udldrmfb: udl_get_edid: drop unneeded i--
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / mm / ksm.c
blob51573858938d1435a5b74c2c36dbd833b5c27837
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>
38 #include <linux/oom.h>
40 #include <asm/tlbflush.h>
41 #include "internal.h"
44 * A few notes about the KSM scanning process,
45 * to make it easier to understand the data structures below:
47 * In order to reduce excessive scanning, KSM sorts the memory pages by their
48 * contents into a data structure that holds pointers to the pages' locations.
50 * Since the contents of the pages may change at any moment, KSM cannot just
51 * insert the pages into a normal sorted tree and expect it to find anything.
52 * Therefore KSM uses two data structures - the stable and the unstable tree.
54 * The stable tree holds pointers to all the merged pages (ksm pages), sorted
55 * by their contents. Because each such page is write-protected, searching on
56 * this tree is fully assured to be working (except when pages are unmapped),
57 * and therefore this tree is called the stable tree.
59 * In addition to the stable tree, KSM uses a second data structure called the
60 * unstable tree: this tree holds pointers to pages which have been found to
61 * be "unchanged for a period of time". The unstable tree sorts these pages
62 * by their contents, but since they are not write-protected, KSM cannot rely
63 * upon the unstable tree to work correctly - the unstable tree is liable to
64 * be corrupted as its contents are modified, and so it is called unstable.
66 * KSM solves this problem by several techniques:
68 * 1) The unstable tree is flushed every time KSM completes scanning all
69 * memory areas, and then the tree is rebuilt again from the beginning.
70 * 2) KSM will only insert into the unstable tree, pages whose hash value
71 * has not changed since the previous scan of all memory areas.
72 * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
73 * colors of the nodes and not on their contents, assuring that even when
74 * the tree gets "corrupted" it won't get out of balance, so scanning time
75 * remains the same (also, searching and inserting nodes in an rbtree uses
76 * the same algorithm, so we have no overhead when we flush and rebuild).
77 * 4) KSM never flushes the stable tree, which means that even if it were to
78 * take 10 attempts to find a page in the unstable tree, once it is found,
79 * it is secured in the stable tree. (When we scan a new page, we first
80 * compare it against the stable tree, and then against the unstable tree.)
83 /**
84 * struct mm_slot - ksm information per mm that is being scanned
85 * @link: link to the mm_slots hash list
86 * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
87 * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
88 * @mm: the mm that this information is valid for
90 struct mm_slot {
91 struct hlist_node link;
92 struct list_head mm_list;
93 struct rmap_item *rmap_list;
94 struct mm_struct *mm;
97 /**
98 * struct ksm_scan - cursor for scanning
99 * @mm_slot: the current mm_slot we are scanning
100 * @address: the next address inside that to be scanned
101 * @rmap_list: link to the next rmap to be scanned in the rmap_list
102 * @seqnr: count of completed full scans (needed when removing unstable node)
104 * There is only the one ksm_scan instance of this cursor structure.
106 struct ksm_scan {
107 struct mm_slot *mm_slot;
108 unsigned long address;
109 struct rmap_item **rmap_list;
110 unsigned long seqnr;
114 * struct stable_node - node of the stable rbtree
115 * @node: rb node of this ksm page in the stable tree
116 * @hlist: hlist head of rmap_items using this ksm page
117 * @kpfn: page frame number of this ksm page
119 struct stable_node {
120 struct rb_node node;
121 struct hlist_head hlist;
122 unsigned long kpfn;
126 * struct rmap_item - reverse mapping item for virtual addresses
127 * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
128 * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
129 * @mm: the memory structure this rmap_item is pointing into
130 * @address: the virtual address this rmap_item tracks (+ flags in low bits)
131 * @oldchecksum: previous checksum of the page at that virtual address
132 * @node: rb node of this rmap_item in the unstable tree
133 * @head: pointer to stable_node heading this list in the stable tree
134 * @hlist: link into hlist of rmap_items hanging off that stable_node
136 struct rmap_item {
137 struct rmap_item *rmap_list;
138 struct anon_vma *anon_vma; /* when stable */
139 struct mm_struct *mm;
140 unsigned long address; /* + low bits used for flags below */
141 unsigned int oldchecksum; /* when unstable */
142 union {
143 struct rb_node node; /* when node of unstable tree */
144 struct { /* when listed from stable tree */
145 struct stable_node *head;
146 struct hlist_node hlist;
151 #define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */
152 #define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */
153 #define STABLE_FLAG 0x200 /* is listed from the stable tree */
155 /* The stable and unstable tree heads */
156 static struct rb_root root_stable_tree = RB_ROOT;
157 static struct rb_root root_unstable_tree = RB_ROOT;
159 #define MM_SLOTS_HASH_SHIFT 10
160 #define MM_SLOTS_HASH_HEADS (1 << MM_SLOTS_HASH_SHIFT)
161 static struct hlist_head mm_slots_hash[MM_SLOTS_HASH_HEADS];
163 static struct mm_slot ksm_mm_head = {
164 .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
166 static struct ksm_scan ksm_scan = {
167 .mm_slot = &ksm_mm_head,
170 static struct kmem_cache *rmap_item_cache;
171 static struct kmem_cache *stable_node_cache;
172 static struct kmem_cache *mm_slot_cache;
174 /* The number of nodes in the stable tree */
175 static unsigned long ksm_pages_shared;
177 /* The number of page slots additionally sharing those nodes */
178 static unsigned long ksm_pages_sharing;
180 /* The number of nodes in the unstable tree */
181 static unsigned long ksm_pages_unshared;
183 /* The number of rmap_items in use: to calculate pages_volatile */
184 static unsigned long ksm_rmap_items;
186 /* Number of pages ksmd should scan in one batch */
187 static unsigned int ksm_thread_pages_to_scan = 100;
189 /* Milliseconds ksmd should sleep between batches */
190 static unsigned int ksm_thread_sleep_millisecs = 20;
192 #define KSM_RUN_STOP 0
193 #define KSM_RUN_MERGE 1
194 #define KSM_RUN_UNMERGE 2
195 static unsigned int ksm_run = KSM_RUN_STOP;
197 static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
198 static DEFINE_MUTEX(ksm_thread_mutex);
199 static DEFINE_SPINLOCK(ksm_mmlist_lock);
201 #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
202 sizeof(struct __struct), __alignof__(struct __struct),\
203 (__flags), NULL)
205 static int __init ksm_slab_init(void)
207 rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
208 if (!rmap_item_cache)
209 goto out;
211 stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
212 if (!stable_node_cache)
213 goto out_free1;
215 mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
216 if (!mm_slot_cache)
217 goto out_free2;
219 return 0;
221 out_free2:
222 kmem_cache_destroy(stable_node_cache);
223 out_free1:
224 kmem_cache_destroy(rmap_item_cache);
225 out:
226 return -ENOMEM;
229 static void __init ksm_slab_free(void)
231 kmem_cache_destroy(mm_slot_cache);
232 kmem_cache_destroy(stable_node_cache);
233 kmem_cache_destroy(rmap_item_cache);
234 mm_slot_cache = NULL;
237 static inline struct rmap_item *alloc_rmap_item(void)
239 struct rmap_item *rmap_item;
241 rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL);
242 if (rmap_item)
243 ksm_rmap_items++;
244 return rmap_item;
247 static inline void free_rmap_item(struct rmap_item *rmap_item)
249 ksm_rmap_items--;
250 rmap_item->mm = NULL; /* debug safety */
251 kmem_cache_free(rmap_item_cache, rmap_item);
254 static inline struct stable_node *alloc_stable_node(void)
256 return kmem_cache_alloc(stable_node_cache, GFP_KERNEL);
259 static inline void free_stable_node(struct stable_node *stable_node)
261 kmem_cache_free(stable_node_cache, stable_node);
264 static inline struct mm_slot *alloc_mm_slot(void)
266 if (!mm_slot_cache) /* initialization failed */
267 return NULL;
268 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
271 static inline void free_mm_slot(struct mm_slot *mm_slot)
273 kmem_cache_free(mm_slot_cache, mm_slot);
276 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
278 struct mm_slot *mm_slot;
279 struct hlist_head *bucket;
280 struct hlist_node *node;
282 bucket = &mm_slots_hash[hash_ptr(mm, MM_SLOTS_HASH_SHIFT)];
283 hlist_for_each_entry(mm_slot, node, bucket, link) {
284 if (mm == mm_slot->mm)
285 return mm_slot;
287 return NULL;
290 static void insert_to_mm_slots_hash(struct mm_struct *mm,
291 struct mm_slot *mm_slot)
293 struct hlist_head *bucket;
295 bucket = &mm_slots_hash[hash_ptr(mm, MM_SLOTS_HASH_SHIFT)];
296 mm_slot->mm = mm;
297 hlist_add_head(&mm_slot->link, bucket);
300 static inline int in_stable_tree(struct rmap_item *rmap_item)
302 return rmap_item->address & STABLE_FLAG;
306 * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
307 * page tables after it has passed through ksm_exit() - which, if necessary,
308 * takes mmap_sem briefly to serialize against them. ksm_exit() does not set
309 * a special flag: they can just back out as soon as mm_users goes to zero.
