blob: b1873cf03ed986bcb062259da8f1c4093a97160b [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 Dickins4035c07a2009-12-14 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 Dickins4035c07a2009-12-14 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 Dickins4035c07a2009-12-14 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 Dickins4035c07a2009-12-14 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 Dickins4035c07a2009-12-14 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 Dickins4035c07a2009-12-14 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 Dickins4035c07a2009-12-14 17:59:27 -0800534 unlock_page(page);
535 put_page(page);
Hugh Dickins08beca42009-12-14 17:59:21 -0800536
Hugh Dickins4035c07a2009-12-14 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 Dickins4035c07a2009-12-14 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
715 if (pte_write(*ptep)) {
716 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 }
738 entry = pte_wrprotect(entry);
739 set_pte_at_notify(mm, addr, ptep, entry);
740 }
741 *orig_pte = *ptep;
742 err = 0;
743
744out_unlock:
745 pte_unmap_unlock(ptep, ptl);
746out:
747 return err;
748}
749
750/**
751 * replace_page - replace page in vma by new ksm page
Hugh Dickins8dd35572009-12-14 17:59:18 -0800752 * @vma: vma that holds the pte pointing to page
753 * @page: the page we are replacing by kpage
754 * @kpage: the ksm page we replace page by
Izik Eidus31dbd012009-09-21 17:02:03 -0700755 * @orig_pte: the original value of the pte
756 *
757 * Returns 0 on success, -EFAULT on failure.
758 */
Hugh Dickins8dd35572009-12-14 17:59:18 -0800759static int replace_page(struct vm_area_struct *vma, struct page *page,
760 struct page *kpage, pte_t orig_pte)
Izik Eidus31dbd012009-09-21 17:02:03 -0700761{
762 struct mm_struct *mm = vma->vm_mm;
763 pgd_t *pgd;
764 pud_t *pud;
765 pmd_t *pmd;
766 pte_t *ptep;
767 spinlock_t *ptl;
768 unsigned long addr;
Izik Eidus31dbd012009-09-21 17:02:03 -0700769 int err = -EFAULT;
770
Hugh Dickins8dd35572009-12-14 17:59:18 -0800771 addr = page_address_in_vma(page, vma);
Izik Eidus31dbd012009-09-21 17:02:03 -0700772 if (addr == -EFAULT)
773 goto out;
774
775 pgd = pgd_offset(mm, addr);
776 if (!pgd_present(*pgd))
777 goto out;
778
779 pud = pud_offset(pgd, addr);
780 if (!pud_present(*pud))
781 goto out;
782
783 pmd = pmd_offset(pud, addr);
784 if (!pmd_present(*pmd))
785 goto out;
786
787 ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
788 if (!pte_same(*ptep, orig_pte)) {
789 pte_unmap_unlock(ptep, ptl);
790 goto out;
791 }
792
Hugh Dickins8dd35572009-12-14 17:59:18 -0800793 get_page(kpage);
Hugh Dickins5ad64682009-12-14 17:59:24 -0800794 page_add_anon_rmap(kpage, vma, addr);
Izik Eidus31dbd012009-09-21 17:02:03 -0700795
796 flush_cache_page(vma, addr, pte_pfn(*ptep));
797 ptep_clear_flush(vma, addr, ptep);
Hugh Dickins8dd35572009-12-14 17:59:18 -0800798 set_pte_at_notify(mm, addr, ptep, mk_pte(kpage, vma->vm_page_prot));
Izik Eidus31dbd012009-09-21 17:02:03 -0700799
Hugh Dickins8dd35572009-12-14 17:59:18 -0800800 page_remove_rmap(page);
801 put_page(page);
Izik Eidus31dbd012009-09-21 17:02:03 -0700802
803 pte_unmap_unlock(ptep, ptl);
804 err = 0;
805out:
806 return err;
807}
808
809/*
810 * try_to_merge_one_page - take two pages and merge them into one
Hugh Dickins8dd35572009-12-14 17:59:18 -0800811 * @vma: the vma that holds the pte pointing to page
812 * @page: the PageAnon page that we want to replace with kpage
Hugh Dickins80e148222009-12-14 17:59:29 -0800813 * @kpage: the PageKsm page that we want to map instead of page,
814 * or NULL the first time when we want to use page as kpage.
Izik Eidus31dbd012009-09-21 17:02:03 -0700815 *
816 * This function returns 0 if the pages were merged, -EFAULT otherwise.
