blob: 1925ffbfb27f00ac3d3d262ce0ea1c8aaa3a117f [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>
KAMEZAWA Hiroyuki4e5f01c2012-01-12 17:18:58 -080031#include <linux/memcontrol.h>
Izik Eidus31dbd012009-09-21 17:02:03 -070032#include <linux/rbtree.h>
Hugh Dickins62b61f62009-12-14 17:59:33 -080033#include <linux/memory.h>
Izik Eidus31dbd012009-09-21 17:02:03 -070034#include <linux/mmu_notifier.h>
Izik Eidus2c6854f2009-09-23 15:56:04 -070035#include <linux/swap.h>
Hugh Dickinsf8af4da2009-09-21 17:01:57 -070036#include <linux/ksm.h>
Lai Jiangshand9f89842010-08-09 17:20:02 -070037#include <linux/hash.h>
Andrea Arcangeli878aee72011-01-13 15:47:10 -080038#include <linux/freezer.h>
David Rientjes72788c32011-05-24 17:11:40 -070039#include <linux/oom.h>
Hugh Dickinsf8af4da2009-09-21 17:01:57 -070040
Izik Eidus31dbd012009-09-21 17:02:03 -070041#include <asm/tlbflush.h>
Hugh Dickins73848b42009-12-14 17:59:22 -080042#include "internal.h"
Izik Eidus31dbd012009-09-21 17:02:03 -070043
44/*
45 * A few notes about the KSM scanning process,
46 * to make it easier to understand the data structures below:
47 *
48 * In order to reduce excessive scanning, KSM sorts the memory pages by their
49 * contents into a data structure that holds pointers to the pages' locations.
50 *
51 * Since the contents of the pages may change at any moment, KSM cannot just
52 * insert the pages into a normal sorted tree and expect it to find anything.
53 * Therefore KSM uses two data structures - the stable and the unstable tree.
54 *
55 * The stable tree holds pointers to all the merged pages (ksm pages), sorted
56 * by their contents. Because each such page is write-protected, searching on
57 * this tree is fully assured to be working (except when pages are unmapped),
58 * and therefore this tree is called the stable tree.
59 *
60 * In addition to the stable tree, KSM uses a second data structure called the
61 * unstable tree: this tree holds pointers to pages which have been found to
62 * be "unchanged for a period of time". The unstable tree sorts these pages
63 * by their contents, but since they are not write-protected, KSM cannot rely
64 * upon the unstable tree to work correctly - the unstable tree is liable to
65 * be corrupted as its contents are modified, and so it is called unstable.
66 *
67 * KSM solves this problem by several techniques:
68 *
69 * 1) The unstable tree is flushed every time KSM completes scanning all
70 * memory areas, and then the tree is rebuilt again from the beginning.
71 * 2) KSM will only insert into the unstable tree, pages whose hash value
72 * has not changed since the previous scan of all memory areas.
73 * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
74 * colors of the nodes and not on their contents, assuring that even when
75 * the tree gets "corrupted" it won't get out of balance, so scanning time
76 * remains the same (also, searching and inserting nodes in an rbtree uses
77 * the same algorithm, so we have no overhead when we flush and rebuild).
78 * 4) KSM never flushes the stable tree, which means that even if it were to
79 * take 10 attempts to find a page in the unstable tree, once it is found,
80 * it is secured in the stable tree. (When we scan a new page, we first
81 * compare it against the stable tree, and then against the unstable tree.)
82 */
83
84/**
85 * struct mm_slot - ksm information per mm that is being scanned
86 * @link: link to the mm_slots hash list
87 * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
Hugh Dickins6514d512009-12-14 17:59:19 -080088 * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
Izik Eidus31dbd012009-09-21 17:02:03 -070089 * @mm: the mm that this information is valid for
90 */
91struct mm_slot {
92 struct hlist_node link;
93 struct list_head mm_list;
Hugh Dickins6514d512009-12-14 17:59:19 -080094 struct rmap_item *rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -070095 struct mm_struct *mm;
96};
97
98/**
99 * struct ksm_scan - cursor for scanning
100 * @mm_slot: the current mm_slot we are scanning
101 * @address: the next address inside that to be scanned
Hugh Dickins6514d512009-12-14 17:59:19 -0800102 * @rmap_list: link to the next rmap to be scanned in the rmap_list
Izik Eidus31dbd012009-09-21 17:02:03 -0700103 * @seqnr: count of completed full scans (needed when removing unstable node)
104 *
105 * There is only the one ksm_scan instance of this cursor structure.
106 */
107struct ksm_scan {
108 struct mm_slot *mm_slot;
109 unsigned long address;
Hugh Dickins6514d512009-12-14 17:59:19 -0800110 struct rmap_item **rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -0700111 unsigned long seqnr;
112};
113
114/**
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800115 * struct stable_node - node of the stable rbtree
116 * @node: rb node of this ksm page in the stable tree
117 * @hlist: hlist head of rmap_items using this ksm page
Hugh Dickins62b61f62009-12-14 17:59:33 -0800118 * @kpfn: page frame number of this ksm page
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800119 */
120struct stable_node {
121 struct rb_node node;
122 struct hlist_head hlist;
Hugh Dickins62b61f62009-12-14 17:59:33 -0800123 unsigned long kpfn;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800124};
125
126/**
Izik Eidus31dbd012009-09-21 17:02:03 -0700127 * struct rmap_item - reverse mapping item for virtual addresses
Hugh Dickins6514d512009-12-14 17:59:19 -0800128 * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
Hugh Dickinsdb114b82009-12-14 17:59:25 -0800129 * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
Izik Eidus31dbd012009-09-21 17:02:03 -0700130 * @mm: the memory structure this rmap_item is pointing into
131 * @address: the virtual address this rmap_item tracks (+ flags in low bits)
132 * @oldchecksum: previous checksum of the page at that virtual address
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800133 * @node: rb node of this rmap_item in the unstable tree
134 * @head: pointer to stable_node heading this list in the stable tree
135 * @hlist: link into hlist of rmap_items hanging off that stable_node
Izik Eidus31dbd012009-09-21 17:02:03 -0700136 */
137struct rmap_item {
Hugh Dickins6514d512009-12-14 17:59:19 -0800138 struct rmap_item *rmap_list;
Hugh Dickinsdb114b82009-12-14 17:59:25 -0800139 struct anon_vma *anon_vma; /* when stable */
Izik Eidus31dbd012009-09-21 17:02:03 -0700140 struct mm_struct *mm;
141 unsigned long address; /* + low bits used for flags below */
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800142 unsigned int oldchecksum; /* when unstable */
Izik Eidus31dbd012009-09-21 17:02:03 -0700143 union {
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800144 struct rb_node node; /* when node of unstable tree */
145 struct { /* when listed from stable tree */
146 struct stable_node *head;
147 struct hlist_node hlist;
148 };
Izik Eidus31dbd012009-09-21 17:02:03 -0700149 };
150};
151
152#define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800153#define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */
154#define STABLE_FLAG 0x200 /* is listed from the stable tree */
Izik Eidus31dbd012009-09-21 17:02:03 -0700155
156/* The stable and unstable tree heads */
157static struct rb_root root_stable_tree = RB_ROOT;
158static struct rb_root root_unstable_tree = RB_ROOT;
159
Lai Jiangshand9f89842010-08-09 17:20:02 -0700160#define MM_SLOTS_HASH_SHIFT 10
161#define MM_SLOTS_HASH_HEADS (1 << MM_SLOTS_HASH_SHIFT)
162static struct hlist_head mm_slots_hash[MM_SLOTS_HASH_HEADS];
Izik Eidus31dbd012009-09-21 17:02:03 -0700163
164static struct mm_slot ksm_mm_head = {
165 .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
166};
167static struct ksm_scan ksm_scan = {
168 .mm_slot = &ksm_mm_head,
169};
170
171static struct kmem_cache *rmap_item_cache;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800172static struct kmem_cache *stable_node_cache;
Izik Eidus31dbd012009-09-21 17:02:03 -0700173static struct kmem_cache *mm_slot_cache;
174
175/* The number of nodes in the stable tree */
Hugh Dickinsb4028262009-09-21 17:02:09 -0700176static unsigned long ksm_pages_shared;
Izik Eidus31dbd012009-09-21 17:02:03 -0700177
Hugh Dickinse178dfd2009-09-21 17:02:10 -0700178/* The number of page slots additionally sharing those nodes */
Hugh Dickinsb4028262009-09-21 17:02:09 -0700179static unsigned long ksm_pages_sharing;
Izik Eidus31dbd012009-09-21 17:02:03 -0700180
Hugh Dickins473b0ce2009-09-21 17:02:11 -0700181/* The number of nodes in the unstable tree */
182static unsigned long ksm_pages_unshared;
183
184/* The number of rmap_items in use: to calculate pages_volatile */
185static unsigned long ksm_rmap_items;
186
Izik Eidus31dbd012009-09-21 17:02:03 -0700187/* Number of pages ksmd should scan in one batch */
Izik Eidus2c6854f2009-09-23 15:56:04 -0700188static unsigned int ksm_thread_pages_to_scan = 100;
Izik Eidus31dbd012009-09-21 17:02:03 -0700189
190/* Milliseconds ksmd should sleep between batches */
Hugh Dickins2ffd8672009-09-21 17:02:23 -0700191static unsigned int ksm_thread_sleep_millisecs = 20;
Izik Eidus31dbd012009-09-21 17:02:03 -0700192
193#define KSM_RUN_STOP 0
194#define KSM_RUN_MERGE 1
195#define KSM_RUN_UNMERGE 2
Izik Eidus2c6854f2009-09-23 15:56:04 -0700196static unsigned int ksm_run = KSM_RUN_STOP;
Izik Eidus31dbd012009-09-21 17:02:03 -0700197
198static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
199static DEFINE_MUTEX(ksm_thread_mutex);
200static DEFINE_SPINLOCK(ksm_mmlist_lock);
201
202#define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
203 sizeof(struct __struct), __alignof__(struct __struct),\
204 (__flags), NULL)
205
206static int __init ksm_slab_init(void)
207{
208 rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
209 if (!rmap_item_cache)
210 goto out;
211
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800212 stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
213 if (!stable_node_cache)
214 goto out_free1;
215
Izik Eidus31dbd012009-09-21 17:02:03 -0700216 mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
217 if (!mm_slot_cache)
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800218 goto out_free2;
Izik Eidus31dbd012009-09-21 17:02:03 -0700219
220 return 0;
221
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800222out_free2:
223 kmem_cache_destroy(stable_node_cache);
224out_free1:
Izik Eidus31dbd012009-09-21 17:02:03 -0700225 kmem_cache_destroy(rmap_item_cache);
226out:
227 return -ENOMEM;
228}
229
230static void __init ksm_slab_free(void)
231{
232 kmem_cache_destroy(mm_slot_cache);
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800233 kmem_cache_destroy(stable_node_cache);
Izik Eidus31dbd012009-09-21 17:02:03 -0700234 kmem_cache_destroy(rmap_item_cache);
235 mm_slot_cache = NULL;
236}
237
238static inline struct rmap_item *alloc_rmap_item(void)
239{
Hugh Dickins473b0ce2009-09-21 17:02:11 -0700240 struct rmap_item *rmap_item;
241
242 rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL);
243 if (rmap_item)
244 ksm_rmap_items++;
245 return rmap_item;
Izik Eidus31dbd012009-09-21 17:02:03 -0700246}
247
248static inline void free_rmap_item(struct rmap_item *rmap_item)
249{
Hugh Dickins473b0ce2009-09-21 17:02:11 -0700250 ksm_rmap_items--;
Izik Eidus31dbd012009-09-21 17:02:03 -0700251 rmap_item->mm = NULL; /* debug safety */
252 kmem_cache_free(rmap_item_cache, rmap_item);
253}
254
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800255static inline struct stable_node *alloc_stable_node(void)
256{
257 return kmem_cache_alloc(stable_node_cache, GFP_KERNEL);
258}
259
260static inline void free_stable_node(struct stable_node *stable_node)
261{
262 kmem_cache_free(stable_node_cache, stable_node);
263}
264
Izik Eidus31dbd012009-09-21 17:02:03 -0700265static inline struct mm_slot *alloc_mm_slot(void)
266{
267 if (!mm_slot_cache) /* initialization failed */
268 return NULL;
269 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
270}
271
272static inline void free_mm_slot(struct mm_slot *mm_slot)
273{
274 kmem_cache_free(mm_slot_cache, mm_slot);
275}
276
Izik Eidus31dbd012009-09-21 17:02:03 -0700277static struct mm_slot *get_mm_slot(struct mm_struct *mm)
278{
279 struct mm_slot *mm_slot;
280 struct hlist_head *bucket;
281 struct hlist_node *node;
282
Lai Jiangshand9f89842010-08-09 17:20:02 -0700283 bucket = &mm_slots_hash[hash_ptr(mm, MM_SLOTS_HASH_SHIFT)];
Izik Eidus31dbd012009-09-21 17:02:03 -0700284 hlist_for_each_entry(mm_slot, node, bucket, link) {
285 if (mm == mm_slot->mm)
286 return mm_slot;
287 }
288 return NULL;
289}
290
291static void insert_to_mm_slots_hash(struct mm_struct *mm,
292 struct mm_slot *mm_slot)
293{
294 struct hlist_head *bucket;
295
Lai Jiangshand9f89842010-08-09 17:20:02 -0700296 bucket = &mm_slots_hash[hash_ptr(mm, MM_SLOTS_HASH_SHIFT)];
Izik Eidus31dbd012009-09-21 17:02:03 -0700297 mm_slot->mm = mm;
Izik Eidus31dbd012009-09-21 17:02:03 -0700298 hlist_add_head(&mm_slot->link, bucket);
299}
300
301static inline int in_stable_tree(struct rmap_item *rmap_item)
302{
303 return rmap_item->address & STABLE_FLAG;
304}
305
306/*
Hugh Dickinsa913e182009-09-21 17:02:26 -0700307 * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
308 * page tables after it has passed through ksm_exit() - which, if necessary,
309 * takes mmap_sem briefly to serialize against them. ksm_exit() does not set
310 * a special flag: they can just back out as soon as mm_users goes to zero.
