blob: 942dfc73a2ff89c3c7c96b3f9b1838c99ec16d02 [file] [log] [blame]
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001/*
Izik Eidus31dbd012009-09-21 17:02:03 -07002 * 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 *
Izik Eidus36b25282009-09-21 17:02:06 -07007 * Copyright (C) 2008-2009 Red Hat, Inc.
Izik Eidus31dbd012009-09-21 17:02:03 -07008 * Authors:
9 * Izik Eidus
10 * Andrea Arcangeli
11 * Chris Wright
Izik Eidus36b25282009-09-21 17:02:06 -070012 * Hugh Dickins
Izik Eidus31dbd012009-09-21 17:02:03 -070013 *
14 * This work is licensed under the terms of the GNU GPL, version 2.
Hugh Dickinsf8af4da2009-09-21 17:01:57 -070015 */
16
17#include <linux/errno.h>
Izik Eidus31dbd012009-09-21 17:02:03 -070018#include <linux/mm.h>
19#include <linux/fs.h>
Hugh Dickinsf8af4da2009-09-21 17:01:57 -070020#include <linux/mman.h>
Izik Eidus31dbd012009-09-21 17:02:03 -070021#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>
Hugh Dickins62b61f62009-12-14 17:59:33 -080032#include <linux/memory.h>
Izik Eidus31dbd012009-09-21 17:02:03 -070033#include <linux/mmu_notifier.h>
Izik Eidus2c6854f2009-09-23 15:56:04 -070034#include <linux/swap.h>
Hugh Dickinsf8af4da2009-09-21 17:01:57 -070035#include <linux/ksm.h>
Lai Jiangshand9f89842010-08-09 17:20:02 -070036#include <linux/hash.h>
Andrea Arcangeli878aee72011-01-13 15:47:10 -080037#include <linux/freezer.h>
Hugh Dickinsf8af4da2009-09-21 17:01:57 -070038
Izik Eidus31dbd012009-09-21 17:02:03 -070039#include <asm/tlbflush.h>
Hugh Dickins73848b42009-12-14 17:59:22 -080040#include "internal.h"
Izik Eidus31dbd012009-09-21 17:02:03 -070041
42/*
43 * A few notes about the KSM scanning process,
44 * to make it easier to understand the data structures below:
45 *
46 * In order to reduce excessive scanning, KSM sorts the memory pages by their
47 * contents into a data structure that holds pointers to the pages' locations.
48 *
49 * Since the contents of the pages may change at any moment, KSM cannot just
50 * insert the pages into a normal sorted tree and expect it to find anything.
51 * Therefore KSM uses two data structures - the stable and the unstable tree.
52 *
53 * The stable tree holds pointers to all the merged pages (ksm pages), sorted
54 * by their contents. Because each such page is write-protected, searching on
55 * this tree is fully assured to be working (except when pages are unmapped),
56 * and therefore this tree is called the stable tree.
57 *
58 * In addition to the stable tree, KSM uses a second data structure called the
59 * unstable tree: this tree holds pointers to pages which have been found to
60 * be "unchanged for a period of time". The unstable tree sorts these pages
61 * by their contents, but since they are not write-protected, KSM cannot rely
62 * upon the unstable tree to work correctly - the unstable tree is liable to
63 * be corrupted as its contents are modified, and so it is called unstable.
64 *
65 * KSM solves this problem by several techniques:
66 *
67 * 1) The unstable tree is flushed every time KSM completes scanning all
68 * memory areas, and then the tree is rebuilt again from the beginning.
69 * 2) KSM will only insert into the unstable tree, pages whose hash value
70 * has not changed since the previous scan of all memory areas.
71 * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
72 * colors of the nodes and not on their contents, assuring that even when
73 * the tree gets "corrupted" it won't get out of balance, so scanning time
74 * remains the same (also, searching and inserting nodes in an rbtree uses
75 * the same algorithm, so we have no overhead when we flush and rebuild).
76 * 4) KSM never flushes the stable tree, which means that even if it were to
77 * take 10 attempts to find a page in the unstable tree, once it is found,
78 * it is secured in the stable tree. (When we scan a new page, we first
79 * compare it against the stable tree, and then against the unstable tree.)
80 */
81
82/**
83 * struct mm_slot - ksm information per mm that is being scanned
84 * @link: link to the mm_slots hash list
85 * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
Hugh Dickins6514d512009-12-14 17:59:19 -080086 * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
Izik Eidus31dbd012009-09-21 17:02:03 -070087 * @mm: the mm that this information is valid for
88 */
89struct mm_slot {
90 struct hlist_node link;
91 struct list_head mm_list;
Hugh Dickins6514d512009-12-14 17:59:19 -080092 struct rmap_item *rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -070093 struct mm_struct *mm;
94};
95
96/**
97 * struct ksm_scan - cursor for scanning
98 * @mm_slot: the current mm_slot we are scanning
99 * @address: the next address inside that to be scanned
Hugh Dickins6514d512009-12-14 17:59:19 -0800100 * @rmap_list: link to the next rmap to be scanned in the rmap_list
Izik Eidus31dbd012009-09-21 17:02:03 -0700101 * @seqnr: count of completed full scans (needed when removing unstable node)
102 *
103 * There is only the one ksm_scan instance of this cursor structure.
104 */
105struct ksm_scan {
106 struct mm_slot *mm_slot;
107 unsigned long address;
Hugh Dickins6514d512009-12-14 17:59:19 -0800108 struct rmap_item **rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -0700109 unsigned long seqnr;
110};
111
112/**
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800113 * struct stable_node - node of the stable rbtree
114 * @node: rb node of this ksm page in the stable tree
115 * @hlist: hlist head of rmap_items using this ksm page
Hugh Dickins62b61f62009-12-14 17:59:33 -0800116 * @kpfn: page frame number of this ksm page
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800117 */
118struct stable_node {
119 struct rb_node node;
120 struct hlist_head hlist;
Hugh Dickins62b61f62009-12-14 17:59:33 -0800121 unsigned long kpfn;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800122};
123
124/**
Izik Eidus31dbd012009-09-21 17:02:03 -0700125 * struct rmap_item - reverse mapping item for virtual addresses
Hugh Dickins6514d512009-12-14 17:59:19 -0800126 * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
Hugh Dickinsdb114b82009-12-14 17:59:25 -0800127 * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
Izik Eidus31dbd012009-09-21 17:02:03 -0700128 * @mm: the memory structure this rmap_item is pointing into
129 * @address: the virtual address this rmap_item tracks (+ flags in low bits)
130 * @oldchecksum: previous checksum of the page at that virtual address
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800131 * @node: rb node of this rmap_item in the unstable tree
132 * @head: pointer to stable_node heading this list in the stable tree
133 * @hlist: link into hlist of rmap_items hanging off that stable_node
Izik Eidus31dbd012009-09-21 17:02:03 -0700134 */
135struct rmap_item {
Hugh Dickins6514d512009-12-14 17:59:19 -0800136 struct rmap_item *rmap_list;
Hugh Dickinsdb114b82009-12-14 17:59:25 -0800137 struct anon_vma *anon_vma; /* when stable */
Izik Eidus31dbd012009-09-21 17:02:03 -0700138 struct mm_struct *mm;
139 unsigned long address; /* + low bits used for flags below */
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800140 unsigned int oldchecksum; /* when unstable */
Izik Eidus31dbd012009-09-21 17:02:03 -0700141 union {
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800142 struct rb_node node; /* when node of unstable tree */
143 struct { /* when listed from stable tree */
144 struct stable_node *head;
145 struct hlist_node hlist;
146 };
Izik Eidus31dbd012009-09-21 17:02:03 -0700147 };
148};
149
150#define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800151#define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */
152#define STABLE_FLAG 0x200 /* is listed from the stable tree */
Izik Eidus31dbd012009-09-21 17:02:03 -0700153
154/* The stable and unstable tree heads */
155static struct rb_root root_stable_tree = RB_ROOT;
156static struct rb_root root_unstable_tree = RB_ROOT;
157
Lai Jiangshand9f89842010-08-09 17:20:02 -0700158#define MM_SLOTS_HASH_SHIFT 10
159#define MM_SLOTS_HASH_HEADS (1 << MM_SLOTS_HASH_SHIFT)
160static struct hlist_head mm_slots_hash[MM_SLOTS_HASH_HEADS];
Izik Eidus31dbd012009-09-21 17:02:03 -0700161
162static struct mm_slot ksm_mm_head = {
163 .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
164};
165static struct ksm_scan ksm_scan = {
166 .mm_slot = &ksm_mm_head,
167};
168
169static struct kmem_cache *rmap_item_cache;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800170static struct kmem_cache *stable_node_cache;
Izik Eidus31dbd012009-09-21 17:02:03 -0700171static struct kmem_cache *mm_slot_cache;
172
173/* The number of nodes in the stable tree */
Hugh Dickinsb4028262009-09-21 17:02:09 -0700174static unsigned long ksm_pages_shared;
Izik Eidus31dbd012009-09-21 17:02:03 -0700175
Hugh Dickinse178dfd2009-09-21 17:02:10 -0700176/* The number of page slots additionally sharing those nodes */
Hugh Dickinsb4028262009-09-21 17:02:09 -0700177static unsigned long ksm_pages_sharing;
Izik Eidus31dbd012009-09-21 17:02:03 -0700178
Hugh Dickins473b0ce2009-09-21 17:02:11 -0700179/* The number of nodes in the unstable tree */
180static unsigned long ksm_pages_unshared;
181
182/* The number of rmap_items in use: to calculate pages_volatile */
183static unsigned long ksm_rmap_items;
184
Izik Eidus31dbd012009-09-21 17:02:03 -0700185/* Number of pages ksmd should scan in one batch */
Izik Eidus2c6854f2009-09-23 15:56:04 -0700186static unsigned int ksm_thread_pages_to_scan = 100;
Izik Eidus31dbd012009-09-21 17:02:03 -0700187
188/* Milliseconds ksmd should sleep between batches */
Hugh Dickins2ffd8672009-09-21 17:02:23 -0700189static unsigned int ksm_thread_sleep_millisecs = 20;
Izik Eidus31dbd012009-09-21 17:02:03 -0700190
191#define KSM_RUN_STOP 0
192#define KSM_RUN_MERGE 1
193#define KSM_RUN_UNMERGE 2
Izik Eidus2c6854f2009-09-23 15:56:04 -0700194static unsigned int ksm_run = KSM_RUN_STOP;
Izik Eidus31dbd012009-09-21 17:02:03 -0700195
196static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
197static DEFINE_MUTEX(ksm_thread_mutex);
198static DEFINE_SPINLOCK(ksm_mmlist_lock);
199
200#define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
201 sizeof(struct __struct), __alignof__(struct __struct),\
202 (__flags), NULL)
203
204static int __init ksm_slab_init(void)
205{
206 rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
207 if (!rmap_item_cache)
208 goto out;
209
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800210 stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
211 if (!stable_node_cache)
212 goto out_free1;
213
Izik Eidus31dbd012009-09-21 17:02:03 -0700214 mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
215 if (!mm_slot_cache)
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800216 goto out_free2;
Izik Eidus31dbd012009-09-21 17:02:03 -0700217
218 return 0;
219
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800220out_free2:
221 kmem_cache_destroy(stable_node_cache);
222out_free1:
Izik Eidus31dbd012009-09-21 17:02:03 -0700223 kmem_cache_destroy(rmap_item_cache);
224out:
225 return -ENOMEM;
226}
227
228static void __init ksm_slab_free(void)
229{
230 kmem_cache_destroy(mm_slot_cache);
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800231 kmem_cache_destroy(stable_node_cache);
Izik Eidus31dbd012009-09-21 17:02:03 -0700232 kmem_cache_destroy(rmap_item_cache);
233 mm_slot_cache = NULL;
234}
235
236static inline struct rmap_item *alloc_rmap_item(void)
237{
Hugh Dickins473b0ce2009-09-21 17:02:11 -0700238 struct rmap_item *rmap_item;
239
240 rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL);
241 if (rmap_item)
242 ksm_rmap_items++;
243 return rmap_item;
Izik Eidus31dbd012009-09-21 17:02:03 -0700244}
245
246static inline void free_rmap_item(struct rmap_item *rmap_item)
247{
Hugh Dickins473b0ce2009-09-21 17:02:11 -0700248 ksm_rmap_items--;
Izik Eidus31dbd012009-09-21 17:02:03 -0700249 rmap_item->mm = NULL; /* debug safety */
250 kmem_cache_free(rmap_item_cache, rmap_item);
251}
252
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800253static inline struct stable_node *alloc_stable_node(void)
254{
255 return kmem_cache_alloc(stable_node_cache, GFP_KERNEL);
256}
257
258static inline void free_stable_node(struct stable_node *stable_node)
259{
260 kmem_cache_free(stable_node_cache, stable_node);
261}
262
Izik Eidus31dbd012009-09-21 17:02:03 -0700263static inline struct mm_slot *alloc_mm_slot(void)
264{
265 if (!mm_slot_cache) /* initialization failed */
266 return NULL;
267 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
268}
269
270static inline void free_mm_slot(struct mm_slot *mm_slot)
271{
272 kmem_cache_free(mm_slot_cache, mm_slot);
273}
274
Izik Eidus31dbd012009-09-21 17:02:03 -0700275static struct mm_slot *get_mm_slot(struct mm_struct *mm)
276{
277 struct mm_slot *mm_slot;
278 struct hlist_head *bucket;
279 struct hlist_node *node;
280
Lai Jiangshand9f89842010-08-09 17:20:02 -0700281 bucket = &mm_slots_hash[hash_ptr(mm, MM_SLOTS_HASH_SHIFT)];
Izik Eidus31dbd012009-09-21 17:02:03 -0700282 hlist_for_each_entry(mm_slot, node, bucket, link) {
283 if (mm == mm_slot->mm)
284 return mm_slot;
285 }
286 return NULL;
287}
288
289static void insert_to_mm_slots_hash(struct mm_struct *mm,
290 struct mm_slot *mm_slot)
291{
292 struct hlist_head *bucket;
293
Lai Jiangshand9f89842010-08-09 17:20:02 -0700294 bucket = &mm_slots_hash[hash_ptr(mm, MM_SLOTS_HASH_SHIFT)];
Izik Eidus31dbd012009-09-21 17:02:03 -0700295 mm_slot->mm = mm;
Izik Eidus31dbd012009-09-21 17:02:03 -0700296 hlist_add_head(&mm_slot->link, bucket);
297}
298
299static inline int in_stable_tree(struct rmap_item *rmap_item)
300{
301 return rmap_item->address & STABLE_FLAG;
302}
303
304/*
Hugh Dickinsa913e182009-09-21 17:02:26 -0700305 * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
306 * page tables after it has passed through ksm_exit() - which, if necessary,
307 * takes mmap_sem briefly to serialize against them. ksm_exit() does not set
308 * a special flag: they can just back out as soon as mm_users goes to zero.
