Loading...
1/*
2 * Memory merging support.
3 *
4 * This code enables dynamic sharing of identical pages found in different
5 * memory areas, even if they are not shared by fork()
6 *
7 * Copyright (C) 2008-2009 Red Hat, Inc.
8 * Authors:
9 * Izik Eidus
10 * Andrea Arcangeli
11 * Chris Wright
12 * Hugh Dickins
13 *
14 * This work is licensed under the terms of the GNU GPL, version 2.
15 */
16
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>
39
40#include <asm/tlbflush.h>
41#include "internal.h"
42
43/*
44 * A few notes about the KSM scanning process,
45 * to make it easier to understand the data structures below:
46 *
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.
49 *
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.
53 *
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.
58 *
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.
65 *
66 * KSM solves this problem by several techniques:
67 *
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.)
81 */
82
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
89 */
90struct mm_slot {
91 struct hlist_node link;
92 struct list_head mm_list;
93 struct rmap_item *rmap_list;
94 struct mm_struct *mm;
95};
96
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)
103 *
104 * There is only the one ksm_scan instance of this cursor structure.
105 */
106struct ksm_scan {
107 struct mm_slot *mm_slot;
108 unsigned long address;
109 struct rmap_item **rmap_list;
110 unsigned long seqnr;
111};
112
113/**
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
118 */
119struct stable_node {
120 struct rb_node node;
121 struct hlist_head hlist;
122 unsigned long kpfn;
123};
124
125/**
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
135 */
136struct 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;
147 };
148 };
149};
150
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 */
154
155/* The stable and unstable tree heads */
156static struct rb_root root_stable_tree = RB_ROOT;
157static struct rb_root root_unstable_tree = RB_ROOT;
158
159#define MM_SLOTS_HASH_SHIFT 10
160#define MM_SLOTS_HASH_HEADS (1 << MM_SLOTS_HASH_SHIFT)
161static struct hlist_head mm_slots_hash[MM_SLOTS_HASH_HEADS];
162
163static struct mm_slot ksm_mm_head = {
164 .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
165};
166static struct ksm_scan ksm_scan = {
167 .mm_slot = &ksm_mm_head,
168};
169
170static struct kmem_cache *rmap_item_cache;
171static struct kmem_cache *stable_node_cache;
172static struct kmem_cache *mm_slot_cache;
173
174/* The number of nodes in the stable tree */
175static unsigned long ksm_pages_shared;
176
177/* The number of page slots additionally sharing those nodes */
178static unsigned long ksm_pages_sharing;
179
180/* The number of nodes in the unstable tree */
181static unsigned long ksm_pages_unshared;
182
183/* The number of rmap_items in use: to calculate pages_volatile */
184static unsigned long ksm_rmap_items;
185
186/* Number of pages ksmd should scan in one batch */
187static unsigned int ksm_thread_pages_to_scan = 100;
188
189/* Milliseconds ksmd should sleep between batches */
190static unsigned int ksm_thread_sleep_millisecs = 20;
191
192#define KSM_RUN_STOP 0
193#define KSM_RUN_MERGE 1
194#define KSM_RUN_UNMERGE 2
195static unsigned int ksm_run = KSM_RUN_STOP;
196
197static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
198static DEFINE_MUTEX(ksm_thread_mutex);
199static DEFINE_SPINLOCK(ksm_mmlist_lock);
200
201#define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
202 sizeof(struct __struct), __alignof__(struct __struct),\
203 (__flags), NULL)
204
205static int __init ksm_slab_init(void)
206{
207 rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
208 if (!rmap_item_cache)
209 goto out;
210
211 stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
212 if (!stable_node_cache)
213 goto out_free1;
214
215 mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
216 if (!mm_slot_cache)
217 goto out_free2;
218
219 return 0;
220
221out_free2:
222 kmem_cache_destroy(stable_node_cache);
223out_free1:
224 kmem_cache_destroy(rmap_item_cache);
225out:
226 return -ENOMEM;
227}
228
229static void __init ksm_slab_free(void)
230{
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;
235}
236
237static inline struct rmap_item *alloc_rmap_item(void)
238{
239 struct rmap_item *rmap_item;
240
241 rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL);
242 if (rmap_item)
243 ksm_rmap_items++;
244 return rmap_item;
245}
246
247static inline void free_rmap_item(struct rmap_item *rmap_item)
248{
249 ksm_rmap_items--;
250 rmap_item->mm = NULL; /* debug safety */
251 kmem_cache_free(rmap_item_cache, rmap_item);
252}
253
254static inline struct stable_node *alloc_stable_node(void)
255{
256 return kmem_cache_alloc(stable_node_cache, GFP_KERNEL);
257}
258
259static inline void free_stable_node(struct stable_node *stable_node)
260{
261 kmem_cache_free(stable_node_cache, stable_node);
262}
263
264static inline struct mm_slot *alloc_mm_slot(void)
265{
266 if (!mm_slot_cache) /* initialization failed */
267 return NULL;
268 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
269}
270
271static inline void free_mm_slot(struct mm_slot *mm_slot)
272{
273 kmem_cache_free(mm_slot_cache, mm_slot);
274}
275
276static struct mm_slot *get_mm_slot(struct mm_struct *mm)
277{
278 struct mm_slot *mm_slot;
279 struct hlist_head *bucket;
280 struct hlist_node *node;
281
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;
286 }
287 return NULL;
288}
289
290static void insert_to_mm_slots_hash(struct mm_struct *mm,
291 struct mm_slot *mm_slot)
292{
293 struct hlist_head *bucket;
294
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);
298}
299
300static inline int in_stable_tree(struct rmap_item *rmap_item)
301{
302 return rmap_item->address & STABLE_FLAG;
303}
304
305/*
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.
312 */
313static inline bool ksm_test_exit(struct mm_struct *mm)
314{
315 return atomic_read(&mm->mm_users) == 0;
316}
317
318/*
319 * We use break_ksm to break COW on a ksm page: it's a stripped down
320 *
321 * if (get_user_pages(current, mm, addr, 1, 1, 1, &page, NULL) == 1)
322 * put_page(page);
323 *
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.
328 */
329static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
330{
331 struct page *page;
332 int ret = 0;
333
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)));
346 /*
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.
349 *
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.
353 *
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.
357 *
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.
362 *
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.
367 *
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.
373 */
374 return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
375}
376
377static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm,
378 unsigned long addr)
379{
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;
389}
390
391static void break_cow(struct rmap_item *rmap_item)
392{
393 struct mm_struct *mm = rmap_item->mm;
394 unsigned long addr = rmap_item->address;
395 struct vm_area_struct *vma;
396
397 /*
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.
400 */
401 put_anon_vma(rmap_item->anon_vma);
402
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);
408}
409
410static struct page *page_trans_compound_anon(struct page *page)
411{
412 if (PageTransCompound(page)) {
413 struct page *head = compound_trans_head(page);
414 /*
415 * head may actually be splitted and freed from under
416 * us but it's ok here.