310 * ksm_test_exit() is used throughout to make this test for exit: in some
311 * places for correctness, in some places just to avoid unnecessary work.
313 static inline bool ksm_test_exit(struct mm_struct *mm)
315 return atomic_read(&mm->mm_users) == 0;
319 * We use break_ksm to break COW on a ksm page: it's a stripped down
321 * if (get_user_pages(current, mm, addr, 1, 1, 1, &page, NULL) == 1)
322 * put_page(page);
324 * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
325 * in case the application has unmapped and remapped mm,addr meanwhile.
326 * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP
327 * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
329 static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
331 struct page *page;
332 int ret = 0;
334 do {
335 cond_resched();
336 page = follow_page(vma, addr, FOLL_GET);
337 if (IS_ERR_OR_NULL(page))
338 break;
339 if (PageKsm(page))
340 ret = handle_mm_fault(vma->vm_mm, vma, addr,
341 FAULT_FLAG_WRITE);
342 else
343 ret = VM_FAULT_WRITE;
344 put_page(page);
345 } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_OOM)));
347 * We must loop because handle_mm_fault() may back out if there's
348 * any difficulty e.g. if pte accessed bit gets updated concurrently.
350 * VM_FAULT_WRITE is what we have been hoping for: it indicates that
351 * COW has been broken, even if the vma does not permit VM_WRITE;
352 * but note that a concurrent fault might break PageKsm for us.
354 * VM_FAULT_SIGBUS could occur if we race with truncation of the
355 * backing file, which also invalidates anonymous pages: that's
356 * okay, that truncation will have unmapped the PageKsm for us.
358 * VM_FAULT_OOM: at the time of writing (late July 2009), setting
359 * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
360 * current task has TIF_MEMDIE set, and will be OOM killed on return
361 * to user; and ksmd, having no mm, would never be chosen for that.
363 * But if the mm is in a limited mem_cgroup, then the fault may fail
364 * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
365 * even ksmd can fail in this way - though it's usually breaking ksm
366 * just to undo a merge it made a moment before, so unlikely to oom.
368 * That's a pity: we might therefore have more kernel pages allocated
369 * than we're counting as nodes in the stable tree; but ksm_do_scan
370 * will retry to break_cow on each pass, so should recover the page
371 * in due course. The important thing is to not let VM_MERGEABLE
372 * be cleared while any such pages might remain in the area.
374 return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
377 static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm,
378 unsigned long addr)
380 struct vm_area_struct *vma;
381 if (ksm_test_exit(mm))
382 return NULL;
383 vma = find_vma(mm, addr);
384 if (!vma || vma->vm_start > addr)
385 return NULL;
386 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
387 return NULL;
388 return vma;
391 static void break_cow(struct rmap_item *rmap_item)
393 struct mm_struct *mm = rmap_item->mm;
394 unsigned long addr = rmap_item->address;
395 struct vm_area_struct *vma;
398 * It is not an accident that whenever we want to break COW
399 * to undo, we also need to drop a reference to the anon_vma.
401 put_anon_vma(rmap_item->anon_vma);
403 down_read(&mm->mmap_sem);
404 vma = find_mergeable_vma(mm, addr);
405 if (vma)
406 break_ksm(vma, addr);
407 up_read(&mm->mmap_sem);
410 static struct page *page_trans_compound_anon(struct page *page)
412 if (PageTransCompound(page)) {
413 struct page *head = compound_trans_head(page);
415 * head may actually be splitted and freed from under
416 * us but it's ok here.
418 if (PageAnon(head))
419 return head;
421 return NULL;
424 static struct page *get_mergeable_page(struct rmap_item *rmap_item)
426 struct mm_struct *mm = rmap_item->mm;
427 unsigned long addr = rmap_item->address;
428 struct vm_area_struct *vma;
429 struct page *page;
431 down_read(&mm->mmap_sem);
432 vma = find_mergeable_vma(mm, addr);
433 if (!vma)
434 goto out;
436 page = follow_page(vma, addr, FOLL_GET);
437 if (IS_ERR_OR_NULL(page))
438 goto out;
439 if (PageAnon(page) || page_trans_compound_anon(page)) {
440 flush_anon_page(vma, page, addr);
441 flush_dcache_page(page);
442 } else {
443 put_page(page);
444 out: page = NULL;
446 up_read(&mm->mmap_sem);
447 return page;
450 static void remove_node_from_stable_tree(struct stable_node *stable_node)
452 struct rmap_item *rmap_item;
453 struct hlist_node *hlist;
455 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
456 if (rmap_item->hlist.next)
457 ksm_pages_sharing--;
458 else
459 ksm_pages_shared--;
460 put_anon_vma(rmap_item->anon_vma);
461 rmap_item->address &= PAGE_MASK;
462 cond_resched();
465 rb_erase(&stable_node->node, &root_stable_tree);
466 free_stable_node(stable_node);
470 * get_ksm_page: checks if the page indicated by the stable node
471 * is still its ksm page, despite having held no reference to it.
472 * In which case we can trust the content of the page, and it
473 * returns the gotten page; but if the page has now been zapped,
474 * remove the stale node from the stable tree and return NULL.
476 * You would expect the stable_node to hold a reference to the ksm page.
477 * But if it increments the page's count, swapping out has to wait for
478 * ksmd to come around again before it can free the page, which may take
479 * seconds or even minutes: much too unresponsive. So instead we use a
480 * "keyhole reference": access to the ksm page from the stable node peeps
481 * out through its keyhole to see if that page still holds the right key,
482 * pointing back to this stable node. This relies on freeing a PageAnon
483 * page to reset its page->mapping to NULL, and relies on no other use of
484 * a page to put something that might look like our key in page->mapping.
486 * include/linux/pagemap.h page_cache_get_speculative() is a good reference,
487 * but this is different - made simpler by ksm_thread_mutex being held, but
488 * interesting for assuming that no other use of the struct page could ever
489 * put our expected_mapping into page->mapping (or a field of the union which
490 * coincides with page->mapping). The RCU calls are not for KSM at all, but
491 * to keep the page_count protocol described with page_cache_get_speculative.
493 * Note: it is possible that get_ksm_page() will return NULL one moment,
494 * then page the next, if the page is in between page_freeze_refs() and
495 * page_unfreeze_refs(): this shouldn't be a problem anywhere, the page
496 * is on its way to being freed; but it is an anomaly to bear in mind.
