blob: 43bc893470b40512e26bdd129fb2194708bfbbdc [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>
Hugh Dickinsf8af4da2009-09-21 17:01:57 -070037
Izik Eidus31dbd012009-09-21 17:02:03 -070038#include <asm/tlbflush.h>
Hugh Dickins73848b42009-12-14 17:59:22 -080039#include "internal.h"
Izik Eidus31dbd012009-09-21 17:02:03 -070040
41/*
42 * A few notes about the KSM scanning process,
43 * to make it easier to understand the data structures below:
44 *
45 * In order to reduce excessive scanning, KSM sorts the memory pages by their
46 * contents into a data structure that holds pointers to the pages' locations.
47 *
48 * Since the contents of the pages may change at any moment, KSM cannot just
49 * insert the pages into a normal sorted tree and expect it to find anything.
50 * Therefore KSM uses two data structures - the stable and the unstable tree.
51 *
52 * The stable tree holds pointers to all the merged pages (ksm pages), sorted
53 * by their contents. Because each such page is write-protected, searching on
54 * this tree is fully assured to be working (except when pages are unmapped),
55 * and therefore this tree is called the stable tree.
56 *
57 * In addition to the stable tree, KSM uses a second data structure called the
58 * unstable tree: this tree holds pointers to pages which have been found to
59 * be "unchanged for a period of time". The unstable tree sorts these pages
60 * by their contents, but since they are not write-protected, KSM cannot rely
61 * upon the unstable tree to work correctly - the unstable tree is liable to
62 * be corrupted as its contents are modified, and so it is called unstable.
63 *
64 * KSM solves this problem by several techniques:
65 *
66 * 1) The unstable tree is flushed every time KSM completes scanning all
67 * memory areas, and then the tree is rebuilt again from the beginning.
68 * 2) KSM will only insert into the unstable tree, pages whose hash value
69 * has not changed since the previous scan of all memory areas.
70 * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
71 * colors of the nodes and not on their contents, assuring that even when
72 * the tree gets "corrupted" it won't get out of balance, so scanning time
73 * remains the same (also, searching and inserting nodes in an rbtree uses
74 * the same algorithm, so we have no overhead when we flush and rebuild).
75 * 4) KSM never flushes the stable tree, which means that even if it were to
76 * take 10 attempts to find a page in the unstable tree, once it is found,
77 * it is secured in the stable tree. (When we scan a new page, we first
78 * compare it against the stable tree, and then against the unstable tree.)
79 */
80
81/**
82 * struct mm_slot - ksm information per mm that is being scanned
83 * @link: link to the mm_slots hash list
84 * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
Hugh Dickins6514d512009-12-14 17:59:19 -080085 * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
Izik Eidus31dbd012009-09-21 17:02:03 -070086 * @mm: the mm that this information is valid for
87 */
88struct mm_slot {
89 struct hlist_node link;
90 struct list_head mm_list;
Hugh Dickins6514d512009-12-14 17:59:19 -080091 struct rmap_item *rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -070092 struct mm_struct *mm;
93};
94
95/**
96 * struct ksm_scan - cursor for scanning
97 * @mm_slot: the current mm_slot we are scanning
98 * @address: the next address inside that to be scanned
Hugh Dickins6514d512009-12-14 17:59:19 -080099 * @rmap_list: link to the next rmap to be scanned in the rmap_list
Izik Eidus31dbd012009-09-21 17:02:03 -0700100 * @seqnr: count of completed full scans (needed when removing unstable node)
101 *
102 * There is only the one ksm_scan instance of this cursor structure.
103 */
104struct ksm_scan {
105 struct mm_slot *mm_slot;
106 unsigned long address;
Hugh Dickins6514d512009-12-14 17:59:19 -0800107 struct rmap_item **rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -0700108 unsigned long seqnr;
109};
110
111/**
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800112 * struct stable_node - node of the stable rbtree
113 * @node: rb node of this ksm page in the stable tree
114 * @hlist: hlist head of rmap_items using this ksm page
Hugh Dickins62b61f62009-12-14 17:59:33 -0800115 * @kpfn: page frame number of this ksm page
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800116 */
117struct stable_node {
118 struct rb_node node;
119 struct hlist_head hlist;
Hugh Dickins62b61f62009-12-14 17:59:33 -0800120 unsigned long kpfn;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800121};
122
123/**
Izik Eidus31dbd012009-09-21 17:02:03 -0700124 * struct rmap_item - reverse mapping item for virtual addresses
Hugh Dickins6514d512009-12-14 17:59:19 -0800125 * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
Hugh Dickinsdb114b82009-12-14 17:59:25 -0800126 * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
Izik Eidus31dbd012009-09-21 17:02:03 -0700127 * @mm: the memory structure this rmap_item is pointing into
128 * @address: the virtual address this rmap_item tracks (+ flags in low bits)
129 * @oldchecksum: previous checksum of the page at that virtual address
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800130 * @node: rb node of this rmap_item in the unstable tree
131 * @head: pointer to stable_node heading this list in the stable tree
132 * @hlist: link into hlist of rmap_items hanging off that stable_node
Izik Eidus31dbd012009-09-21 17:02:03 -0700133 */
134struct rmap_item {
Hugh Dickins6514d512009-12-14 17:59:19 -0800135 struct rmap_item *rmap_list;
Hugh Dickinsdb114b82009-12-14 17:59:25 -0800136 struct anon_vma *anon_vma; /* when stable */
Izik Eidus31dbd012009-09-21 17:02:03 -0700137 struct mm_struct *mm;
138 unsigned long address; /* + low bits used for flags below */
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800139 unsigned int oldchecksum; /* when unstable */
Izik Eidus31dbd012009-09-21 17:02:03 -0700140 union {
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800141 struct rb_node node; /* when node of unstable tree */
142 struct { /* when listed from stable tree */
143 struct stable_node *head;
144 struct hlist_node hlist;
145 };
Izik Eidus31dbd012009-09-21 17:02:03 -0700146 };
147};
148
149#define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800150#define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */
151#define STABLE_FLAG 0x200 /* is listed from the stable tree */
Izik Eidus31dbd012009-09-21 17:02:03 -0700152
153/* The stable and unstable tree heads */
154static struct rb_root root_stable_tree = RB_ROOT;
155static struct rb_root root_unstable_tree = RB_ROOT;
156
Lai Jiangshand9f89842010-08-09 17:20:02 -0700157#define MM_SLOTS_HASH_SHIFT 10
158#define MM_SLOTS_HASH_HEADS (1 << MM_SLOTS_HASH_SHIFT)
159static struct hlist_head mm_slots_hash[MM_SLOTS_HASH_HEADS];
Izik Eidus31dbd012009-09-21 17:02:03 -0700160
161static struct mm_slot ksm_mm_head = {
162 .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
163};
164static struct ksm_scan ksm_scan = {
165 .mm_slot = &ksm_mm_head,
166};
167
168static struct kmem_cache *rmap_item_cache;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800169static struct kmem_cache *stable_node_cache;
Izik Eidus31dbd012009-09-21 17:02:03 -0700170static struct kmem_cache *mm_slot_cache;
171
172/* The number of nodes in the stable tree */
Hugh Dickinsb4028262009-09-21 17:02:09 -0700173static unsigned long ksm_pages_shared;
Izik Eidus31dbd012009-09-21 17:02:03 -0700174
Hugh Dickinse178dfd2009-09-21 17:02:10 -0700175/* The number of page slots additionally sharing those nodes */
Hugh Dickinsb4028262009-09-21 17:02:09 -0700176static unsigned long ksm_pages_sharing;
Izik Eidus31dbd012009-09-21 17:02:03 -0700177
Hugh Dickins473b0ce2009-09-21 17:02:11 -0700178/* The number of nodes in the unstable tree */
179static unsigned long ksm_pages_unshared;
180
181/* The number of rmap_items in use: to calculate pages_volatile */
182static unsigned long ksm_rmap_items;
183
Izik Eidus31dbd012009-09-21 17:02:03 -0700184/* Number of pages ksmd should scan in one batch */
Izik Eidus2c6854f2009-09-23 15:56:04 -0700185static unsigned int ksm_thread_pages_to_scan = 100;
Izik Eidus31dbd012009-09-21 17:02:03 -0700186
187/* Milliseconds ksmd should sleep between batches */
Hugh Dickins2ffd8672009-09-21 17:02:23 -0700188static unsigned int ksm_thread_sleep_millisecs = 20;
Izik Eidus31dbd012009-09-21 17:02:03 -0700189
190#define KSM_RUN_STOP 0
191#define KSM_RUN_MERGE 1
192#define KSM_RUN_UNMERGE 2
Izik Eidus2c6854f2009-09-23 15:56:04 -0700193static unsigned int ksm_run = KSM_RUN_STOP;
Izik Eidus31dbd012009-09-21 17:02:03 -0700194
195static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
196static DEFINE_MUTEX(ksm_thread_mutex);
197static DEFINE_SPINLOCK(ksm_mmlist_lock);
198
199#define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
200 sizeof(struct __struct), __alignof__(struct __struct),\
201 (__flags), NULL)
202
203static int __init ksm_slab_init(void)
204{
205 rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
206 if (!rmap_item_cache)
207 goto out;
208
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800209 stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
210 if (!stable_node_cache)
211 goto out_free1;
212
Izik Eidus31dbd012009-09-21 17:02:03 -0700213 mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
214 if (!mm_slot_cache)
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800215 goto out_free2;
Izik Eidus31dbd012009-09-21 17:02:03 -0700216
217 return 0;
218
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800219out_free2:
220 kmem_cache_destroy(stable_node_cache);
221out_free1:
Izik Eidus31dbd012009-09-21 17:02:03 -0700222 kmem_cache_destroy(rmap_item_cache);
223out:
224 return -ENOMEM;
225}
226
227static void __init ksm_slab_free(void)
228{
229 kmem_cache_destroy(mm_slot_cache);
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800230 kmem_cache_destroy(stable_node_cache);
Izik Eidus31dbd012009-09-21 17:02:03 -0700231 kmem_cache_destroy(rmap_item_cache);
232 mm_slot_cache = NULL;
233}
234
235static inline struct rmap_item *alloc_rmap_item(void)
236{
Hugh Dickins473b0ce2009-09-21 17:02:11 -0700237 struct rmap_item *rmap_item;
238
239 rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL);
240 if (rmap_item)
241 ksm_rmap_items++;
242 return rmap_item;
Izik Eidus31dbd012009-09-21 17:02:03 -0700243}
244
245static inline void free_rmap_item(struct rmap_item *rmap_item)
246{
Hugh Dickins473b0ce2009-09-21 17:02:11 -0700247 ksm_rmap_items--;
Izik Eidus31dbd012009-09-21 17:02:03 -0700248 rmap_item->mm = NULL; /* debug safety */
249 kmem_cache_free(rmap_item_cache, rmap_item);
250}
251
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800252static inline struct stable_node *alloc_stable_node(void)
253{
254 return kmem_cache_alloc(stable_node_cache, GFP_KERNEL);
255}
256
257static inline void free_stable_node(struct stable_node *stable_node)
258{
259 kmem_cache_free(stable_node_cache, stable_node);
260}
261
Izik Eidus31dbd012009-09-21 17:02:03 -0700262static inline struct mm_slot *alloc_mm_slot(void)
263{
264 if (!mm_slot_cache) /* initialization failed */
265 return NULL;
266 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
267}
268
269static inline void free_mm_slot(struct mm_slot *mm_slot)
270{
271 kmem_cache_free(mm_slot_cache, mm_slot);
272}
273
Izik Eidus31dbd012009-09-21 17:02:03 -0700274static struct mm_slot *get_mm_slot(struct mm_struct *mm)
275{
276 struct mm_slot *mm_slot;
277 struct hlist_head *bucket;
278 struct hlist_node *node;
279
Lai Jiangshand9f89842010-08-09 17:20:02 -0700280 bucket = &mm_slots_hash[hash_ptr(mm, MM_SLOTS_HASH_SHIFT)];
Izik Eidus31dbd012009-09-21 17:02:03 -0700281 hlist_for_each_entry(mm_slot, node, bucket, link) {
282 if (mm == mm_slot->mm)
283 return mm_slot;
284 }
285 return NULL;
286}
287
288static void insert_to_mm_slots_hash(struct mm_struct *mm,
289 struct mm_slot *mm_slot)
290{
291 struct hlist_head *bucket;
292
Lai Jiangshand9f89842010-08-09 17:20:02 -0700293 bucket = &mm_slots_hash[hash_ptr(mm, MM_SLOTS_HASH_SHIFT)];
Izik Eidus31dbd012009-09-21 17:02:03 -0700294 mm_slot->mm = mm;
Izik Eidus31dbd012009-09-21 17:02:03 -0700295 hlist_add_head(&mm_slot->link, bucket);
296}
297
298static inline int in_stable_tree(struct rmap_item *rmap_item)
299{
300 return rmap_item->address & STABLE_FLAG;
301}
302
Hugh Dickinsdb114b82009-12-14 17:59:25 -0800303static void hold_anon_vma(struct rmap_item *rmap_item,
304 struct anon_vma *anon_vma)
305{
306 rmap_item->anon_vma = anon_vma;
Rik van Riel76545062010-08-09 17:18:41 -0700307 get_anon_vma(anon_vma);
Hugh Dickinsdb114b82009-12-14 17:59:25 -0800308}
309
Rik van Riel76545062010-08-09 17:18:41 -0700310static void ksm_drop_anon_vma(struct rmap_item *rmap_item)
Hugh Dickinsdb114b82009-12-14 17:59:25 -0800311{
312 struct anon_vma *anon_vma = rmap_item->anon_vma;
313
Rik van Riel76545062010-08-09 17:18:41 -0700314 drop_anon_vma(anon_vma);
Hugh Dickinsdb114b82009-12-14 17:59:25 -0800315}
316
Izik Eidus31dbd012009-09-21 17:02:03 -0700317/*
Hugh Dickinsa913e182009-09-21 17:02:26 -0700318 * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
319 * page tables after it has passed through ksm_exit() - which, if necessary,
320 * takes mmap_sem briefly to serialize against them. ksm_exit() does not set
321 * a special flag: they can just back out as soon as mm_users goes to zero.