417 */
418 if (PageAnon(head))
419 return head;
420 }
421 return NULL;
422}
423
424static struct page *get_mergeable_page(struct rmap_item *rmap_item)
425{
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;
430
431 down_read(&mm->mmap_sem);
432 vma = find_mergeable_vma(mm, addr);
433 if (!vma)
434 goto out;
435
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);
444out: page = NULL;
445 }
446 up_read(&mm->mmap_sem);
447 return page;
448}
449
450static void remove_node_from_stable_tree(struct stable_node *stable_node)
451{
452 struct rmap_item *rmap_item;
453 struct hlist_node *hlist;
454
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();
463 }
464
465 rb_erase(&stable_node->node, &root_stable_tree);
466 free_stable_node(stable_node);
467}
468
469/*
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.
475 *
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.
485 *
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.
492 *
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.
497 */
498static struct page *get_ksm_page(struct stable_node *stable_node)
499{
500 struct page *page;
501 void *expected_mapping;
502
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;
514 }
515 rcu_read_unlock();
516 return page;
517stale:
518 rcu_read_unlock();
519 remove_node_from_stable_tree(stable_node);
520 return NULL;
521}
522
523/*
524 * Removing rmap_item from stable or unstable tree.
525 * This function will clean the information from the stable/unstable tree.
526 */
527static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
528{
529 if (rmap_item->address & STABLE_FLAG) {
530 struct stable_node *stable_node;
531 struct page *page;
532
533 stable_node = rmap_item->head;
534 page = get_ksm_page(stable_node);
535 if (!page)
536 goto out;
537
538 lock_page(page);
539 hlist_del(&rmap_item->hlist);
540 unlock_page(page);
541 put_page(page);
542
543 if (stable_node->hlist.first)
544 ksm_pages_sharing--;
545 else
546 ksm_pages_shared--;
547
548 put_anon_vma(rmap_item->anon_vma);
549 rmap_item->address &= PAGE_MASK;
550
551 } else if (rmap_item->address & UNSTABLE_FLAG) {
552 unsigned char age;
553 /*
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.
559 */
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);
564
565 ksm_pages_unshared--;
566 rmap_item->address &= PAGE_MASK;
567 }
568out:
569 cond_resched(); /* we're called from many long loops */
570}
571
572static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
573 struct rmap_item **rmap_list)
574{
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);
580 }
581}
582
583/*
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).
590 *
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.
595 */
596static int unmerge_ksm_pages(struct vm_area_struct *vma,
597 unsigned long start, unsigned long end)
598{
599 unsigned long addr;
600 int err = 0;
601
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);
609 }
610 return err;
611}
612
613#ifdef CONFIG_SYSFS
614/*
615 * Only called through the sysfs control interface:
616 */
617static int unmerge_and_remove_all_rmap_items(void)
618{
619 struct mm_slot *mm_slot;
620 struct mm_struct *mm;
621 struct vm_area_struct *vma;
622 int err = 0;
623
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);
628
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;
642 }
643
644 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
645
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);
653
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);
661 }
662 }
663
664 ksm_scan.seqnr = 0;
665 return 0;
666
667error:
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;
673}
674#endif /* CONFIG_SYSFS */
675
676static u32 calc_checksum(struct page *page)
677{
678 u32 checksum;
679 void *addr = kmap_atomic(page);
680 checksum = jhash2(addr, PAGE_SIZE / 4, 17);
681 kunmap_atomic(addr);
682 return checksum;
683}
684
685static int memcmp_pages(struct page *page1, struct page *page2)
686{
687 char *addr1, *addr2;
688 int ret;
689
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;
696}
697
698static inline int pages_identical(struct page *page1, struct page *page2)
699{
700 return !memcmp_pages(page1, page2);
701}
702
703static int write_protect_page(struct vm_area_struct *vma, struct page *page,
704 pte_t *orig_pte)
705{
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
713 addr = page_address_in_vma(page, vma);
714 if (addr == -EFAULT)
715 goto out;
716
717 BUG_ON(PageTransCompound(page));
718 ptep = page_check_address(page, mm, addr, &ptl, 0);
719 if (!ptep)
720 goto out;
721
722 if (pte_write(*ptep) || pte_dirty(*ptep)) {
723 pte_t entry;
724
725 swapped = PageSwapCache(page);
726 flush_cache_page(vma, addr, page_to_pfn(page));
727 /*
728 * Ok this is tricky, when get_user_pages_fast() run it doesn't
729 * take any lock, therefore the check that we are going to make
730 * with the pagecount against the mapcount is racey and
731 * O_DIRECT can happen right after the check.
732 * So we clear the pte and flush the tlb before the check
733 * this assure us that no O_DIRECT can happen after the check
734 * or in the middle of the check.
735 */
736 entry = ptep_clear_flush(vma, addr, ptep);
737 /*
738 * Check that no O_DIRECT or similar I/O is in progress on the
739 * page
740 */
741 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
742 set_pte_at(mm, addr, ptep, entry);
743 goto out_unlock;
744 }
745 if (pte_dirty(entry))
746 set_page_dirty(page);
747 entry = pte_mkclean(pte_wrprotect(entry));
748 set_pte_at_notify(mm, addr, ptep, entry);
749 }
750 *orig_pte = *ptep;
751 err = 0;
752
753out_unlock:
754 pte_unmap_unlock(ptep, ptl);
755out:
756 return err;
757}
758
759/**
760 * replace_page - replace page in vma by new ksm page
761 * @vma: vma that holds the pte pointing to page
762 * @page: the page we are replacing by kpage
763 * @kpage: the ksm page we replace page by
764 * @orig_pte: the original value of the pte
765 *
766 * Returns 0 on success, -EFAULT on failure.
767 */
768static int replace_page(struct vm_area_struct *vma, struct page *page,
769 struct page *kpage, pte_t orig_pte)
770{
771 struct mm_struct *mm = vma->vm_mm;
772 pgd_t *pgd;
773 pud_t *pud;
774 pmd_t *pmd;
775 pte_t *ptep;
776 spinlock_t *ptl;
777 unsigned long addr;
778 int err = -EFAULT;
779
780 addr = page_address_in_vma(page, vma);
781 if (addr == -EFAULT)
782 goto out;
783
784 pgd = pgd_offset(mm, addr);
785 if (!pgd_present(*pgd))
786 goto out;
787
788 pud = pud_offset(pgd, addr);
789 if (!pud_present(*pud))
790 goto out;
791
792 pmd = pmd_offset(pud, addr);
793 BUG_ON(pmd_trans_huge(*pmd));
794 if (!pmd_present(*pmd))
795 goto out;
796
797 ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
798 if (!pte_same(*ptep, orig_pte)) {
799 pte_unmap_unlock(ptep, ptl);
800 goto out;
801 }
802
803 get_page(kpage);
804 page_add_anon_rmap(kpage, vma, addr);
805
806 flush_cache_page(vma, addr, pte_pfn(*ptep));
807 ptep_clear_flush(vma, addr, ptep);
808 set_pte_at_notify(mm, addr, ptep, mk_pte(kpage, vma->vm_page_prot));
809
810 page_remove_rmap(page);
811 if (!page_mapped(page))
812 try_to_free_swap(page);
813 put_page(page);
814
815 pte_unmap_unlock(ptep, ptl);
816 err = 0;
817out:
818 return err;
819}
820
821static int page_trans_compound_anon_split(struct page *page)
822{
823 int ret = 0;
824 struct page *transhuge_head = page_trans_compound_anon(page);
825 if (transhuge_head) {
826 /* Get the reference on the head to split it. */
827 if (get_page_unless_zero(transhuge_head)) {
828 /*
829 * Recheck we got the reference while the head
830 * was still anonymous.