311 * ksm_test_exit() is used throughout to make this test for exit: in some
312 * places for correctness, in some places just to avoid unnecessary work.
313 */
314static inline bool ksm_test_exit(struct mm_struct *mm)
315{
316 return atomic_read(&mm->mm_users) == 0;
317}
318
319/*
Izik Eidus31dbd012009-09-21 17:02:03 -0700320 * We use break_ksm to break COW on a ksm page: it's a stripped down
321 *
322 * if (get_user_pages(current, mm, addr, 1, 1, 1, &page, NULL) == 1)
323 * put_page(page);
324 *
325 * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
326 * in case the application has unmapped and remapped mm,addr meanwhile.
327 * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP
328 * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
329 */
Hugh Dickinsd952b792009-09-21 17:02:16 -0700330static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
Izik Eidus31dbd012009-09-21 17:02:03 -0700331{
332 struct page *page;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700333 int ret = 0;
Izik Eidus31dbd012009-09-21 17:02:03 -0700334
335 do {
336 cond_resched();
337 page = follow_page(vma, addr, FOLL_GET);
Dan Carpenter22eccdd2010-04-23 13:18:10 -0400338 if (IS_ERR_OR_NULL(page))
Izik Eidus31dbd012009-09-21 17:02:03 -0700339 break;
340 if (PageKsm(page))
341 ret = handle_mm_fault(vma->vm_mm, vma, addr,
342 FAULT_FLAG_WRITE);
343 else
344 ret = VM_FAULT_WRITE;
345 put_page(page);
Hugh Dickinsd952b792009-09-21 17:02:16 -0700346 } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_OOM)));
347 /*
348 * We must loop because handle_mm_fault() may back out if there's
349 * any difficulty e.g. if pte accessed bit gets updated concurrently.
350 *
351 * VM_FAULT_WRITE is what we have been hoping for: it indicates that
352 * COW has been broken, even if the vma does not permit VM_WRITE;
353 * but note that a concurrent fault might break PageKsm for us.
354 *
355 * VM_FAULT_SIGBUS could occur if we race with truncation of the
356 * backing file, which also invalidates anonymous pages: that's
357 * okay, that truncation will have unmapped the PageKsm for us.
358 *
359 * VM_FAULT_OOM: at the time of writing (late July 2009), setting
360 * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
361 * current task has TIF_MEMDIE set, and will be OOM killed on return
362 * to user; and ksmd, having no mm, would never be chosen for that.
363 *
364 * But if the mm is in a limited mem_cgroup, then the fault may fail
365 * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
366 * even ksmd can fail in this way - though it's usually breaking ksm
367 * just to undo a merge it made a moment before, so unlikely to oom.
368 *
369 * That's a pity: we might therefore have more kernel pages allocated
370 * than we're counting as nodes in the stable tree; but ksm_do_scan
371 * will retry to break_cow on each pass, so should recover the page
372 * in due course. The important thing is to not let VM_MERGEABLE
373 * be cleared while any such pages might remain in the area.
374 */
375 return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
Izik Eidus31dbd012009-09-21 17:02:03 -0700376}
377
Hugh Dickins8dd35572009-12-14 17:59:18 -0800378static void break_cow(struct rmap_item *rmap_item)
Izik Eidus31dbd012009-09-21 17:02:03 -0700379{
Hugh Dickins8dd35572009-12-14 17:59:18 -0800380 struct mm_struct *mm = rmap_item->mm;
381 unsigned long addr = rmap_item->address;
Izik Eidus31dbd012009-09-21 17:02:03 -0700382 struct vm_area_struct *vma;
383
Hugh Dickins4035c072009-12-14 17:59:27 -0800384 /*
385 * It is not an accident that whenever we want to break COW
386 * to undo, we also need to drop a reference to the anon_vma.
387 */
Peter Zijlstra9e601092011-03-22 16:32:46 -0700388 put_anon_vma(rmap_item->anon_vma);
Hugh Dickins4035c072009-12-14 17:59:27 -0800389
Hugh Dickins81464e302009-09-21 17:02:15 -0700390 down_read(&mm->mmap_sem);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700391 if (ksm_test_exit(mm))
392 goto out;
Izik Eidus31dbd012009-09-21 17:02:03 -0700393 vma = find_vma(mm, addr);
394 if (!vma || vma->vm_start > addr)
Hugh Dickins81464e302009-09-21 17:02:15 -0700395 goto out;
Izik Eidus31dbd012009-09-21 17:02:03 -0700396 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
Hugh Dickins81464e302009-09-21 17:02:15 -0700397 goto out;
Izik Eidus31dbd012009-09-21 17:02:03 -0700398 break_ksm(vma, addr);
Hugh Dickins81464e302009-09-21 17:02:15 -0700399out:
Izik Eidus31dbd012009-09-21 17:02:03 -0700400 up_read(&mm->mmap_sem);
401}
402
Andrea Arcangeli29ad7682011-01-13 15:47:19 -0800403static struct page *page_trans_compound_anon(struct page *page)
404{
405 if (PageTransCompound(page)) {
Andrea Arcangeli22e5c472011-01-13 15:47:20 -0800406 struct page *head = compound_trans_head(page);
Andrea Arcangeli29ad7682011-01-13 15:47:19 -0800407 /*
Andrea Arcangeli22e5c472011-01-13 15:47:20 -0800408 * head may actually be splitted and freed from under
409 * us but it's ok here.
Andrea Arcangeli29ad7682011-01-13 15:47:19 -0800410 */
Andrea Arcangeli29ad7682011-01-13 15:47:19 -0800411 if (PageAnon(head))
412 return head;
413 }
414 return NULL;
415}
416
Izik Eidus31dbd012009-09-21 17:02:03 -0700417static struct page *get_mergeable_page(struct rmap_item *rmap_item)
418{
419 struct mm_struct *mm = rmap_item->mm;
420 unsigned long addr = rmap_item->address;
421 struct vm_area_struct *vma;
422 struct page *page;
423
424 down_read(&mm->mmap_sem);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700425 if (ksm_test_exit(mm))
426 goto out;
Izik Eidus31dbd012009-09-21 17:02:03 -0700427 vma = find_vma(mm, addr);
428 if (!vma || vma->vm_start > addr)
429 goto out;
430 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
431 goto out;
432
433 page = follow_page(vma, addr, FOLL_GET);
Dan Carpenter22eccdd2010-04-23 13:18:10 -0400434 if (IS_ERR_OR_NULL(page))
Izik Eidus31dbd012009-09-21 17:02:03 -0700435 goto out;
Andrea Arcangeli29ad7682011-01-13 15:47:19 -0800436 if (PageAnon(page) || page_trans_compound_anon(page)) {
Izik Eidus31dbd012009-09-21 17:02:03 -0700437 flush_anon_page(vma, page, addr);
438 flush_dcache_page(page);
439 } else {
440 put_page(page);
441out: page = NULL;
442 }
443 up_read(&mm->mmap_sem);
444 return page;
445}
446
Hugh Dickins4035c072009-12-14 17:59:27 -0800447static void remove_node_from_stable_tree(struct stable_node *stable_node)
448{
449 struct rmap_item *rmap_item;
450 struct hlist_node *hlist;
451
452 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
453 if (rmap_item->hlist.next)
454 ksm_pages_sharing--;
455 else
456 ksm_pages_shared--;
Peter Zijlstra9e601092011-03-22 16:32:46 -0700457 put_anon_vma(rmap_item->anon_vma);
Hugh Dickins4035c072009-12-14 17:59:27 -0800458 rmap_item->address &= PAGE_MASK;
459 cond_resched();
460 }
461
462 rb_erase(&stable_node->node, &root_stable_tree);
463 free_stable_node(stable_node);
464}
465
466/*
467 * get_ksm_page: checks if the page indicated by the stable node
468 * is still its ksm page, despite having held no reference to it.
469 * In which case we can trust the content of the page, and it
470 * returns the gotten page; but if the page has now been zapped,
471 * remove the stale node from the stable tree and return NULL.
472 *
473 * You would expect the stable_node to hold a reference to the ksm page.
474 * But if it increments the page's count, swapping out has to wait for
475 * ksmd to come around again before it can free the page, which may take
476 * seconds or even minutes: much too unresponsive. So instead we use a
477 * "keyhole reference": access to the ksm page from the stable node peeps
478 * out through its keyhole to see if that page still holds the right key,
479 * pointing back to this stable node. This relies on freeing a PageAnon
480 * page to reset its page->mapping to NULL, and relies on no other use of
481 * a page to put something that might look like our key in page->mapping.
482 *
483 * include/linux/pagemap.h page_cache_get_speculative() is a good reference,
484 * but this is different - made simpler by ksm_thread_mutex being held, but
485 * interesting for assuming that no other use of the struct page could ever
486 * put our expected_mapping into page->mapping (or a field of the union which
487 * coincides with page->mapping). The RCU calls are not for KSM at all, but
488 * to keep the page_count protocol described with page_cache_get_speculative.
489 *
490 * Note: it is possible that get_ksm_page() will return NULL one moment,
491 * then page the next, if the page is in between page_freeze_refs() and
492 * page_unfreeze_refs(): this shouldn't be a problem anywhere, the page
493 * is on its way to being freed; but it is an anomaly to bear in mind.
494 */
495static struct page *get_ksm_page(struct stable_node *stable_node)
496{
497 struct page *page;
498 void *expected_mapping;
499
Hugh Dickins62b61f62009-12-14 17:59:33 -0800500 page = pfn_to_page(stable_node->kpfn);
Hugh Dickins4035c072009-12-14 17:59:27 -0800501 expected_mapping = (void *)stable_node +
502 (PAGE_MAPPING_ANON | PAGE_MAPPING_KSM);
503 rcu_read_lock();
504 if (page->mapping != expected_mapping)
505 goto stale;
506 if (!get_page_unless_zero(page))
507 goto stale;
508 if (page->mapping != expected_mapping) {
509 put_page(page);
510 goto stale;
511 }
512 rcu_read_unlock();
513 return page;
514stale:
515 rcu_read_unlock();
516 remove_node_from_stable_tree(stable_node);
517 return NULL;
518}
519
Izik Eidus31dbd012009-09-21 17:02:03 -0700520/*
Izik Eidus31dbd012009-09-21 17:02:03 -0700521 * Removing rmap_item from stable or unstable tree.
522 * This function will clean the information from the stable/unstable tree.