817 */
818static int try_to_merge_one_page(struct vm_area_struct *vma,
Hugh Dickins8dd35572009-12-14 17:59:18 -0800819 struct page *page, struct page *kpage)
Izik Eidus31dbd012009-09-21 17:02:03 -0700820{
821 pte_t orig_pte = __pte(0);
822 int err = -EFAULT;
823
Hugh Dickinsdb114b82009-12-14 17:59:25 -0800824 if (page == kpage) /* ksm page forked */
825 return 0;
826
Izik Eidus31dbd012009-09-21 17:02:03 -0700827 if (!(vma->vm_flags & VM_MERGEABLE))
828 goto out;
Hugh Dickins8dd35572009-12-14 17:59:18 -0800829 if (!PageAnon(page))
Izik Eidus31dbd012009-09-21 17:02:03 -0700830 goto out;
831
Izik Eidus31dbd012009-09-21 17:02:03 -0700832 /*
833 * We need the page lock to read a stable PageSwapCache in
834 * write_protect_page(). We use trylock_page() instead of
835 * lock_page() because we don't want to wait here - we
836 * prefer to continue scanning and merging different pages,
837 * then come back to this page when it is unlocked.
838 */
Hugh Dickins8dd35572009-12-14 17:59:18 -0800839 if (!trylock_page(page))
Hugh Dickins31e855e2009-12-14 17:59:17 -0800840 goto out;
Izik Eidus31dbd012009-09-21 17:02:03 -0700841 /*
842 * If this anonymous page is mapped only here, its pte may need
843 * to be write-protected. If it's mapped elsewhere, all of its
844 * ptes are necessarily already write-protected. But in either
845 * case, we need to lock and check page_count is not raised.
846 */
Hugh Dickins80e148222009-12-14 17:59:29 -0800847 if (write_protect_page(vma, page, &orig_pte) == 0) {
848 if (!kpage) {
849 /*
850 * While we hold page lock, upgrade page from
851 * PageAnon+anon_vma to PageKsm+NULL stable_node:
852 * stable_tree_insert() will update stable_node.
853 */
854 set_page_stable_node(page, NULL);
855 mark_page_accessed(page);
856 err = 0;
857 } else if (pages_identical(page, kpage))
858 err = replace_page(vma, page, kpage, orig_pte);
859 }
Izik Eidus31dbd012009-09-21 17:02:03 -0700860
Hugh Dickins80e148222009-12-14 17:59:29 -0800861 if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
Hugh Dickins73848b42009-12-14 17:59:22 -0800862 munlock_vma_page(page);
Hugh Dickins5ad64682009-12-14 17:59:24 -0800863 if (!PageMlocked(kpage)) {
864 unlock_page(page);
Hugh Dickins5ad64682009-12-14 17:59:24 -0800865 lock_page(kpage);
866 mlock_vma_page(kpage);
867 page = kpage; /* for final unlock */
868 }
869 }
Hugh Dickins73848b42009-12-14 17:59:22 -0800870
Hugh Dickins8dd35572009-12-14 17:59:18 -0800871 unlock_page(page);
Izik Eidus31dbd012009-09-21 17:02:03 -0700872out:
873 return err;
874}
875
876/*
Hugh Dickins81464e302009-09-21 17:02:15 -0700877 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
878 * but no new kernel page is allocated: kpage must already be a ksm page.
Hugh Dickins8dd35572009-12-14 17:59:18 -0800879 *
880 * This function returns 0 if the pages were merged, -EFAULT otherwise.
Hugh Dickins81464e302009-09-21 17:02:15 -0700881 */
Hugh Dickins8dd35572009-12-14 17:59:18 -0800882static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
883 struct page *page, struct page *kpage)
Hugh Dickins81464e302009-09-21 17:02:15 -0700884{
Hugh Dickins8dd35572009-12-14 17:59:18 -0800885 struct mm_struct *mm = rmap_item->mm;
Hugh Dickins81464e302009-09-21 17:02:15 -0700886 struct vm_area_struct *vma;
887 int err = -EFAULT;
888
Hugh Dickins8dd35572009-12-14 17:59:18 -0800889 down_read(&mm->mmap_sem);
890 if (ksm_test_exit(mm))
891 goto out;
892 vma = find_vma(mm, rmap_item->address);
893 if (!vma || vma->vm_start > rmap_item->address)
Hugh Dickins9ba69292009-09-21 17:02:20 -0700894 goto out;
895
Hugh Dickins8dd35572009-12-14 17:59:18 -0800896 err = try_to_merge_one_page(vma, page, kpage);
Hugh Dickinsdb114b82009-12-14 17:59:25 -0800897 if (err)
898 goto out;
899
900 /* Must get reference to anon_vma while still holding mmap_sem */
901 hold_anon_vma(rmap_item, vma->anon_vma);
Hugh Dickins81464e302009-09-21 17:02:15 -0700902out:
Hugh Dickins8dd35572009-12-14 17:59:18 -0800903 up_read(&mm->mmap_sem);
Hugh Dickins81464e302009-09-21 17:02:15 -0700904 return err;
905}
906
907/*
Izik Eidus31dbd012009-09-21 17:02:03 -0700908 * try_to_merge_two_pages - take two identical pages and prepare them
909 * to be merged into one page.