831 */
832 if (PageAnon(transhuge_head))
833 ret = split_huge_page(transhuge_head);
834 else
835 /*
836 * Retry later if split_huge_page run
837 * from under us.
838 */
839 ret = 1;
840 put_page(transhuge_head);
841 } else
842 /* Retry later if split_huge_page run from under us. */
843 ret = 1;
844 }
845 return ret;
846}
847
848/*
849 * try_to_merge_one_page - take two pages and merge them into one
850 * @vma: the vma that holds the pte pointing to page
851 * @page: the PageAnon page that we want to replace with kpage
852 * @kpage: the PageKsm page that we want to map instead of page,
853 * or NULL the first time when we want to use page as kpage.
854 *
855 * This function returns 0 if the pages were merged, -EFAULT otherwise.
856 */
857static int try_to_merge_one_page(struct vm_area_struct *vma,
858 struct page *page, struct page *kpage)
859{
860 pte_t orig_pte = __pte(0);
861 int err = -EFAULT;
862
863 if (page == kpage) /* ksm page forked */
864 return 0;
865
866 if (!(vma->vm_flags & VM_MERGEABLE))
867 goto out;
868 if (PageTransCompound(page) && page_trans_compound_anon_split(page))
869 goto out;
870 BUG_ON(PageTransCompound(page));
871 if (!PageAnon(page))
872 goto out;
873
874 /*
875 * We need the page lock to read a stable PageSwapCache in
876 * write_protect_page(). We use trylock_page() instead of
877 * lock_page() because we don't want to wait here - we
878 * prefer to continue scanning and merging different pages,
879 * then come back to this page when it is unlocked.
880 */
881 if (!trylock_page(page))
882 goto out;
883 /*
884 * If this anonymous page is mapped only here, its pte may need
885 * to be write-protected. If it's mapped elsewhere, all of its
886 * ptes are necessarily already write-protected. But in either
887 * case, we need to lock and check page_count is not raised.
888 */
889 if (write_protect_page(vma, page, &orig_pte) == 0) {
890 if (!kpage) {
891 /*
892 * While we hold page lock, upgrade page from
893 * PageAnon+anon_vma to PageKsm+NULL stable_node:
894 * stable_tree_insert() will update stable_node.
895 */
896 set_page_stable_node(page, NULL);
897 mark_page_accessed(page);
898 err = 0;
899 } else if (pages_identical(page, kpage))
900 err = replace_page(vma, page, kpage, orig_pte);
901 }
902
903 if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
904 munlock_vma_page(page);
905 if (!PageMlocked(kpage)) {
906 unlock_page(page);
907 lock_page(kpage);
908 mlock_vma_page(kpage);
909 page = kpage; /* for final unlock */
910 }
911 }
912
913 unlock_page(page);
914out:
915 return err;
916}
917
918/*
919 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
920 * but no new kernel page is allocated: kpage must already be a ksm page.
921 *
922 * This function returns 0 if the pages were merged, -EFAULT otherwise.
923 */
924static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
925 struct page *page, struct page *kpage)
926{
927 struct mm_struct *mm = rmap_item->mm;
928 struct vm_area_struct *vma;
929 int err = -EFAULT;
930
931 down_read(&mm->mmap_sem);
932 if (ksm_test_exit(mm))
933 goto out;
934 vma = find_vma(mm, rmap_item->address);
935 if (!vma || vma->vm_start > rmap_item->address)
936 goto out;
937
938 err = try_to_merge_one_page(vma, page, kpage);
939 if (err)
940 goto out;
941
942 /* Must get reference to anon_vma while still holding mmap_sem */
943 rmap_item->anon_vma = vma->anon_vma;
944 get_anon_vma(vma->anon_vma);
945out:
946 up_read(&mm->mmap_sem);
947 return err;
948}
949
950/*
951 * try_to_merge_two_pages - take two identical pages and prepare them
952 * to be merged into one page.
953 *
954 * This function returns the kpage if we successfully merged two identical
955 * pages into one ksm page, NULL otherwise.
956 *
957 * Note that this function upgrades page to ksm page: if one of the pages
958 * is already a ksm page, try_to_merge_with_ksm_page should be used.
959 */
960static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
961 struct page *page,
962 struct rmap_item *tree_rmap_item,
963 struct page *tree_page)
964{
965 int err;
966
967 err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
968 if (!err) {
969 err = try_to_merge_with_ksm_page(tree_rmap_item,
970 tree_page, page);
971 /*
972 * If that fails, we have a ksm page with only one pte
973 * pointing to it: so break it.
974 */
975 if (err)
976 break_cow(rmap_item);
977 }
978 return err ? NULL : page;
979}
980
981/*
982 * stable_tree_search - search for page inside the stable tree
983 *
984 * This function checks if there is a page inside the stable tree
985 * with identical content to the page that we are scanning right now.
986 *
987 * This function returns the stable tree node of identical content if found,
988 * NULL otherwise.
989 */
990static struct page *stable_tree_search(struct page *page)
991{
992 struct rb_node *node = root_stable_tree.rb_node;
993 struct stable_node *stable_node;
994
995 stable_node = page_stable_node(page);
996 if (stable_node) { /* ksm page forked */
997 get_page(page);
998 return page;
999 }
1000
1001 while (node) {
1002 struct page *tree_page;
1003 int ret;
1004
1005 cond_resched();
1006 stable_node = rb_entry(node, struct stable_node, node);
1007 tree_page = get_ksm_page(stable_node);
1008 if (!tree_page)
1009 return NULL;
1010
1011 ret = memcmp_pages(page, tree_page);
1012
1013 if (ret < 0) {
1014 put_page(tree_page);
1015 node = node->rb_left;
1016 } else if (ret > 0) {
1017 put_page(tree_page);
1018 node = node->rb_right;
1019 } else
1020 return tree_page;
1021 }
1022
1023 return NULL;
1024}
1025
1026/*
1027 * stable_tree_insert - insert rmap_item pointing to new ksm page
1028 * into the stable tree.
1029 *
1030 * This function returns the stable tree node just allocated on success,
1031 * NULL otherwise.