523 */
524static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
525{
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800526 if (rmap_item->address & STABLE_FLAG) {
527 struct stable_node *stable_node;
Hugh Dickins5ad64682009-12-14 17:59:24 -0800528 struct page *page;
Izik Eidus31dbd012009-09-21 17:02:03 -0700529
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800530 stable_node = rmap_item->head;
Hugh Dickins4035c072009-12-14 17:59:27 -0800531 page = get_ksm_page(stable_node);
532 if (!page)
533 goto out;
534
Hugh Dickins5ad64682009-12-14 17:59:24 -0800535 lock_page(page);
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800536 hlist_del(&rmap_item->hlist);
Hugh Dickins4035c072009-12-14 17:59:27 -0800537 unlock_page(page);
538 put_page(page);
Hugh Dickins08beca42009-12-14 17:59:21 -0800539
Hugh Dickins4035c072009-12-14 17:59:27 -0800540 if (stable_node->hlist.first)
541 ksm_pages_sharing--;
542 else
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800543 ksm_pages_shared--;
Izik Eidus31dbd012009-09-21 17:02:03 -0700544
Peter Zijlstra9e601092011-03-22 16:32:46 -0700545 put_anon_vma(rmap_item->anon_vma);
Hugh Dickins93d17712009-12-14 17:59:16 -0800546 rmap_item->address &= PAGE_MASK;
Izik Eidus31dbd012009-09-21 17:02:03 -0700547
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800548 } else if (rmap_item->address & UNSTABLE_FLAG) {
Izik Eidus31dbd012009-09-21 17:02:03 -0700549 unsigned char age;
550 /*
Hugh Dickins9ba69292009-09-21 17:02:20 -0700551 * Usually ksmd can and must skip the rb_erase, because
Izik Eidus31dbd012009-09-21 17:02:03 -0700552 * root_unstable_tree was already reset to RB_ROOT.
Hugh Dickins9ba69292009-09-21 17:02:20 -0700553 * But be careful when an mm is exiting: do the rb_erase
554 * if this rmap_item was inserted by this scan, rather
555 * than left over from before.
Izik Eidus31dbd012009-09-21 17:02:03 -0700556 */
557 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
Hugh Dickinscd551f92009-09-21 17:02:17 -0700558 BUG_ON(age > 1);
Izik Eidus31dbd012009-09-21 17:02:03 -0700559 if (!age)
560 rb_erase(&rmap_item->node, &root_unstable_tree);
Izik Eidus31dbd012009-09-21 17:02:03 -0700561
Hugh Dickins93d17712009-12-14 17:59:16 -0800562 ksm_pages_unshared--;
563 rmap_item->address &= PAGE_MASK;
564 }
Hugh Dickins4035c072009-12-14 17:59:27 -0800565out:
Izik Eidus31dbd012009-09-21 17:02:03 -0700566 cond_resched(); /* we're called from many long loops */
567}
568
Izik Eidus31dbd012009-09-21 17:02:03 -0700569static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
Hugh Dickins6514d512009-12-14 17:59:19 -0800570 struct rmap_item **rmap_list)
Izik Eidus31dbd012009-09-21 17:02:03 -0700571{
Hugh Dickins6514d512009-12-14 17:59:19 -0800572 while (*rmap_list) {
573 struct rmap_item *rmap_item = *rmap_list;
574 *rmap_list = rmap_item->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -0700575 remove_rmap_item_from_tree(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -0700576 free_rmap_item(rmap_item);
577 }
578}
579
580/*
581 * Though it's very tempting to unmerge in_stable_tree(rmap_item)s rather
582 * than check every pte of a given vma, the locking doesn't quite work for
583 * that - an rmap_item is assigned to the stable tree after inserting ksm
584 * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing
585 * rmap_items from parent to child at fork time (so as not to waste time
586 * if exit comes before the next scan reaches it).
Hugh Dickins81464e302009-09-21 17:02:15 -0700587 *
588 * Similarly, although we'd like to remove rmap_items (so updating counts
589 * and freeing memory) when unmerging an area, it's easier to leave that
590 * to the next pass of ksmd - consider, for example, how ksmd might be
591 * in cmp_and_merge_page on one of the rmap_items we would be removing.
Izik Eidus31dbd012009-09-21 17:02:03 -0700592 */
Hugh Dickinsd952b792009-09-21 17:02:16 -0700593static int unmerge_ksm_pages(struct vm_area_struct *vma,
594 unsigned long start, unsigned long end)
Izik Eidus31dbd012009-09-21 17:02:03 -0700595{
596 unsigned long addr;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700597 int err = 0;
Izik Eidus31dbd012009-09-21 17:02:03 -0700598
Hugh Dickinsd952b792009-09-21 17:02:16 -0700599 for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
Hugh Dickins9ba69292009-09-21 17:02:20 -0700600 if (ksm_test_exit(vma->vm_mm))
601 break;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700602 if (signal_pending(current))
603 err = -ERESTARTSYS;
604 else
605 err = break_ksm(vma, addr);
606 }
607 return err;
Izik Eidus31dbd012009-09-21 17:02:03 -0700608}
609
Hugh Dickins2ffd8672009-09-21 17:02:23 -0700610#ifdef CONFIG_SYSFS
611/*
612 * Only called through the sysfs control interface:
613 */
Hugh Dickinsd952b792009-09-21 17:02:16 -0700614static int unmerge_and_remove_all_rmap_items(void)
Izik Eidus31dbd012009-09-21 17:02:03 -0700615{
616 struct mm_slot *mm_slot;
617 struct mm_struct *mm;
618 struct vm_area_struct *vma;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700619 int err = 0;
Izik Eidus31dbd012009-09-21 17:02:03 -0700620
Hugh Dickinsd952b792009-09-21 17:02:16 -0700621 spin_lock(&ksm_mmlist_lock);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700622 ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
Hugh Dickinsd952b792009-09-21 17:02:16 -0700623 struct mm_slot, mm_list);
624 spin_unlock(&ksm_mmlist_lock);
625
Hugh Dickins9ba69292009-09-21 17:02:20 -0700626 for (mm_slot = ksm_scan.mm_slot;
627 mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
Izik Eidus31dbd012009-09-21 17:02:03 -0700628 mm = mm_slot->mm;
629 down_read(&mm->mmap_sem);
630 for (vma = mm->mmap; vma; vma = vma->vm_next) {
Hugh Dickins9ba69292009-09-21 17:02:20 -0700631 if (ksm_test_exit(mm))
632 break;
Izik Eidus31dbd012009-09-21 17:02:03 -0700633 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
634 continue;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700635 err = unmerge_ksm_pages(vma,
636 vma->vm_start, vma->vm_end);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700637 if (err)
638 goto error;
Izik Eidus31dbd012009-09-21 17:02:03 -0700639 }
Hugh Dickins9ba69292009-09-21 17:02:20 -0700640
Hugh Dickins6514d512009-12-14 17:59:19 -0800641 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
Hugh Dickinsd952b792009-09-21 17:02:16 -0700642
643 spin_lock(&ksm_mmlist_lock);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700644 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
Hugh Dickinsd952b792009-09-21 17:02:16 -0700645 struct mm_slot, mm_list);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700646 if (ksm_test_exit(mm)) {
647 hlist_del(&mm_slot->link);
648 list_del(&mm_slot->mm_list);
649 spin_unlock(&ksm_mmlist_lock);
650
651 free_mm_slot(mm_slot);
652 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
653 up_read(&mm->mmap_sem);
654 mmdrop(mm);
655 } else {
656 spin_unlock(&ksm_mmlist_lock);
657 up_read(&mm->mmap_sem);
658 }
Izik Eidus31dbd012009-09-21 17:02:03 -0700659 }
660
Hugh Dickinsd952b792009-09-21 17:02:16 -0700661 ksm_scan.seqnr = 0;
Hugh Dickins9ba69292009-09-21 17:02:20 -0700662 return 0;
663
664error:
665 up_read(&mm->mmap_sem);
Izik Eidus31dbd012009-09-21 17:02:03 -0700666 spin_lock(&ksm_mmlist_lock);
Hugh Dickinsd952b792009-09-21 17:02:16 -0700667 ksm_scan.mm_slot = &ksm_mm_head;
Izik Eidus31dbd012009-09-21 17:02:03 -0700668 spin_unlock(&ksm_mmlist_lock);
Hugh Dickinsd952b792009-09-21 17:02:16 -0700669 return err;
Izik Eidus31dbd012009-09-21 17:02:03 -0700670}
Hugh Dickins2ffd8672009-09-21 17:02:23 -0700671#endif /* CONFIG_SYSFS */
Izik Eidus31dbd012009-09-21 17:02:03 -0700672
Izik Eidus31dbd012009-09-21 17:02:03 -0700673static u32 calc_checksum(struct page *page)
674{
675 u32 checksum;
676 void *addr = kmap_atomic(page, KM_USER0);
677 checksum = jhash2(addr, PAGE_SIZE / 4, 17);
678 kunmap_atomic(addr, KM_USER0);
679 return checksum;
680}
681
682static int memcmp_pages(struct page *page1, struct page *page2)
683{
684 char *addr1, *addr2;
685 int ret;
686
687 addr1 = kmap_atomic(page1, KM_USER0);
688 addr2 = kmap_atomic(page2, KM_USER1);
689 ret = memcmp(addr1, addr2, PAGE_SIZE);
690 kunmap_atomic(addr2, KM_USER1);
691 kunmap_atomic(addr1, KM_USER0);
692 return ret;
693}
694
695static inline int pages_identical(struct page *page1, struct page *page2)
696{
697 return !memcmp_pages(page1, page2);
698}
699
700static int write_protect_page(struct vm_area_struct *vma, struct page *page,
701 pte_t *orig_pte)
702{
703 struct mm_struct *mm = vma->vm_mm;
704 unsigned long addr;
705 pte_t *ptep;
706 spinlock_t *ptl;
707 int swapped;
708 int err = -EFAULT;
709
710 addr = page_address_in_vma(page, vma);
711 if (addr == -EFAULT)
712 goto out;
713
Andrea Arcangeli29ad7682011-01-13 15:47:19 -0800714 BUG_ON(PageTransCompound(page));
Izik Eidus31dbd012009-09-21 17:02:03 -0700715 ptep = page_check_address(page, mm, addr, &ptl, 0);
716 if (!ptep)
717 goto out;
718
Hugh Dickins4e316352010-10-02 17:49:08 -0700719 if (pte_write(*ptep) || pte_dirty(*ptep)) {
Izik Eidus31dbd012009-09-21 17:02:03 -0700720 pte_t entry;
721
722 swapped = PageSwapCache(page);
723 flush_cache_page(vma, addr, page_to_pfn(page));
724 /*
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300725 * Ok this is tricky, when get_user_pages_fast() run it doesn't
Izik Eidus31dbd012009-09-21 17:02:03 -0700726 * take any lock, therefore the check that we are going to make
727 * with the pagecount against the mapcount is racey and
728 * O_DIRECT can happen right after the check.
729 * So we clear the pte and flush the tlb before the check
730 * this assure us that no O_DIRECT can happen after the check
731 * or in the middle of the check.
732 */
733 entry = ptep_clear_flush(vma, addr, ptep);
734 /*
735 * Check that no O_DIRECT or similar I/O is in progress on the
736 * page
737 */
Hugh Dickins31e855e2009-12-14 17:59:17 -0800738 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
Robin Holtcb532372010-03-23 13:35:26 -0700739 set_pte_at(mm, addr, ptep, entry);
Izik Eidus31dbd012009-09-21 17:02:03 -0700740 goto out_unlock;
741 }
Hugh Dickins4e316352010-10-02 17:49:08 -0700742 if (pte_dirty(entry))
743 set_page_dirty(page);
744 entry = pte_mkclean(pte_wrprotect(entry));
Izik Eidus31dbd012009-09-21 17:02:03 -0700745 set_pte_at_notify(mm, addr, ptep, entry);
746 }
747 *orig_pte = *ptep;
748 err = 0;
749
750out_unlock:
751 pte_unmap_unlock(ptep, ptl);
752out:
753 return err;
754}
755
756/**
757 * replace_page - replace page in vma by new ksm page
Hugh Dickins8dd35572009-12-14 17:59:18 -0800758 * @vma: vma that holds the pte pointing to page
759 * @page: the page we are replacing by kpage
760 * @kpage: the ksm page we replace page by
Izik Eidus31dbd012009-09-21 17:02:03 -0700761 * @orig_pte: the original value of the pte
762 *
763 * Returns 0 on success, -EFAULT on failure.