498 static struct page *get_ksm_page(struct stable_node *stable_node)
500 struct page *page;
501 void *expected_mapping;
503 page = pfn_to_page(stable_node->kpfn);
504 expected_mapping = (void *)stable_node +
505 (PAGE_MAPPING_ANON | PAGE_MAPPING_KSM);
506 rcu_read_lock();
507 if (page->mapping != expected_mapping)
508 goto stale;
509 if (!get_page_unless_zero(page))
510 goto stale;
511 if (page->mapping != expected_mapping) {
512 put_page(page);
513 goto stale;
515 rcu_read_unlock();
516 return page;
517 stale:
518 rcu_read_unlock();
519 remove_node_from_stable_tree(stable_node);
520 return NULL;
524 * Removing rmap_item from stable or unstable tree.
525 * This function will clean the information from the stable/unstable tree.
527 static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
529 if (rmap_item->address & STABLE_FLAG) {
530 struct stable_node *stable_node;
531 struct page *page;
533 stable_node = rmap_item->head;
534 page = get_ksm_page(stable_node);
535 if (!page)
536 goto out;
538 lock_page(page);
539 hlist_del(&rmap_item->hlist);
540 unlock_page(page);
541 put_page(page);
543 if (stable_node->hlist.first)
544 ksm_pages_sharing--;
545 else
546 ksm_pages_shared--;
548 put_anon_vma(rmap_item->anon_vma);
549 rmap_item->address &= PAGE_MASK;
551 } else if (rmap_item->address & UNSTABLE_FLAG) {
552 unsigned char age;
554 * Usually ksmd can and must skip the rb_erase, because
555 * root_unstable_tree was already reset to RB_ROOT.
556 * But be careful when an mm is exiting: do the rb_erase
557 * if this rmap_item was inserted by this scan, rather
558 * than left over from before.
560 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
561 BUG_ON(age > 1);
562 if (!age)
563 rb_erase(&rmap_item->node, &root_unstable_tree);
565 ksm_pages_unshared--;
566 rmap_item->address &= PAGE_MASK;
568 out:
569 cond_resched(); /* we're called from many long loops */
572 static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
573 struct rmap_item **rmap_list)
575 while (*rmap_list) {
576 struct rmap_item *rmap_item = *rmap_list;
577 *rmap_list = rmap_item->rmap_list;
578 remove_rmap_item_from_tree(rmap_item);
579 free_rmap_item(rmap_item);
584 * Though it's very tempting to unmerge in_stable_tree(rmap_item)s rather
585 * than check every pte of a given vma, the locking doesn't quite work for
586 * that - an rmap_item is assigned to the stable tree after inserting ksm
587 * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing
588 * rmap_items from parent to child at fork time (so as not to waste time
589 * if exit comes before the next scan reaches it).
591 * Similarly, although we'd like to remove rmap_items (so updating counts
592 * and freeing memory) when unmerging an area, it's easier to leave that
593 * to the next pass of ksmd - consider, for example, how ksmd might be
594 * in cmp_and_merge_page on one of the rmap_items we would be removing.
596 static int unmerge_ksm_pages(struct vm_area_struct *vma,
597 unsigned long start, unsigned long end)
599 unsigned long addr;
600 int err = 0;
602 for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
603 if (ksm_test_exit(vma->vm_mm))
604 break;
605 if (signal_pending(current))
606 err = -ERESTARTSYS;
607 else
608 err = break_ksm(vma, addr);
610 return err;
613 #ifdef CONFIG_SYSFS
615 * Only called through the sysfs control interface:
617 static int unmerge_and_remove_all_rmap_items(void)
619 struct mm_slot *mm_slot;
620 struct mm_struct *mm;
621 struct vm_area_struct *vma;
622 int err = 0;
624 spin_lock(&ksm_mmlist_lock);
625 ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
626 struct mm_slot, mm_list);
627 spin_unlock(&ksm_mmlist_lock);
629 for (mm_slot = ksm_scan.mm_slot;
630 mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
631 mm = mm_slot->mm;
632 down_read(&mm->mmap_sem);
633 for (vma = mm->mmap; vma; vma = vma->vm_next) {
634 if (ksm_test_exit(mm))
635 break;
636 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
637 continue;
638 err = unmerge_ksm_pages(vma,
639 vma->vm_start, vma->vm_end);
640 if (err)
641 goto error;
644 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
646 spin_lock(&ksm_mmlist_lock);
647 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
648 struct mm_slot, mm_list);
649 if (ksm_test_exit(mm)) {
650 hlist_del(&mm_slot->link);
651 list_del(&mm_slot->mm_list);
652 spin_unlock(&ksm_mmlist_lock);
654 free_mm_slot(mm_slot);
655 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
656 up_read(&mm->mmap_sem);
657 mmdrop(mm);
658 } else {
659 spin_unlock(&ksm_mmlist_lock);
660 up_read(&mm->mmap_sem);
664 ksm_scan.seqnr = 0;
665 return 0;
667 error:
668 up_read(&mm->mmap_sem);
669 spin_lock(&ksm_mmlist_lock);
670 ksm_scan.mm_slot = &ksm_mm_head;
671 spin_unlock(&ksm_mmlist_lock);
672 return err;
674 #endif /* CONFIG_SYSFS */
676 static u32 calc_checksum(struct page *page)
678 u32 checksum;
679 void *addr = kmap_atomic(page);
680 checksum = jhash2(addr, PAGE_SIZE / 4, 17);
681 kunmap_atomic(addr);
682 return checksum;
685 static int memcmp_pages(struct page *page1, struct page *page2)
687 char *addr1, *addr2;
688 int ret;
690 addr1 = kmap_atomic(page1);
691 addr2 = kmap_atomic(page2);
692 ret = memcmp(addr1, addr2, PAGE_SIZE);
693 kunmap_atomic(addr2);
694 kunmap_atomic(addr1);
695 return ret;
698 static inline int pages_identical(struct page *page1, struct page *page2)
700 return !memcmp_pages(page1, page2);
703 static int write_protect_page(struct vm_area_struct *vma, struct page *page,
704 pte_t *orig_pte)
706 struct mm_struct *mm = vma->vm_mm;
707 unsigned long addr;
708 pte_t *ptep;
709 spinlock_t *ptl;
710 int swapped;
711 int err = -EFAULT;
712 unsigned long mmun_start; /* For mmu_notifiers */
713 unsigned long mmun_end; /* For mmu_notifiers */
715 addr = page_address_in_vma(page, vma);
716 if (addr == -EFAULT)
717 goto out;
719 BUG_ON(PageTransCompound(page));
721 mmun_start = addr;
722 mmun_end = addr + PAGE_SIZE;
723 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
725 ptep = page_check_address(page, mm, addr, &ptl, 0);
726 if (!ptep)
727 goto out_mn;
729 if (pte_write(*ptep) || pte_dirty(*ptep)) {
730 pte_t entry;
732 swapped = PageSwapCache(page);
733 flush_cache_page(vma, addr, page_to_pfn(page));
735 * Ok this is tricky, when get_user_pages_fast() run it doesn't
736 * take any lock, therefore the check that we are going to make
737 * with the pagecount against the mapcount is racey and
738 * O_DIRECT can happen right after the check.
739 * So we clear the pte and flush the tlb before the check
740 * this assure us that no O_DIRECT can happen after the check
741 * or in the middle of the check.
743 entry = ptep_clear_flush(vma, addr, ptep);
745 * Check that no O_DIRECT or similar I/O is in progress on the
746 * page
748 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
749 set_pte_at(mm, addr, ptep, entry);
750 goto out_unlock;
752 if (pte_dirty(entry))
753 set_page_dirty(page);
754 entry = pte_mkclean(pte_wrprotect(entry));
755 set_pte_at_notify(mm, addr, ptep, entry);
757 *orig_pte = *ptep;
758 err = 0;
760 out_unlock:
761 pte_unmap_unlock(ptep, ptl);
762 out_mn:
763 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
764 out:
765 return err;
769 * replace_page - replace page in vma by new ksm page
770 * @vma: vma that holds the pte pointing to page
771 * @page: the page we are replacing by kpage
772 * @kpage: the ksm page we replace page by
773 * @orig_pte: the original value of the pte
775 * Returns 0 on success, -EFAULT on failure.