309 * ksm_test_exit() is used throughout to make this test for exit: in some
310 * places for correctness, in some places just to avoid unnecessary work.
311 */
312static inline bool ksm_test_exit(struct mm_struct *mm)
313{
314 return atomic_read(&mm->mm_users) == 0;
315}
316
317/*
Izik Eidus31dbd012009-09-21 17:02:03 -0700318 * We use break_ksm to break COW on a ksm page: it's a stripped down
319 *
320 * if (get_user_pages(current, mm, addr, 1, 1, 1, &page, NULL) == 1)
321 * put_page(page);
322 *
323 * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
324 * in case the application has unmapped and remapped mm,addr meanwhile.
325 * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP
326 * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
327 */
Hugh Dickinsd952b792009-09-21 17:02:16 -0700328static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
Izik Eidus31dbd012009-09-21 17:02:03 -0700329{
330 struct page *page;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700331 int ret = 0;
Izik Eidus31dbd012009-09-21 17:02:03 -0700332
333 do {
334 cond_resched();
335 page = follow_page(vma, addr, FOLL_GET);
Dan Carpenter22eccdd2010-04-23 13:18:10 -0400336 if (IS_ERR_OR_NULL(page))
Izik Eidus31dbd012009-09-21 17:02:03 -0700337 break;
338 if (PageKsm(page))
339 ret = handle_mm_fault(vma->vm_mm, vma, addr,
340 FAULT_FLAG_WRITE);
341 else
342 ret = VM_FAULT_WRITE;
343 put_page(page);
Hugh Dickinsd952b792009-09-21 17:02:16 -0700344 } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_OOM)));
345 /*
346 * We must loop because handle_mm_fault() may back out if there's
347 * any difficulty e.g. if pte accessed bit gets updated concurrently.
348 *
349 * VM_FAULT_WRITE is what we have been hoping for: it indicates that
350 * COW has been broken, even if the vma does not permit VM_WRITE;
351 * but note that a concurrent fault might break PageKsm for us.
352 *
353 * VM_FAULT_SIGBUS could occur if we race with truncation of the
354 * backing file, which also invalidates anonymous pages: that's
355 * okay, that truncation will have unmapped the PageKsm for us.
356 *
357 * VM_FAULT_OOM: at the time of writing (late July 2009), setting
358 * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
359 * current task has TIF_MEMDIE set, and will be OOM killed on return
360 * to user; and ksmd, having no mm, would never be chosen for that.
361 *
362 * But if the mm is in a limited mem_cgroup, then the fault may fail
363 * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
364 * even ksmd can fail in this way - though it's usually breaking ksm
365 * just to undo a merge it made a moment before, so unlikely to oom.
366 *
367 * That's a pity: we might therefore have more kernel pages allocated
368 * than we're counting as nodes in the stable tree; but ksm_do_scan
369 * will retry to break_cow on each pass, so should recover the page
370 * in due course. The important thing is to not let VM_MERGEABLE
371 * be cleared while any such pages might remain in the area.
372 */
373 return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
Izik Eidus31dbd012009-09-21 17:02:03 -0700374}
375
Hugh Dickins8dd35572009-12-14 17:59:18 -0800376static void break_cow(struct rmap_item *rmap_item)
Izik Eidus31dbd012009-09-21 17:02:03 -0700377{
Hugh Dickins8dd35572009-12-14 17:59:18 -0800378 struct mm_struct *mm = rmap_item->mm;
379 unsigned long addr = rmap_item->address;
Izik Eidus31dbd012009-09-21 17:02:03 -0700380 struct vm_area_struct *vma;
381
Hugh Dickins4035c072009-12-14 17:59:27 -0800382 /*
383 * It is not an accident that whenever we want to break COW
384 * to undo, we also need to drop a reference to the anon_vma.
385 */
Peter Zijlstra9e601092011-03-22 16:32:46 -0700386 put_anon_vma(rmap_item->anon_vma);
Hugh Dickins4035c072009-12-14 17:59:27 -0800387
Hugh Dickins81464e302009-09-21 17:02:15 -0700388 down_read(&mm->mmap_sem);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700389 if (ksm_test_exit(mm))
390 goto out;
Izik Eidus31dbd012009-09-21 17:02:03 -0700391 vma = find_vma(mm, addr);
392 if (!vma || vma->vm_start > addr)
Hugh Dickins81464e302009-09-21 17:02:15 -0700393 goto out;
Izik Eidus31dbd012009-09-21 17:02:03 -0700394 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
Hugh Dickins81464e302009-09-21 17:02:15 -0700395 goto out;
Izik Eidus31dbd012009-09-21 17:02:03 -0700396 break_ksm(vma, addr);
Hugh Dickins81464e302009-09-21 17:02:15 -0700397out:
Izik Eidus31dbd012009-09-21 17:02:03 -0700398 up_read(&mm->mmap_sem);
399}
400
Andrea Arcangeli29ad7682011-01-13 15:47:19 -0800401static struct page *page_trans_compound_anon(struct page *page)
402{
403 if (PageTransCompound(page)) {
Andrea Arcangeli22e5c472011-01-13 15:47:20 -0800404 struct page *head = compound_trans_head(page);
Andrea Arcangeli29ad7682011-01-13 15:47:19 -0800405 /*
Andrea Arcangeli22e5c472011-01-13 15:47:20 -0800406 * head may actually be splitted and freed from under
407 * us but it's ok here.
Andrea Arcangeli29ad7682011-01-13 15:47:19 -0800408 */
Andrea Arcangeli29ad7682011-01-13 15:47:19 -0800409 if (PageAnon(head))
410 return head;
411 }
412 return NULL;
413}
414
Izik Eidus31dbd012009-09-21 17:02:03 -0700415static struct page *get_mergeable_page(struct rmap_item *rmap_item)
416{
417 struct mm_struct *mm = rmap_item->mm;
418 unsigned long addr = rmap_item->address;
419 struct vm_area_struct *vma;
420 struct page *page;
421
422 down_read(&mm->mmap_sem);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700423 if (ksm_test_exit(mm))
424 goto out;
Izik Eidus31dbd012009-09-21 17:02:03 -0700425 vma = find_vma(mm, addr);
426 if (!vma || vma->vm_start > addr)
427 goto out;
428 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
429 goto out;
430
431 page = follow_page(vma, addr, FOLL_GET);
Dan Carpenter22eccdd2010-04-23 13:18:10 -0400432 if (IS_ERR_OR_NULL(page))
Izik Eidus31dbd012009-09-21 17:02:03 -0700433 goto out;
Andrea Arcangeli29ad7682011-01-13 15:47:19 -0800434 if (PageAnon(page) || page_trans_compound_anon(page)) {
Izik Eidus31dbd012009-09-21 17:02:03 -0700435 flush_anon_page(vma, page, addr);
436 flush_dcache_page(page);
437 } else {
438 put_page(page);
439out: page = NULL;
440 }
441 up_read(&mm->mmap_sem);
442 return page;
443}
444
Hugh Dickins4035c072009-12-14 17:59:27 -0800445static void remove_node_from_stable_tree(struct stable_node *stable_node)
446{
447 struct rmap_item *rmap_item;
448 struct hlist_node *hlist;
449
450 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
451 if (rmap_item->hlist.next)
452 ksm_pages_sharing--;
453 else
454 ksm_pages_shared--;
Peter Zijlstra9e601092011-03-22 16:32:46 -0700455 put_anon_vma(rmap_item->anon_vma);
Hugh Dickins4035c072009-12-14 17:59:27 -0800456 rmap_item->address &= PAGE_MASK;
457 cond_resched();
458 }
459
460 rb_erase(&stable_node->node, &root_stable_tree);
461 free_stable_node(stable_node);
462}
463
464/*
465 * get_ksm_page: checks if the page indicated by the stable node
466 * is still its ksm page, despite having held no reference to it.
467 * In which case we can trust the content of the page, and it
468 * returns the gotten page; but if the page has now been zapped,
469 * remove the stale node from the stable tree and return NULL.
470 *
471 * You would expect the stable_node to hold a reference to the ksm page.
472 * But if it increments the page's count, swapping out has to wait for
473 * ksmd to come around again before it can free the page, which may take
474 * seconds or even minutes: much too unresponsive. So instead we use a
475 * "keyhole reference": access to the ksm page from the stable node peeps
476 * out through its keyhole to see if that page still holds the right key,
477 * pointing back to this stable node. This relies on freeing a PageAnon
478 * page to reset its page->mapping to NULL, and relies on no other use of
479 * a page to put something that might look like our key in page->mapping.
480 *
481 * include/linux/pagemap.h page_cache_get_speculative() is a good reference,
482 * but this is different - made simpler by ksm_thread_mutex being held, but
483 * interesting for assuming that no other use of the struct page could ever
484 * put our expected_mapping into page->mapping (or a field of the union which
485 * coincides with page->mapping). The RCU calls are not for KSM at all, but
486 * to keep the page_count protocol described with page_cache_get_speculative.
487 *
488 * Note: it is possible that get_ksm_page() will return NULL one moment,
489 * then page the next, if the page is in between page_freeze_refs() and
490 * page_unfreeze_refs(): this shouldn't be a problem anywhere, the page
491 * is on its way to being freed; but it is an anomaly to bear in mind.
492 */
493static struct page *get_ksm_page(struct stable_node *stable_node)
494{
495 struct page *page;
496 void *expected_mapping;
497
Hugh Dickins62b61f62009-12-14 17:59:33 -0800498 page = pfn_to_page(stable_node->kpfn);
Hugh Dickins4035c072009-12-14 17:59:27 -0800499 expected_mapping = (void *)stable_node +
500 (PAGE_MAPPING_ANON | PAGE_MAPPING_KSM);
501 rcu_read_lock();
502 if (page->mapping != expected_mapping)
503 goto stale;
504 if (!get_page_unless_zero(page))
505 goto stale;
506 if (page->mapping != expected_mapping) {
507 put_page(page);
508 goto stale;
509 }
510 rcu_read_unlock();
511 return page;
512stale:
513 rcu_read_unlock();
514 remove_node_from_stable_tree(stable_node);
515 return NULL;
516}
517
Izik Eidus31dbd012009-09-21 17:02:03 -0700518/*
Izik Eidus31dbd012009-09-21 17:02:03 -0700519 * Removing rmap_item from stable or unstable tree.
