpage);
1194 put_page(kpage);
1195 return;
1199 * If the hash value of the page has changed from the last time
1200 * we calculated it, this page is changing frequently: therefore we
1201 * don't want to insert it in the unstable tree, and we don't want
1202 * to waste our time searching for something identical to it there.
1204 checksum = calc_checksum(page);
1205 if (rmap_item->oldchecksum != checksum) {
1206 rmap_item->oldchecksum = checksum;
1207 return;
1210 tree_rmap_item =
1211 unstable_tree_search_insert(rmap_item, page, &tree_page);
1212 if (tree_rmap_item) {
1213 kpage = try_to_merge_two_pages(rmap_item, page,
1214 tree_rmap_item, tree_page);
1215 put_page(tree_page);
1217 * As soon as we merge this page, we want to remove the
1218 * rmap_item of the page we have merged with from the unstable
1219 * tree, and insert it instead as new node in the stable tree.
1221 if (kpage) {
1222 remove_rmap_item_from_tree(tree_rmap_item);
1224 lock_page(kpage);
1225 stable_node = stable_tree_insert(kpage);
1226 if (stable_node) {
1227 stable_tree_append(tree_rmap_item, stable_node);
1228 stable_tree_append(rmap_item, stable_node);
1230 unlock_page(kpage);
1233 * If we fail to insert the page into the stable tree,
1234 * we will have 2 virtual addresses that are pointing
1235 * to a ksm page left outside the stable tree,
1236 * in which case we need to break_cow on both.
1238 if (!stable_node) {
1239 break_cow(tree_rmap_item);
1240 break_cow(rmap_item);
1246 static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
1247 struct rmap_item **rmap_list,
1248 unsigned long addr)
1250 struct rmap_item *rmap_item;
1252 while (*rmap_list) {
1253 rmap_item = *rmap_list;
1254 if ((rmap_item->address & PAGE_MASK) == addr)
1255 return rmap_item;
1256 if (rmap_item->address > addr)
1257 break;
1258 *rmap_list = rmap_item->rmap_list;
1259 remove_rmap_item_from_tree(rmap_item);
1260 free_rmap_item(rmap_item);
1263 rmap_item = alloc_rmap_item();
1264 if (rmap_item) {
1265 /* It has already been zeroed */
1266 rmap_item->mm = mm_slot->mm;
1267 rmap_item->address = addr;
1268 rmap_item->rmap_list = *rmap_list;
1269 *rmap_list = rmap_item;
1271 return rmap_item;
1274 static struct rmap_item *scan_get_next_rmap_item(struct page **page)
1276 struct mm_struct *mm;
1277 struct mm_slot *slot;
1278 struct vm_area_struct *vma;
1279 struct rmap_item *rmap_item;
1281 if (list_empty(&ksm_mm_head.mm_list))
1282 return NULL;
1284 slot = ksm_scan.mm_slot;
1285 if (slot == &ksm_mm_head) {
1287 * A number of pages can hang around indefinitely on per-cpu
1288 * pagevecs, raised page count preventing write_protect_page
1289 * from merging them. Though it doesn't really matter much,
1290 * it is puzzling to see some stuck in pages_volatile until
1291 * other activity jostles them out, and they also prevented
1292 * LTP's KSM test from succeeding deterministically; so drain
1293 * them here (here rather than on entry to ksm_do_scan(),
1294 * so we don't IPI too often when pages_to_scan is set low).