1032 */
1033static struct stable_node *stable_tree_insert(struct page *kpage)
1034{
1035 struct rb_node **new = &root_stable_tree.rb_node;
1036 struct rb_node *parent = NULL;
1037 struct stable_node *stable_node;
1038
1039 while (*new) {
1040 struct page *tree_page;
1041 int ret;
1042
1043 cond_resched();
1044 stable_node = rb_entry(*new, struct stable_node, node);
1045 tree_page = get_ksm_page(stable_node);
1046 if (!tree_page)
1047 return NULL;
1048
1049 ret = memcmp_pages(kpage, tree_page);
1050 put_page(tree_page);
1051
1052 parent = *new;
1053 if (ret < 0)
1054 new = &parent->rb_left;
1055 else if (ret > 0)
1056 new = &parent->rb_right;
1057 else {
1058 /*
1059 * It is not a bug that stable_tree_search() didn't
1060 * find this node: because at that time our page was
1061 * not yet write-protected, so may have changed since.
1062 */
1063 return NULL;
1064 }
1065 }
1066
1067 stable_node = alloc_stable_node();
1068 if (!stable_node)
1069 return NULL;
1070
1071 rb_link_node(&stable_node->node, parent, new);
1072 rb_insert_color(&stable_node->node, &root_stable_tree);
1073
1074 INIT_HLIST_HEAD(&stable_node->hlist);
1075
1076 stable_node->kpfn = page_to_pfn(kpage);
1077 set_page_stable_node(kpage, stable_node);
1078
1079 return stable_node;
1080}
1081
1082/*
1083 * unstable_tree_search_insert - search for identical page,
1084 * else insert rmap_item into the unstable tree.
1085 *
1086 * This function searches for a page in the unstable tree identical to the
1087 * page currently being scanned; and if no identical page is found in the
1088 * tree, we insert rmap_item as a new object into the unstable tree.
1089 *
1090 * This function returns pointer to rmap_item found to be identical
1091 * to the currently scanned page, NULL otherwise.
1092 *
1093 * This function does both searching and inserting, because they share
1094 * the same walking algorithm in an rbtree.
1095 */
1096static
1097struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1098 struct page *page,
1099 struct page **tree_pagep)
1100
1101{
1102 struct rb_node **new = &root_unstable_tree.rb_node;
1103 struct rb_node *parent = NULL;
1104
1105 while (*new) {
1106 struct rmap_item *tree_rmap_item;
1107 struct page *tree_page;
1108 int ret;
1109
1110 cond_resched();
1111 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
1112 tree_page = get_mergeable_page(tree_rmap_item);
1113 if (IS_ERR_OR_NULL(tree_page))
1114 return NULL;
1115
1116 /*
1117 * Don't substitute a ksm page for a forked page.
1118 */
1119 if (page == tree_page) {
1120 put_page(tree_page);
1121 return NULL;
1122 }
1123
1124 ret = memcmp_pages(page, tree_page);
1125
1126 parent = *new;
1127 if (ret < 0) {
1128 put_page(tree_page);
1129 new = &parent->rb_left;
1130 } else if (ret > 0) {
1131 put_page(tree_page);
1132 new = &parent->rb_right;
1133 } else {
1134 *tree_pagep = tree_page;
1135 return tree_rmap_item;
1136 }
1137 }
1138
1139 rmap_item->address |= UNSTABLE_FLAG;
1140 rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1141 rb_link_node(&rmap_item->node, parent, new);
1142 rb_insert_color(&rmap_item->node, &root_unstable_tree);
1143
1144 ksm_pages_unshared++;
1145 return NULL;
1146}
1147
1148/*
1149 * stable_tree_append - add another rmap_item to the linked list of
1150 * rmap_items hanging off a given node of the stable tree, all sharing
1151 * the same ksm page.
1152 */
1153static void stable_tree_append(struct rmap_item *rmap_item,
1154 struct stable_node *stable_node)
1155{
1156 rmap_item->head = stable_node;
1157 rmap_item->address |= STABLE_FLAG;
1158 hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
1159
1160 if (rmap_item->hlist.next)
1161 ksm_pages_sharing++;
1162 else
1163 ksm_pages_shared++;
1164}
1165
1166/*
1167 * cmp_and_merge_page - first see if page can be merged into the stable tree;
1168 * if not, compare checksum to previous and if it's the same, see if page can
1169 * be inserted into the unstable tree, or merged with a page already there and
1170 * both transferred to the stable tree.
1171 *
1172 * @page: the page that we are searching identical page to.
1173 * @rmap_item: the reverse mapping into the virtual address of this page
1174 */
1175static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1176{
1177 struct rmap_item *tree_rmap_item;
1178 struct page *tree_page = NULL;
1179 struct stable_node *stable_node;
1180 struct page *kpage;
1181 unsigned int checksum;
1182 int err;
1183
1184 remove_rmap_item_from_tree(rmap_item);
1185
1186 /* We first start with searching the page inside the stable tree */
1187 kpage = stable_tree_search(page);
1188 if (kpage) {
1189 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
1190 if (!err) {
1191 /*
1192 * The page was successfully merged:
1193 * add its rmap_item to the stable tree.
1194 */
1195 lock_page(kpage);
1196 stable_tree_append(rmap_item, page_stable_node(kpage));
1197 unlock_page(kpage);
1198 }
1199 put_page(kpage);
1200 return;
1201 }
1202
1203 /*
1204 * If the hash value of the page has changed from the last time
1205 * we calculated it, this page is changing frequently: therefore we
1206 * don't want to insert it in the unstable tree, and we don't want
1207 * to waste our time searching for something identical to it there.
1208 */
1209 checksum = calc_checksum(page);
1210 if (rmap_item->oldchecksum != checksum) {
1211 rmap_item->oldchecksum = checksum;
1212 return;
1213 }
1214
1215 tree_rmap_item =
1216 unstable_tree_search_insert(rmap_item, page, &tree_page);
1217 if (tree_rmap_item) {
1218 kpage = try_to_merge_two_pages(rmap_item, page,
1219 tree_rmap_item, tree_page);
1220 put_page(tree_page);
1221 /*
1222 * As soon as we merge this page, we want to remove the
1223 * rmap_item of the page we have merged with from the unstable
1224 * tree, and insert it instead as new node in the stable tree.
1225 */
1226 if (kpage) {
1227 remove_rmap_item_from_tree(tree_rmap_item);
1228
1229 lock_page(kpage);
1230 stable_node = stable_tree_insert(kpage);
1231 if (stable_node) {
1232 stable_tree_append(tree_rmap_item, stable_node);
1233 stable_tree_append(rmap_item, stable_node);
1234 }
1235 unlock_page(kpage);
1236
1237 /*
1238 * If we fail to insert the page into the stable tree,
1239 * we will have 2 virtual addresses that are pointing
1240 * to a ksm page left outside the stable tree,
1241 * in which case we need to break_cow on both.