520 * This function will clean the information from the stable/unstable tree.
521 */
522static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
523{
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800524 if (rmap_item->address & STABLE_FLAG) {
525 struct stable_node *stable_node;
Hugh Dickins5ad64682009-12-14 17:59:24 -0800526 struct page *page;
Izik Eidus31dbd012009-09-21 17:02:03 -0700527
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800528 stable_node = rmap_item->head;
Hugh Dickins4035c072009-12-14 17:59:27 -0800529 page = get_ksm_page(stable_node);
530 if (!page)
531 goto out;
532
Hugh Dickins5ad64682009-12-14 17:59:24 -0800533 lock_page(page);
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800534 hlist_del(&rmap_item->hlist);
Hugh Dickins4035c072009-12-14 17:59:27 -0800535 unlock_page(page);
536 put_page(page);
Hugh Dickins08beca42009-12-14 17:59:21 -0800537
Hugh Dickins4035c072009-12-14 17:59:27 -0800538 if (stable_node->hlist.first)
539 ksm_pages_sharing--;
540 else
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800541 ksm_pages_shared--;
Izik Eidus31dbd012009-09-21 17:02:03 -0700542
Peter Zijlstra9e601092011-03-22 16:32:46 -0700543 put_anon_vma(rmap_item->anon_vma);
Hugh Dickins93d17712009-12-14 17:59:16 -0800544 rmap_item->address &= PAGE_MASK;
Izik Eidus31dbd012009-09-21 17:02:03 -0700545
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800546 } else if (rmap_item->address & UNSTABLE_FLAG) {
Izik Eidus31dbd012009-09-21 17:02:03 -0700547 unsigned char age;
548 /*
Hugh Dickins9ba69292009-09-21 17:02:20 -0700549 * Usually ksmd can and must skip the rb_erase, because
Izik Eidus31dbd012009-09-21 17:02:03 -0700550 * root_unstable_tree was already reset to RB_ROOT.
Hugh Dickins9ba69292009-09-21 17:02:20 -0700551 * But be careful when an mm is exiting: do the rb_erase
552 * if this rmap_item was inserted by this scan, rather
553 * than left over from before.
Izik Eidus31dbd012009-09-21 17:02:03 -0700554 */
555 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
Hugh Dickinscd551f92009-09-21 17:02:17 -0700556 BUG_ON(age > 1);
Izik Eidus31dbd012009-09-21 17:02:03 -0700557 if (!age)
558 rb_erase(&rmap_item->node, &root_unstable_tree);
Izik Eidus31dbd012009-09-21 17:02:03 -0700559
Hugh Dickins93d17712009-12-14 17:59:16 -0800560 ksm_pages_unshared--;
561 rmap_item->address &= PAGE_MASK;
562 }
Hugh Dickins4035c072009-12-14 17:59:27 -0800563out:
Izik Eidus31dbd012009-09-21 17:02:03 -0700564 cond_resched(); /* we're called from many long loops */
565}
566
Izik Eidus31dbd012009-09-21 17:02:03 -0700567static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
Hugh Dickins6514d512009-12-14 17:59:19 -0800568 struct rmap_item **rmap_list)
Izik Eidus31dbd012009-09-21 17:02:03 -0700569{
Hugh Dickins6514d512009-12-14 17:59:19 -0800570 while (*rmap_list) {
571 struct rmap_item *rmap_item = *rmap_list;
572 *rmap_list = rmap_item->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -0700573 remove_rmap_item_from_tree(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -0700574 free_rmap_item(rmap_item);
575 }
576}
577
578/*
579 * Though it's very tempting to unmerge in_stable_tree(rmap_item)s rather
580 * than check every pte of a given vma, the locking doesn't quite work for
581 * that - an rmap_item is assigned to the stable tree after inserting ksm
582 * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing
583 * rmap_items from parent to child at fork time (so as not to waste time
584 * if exit comes before the next scan reaches it).
Hugh Dickins81464e302009-09-21 17:02:15 -0700585 *
586 * Similarly, although we'd like to remove rmap_items (so updating counts
587 * and freeing memory) when unmerging an area, it's easier to leave that
588 * to the next pass of ksmd - consider, for example, how ksmd might be
589 * in cmp_and_merge_page on one of the rmap_items we would be removing.
Izik Eidus31dbd012009-09-21 17:02:03 -0700590 */
Hugh Dickinsd952b792009-09-21 17:02:16 -0700591static int unmerge_ksm_pages(struct vm_area_struct *vma,
592 unsigned long start, unsigned long end)
Izik Eidus31dbd012009-09-21 17:02:03 -0700593{
594 unsigned long addr;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700595 int err = 0;
Izik Eidus31dbd012009-09-21 17:02:03 -0700596
Hugh Dickinsd952b792009-09-21 17:02:16 -0700597 for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
Hugh Dickins9ba69292009-09-21 17:02:20 -0700598 if (ksm_test_exit(vma->vm_mm))
599 break;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700600 if (signal_pending(current))
601 err = -ERESTARTSYS;
602 else
603 err = break_ksm(vma, addr);
604 }
605 return err;
Izik Eidus31dbd012009-09-21 17:02:03 -0700606}
607
Hugh Dickins2ffd8672009-09-21 17:02:23 -0700608#ifdef CONFIG_SYSFS
609/*
610 * Only called through the sysfs control interface:
611 */
Hugh Dickinsd952b792009-09-21 17:02:16 -0700612static int unmerge_and_remove_all_rmap_items(void)
Izik Eidus31dbd012009-09-21 17:02:03 -0700613{
614 struct mm_slot *mm_slot;
615 struct mm_struct *mm;
616 struct vm_area_struct *vma;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700617 int err = 0;
Izik Eidus31dbd012009-09-21 17:02:03 -0700618
Hugh Dickinsd952b792009-09-21 17:02:16 -0700619 spin_lock(&ksm_mmlist_lock);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700620 ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
Hugh Dickinsd952b792009-09-21 17:02:16 -0700621 struct mm_slot, mm_list);
622 spin_unlock(&ksm_mmlist_lock);
623
Hugh Dickins9ba69292009-09-21 17:02:20 -0700624 for (mm_slot = ksm_scan.mm_slot;
625 mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
Izik Eidus31dbd012009-09-21 17:02:03 -0700626 mm = mm_slot->mm;
627 down_read(&mm->mmap_sem);
628 for (vma = mm->mmap; vma; vma = vma->vm_next) {
Hugh Dickins9ba69292009-09-21 17:02:20 -0700629 if (ksm_test_exit(mm))
630 break;
Izik Eidus31dbd012009-09-21 17:02:03 -0700631 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
632 continue;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700633 err = unmerge_ksm_pages(vma,
634 vma->vm_start, vma->vm_end);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700635 if (err)
636 goto error;
Izik Eidus31dbd012009-09-21 17:02:03 -0700637 }
Hugh Dickins9ba69292009-09-21 17:02:20 -0700638
Hugh Dickins6514d512009-12-14 17:59:19 -0800639 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
Hugh Dickinsd952b792009-09-21 17:02:16 -0700640
641 spin_lock(&ksm_mmlist_lock);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700642 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
Hugh Dickinsd952b792009-09-21 17:02:16 -0700643 struct mm_slot, mm_list);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700644 if (ksm_test_exit(mm)) {
645 hlist_del(&mm_slot->link);
646 list_del(&mm_slot->mm_list);
647 spin_unlock(&ksm_mmlist_lock);
648
649 free_mm_slot(mm_slot);
650 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
651 up_read(&mm->mmap_sem);
652 mmdrop(mm);
653 } else {
654 spin_unlock(&ksm_mmlist_lock);
655 up_read(&mm->mmap_sem);
656 }
Izik Eidus31dbd012009-09-21 17:02:03 -0700657 }
658
Hugh Dickinsd952b792009-09-21 17:02:16 -0700659 ksm_scan.seqnr = 0;
Hugh Dickins9ba69292009-09-21 17:02:20 -0700660 return 0;
661
662error:
663 up_read(&mm->mmap_sem);
Izik Eidus31dbd012009-09-21 17:02:03 -0700664 spin_lock(&ksm_mmlist_lock);
Hugh Dickinsd952b792009-09-21 17:02:16 -0700665 ksm_scan.mm_slot = &ksm_mm_head;
Izik Eidus31dbd012009-09-21 17:02:03 -0700666 spin_unlock(&ksm_mmlist_lock);
Hugh Dickinsd952b792009-09-21 17:02:16 -0700667 return err;
Izik Eidus31dbd012009-09-21 17:02:03 -0700668}
Hugh Dickins2ffd8672009-09-21 17:02:23 -0700669#endif /* CONFIG_SYSFS */
Izik Eidus31dbd012009-09-21 17:02:03 -0700670
Izik Eidus31dbd012009-09-21 17:02:03 -0700671static u32 calc_checksum(struct page *page)
672{
673 u32 checksum;
674 void *addr = kmap_atomic(page, KM_USER0);
675 checksum = jhash2(addr, PAGE_SIZE / 4, 17);
676 kunmap_atomic(addr, KM_USER0);
677 return checksum;
678}
679
680static int memcmp_pages(struct page *page1, struct page *page2)
681{
682 char *addr1, *addr2;
683 int ret;
684
685 addr1 = kmap_atomic(page1, KM_USER0);
686 addr2 = kmap_atomic(page2, KM_USER1);
687 ret = memcmp(addr1, addr2, PAGE_SIZE);
688 kunmap_atomic(addr2, KM_USER1);
689 kunmap_atomic(addr1, KM_USER0);
690 return ret;
691}
692
693static inline int pages_identical(struct page *page1, struct page *page2)
694{
695 return !memcmp_pages(page1, page2);
696}
697
698static int write_protect_page(struct vm_area_struct *vma, struct page *page,
699 pte_t *orig_pte)
700{
701 struct mm_struct *mm = vma->vm_mm;
702 unsigned long addr;
703 pte_t *ptep;
704 spinlock_t *ptl;
705 int swapped;
706 int err = -EFAULT;
707
708 addr = page_address_in_vma(page, vma);
709 if (addr == -EFAULT)
710 goto out;
711
Andrea Arcangeli29ad7682011-01-13 15:47:19 -0800712 BUG_ON(PageTransCompound(page));
Izik Eidus31dbd012009-09-21 17:02:03 -0700713 ptep = page_check_address(page, mm, addr, &ptl, 0);
714 if (!ptep)
715 goto out;
716
Hugh Dickins4e316352010-10-02 17:49:08 -0700717 if (pte_write(*ptep) || pte_dirty(*ptep)) {
Izik Eidus31dbd012009-09-21 17:02:03 -0700718 pte_t entry;
719
720 swapped = PageSwapCache(page);
721 flush_cache_page(vma, addr, page_to_pfn(page));
722 /*
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300723 * Ok this is tricky, when get_user_pages_fast() run it doesn't
Izik Eidus31dbd012009-09-21 17:02:03 -0700724 * take any lock, therefore the check that we are going to make
725 * with the pagecount against the mapcount is racey and
726 * O_DIRECT can happen right after the check.
727 * So we clear the pte and flush the tlb before the check
728 * this assure us that no O_DIRECT can happen after the check
729 * or in the middle of the check.
730 */
731 entry = ptep_clear_flush(vma, addr, ptep);
732 /*
733 * Check that no O_DIRECT or similar I/O is in progress on the
734 * page
735 */
Hugh Dickins31e855e2009-12-14 17:59:17 -0800736 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
Robin Holtcb532372010-03-23 13:35:26 -0700737 set_pte_at(mm, addr, ptep, entry);
Izik Eidus31dbd012009-09-21 17:02:03 -0700738 goto out_unlock;
739 }
Hugh Dickins4e316352010-10-02 17:49:08 -0700740 if (pte_dirty(entry))
741 set_page_dirty(page);
742 entry = pte_mkclean(pte_wrprotect(entry));
Izik Eidus31dbd012009-09-21 17:02:03 -0700743 set_pte_at_notify(mm, addr, ptep, entry);
744 }
745 *orig_pte = *ptep;
746 err = 0;
747
748out_unlock:
749 pte_unmap_unlock(ptep, ptl);
750out:
751 return err;
752}
753
754/**
755 * replace_page - replace page in vma by new ksm page
Hugh Dickins8dd35572009-12-14 17:59:18 -0800756 * @vma: vma that holds the pte pointing to page
757 * @page: the page we are replacing by kpage
758 * @kpage: the ksm page we replace page by
Izik Eidus31dbd012009-09-21 17:02:03 -0700759 * @orig_pte: the original value of the pte
760 *
761 * Returns 0 on success, -EFAULT on failure.