322 * ksm_test_exit() is used throughout to make this test for exit: in some
323 * places for correctness, in some places just to avoid unnecessary work.
324 */
325static inline bool ksm_test_exit(struct mm_struct *mm)
326{
327 return atomic_read(&mm->mm_users) == 0;
328}
329
330/*
Izik Eidus31dbd012009-09-21 17:02:03 -0700331 * We use break_ksm to break COW on a ksm page: it's a stripped down
332 *
333 * if (get_user_pages(current, mm, addr, 1, 1, 1, &page, NULL) == 1)
334 * put_page(page);
335 *
336 * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
337 * in case the application has unmapped and remapped mm,addr meanwhile.
338 * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP
339 * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
340 */
Hugh Dickinsd952b792009-09-21 17:02:16 -0700341static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
Izik Eidus31dbd012009-09-21 17:02:03 -0700342{
343 struct page *page;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700344 int ret = 0;
Izik Eidus31dbd012009-09-21 17:02:03 -0700345
346 do {
347 cond_resched();
348 page = follow_page(vma, addr, FOLL_GET);
Dan Carpenter22eccdd2010-04-23 13:18:10 -0400349 if (IS_ERR_OR_NULL(page))
Izik Eidus31dbd012009-09-21 17:02:03 -0700350 break;
351 if (PageKsm(page))
352 ret = handle_mm_fault(vma->vm_mm, vma, addr,
353 FAULT_FLAG_WRITE);
354 else
355 ret = VM_FAULT_WRITE;
356 put_page(page);
Hugh Dickinsd952b792009-09-21 17:02:16 -0700357 } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_OOM)));
358 /*
359 * We must loop because handle_mm_fault() may back out if there's
360 * any difficulty e.g. if pte accessed bit gets updated concurrently.
361 *
362 * VM_FAULT_WRITE is what we have been hoping for: it indicates that
363 * COW has been broken, even if the vma does not permit VM_WRITE;
364 * but note that a concurrent fault might break PageKsm for us.
365 *
366 * VM_FAULT_SIGBUS could occur if we race with truncation of the
367 * backing file, which also invalidates anonymous pages: that's
368 * okay, that truncation will have unmapped the PageKsm for us.
369 *
370 * VM_FAULT_OOM: at the time of writing (late July 2009), setting
371 * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
372 * current task has TIF_MEMDIE set, and will be OOM killed on return
373 * to user; and ksmd, having no mm, would never be chosen for that.
374 *
375 * But if the mm is in a limited mem_cgroup, then the fault may fail
376 * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
377 * even ksmd can fail in this way - though it's usually breaking ksm
378 * just to undo a merge it made a moment before, so unlikely to oom.
379 *
380 * That's a pity: we might therefore have more kernel pages allocated
381 * than we're counting as nodes in the stable tree; but ksm_do_scan
382 * will retry to break_cow on each pass, so should recover the page
383 * in due course. The important thing is to not let VM_MERGEABLE
384 * be cleared while any such pages might remain in the area.
385 */
386 return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
Izik Eidus31dbd012009-09-21 17:02:03 -0700387}
388
Hugh Dickins8dd35572009-12-14 17:59:18 -0800389static void break_cow(struct rmap_item *rmap_item)
Izik Eidus31dbd012009-09-21 17:02:03 -0700390{
Hugh Dickins8dd35572009-12-14 17:59:18 -0800391 struct mm_struct *mm = rmap_item->mm;
392 unsigned long addr = rmap_item->address;
Izik Eidus31dbd012009-09-21 17:02:03 -0700393 struct vm_area_struct *vma;
394
Hugh Dickins4035c072009-12-14 17:59:27 -0800395 /*
396 * It is not an accident that whenever we want to break COW
397 * to undo, we also need to drop a reference to the anon_vma.
398 */
Rik van Riel76545062010-08-09 17:18:41 -0700399 ksm_drop_anon_vma(rmap_item);
Hugh Dickins4035c072009-12-14 17:59:27 -0800400
Hugh Dickins81464e302009-09-21 17:02:15 -0700401 down_read(&mm->mmap_sem);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700402 if (ksm_test_exit(mm))
403 goto out;
Izik Eidus31dbd012009-09-21 17:02:03 -0700404 vma = find_vma(mm, addr);
405 if (!vma || vma->vm_start > addr)
Hugh Dickins81464e302009-09-21 17:02:15 -0700406 goto out;
Izik Eidus31dbd012009-09-21 17:02:03 -0700407 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
Hugh Dickins81464e302009-09-21 17:02:15 -0700408 goto out;
Izik Eidus31dbd012009-09-21 17:02:03 -0700409 break_ksm(vma, addr);
Hugh Dickins81464e302009-09-21 17:02:15 -0700410out:
Izik Eidus31dbd012009-09-21 17:02:03 -0700411 up_read(&mm->mmap_sem);
412}
413
414static struct page *get_mergeable_page(struct rmap_item *rmap_item)
415{
416 struct mm_struct *mm = rmap_item->mm;
417 unsigned long addr = rmap_item->address;
418 struct vm_area_struct *vma;
419 struct page *page;
420
421 down_read(&mm->mmap_sem);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700422 if (ksm_test_exit(mm))
423 goto out;
Izik Eidus31dbd012009-09-21 17:02:03 -0700424 vma = find_vma(mm, addr);
425 if (!vma || vma->vm_start > addr)
426 goto out;
427 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
428 goto out;
429
430 page = follow_page(vma, addr, FOLL_GET);
Dan Carpenter22eccdd2010-04-23 13:18:10 -0400431 if (IS_ERR_OR_NULL(page))
Izik Eidus31dbd012009-09-21 17:02:03 -0700432 goto out;
433 if (PageAnon(page)) {
434 flush_anon_page(vma, page, addr);
435 flush_dcache_page(page);
436 } else {
437 put_page(page);
438out: page = NULL;
439 }
440 up_read(&mm->mmap_sem);
441 return page;
442}
443
Hugh Dickins4035c072009-12-14 17:59:27 -0800444static void remove_node_from_stable_tree(struct stable_node *stable_node)
445{
446 struct rmap_item *rmap_item;
447 struct hlist_node *hlist;
448
449 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
450 if (rmap_item->hlist.next)
451 ksm_pages_sharing--;
452 else
453 ksm_pages_shared--;
Rik van Riel76545062010-08-09 17:18:41 -0700454 ksm_drop_anon_vma(rmap_item);
Hugh Dickins4035c072009-12-14 17:59:27 -0800455 rmap_item->address &= PAGE_MASK;
456 cond_resched();
457 }
458
459 rb_erase(&stable_node->node, &root_stable_tree);
460 free_stable_node(stable_node);
461}
462
463/*
464 * get_ksm_page: checks if the page indicated by the stable node
465 * is still its ksm page, despite having held no reference to it.
466 * In which case we can trust the content of the page, and it
467 * returns the gotten page; but if the page has now been zapped,
468 * remove the stale node from the stable tree and return NULL.
469 *
470 * You would expect the stable_node to hold a reference to the ksm page.
471 * But if it increments the page's count, swapping out has to wait for
472 * ksmd to come around again before it can free the page, which may take
473 * seconds or even minutes: much too unresponsive. So instead we use a
474 * "keyhole reference": access to the ksm page from the stable node peeps
475 * out through its keyhole to see if that page still holds the right key,
476 * pointing back to this stable node. This relies on freeing a PageAnon
477 * page to reset its page->mapping to NULL, and relies on no other use of
478 * a page to put something that might look like our key in page->mapping.
479 *
480 * include/linux/pagemap.h page_cache_get_speculative() is a good reference,
481 * but this is different - made simpler by ksm_thread_mutex being held, but
482 * interesting for assuming that no other use of the struct page could ever
483 * put our expected_mapping into page->mapping (or a field of the union which
484 * coincides with page->mapping). The RCU calls are not for KSM at all, but
485 * to keep the page_count protocol described with page_cache_get_speculative.
486 *
487 * Note: it is possible that get_ksm_page() will return NULL one moment,
488 * then page the next, if the page is in between page_freeze_refs() and
489 * page_unfreeze_refs(): this shouldn't be a problem anywhere, the page
490 * is on its way to being freed; but it is an anomaly to bear in mind.
491 */
492static struct page *get_ksm_page(struct stable_node *stable_node)
493{
494 struct page *page;
495 void *expected_mapping;
496
Hugh Dickins62b61f62009-12-14 17:59:33 -0800497 page = pfn_to_page(stable_node->kpfn);
Hugh Dickins4035c072009-12-14 17:59:27 -0800498 expected_mapping = (void *)stable_node +
499 (PAGE_MAPPING_ANON | PAGE_MAPPING_KSM);
500 rcu_read_lock();
501 if (page->mapping != expected_mapping)
502 goto stale;
503 if (!get_page_unless_zero(page))
504 goto stale;
505 if (page->mapping != expected_mapping) {
506 put_page(page);
507 goto stale;
508 }
509 rcu_read_unlock();
510 return page;
511stale:
512 rcu_read_unlock();
513 remove_node_from_stable_tree(stable_node);
514 return NULL;
515}
516
Izik Eidus31dbd012009-09-21 17:02:03 -0700517/*
Izik Eidus31dbd012009-09-21 17:02:03 -0700518 * Removing rmap_item from stable or unstable tree.