764 */
Hugh Dickins8dd35572009-12-14 17:59:18 -0800765static int replace_page(struct vm_area_struct *vma, struct page *page,
766 struct page *kpage, pte_t orig_pte)
Izik Eidus31dbd012009-09-21 17:02:03 -0700767{
768 struct mm_struct *mm = vma->vm_mm;
769 pgd_t *pgd;
770 pud_t *pud;
771 pmd_t *pmd;
772 pte_t *ptep;
773 spinlock_t *ptl;
774 unsigned long addr;
Izik Eidus31dbd012009-09-21 17:02:03 -0700775 int err = -EFAULT;
776
Hugh Dickins8dd35572009-12-14 17:59:18 -0800777 addr = page_address_in_vma(page, vma);
Izik Eidus31dbd012009-09-21 17:02:03 -0700778 if (addr == -EFAULT)
779 goto out;
780
781 pgd = pgd_offset(mm, addr);
782 if (!pgd_present(*pgd))
783 goto out;
784
785 pud = pud_offset(pgd, addr);
786 if (!pud_present(*pud))
787 goto out;
788
789 pmd = pmd_offset(pud, addr);
Andrea Arcangeli29ad7682011-01-13 15:47:19 -0800790 BUG_ON(pmd_trans_huge(*pmd));
Izik Eidus31dbd012009-09-21 17:02:03 -0700791 if (!pmd_present(*pmd))
792 goto out;
793
794 ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
795 if (!pte_same(*ptep, orig_pte)) {
796 pte_unmap_unlock(ptep, ptl);
797 goto out;
798 }
799
Hugh Dickins8dd35572009-12-14 17:59:18 -0800800 get_page(kpage);
Hugh Dickins5ad64682009-12-14 17:59:24 -0800801 page_add_anon_rmap(kpage, vma, addr);
Izik Eidus31dbd012009-09-21 17:02:03 -0700802
803 flush_cache_page(vma, addr, pte_pfn(*ptep));
804 ptep_clear_flush(vma, addr, ptep);
Hugh Dickins8dd35572009-12-14 17:59:18 -0800805 set_pte_at_notify(mm, addr, ptep, mk_pte(kpage, vma->vm_page_prot));
Izik Eidus31dbd012009-09-21 17:02:03 -0700806
Hugh Dickins8dd35572009-12-14 17:59:18 -0800807 page_remove_rmap(page);
Hugh Dickinsae52a2a2011-01-13 15:46:28 -0800808 if (!page_mapped(page))
809 try_to_free_swap(page);
Hugh Dickins8dd35572009-12-14 17:59:18 -0800810 put_page(page);
Izik Eidus31dbd012009-09-21 17:02:03 -0700811
812 pte_unmap_unlock(ptep, ptl);
813 err = 0;
814out:
815 return err;
816}
817
Andrea Arcangeli29ad7682011-01-13 15:47:19 -0800818static int page_trans_compound_anon_split(struct page *page)
819{
820 int ret = 0;
821 struct page *transhuge_head = page_trans_compound_anon(page);
822 if (transhuge_head) {
823 /* Get the reference on the head to split it. */
824 if (get_page_unless_zero(transhuge_head)) {
825 /*
826 * Recheck we got the reference while the head
827 * was still anonymous.
828 */
829 if (PageAnon(transhuge_head))
830 ret = split_huge_page(transhuge_head);
831 else
832 /*
833 * Retry later if split_huge_page run
834 * from under us.
835 */
836 ret = 1;
837 put_page(transhuge_head);
838 } else
839 /* Retry later if split_huge_page run from under us. */
840 ret = 1;
841 }
842 return ret;
843}
844
Izik Eidus31dbd012009-09-21 17:02:03 -0700845/*
846 * try_to_merge_one_page - take two pages and merge them into one
Hugh Dickins8dd35572009-12-14 17:59:18 -0800847 * @vma: the vma that holds the pte pointing to page
848 * @page: the PageAnon page that we want to replace with kpage
Hugh Dickins80e14822009-12-14 17:59:29 -0800849 * @kpage: the PageKsm page that we want to map instead of page,
850 * or NULL the first time when we want to use page as kpage.
Izik Eidus31dbd012009-09-21 17:02:03 -0700851 *
852 * This function returns 0 if the pages were merged, -EFAULT otherwise.
853 */
854static int try_to_merge_one_page(struct vm_area_struct *vma,
Hugh Dickins8dd35572009-12-14 17:59:18 -0800855 struct page *page, struct page *kpage)
Izik Eidus31dbd012009-09-21 17:02:03 -0700856{
857 pte_t orig_pte = __pte(0);
858 int err = -EFAULT;
859
Hugh Dickinsdb114b82009-12-14 17:59:25 -0800860 if (page == kpage) /* ksm page forked */
861 return 0;
862
Izik Eidus31dbd012009-09-21 17:02:03 -0700863 if (!(vma->vm_flags & VM_MERGEABLE))
864 goto out;
Andrea Arcangeli29ad7682011-01-13 15:47:19 -0800865 if (PageTransCompound(page) && page_trans_compound_anon_split(page))
866 goto out;
867 BUG_ON(PageTransCompound(page));
Hugh Dickins8dd35572009-12-14 17:59:18 -0800868 if (!PageAnon(page))
Izik Eidus31dbd012009-09-21 17:02:03 -0700869 goto out;
870
Izik Eidus31dbd012009-09-21 17:02:03 -0700871 /*
872 * We need the page lock to read a stable PageSwapCache in
873 * write_protect_page(). We use trylock_page() instead of
874 * lock_page() because we don't want to wait here - we
875 * prefer to continue scanning and merging different pages,
876 * then come back to this page when it is unlocked.
877 */
Hugh Dickins8dd35572009-12-14 17:59:18 -0800878 if (!trylock_page(page))
Hugh Dickins31e855e2009-12-14 17:59:17 -0800879 goto out;
Izik Eidus31dbd012009-09-21 17:02:03 -0700880 /*
881 * If this anonymous page is mapped only here, its pte may need
882 * to be write-protected. If it's mapped elsewhere, all of its
883 * ptes are necessarily already write-protected. But in either
884 * case, we need to lock and check page_count is not raised.
885 */
Hugh Dickins80e14822009-12-14 17:59:29 -0800886 if (write_protect_page(vma, page, &orig_pte) == 0) {
887 if (!kpage) {
888 /*
889 * While we hold page lock, upgrade page from
890 * PageAnon+anon_vma to PageKsm+NULL stable_node:
891 * stable_tree_insert() will update stable_node.
892 */
893 set_page_stable_node(page, NULL);
894 mark_page_accessed(page);
895 err = 0;
896 } else if (pages_identical(page, kpage))
897 err = replace_page(vma, page, kpage, orig_pte);
898 }
Izik Eidus31dbd012009-09-21 17:02:03 -0700899
Hugh Dickins80e14822009-12-14 17:59:29 -0800900 if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
Hugh Dickins73848b42009-12-14 17:59:22 -0800901 munlock_vma_page(page);
Hugh Dickins5ad64682009-12-14 17:59:24 -0800902 if (!PageMlocked(kpage)) {
903 unlock_page(page);
Hugh Dickins5ad64682009-12-14 17:59:24 -0800904 lock_page(kpage);
905 mlock_vma_page(kpage);
906 page = kpage; /* for final unlock */
907 }
908 }
Hugh Dickins73848b42009-12-14 17:59:22 -0800909
Hugh Dickins8dd35572009-12-14 17:59:18 -0800910 unlock_page(page);
Izik Eidus31dbd012009-09-21 17:02:03 -0700911out:
912 return err;
913}
914
915/*
Hugh Dickins81464e302009-09-21 17:02:15 -0700916 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
917 * but no new kernel page is allocated: kpage must already be a ksm page.
Hugh Dickins8dd35572009-12-14 17:59:18 -0800918 *
919 * This function returns 0 if the pages were merged, -EFAULT otherwise.
Hugh Dickins81464e302009-09-21 17:02:15 -0700920 */
Hugh Dickins8dd35572009-12-14 17:59:18 -0800921static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
922 struct page *page, struct page *kpage)
Hugh Dickins81464e302009-09-21 17:02:15 -0700923{
Hugh Dickins8dd35572009-12-14 17:59:18 -0800924 struct mm_struct *mm = rmap_item->mm;
Hugh Dickins81464e302009-09-21 17:02:15 -0700925 struct vm_area_struct *vma;
926 int err = -EFAULT;
927
Hugh Dickins8dd35572009-12-14 17:59:18 -0800928 down_read(&mm->mmap_sem);
929 if (ksm_test_exit(mm))
930 goto out;
931 vma = find_vma(mm, rmap_item->address);
932 if (!vma || vma->vm_start > rmap_item->address)
Hugh Dickins9ba69292009-09-21 17:02:20 -0700933 goto out;
934
Hugh Dickins8dd35572009-12-14 17:59:18 -0800935 err = try_to_merge_one_page(vma, page, kpage);
Hugh Dickinsdb114b82009-12-14 17:59:25 -0800936 if (err)
937 goto out;
938
939 /* Must get reference to anon_vma while still holding mmap_sem */
Peter Zijlstra9e601092011-03-22 16:32:46 -0700940 rmap_item->anon_vma = vma->anon_vma;
941 get_anon_vma(vma->anon_vma);
Hugh Dickins81464e302009-09-21 17:02:15 -0700942out:
Hugh Dickins8dd35572009-12-14 17:59:18 -0800943 up_read(&mm->mmap_sem);
Hugh Dickins81464e302009-09-21 17:02:15 -0700944 return err;
945}
946
947/*
Izik Eidus31dbd012009-09-21 17:02:03 -0700948 * try_to_merge_two_pages - take two identical pages and prepare them
949 * to be merged into one page.
950 *
Hugh Dickins8dd35572009-12-14 17:59:18 -0800951 * This function returns the kpage if we successfully merged two identical
952 * pages into one ksm page, NULL otherwise.
Izik Eidus31dbd012009-09-21 17:02:03 -0700953 *
Hugh Dickins80e14822009-12-14 17:59:29 -0800954 * Note that this function upgrades page to ksm page: if one of the pages
Izik Eidus31dbd012009-09-21 17:02:03 -0700955 * is already a ksm page, try_to_merge_with_ksm_page should be used.
956 */
Hugh Dickins8dd35572009-12-14 17:59:18 -0800957static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
958 struct page *page,
959 struct rmap_item *tree_rmap_item,
960 struct page *tree_page)
Izik Eidus31dbd012009-09-21 17:02:03 -0700961{
Hugh Dickins80e14822009-12-14 17:59:29 -0800962 int err;
Izik Eidus31dbd012009-09-21 17:02:03 -0700963
Hugh Dickins80e14822009-12-14 17:59:29 -0800964 err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
Izik Eidus31dbd012009-09-21 17:02:03 -0700965 if (!err) {
Hugh Dickins8dd35572009-12-14 17:59:18 -0800966 err = try_to_merge_with_ksm_page(tree_rmap_item,
Hugh Dickins80e14822009-12-14 17:59:29 -0800967 tree_page, page);
Izik Eidus31dbd012009-09-21 17:02:03 -0700968 /*
Hugh Dickins81464e302009-09-21 17:02:15 -0700969 * If that fails, we have a ksm page with only one pte
970 * pointing to it: so break it.
Izik Eidus31dbd012009-09-21 17:02:03 -0700971 */
Hugh Dickins4035c072009-12-14 17:59:27 -0800972 if (err)
Hugh Dickins8dd35572009-12-14 17:59:18 -0800973 break_cow(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -0700974 }
Hugh Dickins80e14822009-12-14 17:59:29 -0800975 return err ? NULL : page;
Izik Eidus31dbd012009-09-21 17:02:03 -0700976}
977
978/*
Hugh Dickins8dd35572009-12-14 17:59:18 -0800979 * stable_tree_search - search for page inside the stable tree
Izik Eidus31dbd012009-09-21 17:02:03 -0700980 *
981 * This function checks if there is a page inside the stable tree
982 * with identical content to the page that we are scanning right now.
983 *
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800984 * This function returns the stable tree node of identical content if found,
Izik Eidus31dbd012009-09-21 17:02:03 -0700985 * NULL otherwise.