777 static int replace_page(struct vm_area_struct *vma, struct page *page,
778 struct page *kpage, pte_t orig_pte)
780 struct mm_struct *mm = vma->vm_mm;
781 pmd_t *pmd;
782 pte_t *ptep;
783 spinlock_t *ptl;
784 unsigned long addr;
785 int err = -EFAULT;
786 unsigned long mmun_start; /* For mmu_notifiers */
787 unsigned long mmun_end; /* For mmu_notifiers */
789 addr = page_address_in_vma(page, vma);
790 if (addr == -EFAULT)
791 goto out;
793 pmd = mm_find_pmd(mm, addr);
794 if (!pmd)
795 goto out;
796 BUG_ON(pmd_trans_huge(*pmd));
798 mmun_start = addr;
799 mmun_end = addr + PAGE_SIZE;
800 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
802 ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
803 if (!pte_same(*ptep, orig_pte)) {
804 pte_unmap_unlock(ptep, ptl);
805 goto out_mn;
808 get_page(kpage);
809 page_add_anon_rmap(kpage, vma, addr);
811 flush_cache_page(vma, addr, pte_pfn(*ptep));
812 ptep_clear_flush(vma, addr, ptep);
813 set_pte_at_notify(mm, addr, ptep, mk_pte(kpage, vma->vm_page_prot));
815 page_remove_rmap(page);
816 if (!page_mapped(page))
817 try_to_free_swap(page);
818 put_page(page);
820 pte_unmap_unlock(ptep, ptl);
821 err = 0;
822 out_mn:
823 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
824 out:
825 return err;
828 static int page_trans_compound_anon_split(struct page *page)
830 int ret = 0;
831 struct page *transhuge_head = page_trans_compound_anon(page);
832 if (transhuge_head) {
833 /* Get the reference on the head to split it. */
834 if (get_page_unless_zero(transhuge_head)) {
836 * Recheck we got the reference while the head
837 * was still anonymous.
839 if (PageAnon(transhuge_head))
840 ret = split_huge_page(transhuge_head);
841 else
843 * Retry later if split_huge_page run
844 * from under us.
846 ret = 1;
847 put_page(transhuge_head);
848 } else
849 /* Retry later if split_huge_page run from under us. */
850 ret = 1;
852 return ret;
856 * try_to_merge_one_page - take two pages and merge them into one
857 * @vma: the vma that holds the pte pointing to page
858 * @page: the PageAnon page that we want to replace with kpage
859 * @kpage: the PageKsm page that we want to map instead of page,
860 * or NULL the first time when we want to use page as kpage.
862 * This function returns 0 if the pages were merged, -EFAULT otherwise.
864 static int try_to_merge_one_page(struct vm_area_struct *vma,
865 struct page *page, struct page *kpage)
867 pte_t orig_pte = __pte(0);
868 int err = -EFAULT;
870 if (page == kpage) /* ksm page forked */
871 return 0;
873 if (!(vma->vm_flags & VM_MERGEABLE))
874 goto out;
875 if (PageTransCompound(page) && page_trans_compound_anon_split(page))
876 goto out;
877 BUG_ON(PageTransCompound(page));
878 if (!PageAnon(page))
879 goto out;
882 * We need the page lock to read a stable PageSwapCache in
883 * write_protect_page(). We use trylock_page() instead of
884 * lock_page() because we don't want to wait here - we
885 * prefer to continue scanning and merging different pages,
886 * then come back to this page when it is unlocked.
888 if (!trylock_page(page))
889 goto out;
891 * If this anonymous page is mapped only here, its pte may need
892 * to be write-protected. If it's mapped elsewhere, all of its
893 * ptes are necessarily already write-protected. But in either
894 * case, we need to lock and check page_count is not raised.
896 if (write_protect_page(vma, page, &orig_pte) == 0) {
897 if (!kpage) {
899 * While we hold page lock, upgrade page from
900 * PageAnon+anon_vma to PageKsm+NULL stable_node:
901 * stable_tree_insert() will update stable_node.
903 set_page_stable_node(page, NULL);
904 mark_page_accessed(page);
905 err = 0;
906 } else if (pages_identical(page, kpage))
907 err = replace_page(vma, page, kpage, orig_pte);
910 if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
911 munlock_vma_page(page);
912 if (!PageMlocked(kpage)) {
913 unlock_page(page);
914 lock_page(kpage);
915 mlock_vma_page(kpage);
916 page = kpage; /* for final unlock */
920 unlock_page(page);
921 out:
922 return err;
926 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
927 * but no new kernel page is allocated: kpage must already be a ksm page.
929 * This function returns 0 if the pages were merged, -EFAULT otherwise.
931 static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
932 struct page *page, struct page *kpage)
934 struct mm_struct *mm = rmap_item->mm;
935 struct vm_area_struct *vma;
936 int err = -EFAULT;
938 down_read(&mm->mmap_sem);
939 if (ksm_test_exit(mm))
940 goto out;
941 vma = find_vma(mm, rmap_item->address);
942 if (!vma || vma->vm_start > rmap_item->address)
943 goto out;
945 err = try_to_merge_one_page(vma, page, kpage);
946 if (err)
947 goto out;
949 /* Must get reference to anon_vma while still holding mmap_sem */
950 rmap_item->anon_vma = vma->anon_vma;
951 get_anon_vma(vma->anon_vma);
952 out:
953 up_read(&mm->mmap_sem);
954 return err;
958 * try_to_merge_two_pages - take two identical pages and prepare them
959 * to be merged into one page.
961 * This function returns the kpage if we successfully merged two identical
962 * pages into one ksm page, NULL otherwise.
964 * Note that this function upgrades page to ksm page: if one of the pages
965 * is already a ksm page, try_to_merge_with_ksm_page should be used.
967 static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
968 struct page *page,
969 struct rmap_item *tree_rmap_item,
970 struct page *tree_page)
972 int err;
974 err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
975 if (!err) {
976 err = try_to_merge_with_ksm_page(tree_rmap_item,
977 tree_page, page);
979 * If that fails, we have a ksm page with only one pte
980 * pointing to it: so break it.
982 if (err)
983 break_cow(rmap_item);
985 return err ? NULL : page;
989 * stable_tree_search - search for page inside the stable tree
991 * This function checks if there is a page inside the stable tree
992 * with identical content to the page that we are scanning right now.
994 * This function returns the stable tree node of identical content if found,
995 * NULL otherwise.
997 static struct page *stable_tree_search(struct page *page)
999 struct rb_node *node = root_stable_tree.rb_node;
1000 struct stable_node *stable_node;
1002 stable_node = page_stable_node(page);
1003 if (stable_node) { /* ksm page forked */
1004 get_page(page);
1005 return page;
1008 while (node) {
1009 struct page *tree_page;
1010 int ret;
1012 cond_resched();
1013 stable_node = rb_entry(node, struct stable_node, node);
1014 tree_page = get_ksm_page(stable_node);
1015 if (!tree_page)
1016 return NULL;
1018 ret = memcmp_pages(page, tree_page);
1020 if (ret < 0) {
1021 put_page(tree_page);
1022 node = node->rb_left;
1023 } else if (ret > 0) {
1024 put_page(tree_page);
1025 node = node->rb_right;
1026 } else
1027 return tree_page;
1030 return NULL;
1034 * stable_tree_insert - insert rmap_item pointing to new ksm page
1035 * into the stable tree.
1037 * This function returns the stable tree node just allocated on success,
1038 * NULL otherwise.