519 * This function will clean the information from the stable/unstable tree.
520 */
521static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
522{
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800523 if (rmap_item->address & STABLE_FLAG) {
524 struct stable_node *stable_node;
Hugh Dickins5ad64682009-12-14 17:59:24 -0800525 struct page *page;
Izik Eidus31dbd012009-09-21 17:02:03 -0700526
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800527 stable_node = rmap_item->head;
Hugh Dickins4035c072009-12-14 17:59:27 -0800528 page = get_ksm_page(stable_node);
529 if (!page)
530 goto out;
531
Hugh Dickins5ad64682009-12-14 17:59:24 -0800532 lock_page(page);
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800533 hlist_del(&rmap_item->hlist);
Hugh Dickins4035c072009-12-14 17:59:27 -0800534 unlock_page(page);
535 put_page(page);
Hugh Dickins08beca42009-12-14 17:59:21 -0800536
Hugh Dickins4035c072009-12-14 17:59:27 -0800537 if (stable_node->hlist.first)
538 ksm_pages_sharing--;
539 else
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800540 ksm_pages_shared--;
Izik Eidus31dbd012009-09-21 17:02:03 -0700541
Rik van Riel76545062010-08-09 17:18:41 -0700542 ksm_drop_anon_vma(rmap_item);
Hugh Dickins93d17712009-12-14 17:59:16 -0800543 rmap_item->address &= PAGE_MASK;
Izik Eidus31dbd012009-09-21 17:02:03 -0700544
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800545 } else if (rmap_item->address & UNSTABLE_FLAG) {
Izik Eidus31dbd012009-09-21 17:02:03 -0700546 unsigned char age;
547 /*
Hugh Dickins9ba69292009-09-21 17:02:20 -0700548 * Usually ksmd can and must skip the rb_erase, because
Izik Eidus31dbd012009-09-21 17:02:03 -0700549 * root_unstable_tree was already reset to RB_ROOT.
Hugh Dickins9ba69292009-09-21 17:02:20 -0700550 * But be careful when an mm is exiting: do the rb_erase
551 * if this rmap_item was inserted by this scan, rather
552 * than left over from before.
Izik Eidus31dbd012009-09-21 17:02:03 -0700553 */
554 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
Hugh Dickinscd551f92009-09-21 17:02:17 -0700555 BUG_ON(age > 1);
Izik Eidus31dbd012009-09-21 17:02:03 -0700556 if (!age)
557 rb_erase(&rmap_item->node, &root_unstable_tree);
Izik Eidus31dbd012009-09-21 17:02:03 -0700558
Hugh Dickins93d17712009-12-14 17:59:16 -0800559 ksm_pages_unshared--;
560 rmap_item->address &= PAGE_MASK;
561 }
Hugh Dickins4035c072009-12-14 17:59:27 -0800562out:
Izik Eidus31dbd012009-09-21 17:02:03 -0700563 cond_resched(); /* we're called from many long loops */
564}
565
Izik Eidus31dbd012009-09-21 17:02:03 -0700566static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
Hugh Dickins6514d512009-12-14 17:59:19 -0800567 struct rmap_item **rmap_list)
Izik Eidus31dbd012009-09-21 17:02:03 -0700568{
Hugh Dickins6514d512009-12-14 17:59:19 -0800569 while (*rmap_list) {
570 struct rmap_item *rmap_item = *rmap_list;
571 *rmap_list = rmap_item->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -0700572 remove_rmap_item_from_tree(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -0700573 free_rmap_item(rmap_item);
574 }
575}
576
577/*
578 * Though it's very tempting to unmerge in_stable_tree(rmap_item)s rather
579 * than check every pte of a given vma, the locking doesn't quite work for
580 * that - an rmap_item is assigned to the stable tree after inserting ksm
581 * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing
582 * rmap_items from parent to child at fork time (so as not to waste time
583 * if exit comes before the next scan reaches it).
Hugh Dickins81464e302009-09-21 17:02:15 -0700584 *
585 * Similarly, although we'd like to remove rmap_items (so updating counts
586 * and freeing memory) when unmerging an area, it's easier to leave that
587 * to the next pass of ksmd - consider, for example, how ksmd might be
588 * in cmp_and_merge_page on one of the rmap_items we would be removing.
Izik Eidus31dbd012009-09-21 17:02:03 -0700589 */
Hugh Dickinsd952b792009-09-21 17:02:16 -0700590static int unmerge_ksm_pages(struct vm_area_struct *vma,
591 unsigned long start, unsigned long end)
Izik Eidus31dbd012009-09-21 17:02:03 -0700592{
593 unsigned long addr;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700594 int err = 0;
Izik Eidus31dbd012009-09-21 17:02:03 -0700595
Hugh Dickinsd952b792009-09-21 17:02:16 -0700596 for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
Hugh Dickins9ba69292009-09-21 17:02:20 -0700597 if (ksm_test_exit(vma->vm_mm))
598 break;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700599 if (signal_pending(current))
600 err = -ERESTARTSYS;
601 else
602 err = break_ksm(vma, addr);
603 }
604 return err;
Izik Eidus31dbd012009-09-21 17:02:03 -0700605}
606
Hugh Dickins2ffd8672009-09-21 17:02:23 -0700607#ifdef CONFIG_SYSFS
608/*
609 * Only called through the sysfs control interface:
610 */
Hugh Dickinsd952b792009-09-21 17:02:16 -0700611static int unmerge_and_remove_all_rmap_items(void)
Izik Eidus31dbd012009-09-21 17:02:03 -0700612{
613 struct mm_slot *mm_slot;
614 struct mm_struct *mm;
615 struct vm_area_struct *vma;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700616 int err = 0;
Izik Eidus31dbd012009-09-21 17:02:03 -0700617
Hugh Dickinsd952b792009-09-21 17:02:16 -0700618 spin_lock(&ksm_mmlist_lock);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700619 ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
Hugh Dickinsd952b792009-09-21 17:02:16 -0700620 struct mm_slot, mm_list);
621 spin_unlock(&ksm_mmlist_lock);
622
Hugh Dickins9ba69292009-09-21 17:02:20 -0700623 for (mm_slot = ksm_scan.mm_slot;
624 mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
Izik Eidus31dbd012009-09-21 17:02:03 -0700625 mm = mm_slot->mm;
626 down_read(&mm->mmap_sem);
627 for (vma = mm->mmap; vma; vma = vma->vm_next) {
Hugh Dickins9ba69292009-09-21 17:02:20 -0700628 if (ksm_test_exit(mm))
629 break;
Izik Eidus31dbd012009-09-21 17:02:03 -0700630 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
631 continue;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700632 err = unmerge_ksm_pages(vma,
633 vma->vm_start, vma->vm_end);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700634 if (err)
635 goto error;
Izik Eidus31dbd012009-09-21 17:02:03 -0700636 }
Hugh Dickins9ba69292009-09-21 17:02:20 -0700637
Hugh Dickins6514d512009-12-14 17:59:19 -0800638 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
Hugh Dickinsd952b792009-09-21 17:02:16 -0700639
640 spin_lock(&ksm_mmlist_lock);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700641 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
Hugh Dickinsd952b792009-09-21 17:02:16 -0700642 struct mm_slot, mm_list);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700643 if (ksm_test_exit(mm)) {
644 hlist_del(&mm_slot->link);
645 list_del(&mm_slot->mm_list);
646 spin_unlock(&ksm_mmlist_lock);
647
648 free_mm_slot(mm_slot);
649 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
650 up_read(&mm->mmap_sem);
651 mmdrop(mm);
652 } else {
653 spin_unlock(&ksm_mmlist_lock);
654 up_read(&mm->mmap_sem);
655 }
Izik Eidus31dbd012009-09-21 17:02:03 -0700656 }
657
Hugh Dickinsd952b792009-09-21 17:02:16 -0700658 ksm_scan.seqnr = 0;
Hugh Dickins9ba69292009-09-21 17:02:20 -0700659 return 0;
660
661error:
662 up_read(&mm->mmap_sem);
Izik Eidus31dbd012009-09-21 17:02:03 -0700663 spin_lock(&ksm_mmlist_lock);
Hugh Dickinsd952b792009-09-21 17:02:16 -0700664 ksm_scan.mm_slot = &ksm_mm_head;
Izik Eidus31dbd012009-09-21 17:02:03 -0700665 spin_unlock(&ksm_mmlist_lock);
Hugh Dickinsd952b792009-09-21 17:02:16 -0700666 return err;
Izik Eidus31dbd012009-09-21 17:02:03 -0700667}
Hugh Dickins2ffd8672009-09-21 17:02:23 -0700668#endif /* CONFIG_SYSFS */
Izik Eidus31dbd012009-09-21 17:02:03 -0700669
Izik Eidus31dbd012009-09-21 17:02:03 -0700670static u32 calc_checksum(struct page *page)
671{
672 u32 checksum;
673 void *addr = kmap_atomic(page, KM_USER0);
674 checksum = jhash2(addr, PAGE_SIZE / 4, 17);
675 kunmap_atomic(addr, KM_USER0);
676 return checksum;
677}
678
679static int memcmp_pages(struct page *page1, struct page *page2)
680{
681 char *addr1, *addr2;
682 int ret;
683
684 addr1 = kmap_atomic(page1, KM_USER0);
685 addr2 = kmap_atomic(page2, KM_USER1);
686 ret = memcmp(addr1, addr2, PAGE_SIZE);
687 kunmap_atomic(addr2, KM_USER1);
688 kunmap_atomic(addr1, KM_USER0);
689 return ret;
690}
691
692static inline int pages_identical(struct page *page1, struct page *page2)
693{
694 return !memcmp_pages(page1, page2);
695}
696
697static int write_protect_page(struct vm_area_struct *vma, struct page *page,
698 pte_t *orig_pte)
699{
700 struct mm_struct *mm = vma->vm_mm;
701 unsigned long addr;
702 pte_t *ptep;
703 spinlock_t *ptl;
704 int swapped;
705 int err = -EFAULT;
706
707 addr = page_address_in_vma(page, vma);
708 if (addr == -EFAULT)
709 goto out;
710
711 ptep = page_check_address(page, mm, addr, &ptl, 0);
712 if (!ptep)
713 goto out;
714
Hugh Dickins4e316352010-10-02 17:49:08 -0700715 if (pte_write(*ptep) || pte_dirty(*ptep)) {
Izik Eidus31dbd012009-09-21 17:02:03 -0700716 pte_t entry;
717
718 swapped = PageSwapCache(page);
719 flush_cache_page(vma, addr, page_to_pfn(page));
720 /*
721 * Ok this is tricky, when get_user_pages_fast() run it doesnt
722 * take any lock, therefore the check that we are going to make
723 * with the pagecount against the mapcount is racey and
724 * O_DIRECT can happen right after the check.
725 * So we clear the pte and flush the tlb before the check
726 * this assure us that no O_DIRECT can happen after the check
727 * or in the middle of the check.
728 */
729 entry = ptep_clear_flush(vma, addr, ptep);
730 /*
731 * Check that no O_DIRECT or similar I/O is in progress on the
732 * page
733 */
Hugh Dickins31e855e2009-12-14 17:59:17 -0800734 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
Robin Holtcb532372010-03-23 13:35:26 -0700735 set_pte_at(mm, addr, ptep, entry);
Izik Eidus31dbd012009-09-21 17:02:03 -0700736 goto out_unlock;
737 }
Hugh Dickins4e316352010-10-02 17:49:08 -0700738 if (pte_dirty(entry))
739 set_page_dirty(page);
740 entry = pte_mkclean(pte_wrprotect(entry));
Izik Eidus31dbd012009-09-21 17:02:03 -0700741 set_pte_at_notify(mm, addr, ptep, entry);
742 }
743 *orig_pte = *ptep;
744 err = 0;
745
746out_unlock:
747 pte_unmap_unlock(ptep, ptl);
748out:
749 return err;
750}
751
752/**
753 * replace_page - replace page in vma by new ksm page
Hugh Dickins8dd35572009-12-14 17:59:18 -0800754 * @vma: vma that holds the pte pointing to page
755 * @page: the page we are replacing by kpage
756 * @kpage: the ksm page we replace page by
Izik Eidus31dbd012009-09-21 17:02:03 -0700757 * @orig_pte: the original value of the pte
758 *
759 * Returns 0 on success, -EFAULT on failure.