762 */
Hugh Dickins8dd35572009-12-14 17:59:18 -0800763static int replace_page(struct vm_area_struct *vma, struct page *page,
764 struct page *kpage, pte_t orig_pte)
Izik Eidus31dbd012009-09-21 17:02:03 -0700765{
766 struct mm_struct *mm = vma->vm_mm;
767 pgd_t *pgd;
768 pud_t *pud;
769 pmd_t *pmd;
770 pte_t *ptep;
771 spinlock_t *ptl;
772 unsigned long addr;
Izik Eidus31dbd012009-09-21 17:02:03 -0700773 int err = -EFAULT;
774
Hugh Dickins8dd35572009-12-14 17:59:18 -0800775 addr = page_address_in_vma(page, vma);
Izik Eidus31dbd012009-09-21 17:02:03 -0700776 if (addr == -EFAULT)
777 goto out;
778
779 pgd = pgd_offset(mm, addr);
780 if (!pgd_present(*pgd))
781 goto out;
782
783 pud = pud_offset(pgd, addr);
784 if (!pud_present(*pud))
785 goto out;
786
787 pmd = pmd_offset(pud, addr);
Andrea Arcangeli29ad7682011-01-13 15:47:19 -0800788 BUG_ON(pmd_trans_huge(*pmd));
Izik Eidus31dbd012009-09-21 17:02:03 -0700789 if (!pmd_present(*pmd))
790 goto out;
791
792 ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
793 if (!pte_same(*ptep, orig_pte)) {
794 pte_unmap_unlock(ptep, ptl);
795 goto out;
796 }
797
Hugh Dickins8dd35572009-12-14 17:59:18 -0800798 get_page(kpage);
Hugh Dickins5ad64682009-12-14 17:59:24 -0800799 page_add_anon_rmap(kpage, vma, addr);
Izik Eidus31dbd012009-09-21 17:02:03 -0700800
801 flush_cache_page(vma, addr, pte_pfn(*ptep));
802 ptep_clear_flush(vma, addr, ptep);
Hugh Dickins8dd35572009-12-14 17:59:18 -0800803 set_pte_at_notify(mm, addr, ptep, mk_pte(kpage, vma->vm_page_prot));
Izik Eidus31dbd012009-09-21 17:02:03 -0700804
Hugh Dickins8dd35572009-12-14 17:59:18 -0800805 page_remove_rmap(page);
Hugh Dickinsae52a2a2011-01-13 15:46:28 -0800806 if (!page_mapped(page))
807 try_to_free_swap(page);
Hugh Dickins8dd35572009-12-14 17:59:18 -0800808 put_page(page);
Izik Eidus31dbd012009-09-21 17:02:03 -0700809
810 pte_unmap_unlock(ptep, ptl);
811 err = 0;
812out:
813 return err;
814}
815
Andrea Arcangeli29ad7682011-01-13 15:47:19 -0800816static int page_trans_compound_anon_split(struct page *page)
817{
818 int ret = 0;
819 struct page *transhuge_head = page_trans_compound_anon(page);
820 if (transhuge_head) {
821 /* Get the reference on the head to split it. */
822 if (get_page_unless_zero(transhuge_head)) {
823 /*
824 * Recheck we got the reference while the head
825 * was still anonymous.
826 */
827 if (PageAnon(transhuge_head))
828 ret = split_huge_page(transhuge_head);
829 else
830 /*
831 * Retry later if split_huge_page run
832 * from under us.
833 */
834 ret = 1;
835 put_page(transhuge_head);
836 } else
837 /* Retry later if split_huge_page run from under us. */
838 ret = 1;
839 }
840 return ret;
841}
842
Izik Eidus31dbd012009-09-21 17:02:03 -0700843/*
844 * try_to_merge_one_page - take two pages and merge them into one
Hugh Dickins8dd35572009-12-14 17:59:18 -0800845 * @vma: the vma that holds the pte pointing to page
846 * @page: the PageAnon page that we want to replace with kpage
Hugh Dickins80e14822009-12-14 17:59:29 -0800847 * @kpage: the PageKsm page that we want to map instead of page,
848 * or NULL the first time when we want to use page as kpage.
Izik Eidus31dbd012009-09-21 17:02:03 -0700849 *
850 * This function returns 0 if the pages were merged, -EFAULT otherwise.
851 */
852static int try_to_merge_one_page(struct vm_area_struct *vma,
Hugh Dickins8dd35572009-12-14 17:59:18 -0800853 struct page *page, struct page *kpage)
Izik Eidus31dbd012009-09-21 17:02:03 -0700854{
855 pte_t orig_pte = __pte(0);
856 int err = -EFAULT;
857
Hugh Dickinsdb114b82009-12-14 17:59:25 -0800858 if (page == kpage) /* ksm page forked */
859 return 0;
860
Izik Eidus31dbd012009-09-21 17:02:03 -0700861 if (!(vma->vm_flags & VM_MERGEABLE))
862 goto out;
Andrea Arcangeli29ad7682011-01-13 15:47:19 -0800863 if (PageTransCompound(page) && page_trans_compound_anon_split(page))
864 goto out;
865 BUG_ON(PageTransCompound(page));
Hugh Dickins8dd35572009-12-14 17:59:18 -0800866 if (!PageAnon(page))
Izik Eidus31dbd012009-09-21 17:02:03 -0700867 goto out;
868
Izik Eidus31dbd012009-09-21 17:02:03 -0700869 /*
870 * We need the page lock to read a stable PageSwapCache in
871 * write_protect_page(). We use trylock_page() instead of
872 * lock_page() because we don't want to wait here - we
873 * prefer to continue scanning and merging different pages,
874 * then come back to this page when it is unlocked.
875 */
Hugh Dickins8dd35572009-12-14 17:59:18 -0800876 if (!trylock_page(page))
Hugh Dickins31e855e2009-12-14 17:59:17 -0800877 goto out;
Izik Eidus31dbd012009-09-21 17:02:03 -0700878 /*
879 * If this anonymous page is mapped only here, its pte may need
880 * to be write-protected. If it's mapped elsewhere, all of its
881 * ptes are necessarily already write-protected. But in either
882 * case, we need to lock and check page_count is not raised.
883 */
Hugh Dickins80e14822009-12-14 17:59:29 -0800884 if (write_protect_page(vma, page, &orig_pte) == 0) {
885 if (!kpage) {
886 /*
887 * While we hold page lock, upgrade page from
888 * PageAnon+anon_vma to PageKsm+NULL stable_node:
889 * stable_tree_insert() will update stable_node.
890 */
891 set_page_stable_node(page, NULL);
892 mark_page_accessed(page);
893 err = 0;
894 } else if (pages_identical(page, kpage))
895 err = replace_page(vma, page, kpage, orig_pte);
896 }
Izik Eidus31dbd012009-09-21 17:02:03 -0700897
Hugh Dickins80e14822009-12-14 17:59:29 -0800898 if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
Hugh Dickins73848b42009-12-14 17:59:22 -0800899 munlock_vma_page(page);
Hugh Dickins5ad64682009-12-14 17:59:24 -0800900 if (!PageMlocked(kpage)) {
901 unlock_page(page);
Hugh Dickins5ad64682009-12-14 17:59:24 -0800902 lock_page(kpage);
903 mlock_vma_page(kpage);
904 page = kpage; /* for final unlock */
905 }
906 }
Hugh Dickins73848b42009-12-14 17:59:22 -0800907
Hugh Dickins8dd35572009-12-14 17:59:18 -0800908 unlock_page(page);
Izik Eidus31dbd012009-09-21 17:02:03 -0700909out:
910 return err;
911}
912
913/*
Hugh Dickins81464e302009-09-21 17:02:15 -0700914 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
915 * but no new kernel page is allocated: kpage must already be a ksm page.
Hugh Dickins8dd35572009-12-14 17:59:18 -0800916 *
917 * This function returns 0 if the pages were merged, -EFAULT otherwise.
Hugh Dickins81464e302009-09-21 17:02:15 -0700918 */
Hugh Dickins8dd35572009-12-14 17:59:18 -0800919static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
920 struct page *page, struct page *kpage)
Hugh Dickins81464e302009-09-21 17:02:15 -0700921{
Hugh Dickins8dd35572009-12-14 17:59:18 -0800922 struct mm_struct *mm = rmap_item->mm;
Hugh Dickins81464e302009-09-21 17:02:15 -0700923 struct vm_area_struct *vma;
924 int err = -EFAULT;
925
Hugh Dickins8dd35572009-12-14 17:59:18 -0800926 down_read(&mm->mmap_sem);
927 if (ksm_test_exit(mm))
928 goto out;
929 vma = find_vma(mm, rmap_item->address);
930 if (!vma || vma->vm_start > rmap_item->address)
Hugh Dickins9ba69292009-09-21 17:02:20 -0700931 goto out;
932
Hugh Dickins8dd35572009-12-14 17:59:18 -0800933 err = try_to_merge_one_page(vma, page, kpage);
Hugh Dickinsdb114b82009-12-14 17:59:25 -0800934 if (err)
935 goto out;
936
937 /* Must get reference to anon_vma while still holding mmap_sem */
Peter Zijlstra9e601092011-03-22 16:32:46 -0700938 rmap_item->anon_vma = vma->anon_vma;
939 get_anon_vma(vma->anon_vma);
Hugh Dickins81464e302009-09-21 17:02:15 -0700940out:
Hugh Dickins8dd35572009-12-14 17:59:18 -0800941 up_read(&mm->mmap_sem);
Hugh Dickins81464e302009-09-21 17:02:15 -0700942 return err;
943}
944
945/*
Izik Eidus31dbd012009-09-21 17:02:03 -0700946 * try_to_merge_two_pages - take two identical pages and prepare them
947 * to be merged into one page.
948 *
Hugh Dickins8dd35572009-12-14 17:59:18 -0800949 * This function returns the kpage if we successfully merged two identical
950 * pages into one ksm page, NULL otherwise.
Izik Eidus31dbd012009-09-21 17:02:03 -0700951 *
Hugh Dickins80e14822009-12-14 17:59:29 -0800952 * Note that this function upgrades page to ksm page: if one of the pages
Izik Eidus31dbd012009-09-21 17:02:03 -0700953 * is already a ksm page, try_to_merge_with_ksm_page should be used.
954 */
Hugh Dickins8dd35572009-12-14 17:59:18 -0800955static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
956 struct page *page,
957 struct rmap_item *tree_rmap_item,
958 struct page *tree_page)
Izik Eidus31dbd012009-09-21 17:02:03 -0700959{
Hugh Dickins80e14822009-12-14 17:59:29 -0800960 int err;
Izik Eidus31dbd012009-09-21 17:02:03 -0700961
Hugh Dickins80e14822009-12-14 17:59:29 -0800962 err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
Izik Eidus31dbd012009-09-21 17:02:03 -0700963 if (!err) {
Hugh Dickins8dd35572009-12-14 17:59:18 -0800964 err = try_to_merge_with_ksm_page(tree_rmap_item,
Hugh Dickins80e14822009-12-14 17:59:29 -0800965 tree_page, page);
Izik Eidus31dbd012009-09-21 17:02:03 -0700966 /*
Hugh Dickins81464e302009-09-21 17:02:15 -0700967 * If that fails, we have a ksm page with only one pte
968 * pointing to it: so break it.
Izik Eidus31dbd012009-09-21 17:02:03 -0700969 */
Hugh Dickins4035c072009-12-14 17:59:27 -0800970 if (err)
Hugh Dickins8dd35572009-12-14 17:59:18 -0800971 break_cow(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -0700972 }
Hugh Dickins80e14822009-12-14 17:59:29 -0800973 return err ? NULL : page;
Izik Eidus31dbd012009-09-21 17:02:03 -0700974}
975
976/*
Hugh Dickins8dd35572009-12-14 17:59:18 -0800977 * stable_tree_search - search for page inside the stable tree
Izik Eidus31dbd012009-09-21 17:02:03 -0700978 *
979 * This function checks if there is a page inside the stable tree
980 * with identical content to the page that we are scanning right now.