1040 static struct stable_node *stable_tree_insert(struct page *kpage)
1042 struct rb_node **new = &root_stable_tree.rb_node;
1043 struct rb_node *parent = NULL;
1044 struct stable_node *stable_node;
1046 while (*new) {
1047 struct page *tree_page;
1048 int ret;
1050 cond_resched();
1051 stable_node = rb_entry(*new, struct stable_node, node);
1052 tree_page = get_ksm_page(stable_node);
1053 if (!tree_page)
1054 return NULL;
1056 ret = memcmp_pages(kpage, tree_page);
1057 put_page(tree_page);
1059 parent = *new;
1060 if (ret < 0)
1061 new = &parent->rb_left;
1062 else if (ret > 0)
1063 new = &parent->rb_right;
1064 else {
1066 * It is not a bug that stable_tree_search() didn't
1067 * find this node: because at that time our page was
1068 * not yet write-protected, so may have changed since.
1070 return NULL;
1074 stable_node = alloc_stable_node();
1075 if (!stable_node)
1076 return NULL;
1078 rb_link_node(&stable_node->node, parent, new);
1079 rb_insert_color(&stable_node->node, &root_stable_tree);
1081 INIT_HLIST_HEAD(&stable_node->hlist);
1083 stable_node->kpfn = page_to_pfn(kpage);
1084 set_page_stable_node(kpage, stable_node);
1086 return stable_node;
1090 * unstable_tree_search_insert - search for identical page,
1091 * else insert rmap_item into the unstable tree.
1093 * This function searches for a page in the unstable tree identical to the
1094 * page currently being scanned; and if no identical page is found in the
1095 * tree, we insert rmap_item as a new object into the unstable tree.
1097 * This function returns pointer to rmap_item found to be identical
1098 * to the currently scanned page, NULL otherwise.
1100 * This function does both searching and inserting, because they share
1101 * the same walking algorithm in an rbtree.
1103 static
1104 struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1105 struct page *page,
1106 struct page **tree_pagep)
1109 struct rb_node **new = &root_unstable_tree.rb_node;
1110 struct rb_node *parent = NULL;
1112 while (*new) {
1113 struct rmap_item *tree_rmap_item;
1114 struct page *tree_page;
1115 int ret;
1117 cond_resched();
1118 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
1119 tree_page = get_mergeable_page(tree_rmap_item);
1120 if (IS_ERR_OR_NULL(tree_page))
1121 return NULL;
1124 * Don't substitute a ksm page for a forked page.
1126 if (page == tree_page) {
1127 put_page(tree_page);
1128 return NULL;
1131 ret = memcmp_pages(page, tree_page);
1133 parent = *new;
1134 if (ret < 0) {
1135 put_page(tree_page);
1136 new = &parent->rb_left;
1137 } else if (ret > 0) {
1138 put_page(tree_page);
1139 new = &parent->rb_right;
1140 } else {
1141 *tree_pagep = tree_page;
1142 return tree_rmap_item;
1146 rmap_item->address |= UNSTABLE_FLAG;
1147 rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1148 rb_link_node(&rmap_item->node, parent, new);
1149 rb_insert_color(&rmap_item->node, &root_unstable_tree);
1151 ksm_pages_unshared++;
1152 return NULL;
1156 * stable_tree_append - add another rmap_item to the linked list of
1157 * rmap_items hanging off a given node of the stable tree, all sharing
1158 * the same ksm page.
1160 static void stable_tree_append(struct rmap_item *rmap_item,
1161 struct stable_node *stable_node)
1163 rmap_item->head = stable_node;
1164 rmap_item->address |= STABLE_FLAG;
1165 hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
1167 if (rmap_item->hlist.next)
1168 ksm_pages_sharing++;
1169 else
1170 ksm_pages_shared++;
1174 * cmp_and_merge_page - first see if page can be merged into the stable tree;
1175 * if not, compare checksum to previous and if it's the same, see if page can
1176 * be inserted into the unstable tree, or merged with a page already there and
1177 * both transferred to the stable tree.
1179 * @page: the page that we are searching identical page to.
1180 * @rmap_item: the reverse mapping into the virtual address of this page
1182 static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1184 struct rmap_item *tree_rmap_item;
1185 struct page *tree_page = NULL;
1186 struct stable_node *stable_node;
1187 struct page *kpage;
1188 unsigned int checksum;
1189 int err;
1191 remove_rmap_item_from_tree(rmap_item);
1193 /* We first start with searching the page inside the stable tree */
1194 kpage = stable_tree_search(page);
1195 if (kpage) {
1196 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
1197 if (!err) {
1199 * The page was successfully merged:
1200 * add its rmap_item to the stable tree.
1202 lock_page(kpage);
1203 stable_tree_append(rmap_item, page_stable_node(kpage));
1204 unlock_page(kpage);
1206 put_page(kpage);
1207 return;
1211 * If the hash value of the page has changed from the last time
1212 * we calculated it, this page is changing frequently: therefore we
1213 * don't want to insert it in the unstable tree, and we don't want
1214 * to waste our time searching for something identical to it there.
1216 checksum = calc_checksum(page);
1217 if (rmap_item->oldchecksum != checksum) {
1218 rmap_item->oldchecksum = checksum;
1219 return;
1222 tree_rmap_item =
1223 unstable_tree_search_insert(rmap_item, page, &tree_page);
1224 if (tree_rmap_item) {
1225 kpage = try_to_merge_two_pages(rmap_item, page,
1226 tree_rmap_item, tree_page);
1227 put_page(tree_page);
1229 * As soon as we merge this page, we want to remove the
1230 * rmap_item of the page we have merged with from the unstable
1231 * tree, and insert it instead as new node in the stable tree.
1233 if (kpage) {
1234 remove_rmap_item_from_tree(tree_rmap_item);
1236 lock_page(kpage);
1237 stable_node = stable_tree_insert(kpage);
1238 if (stable_node) {
1239 stable_tree_append(tree_rmap_item, stable_node);
1240 stable_tree_append(rmap_item, stable_node);
1242 unlock_page(kpage);
1245 * If we fail to insert the page into the stable tree,
1246 * we will have 2 virtual addresses that are pointing
1247 * to a ksm page left outside the stable tree,
1248 * in which case we need to break_cow on both.
1250 if (!stable_node) {
1251 break_cow(tree_rmap_item);
1252 break_cow(rmap_item);
1258 static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
1259 struct rmap_item **rmap_list,
1260 unsigned long addr)
1262 struct rmap_item *rmap_item;
1264 while (*rmap_list) {
1265 rmap_item = *rmap_list;
1266 if ((rmap_item->address & PAGE_MASK) == addr)
1267 return rmap_item;
1268 if (rmap_item->address > addr)
1269 break;
1270 *rmap_list = rmap_item->rmap_list;
1271 remove_rmap_item_from_tree(rmap_item);
1272 free_rmap_item(rmap_item);
1275 rmap_item = alloc_rmap_item();
1276 if (rmap_item) {
1277 /* It has already been zeroed */
1278 rmap_item->mm = mm_slot->mm;
1279 rmap_item->address = addr;
1280 rmap_item->rmap_list = *rmap_list;
1281 *rmap_list = rmap_item;
1283 return rmap_item;
1286 static struct rmap_item *scan_get_next_rmap_item(struct page **page)
1288 struct mm_struct *mm;
1289 struct mm_slot *slot;
1290 struct vm_area_struct *vma;
1291 struct rmap_item *rmap_item;
1293 if (list_empty(&ksm_mm_head.mm_list))
1294 return NULL;
1296 slot = ksm_scan.mm_slot;
1297 if (slot == &ksm_mm_head) {
1299 * A number of pages can hang around indefinitely on per-cpu
1300 * pagevecs, raised page count preventing write_protect_page
1301 * from merging them. Though it doesn't really matter much,
1302 * it is puzzling to see some stuck in pages_volatile until
1303 * other activity jostles them out, and they also prevented
1304 * LTP's KSM test from succeeding deterministically; so drain
1305 * them here (here rather than on entry to ksm_do_scan(),
1306 * so we don't IPI too often when pages_to_scan is set low).