760 */
Hugh Dickins8dd35572009-12-14 17:59:18 -0800761static int replace_page(struct vm_area_struct *vma, struct page *page,
762 struct page *kpage, pte_t orig_pte)
Izik Eidus31dbd012009-09-21 17:02:03 -0700763{
764 struct mm_struct *mm = vma->vm_mm;
765 pgd_t *pgd;
766 pud_t *pud;
767 pmd_t *pmd;
768 pte_t *ptep;
769 spinlock_t *ptl;
770 unsigned long addr;
Izik Eidus31dbd012009-09-21 17:02:03 -0700771 int err = -EFAULT;
772
Hugh Dickins8dd35572009-12-14 17:59:18 -0800773 addr = page_address_in_vma(page, vma);
Izik Eidus31dbd012009-09-21 17:02:03 -0700774 if (addr == -EFAULT)
775 goto out;
776
777 pgd = pgd_offset(mm, addr);
778 if (!pgd_present(*pgd))
779 goto out;
780
781 pud = pud_offset(pgd, addr);
782 if (!pud_present(*pud))
783 goto out;
784
785 pmd = pmd_offset(pud, addr);
786 if (!pmd_present(*pmd))
787 goto out;
788
789 ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
790 if (!pte_same(*ptep, orig_pte)) {
791 pte_unmap_unlock(ptep, ptl);
792 goto out;
793 }
794
Hugh Dickins8dd35572009-12-14 17:59:18 -0800795 get_page(kpage);
Hugh Dickins5ad64682009-12-14 17:59:24 -0800796 page_add_anon_rmap(kpage, vma, addr);
Izik Eidus31dbd012009-09-21 17:02:03 -0700797
798 flush_cache_page(vma, addr, pte_pfn(*ptep));
799 ptep_clear_flush(vma, addr, ptep);
Hugh Dickins8dd35572009-12-14 17:59:18 -0800800 set_pte_at_notify(mm, addr, ptep, mk_pte(kpage, vma->vm_page_prot));
Izik Eidus31dbd012009-09-21 17:02:03 -0700801
Hugh Dickins8dd35572009-12-14 17:59:18 -0800802 page_remove_rmap(page);
803 put_page(page);
Izik Eidus31dbd012009-09-21 17:02:03 -0700804
805 pte_unmap_unlock(ptep, ptl);
806 err = 0;
807out:
808 return err;
809}
810
811/*
812 * try_to_merge_one_page - take two pages and merge them into one
Hugh Dickins8dd35572009-12-14 17:59:18 -0800813 * @vma: the vma that holds the pte pointing to page
814 * @page: the PageAnon page that we want to replace with kpage
Hugh Dickins80e14822009-12-14 17:59:29 -0800815 * @kpage: the PageKsm page that we want to map instead of page,
816 * or NULL the first time when we want to use page as kpage.
Izik Eidus31dbd012009-09-21 17:02:03 -0700817 *
818 * This function returns 0 if the pages were merged, -EFAULT otherwise.
819 */
820static int try_to_merge_one_page(struct vm_area_struct *vma,
Hugh Dickins8dd35572009-12-14 17:59:18 -0800821 struct page *page, struct page *kpage)
Izik Eidus31dbd012009-09-21 17:02:03 -0700822{
823 pte_t orig_pte = __pte(0);
824 int err = -EFAULT;
825
Hugh Dickinsdb114b82009-12-14 17:59:25 -0800826 if (page == kpage) /* ksm page forked */
827 return 0;
828
Izik Eidus31dbd012009-09-21 17:02:03 -0700829 if (!(vma->vm_flags & VM_MERGEABLE))
830 goto out;
Hugh Dickins8dd35572009-12-14 17:59:18 -0800831 if (!PageAnon(page))
Izik Eidus31dbd012009-09-21 17:02:03 -0700832 goto out;
833
Izik Eidus31dbd012009-09-21 17:02:03 -0700834 /*
835 * We need the page lock to read a stable PageSwapCache in
836 * write_protect_page(). We use trylock_page() instead of
837 * lock_page() because we don't want to wait here - we
838 * prefer to continue scanning and merging different pages,
839 * then come back to this page when it is unlocked.
840 */
Hugh Dickins8dd35572009-12-14 17:59:18 -0800841 if (!trylock_page(page))
Hugh Dickins31e855e2009-12-14 17:59:17 -0800842 goto out;
Izik Eidus31dbd012009-09-21 17:02:03 -0700843 /*
844 * If this anonymous page is mapped only here, its pte may need
845 * to be write-protected. If it's mapped elsewhere, all of its
846 * ptes are necessarily already write-protected. But in either
847 * case, we need to lock and check page_count is not raised.
848 */
Hugh Dickins80e14822009-12-14 17:59:29 -0800849 if (write_protect_page(vma, page, &orig_pte) == 0) {
850 if (!kpage) {
851 /*
852 * While we hold page lock, upgrade page from
853 * PageAnon+anon_vma to PageKsm+NULL stable_node:
854 * stable_tree_insert() will update stable_node.
855 */
856 set_page_stable_node(page, NULL);
857 mark_page_accessed(page);
858 err = 0;
859 } else if (pages_identical(page, kpage))
860 err = replace_page(vma, page, kpage, orig_pte);
861 }
Izik Eidus31dbd012009-09-21 17:02:03 -0700862
Hugh Dickins80e14822009-12-14 17:59:29 -0800863 if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
Hugh Dickins73848b42009-12-14 17:59:22 -0800864 munlock_vma_page(page);
Hugh Dickins5ad64682009-12-14 17:59:24 -0800865 if (!PageMlocked(kpage)) {
866 unlock_page(page);
Hugh Dickins5ad64682009-12-14 17:59:24 -0800867 lock_page(kpage);
868 mlock_vma_page(kpage);
869 page = kpage; /* for final unlock */
870 }
871 }
Hugh Dickins73848b42009-12-14 17:59:22 -0800872
Hugh Dickins8dd35572009-12-14 17:59:18 -0800873 unlock_page(page);
Izik Eidus31dbd012009-09-21 17:02:03 -0700874out:
875 return err;
876}
877
878/*
Hugh Dickins81464e302009-09-21 17:02:15 -0700879 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
880 * but no new kernel page is allocated: kpage must already be a ksm page.
Hugh Dickins8dd35572009-12-14 17:59:18 -0800881 *
882 * This function returns 0 if the pages were merged, -EFAULT otherwise.
Hugh Dickins81464e302009-09-21 17:02:15 -0700883 */
Hugh Dickins8dd35572009-12-14 17:59:18 -0800884static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
885 struct page *page, struct page *kpage)
Hugh Dickins81464e302009-09-21 17:02:15 -0700886{
Hugh Dickins8dd35572009-12-14 17:59:18 -0800887 struct mm_struct *mm = rmap_item->mm;
Hugh Dickins81464e302009-09-21 17:02:15 -0700888 struct vm_area_struct *vma;
889 int err = -EFAULT;
890
Hugh Dickins8dd35572009-12-14 17:59:18 -0800891 down_read(&mm->mmap_sem);
892 if (ksm_test_exit(mm))
893 goto out;
894 vma = find_vma(mm, rmap_item->address);
895 if (!vma || vma->vm_start > rmap_item->address)
Hugh Dickins9ba69292009-09-21 17:02:20 -0700896 goto out;
897
Hugh Dickins8dd35572009-12-14 17:59:18 -0800898 err = try_to_merge_one_page(vma, page, kpage);
Hugh Dickinsdb114b82009-12-14 17:59:25 -0800899 if (err)
900 goto out;
901
902 /* Must get reference to anon_vma while still holding mmap_sem */
903 hold_anon_vma(rmap_item, vma->anon_vma);
Hugh Dickins81464e302009-09-21 17:02:15 -0700904out:
Hugh Dickins8dd35572009-12-14 17:59:18 -0800905 up_read(&mm->mmap_sem);
Hugh Dickins81464e302009-09-21 17:02:15 -0700906 return err;
907}
908
909/*
Izik Eidus31dbd012009-09-21 17:02:03 -0700910 * try_to_merge_two_pages - take two identical pages and prepare them
911 * to be merged into one page.
912 *
Hugh Dickins8dd35572009-12-14 17:59:18 -0800913 * This function returns the kpage if we successfully merged two identical
914 * pages into one ksm page, NULL otherwise.
Izik Eidus31dbd012009-09-21 17:02:03 -0700915 *
Hugh Dickins80e14822009-12-14 17:59:29 -0800916 * Note that this function upgrades page to ksm page: if one of the pages
Izik Eidus31dbd012009-09-21 17:02:03 -0700917 * is already a ksm page, try_to_merge_with_ksm_page should be used.
918 */
Hugh Dickins8dd35572009-12-14 17:59:18 -0800919static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
920 struct page *page,
921 struct rmap_item *tree_rmap_item,
922 struct page *tree_page)
Izik Eidus31dbd012009-09-21 17:02:03 -0700923{
Hugh Dickins80e14822009-12-14 17:59:29 -0800924 int err;
Izik Eidus31dbd012009-09-21 17:02:03 -0700925
Hugh Dickins80e14822009-12-14 17:59:29 -0800926 err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
Izik Eidus31dbd012009-09-21 17:02:03 -0700927 if (!err) {
Hugh Dickins8dd35572009-12-14 17:59:18 -0800928 err = try_to_merge_with_ksm_page(tree_rmap_item,
Hugh Dickins80e14822009-12-14 17:59:29 -0800929 tree_page, page);
Izik Eidus31dbd012009-09-21 17:02:03 -0700930 /*
Hugh Dickins81464e302009-09-21 17:02:15 -0700931 * If that fails, we have a ksm page with only one pte
932 * pointing to it: so break it.
Izik Eidus31dbd012009-09-21 17:02:03 -0700933 */
Hugh Dickins4035c072009-12-14 17:59:27 -0800934 if (err)
Hugh Dickins8dd35572009-12-14 17:59:18 -0800935 break_cow(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -0700936 }
Hugh Dickins80e14822009-12-14 17:59:29 -0800937 return err ? NULL : page;
Izik Eidus31dbd012009-09-21 17:02:03 -0700938}
939
940/*
Hugh Dickins8dd35572009-12-14 17:59:18 -0800941 * stable_tree_search - search for page inside the stable tree
Izik Eidus31dbd012009-09-21 17:02:03 -0700942 *
943 * This function checks if there is a page inside the stable tree
944 * with identical content to the page that we are scanning right now.
945 *
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800946 * This function returns the stable tree node of identical content if found,
Izik Eidus31dbd012009-09-21 17:02:03 -0700947 * NULL otherwise.