1296 lru_add_drain_all();
1298 root_unstable_tree = RB_ROOT;
1300 spin_lock(&ksm_mmlist_lock);
1301 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1302 ksm_scan.mm_slot = slot;
1303 spin_unlock(&ksm_mmlist_lock);
1304 next_mm:
1305 ksm_scan.address = 0;
1306 ksm_scan.rmap_list = &slot->rmap_list;
1309 mm = slot->mm;
1310 down_read(&mm->mmap_sem);
1311 if (ksm_test_exit(mm))
1312 vma = NULL;
1313 else
1314 vma = find_vma(mm, ksm_scan.address);
1316 for (; vma; vma = vma->vm_next) {
1317 if (!(vma->vm_flags & VM_MERGEABLE))
1318 continue;
1319 if (ksm_scan.address < vma->vm_start)
1320 ksm_scan.address = vma->vm_start;
1321 if (!vma->anon_vma)
1322 ksm_scan.address = vma->vm_end;
1324 while (ksm_scan.address < vma->vm_end) {
1325 if (ksm_test_exit(mm))
1326 break;
1327 *page = follow_page(vma, ksm_scan.address, FOLL_GET);
1328 if (IS_ERR_OR_NULL(*page)) {
1329 ksm_scan.address += PAGE_SIZE;
1330 cond_resched();
1331 continue;
1333 if (PageAnon(*page) ||
1334 page_trans_compound_anon(*page)) {
1335 flush_anon_page(vma, *page, ksm_scan.address);
1336 flush_dcache_page(*page);
1337 rmap_item = get_next_rmap_item(slot,
1338 ksm_scan.rmap_list, ksm_scan.address);
1339 if (rmap_item) {
1340 ksm_scan.rmap_list =
1341 &rmap_item->rmap_list;
1342 ksm_scan.address += PAGE_SIZE;
1343 } else
1344 put_page(*page);
1345 up_read(&mm->mmap_sem);
1346 return rmap_item;
1348 put_page(*page);
1349 ksm_scan.address += PAGE_SIZE;
1350 cond_resched();
1354 if (ksm_test_exit(mm)) {
1355 ksm_scan.address = 0;
1356 ksm_scan.rmap_list = &slot->rmap_list;
1359 * Nuke all the rmap_items that are above this current rmap:
1360 * because there were no VM_MERGEABLE vmas with such addresses.
1362 remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
1364 spin_lock(&ksm_mmlist_lock);
1365 ksm_scan.mm_slot = list_entry(slot->mm_list.next,
1366 struct mm_slot, mm_list);
1367 if (ksm_scan.address == 0) {
1369 * We've completed a full scan of all vmas, holding mmap_sem
1370 * throughout, and found no VM_MERGEABLE: so do the same as
1371 * __ksm_exit does to remove this mm from all our lists now.
1372 * This applies either when cleaning up after __ksm_exit
1373 * (but beware: we can reach here even before __ksm_exit),
1374 * or when all VM_MERGEABLE areas have been unmapped (and
1375 * mmap_sem then protects against race with MADV_MERGEABLE).
1377 hlist_del(&slot->link);
1378 list_del(&slot->mm_list);
1379 spin_unlock(&ksm_mmlist_lock);
1381 free_mm_slot(slot);
1382 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1383 up_read(&mm->mmap_sem);
1384 mmdrop(mm);
1385 } else {
1386 spin_unlock(&ksm_mmlist_lock);
1387 up_read(&mm->mmap_sem);
1390 /* Repeat until we've completed scanning the whole list */
1391 slot = ksm_scan.mm_slot;
1392 if (slot != &ksm_mm_head)
1393 goto next_mm;
1395 ksm_scan.seqnr++;
1396 return NULL;
1400 * ksm_do_scan - the ksm scanner main worker function.
1401 * @scan_npages - number of pages we want to scan before we return.
1403 static void ksm_do_scan(unsigned int scan_npages)
1405 struct rmap_item *rmap_item;
1406 struct page *uninitialized_var(page);
1408 while (scan_npages-- && likely(!freezing(current))) {
1409 cond_resched();
1410 rmap_item = scan_get_next_rmap_item(&page);
1411 if (!rmap_item)
1412 return;
1413 if (!PageKsm(page) || !in_stable_tree(rmap_item))
1414 cmp_and_merge_page(page, rmap_item);
1415 put_page(page);
1419 static int ksmd_should_run(void)
1421 return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
1424 static int ksm_scan_thread(void *nothing)
1426 set_freezable();
1427 set_user_nice(current, 5);
1429 while (!kthread_should_stop()) {
1430 mutex_lock(&ksm_thread_mutex);
1431 if (ksmd_should_run())
1432 ksm_do_scan(ksm_thread_pages_to_scan);
1433 mutex_unlock(&ksm_thread_mutex);
1435 try_to_freeze();
1437 if (ksmd_should_run()) {
1438 schedule_timeout_interruptible(
1439 msecs_to_jiffies(ksm_thread_sleep_millisecs));
1440 } else {
1441 wait_event_freezable(ksm_thread_wait,
1442 ksmd_should_run() || kthread_should_stop());
1445 return 0;
1448 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1449 unsigned long end, int advice, unsigned long *vm_flags)
1451 struct mm_struct *mm = vma->vm_mm;
1452 int err;
1454 switch (advice) {
1455 case MADV_MERGEABLE:
1457 * Be somewhat over-protective for now!