1242 */
1243 if (!stable_node) {
1244 break_cow(tree_rmap_item);
1245 break_cow(rmap_item);
1246 }
1247 }
1248 }
1249}
1250
1251static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
1252 struct rmap_item **rmap_list,
1253 unsigned long addr)
1254{
1255 struct rmap_item *rmap_item;
1256
1257 while (*rmap_list) {
1258 rmap_item = *rmap_list;
1259 if ((rmap_item->address & PAGE_MASK) == addr)
1260 return rmap_item;
1261 if (rmap_item->address > addr)
1262 break;
1263 *rmap_list = rmap_item->rmap_list;
1264 remove_rmap_item_from_tree(rmap_item);
1265 free_rmap_item(rmap_item);
1266 }
1267
1268 rmap_item = alloc_rmap_item();
1269 if (rmap_item) {
1270 /* It has already been zeroed */
1271 rmap_item->mm = mm_slot->mm;
1272 rmap_item->address = addr;
1273 rmap_item->rmap_list = *rmap_list;
1274 *rmap_list = rmap_item;
1275 }
1276 return rmap_item;
1277}
1278
1279static struct rmap_item *scan_get_next_rmap_item(struct page **page)
1280{
1281 struct mm_struct *mm;
1282 struct mm_slot *slot;
1283 struct vm_area_struct *vma;
1284 struct rmap_item *rmap_item;
1285
1286 if (list_empty(&ksm_mm_head.mm_list))
1287 return NULL;
1288
1289 slot = ksm_scan.mm_slot;
1290 if (slot == &ksm_mm_head) {
1291 /*
1292 * A number of pages can hang around indefinitely on per-cpu
1293 * pagevecs, raised page count preventing write_protect_page
1294 * from merging them. Though it doesn't really matter much,
1295 * it is puzzling to see some stuck in pages_volatile until
1296 * other activity jostles them out, and they also prevented
1297 * LTP's KSM test from succeeding deterministically; so drain
1298 * them here (here rather than on entry to ksm_do_scan(),
1299 * so we don't IPI too often when pages_to_scan is set low).
1300 */
1301 lru_add_drain_all();
1302
1303 root_unstable_tree = RB_ROOT;
1304
1305 spin_lock(&ksm_mmlist_lock);
1306 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1307 ksm_scan.mm_slot = slot;
1308 spin_unlock(&ksm_mmlist_lock);
1309 /*
1310 * Although we tested list_empty() above, a racing __ksm_exit
1311 * of the last mm on the list may have removed it since then.
1312 */
1313 if (slot == &ksm_mm_head)
1314 return NULL;
1315next_mm:
1316 ksm_scan.address = 0;
1317 ksm_scan.rmap_list = &slot->rmap_list;
1318 }
1319
1320 mm = slot->mm;
1321 down_read(&mm->mmap_sem);
1322 if (ksm_test_exit(mm))
1323 vma = NULL;
1324 else
1325 vma = find_vma(mm, ksm_scan.address);
1326
1327 for (; vma; vma = vma->vm_next) {
1328 if (!(vma->vm_flags & VM_MERGEABLE))
1329 continue;
1330 if (ksm_scan.address < vma->vm_start)
1331 ksm_scan.address = vma->vm_start;
1332 if (!vma->anon_vma)
1333 ksm_scan.address = vma->vm_end;
1334
1335 while (ksm_scan.address < vma->vm_end) {
1336 if (ksm_test_exit(mm))
1337 break;
1338 *page = follow_page(vma, ksm_scan.address, FOLL_GET);
1339 if (IS_ERR_OR_NULL(*page)) {
1340 ksm_scan.address += PAGE_SIZE;
1341 cond_resched();
1342 continue;
1343 }
1344 if (PageAnon(*page) ||
1345 page_trans_compound_anon(*page)) {
1346 flush_anon_page(vma, *page, ksm_scan.address);
1347 flush_dcache_page(*page);
1348 rmap_item = get_next_rmap_item(slot,
1349 ksm_scan.rmap_list, ksm_scan.address);
1350 if (rmap_item) {
1351 ksm_scan.rmap_list =
1352 &rmap_item->rmap_list;
1353 ksm_scan.address += PAGE_SIZE;
1354 } else
1355 put_page(*page);
1356 up_read(&mm->mmap_sem);
1357 return rmap_item;
1358 }
1359 put_page(*page);
1360 ksm_scan.address += PAGE_SIZE;
1361 cond_resched();
1362 }
1363 }
1364
1365 if (ksm_test_exit(mm)) {
1366 ksm_scan.address = 0;
1367 ksm_scan.rmap_list = &slot->rmap_list;
1368 }
1369 /*
1370 * Nuke all the rmap_items that are above this current rmap:
1371 * because there were no VM_MERGEABLE vmas with such addresses.
1372 */
1373 remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
1374
1375 spin_lock(&ksm_mmlist_lock);
1376 ksm_scan.mm_slot = list_entry(slot->mm_list.next,
1377 struct mm_slot, mm_list);
1378 if (ksm_scan.address == 0) {
1379 /*
1380 * We've completed a full scan of all vmas, holding mmap_sem
1381 * throughout, and found no VM_MERGEABLE: so do the same as
1382 * __ksm_exit does to remove this mm from all our lists now.
1383 * This applies either when cleaning up after __ksm_exit
1384 * (but beware: we can reach here even before __ksm_exit),
1385 * or when all VM_MERGEABLE areas have been unmapped (and
1386 * mmap_sem then protects against race with MADV_MERGEABLE).
1387 */
1388 hlist_del(&slot->link);
1389 list_del(&slot->mm_list);
1390 spin_unlock(&ksm_mmlist_lock);
1391
1392 free_mm_slot(slot);
1393 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1394 up_read(&mm->mmap_sem);
1395 mmdrop(mm);
1396 } else {
1397 spin_unlock(&ksm_mmlist_lock);
1398 up_read(&mm->mmap_sem);
1399 }
1400
1401 /* Repeat until we've completed scanning the whole list */
1402 slot = ksm_scan.mm_slot;
1403 if (slot != &ksm_mm_head)
1404 goto next_mm;
1405
1406 ksm_scan.seqnr++;
1407 return NULL;
1408}
1409
1410/**
1411 * ksm_do_scan - the ksm scanner main worker function.
1412 * @scan_npages - number of pages we want to scan before we return.