910 *
Hugh Dickins8dd35572009-12-14 17:59:18 -0800911 * This function returns the kpage if we successfully merged two identical
912 * pages into one ksm page, NULL otherwise.
Izik Eidus31dbd012009-09-21 17:02:03 -0700913 *
Hugh Dickins80e148222009-12-14 17:59:29 -0800914 * Note that this function upgrades page to ksm page: if one of the pages
Izik Eidus31dbd012009-09-21 17:02:03 -0700915 * is already a ksm page, try_to_merge_with_ksm_page should be used.
916 */
Hugh Dickins8dd35572009-12-14 17:59:18 -0800917static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
918 struct page *page,
919 struct rmap_item *tree_rmap_item,
920 struct page *tree_page)
Izik Eidus31dbd012009-09-21 17:02:03 -0700921{
Hugh Dickins80e148222009-12-14 17:59:29 -0800922 int err;
Izik Eidus31dbd012009-09-21 17:02:03 -0700923
Hugh Dickins80e148222009-12-14 17:59:29 -0800924 err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
Izik Eidus31dbd012009-09-21 17:02:03 -0700925 if (!err) {
Hugh Dickins8dd35572009-12-14 17:59:18 -0800926 err = try_to_merge_with_ksm_page(tree_rmap_item,
Hugh Dickins80e148222009-12-14 17:59:29 -0800927 tree_page, page);
Izik Eidus31dbd012009-09-21 17:02:03 -0700928 /*
Hugh Dickins81464e302009-09-21 17:02:15 -0700929 * If that fails, we have a ksm page with only one pte
930 * pointing to it: so break it.
Izik Eidus31dbd012009-09-21 17:02:03 -0700931 */
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800932 if (err)
Hugh Dickins8dd35572009-12-14 17:59:18 -0800933 break_cow(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -0700934 }
Hugh Dickins80e148222009-12-14 17:59:29 -0800935 return err ? NULL : page;
Izik Eidus31dbd012009-09-21 17:02:03 -0700936}
937
938/*
Hugh Dickins8dd35572009-12-14 17:59:18 -0800939 * stable_tree_search - search for page inside the stable tree
Izik Eidus31dbd012009-09-21 17:02:03 -0700940 *
941 * This function checks if there is a page inside the stable tree
942 * with identical content to the page that we are scanning right now.
943 *
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800944 * This function returns the stable tree node of identical content if found,
Izik Eidus31dbd012009-09-21 17:02:03 -0700945 * NULL otherwise.
946 */
Hugh Dickins62b61f62009-12-14 17:59:33 -0800947static struct page *stable_tree_search(struct page *page)
Izik Eidus31dbd012009-09-21 17:02:03 -0700948{
949 struct rb_node *node = root_stable_tree.rb_node;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800950 struct stable_node *stable_node;
Izik Eidus31dbd012009-09-21 17:02:03 -0700951
Hugh Dickins08beca42009-12-14 17:59:21 -0800952 stable_node = page_stable_node(page);
953 if (stable_node) { /* ksm page forked */
954 get_page(page);
Hugh Dickins62b61f62009-12-14 17:59:33 -0800955 return page;
Hugh Dickins08beca42009-12-14 17:59:21 -0800956 }
957
Izik Eidus31dbd012009-09-21 17:02:03 -0700958 while (node) {
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800959 struct page *tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -0700960 int ret;
961
Hugh Dickins08beca42009-12-14 17:59:21 -0800962 cond_resched();
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800963 stable_node = rb_entry(node, struct stable_node, node);
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800964 tree_page = get_ksm_page(stable_node);
965 if (!tree_page)
966 return NULL;
Izik Eidus31dbd012009-09-21 17:02:03 -0700967
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800968 ret = memcmp_pages(page, tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -0700969
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800970 if (ret < 0) {
971 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -0700972 node = node->rb_left;
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800973 } else if (ret > 0) {
974 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -0700975 node = node->rb_right;
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800976 } else
Hugh Dickins62b61f62009-12-14 17:59:33 -0800977 return tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -0700978 }
979
980 return NULL;
981}
982
983/*
984 * stable_tree_insert - insert rmap_item pointing to new ksm page
985 * into the stable tree.