1308 lru_add_drain_all();
1310 root_unstable_tree = RB_ROOT;
1312 spin_lock(&ksm_mmlist_lock);
1313 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1314 ksm_scan.mm_slot = slot;
1315 spin_unlock(&ksm_mmlist_lock);
1317 * Although we tested list_empty() above, a racing __ksm_exit
1318 * of the last mm on the list may have removed it since then.
1320 if (slot == &ksm_mm_head)
1321 return NULL;
1322 next_mm:
1323 ksm_scan.address = 0;
1324 ksm_scan.rmap_list = &slot->rmap_list;
1327 mm = slot->mm;
1328 down_read(&mm->mmap_sem);
1329 if (ksm_test_exit(mm))
1330 vma = NULL;
1331 else
1332 vma = find_vma(mm, ksm_scan.address);
1334 for (; vma; vma = vma->vm_next) {
1335 if (!(vma->vm_flags & VM_MERGEABLE))
1336 continue;
1337 if (ksm_scan.address < vma->vm_start)
1338 ksm_scan.address = vma->vm_start;
1339 if (!vma->anon_vma)
1340 ksm_scan.address = vma->vm_end;
1342 while (ksm_scan.address < vma->vm_end) {
1343 if (ksm_test_exit(mm))
1344 break;
1345 *page = follow_page(vma, ksm_scan.address, FOLL_GET);
1346 if (IS_ERR_OR_NULL(*page)) {
1347 ksm_scan.address += PAGE_SIZE;
1348 cond_resched();
1349 continue;
1351 if (PageAnon(*page) ||
1352 page_trans_compound_anon(*page)) {
1353 flush_anon_page(vma, *page, ksm_scan.address);
1354 flush_dcache_page(*page);
1355 rmap_item = get_next_rmap_item(slot,
1356 ksm_scan.rmap_list, ksm_scan.address);
1357 if (rmap_item) {
1358 ksm_scan.rmap_list =
1359 &rmap_item->rmap_list;
1360 ksm_scan.address += PAGE_SIZE;
1361 } else
1362 put_page(*page);
1363 up_read(&mm->mmap_sem);
1364 return rmap_item;
1366 put_page(*page);
1367 ksm_scan.address += PAGE_SIZE;
1368 cond_resched();
1372 if (ksm_test_exit(mm)) {
1373 ksm_scan.address = 0;
1374 ksm_scan.rmap_list = &slot->rmap_list;
1377 * Nuke all the rmap_items that are above this current rmap:
1378 * because there were no VM_MERGEABLE vmas with such addresses.
1380 remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
1382 spin_lock(&ksm_mmlist_lock);
1383 ksm_scan.mm_slot = list_entry(slot->mm_list.next,
1384 struct mm_slot, mm_list);
1385 if (ksm_scan.address == 0) {
1387 * We've completed a full scan of all vmas, holding mmap_sem
1388 * throughout, and found no VM_MERGEABLE: so do the same as
1389 * __ksm_exit does to remove this mm from all our lists now.
1390 * This applies either when cleaning up after __ksm_exit
1391 * (but beware: we can reach here even before __ksm_exit),
1392 * or when all VM_MERGEABLE areas have been unmapped (and
1393 * mmap_sem then protects against race with MADV_MERGEABLE).
1395 hlist_del(&slot->link);
1396 list_del(&slot->mm_list);
1397 spin_unlock(&ksm_mmlist_lock);
1399 free_mm_slot(slot);
1400 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1401 up_read(&mm->mmap_sem);
1402 mmdrop(mm);
1403 } else {
1404 spin_unlock(&ksm_mmlist_lock);
1405 up_read(&mm->mmap_sem);
1408 /* Repeat until we've completed scanning the whole list */
1409 slot = ksm_scan.mm_slot;
1410 if (slot != &ksm_mm_head)
1411 goto next_mm;
1413 ksm_scan.seqnr++;
1414 return NULL;
1418 * ksm_do_scan - the ksm scanner main worker function.
1419 * @scan_npages - number of pages we want to scan before we return.
1421 static void ksm_do_scan(unsigned int scan_npages)
1423 struct rmap_item *rmap_item;
1424 struct page *uninitialized_var(page);
1426 while (scan_npages-- && likely(!freezing(current))) {
1427 cond_resched();
1428 rmap_item = scan_get_next_rmap_item(&page);
1429 if (!rmap_item)
1430 return;
1431 if (!PageKsm(page) || !in_stable_tree(rmap_item))
1432 cmp_and_merge_page(page, rmap_item);
1433 put_page(page);
1437 static int ksmd_should_run(void)
1439 return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
1442 static int ksm_scan_thread(void *nothing)
1444 set_freezable();
1445 set_user_nice(current, 5);
1447 while (!kthread_should_stop()) {
1448 mutex_lock(&ksm_thread_mutex);
1449 if (ksmd_should_run())
1450 ksm_do_scan(ksm_thread_pages_to_scan);
1451 mutex_unlock(&ksm_thread_mutex);
1453 try_to_freeze();
1455 if (ksmd_should_run()) {
1456 schedule_timeout_interruptible(
1457 msecs_to_jiffies(ksm_thread_sleep_millisecs));
1458 } else {
1459 wait_event_freezable(ksm_thread_wait,
1460 ksmd_should_run() || kthread_should_stop());
1463 return 0;
1466 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1467 unsigned long end, int advice, unsigned long *vm_flags)
1469 struct mm_struct *mm = vma->vm_mm;
1470 int err;
1472 switch (advice) {
1473 case MADV_MERGEABLE:
1475 * Be somewhat over-protective for now!
1477 if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE |
1478 VM_PFNMAP | VM_IO | VM_DONTEXPAND |
1479 VM_HUGETLB | VM_NONLINEAR | VM_MIXEDMAP))
1480 return 0; /* just ignore the advice */
1482 #ifdef VM_SAO
1483 if (*vm_flags & VM_SAO)
1484 return 0;
1485 #endif
1487 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
1488 err = __ksm_enter(mm);
1489 if (err)
1490 return err;
1493 *vm_flags |= VM_MERGEABLE;
1494 break;
1496 case MADV_UNMERGEABLE:
1497 if (!(*vm_flags & VM_MERGEABLE))
1498 return 0; /* just ignore the advice */
1500 if (vma->anon_vma) {
1501 err = unmerge_ksm_pages(vma, start, end);
1502 if (err)
1503 return err;
1506 *vm_flags &= ~VM_MERGEABLE;
1507 break;
1510 return 0;
1513 int __ksm_enter(struct mm_struct *mm)
1515 struct mm_slot *mm_slot;
1516 int needs_wakeup;
1518 mm_slot = alloc_mm_slot();
1519 if (!mm_slot)
1520 return -ENOMEM;
1522 /* Check ksm_run too? Would need tighter locking */
1523 needs_wakeup = list_empty(&ksm_mm_head.mm_list);
1525 spin_lock(&ksm_mmlist_lock);
1526 insert_to_mm_slots_hash(mm, mm_slot);
1528 * Insert just behind the scanning cursor, to let the area settle
1529 * down a little; when fork is followed by immediate exec, we don't
1530 * want ksmd to waste time setting up and tearing down an rmap_list.
1532 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
1533 spin_unlock(&ksm_mmlist_lock);
1535 set_bit(MMF_VM_MERGEABLE, &mm->flags);
1536 atomic_inc(&mm->mm_count);
1538 if (needs_wakeup)
1539 wake_up_interruptible(&ksm_thread_wait);
1541 return 0;
1544 void __ksm_exit(struct mm_struct *mm)
1546 struct mm_slot *mm_slot;
1547 int easy_to_free = 0;
1550 * This process is exiting: if it's straightforward (as is the
1551 * case when ksmd was never running), free mm_slot immediately.