1413 */
1414static void ksm_do_scan(unsigned int scan_npages)
1415{
1416 struct rmap_item *rmap_item;
1417 struct page *uninitialized_var(page);
1418
1419 while (scan_npages-- && likely(!freezing(current))) {
1420 cond_resched();
1421 rmap_item = scan_get_next_rmap_item(&page);
1422 if (!rmap_item)
1423 return;
1424 if (!PageKsm(page) || !in_stable_tree(rmap_item))
1425 cmp_and_merge_page(page, rmap_item);
1426 put_page(page);
1427 }
1428}
1429
1430static int ksmd_should_run(void)
1431{
1432 return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
1433}
1434
1435static int ksm_scan_thread(void *nothing)
1436{
1437 set_freezable();
1438 set_user_nice(current, 5);
1439
1440 while (!kthread_should_stop()) {
1441 mutex_lock(&ksm_thread_mutex);
1442 if (ksmd_should_run())
1443 ksm_do_scan(ksm_thread_pages_to_scan);
1444 mutex_unlock(&ksm_thread_mutex);
1445
1446 try_to_freeze();
1447
1448 if (ksmd_should_run()) {
1449 schedule_timeout_interruptible(
1450 msecs_to_jiffies(ksm_thread_sleep_millisecs));
1451 } else {
1452 wait_event_freezable(ksm_thread_wait,
1453 ksmd_should_run() || kthread_should_stop());
1454 }
1455 }
1456 return 0;
1457}
1458
1459int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1460 unsigned long end, int advice, unsigned long *vm_flags)
1461{
1462 struct mm_struct *mm = vma->vm_mm;
1463 int err;
1464
1465 switch (advice) {
1466 case MADV_MERGEABLE:
1467 /*
1468 * Be somewhat over-protective for now!
1469 */
1470 if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE |
1471 VM_PFNMAP | VM_IO | VM_DONTEXPAND |
1472 VM_RESERVED | VM_HUGETLB | VM_INSERTPAGE |
1473 VM_NONLINEAR | VM_MIXEDMAP | VM_SAO))
1474 return 0; /* just ignore the advice */
1475
1476 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
1477 err = __ksm_enter(mm);
1478 if (err)
1479 return err;
1480 }
1481
1482 *vm_flags |= VM_MERGEABLE;
1483 break;
1484
1485 case MADV_UNMERGEABLE:
1486 if (!(*vm_flags & VM_MERGEABLE))
1487 return 0; /* just ignore the advice */
1488
1489 if (vma->anon_vma) {
1490 err = unmerge_ksm_pages(vma, start, end);
1491 if (err)
1492 return err;
1493 }
1494
1495 *vm_flags &= ~VM_MERGEABLE;
1496 break;
1497 }
1498
1499 return 0;
1500}
1501
1502int __ksm_enter(struct mm_struct *mm)
1503{
1504 struct mm_slot *mm_slot;
1505 int needs_wakeup;
1506
1507 mm_slot = alloc_mm_slot();
1508 if (!mm_slot)
1509 return -ENOMEM;
1510
1511 /* Check ksm_run too? Would need tighter locking */
1512 needs_wakeup = list_empty(&ksm_mm_head.mm_list);
1513
1514 spin_lock(&ksm_mmlist_lock);
1515 insert_to_mm_slots_hash(mm, mm_slot);
1516 /*
1517 * Insert just behind the scanning cursor, to let the area settle
1518 * down a little; when fork is followed by immediate exec, we don't
1519 * want ksmd to waste time setting up and tearing down an rmap_list.
1520 */
1521 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
1522 spin_unlock(&ksm_mmlist_lock);
1523
1524 set_bit(MMF_VM_MERGEABLE, &mm->flags);
1525 atomic_inc(&mm->mm_count);
1526
1527 if (needs_wakeup)
1528 wake_up_interruptible(&ksm_thread_wait);
1529
1530 return 0;
1531}
1532
1533void __ksm_exit(struct mm_struct *mm)
1534{
1535 struct mm_slot *mm_slot;
1536 int easy_to_free = 0;
1537
1538 /*
1539 * This process is exiting: if it's straightforward (as is the
1540 * case when ksmd was never running), free mm_slot immediately.
1541 * But if it's at the cursor or has rmap_items linked to it, use
1542 * mmap_sem to synchronize with any break_cows before pagetables
1543 * are freed, and leave the mm_slot on the list for ksmd to free.
1544 * Beware: ksm may already have noticed it exiting and freed the slot.
1545 */
1546
1547 spin_lock(&ksm_mmlist_lock);
1548 mm_slot = get_mm_slot(mm);
1549 if (mm_slot && ksm_scan.mm_slot != mm_slot) {
1550 if (!mm_slot->rmap_list) {
1551 hlist_del(&mm_slot->link);
1552 list_del(&mm_slot->mm_list);
1553 easy_to_free = 1;
1554 } else {
1555 list_move(&mm_slot->mm_list,
1556 &ksm_scan.mm_slot->mm_list);
1557 }
1558 }
1559 spin_unlock(&ksm_mmlist_lock);
1560
1561 if (easy_to_free) {
1562 free_mm_slot(mm_slot);
1563 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1564 mmdrop(mm);
1565 } else if (mm_slot) {
1566 down_write(&mm->mmap_sem);
1567 up_write(&mm->mmap_sem);
1568 }
1569}
1570
1571struct page *ksm_does_need_to_copy(struct page *page,
1572 struct vm_area_struct *vma, unsigned long address)
1573{
1574 struct page *new_page;
1575
1576 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1577 if (new_page) {
1578 copy_user_highpage(new_page, page, address, vma);
1579
1580 SetPageDirty(new_page);
1581 __SetPageUptodate(new_page);
1582 SetPageSwapBacked(new_page);
1583 __set_page_locked(new_page);
1584
1585 if (page_evictable(new_page, vma))
1586 lru_cache_add_lru(new_page, LRU_ACTIVE_ANON);
1587 else
1588 add_page_to_unevictable_list(new_page);
1589 }
1590
1591 return new_page;
1592}
1593
1594int page_referenced_ksm(struct page *page, struct mem_cgroup *memcg,
1595 unsigned long *vm_flags)
1596{
1597 struct stable_node *stable_node;
1598 struct rmap_item *rmap_item;
1599 struct hlist_node *hlist;
1600 unsigned int mapcount = page_mapcount(page);
1601 int referenced = 0;
1602 int search_new_forks = 0;
1603
1604 VM_BUG_ON(!PageKsm(page));
1605 VM_BUG_ON(!PageLocked(page));
1606
1607 stable_node = page_stable_node(page);
1608 if (!stable_node)
1609 return 0;
1610again:
1611 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1612 struct anon_vma *anon_vma = rmap_item->anon_vma;
1613 struct anon_vma_chain *vmac;
1614 struct vm_area_struct *vma;
1615
1616 anon_vma_lock(anon_vma);
1617 list_for_each_entry(vmac, &anon_vma->head, same_anon_vma) {
1618 vma = vmac->vma;
1619 if (rmap_item->address < vma->vm_start ||
1620 rmap_item->address >= vma->vm_end)
1621 continue;
1622 /*
1623 * Initially we examine only the vma which covers this
1624 * rmap_item; but later, if there is still work to do,
1625 * we examine covering vmas in other mms: in case they
1626 * were forked from the original since ksmd passed.