986 *
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800987 * This function returns the stable tree node just allocated on success,
988 * NULL otherwise.
Izik Eidus31dbd012009-09-21 17:02:03 -0700989 */
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800990static struct stable_node *stable_tree_insert(struct page *kpage)
Izik Eidus31dbd012009-09-21 17:02:03 -0700991{
992 struct rb_node **new = &root_stable_tree.rb_node;
993 struct rb_node *parent = NULL;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800994 struct stable_node *stable_node;
Izik Eidus31dbd012009-09-21 17:02:03 -0700995
996 while (*new) {
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800997 struct page *tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -0700998 int ret;
999
Hugh Dickins08beca42009-12-14 17:59:21 -08001000 cond_resched();
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001001 stable_node = rb_entry(*new, struct stable_node, node);
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001002 tree_page = get_ksm_page(stable_node);
1003 if (!tree_page)
1004 return NULL;
Izik Eidus31dbd012009-09-21 17:02:03 -07001005
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001006 ret = memcmp_pages(kpage, tree_page);
1007 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001008
1009 parent = *new;
1010 if (ret < 0)
1011 new = &parent->rb_left;
1012 else if (ret > 0)
1013 new = &parent->rb_right;
1014 else {
1015 /*
1016 * It is not a bug that stable_tree_search() didn't
1017 * find this node: because at that time our page was
1018 * not yet write-protected, so may have changed since.
1019 */
1020 return NULL;
1021 }
1022 }
1023
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001024 stable_node = alloc_stable_node();
1025 if (!stable_node)
1026 return NULL;
Izik Eidus31dbd012009-09-21 17:02:03 -07001027
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001028 rb_link_node(&stable_node->node, parent, new);
1029 rb_insert_color(&stable_node->node, &root_stable_tree);
1030
1031 INIT_HLIST_HEAD(&stable_node->hlist);
1032
Hugh Dickins62b61f62009-12-14 17:59:33 -08001033 stable_node->kpfn = page_to_pfn(kpage);
Hugh Dickins08beca42009-12-14 17:59:21 -08001034 set_page_stable_node(kpage, stable_node);
1035
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001036 return stable_node;
Izik Eidus31dbd012009-09-21 17:02:03 -07001037}
1038
1039/*
Hugh Dickins8dd35572009-12-14 17:59:18 -08001040 * unstable_tree_search_insert - search for identical page,
1041 * else insert rmap_item into the unstable tree.
Izik Eidus31dbd012009-09-21 17:02:03 -07001042 *
1043 * This function searches for a page in the unstable tree identical to the
1044 * page currently being scanned; and if no identical page is found in the
1045 * tree, we insert rmap_item as a new object into the unstable tree.
1046 *
1047 * This function returns pointer to rmap_item found to be identical
1048 * to the currently scanned page, NULL otherwise.
1049 *
1050 * This function does both searching and inserting, because they share
1051 * the same walking algorithm in an rbtree.
1052 */
Hugh Dickins8dd35572009-12-14 17:59:18 -08001053static
1054struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1055 struct page *page,
1056 struct page **tree_pagep)
1057
Izik Eidus31dbd012009-09-21 17:02:03 -07001058{
1059 struct rb_node **new = &root_unstable_tree.rb_node;
1060 struct rb_node *parent = NULL;
1061
1062 while (*new) {
1063 struct rmap_item *tree_rmap_item;
Hugh Dickins8dd35572009-12-14 17:59:18 -08001064 struct page *tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001065 int ret;
1066
Hugh Dickinsd178f272009-11-09 15:58:23 +00001067 cond_resched();
Izik Eidus31dbd012009-09-21 17:02:03 -07001068 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
Hugh Dickins8dd35572009-12-14 17:59:18 -08001069 tree_page = get_mergeable_page(tree_rmap_item);
Dan Carpenter22eccdd2010-04-23 13:18:10 -04001070 if (IS_ERR_OR_NULL(tree_page))
Izik Eidus31dbd012009-09-21 17:02:03 -07001071 return NULL;
1072
1073 /*
Hugh Dickins8dd35572009-12-14 17:59:18 -08001074 * Don't substitute a ksm page for a forked page.