1552 * But if it's at the cursor or has rmap_items linked to it, use
1553 * mmap_sem to synchronize with any break_cows before pagetables
1554 * are freed, and leave the mm_slot on the list for ksmd to free.
1555 * Beware: ksm may already have noticed it exiting and freed the slot.
1558 spin_lock(&ksm_mmlist_lock);
1559 mm_slot = get_mm_slot(mm);
1560 if (mm_slot && ksm_scan.mm_slot != mm_slot) {
1561 if (!mm_slot->rmap_list) {
1562 hlist_del(&mm_slot->link);
1563 list_del(&mm_slot->mm_list);
1564 easy_to_free = 1;
1565 } else {
1566 list_move(&mm_slot->mm_list,
1567 &ksm_scan.mm_slot->mm_list);
1570 spin_unlock(&ksm_mmlist_lock);
1572 if (easy_to_free) {
1573 free_mm_slot(mm_slot);
1574 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1575 mmdrop(mm);
1576 } else if (mm_slot) {
1577 down_write(&mm->mmap_sem);
1578 up_write(&mm->mmap_sem);
1582 struct page *ksm_does_need_to_copy(struct page *page,
1583 struct vm_area_struct *vma, unsigned long address)
1585 struct page *new_page;
1587 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1588 if (new_page) {
1589 copy_user_highpage(new_page, page, address, vma);
1591 SetPageDirty(new_page);
1592 __SetPageUptodate(new_page);
1593 SetPageSwapBacked(new_page);
1594 __set_page_locked(new_page);
1596 if (!mlocked_vma_newpage(vma, new_page))
1597 lru_cache_add_lru(new_page, LRU_ACTIVE_ANON);
1598 else
1599 add_page_to_unevictable_list(new_page);
1602 return new_page;
1605 int page_referenced_ksm(struct page *page, struct mem_cgroup *memcg,
1606 unsigned long *vm_flags)
1608 struct stable_node *stable_node;
1609 struct rmap_item *rmap_item;
1610 struct hlist_node *hlist;
1611 unsigned int mapcount = page_mapcount(page);
1612 int referenced = 0;
1613 int search_new_forks = 0;
1615 VM_BUG_ON(!PageKsm(page));
1616 VM_BUG_ON(!PageLocked(page));
1618 stable_node = page_stable_node(page);
1619 if (!stable_node)
1620 return 0;
1621 again:
1622 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1623 struct anon_vma *anon_vma = rmap_item->anon_vma;
1624 struct anon_vma_chain *vmac;
1625 struct vm_area_struct *vma;
1627 anon_vma_lock_read(anon_vma);
1628 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
1629 0, ULONG_MAX) {
1630 vma = vmac->vma;
1631 if (rmap_item->address < vma->vm_start ||
1632 rmap_item->address >= vma->vm_end)
1633 continue;
1635 * Initially we examine only the vma which covers this
1636 * rmap_item; but later, if there is still work to do,
1637 * we examine covering vmas in other mms: in case they
1638 * were forked from the original since ksmd passed.
1640 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1641 continue;
1643 if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
1644 continue;
1646 referenced += page_referenced_one(page, vma,
1647 rmap_item->address, &mapcount, vm_flags);
1648 if (!search_new_forks || !mapcount)
1649 break;
1651 anon_vma_unlock_read(anon_vma);
1652 if (!mapcount)
1653 goto out;
1655 if (!search_new_forks++)
1656 goto again;
1657 out:
1658 return referenced;
1661 int try_to_unmap_ksm(struct page *page, enum ttu_flags flags)
1663 struct stable_node *stable_node;
1664 struct hlist_node *hlist;
1665 struct rmap_item *rmap_item;
1666 int ret = SWAP_AGAIN;
1667 int search_new_forks = 0;
1669 VM_BUG_ON(!PageKsm(page));
1670 VM_BUG_ON(!PageLocked(page));
1672 stable_node = page_stable_node(page);
1673 if (!stable_node)
1674 return SWAP_FAIL;
1675 again:
1676 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1677 struct anon_vma *anon_vma = rmap_item->anon_vma;
1678 struct anon_vma_chain *vmac;
1679 struct vm_area_struct *vma;
1681 anon_vma_lock_read(anon_vma);
1682 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
1683 0, ULONG_MAX) {
1684 vma = vmac->vma;
1685 if (rmap_item->address < vma->vm_start ||
1686 rmap_item->address >= vma->vm_end)
1687 continue;
1689 * Initially we examine only the vma which covers this
1690 * rmap_item; but later, if there is still work to do,
1691 * we examine covering vmas in other mms: in case they
1692 * were forked from the original since ksmd passed.
1694 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1695 continue;
1697 ret = try_to_unmap_one(page, vma,
1698 rmap_item->address, flags);
1699 if (ret != SWAP_AGAIN || !page_mapped(page)) {
1700 anon_vma_unlock_read(anon_vma);
1701 goto out;
1704 anon_vma_unlock_read(anon_vma);
1706 if (!search_new_forks++)
1707 goto again;
1708 out:
1709 return ret;
1712 #ifdef CONFIG_MIGRATION
1713 int rmap_walk_ksm(struct page *page, int (*rmap_one)(struct page *,
1714 struct vm_area_struct *, unsigned long, void *), void *arg)
1716 struct stable_node *stable_node;
1717 struct hlist_node *hlist;
1718 struct rmap_item *rmap_item;
1719 int ret = SWAP_AGAIN;
1720 int search_new_forks = 0;
1722 VM_BUG_ON(!PageKsm(page));
1723 VM_BUG_ON(!PageLocked(page));
1725 stable_node = page_stable_node(page);
1726 if (!stable_node)
1727 return ret;
1728 again:
1729 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1730 struct anon_vma *anon_vma = rmap_item->anon_vma;
1731 struct anon_vma_chain *vmac;
1732 struct vm_area_struct *vma;
1734 anon_vma_lock_read(anon_vma);
1735 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
1736 0, ULONG_MAX) {
1737 vma = vmac->vma;
1738 if (rmap_item->address < vma->vm_start ||
1739 rmap_item->address >= vma->vm_end)
1740 continue;
1742 * Initially we examine only the vma which covers this
1743 * rmap_item; but later, if there is still work to do,
1744 * we examine covering vmas in other mms: in case they
1745 * were forked from the original since ksmd passed.
1747 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1748 continue;
1750 ret = rmap_one(page, vma, rmap_item->address, arg);
1751 if (ret != SWAP_AGAIN) {
1752 anon_vma_unlock_read(anon_vma);
1753 goto out;
1756 anon_vma_unlock_read(anon_vma);
1758 if (!search_new_forks++)
1759 goto again;
1760 out:
1761 return ret;
1764 void ksm_migrate_page(struct page *newpage, struct page *oldpage)
1766 struct stable_node *stable_node;
1768 VM_BUG_ON(!PageLocked(oldpage));
1769 VM_BUG_ON(!PageLocked(newpage));
1770 VM_BUG_ON(newpage->mapping != oldpage->mapping);
1772 stable_node = page_stable_node(newpage);
1773 if (stable_node) {
1774 VM_BUG_ON(stable_node->kpfn != page_to_pfn(oldpage));
1775 stable_node->kpfn = page_to_pfn(newpage);
1778 #endif /* CONFIG_MIGRATION */
1780 #ifdef CONFIG_MEMORY_HOTREMOVE
1781 static struct stable_node *ksm_check_stable_tree(unsigned long start_pfn,
1782 unsigned long end_pfn)
1784 struct rb_node *node;
1786 for (node = rb_first(&root_stable_tree); node; node = rb_next(node)) {
1787 struct stable_node *stable_node;
1789 stable_node = rb_entry(node, struct stable_node, node);
1790 if (stable_node->kpfn >= start_pfn &&
1791 stable_node->kpfn < end_pfn)
1792 return stable_node;
1794 return NULL;
1797 static int ksm_memory_callback(struct notifier_block *self,
1798 unsigned long action, void *arg)
1800 struct memory_notify *mn = arg;
1801 struct stable_node *stable_node;
1803 switch (action) {
1804 case MEM_GOING_OFFLINE:
1806 * Keep it very simple for now: just lock out ksmd and
1807 * MADV_UNMERGEABLE while any memory is going offline.