981 *
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800982 * This function returns the stable tree node of identical content if found,
Izik Eidus31dbd012009-09-21 17:02:03 -0700983 * NULL otherwise.
984 */
Hugh Dickins62b61f62009-12-14 17:59:33 -0800985static struct page *stable_tree_search(struct page *page)
Izik Eidus31dbd012009-09-21 17:02:03 -0700986{
987 struct rb_node *node = root_stable_tree.rb_node;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800988 struct stable_node *stable_node;
Izik Eidus31dbd012009-09-21 17:02:03 -0700989
Hugh Dickins08beca42009-12-14 17:59:21 -0800990 stable_node = page_stable_node(page);
991 if (stable_node) { /* ksm page forked */
992 get_page(page);
Hugh Dickins62b61f62009-12-14 17:59:33 -0800993 return page;
Hugh Dickins08beca42009-12-14 17:59:21 -0800994 }
995
Izik Eidus31dbd012009-09-21 17:02:03 -0700996 while (node) {
Hugh Dickins4035c072009-12-14 17:59:27 -0800997 struct page *tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -0700998 int ret;
999
Hugh Dickins08beca42009-12-14 17:59:21 -08001000 cond_resched();
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001001 stable_node = rb_entry(node, struct stable_node, node);
Hugh Dickins4035c072009-12-14 17:59:27 -08001002 tree_page = get_ksm_page(stable_node);
1003 if (!tree_page)
1004 return NULL;
Izik Eidus31dbd012009-09-21 17:02:03 -07001005
Hugh Dickins4035c072009-12-14 17:59:27 -08001006 ret = memcmp_pages(page, tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001007
Hugh Dickins4035c072009-12-14 17:59:27 -08001008 if (ret < 0) {
1009 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001010 node = node->rb_left;
Hugh Dickins4035c072009-12-14 17:59:27 -08001011 } else if (ret > 0) {
1012 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001013 node = node->rb_right;
Hugh Dickins4035c072009-12-14 17:59:27 -08001014 } else
Hugh Dickins62b61f62009-12-14 17:59:33 -08001015 return tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001016 }
1017
1018 return NULL;
1019}
1020
1021/*
1022 * stable_tree_insert - insert rmap_item pointing to new ksm page
1023 * into the stable tree.
1024 *
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001025 * This function returns the stable tree node just allocated on success,
1026 * NULL otherwise.
Izik Eidus31dbd012009-09-21 17:02:03 -07001027 */
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001028static struct stable_node *stable_tree_insert(struct page *kpage)
Izik Eidus31dbd012009-09-21 17:02:03 -07001029{
1030 struct rb_node **new = &root_stable_tree.rb_node;
1031 struct rb_node *parent = NULL;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001032 struct stable_node *stable_node;
Izik Eidus31dbd012009-09-21 17:02:03 -07001033
1034 while (*new) {
Hugh Dickins4035c072009-12-14 17:59:27 -08001035 struct page *tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001036 int ret;
1037
Hugh Dickins08beca42009-12-14 17:59:21 -08001038 cond_resched();
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001039 stable_node = rb_entry(*new, struct stable_node, node);
Hugh Dickins4035c072009-12-14 17:59:27 -08001040 tree_page = get_ksm_page(stable_node);
1041 if (!tree_page)
1042 return NULL;
Izik Eidus31dbd012009-09-21 17:02:03 -07001043
Hugh Dickins4035c072009-12-14 17:59:27 -08001044 ret = memcmp_pages(kpage, tree_page);
1045 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001046
1047 parent = *new;
1048 if (ret < 0)
1049 new = &parent->rb_left;
1050 else if (ret > 0)
1051 new = &parent->rb_right;
1052 else {
1053 /*
1054 * It is not a bug that stable_tree_search() didn't
1055 * find this node: because at that time our page was
1056 * not yet write-protected, so may have changed since.
1057 */
1058 return NULL;
1059 }
1060 }
1061
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001062 stable_node = alloc_stable_node();
1063 if (!stable_node)
1064 return NULL;
Izik Eidus31dbd012009-09-21 17:02:03 -07001065
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001066 rb_link_node(&stable_node->node, parent, new);
1067 rb_insert_color(&stable_node->node, &root_stable_tree);
1068
1069 INIT_HLIST_HEAD(&stable_node->hlist);
1070
Hugh Dickins62b61f62009-12-14 17:59:33 -08001071 stable_node->kpfn = page_to_pfn(kpage);
Hugh Dickins08beca42009-12-14 17:59:21 -08001072 set_page_stable_node(kpage, stable_node);
1073
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001074 return stable_node;
Izik Eidus31dbd012009-09-21 17:02:03 -07001075}
1076
1077/*
Hugh Dickins8dd35572009-12-14 17:59:18 -08001078 * unstable_tree_search_insert - search for identical page,
1079 * else insert rmap_item into the unstable tree.
Izik Eidus31dbd012009-09-21 17:02:03 -07001080 *
1081 * This function searches for a page in the unstable tree identical to the
1082 * page currently being scanned; and if no identical page is found in the
1083 * tree, we insert rmap_item as a new object into the unstable tree.
1084 *
1085 * This function returns pointer to rmap_item found to be identical
1086 * to the currently scanned page, NULL otherwise.
1087 *
1088 * This function does both searching and inserting, because they share
1089 * the same walking algorithm in an rbtree.
1090 */
Hugh Dickins8dd35572009-12-14 17:59:18 -08001091static
1092struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1093 struct page *page,
1094 struct page **tree_pagep)
1095
Izik Eidus31dbd012009-09-21 17:02:03 -07001096{
1097 struct rb_node **new = &root_unstable_tree.rb_node;
1098 struct rb_node *parent = NULL;
1099
1100 while (*new) {
1101 struct rmap_item *tree_rmap_item;
Hugh Dickins8dd35572009-12-14 17:59:18 -08001102 struct page *tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001103 int ret;
1104
Hugh Dickinsd178f272009-11-09 15:58:23 +00001105 cond_resched();
Izik Eidus31dbd012009-09-21 17:02:03 -07001106 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
Hugh Dickins8dd35572009-12-14 17:59:18 -08001107 tree_page = get_mergeable_page(tree_rmap_item);
Dan Carpenter22eccdd2010-04-23 13:18:10 -04001108 if (IS_ERR_OR_NULL(tree_page))
Izik Eidus31dbd012009-09-21 17:02:03 -07001109 return NULL;
1110
1111 /*
Hugh Dickins8dd35572009-12-14 17:59:18 -08001112 * Don't substitute a ksm page for a forked page.
Izik Eidus31dbd012009-09-21 17:02:03 -07001113 */
Hugh Dickins8dd35572009-12-14 17:59:18 -08001114 if (page == tree_page) {
1115 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001116 return NULL;
1117 }
1118
Hugh Dickins8dd35572009-12-14 17:59:18 -08001119 ret = memcmp_pages(page, tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001120
1121 parent = *new;
1122 if (ret < 0) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001123 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001124 new = &parent->rb_left;
1125 } else if (ret > 0) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001126 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001127 new = &parent->rb_right;
1128 } else {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001129 *tree_pagep = tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001130 return tree_rmap_item;
1131 }
1132 }
1133
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001134 rmap_item->address |= UNSTABLE_FLAG;
Izik Eidus31dbd012009-09-21 17:02:03 -07001135 rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1136 rb_link_node(&rmap_item->node, parent, new);
1137 rb_insert_color(&rmap_item->node, &root_unstable_tree);
1138
Hugh Dickins473b0ce2009-09-21 17:02:11 -07001139 ksm_pages_unshared++;
Izik Eidus31dbd012009-09-21 17:02:03 -07001140 return NULL;
1141}
1142
1143/*
1144 * stable_tree_append - add another rmap_item to the linked list of
1145 * rmap_items hanging off a given node of the stable tree, all sharing
1146 * the same ksm page.
1147 */
1148static void stable_tree_append(struct rmap_item *rmap_item,
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001149 struct stable_node *stable_node)
Izik Eidus31dbd012009-09-21 17:02:03 -07001150{
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001151 rmap_item->head = stable_node;
Izik Eidus31dbd012009-09-21 17:02:03 -07001152 rmap_item->address |= STABLE_FLAG;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001153 hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
Hugh Dickinse178dfd2009-09-21 17:02:10 -07001154
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001155 if (rmap_item->hlist.next)
1156 ksm_pages_sharing++;
1157 else
1158 ksm_pages_shared++;
Izik Eidus31dbd012009-09-21 17:02:03 -07001159}
1160
1161/*
Hugh Dickins81464e302009-09-21 17:02:15 -07001162 * cmp_and_merge_page - first see if page can be merged into the stable tree;
1163 * if not, compare checksum to previous and if it's the same, see if page can
1164 * be inserted into the unstable tree, or merged with a page already there and
1165 * both transferred to the stable tree.
Izik Eidus31dbd012009-09-21 17:02:03 -07001166 *
1167 * @page: the page that we are searching identical page to.
1168 * @rmap_item: the reverse mapping into the virtual address of this page
1169 */
1170static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1171{
Izik Eidus31dbd012009-09-21 17:02:03 -07001172 struct rmap_item *tree_rmap_item;
Hugh Dickins8dd35572009-12-14 17:59:18 -08001173 struct page *tree_page = NULL;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001174 struct stable_node *stable_node;
Hugh Dickins8dd35572009-12-14 17:59:18 -08001175 struct page *kpage;
Izik Eidus31dbd012009-09-21 17:02:03 -07001176 unsigned int checksum;
1177 int err;
1178
Hugh Dickins93d17712009-12-14 17:59:16 -08001179 remove_rmap_item_from_tree(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -07001180
1181 /* We first start with searching the page inside the stable tree */
Hugh Dickins62b61f62009-12-14 17:59:33 -08001182 kpage = stable_tree_search(page);
1183 if (kpage) {
Hugh Dickins08beca42009-12-14 17:59:21 -08001184 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
Izik Eidus31dbd012009-09-21 17:02:03 -07001185 if (!err) {
1186 /*
1187 * The page was successfully merged:
1188 * add its rmap_item to the stable tree.
1189 */
Hugh Dickins5ad64682009-12-14 17:59:24 -08001190 lock_page(kpage);
Hugh Dickins62b61f62009-12-14 17:59:33 -08001191 stable_tree_append(rmap_item, page_stable_node(kpage));
Hugh Dickins5ad64682009-12-14 17:59:24 -08001192 unlock_page(kpage);
Izik Eidus31dbd012009-09-21 17:02:03 -07001193 }
Hugh Dickins8dd35572009-12-14 17:59:18 -08001194 put_page(kpage);
Izik Eidus31dbd012009-09-21 17:02:03 -07001195 return;
1196 }
1197
1198 /*
Hugh Dickins4035c072009-12-14 17:59:27 -08001199 * If the hash value of the page has changed from the last time
1200 * we calculated it, this page is changing frequently: therefore we
1201 * don't want to insert it in the unstable tree, and we don't want
1202 * to waste our time searching for something identical to it there.
Izik Eidus31dbd012009-09-21 17:02:03 -07001203 */
1204 checksum = calc_checksum(page);
1205 if (rmap_item->oldchecksum != checksum) {
1206 rmap_item->oldchecksum = checksum;
1207 return;
1208 }
1209
Hugh Dickins8dd35572009-12-14 17:59:18 -08001210 tree_rmap_item =
1211 unstable_tree_search_insert(rmap_item, page, &tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001212 if (tree_rmap_item) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001213 kpage = try_to_merge_two_pages(rmap_item, page,
1214 tree_rmap_item, tree_page);
1215 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001216 /*
1217 * As soon as we merge this page, we want to remove the
1218 * rmap_item of the page we have merged with from the unstable
1219 * tree, and insert it instead as new node in the stable tree.