Izik Eidus31dbd012009-09-21 17:02:03 -07001075 */
Hugh Dickins8dd35572009-12-14 17:59:18 -08001076 if (page == tree_page) {
1077 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001078 return NULL;
1079 }
1080
Hugh Dickins8dd35572009-12-14 17:59:18 -08001081 ret = memcmp_pages(page, tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001082
1083 parent = *new;
1084 if (ret < 0) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001085 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001086 new = &parent->rb_left;
1087 } else if (ret > 0) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001088 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001089 new = &parent->rb_right;
1090 } else {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001091 *tree_pagep = tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001092 return tree_rmap_item;
1093 }
1094 }
1095
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001096 rmap_item->address |= UNSTABLE_FLAG;
Izik Eidus31dbd012009-09-21 17:02:03 -07001097 rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1098 rb_link_node(&rmap_item->node, parent, new);
1099 rb_insert_color(&rmap_item->node, &root_unstable_tree);
1100
Hugh Dickins473b0ce2009-09-21 17:02:11 -07001101 ksm_pages_unshared++;
Izik Eidus31dbd012009-09-21 17:02:03 -07001102 return NULL;
1103}
1104
1105/*
1106 * stable_tree_append - add another rmap_item to the linked list of
1107 * rmap_items hanging off a given node of the stable tree, all sharing
1108 * the same ksm page.
1109 */
1110static void stable_tree_append(struct rmap_item *rmap_item,
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001111 struct stable_node *stable_node)
Izik Eidus31dbd012009-09-21 17:02:03 -07001112{
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001113 rmap_item->head = stable_node;
Izik Eidus31dbd012009-09-21 17:02:03 -07001114 rmap_item->address |= STABLE_FLAG;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001115 hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
Hugh Dickinse178dfd2009-09-21 17:02:10 -07001116
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001117 if (rmap_item->hlist.next)
1118 ksm_pages_sharing++;
1119 else
1120 ksm_pages_shared++;
Izik Eidus31dbd012009-09-21 17:02:03 -07001121}
1122
1123/*
Hugh Dickins81464e302009-09-21 17:02:15 -07001124 * cmp_and_merge_page - first see if page can be merged into the stable tree;
1125 * if not, compare checksum to previous and if it's the same, see if page can
1126 * be inserted into the unstable tree, or merged with a page already there and
1127 * both transferred to the stable tree.
Izik Eidus31dbd012009-09-21 17:02:03 -07001128 *
1129 * @page: the page that we are searching identical page to.
1130 * @rmap_item: the reverse mapping into the virtual address of this page
1131 */
1132static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1133{
Izik Eidus31dbd012009-09-21 17:02:03 -07001134 struct rmap_item *tree_rmap_item;
Hugh Dickins8dd35572009-12-14 17:59:18 -08001135 struct page *tree_page = NULL;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001136 struct stable_node *stable_node;
Hugh Dickins8dd35572009-12-14 17:59:18 -08001137 struct page *kpage;
Izik Eidus31dbd012009-09-21 17:02:03 -07001138 unsigned int checksum;
1139 int err;
1140
Hugh Dickins93d17712009-12-14 17:59:16 -08001141 remove_rmap_item_from_tree(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -07001142
1143 /* We first start with searching the page inside the stable tree */
Hugh Dickins62b61f62009-12-14 17:59:33 -08001144 kpage = stable_tree_search(page);
1145 if (kpage) {
Hugh Dickins08beca42009-12-14 17:59:21 -08001146 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
Izik Eidus31dbd012009-09-21 17:02:03 -07001147 if (!err) {
1148 /*
1149 * The page was successfully merged:
1150 * add its rmap_item to the stable tree.
1151 */
Hugh Dickins5ad64682009-12-14 17:59:24 -08001152 lock_page(kpage);
Hugh Dickins62b61f62009-12-14 17:59:33 -08001153 stable_tree_append(rmap_item, page_stable_node(kpage));
Hugh Dickins5ad64682009-12-14 17:59:24 -08001154 unlock_page(kpage);
Izik Eidus31dbd012009-09-21 17:02:03 -07001155 }
Hugh Dickins8dd35572009-12-14 17:59:18 -08001156 put_page(kpage);
Izik Eidus31dbd012009-09-21 17:02:03 -07001157 return;
1158 }
1159
1160 /*
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001161 * If the hash value of the page has changed from the last time
1162 * we calculated it, this page is changing frequently: therefore we
1163 * don't want to insert it in the unstable tree, and we don't want
1164 * to waste our time searching for something identical to it there.