1459 if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE |
1460 VM_PFNMAP | VM_IO | VM_DONTEXPAND |
1461 VM_RESERVED | VM_HUGETLB | VM_INSERTPAGE |
1462 VM_NONLINEAR | VM_MIXEDMAP | VM_SAO))
1463 return 0; /* just ignore the advice */
1465 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
1466 err = __ksm_enter(mm);
1467 if (err)
1468 return err;
1471 *vm_flags |= VM_MERGEABLE;
1472 break;
1474 case MADV_UNMERGEABLE:
1475 if (!(*vm_flags & VM_MERGEABLE))
1476 return 0; /* just ignore the advice */
1478 if (vma->anon_vma) {
1479 err = unmerge_ksm_pages(vma, start, end);
1480 if (err)
1481 return err;
1484 *vm_flags &= ~VM_MERGEABLE;
1485 break;
1488 return 0;
1491 int __ksm_enter(struct mm_struct *mm)
1493 struct mm_slot *mm_slot;
1494 int needs_wakeup;
1496 mm_slot = alloc_mm_slot();
1497 if (!mm_slot)
1498 return -ENOMEM;
1500 /* Check ksm_run too? Would need tighter locking */
1501 needs_wakeup = list_empty(&ksm_mm_head.mm_list);
1503 spin_lock(&ksm_mmlist_lock);
1504 insert_to_mm_slots_hash(mm, mm_slot);
1506 * Insert just behind the scanning cursor, to let the area settle
1507 * down a little; when fork is followed by immediate exec, we don't
1508 * want ksmd to waste time setting up and tearing down an rmap_list.
1510 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
1511 spin_unlock(&ksm_mmlist_lock);
1513 set_bit(MMF_VM_MERGEABLE, &mm->flags);
1514 atomic_inc(&mm->mm_count);
1516 if (needs_wakeup)
1517 wake_up_interruptible(&ksm_thread_wait);
1519 return 0;
1522 void __ksm_exit(struct mm_struct *mm)
1524 struct mm_slot *mm_slot;
1525 int easy_to_free = 0;
1528 * This process is exiting: if it's straightforward (as is the
1529 * case when ksmd was never running), free mm_slot immediately.
1530 * But if it's at the cursor or has rmap_items linked to it, use
1531 * mmap_sem to synchronize with any break_cows before pagetables
1532 * are freed, and leave the mm_slot on the list for ksmd to free.
1533 * Beware: ksm may already have noticed it exiting and freed the slot.
1536 spin_lock(&ksm_mmlist_lock);
1537 mm_slot = get_mm_slot(mm);
1538 if (mm_slot && ksm_scan.mm_slot != mm_slot) {
1539 if (!mm_slot->rmap_list) {
1540 hlist_del(&mm_slot->link);
1541 list_del(&mm_slot->mm_list);
1542 easy_to_free = 1;
1543 } else {
1544 list_move(&mm_slot->mm_list,
1545 &ksm_scan.mm_slot->mm_list);
1548 spin_unlock(&ksm_mmlist_lock);
1550 if (easy_to_free) {
1551 free_mm_slot(mm_slot);
1552 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1553 mmdrop(mm);
1554 } else if (mm_slot) {
1555 down_write(&mm->mmap_sem);
1556 up_write(&mm->mmap_sem);
1560 struct page *ksm_does_need_to_copy(struct page *page,
1561 struct vm_area_struct *vma, unsigned long address)
1563 struct page *new_page;
1565 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1566 if (new_page) {
1567 copy_user_highpage(new_page, page, address, vma);
1569 SetPageDirty(new_page);
1570 __SetPageUptodate(new_page);
1571 SetPageSwapBacked(new_page);
1572 __set_page_locked(new_page);
1574 if (page_evictable(new_page, vma))
1575 lru_cache_add_lru(new_page, LRU_ACTIVE_ANON);
1576 else
1577 add_page_to_unevictable_list(new_page);
1580 return new_page;
1583 int page_referenced_ksm(struct page *page, struct mem_cgroup *memcg,
1584 unsigned long *vm_flags)
1586 struct stable_node *stable_node;
1587 struct rmap_item *rmap_item;
1588 struct hlist_node *hlist;
1589 unsigned int mapcount = page_mapcount(page);
1590 int referenced = 0;
1591 int search_new_forks = 0;
1593 VM_BUG_ON(!PageKsm(page));
1594 VM_BUG_ON(!PageLocked(page));
1596 stable_node = page_stable_node(page);
1597 if (!stable_node)
1598 return 0;
1599 again:
1600 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1601 struct anon_vma *anon_vma = rmap_item->anon_vma;
1602 struct anon_vma_chain *vmac;
1603 struct vm_area_struct *vma;
1605 anon_vma_lock(anon_vma);
1606 list_for_each_entry(vmac, &anon_vma->head, same_anon_vma) {
1607 vma = vmac->vma;
1608 if (rmap_item->address < vma->vm_start ||
1609 rmap_item->address >= vma->vm_end)
1610 continue;
1612 * Initially we examine only the vma which covers this
1613 * rmap_item; but later, if there is still work to do,
1614 * we examine covering vmas in other mms: in case they
1615 * were forked from the original since ksmd passed.