1220 */
Hugh Dickins8dd35572009-12-14 17:59:18 -08001221 if (kpage) {
Hugh Dickins93d17712009-12-14 17:59:16 -08001222 remove_rmap_item_from_tree(tree_rmap_item);
Hugh Dickins473b0ce2009-09-21 17:02:11 -07001223
Hugh Dickins5ad64682009-12-14 17:59:24 -08001224 lock_page(kpage);
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001225 stable_node = stable_tree_insert(kpage);
1226 if (stable_node) {
1227 stable_tree_append(tree_rmap_item, stable_node);
1228 stable_tree_append(rmap_item, stable_node);
1229 }
Hugh Dickins5ad64682009-12-14 17:59:24 -08001230 unlock_page(kpage);
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001231
Izik Eidus31dbd012009-09-21 17:02:03 -07001232 /*
1233 * If we fail to insert the page into the stable tree,
1234 * we will have 2 virtual addresses that are pointing
1235 * to a ksm page left outside the stable tree,
1236 * in which case we need to break_cow on both.
1237 */
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001238 if (!stable_node) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001239 break_cow(tree_rmap_item);
1240 break_cow(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -07001241 }
1242 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001243 }
1244}
1245
1246static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
Hugh Dickins6514d512009-12-14 17:59:19 -08001247 struct rmap_item **rmap_list,
Izik Eidus31dbd012009-09-21 17:02:03 -07001248 unsigned long addr)
1249{
1250 struct rmap_item *rmap_item;
1251
Hugh Dickins6514d512009-12-14 17:59:19 -08001252 while (*rmap_list) {
1253 rmap_item = *rmap_list;
Hugh Dickins93d17712009-12-14 17:59:16 -08001254 if ((rmap_item->address & PAGE_MASK) == addr)
Izik Eidus31dbd012009-09-21 17:02:03 -07001255 return rmap_item;
Izik Eidus31dbd012009-09-21 17:02:03 -07001256 if (rmap_item->address > addr)
1257 break;
Hugh Dickins6514d512009-12-14 17:59:19 -08001258 *rmap_list = rmap_item->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -07001259 remove_rmap_item_from_tree(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -07001260 free_rmap_item(rmap_item);
1261 }
1262
1263 rmap_item = alloc_rmap_item();
1264 if (rmap_item) {
1265 /* It has already been zeroed */
1266 rmap_item->mm = mm_slot->mm;
1267 rmap_item->address = addr;
Hugh Dickins6514d512009-12-14 17:59:19 -08001268 rmap_item->rmap_list = *rmap_list;
1269 *rmap_list = rmap_item;
Izik Eidus31dbd012009-09-21 17:02:03 -07001270 }
1271 return rmap_item;
1272}
1273
1274static struct rmap_item *scan_get_next_rmap_item(struct page **page)
1275{
1276 struct mm_struct *mm;
1277 struct mm_slot *slot;
1278 struct vm_area_struct *vma;
1279 struct rmap_item *rmap_item;
1280
1281 if (list_empty(&ksm_mm_head.mm_list))
1282 return NULL;
1283
1284 slot = ksm_scan.mm_slot;
1285 if (slot == &ksm_mm_head) {
Hugh Dickins2919bfd2011-01-13 15:47:29 -08001286 /*
1287 * A number of pages can hang around indefinitely on per-cpu
1288 * pagevecs, raised page count preventing write_protect_page
1289 * from merging them. Though it doesn't really matter much,
1290 * it is puzzling to see some stuck in pages_volatile until
1291 * other activity jostles them out, and they also prevented
1292 * LTP's KSM test from succeeding deterministically; so drain
1293 * them here (here rather than on entry to ksm_do_scan(),
1294 * so we don't IPI too often when pages_to_scan is set low).
1295 */
1296 lru_add_drain_all();
1297
Izik Eidus31dbd012009-09-21 17:02:03 -07001298 root_unstable_tree = RB_ROOT;
1299
1300 spin_lock(&ksm_mmlist_lock);
1301 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1302 ksm_scan.mm_slot = slot;
1303 spin_unlock(&ksm_mmlist_lock);
1304next_mm:
1305 ksm_scan.address = 0;
Hugh Dickins6514d512009-12-14 17:59:19 -08001306 ksm_scan.rmap_list = &slot->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -07001307 }
1308
1309 mm = slot->mm;
1310 down_read(&mm->mmap_sem);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001311 if (ksm_test_exit(mm))
1312 vma = NULL;
1313 else
1314 vma = find_vma(mm, ksm_scan.address);
1315
1316 for (; vma; vma = vma->vm_next) {
Izik Eidus31dbd012009-09-21 17:02:03 -07001317 if (!(vma->vm_flags & VM_MERGEABLE))
1318 continue;
1319 if (ksm_scan.address < vma->vm_start)
1320 ksm_scan.address = vma->vm_start;
1321 if (!vma->anon_vma)
1322 ksm_scan.address = vma->vm_end;
1323
1324 while (ksm_scan.address < vma->vm_end) {
Hugh Dickins9ba69292009-09-21 17:02:20 -07001325 if (ksm_test_exit(mm))
1326 break;
Izik Eidus31dbd012009-09-21 17:02:03 -07001327 *page = follow_page(vma, ksm_scan.address, FOLL_GET);
Andrea Arcangeli21ae5b02011-01-13 15:47:00 -08001328 if (IS_ERR_OR_NULL(*page)) {
1329 ksm_scan.address += PAGE_SIZE;
1330 cond_resched();
1331 continue;
1332 }
Andrea Arcangeli29ad7682011-01-13 15:47:19 -08001333 if (PageAnon(*page) ||
1334 page_trans_compound_anon(*page)) {
Izik Eidus31dbd012009-09-21 17:02:03 -07001335 flush_anon_page(vma, *page, ksm_scan.address);
1336 flush_dcache_page(*page);
1337 rmap_item = get_next_rmap_item(slot,
Hugh Dickins6514d512009-12-14 17:59:19 -08001338 ksm_scan.rmap_list, ksm_scan.address);
Izik Eidus31dbd012009-09-21 17:02:03 -07001339 if (rmap_item) {
Hugh Dickins6514d512009-12-14 17:59:19 -08001340 ksm_scan.rmap_list =
1341 &rmap_item->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -07001342 ksm_scan.address += PAGE_SIZE;
1343 } else
1344 put_page(*page);
1345 up_read(&mm->mmap_sem);
1346 return rmap_item;
1347 }
Andrea Arcangeli21ae5b02011-01-13 15:47:00 -08001348 put_page(*page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001349 ksm_scan.address += PAGE_SIZE;
1350 cond_resched();
1351 }
1352 }
1353
Hugh Dickins9ba69292009-09-21 17:02:20 -07001354 if (ksm_test_exit(mm)) {
1355 ksm_scan.address = 0;
Hugh Dickins6514d512009-12-14 17:59:19 -08001356 ksm_scan.rmap_list = &slot->rmap_list;
Hugh Dickins9ba69292009-09-21 17:02:20 -07001357 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001358 /*
1359 * Nuke all the rmap_items that are above this current rmap:
1360 * because there were no VM_MERGEABLE vmas with such addresses.
1361 */
Hugh Dickins6514d512009-12-14 17:59:19 -08001362 remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
Izik Eidus31dbd012009-09-21 17:02:03 -07001363
1364 spin_lock(&ksm_mmlist_lock);
Hugh Dickinscd551f92009-09-21 17:02:17 -07001365 ksm_scan.mm_slot = list_entry(slot->mm_list.next,
1366 struct mm_slot, mm_list);
1367 if (ksm_scan.address == 0) {
1368 /*
1369 * We've completed a full scan of all vmas, holding mmap_sem
1370 * throughout, and found no VM_MERGEABLE: so do the same as
1371 * __ksm_exit does to remove this mm from all our lists now.
Hugh Dickins9ba69292009-09-21 17:02:20 -07001372 * This applies either when cleaning up after __ksm_exit
1373 * (but beware: we can reach here even before __ksm_exit),
1374 * or when all VM_MERGEABLE areas have been unmapped (and
1375 * mmap_sem then protects against race with MADV_MERGEABLE).
Hugh Dickinscd551f92009-09-21 17:02:17 -07001376 */
1377 hlist_del(&slot->link);
1378 list_del(&slot->mm_list);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001379 spin_unlock(&ksm_mmlist_lock);
1380
Hugh Dickinscd551f92009-09-21 17:02:17 -07001381 free_mm_slot(slot);
1382 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001383 up_read(&mm->mmap_sem);
1384 mmdrop(mm);
1385 } else {
1386 spin_unlock(&ksm_mmlist_lock);
1387 up_read(&mm->mmap_sem);
Hugh Dickinscd551f92009-09-21 17:02:17 -07001388 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001389
1390 /* Repeat until we've completed scanning the whole list */
Hugh Dickinscd551f92009-09-21 17:02:17 -07001391 slot = ksm_scan.mm_slot;
Izik Eidus31dbd012009-09-21 17:02:03 -07001392 if (slot != &ksm_mm_head)
1393 goto next_mm;
1394
Izik Eidus31dbd012009-09-21 17:02:03 -07001395 ksm_scan.seqnr++;
1396 return NULL;
1397}
1398
1399/**
1400 * ksm_do_scan - the ksm scanner main worker function.
1401 * @scan_npages - number of pages we want to scan before we return.
1402 */
1403static void ksm_do_scan(unsigned int scan_npages)
1404{
1405 struct rmap_item *rmap_item;
Dan Carpenter22eccdd2010-04-23 13:18:10 -04001406 struct page *uninitialized_var(page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001407
Andrea Arcangeli878aee72011-01-13 15:47:10 -08001408 while (scan_npages-- && likely(!freezing(current))) {
Izik Eidus31dbd012009-09-21 17:02:03 -07001409 cond_resched();
1410 rmap_item = scan_get_next_rmap_item(&page);
1411 if (!rmap_item)
1412 return;
1413 if (!PageKsm(page) || !in_stable_tree(rmap_item))
1414 cmp_and_merge_page(page, rmap_item);
1415 put_page(page);
1416 }
1417}
1418
Hugh Dickins6e158382009-09-21 17:02:14 -07001419static int ksmd_should_run(void)
1420{
1421 return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
1422}
1423
Izik Eidus31dbd012009-09-21 17:02:03 -07001424static int ksm_scan_thread(void *nothing)
1425{
Andrea Arcangeli878aee72011-01-13 15:47:10 -08001426 set_freezable();
Izik Eidus339aa622009-09-21 17:02:07 -07001427 set_user_nice(current, 5);
Izik Eidus31dbd012009-09-21 17:02:03 -07001428
1429 while (!kthread_should_stop()) {
Hugh Dickins6e158382009-09-21 17:02:14 -07001430 mutex_lock(&ksm_thread_mutex);
1431 if (ksmd_should_run())
Izik Eidus31dbd012009-09-21 17:02:03 -07001432 ksm_do_scan(ksm_thread_pages_to_scan);
Hugh Dickins6e158382009-09-21 17:02:14 -07001433 mutex_unlock(&ksm_thread_mutex);
1434
Andrea Arcangeli878aee72011-01-13 15:47:10 -08001435 try_to_freeze();
1436
Hugh Dickins6e158382009-09-21 17:02:14 -07001437 if (ksmd_should_run()) {
Izik Eidus31dbd012009-09-21 17:02:03 -07001438 schedule_timeout_interruptible(
1439 msecs_to_jiffies(ksm_thread_sleep_millisecs));
1440 } else {
Andrea Arcangeli878aee72011-01-13 15:47:10 -08001441 wait_event_freezable(ksm_thread_wait,
Hugh Dickins6e158382009-09-21 17:02:14 -07001442 ksmd_should_run() || kthread_should_stop());
Izik Eidus31dbd012009-09-21 17:02:03 -07001443 }
1444 }
1445 return 0;
1446}
1447
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001448int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1449 unsigned long end, int advice, unsigned long *vm_flags)
1450{
1451 struct mm_struct *mm = vma->vm_mm;
Hugh Dickinsd952b792009-09-21 17:02:16 -07001452 int err;
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001453
1454 switch (advice) {
1455 case MADV_MERGEABLE:
1456 /*
1457 * Be somewhat over-protective for now!