Izik Eidus31dbd012009-09-21 17:02:03 -07001165 */
1166 checksum = calc_checksum(page);
1167 if (rmap_item->oldchecksum != checksum) {
1168 rmap_item->oldchecksum = checksum;
1169 return;
1170 }
1171
Hugh Dickins8dd35572009-12-14 17:59:18 -08001172 tree_rmap_item =
1173 unstable_tree_search_insert(rmap_item, page, &tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001174 if (tree_rmap_item) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001175 kpage = try_to_merge_two_pages(rmap_item, page,
1176 tree_rmap_item, tree_page);
1177 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001178 /*
1179 * As soon as we merge this page, we want to remove the
1180 * rmap_item of the page we have merged with from the unstable
1181 * tree, and insert it instead as new node in the stable tree.
1182 */
Hugh Dickins8dd35572009-12-14 17:59:18 -08001183 if (kpage) {
Hugh Dickins93d17712009-12-14 17:59:16 -08001184 remove_rmap_item_from_tree(tree_rmap_item);
Hugh Dickins473b0ce2009-09-21 17:02:11 -07001185
Hugh Dickins5ad64682009-12-14 17:59:24 -08001186 lock_page(kpage);
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001187 stable_node = stable_tree_insert(kpage);
1188 if (stable_node) {
1189 stable_tree_append(tree_rmap_item, stable_node);
1190 stable_tree_append(rmap_item, stable_node);
1191 }
Hugh Dickins5ad64682009-12-14 17:59:24 -08001192 unlock_page(kpage);
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001193
Izik Eidus31dbd012009-09-21 17:02:03 -07001194 /*
1195 * If we fail to insert the page into the stable tree,
1196 * we will have 2 virtual addresses that are pointing
1197 * to a ksm page left outside the stable tree,
1198 * in which case we need to break_cow on both.
1199 */
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001200 if (!stable_node) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001201 break_cow(tree_rmap_item);
1202 break_cow(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -07001203 }
1204 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001205 }
1206}
1207
1208static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
Hugh Dickins6514d512009-12-14 17:59:19 -08001209 struct rmap_item **rmap_list,
Izik Eidus31dbd012009-09-21 17:02:03 -07001210 unsigned long addr)
1211{
1212 struct rmap_item *rmap_item;
1213
Hugh Dickins6514d512009-12-14 17:59:19 -08001214 while (*rmap_list) {
1215 rmap_item = *rmap_list;
Hugh Dickins93d17712009-12-14 17:59:16 -08001216 if ((rmap_item->address & PAGE_MASK) == addr)
Izik Eidus31dbd012009-09-21 17:02:03 -07001217 return rmap_item;
Izik Eidus31dbd012009-09-21 17:02:03 -07001218 if (rmap_item->address > addr)
1219 break;
Hugh Dickins6514d512009-12-14 17:59:19 -08001220 *rmap_list = rmap_item->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -07001221 remove_rmap_item_from_tree(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -07001222 free_rmap_item(rmap_item);
1223 }
1224
1225 rmap_item = alloc_rmap_item();
1226 if (rmap_item) {
1227 /* It has already been zeroed */
1228 rmap_item->mm = mm_slot->mm;
1229 rmap_item->address = addr;
Hugh Dickins6514d512009-12-14 17:59:19 -08001230 rmap_item->rmap_list = *rmap_list;
1231 *rmap_list = rmap_item;
Izik Eidus31dbd012009-09-21 17:02:03 -07001232 }
1233 return rmap_item;
1234}
1235
1236static struct rmap_item *scan_get_next_rmap_item(struct page **page)
1237{
1238 struct mm_struct *mm;
1239 struct mm_slot *slot;
1240 struct vm_area_struct *vma;
1241 struct rmap_item *rmap_item;