948 */
Hugh Dickins62b61f62009-12-14 17:59:33 -0800949static struct page *stable_tree_search(struct page *page)
Izik Eidus31dbd012009-09-21 17:02:03 -0700950{
951 struct rb_node *node = root_stable_tree.rb_node;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800952 struct stable_node *stable_node;
Izik Eidus31dbd012009-09-21 17:02:03 -0700953
Hugh Dickins08beca42009-12-14 17:59:21 -0800954 stable_node = page_stable_node(page);
955 if (stable_node) { /* ksm page forked */
956 get_page(page);
Hugh Dickins62b61f62009-12-14 17:59:33 -0800957 return page;
Hugh Dickins08beca42009-12-14 17:59:21 -0800958 }
959
Izik Eidus31dbd012009-09-21 17:02:03 -0700960 while (node) {
Hugh Dickins4035c072009-12-14 17:59:27 -0800961 struct page *tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -0700962 int ret;
963
Hugh Dickins08beca42009-12-14 17:59:21 -0800964 cond_resched();
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800965 stable_node = rb_entry(node, struct stable_node, node);
Hugh Dickins4035c072009-12-14 17:59:27 -0800966 tree_page = get_ksm_page(stable_node);
967 if (!tree_page)
968 return NULL;
Izik Eidus31dbd012009-09-21 17:02:03 -0700969
Hugh Dickins4035c072009-12-14 17:59:27 -0800970 ret = memcmp_pages(page, tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -0700971
Hugh Dickins4035c072009-12-14 17:59:27 -0800972 if (ret < 0) {
973 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -0700974 node = node->rb_left;
Hugh Dickins4035c072009-12-14 17:59:27 -0800975 } else if (ret > 0) {
976 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -0700977 node = node->rb_right;
Hugh Dickins4035c072009-12-14 17:59:27 -0800978 } else
Hugh Dickins62b61f62009-12-14 17:59:33 -0800979 return tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -0700980 }
981
982 return NULL;
983}
984
985/*
986 * stable_tree_insert - insert rmap_item pointing to new ksm page
987 * into the stable tree.
988 *
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800989 * This function returns the stable tree node just allocated on success,
990 * NULL otherwise.
Izik Eidus31dbd012009-09-21 17:02:03 -0700991 */
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800992static struct stable_node *stable_tree_insert(struct page *kpage)
Izik Eidus31dbd012009-09-21 17:02:03 -0700993{
994 struct rb_node **new = &root_stable_tree.rb_node;
995 struct rb_node *parent = NULL;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800996 struct stable_node *stable_node;
Izik Eidus31dbd012009-09-21 17:02:03 -0700997
998 while (*new) {
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(*new, 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(kpage, tree_page);
1009 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001010
1011 parent = *new;
1012 if (ret < 0)
1013 new = &parent->rb_left;
1014 else if (ret > 0)
1015 new = &parent->rb_right;
1016 else {
1017 /*
1018 * It is not a bug that stable_tree_search() didn't
1019 * find this node: because at that time our page was
1020 * not yet write-protected, so may have changed since.
1021 */
1022 return NULL;
1023 }
1024 }
1025
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001026 stable_node = alloc_stable_node();
1027 if (!stable_node)
1028 return NULL;
Izik Eidus31dbd012009-09-21 17:02:03 -07001029
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001030 rb_link_node(&stable_node->node, parent, new);
1031 rb_insert_color(&stable_node->node, &root_stable_tree);
1032
1033 INIT_HLIST_HEAD(&stable_node->hlist);
1034
Hugh Dickins62b61f62009-12-14 17:59:33 -08001035 stable_node->kpfn = page_to_pfn(kpage);
Hugh Dickins08beca42009-12-14 17:59:21 -08001036 set_page_stable_node(kpage, stable_node);
1037
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001038 return stable_node;
Izik Eidus31dbd012009-09-21 17:02:03 -07001039}
1040
1041/*
Hugh Dickins8dd35572009-12-14 17:59:18 -08001042 * unstable_tree_search_insert - search for identical page,
1043 * else insert rmap_item into the unstable tree.
Izik Eidus31dbd012009-09-21 17:02:03 -07001044 *
1045 * This function searches for a page in the unstable tree identical to the
1046 * page currently being scanned; and if no identical page is found in the
1047 * tree, we insert rmap_item as a new object into the unstable tree.
1048 *
1049 * This function returns pointer to rmap_item found to be identical
1050 * to the currently scanned page, NULL otherwise.
1051 *
1052 * This function does both searching and inserting, because they share
1053 * the same walking algorithm in an rbtree.
1054 */
Hugh Dickins8dd35572009-12-14 17:59:18 -08001055static
1056struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1057 struct page *page,
1058 struct page **tree_pagep)
1059
Izik Eidus31dbd012009-09-21 17:02:03 -07001060{
1061 struct rb_node **new = &root_unstable_tree.rb_node;
1062 struct rb_node *parent = NULL;
1063
1064 while (*new) {
1065 struct rmap_item *tree_rmap_item;
Hugh Dickins8dd35572009-12-14 17:59:18 -08001066 struct page *tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001067 int ret;
1068
Hugh Dickinsd178f272009-11-09 15:58:23 +00001069 cond_resched();
Izik Eidus31dbd012009-09-21 17:02:03 -07001070 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
Hugh Dickins8dd35572009-12-14 17:59:18 -08001071 tree_page = get_mergeable_page(tree_rmap_item);
Dan Carpenter22eccdd2010-04-23 13:18:10 -04001072 if (IS_ERR_OR_NULL(tree_page))
Izik Eidus31dbd012009-09-21 17:02:03 -07001073 return NULL;
1074
1075 /*
Hugh Dickins8dd35572009-12-14 17:59:18 -08001076 * Don't substitute a ksm page for a forked page.
Izik Eidus31dbd012009-09-21 17:02:03 -07001077 */
Hugh Dickins8dd35572009-12-14 17:59:18 -08001078 if (page == tree_page) {
1079 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001080 return NULL;
1081 }
1082
Hugh Dickins8dd35572009-12-14 17:59:18 -08001083 ret = memcmp_pages(page, tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001084
1085 parent = *new;
1086 if (ret < 0) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001087 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001088 new = &parent->rb_left;
1089 } else if (ret > 0) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001090 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001091 new = &parent->rb_right;
1092 } else {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001093 *tree_pagep = tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001094 return tree_rmap_item;
1095 }
1096 }
1097
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001098 rmap_item->address |= UNSTABLE_FLAG;
Izik Eidus31dbd012009-09-21 17:02:03 -07001099 rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1100 rb_link_node(&rmap_item->node, parent, new);
1101 rb_insert_color(&rmap_item->node, &root_unstable_tree);
1102
Hugh Dickins473b0ce2009-09-21 17:02:11 -07001103 ksm_pages_unshared++;
Izik Eidus31dbd012009-09-21 17:02:03 -07001104 return NULL;
1105}
1106
1107/*
1108 * stable_tree_append - add another rmap_item to the linked list of
1109 * rmap_items hanging off a given node of the stable tree, all sharing
1110 * the same ksm page.
1111 */
1112static void stable_tree_append(struct rmap_item *rmap_item,
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001113 struct stable_node *stable_node)
Izik Eidus31dbd012009-09-21 17:02:03 -07001114{
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001115 rmap_item->head = stable_node;
Izik Eidus31dbd012009-09-21 17:02:03 -07001116 rmap_item->address |= STABLE_FLAG;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001117 hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
Hugh Dickinse178dfd2009-09-21 17:02:10 -07001118
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001119 if (rmap_item->hlist.next)
1120 ksm_pages_sharing++;
1121 else
1122 ksm_pages_shared++;
Izik Eidus31dbd012009-09-21 17:02:03 -07001123}
1124
1125/*
Hugh Dickins81464e302009-09-21 17:02:15 -07001126 * cmp_and_merge_page - first see if page can be merged into the stable tree;
1127 * if not, compare checksum to previous and if it's the same, see if page can
1128 * be inserted into the unstable tree, or merged with a page already there and
1129 * both transferred to the stable tree.
Izik Eidus31dbd012009-09-21 17:02:03 -07001130 *
1131 * @page: the page that we are searching identical page to.
1132 * @rmap_item: the reverse mapping into the virtual address of this page
1133 */
1134static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1135{
Izik Eidus31dbd012009-09-21 17:02:03 -07001136 struct rmap_item *tree_rmap_item;
Hugh Dickins8dd35572009-12-14 17:59:18 -08001137 struct page *tree_page = NULL;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001138 struct stable_node *stable_node;
Hugh Dickins8dd35572009-12-14 17:59:18 -08001139 struct page *kpage;
Izik Eidus31dbd012009-09-21 17:02:03 -07001140 unsigned int checksum;
1141 int err;
1142
Hugh Dickins93d17712009-12-14 17:59:16 -08001143 remove_rmap_item_from_tree(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -07001144
1145 /* We first start with searching the page inside the stable tree */
Hugh Dickins62b61f62009-12-14 17:59:33 -08001146 kpage = stable_tree_search(page);
1147 if (kpage) {
Hugh Dickins08beca42009-12-14 17:59:21 -08001148 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
Izik Eidus31dbd012009-09-21 17:02:03 -07001149 if (!err) {
1150 /*
1151 * The page was successfully merged:
1152 * add its rmap_item to the stable tree.
1153 */
Hugh Dickins5ad64682009-12-14 17:59:24 -08001154 lock_page(kpage);
Hugh Dickins62b61f62009-12-14 17:59:33 -08001155 stable_tree_append(rmap_item, page_stable_node(kpage));
Hugh Dickins5ad64682009-12-14 17:59:24 -08001156 unlock_page(kpage);
Izik Eidus31dbd012009-09-21 17:02:03 -07001157 }
Hugh Dickins8dd35572009-12-14 17:59:18 -08001158 put_page(kpage);
Izik Eidus31dbd012009-09-21 17:02:03 -07001159 return;
1160 }
1161
1162 /*
Hugh Dickins4035c072009-12-14 17:59:27 -08001163 * If the hash value of the page has changed from the last time
1164 * we calculated it, this page is changing frequently: therefore we
1165 * don't want to insert it in the unstable tree, and we don't want
1166 * to waste our time searching for something identical to it there.
Izik Eidus31dbd012009-09-21 17:02:03 -07001167 */
1168 checksum = calc_checksum(page);
1169 if (rmap_item->oldchecksum != checksum) {
1170 rmap_item->oldchecksum = checksum;
1171 return;
1172 }
1173
Hugh Dickins8dd35572009-12-14 17:59:18 -08001174 tree_rmap_item =
1175 unstable_tree_search_insert(rmap_item, page, &tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001176 if (tree_rmap_item) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001177 kpage = try_to_merge_two_pages(rmap_item, page,
1178 tree_rmap_item, tree_page);
1179 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001180 /*
1181 * As soon as we merge this page, we want to remove the
1182 * rmap_item of the page we have merged with from the unstable
1183 * tree, and insert it instead as new node in the stable tree.
1184 */
Hugh Dickins8dd35572009-12-14 17:59:18 -08001185 if (kpage) {
Hugh Dickins93d17712009-12-14 17:59:16 -08001186 remove_rmap_item_from_tree(tree_rmap_item);
Hugh Dickins473b0ce2009-09-21 17:02:11 -07001187
Hugh Dickins5ad64682009-12-14 17:59:24 -08001188 lock_page(kpage);
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001189 stable_node = stable_tree_insert(kpage);
1190 if (stable_node) {
1191 stable_tree_append(tree_rmap_item, stable_node);
1192 stable_tree_append(rmap_item, stable_node);
1193 }
Hugh Dickins5ad64682009-12-14 17:59:24 -08001194 unlock_page(kpage);
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001195
Izik Eidus31dbd012009-09-21 17:02:03 -07001196 /*
1197 * If we fail to insert the page into the stable tree,
1198 * we will have 2 virtual addresses that are pointing
1199 * to a ksm page left outside the stable tree,
1200 * in which case we need to break_cow on both.