1458 */
1459 if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE |
1460 VM_PFNMAP | VM_IO | VM_DONTEXPAND |
1461 VM_RESERVED | VM_HUGETLB | VM_INSERTPAGE |
Hugh Dickins5ad64682009-12-14 17:59:24 -08001462 VM_NONLINEAR | VM_MIXEDMAP | VM_SAO))
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001463 return 0; /* just ignore the advice */
1464
Hugh Dickinsd952b792009-09-21 17:02:16 -07001465 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
1466 err = __ksm_enter(mm);
1467 if (err)
1468 return err;
1469 }
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001470
1471 *vm_flags |= VM_MERGEABLE;
1472 break;
1473
1474 case MADV_UNMERGEABLE:
1475 if (!(*vm_flags & VM_MERGEABLE))
1476 return 0; /* just ignore the advice */
1477
Hugh Dickinsd952b792009-09-21 17:02:16 -07001478 if (vma->anon_vma) {
1479 err = unmerge_ksm_pages(vma, start, end);
1480 if (err)
1481 return err;
1482 }
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001483
1484 *vm_flags &= ~VM_MERGEABLE;
1485 break;
1486 }
1487
1488 return 0;
1489}
1490
1491int __ksm_enter(struct mm_struct *mm)
1492{
Hugh Dickins6e158382009-09-21 17:02:14 -07001493 struct mm_slot *mm_slot;
1494 int needs_wakeup;
1495
1496 mm_slot = alloc_mm_slot();
Izik Eidus31dbd012009-09-21 17:02:03 -07001497 if (!mm_slot)
1498 return -ENOMEM;
1499
Hugh Dickins6e158382009-09-21 17:02:14 -07001500 /* Check ksm_run too? Would need tighter locking */
1501 needs_wakeup = list_empty(&ksm_mm_head.mm_list);
1502
Izik Eidus31dbd012009-09-21 17:02:03 -07001503 spin_lock(&ksm_mmlist_lock);
1504 insert_to_mm_slots_hash(mm, mm_slot);
1505 /*
1506 * Insert just behind the scanning cursor, to let the area settle
1507 * down a little; when fork is followed by immediate exec, we don't
1508 * want ksmd to waste time setting up and tearing down an rmap_list.
1509 */
1510 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
1511 spin_unlock(&ksm_mmlist_lock);
1512
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001513 set_bit(MMF_VM_MERGEABLE, &mm->flags);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001514 atomic_inc(&mm->mm_count);
Hugh Dickins6e158382009-09-21 17:02:14 -07001515
1516 if (needs_wakeup)
1517 wake_up_interruptible(&ksm_thread_wait);
1518
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001519 return 0;
1520}
1521
Andrea Arcangeli1c2fb7a2009-09-21 17:02:22 -07001522void __ksm_exit(struct mm_struct *mm)
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001523{
Hugh Dickinscd551f92009-09-21 17:02:17 -07001524 struct mm_slot *mm_slot;
Hugh Dickins9ba69292009-09-21 17:02:20 -07001525 int easy_to_free = 0;
Hugh Dickinscd551f92009-09-21 17:02:17 -07001526
Izik Eidus31dbd012009-09-21 17:02:03 -07001527 /*
Hugh Dickins9ba69292009-09-21 17:02:20 -07001528 * This process is exiting: if it's straightforward (as is the
1529 * case when ksmd was never running), free mm_slot immediately.
1530 * But if it's at the cursor or has rmap_items linked to it, use
1531 * mmap_sem to synchronize with any break_cows before pagetables
1532 * are freed, and leave the mm_slot on the list for ksmd to free.
1533 * Beware: ksm may already have noticed it exiting and freed the slot.
Izik Eidus31dbd012009-09-21 17:02:03 -07001534 */
Hugh Dickins9ba69292009-09-21 17:02:20 -07001535
Hugh Dickinscd551f92009-09-21 17:02:17 -07001536 spin_lock(&ksm_mmlist_lock);
1537 mm_slot = get_mm_slot(mm);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001538 if (mm_slot && ksm_scan.mm_slot != mm_slot) {
Hugh Dickins6514d512009-12-14 17:59:19 -08001539 if (!mm_slot->rmap_list) {
Hugh Dickins9ba69292009-09-21 17:02:20 -07001540 hlist_del(&mm_slot->link);
1541 list_del(&mm_slot->mm_list);
1542 easy_to_free = 1;
1543 } else {
1544 list_move(&mm_slot->mm_list,
1545 &ksm_scan.mm_slot->mm_list);
1546 }
Hugh Dickinscd551f92009-09-21 17:02:17 -07001547 }
Hugh Dickinscd551f92009-09-21 17:02:17 -07001548 spin_unlock(&ksm_mmlist_lock);
1549
Hugh Dickins9ba69292009-09-21 17:02:20 -07001550 if (easy_to_free) {
1551 free_mm_slot(mm_slot);
1552 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1553 mmdrop(mm);
1554 } else if (mm_slot) {
Hugh Dickins9ba69292009-09-21 17:02:20 -07001555 down_write(&mm->mmap_sem);
1556 up_write(&mm->mmap_sem);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001557 }
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001558}
Izik Eidus31dbd012009-09-21 17:02:03 -07001559
Hugh Dickins5ad64682009-12-14 17:59:24 -08001560struct page *ksm_does_need_to_copy(struct page *page,
1561 struct vm_area_struct *vma, unsigned long address)
1562{
1563 struct page *new_page;
1564
Hugh Dickins5ad64682009-12-14 17:59:24 -08001565 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1566 if (new_page) {
1567 copy_user_highpage(new_page, page, address, vma);
1568
1569 SetPageDirty(new_page);
1570 __SetPageUptodate(new_page);
1571 SetPageSwapBacked(new_page);
1572 __set_page_locked(new_page);
1573
1574 if (page_evictable(new_page, vma))
1575 lru_cache_add_lru(new_page, LRU_ACTIVE_ANON);
1576 else
1577 add_page_to_unevictable_list(new_page);
1578 }
1579
Hugh Dickins5ad64682009-12-14 17:59:24 -08001580 return new_page;
1581}
1582
1583int page_referenced_ksm(struct page *page, struct mem_cgroup *memcg,
1584 unsigned long *vm_flags)
1585{
1586 struct stable_node *stable_node;
1587 struct rmap_item *rmap_item;
1588 struct hlist_node *hlist;
1589 unsigned int mapcount = page_mapcount(page);
1590 int referenced = 0;
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001591 int search_new_forks = 0;
Hugh Dickins5ad64682009-12-14 17:59:24 -08001592
1593 VM_BUG_ON(!PageKsm(page));
1594 VM_BUG_ON(!PageLocked(page));
1595
1596 stable_node = page_stable_node(page);
1597 if (!stable_node)
1598 return 0;
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001599again:
Hugh Dickins5ad64682009-12-14 17:59:24 -08001600 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001601 struct anon_vma *anon_vma = rmap_item->anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -08001602 struct anon_vma_chain *vmac;
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001603 struct vm_area_struct *vma;
Hugh Dickins5ad64682009-12-14 17:59:24 -08001604
Rik van Rielcba48b92010-08-09 17:18:38 -07001605 anon_vma_lock(anon_vma);
Rik van Riel5beb4932010-03-05 13:42:07 -08001606 list_for_each_entry(vmac, &anon_vma->head, same_anon_vma) {
1607 vma = vmac->vma;
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001608 if (rmap_item->address < vma->vm_start ||
1609 rmap_item->address >= vma->vm_end)
1610 continue;
1611 /*
1612 * Initially we examine only the vma which covers this
1613 * rmap_item; but later, if there is still work to do,
1614 * we examine covering vmas in other mms: in case they
1615 * were forked from the original since ksmd passed.
1616 */
1617 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1618 continue;
Hugh Dickins5ad64682009-12-14 17:59:24 -08001619
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001620 if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
1621 continue;
1622
1623 referenced += page_referenced_one(page, vma,
Hugh Dickins5ad64682009-12-14 17:59:24 -08001624 rmap_item->address, &mapcount, vm_flags);
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001625 if (!search_new_forks || !mapcount)
1626 break;
1627 }
Rik van Rielcba48b92010-08-09 17:18:38 -07001628 anon_vma_unlock(anon_vma);
Hugh Dickins5ad64682009-12-14 17:59:24 -08001629 if (!mapcount)
1630 goto out;
1631 }
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001632 if (!search_new_forks++)
1633 goto again;
Hugh Dickins5ad64682009-12-14 17:59:24 -08001634out:
Hugh Dickins5ad64682009-12-14 17:59:24 -08001635 return referenced;
1636}
1637
1638int try_to_unmap_ksm(struct page *page, enum ttu_flags flags)
1639{
1640 struct stable_node *stable_node;
1641 struct hlist_node *hlist;
1642 struct rmap_item *rmap_item;
1643 int ret = SWAP_AGAIN;
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001644 int search_new_forks = 0;
Hugh Dickins5ad64682009-12-14 17:59:24 -08001645
1646 VM_BUG_ON(!PageKsm(page));
1647 VM_BUG_ON(!PageLocked(page));
1648
1649 stable_node = page_stable_node(page);
1650 if (!stable_node)
1651 return SWAP_FAIL;
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001652again:
Hugh Dickins5ad64682009-12-14 17:59:24 -08001653 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001654 struct anon_vma *anon_vma = rmap_item->anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -08001655 struct anon_vma_chain *vmac;
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001656 struct vm_area_struct *vma;
Hugh Dickins5ad64682009-12-14 17:59:24 -08001657
Rik van Rielcba48b92010-08-09 17:18:38 -07001658 anon_vma_lock(anon_vma);
Rik van Riel5beb4932010-03-05 13:42:07 -08001659 list_for_each_entry(vmac, &anon_vma->head, same_anon_vma) {
1660 vma = vmac->vma;
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001661 if (rmap_item->address < vma->vm_start ||
1662 rmap_item->address >= vma->vm_end)
1663 continue;
1664 /*
1665 * Initially we examine only the vma which covers this
1666 * rmap_item; but later, if there is still work to do,
1667 * we examine covering vmas in other mms: in case they
1668 * were forked from the original since ksmd passed.
1669 */
1670 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1671 continue;
1672
1673 ret = try_to_unmap_one(page, vma,
1674 rmap_item->address, flags);
1675 if (ret != SWAP_AGAIN || !page_mapped(page)) {
Rik van Rielcba48b92010-08-09 17:18:38 -07001676 anon_vma_unlock(anon_vma);
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001677 goto out;
1678 }
1679 }
Rik van Rielcba48b92010-08-09 17:18:38 -07001680 anon_vma_unlock(anon_vma);
Hugh Dickins5ad64682009-12-14 17:59:24 -08001681 }
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001682 if (!search_new_forks++)
1683 goto again;
Hugh Dickins5ad64682009-12-14 17:59:24 -08001684out:
Hugh Dickins5ad64682009-12-14 17:59:24 -08001685 return ret;
1686}
1687
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001688#ifdef CONFIG_MIGRATION
1689int rmap_walk_ksm(struct page *page, int (*rmap_one)(struct page *,
1690 struct vm_area_struct *, unsigned long, void *), void *arg)
1691{
1692 struct stable_node *stable_node;
1693 struct hlist_node *hlist;
1694 struct rmap_item *rmap_item;
1695 int ret = SWAP_AGAIN;
1696 int search_new_forks = 0;
1697
1698 VM_BUG_ON(!PageKsm(page));
1699 VM_BUG_ON(!PageLocked(page));
1700
1701 stable_node = page_stable_node(page);
1702 if (!stable_node)
1703 return ret;
1704again:
1705 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1706 struct anon_vma *anon_vma = rmap_item->anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -08001707 struct anon_vma_chain *vmac;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001708 struct vm_area_struct *vma;
1709
Rik van Rielcba48b92010-08-09 17:18:38 -07001710 anon_vma_lock(anon_vma);
Rik van Riel5beb4932010-03-05 13:42:07 -08001711 list_for_each_entry(vmac, &anon_vma->head, same_anon_vma) {
1712 vma = vmac->vma;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001713 if (rmap_item->address < vma->vm_start ||
1714 rmap_item->address >= vma->vm_end)
1715 continue;
1716 /*
1717 * Initially we examine only the vma which covers this
1718 * rmap_item; but later, if there is still work to do,
1719 * we examine covering vmas in other mms: in case they
1720 * were forked from the original since ksmd passed.