1242
1243 if (list_empty(&ksm_mm_head.mm_list))
1244 return NULL;
1245
1246 slot = ksm_scan.mm_slot;
1247 if (slot == &ksm_mm_head) {
1248 root_unstable_tree = RB_ROOT;
1249
1250 spin_lock(&ksm_mmlist_lock);
1251 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1252 ksm_scan.mm_slot = slot;
1253 spin_unlock(&ksm_mmlist_lock);
1254next_mm:
1255 ksm_scan.address = 0;
Hugh Dickins6514d512009-12-14 17:59:19 -08001256 ksm_scan.rmap_list = &slot->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -07001257 }
1258
1259 mm = slot->mm;
1260 down_read(&mm->mmap_sem);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001261 if (ksm_test_exit(mm))
1262 vma = NULL;
1263 else
1264 vma = find_vma(mm, ksm_scan.address);
1265
1266 for (; vma; vma = vma->vm_next) {
Izik Eidus31dbd012009-09-21 17:02:03 -07001267 if (!(vma->vm_flags & VM_MERGEABLE))
1268 continue;
1269 if (ksm_scan.address < vma->vm_start)
1270 ksm_scan.address = vma->vm_start;
1271 if (!vma->anon_vma)
1272 ksm_scan.address = vma->vm_end;
1273
1274 while (ksm_scan.address < vma->vm_end) {
Hugh Dickins9ba69292009-09-21 17:02:20 -07001275 if (ksm_test_exit(mm))
1276 break;
Izik Eidus31dbd012009-09-21 17:02:03 -07001277 *page = follow_page(vma, ksm_scan.address, FOLL_GET);
Dan Carpenter22eccdd2010-04-23 13:18:10 -04001278 if (!IS_ERR_OR_NULL(*page) && PageAnon(*page)) {
Izik Eidus31dbd012009-09-21 17:02:03 -07001279 flush_anon_page(vma, *page, ksm_scan.address);
1280 flush_dcache_page(*page);
1281 rmap_item = get_next_rmap_item(slot,
Hugh Dickins6514d512009-12-14 17:59:19 -08001282 ksm_scan.rmap_list, ksm_scan.address);
Izik Eidus31dbd012009-09-21 17:02:03 -07001283 if (rmap_item) {
Hugh Dickins6514d512009-12-14 17:59:19 -08001284 ksm_scan.rmap_list =
1285 &rmap_item->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -07001286 ksm_scan.address += PAGE_SIZE;
1287 } else
1288 put_page(*page);
1289 up_read(&mm->mmap_sem);
1290 return rmap_item;
1291 }
Dan Carpenter22eccdd2010-04-23 13:18:10 -04001292 if (!IS_ERR_OR_NULL(*page))
Izik Eidus31dbd012009-09-21 17:02:03 -07001293 put_page(*page);
1294 ksm_scan.address += PAGE_SIZE;
1295 cond_resched();
1296 }
1297 }
1298
Hugh Dickins9ba69292009-09-21 17:02:20 -07001299 if (ksm_test_exit(mm)) {
1300 ksm_scan.address = 0;
Hugh Dickins6514d512009-12-14 17:59:19 -08001301 ksm_scan.rmap_list = &slot->rmap_list;
Hugh Dickins9ba69292009-09-21 17:02:20 -07001302 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001303 /*
1304 * Nuke all the rmap_items that are above this current rmap:
1305 * because there were no VM_MERGEABLE vmas with such addresses.
1306 */
Hugh Dickins6514d512009-12-14 17:59:19 -08001307 remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
Izik Eidus31dbd012009-09-21 17:02:03 -07001308
1309 spin_lock(&ksm_mmlist_lock);
Hugh Dickinscd551f92009-09-21 17:02:17 -07001310 ksm_scan.mm_slot = list_entry(slot->mm_list.next,
1311 struct mm_slot, mm_list);
1312 if (ksm_scan.address == 0) {
1313 /*
1314 * We've completed a full scan of all vmas, holding mmap_sem
1315 * throughout, and found no VM_MERGEABLE: so do the same as
1316 * __ksm_exit does to remove this mm from all our lists now.
Hugh Dickins9ba69292009-09-21 17:02:20 -07001317 * This applies either when cleaning up after __ksm_exit
1318 * (but beware: we can reach here even before __ksm_exit),
1319 * or when all VM_MERGEABLE areas have been unmapped (and
1320 * mmap_sem then protects against race with MADV_MERGEABLE).