986 */
Hugh Dickins62b61f62009-12-14 17:59:33 -0800987static struct page *stable_tree_search(struct page *page)
Izik Eidus31dbd012009-09-21 17:02:03 -0700988{
989 struct rb_node *node = root_stable_tree.rb_node;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800990 struct stable_node *stable_node;
Izik Eidus31dbd012009-09-21 17:02:03 -0700991
Hugh Dickins08beca42009-12-14 17:59:21 -0800992 stable_node = page_stable_node(page);
993 if (stable_node) { /* ksm page forked */
994 get_page(page);
Hugh Dickins62b61f62009-12-14 17:59:33 -0800995 return page;
Hugh Dickins08beca42009-12-14 17:59:21 -0800996 }
997
Izik Eidus31dbd012009-09-21 17:02:03 -0700998 while (node) {
Hugh Dickins4035c072009-12-14 17:59:27 -0800999 struct page *tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001000 int ret;
1001
Hugh Dickins08beca42009-12-14 17:59:21 -08001002 cond_resched();
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001003 stable_node = rb_entry(node, struct stable_node, node);
Hugh Dickins4035c072009-12-14 17:59:27 -08001004 tree_page = get_ksm_page(stable_node);
1005 if (!tree_page)
1006 return NULL;
Izik Eidus31dbd012009-09-21 17:02:03 -07001007
Hugh Dickins4035c072009-12-14 17:59:27 -08001008 ret = memcmp_pages(page, tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001009
Hugh Dickins4035c072009-12-14 17:59:27 -08001010 if (ret < 0) {
1011 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001012 node = node->rb_left;
Hugh Dickins4035c072009-12-14 17:59:27 -08001013 } else if (ret > 0) {
1014 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001015 node = node->rb_right;
Hugh Dickins4035c072009-12-14 17:59:27 -08001016 } else
Hugh Dickins62b61f62009-12-14 17:59:33 -08001017 return tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001018 }
1019
1020 return NULL;
1021}
1022
1023/*
1024 * stable_tree_insert - insert rmap_item pointing to new ksm page
1025 * into the stable tree.
1026 *
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001027 * This function returns the stable tree node just allocated on success,
1028 * NULL otherwise.
Izik Eidus31dbd012009-09-21 17:02:03 -07001029 */
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001030static struct stable_node *stable_tree_insert(struct page *kpage)
Izik Eidus31dbd012009-09-21 17:02:03 -07001031{
1032 struct rb_node **new = &root_stable_tree.rb_node;
1033 struct rb_node *parent = NULL;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001034 struct stable_node *stable_node;
Izik Eidus31dbd012009-09-21 17:02:03 -07001035
1036 while (*new) {
Hugh Dickins4035c072009-12-14 17:59:27 -08001037 struct page *tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001038 int ret;
1039
Hugh Dickins08beca42009-12-14 17:59:21 -08001040 cond_resched();
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001041 stable_node = rb_entry(*new, struct stable_node, node);
Hugh Dickins4035c072009-12-14 17:59:27 -08001042 tree_page = get_ksm_page(stable_node);
1043 if (!tree_page)
1044 return NULL;
Izik Eidus31dbd012009-09-21 17:02:03 -07001045
Hugh Dickins4035c072009-12-14 17:59:27 -08001046 ret = memcmp_pages(kpage, tree_page);
1047 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001048
1049 parent = *new;
1050 if (ret < 0)
1051 new = &parent->rb_left;
1052 else if (ret > 0)
1053 new = &parent->rb_right;
1054 else {
1055 /*
1056 * It is not a bug that stable_tree_search() didn't
1057 * find this node: because at that time our page was
1058 * not yet write-protected, so may have changed since.
1059 */
1060 return NULL;
1061 }
1062 }
1063
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001064 stable_node = alloc_stable_node();
1065 if (!stable_node)
1066 return NULL;
Izik Eidus31dbd012009-09-21 17:02:03 -07001067
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001068 rb_link_node(&stable_node->node, parent, new);
1069 rb_insert_color(&stable_node->node, &root_stable_tree);
1070
1071 INIT_HLIST_HEAD(&stable_node->hlist);
1072
Hugh Dickins62b61f62009-12-14 17:59:33 -08001073 stable_node->kpfn = page_to_pfn(kpage);
Hugh Dickins08beca42009-12-14 17:59:21 -08001074 set_page_stable_node(kpage, stable_node);
1075
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001076 return stable_node;
Izik Eidus31dbd012009-09-21 17:02:03 -07001077}
1078
1079/*
Hugh Dickins8dd35572009-12-14 17:59:18 -08001080 * unstable_tree_search_insert - search for identical page,
1081 * else insert rmap_item into the unstable tree.
Izik Eidus31dbd012009-09-21 17:02:03 -07001082 *
1083 * This function searches for a page in the unstable tree identical to the
1084 * page currently being scanned; and if no identical page is found in the
1085 * tree, we insert rmap_item as a new object into the unstable tree.
1086 *
1087 * This function returns pointer to rmap_item found to be identical
1088 * to the currently scanned page, NULL otherwise.
1089 *
1090 * This function does both searching and inserting, because they share
1091 * the same walking algorithm in an rbtree.
1092 */
Hugh Dickins8dd35572009-12-14 17:59:18 -08001093static
1094struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1095 struct page *page,
1096 struct page **tree_pagep)
1097
Izik Eidus31dbd012009-09-21 17:02:03 -07001098{
1099 struct rb_node **new = &root_unstable_tree.rb_node;
1100 struct rb_node *parent = NULL;
1101
1102 while (*new) {
1103 struct rmap_item *tree_rmap_item;
Hugh Dickins8dd35572009-12-14 17:59:18 -08001104 struct page *tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001105 int ret;
1106
Hugh Dickinsd178f272009-11-09 15:58:23 +00001107 cond_resched();
Izik Eidus31dbd012009-09-21 17:02:03 -07001108 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
Hugh Dickins8dd35572009-12-14 17:59:18 -08001109 tree_page = get_mergeable_page(tree_rmap_item);
Dan Carpenter22eccdd2010-04-23 13:18:10 -04001110 if (IS_ERR_OR_NULL(tree_page))
Izik Eidus31dbd012009-09-21 17:02:03 -07001111 return NULL;
1112
1113 /*
Hugh Dickins8dd35572009-12-14 17:59:18 -08001114 * Don't substitute a ksm page for a forked page.
Izik Eidus31dbd012009-09-21 17:02:03 -07001115 */
Hugh Dickins8dd35572009-12-14 17:59:18 -08001116 if (page == tree_page) {
1117 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001118 return NULL;
1119 }
1120
Hugh Dickins8dd35572009-12-14 17:59:18 -08001121 ret = memcmp_pages(page, tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001122
1123 parent = *new;
1124 if (ret < 0) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001125 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001126 new = &parent->rb_left;
1127 } else if (ret > 0) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001128 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001129 new = &parent->rb_right;
1130 } else {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001131 *tree_pagep = tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001132 return tree_rmap_item;
1133 }
1134 }
1135
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001136 rmap_item->address |= UNSTABLE_FLAG;
Izik Eidus31dbd012009-09-21 17:02:03 -07001137 rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1138 rb_link_node(&rmap_item->node, parent, new);
1139 rb_insert_color(&rmap_item->node, &root_unstable_tree);
1140
Hugh Dickins473b0ce2009-09-21 17:02:11 -07001141 ksm_pages_unshared++;
Izik Eidus31dbd012009-09-21 17:02:03 -07001142 return NULL;
1143}
1144
1145/*
1146 * stable_tree_append - add another rmap_item to the linked list of
1147 * rmap_items hanging off a given node of the stable tree, all sharing
1148 * the same ksm page.
1149 */
1150static void stable_tree_append(struct rmap_item *rmap_item,
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001151 struct stable_node *stable_node)
Izik Eidus31dbd012009-09-21 17:02:03 -07001152{
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001153 rmap_item->head = stable_node;
Izik Eidus31dbd012009-09-21 17:02:03 -07001154 rmap_item->address |= STABLE_FLAG;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001155 hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
Hugh Dickinse178dfd2009-09-21 17:02:10 -07001156
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001157 if (rmap_item->hlist.next)
1158 ksm_pages_sharing++;
1159 else
1160 ksm_pages_shared++;
Izik Eidus31dbd012009-09-21 17:02:03 -07001161}
1162
1163/*
Hugh Dickins81464e302009-09-21 17:02:15 -07001164 * cmp_and_merge_page - first see if page can be merged into the stable tree;
1165 * if not, compare checksum to previous and if it's the same, see if page can
1166 * be inserted into the unstable tree, or merged with a page already there and
1167 * both transferred to the stable tree.
Izik Eidus31dbd012009-09-21 17:02:03 -07001168 *
1169 * @page: the page that we are searching identical page to.
1170 * @rmap_item: the reverse mapping into the virtual address of this page
1171 */
1172static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1173{
Izik Eidus31dbd012009-09-21 17:02:03 -07001174 struct rmap_item *tree_rmap_item;
Hugh Dickins8dd35572009-12-14 17:59:18 -08001175 struct page *tree_page = NULL;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001176 struct stable_node *stable_node;
Hugh Dickins8dd35572009-12-14 17:59:18 -08001177 struct page *kpage;
Izik Eidus31dbd012009-09-21 17:02:03 -07001178 unsigned int checksum;
1179 int err;
1180
Hugh Dickins93d17712009-12-14 17:59:16 -08001181 remove_rmap_item_from_tree(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -07001182
1183 /* We first start with searching the page inside the stable tree */
Hugh Dickins62b61f62009-12-14 17:59:33 -08001184 kpage = stable_tree_search(page);
1185 if (kpage) {
Hugh Dickins08beca42009-12-14 17:59:21 -08001186 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
Izik Eidus31dbd012009-09-21 17:02:03 -07001187 if (!err) {
1188 /*
1189 * The page was successfully merged:
1190 * add its rmap_item to the stable tree.
1191 */
Hugh Dickins5ad64682009-12-14 17:59:24 -08001192 lock_page(kpage);
Hugh Dickins62b61f62009-12-14 17:59:33 -08001193 stable_tree_append(rmap_item, page_stable_node(kpage));
Hugh Dickins5ad64682009-12-14 17:59:24 -08001194 unlock_page(kpage);
Izik Eidus31dbd012009-09-21 17:02:03 -07001195 }
Hugh Dickins8dd35572009-12-14 17:59:18 -08001196 put_page(kpage);
Izik Eidus31dbd012009-09-21 17:02:03 -07001197 return;
1198 }
1199
1200 /*
Hugh Dickins4035c072009-12-14 17:59:27 -08001201 * If the hash value of the page has changed from the last time
1202 * we calculated it, this page is changing frequently: therefore we
1203 * don't want to insert it in the unstable tree, and we don't want
1204 * to waste our time searching for something identical to it there.
Izik Eidus31dbd012009-09-21 17:02:03 -07001205 */
1206 checksum = calc_checksum(page);
1207 if (rmap_item->oldchecksum != checksum) {
1208 rmap_item->oldchecksum = checksum;
1209 return;
1210 }
1211
Hugh Dickins8dd35572009-12-14 17:59:18 -08001212 tree_rmap_item =
1213 unstable_tree_search_insert(rmap_item, page, &tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001214 if (tree_rmap_item) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001215 kpage = try_to_merge_two_pages(rmap_item, page,
1216 tree_rmap_item, tree_page);
1217 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001218 /*
1219 * As soon as we merge this page, we want to remove the
1220 * rmap_item of the page we have merged with from the unstable
1221 * tree, and insert it instead as new node in the stable tree.
1222 */
Hugh Dickins8dd35572009-12-14 17:59:18 -08001223 if (kpage) {
Hugh Dickins93d17712009-12-14 17:59:16 -08001224 remove_rmap_item_from_tree(tree_rmap_item);
Hugh Dickins473b0ce2009-09-21 17:02:11 -07001225
Hugh Dickins5ad64682009-12-14 17:59:24 -08001226 lock_page(kpage);
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001227 stable_node = stable_tree_insert(kpage);
1228 if (stable_node) {
1229 stable_tree_append(tree_rmap_item, stable_node);
1230 stable_tree_append(rmap_item, stable_node);
1231 }
Hugh Dickins5ad64682009-12-14 17:59:24 -08001232 unlock_page(kpage);
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001233
Izik Eidus31dbd012009-09-21 17:02:03 -07001234 /*
1235 * If we fail to insert the page into the stable tree,
1236 * we will have 2 virtual addresses that are pointing
1237 * to a ksm page left outside the stable tree,
1238 * in which case we need to break_cow on both.