1201 */
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001202 if (!stable_node) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001203 break_cow(tree_rmap_item);
1204 break_cow(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -07001205 }
1206 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001207 }
1208}
1209
1210static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
Hugh Dickins6514d512009-12-14 17:59:19 -08001211 struct rmap_item **rmap_list,
Izik Eidus31dbd012009-09-21 17:02:03 -07001212 unsigned long addr)
1213{
1214 struct rmap_item *rmap_item;
1215
Hugh Dickins6514d512009-12-14 17:59:19 -08001216 while (*rmap_list) {
1217 rmap_item = *rmap_list;
Hugh Dickins93d17712009-12-14 17:59:16 -08001218 if ((rmap_item->address & PAGE_MASK) == addr)
Izik Eidus31dbd012009-09-21 17:02:03 -07001219 return rmap_item;
Izik Eidus31dbd012009-09-21 17:02:03 -07001220 if (rmap_item->address > addr)
1221 break;
Hugh Dickins6514d512009-12-14 17:59:19 -08001222 *rmap_list = rmap_item->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -07001223 remove_rmap_item_from_tree(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -07001224 free_rmap_item(rmap_item);
1225 }
1226
1227 rmap_item = alloc_rmap_item();
1228 if (rmap_item) {
1229 /* It has already been zeroed */
1230 rmap_item->mm = mm_slot->mm;
1231 rmap_item->address = addr;
Hugh Dickins6514d512009-12-14 17:59:19 -08001232 rmap_item->rmap_list = *rmap_list;
1233 *rmap_list = rmap_item;
Izik Eidus31dbd012009-09-21 17:02:03 -07001234 }
1235 return rmap_item;
1236}
1237
1238static struct rmap_item *scan_get_next_rmap_item(struct page **page)
1239{
1240 struct mm_struct *mm;
1241 struct mm_slot *slot;
1242 struct vm_area_struct *vma;
1243 struct rmap_item *rmap_item;
1244
1245 if (list_empty(&ksm_mm_head.mm_list))
1246 return NULL;
1247
1248 slot = ksm_scan.mm_slot;
1249 if (slot == &ksm_mm_head) {
1250 root_unstable_tree = RB_ROOT;
1251
1252 spin_lock(&ksm_mmlist_lock);
1253 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1254 ksm_scan.mm_slot = slot;
1255 spin_unlock(&ksm_mmlist_lock);
1256next_mm:
1257 ksm_scan.address = 0;
Hugh Dickins6514d512009-12-14 17:59:19 -08001258 ksm_scan.rmap_list = &slot->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -07001259 }
1260
1261 mm = slot->mm;
1262 down_read(&mm->mmap_sem);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001263 if (ksm_test_exit(mm))
1264 vma = NULL;
1265 else
1266 vma = find_vma(mm, ksm_scan.address);
1267
1268 for (; vma; vma = vma->vm_next) {
Izik Eidus31dbd012009-09-21 17:02:03 -07001269 if (!(vma->vm_flags & VM_MERGEABLE))
1270 continue;
1271 if (ksm_scan.address < vma->vm_start)
1272 ksm_scan.address = vma->vm_start;
1273 if (!vma->anon_vma)
1274 ksm_scan.address = vma->vm_end;
1275
1276 while (ksm_scan.address < vma->vm_end) {
Hugh Dickins9ba69292009-09-21 17:02:20 -07001277 if (ksm_test_exit(mm))
1278 break;
Izik Eidus31dbd012009-09-21 17:02:03 -07001279 *page = follow_page(vma, ksm_scan.address, FOLL_GET);
Dan Carpenter22eccdd2010-04-23 13:18:10 -04001280 if (!IS_ERR_OR_NULL(*page) && PageAnon(*page)) {
Izik Eidus31dbd012009-09-21 17:02:03 -07001281 flush_anon_page(vma, *page, ksm_scan.address);
1282 flush_dcache_page(*page);
1283 rmap_item = get_next_rmap_item(slot,
Hugh Dickins6514d512009-12-14 17:59:19 -08001284 ksm_scan.rmap_list, ksm_scan.address);
Izik Eidus31dbd012009-09-21 17:02:03 -07001285 if (rmap_item) {
Hugh Dickins6514d512009-12-14 17:59:19 -08001286 ksm_scan.rmap_list =
1287 &rmap_item->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -07001288 ksm_scan.address += PAGE_SIZE;
1289 } else
1290 put_page(*page);
1291 up_read(&mm->mmap_sem);
1292 return rmap_item;
1293 }
Dan Carpenter22eccdd2010-04-23 13:18:10 -04001294 if (!IS_ERR_OR_NULL(*page))
Izik Eidus31dbd012009-09-21 17:02:03 -07001295 put_page(*page);
1296 ksm_scan.address += PAGE_SIZE;
1297 cond_resched();
1298 }
1299 }
1300
Hugh Dickins9ba69292009-09-21 17:02:20 -07001301 if (ksm_test_exit(mm)) {
1302 ksm_scan.address = 0;
Hugh Dickins6514d512009-12-14 17:59:19 -08001303 ksm_scan.rmap_list = &slot->rmap_list;
Hugh Dickins9ba69292009-09-21 17:02:20 -07001304 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001305 /*
1306 * Nuke all the rmap_items that are above this current rmap:
1307 * because there were no VM_MERGEABLE vmas with such addresses.
1308 */
Hugh Dickins6514d512009-12-14 17:59:19 -08001309 remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
Izik Eidus31dbd012009-09-21 17:02:03 -07001310
1311 spin_lock(&ksm_mmlist_lock);
Hugh Dickinscd551f92009-09-21 17:02:17 -07001312 ksm_scan.mm_slot = list_entry(slot->mm_list.next,
1313 struct mm_slot, mm_list);
1314 if (ksm_scan.address == 0) {
1315 /*
1316 * We've completed a full scan of all vmas, holding mmap_sem
1317 * throughout, and found no VM_MERGEABLE: so do the same as
1318 * __ksm_exit does to remove this mm from all our lists now.
Hugh Dickins9ba69292009-09-21 17:02:20 -07001319 * This applies either when cleaning up after __ksm_exit
1320 * (but beware: we can reach here even before __ksm_exit),
1321 * or when all VM_MERGEABLE areas have been unmapped (and
1322 * mmap_sem then protects against race with MADV_MERGEABLE).
Hugh Dickinscd551f92009-09-21 17:02:17 -07001323 */
1324 hlist_del(&slot->link);
1325 list_del(&slot->mm_list);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001326 spin_unlock(&ksm_mmlist_lock);
1327
Hugh Dickinscd551f92009-09-21 17:02:17 -07001328 free_mm_slot(slot);
1329 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001330 up_read(&mm->mmap_sem);
1331 mmdrop(mm);
1332 } else {
1333 spin_unlock(&ksm_mmlist_lock);
1334 up_read(&mm->mmap_sem);
Hugh Dickinscd551f92009-09-21 17:02:17 -07001335 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001336
1337 /* Repeat until we've completed scanning the whole list */
Hugh Dickinscd551f92009-09-21 17:02:17 -07001338 slot = ksm_scan.mm_slot;
Izik Eidus31dbd012009-09-21 17:02:03 -07001339 if (slot != &ksm_mm_head)
1340 goto next_mm;
1341
Izik Eidus31dbd012009-09-21 17:02:03 -07001342 ksm_scan.seqnr++;
1343 return NULL;
1344}
1345
1346/**
1347 * ksm_do_scan - the ksm scanner main worker function.
1348 * @scan_npages - number of pages we want to scan before we return.
1349 */
1350static void ksm_do_scan(unsigned int scan_npages)
1351{
1352 struct rmap_item *rmap_item;
Dan Carpenter22eccdd2010-04-23 13:18:10 -04001353 struct page *uninitialized_var(page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001354
1355 while (scan_npages--) {
1356 cond_resched();
1357 rmap_item = scan_get_next_rmap_item(&page);
1358 if (!rmap_item)
1359 return;
1360 if (!PageKsm(page) || !in_stable_tree(rmap_item))
1361 cmp_and_merge_page(page, rmap_item);
1362 put_page(page);
1363 }
1364}
1365
Hugh Dickins6e158382009-09-21 17:02:14 -07001366static int ksmd_should_run(void)
1367{
1368 return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
1369}
1370
Izik Eidus31dbd012009-09-21 17:02:03 -07001371static int ksm_scan_thread(void *nothing)
1372{
Izik Eidus339aa622009-09-21 17:02:07 -07001373 set_user_nice(current, 5);
Izik Eidus31dbd012009-09-21 17:02:03 -07001374
1375 while (!kthread_should_stop()) {
Hugh Dickins6e158382009-09-21 17:02:14 -07001376 mutex_lock(&ksm_thread_mutex);
1377 if (ksmd_should_run())
Izik Eidus31dbd012009-09-21 17:02:03 -07001378 ksm_do_scan(ksm_thread_pages_to_scan);
Hugh Dickins6e158382009-09-21 17:02:14 -07001379 mutex_unlock(&ksm_thread_mutex);
1380
1381 if (ksmd_should_run()) {
Izik Eidus31dbd012009-09-21 17:02:03 -07001382 schedule_timeout_interruptible(
1383 msecs_to_jiffies(ksm_thread_sleep_millisecs));
1384 } else {
1385 wait_event_interruptible(ksm_thread_wait,
Hugh Dickins6e158382009-09-21 17:02:14 -07001386 ksmd_should_run() || kthread_should_stop());
Izik Eidus31dbd012009-09-21 17:02:03 -07001387 }
1388 }
1389 return 0;
1390}
1391
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001392int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1393 unsigned long end, int advice, unsigned long *vm_flags)
1394{
1395 struct mm_struct *mm = vma->vm_mm;
Hugh Dickinsd952b792009-09-21 17:02:16 -07001396 int err;
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001397
1398 switch (advice) {
1399 case MADV_MERGEABLE:
1400 /*
1401 * Be somewhat over-protective for now!
1402 */
1403 if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE |
1404 VM_PFNMAP | VM_IO | VM_DONTEXPAND |
1405 VM_RESERVED | VM_HUGETLB | VM_INSERTPAGE |
Hugh Dickins5ad64682009-12-14 17:59:24 -08001406 VM_NONLINEAR | VM_MIXEDMAP | VM_SAO))
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001407 return 0; /* just ignore the advice */
1408
Hugh Dickinsd952b792009-09-21 17:02:16 -07001409 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
1410 err = __ksm_enter(mm);
1411 if (err)
1412 return err;
1413 }
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001414
1415 *vm_flags |= VM_MERGEABLE;
1416 break;
1417
1418 case MADV_UNMERGEABLE:
1419 if (!(*vm_flags & VM_MERGEABLE))
1420 return 0; /* just ignore the advice */
1421
Hugh Dickinsd952b792009-09-21 17:02:16 -07001422 if (vma->anon_vma) {
1423 err = unmerge_ksm_pages(vma, start, end);
1424 if (err)
1425 return err;
1426 }
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001427
1428 *vm_flags &= ~VM_MERGEABLE;
1429 break;
1430 }
1431
1432 return 0;
1433}
1434
1435int __ksm_enter(struct mm_struct *mm)
1436{
Hugh Dickins6e158382009-09-21 17:02:14 -07001437 struct mm_slot *mm_slot;
1438 int needs_wakeup;
1439
1440 mm_slot = alloc_mm_slot();
Izik Eidus31dbd012009-09-21 17:02:03 -07001441 if (!mm_slot)
1442 return -ENOMEM;
1443
Hugh Dickins6e158382009-09-21 17:02:14 -07001444 /* Check ksm_run too? Would need tighter locking */
1445 needs_wakeup = list_empty(&ksm_mm_head.mm_list);
1446
Izik Eidus31dbd012009-09-21 17:02:03 -07001447 spin_lock(&ksm_mmlist_lock);
1448 insert_to_mm_slots_hash(mm, mm_slot);
1449 /*
1450 * Insert just behind the scanning cursor, to let the area settle
1451 * down a little; when fork is followed by immediate exec, we don't
1452 * want ksmd to waste time setting up and tearing down an rmap_list.