1617 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1618 continue;
1620 if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
1621 continue;
1623 referenced += page_referenced_one(page, vma,
1624 rmap_item->address, &mapcount, vm_flags);
1625 if (!search_new_forks || !mapcount)
1626 break;
1628 anon_vma_unlock(anon_vma);
1629 if (!mapcount)
1630 goto out;
1632 if (!search_new_forks++)
1633 goto again;
1634 out:
1635 return referenced;
1638 int try_to_unmap_ksm(struct page *page, enum ttu_flags flags)
1640 struct stable_node *stable_node;
1641 struct hlist_node *hlist;
1642 struct rmap_item *rmap_item;
1643 int ret = SWAP_AGAIN;
1644 int search_new_forks = 0;
1646 VM_BUG_ON(!PageKsm(page));
1647 VM_BUG_ON(!PageLocked(page));
1649 stable_node = page_stable_node(page);
1650 if (!stable_node)
1651 return SWAP_FAIL;
1652 again:
1653 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1654 struct anon_vma *anon_vma = rmap_item->anon_vma;
1655 struct anon_vma_chain *vmac;
1656 struct vm_area_struct *vma;
1658 anon_vma_lock(anon_vma);
1659 list_for_each_entry(vmac, &anon_vma->head, same_anon_vma) {
1660 vma = vmac->vma;
1661 if (rmap_item->address < vma->vm_start ||
1662 rmap_item->address >= vma->vm_end)
1663 continue;
1665 * Initially we examine only the vma which covers this
1666 * rmap_item; but later, if there is still work to do,
1667 * we examine covering vmas in other mms: in case they
1668 * were forked from the original since ksmd passed.
1670 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1671 continue;
1673 ret = try_to_unmap_one(page, vma,
1674 rmap_item->address, flags);
1675 if (ret != SWAP_AGAIN || !page_mapped(page)) {
1676 anon_vma_unlock(anon_vma);
1677 goto out;
1680 anon_vma_unlock(anon_vma);
1682 if (!search_new_forks++)
1683 goto again;
1684 out:
1685 return ret;
1688 #ifdef CONFIG_MIGRATION
1689 int rmap_walk_ksm(struct page *page, int (*rmap_one)(struct page *,
1690 struct vm_area_struct *, unsigned long, void *), void *arg)
1692 struct stable_node *stable_node;
1693 struct hlist_node *hlist;
1694 struct rmap_item *rmap_item;
1695 int ret = SWAP_AGAIN;
1696 int search_new_forks = 0;
1698 VM_BUG_ON(!PageKsm(page));
1699 VM_BUG_ON(!PageLocked(page));
1701 stable_node = page_stable_node(page);
1702 if (!stable_node)
1703 return ret;
1704 again:
1705 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1706 struct anon_vma *anon_vma = rmap_item->anon_vma;
1707 struct anon_vma_chain *vmac;
1708 struct vm_area_struct *vma;
1710 anon_vma_lock(anon_vma);
1711 list_for_each_entry(vmac, &anon_vma->head, same_anon_vma) {
1712 vma = vmac->vma;
1713 if (rmap_item->address < vma->vm_start ||
1714 rmap_item->address >= vma->vm_end)
1715 continue;
1717 * Initially we examine only the vma which covers this
1718 * rmap_item; but later, if there is still work to do,
1719 * we examine covering vmas in other mms: in case they
1720 * were forked from the original since ksmd passed.