1239 */
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001240 if (!stable_node) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001241 break_cow(tree_rmap_item);
1242 break_cow(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -07001243 }
1244 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001245 }
1246}
1247
1248static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
Hugh Dickins6514d512009-12-14 17:59:19 -08001249 struct rmap_item **rmap_list,
Izik Eidus31dbd012009-09-21 17:02:03 -07001250 unsigned long addr)
1251{
1252 struct rmap_item *rmap_item;
1253
Hugh Dickins6514d512009-12-14 17:59:19 -08001254 while (*rmap_list) {
1255 rmap_item = *rmap_list;
Hugh Dickins93d17712009-12-14 17:59:16 -08001256 if ((rmap_item->address & PAGE_MASK) == addr)
Izik Eidus31dbd012009-09-21 17:02:03 -07001257 return rmap_item;
Izik Eidus31dbd012009-09-21 17:02:03 -07001258 if (rmap_item->address > addr)
1259 break;
Hugh Dickins6514d512009-12-14 17:59:19 -08001260 *rmap_list = rmap_item->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -07001261 remove_rmap_item_from_tree(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -07001262 free_rmap_item(rmap_item);
1263 }
1264
1265 rmap_item = alloc_rmap_item();
1266 if (rmap_item) {
1267 /* It has already been zeroed */
1268 rmap_item->mm = mm_slot->mm;
1269 rmap_item->address = addr;
Hugh Dickins6514d512009-12-14 17:59:19 -08001270 rmap_item->rmap_list = *rmap_list;
1271 *rmap_list = rmap_item;
Izik Eidus31dbd012009-09-21 17:02:03 -07001272 }
1273 return rmap_item;
1274}
1275
1276static struct rmap_item *scan_get_next_rmap_item(struct page **page)
1277{
1278 struct mm_struct *mm;
1279 struct mm_slot *slot;
1280 struct vm_area_struct *vma;
1281 struct rmap_item *rmap_item;
1282
1283 if (list_empty(&ksm_mm_head.mm_list))
1284 return NULL;
1285
1286 slot = ksm_scan.mm_slot;
1287 if (slot == &ksm_mm_head) {
Hugh Dickins2919bfd2011-01-13 15:47:29 -08001288 /*
1289 * A number of pages can hang around indefinitely on per-cpu
1290 * pagevecs, raised page count preventing write_protect_page
1291 * from merging them. Though it doesn't really matter much,
1292 * it is puzzling to see some stuck in pages_volatile until
1293 * other activity jostles them out, and they also prevented
1294 * LTP's KSM test from succeeding deterministically; so drain
1295 * them here (here rather than on entry to ksm_do_scan(),
1296 * so we don't IPI too often when pages_to_scan is set low).
1297 */
1298 lru_add_drain_all();
1299
Izik Eidus31dbd012009-09-21 17:02:03 -07001300 root_unstable_tree = RB_ROOT;
1301
1302 spin_lock(&ksm_mmlist_lock);
1303 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1304 ksm_scan.mm_slot = slot;
1305 spin_unlock(&ksm_mmlist_lock);
Hugh Dickins2b472612011-06-15 15:08:58 -07001306 /*
1307 * Although we tested list_empty() above, a racing __ksm_exit
1308 * of the last mm on the list may have removed it since then.
1309 */
1310 if (slot == &ksm_mm_head)
1311 return NULL;
Izik Eidus31dbd012009-09-21 17:02:03 -07001312next_mm:
1313 ksm_scan.address = 0;
Hugh Dickins6514d512009-12-14 17:59:19 -08001314 ksm_scan.rmap_list = &slot->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -07001315 }
1316
1317 mm = slot->mm;
1318 down_read(&mm->mmap_sem);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001319 if (ksm_test_exit(mm))
1320 vma = NULL;
1321 else
1322 vma = find_vma(mm, ksm_scan.address);
1323
1324 for (; vma; vma = vma->vm_next) {
Izik Eidus31dbd012009-09-21 17:02:03 -07001325 if (!(vma->vm_flags & VM_MERGEABLE))
1326 continue;
1327 if (ksm_scan.address < vma->vm_start)
1328 ksm_scan.address = vma->vm_start;
1329 if (!vma->anon_vma)
1330 ksm_scan.address = vma->vm_end;
1331
1332 while (ksm_scan.address < vma->vm_end) {
Hugh Dickins9ba69292009-09-21 17:02:20 -07001333 if (ksm_test_exit(mm))
1334 break;
Izik Eidus31dbd012009-09-21 17:02:03 -07001335 *page = follow_page(vma, ksm_scan.address, FOLL_GET);
Andrea Arcangeli21ae5b02011-01-13 15:47:00 -08001336 if (IS_ERR_OR_NULL(*page)) {
1337 ksm_scan.address += PAGE_SIZE;
1338 cond_resched();
1339 continue;
1340 }
Andrea Arcangeli29ad7682011-01-13 15:47:19 -08001341 if (PageAnon(*page) ||
1342 page_trans_compound_anon(*page)) {
Izik Eidus31dbd012009-09-21 17:02:03 -07001343 flush_anon_page(vma, *page, ksm_scan.address);
1344 flush_dcache_page(*page);
1345 rmap_item = get_next_rmap_item(slot,
Hugh Dickins6514d512009-12-14 17:59:19 -08001346 ksm_scan.rmap_list, ksm_scan.address);
Izik Eidus31dbd012009-09-21 17:02:03 -07001347 if (rmap_item) {
Hugh Dickins6514d512009-12-14 17:59:19 -08001348 ksm_scan.rmap_list =
1349 &rmap_item->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -07001350 ksm_scan.address += PAGE_SIZE;
1351 } else
1352 put_page(*page);
1353 up_read(&mm->mmap_sem);
1354 return rmap_item;
1355 }
Andrea Arcangeli21ae5b02011-01-13 15:47:00 -08001356 put_page(*page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001357 ksm_scan.address += PAGE_SIZE;
1358 cond_resched();
1359 }
1360 }
1361
Hugh Dickins9ba69292009-09-21 17:02:20 -07001362 if (ksm_test_exit(mm)) {
1363 ksm_scan.address = 0;
Hugh Dickins6514d512009-12-14 17:59:19 -08001364 ksm_scan.rmap_list = &slot->rmap_list;
Hugh Dickins9ba69292009-09-21 17:02:20 -07001365 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001366 /*
1367 * Nuke all the rmap_items that are above this current rmap:
1368 * because there were no VM_MERGEABLE vmas with such addresses.
1369 */
Hugh Dickins6514d512009-12-14 17:59:19 -08001370 remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
Izik Eidus31dbd012009-09-21 17:02:03 -07001371
1372 spin_lock(&ksm_mmlist_lock);
Hugh Dickinscd551f92009-09-21 17:02:17 -07001373 ksm_scan.mm_slot = list_entry(slot->mm_list.next,
1374 struct mm_slot, mm_list);
1375 if (ksm_scan.address == 0) {
1376 /*
1377 * We've completed a full scan of all vmas, holding mmap_sem
1378 * throughout, and found no VM_MERGEABLE: so do the same as
1379 * __ksm_exit does to remove this mm from all our lists now.
Hugh Dickins9ba69292009-09-21 17:02:20 -07001380 * This applies either when cleaning up after __ksm_exit
1381 * (but beware: we can reach here even before __ksm_exit),
1382 * or when all VM_MERGEABLE areas have been unmapped (and
1383 * mmap_sem then protects against race with MADV_MERGEABLE).
Hugh Dickinscd551f92009-09-21 17:02:17 -07001384 */
1385 hlist_del(&slot->link);
1386 list_del(&slot->mm_list);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001387 spin_unlock(&ksm_mmlist_lock);
1388
Hugh Dickinscd551f92009-09-21 17:02:17 -07001389 free_mm_slot(slot);
1390 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001391 up_read(&mm->mmap_sem);
1392 mmdrop(mm);
1393 } else {
1394 spin_unlock(&ksm_mmlist_lock);
1395 up_read(&mm->mmap_sem);
Hugh Dickinscd551f92009-09-21 17:02:17 -07001396 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001397
1398 /* Repeat until we've completed scanning the whole list */
Hugh Dickinscd551f92009-09-21 17:02:17 -07001399 slot = ksm_scan.mm_slot;
Izik Eidus31dbd012009-09-21 17:02:03 -07001400 if (slot != &ksm_mm_head)
1401 goto next_mm;
1402
Izik Eidus31dbd012009-09-21 17:02:03 -07001403 ksm_scan.seqnr++;
1404 return NULL;
1405}
1406
1407/**
1408 * ksm_do_scan - the ksm scanner main worker function.
1409 * @scan_npages - number of pages we want to scan before we return.
1410 */
1411static void ksm_do_scan(unsigned int scan_npages)
1412{
1413 struct rmap_item *rmap_item;
Dan Carpenter22eccdd2010-04-23 13:18:10 -04001414 struct page *uninitialized_var(page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001415
Andrea Arcangeli878aee72011-01-13 15:47:10 -08001416 while (scan_npages-- && likely(!freezing(current))) {
Izik Eidus31dbd012009-09-21 17:02:03 -07001417 cond_resched();
1418 rmap_item = scan_get_next_rmap_item(&page);
1419 if (!rmap_item)
1420 return;
1421 if (!PageKsm(page) || !in_stable_tree(rmap_item))
1422 cmp_and_merge_page(page, rmap_item);
1423 put_page(page);
1424 }
1425}
1426
Hugh Dickins6e158382009-09-21 17:02:14 -07001427static int ksmd_should_run(void)
1428{
1429 return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
1430}
1431
Izik Eidus31dbd012009-09-21 17:02:03 -07001432static int ksm_scan_thread(void *nothing)
1433{
Andrea Arcangeli878aee72011-01-13 15:47:10 -08001434 set_freezable();
Izik Eidus339aa622009-09-21 17:02:07 -07001435 set_user_nice(current, 5);
Izik Eidus31dbd012009-09-21 17:02:03 -07001436
1437 while (!kthread_should_stop()) {
Hugh Dickins6e158382009-09-21 17:02:14 -07001438 mutex_lock(&ksm_thread_mutex);
1439 if (ksmd_should_run())
Izik Eidus31dbd012009-09-21 17:02:03 -07001440 ksm_do_scan(ksm_thread_pages_to_scan);
Hugh Dickins6e158382009-09-21 17:02:14 -07001441 mutex_unlock(&ksm_thread_mutex);
1442
Andrea Arcangeli878aee72011-01-13 15:47:10 -08001443 try_to_freeze();
1444
Hugh Dickins6e158382009-09-21 17:02:14 -07001445 if (ksmd_should_run()) {
Izik Eidus31dbd012009-09-21 17:02:03 -07001446 schedule_timeout_interruptible(
1447 msecs_to_jiffies(ksm_thread_sleep_millisecs));
1448 } else {
Andrea Arcangeli878aee72011-01-13 15:47:10 -08001449 wait_event_freezable(ksm_thread_wait,
Hugh Dickins6e158382009-09-21 17:02:14 -07001450 ksmd_should_run() || kthread_should_stop());
Izik Eidus31dbd012009-09-21 17:02:03 -07001451 }
1452 }
1453 return 0;
1454}
1455
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001456int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1457 unsigned long end, int advice, unsigned long *vm_flags)
1458{
1459 struct mm_struct *mm = vma->vm_mm;
Hugh Dickinsd952b792009-09-21 17:02:16 -07001460 int err;
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001461
1462 switch (advice) {
1463 case MADV_MERGEABLE:
1464 /*
1465 * Be somewhat over-protective for now!
1466 */
1467 if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE |
1468 VM_PFNMAP | VM_IO | VM_DONTEXPAND |
1469 VM_RESERVED | VM_HUGETLB | VM_INSERTPAGE |
Hugh Dickins5ad64682009-12-14 17:59:24 -08001470 VM_NONLINEAR | VM_MIXEDMAP | VM_SAO))
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001471 return 0; /* just ignore the advice */
1472
Hugh Dickinsd952b792009-09-21 17:02:16 -07001473 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
1474 err = __ksm_enter(mm);
1475 if (err)
1476 return err;
1477 }
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001478
1479 *vm_flags |= VM_MERGEABLE;
1480 break;
1481
1482 case MADV_UNMERGEABLE:
1483 if (!(*vm_flags & VM_MERGEABLE))
1484 return 0; /* just ignore the advice */
1485
Hugh Dickinsd952b792009-09-21 17:02:16 -07001486 if (vma->anon_vma) {
1487 err = unmerge_ksm_pages(vma, start, end);
1488 if (err)
1489 return err;
1490 }
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001491
1492 *vm_flags &= ~VM_MERGEABLE;
1493 break;
1494 }
1495
1496 return 0;
1497}
1498
1499int __ksm_enter(struct mm_struct *mm)
1500{
Hugh Dickins6e158382009-09-21 17:02:14 -07001501 struct mm_slot *mm_slot;
1502 int needs_wakeup;
1503
1504 mm_slot = alloc_mm_slot();
Izik Eidus31dbd012009-09-21 17:02:03 -07001505 if (!mm_slot)
1506 return -ENOMEM;
1507
Hugh Dickins6e158382009-09-21 17:02:14 -07001508 /* Check ksm_run too? Would need tighter locking */
1509 needs_wakeup = list_empty(&ksm_mm_head.mm_list);
1510
Izik Eidus31dbd012009-09-21 17:02:03 -07001511 spin_lock(&ksm_mmlist_lock);
1512 insert_to_mm_slots_hash(mm, mm_slot);
1513 /*
1514 * Insert just behind the scanning cursor, to let the area settle
1515 * down a little; when fork is followed by immediate exec, we don't
1516 * want ksmd to waste time setting up and tearing down an rmap_list.