1627 */
1628 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1629 continue;
1630
1631 if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
1632 continue;
1633
1634 referenced += page_referenced_one(page, vma,
1635 rmap_item->address, &mapcount, vm_flags);
1636 if (!search_new_forks || !mapcount)
1637 break;
1638 }
1639 anon_vma_unlock(anon_vma);
1640 if (!mapcount)
1641 goto out;
1642 }
1643 if (!search_new_forks++)
1644 goto again;
1645out:
1646 return referenced;
1647}
1648
1649int try_to_unmap_ksm(struct page *page, enum ttu_flags flags)
1650{
1651 struct stable_node *stable_node;
1652 struct hlist_node *hlist;
1653 struct rmap_item *rmap_item;
1654 int ret = SWAP_AGAIN;
1655 int search_new_forks = 0;
1656
1657 VM_BUG_ON(!PageKsm(page));
1658 VM_BUG_ON(!PageLocked(page));
1659
1660 stable_node = page_stable_node(page);
1661 if (!stable_node)
1662 return SWAP_FAIL;
1663again:
1664 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1665 struct anon_vma *anon_vma = rmap_item->anon_vma;
1666 struct anon_vma_chain *vmac;
1667 struct vm_area_struct *vma;
1668
1669 anon_vma_lock(anon_vma);
1670 list_for_each_entry(vmac, &anon_vma->head, same_anon_vma) {
1671 vma = vmac->vma;
1672 if (rmap_item->address < vma->vm_start ||
1673 rmap_item->address >= vma->vm_end)
1674 continue;
1675 /*
1676 * Initially we examine only the vma which covers this
1677 * rmap_item; but later, if there is still work to do,
1678 * we examine covering vmas in other mms: in case they
1679 * were forked from the original since ksmd passed.
1680 */
1681 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1682 continue;
1683
1684 ret = try_to_unmap_one(page, vma,
1685 rmap_item->address, flags);
1686 if (ret != SWAP_AGAIN || !page_mapped(page)) {
1687 anon_vma_unlock(anon_vma);
1688 goto out;
1689 }
1690 }
1691 anon_vma_unlock(anon_vma);
1692 }
1693 if (!search_new_forks++)
1694 goto again;
1695out:
1696 return ret;
1697}
1698
1699#ifdef CONFIG_MIGRATION
1700int rmap_walk_ksm(struct page *page, int (*rmap_one)(struct page *,
1701 struct vm_area_struct *, unsigned long, void *), void *arg)
1702{
1703 struct stable_node *stable_node;
1704 struct hlist_node *hlist;
1705 struct rmap_item *rmap_item;
1706 int ret = SWAP_AGAIN;
1707 int search_new_forks = 0;
1708
1709 VM_BUG_ON(!PageKsm(page));
1710 VM_BUG_ON(!PageLocked(page));
1711
1712 stable_node = page_stable_node(page);
1713 if (!stable_node)
1714 return ret;
1715again:
1716 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1717 struct anon_vma *anon_vma = rmap_item->anon_vma;
1718 struct anon_vma_chain *vmac;
1719 struct vm_area_struct *vma;
1720
1721 anon_vma_lock(anon_vma);
1722 list_for_each_entry(vmac, &anon_vma->head, same_anon_vma) {
1723 vma = vmac->vma;
1724 if (rmap_item->address < vma->vm_start ||
1725 rmap_item->address >= vma->vm_end)
1726 continue;
1727 /*
1728 * Initially we examine only the vma which covers this
1729 * rmap_item; but later, if there is still work to do,
1730 * we examine covering vmas in other mms: in case they
1731 * were forked from the original since ksmd passed.
1732 */
1733 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1734 continue;
1735
1736 ret = rmap_one(page, vma, rmap_item->address, arg);
1737 if (ret != SWAP_AGAIN) {
1738 anon_vma_unlock(anon_vma);
1739 goto out;
1740 }
1741 }
1742 anon_vma_unlock(anon_vma);
1743 }
1744 if (!search_new_forks++)
1745 goto again;
1746out:
1747 return ret;
1748}
1749
1750void ksm_migrate_page(struct page *newpage, struct page *oldpage)
1751{
1752 struct stable_node *stable_node;
1753
1754 VM_BUG_ON(!PageLocked(oldpage));
1755 VM_BUG_ON(!PageLocked(newpage));
1756 VM_BUG_ON(newpage->mapping != oldpage->mapping);
1757
1758 stable_node = page_stable_node(newpage);
1759 if (stable_node) {
1760 VM_BUG_ON(stable_node->kpfn != page_to_pfn(oldpage));
1761 stable_node->kpfn = page_to_pfn(newpage);
1762 }
1763}
1764#endif /* CONFIG_MIGRATION */
1765
1766#ifdef CONFIG_MEMORY_HOTREMOVE
1767static struct stable_node *ksm_check_stable_tree(unsigned long start_pfn,
1768 unsigned long end_pfn)
1769{
1770 struct rb_node *node;
1771
1772 for (node = rb_first(&root_stable_tree); node; node = rb_next(node)) {
1773 struct stable_node *stable_node;
1774
1775 stable_node = rb_entry(node, struct stable_node, node);
1776 if (stable_node->kpfn >= start_pfn &&
1777 stable_node->kpfn < end_pfn)
1778 return stable_node;
1779 }
1780 return NULL;
1781}
1782
1783static int ksm_memory_callback(struct notifier_block *self,
1784 unsigned long action, void *arg)
1785{
1786 struct memory_notify *mn = arg;
1787 struct stable_node *stable_node;
1788
1789 switch (action) {
1790 case MEM_GOING_OFFLINE:
1791 /*
1792 * Keep it very simple for now: just lock out ksmd and
1793 * MADV_UNMERGEABLE while any memory is going offline.
1794 * mutex_lock_nested() is necessary because lockdep was alarmed
1795 * that here we take ksm_thread_mutex inside notifier chain
1796 * mutex, and later take notifier chain mutex inside
1797 * ksm_thread_mutex to unlock it. But that's safe because both
1798 * are inside mem_hotplug_mutex.
1799 */
1800 mutex_lock_nested(&ksm_thread_mutex, SINGLE_DEPTH_NESTING);
1801 break;
1802
1803 case MEM_OFFLINE:
1804 /*
1805 * Most of the work is done by page migration; but there might
1806 * be a few stable_nodes left over, still pointing to struct
1807 * pages which have been offlined: prune those from the tree.
1808 */
1809 while ((stable_node = ksm_check_stable_tree(mn->start_pfn,
1810 mn->start_pfn + mn->nr_pages)) != NULL)
1811 remove_node_from_stable_tree(stable_node);
1812 /* fallthrough */
1813
1814 case MEM_CANCEL_OFFLINE:
1815 mutex_unlock(&ksm_thread_mutex);
1816 break;
1817 }
1818 return NOTIFY_OK;
1819}
1820#endif /* CONFIG_MEMORY_HOTREMOVE */
1821
1822#ifdef CONFIG_SYSFS
1823/*
1824 * This all compiles without CONFIG_SYSFS, but is a waste of space.