1453 */
1454 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
1455 spin_unlock(&ksm_mmlist_lock);
1456
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001457 set_bit(MMF_VM_MERGEABLE, &mm->flags);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001458 atomic_inc(&mm->mm_count);
Hugh Dickins6e158382009-09-21 17:02:14 -07001459
1460 if (needs_wakeup)
1461 wake_up_interruptible(&ksm_thread_wait);
1462
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001463 return 0;
1464}
1465
Andrea Arcangeli1c2fb7a2009-09-21 17:02:22 -07001466void __ksm_exit(struct mm_struct *mm)
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001467{
Hugh Dickinscd551f92009-09-21 17:02:17 -07001468 struct mm_slot *mm_slot;
Hugh Dickins9ba69292009-09-21 17:02:20 -07001469 int easy_to_free = 0;
Hugh Dickinscd551f92009-09-21 17:02:17 -07001470
Izik Eidus31dbd012009-09-21 17:02:03 -07001471 /*
Hugh Dickins9ba69292009-09-21 17:02:20 -07001472 * This process is exiting: if it's straightforward (as is the
1473 * case when ksmd was never running), free mm_slot immediately.
1474 * But if it's at the cursor or has rmap_items linked to it, use
1475 * mmap_sem to synchronize with any break_cows before pagetables
1476 * are freed, and leave the mm_slot on the list for ksmd to free.
1477 * Beware: ksm may already have noticed it exiting and freed the slot.
Izik Eidus31dbd012009-09-21 17:02:03 -07001478 */
Hugh Dickins9ba69292009-09-21 17:02:20 -07001479
Hugh Dickinscd551f92009-09-21 17:02:17 -07001480 spin_lock(&ksm_mmlist_lock);
1481 mm_slot = get_mm_slot(mm);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001482 if (mm_slot && ksm_scan.mm_slot != mm_slot) {
Hugh Dickins6514d512009-12-14 17:59:19 -08001483 if (!mm_slot->rmap_list) {
Hugh Dickins9ba69292009-09-21 17:02:20 -07001484 hlist_del(&mm_slot->link);
1485 list_del(&mm_slot->mm_list);
1486 easy_to_free = 1;
1487 } else {
1488 list_move(&mm_slot->mm_list,
1489 &ksm_scan.mm_slot->mm_list);
1490 }
Hugh Dickinscd551f92009-09-21 17:02:17 -07001491 }
Hugh Dickinscd551f92009-09-21 17:02:17 -07001492 spin_unlock(&ksm_mmlist_lock);
1493
Hugh Dickins9ba69292009-09-21 17:02:20 -07001494 if (easy_to_free) {
1495 free_mm_slot(mm_slot);
1496 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1497 mmdrop(mm);
1498 } else if (mm_slot) {
Hugh Dickins9ba69292009-09-21 17:02:20 -07001499 down_write(&mm->mmap_sem);
1500 up_write(&mm->mmap_sem);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001501 }
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001502}
Izik Eidus31dbd012009-09-21 17:02:03 -07001503
Hugh Dickins5ad64682009-12-14 17:59:24 -08001504struct page *ksm_does_need_to_copy(struct page *page,
1505 struct vm_area_struct *vma, unsigned long address)
1506{
1507 struct page *new_page;
1508
Hugh Dickins5ad64682009-12-14 17:59:24 -08001509 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1510 if (new_page) {
1511 copy_user_highpage(new_page, page, address, vma);
1512
1513 SetPageDirty(new_page);
1514 __SetPageUptodate(new_page);
1515 SetPageSwapBacked(new_page);
1516 __set_page_locked(new_page);
1517
1518 if (page_evictable(new_page, vma))
1519 lru_cache_add_lru(new_page, LRU_ACTIVE_ANON);
1520 else
1521 add_page_to_unevictable_list(new_page);
1522 }
1523
Hugh Dickins5ad64682009-12-14 17:59:24 -08001524 return new_page;
1525}
1526
1527int page_referenced_ksm(struct page *page, struct mem_cgroup *memcg,
1528 unsigned long *vm_flags)
1529{
1530 struct stable_node *stable_node;
1531 struct rmap_item *rmap_item;
1532 struct hlist_node *hlist;
1533 unsigned int mapcount = page_mapcount(page);
1534 int referenced = 0;
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001535 int search_new_forks = 0;
Hugh Dickins5ad64682009-12-14 17:59:24 -08001536
1537 VM_BUG_ON(!PageKsm(page));
1538 VM_BUG_ON(!PageLocked(page));
1539
1540 stable_node = page_stable_node(page);
1541 if (!stable_node)
1542 return 0;
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001543again:
Hugh Dickins5ad64682009-12-14 17:59:24 -08001544 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001545 struct anon_vma *anon_vma = rmap_item->anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -08001546 struct anon_vma_chain *vmac;
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001547 struct vm_area_struct *vma;
Hugh Dickins5ad64682009-12-14 17:59:24 -08001548
Rik van Rielcba48b92010-08-09 17:18:38 -07001549 anon_vma_lock(anon_vma);
Rik van Riel5beb4932010-03-05 13:42:07 -08001550 list_for_each_entry(vmac, &anon_vma->head, same_anon_vma) {
1551 vma = vmac->vma;
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001552 if (rmap_item->address < vma->vm_start ||
1553 rmap_item->address >= vma->vm_end)
1554 continue;
1555 /*
1556 * Initially we examine only the vma which covers this
1557 * rmap_item; but later, if there is still work to do,
1558 * we examine covering vmas in other mms: in case they
1559 * were forked from the original since ksmd passed.
1560 */
1561 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1562 continue;
Hugh Dickins5ad64682009-12-14 17:59:24 -08001563
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001564 if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
1565 continue;
1566
1567 referenced += page_referenced_one(page, vma,
Hugh Dickins5ad64682009-12-14 17:59:24 -08001568 rmap_item->address, &mapcount, vm_flags);
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001569 if (!search_new_forks || !mapcount)
1570 break;
1571 }
Rik van Rielcba48b92010-08-09 17:18:38 -07001572 anon_vma_unlock(anon_vma);
Hugh Dickins5ad64682009-12-14 17:59:24 -08001573 if (!mapcount)
1574 goto out;
1575 }
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001576 if (!search_new_forks++)
1577 goto again;
Hugh Dickins5ad64682009-12-14 17:59:24 -08001578out:
Hugh Dickins5ad64682009-12-14 17:59:24 -08001579 return referenced;
1580}
1581
1582int try_to_unmap_ksm(struct page *page, enum ttu_flags flags)
1583{
1584 struct stable_node *stable_node;
1585 struct hlist_node *hlist;
1586 struct rmap_item *rmap_item;
1587 int ret = SWAP_AGAIN;
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001588 int search_new_forks = 0;
Hugh Dickins5ad64682009-12-14 17:59:24 -08001589
1590 VM_BUG_ON(!PageKsm(page));
1591 VM_BUG_ON(!PageLocked(page));
1592
1593 stable_node = page_stable_node(page);
1594 if (!stable_node)
1595 return SWAP_FAIL;
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001596again:
Hugh Dickins5ad64682009-12-14 17:59:24 -08001597 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001598 struct anon_vma *anon_vma = rmap_item->anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -08001599 struct anon_vma_chain *vmac;
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001600 struct vm_area_struct *vma;
Hugh Dickins5ad64682009-12-14 17:59:24 -08001601
Rik van Rielcba48b92010-08-09 17:18:38 -07001602 anon_vma_lock(anon_vma);
Rik van Riel5beb4932010-03-05 13:42:07 -08001603 list_for_each_entry(vmac, &anon_vma->head, same_anon_vma) {
1604 vma = vmac->vma;
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001605 if (rmap_item->address < vma->vm_start ||
1606 rmap_item->address >= vma->vm_end)
1607 continue;
1608 /*
1609 * Initially we examine only the vma which covers this
1610 * rmap_item; but later, if there is still work to do,
1611 * we examine covering vmas in other mms: in case they
1612 * were forked from the original since ksmd passed.
1613 */
1614 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1615 continue;
1616
1617 ret = try_to_unmap_one(page, vma,
1618 rmap_item->address, flags);
1619 if (ret != SWAP_AGAIN || !page_mapped(page)) {
Rik van Rielcba48b92010-08-09 17:18:38 -07001620 anon_vma_unlock(anon_vma);
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001621 goto out;
1622 }
1623 }
Rik van Rielcba48b92010-08-09 17:18:38 -07001624 anon_vma_unlock(anon_vma);
Hugh Dickins5ad64682009-12-14 17:59:24 -08001625 }
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001626 if (!search_new_forks++)
1627 goto again;
Hugh Dickins5ad64682009-12-14 17:59:24 -08001628out:
Hugh Dickins5ad64682009-12-14 17:59:24 -08001629 return ret;
1630}
1631
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001632#ifdef CONFIG_MIGRATION
1633int rmap_walk_ksm(struct page *page, int (*rmap_one)(struct page *,
1634 struct vm_area_struct *, unsigned long, void *), void *arg)
1635{
1636 struct stable_node *stable_node;
1637 struct hlist_node *hlist;
1638 struct rmap_item *rmap_item;
1639 int ret = SWAP_AGAIN;
1640 int search_new_forks = 0;
1641
1642 VM_BUG_ON(!PageKsm(page));
1643 VM_BUG_ON(!PageLocked(page));
1644
1645 stable_node = page_stable_node(page);
1646 if (!stable_node)
1647 return ret;
1648again:
1649 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1650 struct anon_vma *anon_vma = rmap_item->anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -08001651 struct anon_vma_chain *vmac;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001652 struct vm_area_struct *vma;
1653
Rik van Rielcba48b92010-08-09 17:18:38 -07001654 anon_vma_lock(anon_vma);
Rik van Riel5beb4932010-03-05 13:42:07 -08001655 list_for_each_entry(vmac, &anon_vma->head, same_anon_vma) {
1656 vma = vmac->vma;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001657 if (rmap_item->address < vma->vm_start ||
1658 rmap_item->address >= vma->vm_end)
1659 continue;
1660 /*
1661 * Initially we examine only the vma which covers this
1662 * rmap_item; but later, if there is still work to do,
1663 * we examine covering vmas in other mms: in case they
1664 * were forked from the original since ksmd passed.
1665 */
1666 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1667 continue;
1668
1669 ret = rmap_one(page, vma, rmap_item->address, arg);
1670 if (ret != SWAP_AGAIN) {
Rik van Rielcba48b92010-08-09 17:18:38 -07001671 anon_vma_unlock(anon_vma);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001672 goto out;
1673 }
1674 }
Rik van Rielcba48b92010-08-09 17:18:38 -07001675 anon_vma_unlock(anon_vma);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001676 }
1677 if (!search_new_forks++)
1678 goto again;
1679out:
1680 return ret;
1681}
1682
1683void ksm_migrate_page(struct page *newpage, struct page *oldpage)
1684{
1685 struct stable_node *stable_node;
1686
1687 VM_BUG_ON(!PageLocked(oldpage));
1688 VM_BUG_ON(!PageLocked(newpage));
1689 VM_BUG_ON(newpage->mapping != oldpage->mapping);
1690
1691 stable_node = page_stable_node(newpage);
1692 if (stable_node) {
Hugh Dickins62b61f62009-12-14 17:59:33 -08001693 VM_BUG_ON(stable_node->kpfn != page_to_pfn(oldpage));
1694 stable_node->kpfn = page_to_pfn(newpage);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001695 }
1696}
1697#endif /* CONFIG_MIGRATION */
1698
Hugh Dickins62b61f62009-12-14 17:59:33 -08001699#ifdef CONFIG_MEMORY_HOTREMOVE
1700static struct stable_node *ksm_check_stable_tree(unsigned long start_pfn,
1701 unsigned long end_pfn)
1702{
1703 struct rb_node *node;
1704
1705 for (node = rb_first(&root_stable_tree); node; node = rb_next(node)) {
1706 struct stable_node *stable_node;
1707
1708 stable_node = rb_entry(node, struct stable_node, node);
1709 if (stable_node->kpfn >= start_pfn &&
1710 stable_node->kpfn < end_pfn)
1711 return stable_node;
1712 }
1713 return NULL;
1714}
1715
1716static int ksm_memory_callback(struct notifier_block *self,
1717 unsigned long action, void *arg)
1718{
1719 struct memory_notify *mn = arg;
1720 struct stable_node *stable_node;
1721
1722 switch (action) {
1723 case MEM_GOING_OFFLINE:
1724 /*
1725 * Keep it very simple for now: just lock out ksmd and
1726 * MADV_UNMERGEABLE while any memory is going offline.