1517 */
1518 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
1519 spin_unlock(&ksm_mmlist_lock);
1520
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001521 set_bit(MMF_VM_MERGEABLE, &mm->flags);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001522 atomic_inc(&mm->mm_count);
Hugh Dickins6e158382009-09-21 17:02:14 -07001523
1524 if (needs_wakeup)
1525 wake_up_interruptible(&ksm_thread_wait);
1526
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001527 return 0;
1528}
1529
Andrea Arcangeli1c2fb7a2009-09-21 17:02:22 -07001530void __ksm_exit(struct mm_struct *mm)
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001531{
Hugh Dickinscd551f92009-09-21 17:02:17 -07001532 struct mm_slot *mm_slot;
Hugh Dickins9ba69292009-09-21 17:02:20 -07001533 int easy_to_free = 0;
Hugh Dickinscd551f92009-09-21 17:02:17 -07001534
Izik Eidus31dbd012009-09-21 17:02:03 -07001535 /*
Hugh Dickins9ba69292009-09-21 17:02:20 -07001536 * This process is exiting: if it's straightforward (as is the
1537 * case when ksmd was never running), free mm_slot immediately.
1538 * But if it's at the cursor or has rmap_items linked to it, use
1539 * mmap_sem to synchronize with any break_cows before pagetables
1540 * are freed, and leave the mm_slot on the list for ksmd to free.
1541 * Beware: ksm may already have noticed it exiting and freed the slot.
Izik Eidus31dbd012009-09-21 17:02:03 -07001542 */
Hugh Dickins9ba69292009-09-21 17:02:20 -07001543
Hugh Dickinscd551f92009-09-21 17:02:17 -07001544 spin_lock(&ksm_mmlist_lock);
1545 mm_slot = get_mm_slot(mm);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001546 if (mm_slot && ksm_scan.mm_slot != mm_slot) {
Hugh Dickins6514d512009-12-14 17:59:19 -08001547 if (!mm_slot->rmap_list) {
Hugh Dickins9ba69292009-09-21 17:02:20 -07001548 hlist_del(&mm_slot->link);
1549 list_del(&mm_slot->mm_list);
1550 easy_to_free = 1;
1551 } else {
1552 list_move(&mm_slot->mm_list,
1553 &ksm_scan.mm_slot->mm_list);
1554 }
Hugh Dickinscd551f92009-09-21 17:02:17 -07001555 }
Hugh Dickinscd551f92009-09-21 17:02:17 -07001556 spin_unlock(&ksm_mmlist_lock);
1557
Hugh Dickins9ba69292009-09-21 17:02:20 -07001558 if (easy_to_free) {
1559 free_mm_slot(mm_slot);
1560 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1561 mmdrop(mm);
1562 } else if (mm_slot) {
Hugh Dickins9ba69292009-09-21 17:02:20 -07001563 down_write(&mm->mmap_sem);
1564 up_write(&mm->mmap_sem);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001565 }
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001566}
Izik Eidus31dbd012009-09-21 17:02:03 -07001567
Hugh Dickins5ad64682009-12-14 17:59:24 -08001568struct page *ksm_does_need_to_copy(struct page *page,
1569 struct vm_area_struct *vma, unsigned long address)
1570{
1571 struct page *new_page;
1572
Hugh Dickins5ad64682009-12-14 17:59:24 -08001573 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1574 if (new_page) {
KAMEZAWA Hiroyuki4e5f01c2012-01-12 17:18:58 -08001575 /*
1576 * The memcg-specific accounting when moving
1577 * pages around the LRU lists relies on the
1578 * page's owner (memcg) to be valid. Usually,
1579 * pages are assigned to a new owner before
1580 * being put on the LRU list, but since this
1581 * is not the case here, the stale owner from
1582 * a previous allocation cycle must be reset.
1583 */
1584 mem_cgroup_reset_owner(new_page);
Hugh Dickins5ad64682009-12-14 17:59:24 -08001585 copy_user_highpage(new_page, page, address, vma);
1586
1587 SetPageDirty(new_page);
1588 __SetPageUptodate(new_page);
1589 SetPageSwapBacked(new_page);
1590 __set_page_locked(new_page);
1591
1592 if (page_evictable(new_page, vma))
1593 lru_cache_add_lru(new_page, LRU_ACTIVE_ANON);
1594 else
1595 add_page_to_unevictable_list(new_page);
1596 }
1597
Hugh Dickins5ad64682009-12-14 17:59:24 -08001598 return new_page;
1599}
1600
1601int page_referenced_ksm(struct page *page, struct mem_cgroup *memcg,
1602 unsigned long *vm_flags)
1603{
1604 struct stable_node *stable_node;
1605 struct rmap_item *rmap_item;
1606 struct hlist_node *hlist;
1607 unsigned int mapcount = page_mapcount(page);
1608 int referenced = 0;
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001609 int search_new_forks = 0;
Hugh Dickins5ad64682009-12-14 17:59:24 -08001610
1611 VM_BUG_ON(!PageKsm(page));
1612 VM_BUG_ON(!PageLocked(page));
1613
1614 stable_node = page_stable_node(page);
1615 if (!stable_node)
1616 return 0;
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001617again:
Hugh Dickins5ad64682009-12-14 17:59:24 -08001618 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001619 struct anon_vma *anon_vma = rmap_item->anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -08001620 struct anon_vma_chain *vmac;
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001621 struct vm_area_struct *vma;
Hugh Dickins5ad64682009-12-14 17:59:24 -08001622
Rik van Rielcba48b92010-08-09 17:18:38 -07001623 anon_vma_lock(anon_vma);
Rik van Riel5beb4932010-03-05 13:42:07 -08001624 list_for_each_entry(vmac, &anon_vma->head, same_anon_vma) {
1625 vma = vmac->vma;
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001626 if (rmap_item->address < vma->vm_start ||
1627 rmap_item->address >= vma->vm_end)
1628 continue;
1629 /*
1630 * Initially we examine only the vma which covers this
1631 * rmap_item; but later, if there is still work to do,
1632 * we examine covering vmas in other mms: in case they
1633 * were forked from the original since ksmd passed.
1634 */
1635 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1636 continue;
Hugh Dickins5ad64682009-12-14 17:59:24 -08001637
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001638 if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
1639 continue;
1640
1641 referenced += page_referenced_one(page, vma,
Hugh Dickins5ad64682009-12-14 17:59:24 -08001642 rmap_item->address, &mapcount, vm_flags);
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001643 if (!search_new_forks || !mapcount)
1644 break;
1645 }
Rik van Rielcba48b92010-08-09 17:18:38 -07001646 anon_vma_unlock(anon_vma);
Hugh Dickins5ad64682009-12-14 17:59:24 -08001647 if (!mapcount)
1648 goto out;
1649 }
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001650 if (!search_new_forks++)
1651 goto again;
Hugh Dickins5ad64682009-12-14 17:59:24 -08001652out:
Hugh Dickins5ad64682009-12-14 17:59:24 -08001653 return referenced;
1654}
1655
1656int try_to_unmap_ksm(struct page *page, enum ttu_flags flags)
1657{
1658 struct stable_node *stable_node;
1659 struct hlist_node *hlist;
1660 struct rmap_item *rmap_item;
1661 int ret = SWAP_AGAIN;
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001662 int search_new_forks = 0;
Hugh Dickins5ad64682009-12-14 17:59:24 -08001663
1664 VM_BUG_ON(!PageKsm(page));
1665 VM_BUG_ON(!PageLocked(page));
1666
1667 stable_node = page_stable_node(page);
1668 if (!stable_node)
1669 return SWAP_FAIL;
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001670again:
Hugh Dickins5ad64682009-12-14 17:59:24 -08001671 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001672 struct anon_vma *anon_vma = rmap_item->anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -08001673 struct anon_vma_chain *vmac;
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001674 struct vm_area_struct *vma;
Hugh Dickins5ad64682009-12-14 17:59:24 -08001675
Rik van Rielcba48b92010-08-09 17:18:38 -07001676 anon_vma_lock(anon_vma);
Rik van Riel5beb4932010-03-05 13:42:07 -08001677 list_for_each_entry(vmac, &anon_vma->head, same_anon_vma) {
1678 vma = vmac->vma;
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001679 if (rmap_item->address < vma->vm_start ||
1680 rmap_item->address >= vma->vm_end)
1681 continue;
1682 /*
1683 * Initially we examine only the vma which covers this
1684 * rmap_item; but later, if there is still work to do,
1685 * we examine covering vmas in other mms: in case they
1686 * were forked from the original since ksmd passed.
1687 */
1688 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1689 continue;
1690
1691 ret = try_to_unmap_one(page, vma,
1692 rmap_item->address, flags);
1693 if (ret != SWAP_AGAIN || !page_mapped(page)) {
Rik van Rielcba48b92010-08-09 17:18:38 -07001694 anon_vma_unlock(anon_vma);
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001695 goto out;
1696 }
1697 }
Rik van Rielcba48b92010-08-09 17:18:38 -07001698 anon_vma_unlock(anon_vma);
Hugh Dickins5ad64682009-12-14 17:59:24 -08001699 }
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001700 if (!search_new_forks++)
1701 goto again;
Hugh Dickins5ad64682009-12-14 17:59:24 -08001702out:
Hugh Dickins5ad64682009-12-14 17:59:24 -08001703 return ret;
1704}
1705
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001706#ifdef CONFIG_MIGRATION
1707int rmap_walk_ksm(struct page *page, int (*rmap_one)(struct page *,
1708 struct vm_area_struct *, unsigned long, void *), void *arg)
1709{
1710 struct stable_node *stable_node;
1711 struct hlist_node *hlist;
1712 struct rmap_item *rmap_item;
1713 int ret = SWAP_AGAIN;
1714 int search_new_forks = 0;
1715
1716 VM_BUG_ON(!PageKsm(page));
1717 VM_BUG_ON(!PageLocked(page));
1718
1719 stable_node = page_stable_node(page);
1720 if (!stable_node)
1721 return ret;
1722again:
1723 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1724 struct anon_vma *anon_vma = rmap_item->anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -08001725 struct anon_vma_chain *vmac;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001726 struct vm_area_struct *vma;
1727
Rik van Rielcba48b92010-08-09 17:18:38 -07001728 anon_vma_lock(anon_vma);
Rik van Riel5beb4932010-03-05 13:42:07 -08001729 list_for_each_entry(vmac, &anon_vma->head, same_anon_vma) {
1730 vma = vmac->vma;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001731 if (rmap_item->address < vma->vm_start ||
1732 rmap_item->address >= vma->vm_end)
1733 continue;
1734 /*
1735 * Initially we examine only the vma which covers this
1736 * rmap_item; but later, if there is still work to do,
1737 * we examine covering vmas in other mms: in case they
1738 * were forked from the original since ksmd passed.