1722 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1723 continue;
1725 ret = rmap_one(page, vma, rmap_item->address, arg);
1726 if (ret != SWAP_AGAIN) {
1727 anon_vma_unlock(anon_vma);
1728 goto out;
1731 anon_vma_unlock(anon_vma);
1733 if (!search_new_forks++)
1734 goto again;
1735 out:
1736 return ret;
1739 void ksm_migrate_page(struct page *newpage, struct page *oldpage)
1741 struct stable_node *stable_node;
1743 VM_BUG_ON(!PageLocked(oldpage));
1744 VM_BUG_ON(!PageLocked(newpage));
1745 VM_BUG_ON(newpage->mapping != oldpage->mapping);
1747 stable_node = page_stable_node(newpage);
1748 if (stable_node) {
1749 VM_BUG_ON(stable_node->kpfn != page_to_pfn(oldpage));
1750 stable_node->kpfn = page_to_pfn(newpage);
1753 #endif /* CONFIG_MIGRATION */
1755 #ifdef CONFIG_MEMORY_HOTREMOVE
1756 static struct stable_node *ksm_check_stable_tree(unsigned long start_pfn,
1757 unsigned long end_pfn)
1759 struct rb_node *node;
1761 for (node = rb_first(&root_stable_tree); node; node = rb_next(node)) {
1762 struct stable_node *stable_node;
1764 stable_node = rb_entry(node, struct stable_node, node);
1765 if (stable_node->kpfn >= start_pfn &&
1766 stable_node->kpfn < end_pfn)
1767 return stable_node;
1769 return NULL;
1772 static int ksm_memory_callback(struct notifier_block *self,
1773 unsigned long action, void *arg)
1775 struct memory_notify *mn = arg;
1776 struct stable_node *stable_node;
1778 switch (action) {
1779 case MEM_GOING_OFFLINE:
1781 * Keep it very simple for now: just lock out ksmd and
1782 * MADV_UNMERGEABLE while any memory is going offline.
1783 * mutex_lock_nested() is necessary because lockdep was alarmed
1784 * that here we take ksm_thread_mutex inside notifier chain
1785 * mutex, and later take notifier chain mutex inside
1786 * ksm_thread_mutex to unlock it. But that's safe because both
1787 * are inside mem_hotplug_mutex.
1789 mutex_lock_nested(&ksm_thread_mutex, SINGLE_DEPTH_NESTING);
1790 break;
1792 case MEM_OFFLINE:
1794 * Most of the work is done by page migration; but there might
1795 * be a few stable_nodes left over, still pointing to struct
1796 * pages which have been offlined: prune those from the tree.
1798 while ((stable_node = ksm_check_stable_tree(mn->start_pfn,
1799 mn->start_pfn + mn->nr_pages)) != NULL)
1800 remove_node_from_stable_tree(stable_node);
1801 /* fallthrough */
1803 case MEM_CANCEL_OFFLINE:
1804 mutex_unlock(&ksm_thread_mutex);
1805 break;
1807 return NOTIFY_OK;
1809 #endif /* CONFIG_MEMORY_HOTREMOVE */
1811 #ifdef CONFIG_SYSFS
1813 * This all compiles without CONFIG_SYSFS, but is a waste of space.
1816 #define KSM_ATTR_RO(_name) \
1817 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1818 #define KSM_ATTR(_name) \
1819 static struct kobj_attribute _name##_attr = \
1820 __ATTR(_name, 0644, _name##_show, _name##_store)
1822 static ssize_t sleep_millisecs_show(struct kobject *kobj,
1823 struct kobj_attribute *attr, char *buf)
1825 return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
1828 static ssize_t sleep_millisecs_store(struct kobject *kobj,
1829 struct kobj_attribute *attr,
1830 const char *buf, size_t count)
1832 unsigned long msecs;
1833 int err;
1835 err = strict_strtoul(buf, 10, &msecs);
1836 if (err || msecs > UINT_MAX)
1837 return -EINVAL;
1839 ksm_thread_sleep_millisecs = msecs;
1841 return count;
1843 KSM_ATTR(sleep_millisecs);
1845 static ssize_t pages_to_scan_show(struct kobject *kobj,
1846 struct kobj_attribute *attr, char *buf)
1848 return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
1851 static ssize_t pages_to_scan_store(struct kobject *kobj,
1852 struct kobj_attribute *attr,
1853 const char *buf, size_t count)
1855 int err;
1856 unsigned long nr_pages;
1858 err = strict_strtoul(buf, 10, &nr_pages);
1859 if (err || nr_pages > UINT_MAX)
1860 return -EINVAL;
1862 ksm_thread_pages_to_scan = nr_pages;
1864 return count;
1866 KSM_ATTR(pages_to_scan);
1868 static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
1869 char *buf)
1871 return sprintf(buf, "%u\n", ksm_run);
1874 static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
1875 const char *buf, size_t count)
1877 int err;
1878 unsigned long flags;
1880 err = strict_strtoul(buf, 10, &flags);
1881 if (err || flags > UINT_MAX)
1882 return -EINVAL;
1883 if (flags > KSM_RUN_UNMERGE)
1884 return -EINVAL;
1887 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
1888 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
1889 * breaking COW to free the pages_shared (but leaves mm_slots
1890 * on the list for when ksmd may be set running again).