1808 * mutex_lock_nested() is necessary because lockdep was alarmed
1809 * that here we take ksm_thread_mutex inside notifier chain
1810 * mutex, and later take notifier chain mutex inside
1811 * ksm_thread_mutex to unlock it. But that's safe because both
1812 * are inside mem_hotplug_mutex.
1814 mutex_lock_nested(&ksm_thread_mutex, SINGLE_DEPTH_NESTING);
1815 break;
1817 case MEM_OFFLINE:
1819 * Most of the work is done by page migration; but there might
1820 * be a few stable_nodes left over, still pointing to struct
1821 * pages which have been offlined: prune those from the tree.
1823 while ((stable_node = ksm_check_stable_tree(mn->start_pfn,
1824 mn->start_pfn + mn->nr_pages)) != NULL)
1825 remove_node_from_stable_tree(stable_node);
1826 /* fallthrough */
1828 case MEM_CANCEL_OFFLINE:
1829 mutex_unlock(&ksm_thread_mutex);
1830 break;
1832 return NOTIFY_OK;
1834 #endif /* CONFIG_MEMORY_HOTREMOVE */
1836 #ifdef CONFIG_SYSFS
1838 * This all compiles without CONFIG_SYSFS, but is a waste of space.
1841 #define KSM_ATTR_RO(_name) \
1842 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1843 #define KSM_ATTR(_name) \
1844 static struct kobj_attribute _name##_attr = \
1845 __ATTR(_name, 0644, _name##_show, _name##_store)
1847 static ssize_t sleep_millisecs_show(struct kobject *kobj,
1848 struct kobj_attribute *attr, char *buf)
1850 return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
1853 static ssize_t sleep_millisecs_store(struct kobject *kobj,
1854 struct kobj_attribute *attr,
1855 const char *buf, size_t count)
1857 unsigned long msecs;
1858 int err;
1860 err = strict_strtoul(buf, 10, &msecs);
1861 if (err || msecs > UINT_MAX)
1862 return -EINVAL;
1864 ksm_thread_sleep_millisecs = msecs;
1866 return count;
1868 KSM_ATTR(sleep_millisecs);
1870 static ssize_t pages_to_scan_show(struct kobject *kobj,
1871 struct kobj_attribute *attr, char *buf)
1873 return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
1876 static ssize_t pages_to_scan_store(struct kobject *kobj,
1877 struct kobj_attribute *attr,
1878 const char *buf, size_t count)
1880 int err;
1881 unsigned long nr_pages;
1883 err = strict_strtoul(buf, 10, &nr_pages);
1884 if (err || nr_pages > UINT_MAX)
1885 return -EINVAL;
1887 ksm_thread_pages_to_scan = nr_pages;
1889 return count;
1891 KSM_ATTR(pages_to_scan);
1893 static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
1894 char *buf)
1896 return sprintf(buf, "%u\n", ksm_run);
1899 static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
1900 const char *buf, size_t count)
1902 int err;
1903 unsigned long flags;
1905 err = strict_strtoul(buf, 10, &flags);
1906 if (err || flags > UINT_MAX)
1907 return -EINVAL;
1908 if (flags > KSM_RUN_UNMERGE)
1909 return -EINVAL;
1912 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
1913 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
1914 * breaking COW to free the pages_shared (but leaves mm_slots
1915 * on the list for when ksmd may be set running again).
1918 mutex_lock(&ksm_thread_mutex);
1919 if (ksm_run != flags) {
1920 ksm_run = flags;
1921 if (flags & KSM_RUN_UNMERGE) {
1922 set_current_oom_origin();
1923 err = unmerge_and_remove_all_rmap_items();
1924 clear_current_oom_origin();
1925 if (err) {
1926 ksm_run = KSM_RUN_STOP;
1927 count = err;
1931 mutex_unlock(&ksm_thread_mutex);
1933 if (flags & KSM_RUN_MERGE)
1934 wake_up_interruptible(&ksm_thread_wait);
1936 return count;
1938 KSM_ATTR(run);
1940 static ssize_t pages_shared_show(struct kobject *kobj,
1941 struct kobj_attribute *attr, char *buf)
1943 return sprintf(buf, "%lu\n", ksm_pages_shared);
1945 KSM_ATTR_RO(pages_shared);
1947 static ssize_t pages_sharing_show(struct kobject *kobj,
1948 struct kobj_attribute *attr, char *buf)
1950 return sprintf(buf, "%lu\n", ksm_pages_sharing);
1952 KSM_ATTR_RO(pages_sharing);
1954 static ssize_t pages_unshared_show(struct kobject *kobj,
1955 struct kobj_attribute *attr, char *buf)
1957 return sprintf(buf, "%lu\n", ksm_pages_unshared);
1959 KSM_ATTR_RO(pages_unshared);
1961 static ssize_t pages_volatile_show(struct kobject *kobj,
1962 struct kobj_attribute *attr, char *buf)
1964 long ksm_pages_volatile;
1966 ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
1967 - ksm_pages_sharing - ksm_pages_unshared;
1969 * It was not worth any locking to calculate that statistic,
1970 * but it might therefore sometimes be negative: conceal that.
1972 if (ksm_pages_volatile < 0)
1973 ksm_pages_volatile = 0;
1974 return sprintf(buf, "%ld\n", ksm_pages_volatile);
1976 KSM_ATTR_RO(pages_volatile);
1978 static ssize_t full_scans_show(struct kobject *kobj,
1979 struct kobj_attribute *attr, char *buf)
1981 return sprintf(buf, "%lu\n", ksm_scan.seqnr);
1983 KSM_ATTR_RO(full_scans);
1985 static struct attribute *ksm_attrs[] = {
1986 &sleep_millisecs_attr.attr,
1987 &pages_to_scan_attr.attr,
1988 &run_attr.attr,
1989 &pages_shared_attr.attr,
1990 &pages_sharing_attr.attr,
1991 &pages_unshared_attr.attr,
1992 &pages_volatile_attr.attr,
1993 &full_scans_attr.attr,
1994 NULL,
1997 static struct attribute_group ksm_attr_group = {
1998 .attrs = ksm_attrs,
1999 .name = "ksm",
2001 #endif /* CONFIG_SYSFS */
2003 static int __init ksm_init(void)
2005 struct task_struct *ksm_thread;
2006 int err;
2008 err = ksm_slab_init();
2009 if (err)
2010 goto out;
2012 ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
2013 if (IS_ERR(ksm_thread)) {
2014 printk(KERN_ERR "ksm: creating kthread failed\n");
2015 err = PTR_ERR(ksm_thread);
2016 goto out_free;
2019 #ifdef CONFIG_SYSFS
2020 err = sysfs_create_group(mm_kobj, &ksm_attr_group);
2021 if (err) {
2022 printk(KERN_ERR "ksm: register sysfs failed\n");
2023 kthread_stop(ksm_thread);
2024 goto out_free;
2026 #else
2027 ksm_run = KSM_RUN_MERGE; /* no way for user to start it */
2029 #endif /* CONFIG_SYSFS */
2031 #ifdef CONFIG_MEMORY_HOTREMOVE
2033 * Choose a high priority since the callback takes ksm_thread_mutex:
2034 * later callbacks could only be taking locks which nest within that.
2036 hotplug_memory_notifier(ksm_memory_callback, 100);
2037 #endif
2038 return 0;
2040 out_free:
2041 ksm_slab_free();
2042 out:
2043 return err;
2045 module_init(ksm_init)