1825 */
1826
1827#define KSM_ATTR_RO(_name) \
1828 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1829#define KSM_ATTR(_name) \
1830 static struct kobj_attribute _name##_attr = \
1831 __ATTR(_name, 0644, _name##_show, _name##_store)
1832
1833static ssize_t sleep_millisecs_show(struct kobject *kobj,
1834 struct kobj_attribute *attr, char *buf)
1835{
1836 return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
1837}
1838
1839static ssize_t sleep_millisecs_store(struct kobject *kobj,
1840 struct kobj_attribute *attr,
1841 const char *buf, size_t count)
1842{
1843 unsigned long msecs;
1844 int err;
1845
1846 err = strict_strtoul(buf, 10, &msecs);
1847 if (err || msecs > UINT_MAX)
1848 return -EINVAL;
1849
1850 ksm_thread_sleep_millisecs = msecs;
1851
1852 return count;
1853}
1854KSM_ATTR(sleep_millisecs);
1855
1856static ssize_t pages_to_scan_show(struct kobject *kobj,
1857 struct kobj_attribute *attr, char *buf)
1858{
1859 return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
1860}
1861
1862static ssize_t pages_to_scan_store(struct kobject *kobj,
1863 struct kobj_attribute *attr,
1864 const char *buf, size_t count)
1865{
1866 int err;
1867 unsigned long nr_pages;
1868
1869 err = strict_strtoul(buf, 10, &nr_pages);
1870 if (err || nr_pages > UINT_MAX)
1871 return -EINVAL;
1872
1873 ksm_thread_pages_to_scan = nr_pages;
1874
1875 return count;
1876}
1877KSM_ATTR(pages_to_scan);
1878
1879static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
1880 char *buf)
1881{
1882 return sprintf(buf, "%u\n", ksm_run);
1883}
1884
1885static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
1886 const char *buf, size_t count)
1887{
1888 int err;
1889 unsigned long flags;
1890
1891 err = strict_strtoul(buf, 10, &flags);
1892 if (err || flags > UINT_MAX)
1893 return -EINVAL;
1894 if (flags > KSM_RUN_UNMERGE)
1895 return -EINVAL;
1896
1897 /*
1898 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
1899 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
1900 * breaking COW to free the pages_shared (but leaves mm_slots
1901 * on the list for when ksmd may be set running again).
1902 */
1903
1904 mutex_lock(&ksm_thread_mutex);
1905 if (ksm_run != flags) {
1906 ksm_run = flags;
1907 if (flags & KSM_RUN_UNMERGE) {
1908 int oom_score_adj;
1909
1910 oom_score_adj = test_set_oom_score_adj(OOM_SCORE_ADJ_MAX);
1911 err = unmerge_and_remove_all_rmap_items();
1912 compare_swap_oom_score_adj(OOM_SCORE_ADJ_MAX,
1913 oom_score_adj);
1914 if (err) {
1915 ksm_run = KSM_RUN_STOP;
1916 count = err;
1917 }
1918 }
1919 }
1920 mutex_unlock(&ksm_thread_mutex);
1921
1922 if (flags & KSM_RUN_MERGE)
1923 wake_up_interruptible(&ksm_thread_wait);
1924
1925 return count;
1926}
1927KSM_ATTR(run);
1928
1929static ssize_t pages_shared_show(struct kobject *kobj,
1930 struct kobj_attribute *attr, char *buf)
1931{
1932 return sprintf(buf, "%lu\n", ksm_pages_shared);
1933}
1934KSM_ATTR_RO(pages_shared);
1935
1936static ssize_t pages_sharing_show(struct kobject *kobj,
1937 struct kobj_attribute *attr, char *buf)
1938{
1939 return sprintf(buf, "%lu\n", ksm_pages_sharing);
1940}
1941KSM_ATTR_RO(pages_sharing);
1942
1943static ssize_t pages_unshared_show(struct kobject *kobj,
1944 struct kobj_attribute *attr, char *buf)
1945{
1946 return sprintf(buf, "%lu\n", ksm_pages_unshared);
1947}
1948KSM_ATTR_RO(pages_unshared);
1949
1950static ssize_t pages_volatile_show(struct kobject *kobj,
1951 struct kobj_attribute *attr, char *buf)
1952{
1953 long ksm_pages_volatile;
1954
1955 ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
1956 - ksm_pages_sharing - ksm_pages_unshared;
1957 /*
1958 * It was not worth any locking to calculate that statistic,
1959 * but it might therefore sometimes be negative: conceal that.
1960 */
1961 if (ksm_pages_volatile < 0)
1962 ksm_pages_volatile = 0;
1963 return sprintf(buf, "%ld\n", ksm_pages_volatile);
1964}
1965KSM_ATTR_RO(pages_volatile);
1966
1967static ssize_t full_scans_show(struct kobject *kobj,
1968 struct kobj_attribute *attr, char *buf)
1969{
1970 return sprintf(buf, "%lu\n", ksm_scan.seqnr);
1971}
1972KSM_ATTR_RO(full_scans);
1973
1974static struct attribute *ksm_attrs[] = {
1975 &sleep_millisecs_attr.attr,
1976 &pages_to_scan_attr.attr,
1977 &run_attr.attr,
1978 &pages_shared_attr.attr,
1979 &pages_sharing_attr.attr,
1980 &pages_unshared_attr.attr,
1981 &pages_volatile_attr.attr,
1982 &full_scans_attr.attr,
1983 NULL,
1984};
1985
1986static struct attribute_group ksm_attr_group = {
1987 .attrs = ksm_attrs,
1988 .name = "ksm",
1989};
1990#endif /* CONFIG_SYSFS */
1991
1992static int __init ksm_init(void)
1993{
1994 struct task_struct *ksm_thread;
1995 int err;
1996
1997 err = ksm_slab_init();
1998 if (err)
1999 goto out;
2000
2001 ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
2002 if (IS_ERR(ksm_thread)) {
2003 printk(KERN_ERR "ksm: creating kthread failed\n");
2004 err = PTR_ERR(ksm_thread);
2005 goto out_free;
2006 }
2007
2008#ifdef CONFIG_SYSFS
2009 err = sysfs_create_group(mm_kobj, &ksm_attr_group);
2010 if (err) {
2011 printk(KERN_ERR "ksm: register sysfs failed\n");
2012 kthread_stop(ksm_thread);
2013 goto out_free;
2014 }
2015#else
2016 ksm_run = KSM_RUN_MERGE; /* no way for user to start it */
2017
2018#endif /* CONFIG_SYSFS */
2019
2020#ifdef CONFIG_MEMORY_HOTREMOVE
2021 /*
2022 * Choose a high priority since the callback takes ksm_thread_mutex:
2023 * later callbacks could only be taking locks which nest within that.
2024 */
2025 hotplug_memory_notifier(ksm_memory_callback, 100);
2026#endif
2027 return 0;
2028
2029out_free:
2030 ksm_slab_free();
2031out:
2032 return err;
2033}
2034module_init(ksm_init)