1893 mutex_lock(&ksm_thread_mutex);
1894 if (ksm_run != flags) {
1895 ksm_run = flags;
1896 if (flags & KSM_RUN_UNMERGE) {
1897 current->flags |= PF_OOM_ORIGIN;
1898 err = unmerge_and_remove_all_rmap_items();
1899 current->flags &= ~PF_OOM_ORIGIN;
1900 if (err) {
1901 ksm_run = KSM_RUN_STOP;
1902 count = err;
1906 mutex_unlock(&ksm_thread_mutex);
1908 if (flags & KSM_RUN_MERGE)
1909 wake_up_interruptible(&ksm_thread_wait);
1911 return count;
1913 KSM_ATTR(run);
1915 static ssize_t pages_shared_show(struct kobject *kobj,
1916 struct kobj_attribute *attr, char *buf)
1918 return sprintf(buf, "%lu\n", ksm_pages_shared);
1920 KSM_ATTR_RO(pages_shared);
1922 static ssize_t pages_sharing_show(struct kobject *kobj,
1923 struct kobj_attribute *attr, char *buf)
1925 return sprintf(buf, "%lu\n", ksm_pages_sharing);
1927 KSM_ATTR_RO(pages_sharing);
1929 static ssize_t pages_unshared_show(struct kobject *kobj,
1930 struct kobj_attribute *attr, char *buf)
1932 return sprintf(buf, "%lu\n", ksm_pages_unshared);
1934 KSM_ATTR_RO(pages_unshared);
1936 static ssize_t pages_volatile_show(struct kobject *kobj,
1937 struct kobj_attribute *attr, char *buf)
1939 long ksm_pages_volatile;
1941 ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
1942 - ksm_pages_sharing - ksm_pages_unshared;
1944 * It was not worth any locking to calculate that statistic,
1945 * but it might therefore sometimes be negative: conceal that.
1947 if (ksm_pages_volatile < 0)
1948 ksm_pages_volatile = 0;
1949 return sprintf(buf, "%ld\n", ksm_pages_volatile);
1951 KSM_ATTR_RO(pages_volatile);
1953 static ssize_t full_scans_show(struct kobject *kobj,
1954 struct kobj_attribute *attr, char *buf)
1956 return sprintf(buf, "%lu\n", ksm_scan.seqnr);
1958 KSM_ATTR_RO(full_scans);
1960 static struct attribute *ksm_attrs[] = {
1961 &sleep_millisecs_attr.attr,
1962 &pages_to_scan_attr.attr,
1963 &run_attr.attr,
1964 &pages_shared_attr.attr,
1965 &pages_sharing_attr.attr,
1966 &pages_unshared_attr.attr,
1967 &pages_volatile_attr.attr,
1968 &full_scans_attr.attr,
1969 NULL,
1972 static struct attribute_group ksm_attr_group = {
1973 .attrs = ksm_attrs,
1974 .name = "ksm",
1976 #endif /* CONFIG_SYSFS */
1978 static int __init ksm_init(void)
1980 struct task_struct *ksm_thread;
1981 int err;
1983 err = ksm_slab_init();
1984 if (err)
1985 goto out;
1987 ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
1988 if (IS_ERR(ksm_thread)) {
1989 printk(KERN_ERR "ksm: creating kthread failed\n");
1990 err = PTR_ERR(ksm_thread);
1991 goto out_free;
1994 #ifdef CONFIG_SYSFS
1995 err = sysfs_create_group(mm_kobj, &ksm_attr_group);
1996 if (err) {
1997 printk(KERN_ERR "ksm: register sysfs failed\n");
1998 kthread_stop(ksm_thread);
1999 goto out_free;
2001 #else
2002 ksm_run = KSM_RUN_MERGE; /* no way for user to start it */
2004 #endif /* CONFIG_SYSFS */
2006 #ifdef CONFIG_MEMORY_HOTREMOVE
2008 * Choose a high priority since the callback takes ksm_thread_mutex:
2009 * later callbacks could only be taking locks which nest within that.
2011 hotplug_memory_notifier(ksm_memory_callback, 100);
2012 #endif
2013 return 0;
2015 out_free:
2016 ksm_slab_free();
2017 out:
2018 return err;
2020 module_init(ksm_init)