1721 */
1722 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1723 continue;
1724
1725 ret = rmap_one(page, vma, rmap_item->address, arg);
1726 if (ret != SWAP_AGAIN) {
Rik van Rielcba48b92010-08-09 17:18:38 -07001727 anon_vma_unlock(anon_vma);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001728 goto out;
1729 }
1730 }
Rik van Rielcba48b92010-08-09 17:18:38 -07001731 anon_vma_unlock(anon_vma);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001732 }
1733 if (!search_new_forks++)
1734 goto again;
1735out:
1736 return ret;
1737}
1738
1739void ksm_migrate_page(struct page *newpage, struct page *oldpage)
1740{
1741 struct stable_node *stable_node;
1742
1743 VM_BUG_ON(!PageLocked(oldpage));
1744 VM_BUG_ON(!PageLocked(newpage));
1745 VM_BUG_ON(newpage->mapping != oldpage->mapping);
1746
1747 stable_node = page_stable_node(newpage);
1748 if (stable_node) {
Hugh Dickins62b61f62009-12-14 17:59:33 -08001749 VM_BUG_ON(stable_node->kpfn != page_to_pfn(oldpage));
1750 stable_node->kpfn = page_to_pfn(newpage);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001751 }
1752}
1753#endif /* CONFIG_MIGRATION */
1754
Hugh Dickins62b61f62009-12-14 17:59:33 -08001755#ifdef CONFIG_MEMORY_HOTREMOVE
1756static struct stable_node *ksm_check_stable_tree(unsigned long start_pfn,
1757 unsigned long end_pfn)
1758{
1759 struct rb_node *node;
1760
1761 for (node = rb_first(&root_stable_tree); node; node = rb_next(node)) {
1762 struct stable_node *stable_node;
1763
1764 stable_node = rb_entry(node, struct stable_node, node);
1765 if (stable_node->kpfn >= start_pfn &&
1766 stable_node->kpfn < end_pfn)
1767 return stable_node;
1768 }
1769 return NULL;
1770}
1771
1772static int ksm_memory_callback(struct notifier_block *self,
1773 unsigned long action, void *arg)
1774{
1775 struct memory_notify *mn = arg;
1776 struct stable_node *stable_node;
1777
1778 switch (action) {
1779 case MEM_GOING_OFFLINE:
1780 /*
1781 * Keep it very simple for now: just lock out ksmd and
1782 * MADV_UNMERGEABLE while any memory is going offline.
KOSAKI Motohiroa0b0f582010-12-02 14:31:20 -08001783 * mutex_lock_nested() is necessary because lockdep was alarmed
1784 * that here we take ksm_thread_mutex inside notifier chain
1785 * mutex, and later take notifier chain mutex inside
1786 * ksm_thread_mutex to unlock it. But that's safe because both
1787 * are inside mem_hotplug_mutex.
Hugh Dickins62b61f62009-12-14 17:59:33 -08001788 */
KOSAKI Motohiroa0b0f582010-12-02 14:31:20 -08001789 mutex_lock_nested(&ksm_thread_mutex, SINGLE_DEPTH_NESTING);
Hugh Dickins62b61f62009-12-14 17:59:33 -08001790 break;
1791
1792 case MEM_OFFLINE:
1793 /*
1794 * Most of the work is done by page migration; but there might
1795 * be a few stable_nodes left over, still pointing to struct
1796 * pages which have been offlined: prune those from the tree.
1797 */
1798 while ((stable_node = ksm_check_stable_tree(mn->start_pfn,
1799 mn->start_pfn + mn->nr_pages)) != NULL)
1800 remove_node_from_stable_tree(stable_node);
1801 /* fallthrough */
1802
1803 case MEM_CANCEL_OFFLINE:
1804 mutex_unlock(&ksm_thread_mutex);
1805 break;
1806 }
1807 return NOTIFY_OK;
1808}
1809#endif /* CONFIG_MEMORY_HOTREMOVE */
1810
Hugh Dickins2ffd8672009-09-21 17:02:23 -07001811#ifdef CONFIG_SYSFS
1812/*
1813 * This all compiles without CONFIG_SYSFS, but is a waste of space.
1814 */
1815
Izik Eidus31dbd012009-09-21 17:02:03 -07001816#define KSM_ATTR_RO(_name) \
1817 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1818#define KSM_ATTR(_name) \
1819 static struct kobj_attribute _name##_attr = \
1820 __ATTR(_name, 0644, _name##_show, _name##_store)
1821
1822static ssize_t sleep_millisecs_show(struct kobject *kobj,
1823 struct kobj_attribute *attr, char *buf)
1824{
1825 return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
1826}
1827
1828static ssize_t sleep_millisecs_store(struct kobject *kobj,
1829 struct kobj_attribute *attr,
1830 const char *buf, size_t count)
1831{
1832 unsigned long msecs;
1833 int err;
1834
1835 err = strict_strtoul(buf, 10, &msecs);
1836 if (err || msecs > UINT_MAX)
1837 return -EINVAL;
1838
1839 ksm_thread_sleep_millisecs = msecs;
1840
1841 return count;
1842}
1843KSM_ATTR(sleep_millisecs);
1844
1845static ssize_t pages_to_scan_show(struct kobject *kobj,
1846 struct kobj_attribute *attr, char *buf)
1847{
1848 return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
1849}
1850
1851static ssize_t pages_to_scan_store(struct kobject *kobj,
1852 struct kobj_attribute *attr,
1853 const char *buf, size_t count)
1854{
1855 int err;
1856 unsigned long nr_pages;
1857
1858 err = strict_strtoul(buf, 10, &nr_pages);
1859 if (err || nr_pages > UINT_MAX)
1860 return -EINVAL;
1861
1862 ksm_thread_pages_to_scan = nr_pages;
1863
1864 return count;
1865}
1866KSM_ATTR(pages_to_scan);
1867
1868static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
1869 char *buf)
1870{
1871 return sprintf(buf, "%u\n", ksm_run);
1872}
1873
1874static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
1875 const char *buf, size_t count)
1876{
1877 int err;
1878 unsigned long flags;
1879
1880 err = strict_strtoul(buf, 10, &flags);
1881 if (err || flags > UINT_MAX)
1882 return -EINVAL;
1883 if (flags > KSM_RUN_UNMERGE)
1884 return -EINVAL;
1885
1886 /*
1887 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
1888 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
Hugh Dickinsd0f209f2009-12-14 17:59:34 -08001889 * breaking COW to free the pages_shared (but leaves mm_slots
1890 * on the list for when ksmd may be set running again).
Izik Eidus31dbd012009-09-21 17:02:03 -07001891 */
1892
1893 mutex_lock(&ksm_thread_mutex);
1894 if (ksm_run != flags) {
1895 ksm_run = flags;
Hugh Dickinsd952b792009-09-21 17:02:16 -07001896 if (flags & KSM_RUN_UNMERGE) {
Hugh Dickins35451be2009-09-21 17:02:27 -07001897 current->flags |= PF_OOM_ORIGIN;
Hugh Dickinsd952b792009-09-21 17:02:16 -07001898 err = unmerge_and_remove_all_rmap_items();
Hugh Dickins35451be2009-09-21 17:02:27 -07001899 current->flags &= ~PF_OOM_ORIGIN;
Hugh Dickinsd952b792009-09-21 17:02:16 -07001900 if (err) {
1901 ksm_run = KSM_RUN_STOP;
1902 count = err;
1903 }
1904 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001905 }
1906 mutex_unlock(&ksm_thread_mutex);
1907
1908 if (flags & KSM_RUN_MERGE)
1909 wake_up_interruptible(&ksm_thread_wait);
1910
1911 return count;
1912}
1913KSM_ATTR(run);
1914
Hugh Dickinsb4028262009-09-21 17:02:09 -07001915static ssize_t pages_shared_show(struct kobject *kobj,
1916 struct kobj_attribute *attr, char *buf)
1917{
1918 return sprintf(buf, "%lu\n", ksm_pages_shared);
1919}
1920KSM_ATTR_RO(pages_shared);
1921
1922static ssize_t pages_sharing_show(struct kobject *kobj,
1923 struct kobj_attribute *attr, char *buf)
1924{
Hugh Dickinse178dfd2009-09-21 17:02:10 -07001925 return sprintf(buf, "%lu\n", ksm_pages_sharing);
Hugh Dickinsb4028262009-09-21 17:02:09 -07001926}
1927KSM_ATTR_RO(pages_sharing);
1928
Hugh Dickins473b0ce2009-09-21 17:02:11 -07001929static ssize_t pages_unshared_show(struct kobject *kobj,
1930 struct kobj_attribute *attr, char *buf)
1931{
1932 return sprintf(buf, "%lu\n", ksm_pages_unshared);
1933}
1934KSM_ATTR_RO(pages_unshared);
1935
1936static ssize_t pages_volatile_show(struct kobject *kobj,
1937 struct kobj_attribute *attr, char *buf)
1938{
1939 long ksm_pages_volatile;
1940
1941 ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
1942 - ksm_pages_sharing - ksm_pages_unshared;
1943 /*
1944 * It was not worth any locking to calculate that statistic,
1945 * but it might therefore sometimes be negative: conceal that.
1946 */
1947 if (ksm_pages_volatile < 0)
1948 ksm_pages_volatile = 0;
1949 return sprintf(buf, "%ld\n", ksm_pages_volatile);
1950}
1951KSM_ATTR_RO(pages_volatile);
1952
1953static ssize_t full_scans_show(struct kobject *kobj,
1954 struct kobj_attribute *attr, char *buf)
1955{
1956 return sprintf(buf, "%lu\n", ksm_scan.seqnr);
1957}
1958KSM_ATTR_RO(full_scans);
1959
Izik Eidus31dbd012009-09-21 17:02:03 -07001960static struct attribute *ksm_attrs[] = {
1961 &sleep_millisecs_attr.attr,
1962 &pages_to_scan_attr.attr,
1963 &run_attr.attr,
Hugh Dickinsb4028262009-09-21 17:02:09 -07001964 &pages_shared_attr.attr,
1965 &pages_sharing_attr.attr,
Hugh Dickins473b0ce2009-09-21 17:02:11 -07001966 &pages_unshared_attr.attr,
1967 &pages_volatile_attr.attr,
1968 &full_scans_attr.attr,
Izik Eidus31dbd012009-09-21 17:02:03 -07001969 NULL,
1970};
1971
1972static struct attribute_group ksm_attr_group = {
1973 .attrs = ksm_attrs,
1974 .name = "ksm",
1975};
Hugh Dickins2ffd8672009-09-21 17:02:23 -07001976#endif /* CONFIG_SYSFS */
Izik Eidus31dbd012009-09-21 17:02:03 -07001977
1978static int __init ksm_init(void)
1979{
1980 struct task_struct *ksm_thread;
1981 int err;
1982
1983 err = ksm_slab_init();
1984 if (err)
1985 goto out;
1986
Izik Eidus31dbd012009-09-21 17:02:03 -07001987 ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
1988 if (IS_ERR(ksm_thread)) {
1989 printk(KERN_ERR "ksm: creating kthread failed\n");
1990 err = PTR_ERR(ksm_thread);
Lai Jiangshand9f89842010-08-09 17:20:02 -07001991 goto out_free;
Izik Eidus31dbd012009-09-21 17:02:03 -07001992 }
1993
Hugh Dickins2ffd8672009-09-21 17:02:23 -07001994#ifdef CONFIG_SYSFS
Izik Eidus31dbd012009-09-21 17:02:03 -07001995 err = sysfs_create_group(mm_kobj, &ksm_attr_group);
1996 if (err) {
1997 printk(KERN_ERR "ksm: register sysfs failed\n");
Hugh Dickins2ffd8672009-09-21 17:02:23 -07001998 kthread_stop(ksm_thread);
Lai Jiangshand9f89842010-08-09 17:20:02 -07001999 goto out_free;
Izik Eidus31dbd012009-09-21 17:02:03 -07002000 }
Hugh Dickinsc73602a2009-10-07 16:32:22 -07002001#else
2002 ksm_run = KSM_RUN_MERGE; /* no way for user to start it */
2003
Hugh Dickins2ffd8672009-09-21 17:02:23 -07002004#endif /* CONFIG_SYSFS */
Izik Eidus31dbd012009-09-21 17:02:03 -07002005
Hugh Dickins62b61f62009-12-14 17:59:33 -08002006#ifdef CONFIG_MEMORY_HOTREMOVE
2007 /*
2008 * Choose a high priority since the callback takes ksm_thread_mutex:
2009 * later callbacks could only be taking locks which nest within that.
2010 */
2011 hotplug_memory_notifier(ksm_memory_callback, 100);
2012#endif
Izik Eidus31dbd012009-09-21 17:02:03 -07002013 return 0;
2014
Lai Jiangshand9f89842010-08-09 17:20:02 -07002015out_free:
Izik Eidus31dbd012009-09-21 17:02:03 -07002016 ksm_slab_free();
2017out:
2018 return err;
2019}
2020module_init(ksm_init)