KOSAKI Motohiroa0b0f582010-12-02 14:31:20 -08001727 * mutex_lock_nested() is necessary because lockdep was alarmed
1728 * that here we take ksm_thread_mutex inside notifier chain
1729 * mutex, and later take notifier chain mutex inside
1730 * ksm_thread_mutex to unlock it. But that's safe because both
1731 * are inside mem_hotplug_mutex.
Hugh Dickins62b61f62009-12-14 17:59:33 -08001732 */
KOSAKI Motohiroa0b0f582010-12-02 14:31:20 -08001733 mutex_lock_nested(&ksm_thread_mutex, SINGLE_DEPTH_NESTING);
Hugh Dickins62b61f62009-12-14 17:59:33 -08001734 break;
1735
1736 case MEM_OFFLINE:
1737 /*
1738 * Most of the work is done by page migration; but there might
1739 * be a few stable_nodes left over, still pointing to struct
1740 * pages which have been offlined: prune those from the tree.
1741 */
1742 while ((stable_node = ksm_check_stable_tree(mn->start_pfn,
1743 mn->start_pfn + mn->nr_pages)) != NULL)
1744 remove_node_from_stable_tree(stable_node);
1745 /* fallthrough */
1746
1747 case MEM_CANCEL_OFFLINE:
1748 mutex_unlock(&ksm_thread_mutex);
1749 break;
1750 }
1751 return NOTIFY_OK;
1752}
1753#endif /* CONFIG_MEMORY_HOTREMOVE */
1754
Hugh Dickins2ffd8672009-09-21 17:02:23 -07001755#ifdef CONFIG_SYSFS
1756/*
1757 * This all compiles without CONFIG_SYSFS, but is a waste of space.
1758 */
1759
Izik Eidus31dbd012009-09-21 17:02:03 -07001760#define KSM_ATTR_RO(_name) \
1761 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1762#define KSM_ATTR(_name) \
1763 static struct kobj_attribute _name##_attr = \
1764 __ATTR(_name, 0644, _name##_show, _name##_store)
1765
1766static ssize_t sleep_millisecs_show(struct kobject *kobj,
1767 struct kobj_attribute *attr, char *buf)
1768{
1769 return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
1770}
1771
1772static ssize_t sleep_millisecs_store(struct kobject *kobj,
1773 struct kobj_attribute *attr,
1774 const char *buf, size_t count)
1775{
1776 unsigned long msecs;
1777 int err;
1778
1779 err = strict_strtoul(buf, 10, &msecs);
1780 if (err || msecs > UINT_MAX)
1781 return -EINVAL;
1782
1783 ksm_thread_sleep_millisecs = msecs;
1784
1785 return count;
1786}
1787KSM_ATTR(sleep_millisecs);
1788
1789static ssize_t pages_to_scan_show(struct kobject *kobj,
1790 struct kobj_attribute *attr, char *buf)
1791{
1792 return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
1793}
1794
1795static ssize_t pages_to_scan_store(struct kobject *kobj,
1796 struct kobj_attribute *attr,
1797 const char *buf, size_t count)
1798{
1799 int err;
1800 unsigned long nr_pages;
1801
1802 err = strict_strtoul(buf, 10, &nr_pages);
1803 if (err || nr_pages > UINT_MAX)
1804 return -EINVAL;
1805
1806 ksm_thread_pages_to_scan = nr_pages;
1807
1808 return count;
1809}
1810KSM_ATTR(pages_to_scan);
1811
1812static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
1813 char *buf)
1814{
1815 return sprintf(buf, "%u\n", ksm_run);
1816}
1817
1818static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
1819 const char *buf, size_t count)
1820{
1821 int err;
1822 unsigned long flags;
1823
1824 err = strict_strtoul(buf, 10, &flags);
1825 if (err || flags > UINT_MAX)
1826 return -EINVAL;
1827 if (flags > KSM_RUN_UNMERGE)
1828 return -EINVAL;
1829
1830 /*
1831 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
1832 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
Hugh Dickinsd0f209f2009-12-14 17:59:34 -08001833 * breaking COW to free the pages_shared (but leaves mm_slots
1834 * on the list for when ksmd may be set running again).
Izik Eidus31dbd012009-09-21 17:02:03 -07001835 */
1836
1837 mutex_lock(&ksm_thread_mutex);
1838 if (ksm_run != flags) {
1839 ksm_run = flags;
Hugh Dickinsd952b792009-09-21 17:02:16 -07001840 if (flags & KSM_RUN_UNMERGE) {
Hugh Dickins35451be2009-09-21 17:02:27 -07001841 current->flags |= PF_OOM_ORIGIN;
Hugh Dickinsd952b792009-09-21 17:02:16 -07001842 err = unmerge_and_remove_all_rmap_items();
Hugh Dickins35451be2009-09-21 17:02:27 -07001843 current->flags &= ~PF_OOM_ORIGIN;
Hugh Dickinsd952b792009-09-21 17:02:16 -07001844 if (err) {
1845 ksm_run = KSM_RUN_STOP;
1846 count = err;
1847 }
1848 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001849 }
1850 mutex_unlock(&ksm_thread_mutex);
1851
1852 if (flags & KSM_RUN_MERGE)
1853 wake_up_interruptible(&ksm_thread_wait);
1854
1855 return count;
1856}
1857KSM_ATTR(run);
1858
Hugh Dickinsb4028262009-09-21 17:02:09 -07001859static ssize_t pages_shared_show(struct kobject *kobj,
1860 struct kobj_attribute *attr, char *buf)
1861{
1862 return sprintf(buf, "%lu\n", ksm_pages_shared);
1863}
1864KSM_ATTR_RO(pages_shared);
1865
1866static ssize_t pages_sharing_show(struct kobject *kobj,
1867 struct kobj_attribute *attr, char *buf)
1868{
Hugh Dickinse178dfd2009-09-21 17:02:10 -07001869 return sprintf(buf, "%lu\n", ksm_pages_sharing);
Hugh Dickinsb4028262009-09-21 17:02:09 -07001870}
1871KSM_ATTR_RO(pages_sharing);
1872
Hugh Dickins473b0ce2009-09-21 17:02:11 -07001873static ssize_t pages_unshared_show(struct kobject *kobj,
1874 struct kobj_attribute *attr, char *buf)
1875{
1876 return sprintf(buf, "%lu\n", ksm_pages_unshared);
1877}
1878KSM_ATTR_RO(pages_unshared);
1879
1880static ssize_t pages_volatile_show(struct kobject *kobj,
1881 struct kobj_attribute *attr, char *buf)
1882{
1883 long ksm_pages_volatile;
1884
1885 ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
1886 - ksm_pages_sharing - ksm_pages_unshared;
1887 /*
1888 * It was not worth any locking to calculate that statistic,
1889 * but it might therefore sometimes be negative: conceal that.
1890 */
1891 if (ksm_pages_volatile < 0)
1892 ksm_pages_volatile = 0;
1893 return sprintf(buf, "%ld\n", ksm_pages_volatile);
1894}
1895KSM_ATTR_RO(pages_volatile);
1896
1897static ssize_t full_scans_show(struct kobject *kobj,
1898 struct kobj_attribute *attr, char *buf)
1899{
1900 return sprintf(buf, "%lu\n", ksm_scan.seqnr);
1901}
1902KSM_ATTR_RO(full_scans);
1903
Izik Eidus31dbd012009-09-21 17:02:03 -07001904static struct attribute *ksm_attrs[] = {
1905 &sleep_millisecs_attr.attr,
1906 &pages_to_scan_attr.attr,
1907 &run_attr.attr,
Hugh Dickinsb4028262009-09-21 17:02:09 -07001908 &pages_shared_attr.attr,
1909 &pages_sharing_attr.attr,
Hugh Dickins473b0ce2009-09-21 17:02:11 -07001910 &pages_unshared_attr.attr,
1911 &pages_volatile_attr.attr,
1912 &full_scans_attr.attr,
Izik Eidus31dbd012009-09-21 17:02:03 -07001913 NULL,
1914};
1915
1916static struct attribute_group ksm_attr_group = {
1917 .attrs = ksm_attrs,
1918 .name = "ksm",
1919};
Hugh Dickins2ffd8672009-09-21 17:02:23 -07001920#endif /* CONFIG_SYSFS */
Izik Eidus31dbd012009-09-21 17:02:03 -07001921
1922static int __init ksm_init(void)
1923{
1924 struct task_struct *ksm_thread;
1925 int err;
1926
1927 err = ksm_slab_init();
1928 if (err)
1929 goto out;
1930
Izik Eidus31dbd012009-09-21 17:02:03 -07001931 ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
1932 if (IS_ERR(ksm_thread)) {
1933 printk(KERN_ERR "ksm: creating kthread failed\n");
1934 err = PTR_ERR(ksm_thread);
Lai Jiangshand9f89842010-08-09 17:20:02 -07001935 goto out_free;
Izik Eidus31dbd012009-09-21 17:02:03 -07001936 }
1937
Hugh Dickins2ffd8672009-09-21 17:02:23 -07001938#ifdef CONFIG_SYSFS
Izik Eidus31dbd012009-09-21 17:02:03 -07001939 err = sysfs_create_group(mm_kobj, &ksm_attr_group);
1940 if (err) {
1941 printk(KERN_ERR "ksm: register sysfs failed\n");
Hugh Dickins2ffd8672009-09-21 17:02:23 -07001942 kthread_stop(ksm_thread);
Lai Jiangshand9f89842010-08-09 17:20:02 -07001943 goto out_free;
Izik Eidus31dbd012009-09-21 17:02:03 -07001944 }
Hugh Dickinsc73602a2009-10-07 16:32:22 -07001945#else
1946 ksm_run = KSM_RUN_MERGE; /* no way for user to start it */
1947
Hugh Dickins2ffd8672009-09-21 17:02:23 -07001948#endif /* CONFIG_SYSFS */
Izik Eidus31dbd012009-09-21 17:02:03 -07001949
Hugh Dickins62b61f62009-12-14 17:59:33 -08001950#ifdef CONFIG_MEMORY_HOTREMOVE
1951 /*
1952 * Choose a high priority since the callback takes ksm_thread_mutex:
1953 * later callbacks could only be taking locks which nest within that.
1954 */
1955 hotplug_memory_notifier(ksm_memory_callback, 100);
1956#endif
Izik Eidus31dbd012009-09-21 17:02:03 -07001957 return 0;
1958
Lai Jiangshand9f89842010-08-09 17:20:02 -07001959out_free:
Izik Eidus31dbd012009-09-21 17:02:03 -07001960 ksm_slab_free();
1961out:
1962 return err;
1963}
1964module_init(ksm_init)