1739 */
1740 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1741 continue;
1742
1743 ret = rmap_one(page, vma, rmap_item->address, arg);
1744 if (ret != SWAP_AGAIN) {
Rik van Rielcba48b92010-08-09 17:18:38 -07001745 anon_vma_unlock(anon_vma);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001746 goto out;
1747 }
1748 }
Rik van Rielcba48b92010-08-09 17:18:38 -07001749 anon_vma_unlock(anon_vma);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001750 }
1751 if (!search_new_forks++)
1752 goto again;
1753out:
1754 return ret;
1755}
1756
1757void ksm_migrate_page(struct page *newpage, struct page *oldpage)
1758{
1759 struct stable_node *stable_node;
1760
1761 VM_BUG_ON(!PageLocked(oldpage));
1762 VM_BUG_ON(!PageLocked(newpage));
1763 VM_BUG_ON(newpage->mapping != oldpage->mapping);
1764
1765 stable_node = page_stable_node(newpage);
1766 if (stable_node) {
Hugh Dickins62b61f62009-12-14 17:59:33 -08001767 VM_BUG_ON(stable_node->kpfn != page_to_pfn(oldpage));
1768 stable_node->kpfn = page_to_pfn(newpage);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001769 }
1770}
1771#endif /* CONFIG_MIGRATION */
1772
Hugh Dickins62b61f62009-12-14 17:59:33 -08001773#ifdef CONFIG_MEMORY_HOTREMOVE
1774static struct stable_node *ksm_check_stable_tree(unsigned long start_pfn,
1775 unsigned long end_pfn)
1776{
1777 struct rb_node *node;
1778
1779 for (node = rb_first(&root_stable_tree); node; node = rb_next(node)) {
1780 struct stable_node *stable_node;
1781
1782 stable_node = rb_entry(node, struct stable_node, node);
1783 if (stable_node->kpfn >= start_pfn &&
1784 stable_node->kpfn < end_pfn)
1785 return stable_node;
1786 }
1787 return NULL;
1788}
1789
1790static int ksm_memory_callback(struct notifier_block *self,
1791 unsigned long action, void *arg)
1792{
1793 struct memory_notify *mn = arg;
1794 struct stable_node *stable_node;
1795
1796 switch (action) {
1797 case MEM_GOING_OFFLINE:
1798 /*
1799 * Keep it very simple for now: just lock out ksmd and
1800 * MADV_UNMERGEABLE while any memory is going offline.
KOSAKI Motohiroa0b0f582010-12-02 14:31:20 -08001801 * mutex_lock_nested() is necessary because lockdep was alarmed
1802 * that here we take ksm_thread_mutex inside notifier chain
1803 * mutex, and later take notifier chain mutex inside
1804 * ksm_thread_mutex to unlock it. But that's safe because both
1805 * are inside mem_hotplug_mutex.
Hugh Dickins62b61f62009-12-14 17:59:33 -08001806 */
KOSAKI Motohiroa0b0f582010-12-02 14:31:20 -08001807 mutex_lock_nested(&ksm_thread_mutex, SINGLE_DEPTH_NESTING);
Hugh Dickins62b61f62009-12-14 17:59:33 -08001808 break;
1809
1810 case MEM_OFFLINE:
1811 /*
1812 * Most of the work is done by page migration; but there might
1813 * be a few stable_nodes left over, still pointing to struct
1814 * pages which have been offlined: prune those from the tree.
1815 */
1816 while ((stable_node = ksm_check_stable_tree(mn->start_pfn,
1817 mn->start_pfn + mn->nr_pages)) != NULL)
1818 remove_node_from_stable_tree(stable_node);
1819 /* fallthrough */
1820
1821 case MEM_CANCEL_OFFLINE:
1822 mutex_unlock(&ksm_thread_mutex);
1823 break;
1824 }
1825 return NOTIFY_OK;
1826}
1827#endif /* CONFIG_MEMORY_HOTREMOVE */
1828
Hugh Dickins2ffd8672009-09-21 17:02:23 -07001829#ifdef CONFIG_SYSFS
1830/*
1831 * This all compiles without CONFIG_SYSFS, but is a waste of space.
1832 */
1833
Izik Eidus31dbd012009-09-21 17:02:03 -07001834#define KSM_ATTR_RO(_name) \
1835 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1836#define KSM_ATTR(_name) \
1837 static struct kobj_attribute _name##_attr = \
1838 __ATTR(_name, 0644, _name##_show, _name##_store)
1839
1840static ssize_t sleep_millisecs_show(struct kobject *kobj,
1841 struct kobj_attribute *attr, char *buf)
1842{
1843 return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
1844}
1845
1846static ssize_t sleep_millisecs_store(struct kobject *kobj,
1847 struct kobj_attribute *attr,
1848 const char *buf, size_t count)
1849{
1850 unsigned long msecs;
1851 int err;
1852
1853 err = strict_strtoul(buf, 10, &msecs);
1854 if (err || msecs > UINT_MAX)
1855 return -EINVAL;
1856
1857 ksm_thread_sleep_millisecs = msecs;
1858
1859 return count;
1860}
1861KSM_ATTR(sleep_millisecs);
1862
1863static ssize_t pages_to_scan_show(struct kobject *kobj,
1864 struct kobj_attribute *attr, char *buf)
1865{
1866 return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
1867}
1868
1869static ssize_t pages_to_scan_store(struct kobject *kobj,
1870 struct kobj_attribute *attr,
1871 const char *buf, size_t count)
1872{
1873 int err;
1874 unsigned long nr_pages;
1875
1876 err = strict_strtoul(buf, 10, &nr_pages);
1877 if (err || nr_pages > UINT_MAX)
1878 return -EINVAL;
1879
1880 ksm_thread_pages_to_scan = nr_pages;
1881
1882 return count;
1883}
1884KSM_ATTR(pages_to_scan);
1885
1886static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
1887 char *buf)
1888{
1889 return sprintf(buf, "%u\n", ksm_run);
1890}
1891
1892static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
1893 const char *buf, size_t count)
1894{
1895 int err;
1896 unsigned long flags;
1897
1898 err = strict_strtoul(buf, 10, &flags);
1899 if (err || flags > UINT_MAX)
1900 return -EINVAL;
1901 if (flags > KSM_RUN_UNMERGE)
1902 return -EINVAL;
1903
1904 /*
1905 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
1906 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
Hugh Dickinsd0f209f2009-12-14 17:59:34 -08001907 * breaking COW to free the pages_shared (but leaves mm_slots
1908 * on the list for when ksmd may be set running again).
Izik Eidus31dbd012009-09-21 17:02:03 -07001909 */
1910
1911 mutex_lock(&ksm_thread_mutex);
1912 if (ksm_run != flags) {
1913 ksm_run = flags;
Hugh Dickinsd952b792009-09-21 17:02:16 -07001914 if (flags & KSM_RUN_UNMERGE) {
David Rientjes72788c32011-05-24 17:11:40 -07001915 int oom_score_adj;
1916
1917 oom_score_adj = test_set_oom_score_adj(OOM_SCORE_ADJ_MAX);
Hugh Dickinsd952b792009-09-21 17:02:16 -07001918 err = unmerge_and_remove_all_rmap_items();
David Rientjes43362a42011-10-31 17:07:18 -07001919 compare_swap_oom_score_adj(OOM_SCORE_ADJ_MAX,
1920 oom_score_adj);
Hugh Dickinsd952b792009-09-21 17:02:16 -07001921 if (err) {
1922 ksm_run = KSM_RUN_STOP;
1923 count = err;
1924 }
1925 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001926 }
1927 mutex_unlock(&ksm_thread_mutex);
1928
1929 if (flags & KSM_RUN_MERGE)
1930 wake_up_interruptible(&ksm_thread_wait);
1931
1932 return count;
1933}
1934KSM_ATTR(run);
1935
Hugh Dickinsb4028262009-09-21 17:02:09 -07001936static ssize_t pages_shared_show(struct kobject *kobj,
1937 struct kobj_attribute *attr, char *buf)
1938{
1939 return sprintf(buf, "%lu\n", ksm_pages_shared);
1940}
1941KSM_ATTR_RO(pages_shared);
1942
1943static ssize_t pages_sharing_show(struct kobject *kobj,
1944 struct kobj_attribute *attr, char *buf)
1945{
Hugh Dickinse178dfd2009-09-21 17:02:10 -07001946 return sprintf(buf, "%lu\n", ksm_pages_sharing);
Hugh Dickinsb4028262009-09-21 17:02:09 -07001947}
1948KSM_ATTR_RO(pages_sharing);
1949
Hugh Dickins473b0ce2009-09-21 17:02:11 -07001950static ssize_t pages_unshared_show(struct kobject *kobj,
1951 struct kobj_attribute *attr, char *buf)
1952{
1953 return sprintf(buf, "%lu\n", ksm_pages_unshared);
1954}
1955KSM_ATTR_RO(pages_unshared);
1956
1957static ssize_t pages_volatile_show(struct kobject *kobj,
1958 struct kobj_attribute *attr, char *buf)
1959{
1960 long ksm_pages_volatile;
1961
1962 ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
1963 - ksm_pages_sharing - ksm_pages_unshared;
1964 /*
1965 * It was not worth any locking to calculate that statistic,
1966 * but it might therefore sometimes be negative: conceal that.
1967 */
1968 if (ksm_pages_volatile < 0)
1969 ksm_pages_volatile = 0;
1970 return sprintf(buf, "%ld\n", ksm_pages_volatile);
1971}
1972KSM_ATTR_RO(pages_volatile);
1973
1974static ssize_t full_scans_show(struct kobject *kobj,
1975 struct kobj_attribute *attr, char *buf)
1976{
1977 return sprintf(buf, "%lu\n", ksm_scan.seqnr);
1978}
1979KSM_ATTR_RO(full_scans);
1980
Izik Eidus31dbd012009-09-21 17:02:03 -07001981static struct attribute *ksm_attrs[] = {
1982 &sleep_millisecs_attr.attr,
1983 &pages_to_scan_attr.attr,
1984 &run_attr.attr,
Hugh Dickinsb4028262009-09-21 17:02:09 -07001985 &pages_shared_attr.attr,
1986 &pages_sharing_attr.attr,
Hugh Dickins473b0ce2009-09-21 17:02:11 -07001987 &pages_unshared_attr.attr,
1988 &pages_volatile_attr.attr,
1989 &full_scans_attr.attr,
Izik Eidus31dbd012009-09-21 17:02:03 -07001990 NULL,
1991};
1992
1993static struct attribute_group ksm_attr_group = {
1994 .attrs = ksm_attrs,
1995 .name = "ksm",
1996};
Hugh Dickins2ffd8672009-09-21 17:02:23 -07001997#endif /* CONFIG_SYSFS */
Izik Eidus31dbd012009-09-21 17:02:03 -07001998
1999static int __init ksm_init(void)
2000{
2001 struct task_struct *ksm_thread;
2002 int err;
2003
2004 err = ksm_slab_init();
2005 if (err)
2006 goto out;
2007
Izik Eidus31dbd012009-09-21 17:02:03 -07002008 ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
2009 if (IS_ERR(ksm_thread)) {
2010 printk(KERN_ERR "ksm: creating kthread failed\n");
2011 err = PTR_ERR(ksm_thread);
Lai Jiangshand9f89842010-08-09 17:20:02 -07002012 goto out_free;
Izik Eidus31dbd012009-09-21 17:02:03 -07002013 }
2014
Hugh Dickins2ffd8672009-09-21 17:02:23 -07002015#ifdef CONFIG_SYSFS
Izik Eidus31dbd012009-09-21 17:02:03 -07002016 err = sysfs_create_group(mm_kobj, &ksm_attr_group);
2017 if (err) {
2018 printk(KERN_ERR "ksm: register sysfs failed\n");
Hugh Dickins2ffd8672009-09-21 17:02:23 -07002019 kthread_stop(ksm_thread);
Lai Jiangshand9f89842010-08-09 17:20:02 -07002020 goto out_free;
Izik Eidus31dbd012009-09-21 17:02:03 -07002021 }
Hugh Dickinsc73602a2009-10-07 16:32:22 -07002022#else
2023 ksm_run = KSM_RUN_MERGE; /* no way for user to start it */
2024
Hugh Dickins2ffd8672009-09-21 17:02:23 -07002025#endif /* CONFIG_SYSFS */
Izik Eidus31dbd012009-09-21 17:02:03 -07002026
Hugh Dickins62b61f62009-12-14 17:59:33 -08002027#ifdef CONFIG_MEMORY_HOTREMOVE
2028 /*
2029 * Choose a high priority since the callback takes ksm_thread_mutex:
2030 * later callbacks could only be taking locks which nest within that.
2031 */
2032 hotplug_memory_notifier(ksm_memory_callback, 100);
2033#endif
Izik Eidus31dbd012009-09-21 17:02:03 -07002034 return 0;
2035
Lai Jiangshand9f89842010-08-09 17:20:02 -07002036out_free:
Izik Eidus31dbd012009-09-21 17:02:03 -07002037 ksm_slab_free();
2038out:
2039 return err;
2040}
2041module_init(ksm_init)