Hugh Dickinscd551f92009-09-21 17:02:17 -07001321 */
1322 hlist_del(&slot->link);
1323 list_del(&slot->mm_list);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001324 spin_unlock(&ksm_mmlist_lock);
1325
Hugh Dickinscd551f92009-09-21 17:02:17 -07001326 free_mm_slot(slot);
1327 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001328 up_read(&mm->mmap_sem);
1329 mmdrop(mm);
1330 } else {
1331 spin_unlock(&ksm_mmlist_lock);
1332 up_read(&mm->mmap_sem);
Hugh Dickinscd551f92009-09-21 17:02:17 -07001333 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001334
1335 /* Repeat until we've completed scanning the whole list */
Hugh Dickinscd551f92009-09-21 17:02:17 -07001336 slot = ksm_scan.mm_slot;
Izik Eidus31dbd012009-09-21 17:02:03 -07001337 if (slot != &ksm_mm_head)
1338 goto next_mm;
1339
Izik Eidus31dbd012009-09-21 17:02:03 -07001340 ksm_scan.seqnr++;
1341 return NULL;
1342}
1343
1344/**
1345 * ksm_do_scan - the ksm scanner main worker function.
1346 * @scan_npages - number of pages we want to scan before we return.
1347 */
1348static void ksm_do_scan(unsigned int scan_npages)
1349{
1350 struct rmap_item *rmap_item;
Dan Carpenter22eccdd2010-04-23 13:18:10 -04001351 struct page *uninitialized_var(page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001352
1353 while (scan_npages--) {
1354 cond_resched();
1355 rmap_item = scan_get_next_rmap_item(&page);
1356 if (!rmap_item)
1357 return;
1358 if (!PageKsm(page) || !in_stable_tree(rmap_item))
1359 cmp_and_merge_page(page, rmap_item);
1360 put_page(page);
1361 }
1362}
1363
Hugh Dickins6e1583842009-09-21 17:02:14 -07001364static int ksmd_should_run(void)
1365{
1366 return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
1367}
1368
Izik Eidus31dbd012009-09-21 17:02:03 -07001369static int ksm_scan_thread(void *nothing)
1370{
Izik Eidus339aa622009-09-21 17:02:07 -07001371 set_user_nice(current, 5);
Izik Eidus31dbd012009-09-21 17:02:03 -07001372
1373 while (!kthread_should_stop()) {
Hugh Dickins6e1583842009-09-21 17:02:14 -07001374 mutex_lock(&ksm_thread_mutex);
1375 if (ksmd_should_run())
Izik Eidus31dbd012009-09-21 17:02:03 -07001376 ksm_do_scan(ksm_thread_pages_to_scan);
Hugh Dickins6e1583842009-09-21 17:02:14 -07001377 mutex_unlock(&ksm_thread_mutex);
1378
1379 if (ksmd_should_run()) {
Izik Eidus31dbd012009-09-21 17:02:03 -07001380 schedule_timeout_interruptible(
1381 msecs_to_jiffies(ksm_thread_sleep_millisecs));
1382 } else {
1383 wait_event_interruptible(ksm_thread_wait,
Hugh Dickins6e1583842009-09-21 17:02:14 -07001384 ksmd_should_run() || kthread_should_stop());
Izik Eidus31dbd012009-09-21 17:02:03 -07001385 }
1386 }
1387 return 0;
1388}
1389
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001390int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1391 unsigned long end, int advice, unsigned long *vm_flags)
1392{
1393 struct mm_struct *mm = vma->vm_mm;
Hugh Dickinsd952b792009-09-21 17:02:16 -07001394 int err;
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001395
1396 switch (advice) {
1397 case MADV_MERGEABLE:
1398 /*
1399 * Be somewhat over-protective for now!
1400 */
1401 if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE |
1402 VM_PFNMAP | VM_IO | VM_DONTEXPAND |
1403 VM_RESERVED | VM_HUGETLB | VM_INSERTPAGE |
Hugh Dickins5ad64682009-12-14 17:59:24 -08001404 VM_NONLINEAR | VM_MIXEDMAP | VM_SAO))
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001405 return 0; /* just ignore the advice */
1406
Hugh Dickinsd952b792009-09-21 17:02:16 -07001407 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
1408 err = __ksm_enter(mm);
1409 if (err)
1410 return err;
1411 }
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001412
1413 *vm_flags |= VM_MERGEABLE;
1414 break;
1415
1416 case MADV_UNMERGEABLE:
1417 if (!(*vm_flags & VM_MERGEABLE))
1418 return 0; /* just ignore the advice */
1419
Hugh Dickinsd952b792009-09-21 17:02:16 -07001420 if (vma->anon_vma) {
1421 err = unmerge_ksm_pages(vma, start, end);
1422 if (err)
1423 return err;
1424 }
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001425
1426 *vm_flags &= ~VM_MERGEABLE;
1427 break;
1428 }
1429
1430 return 0;
1431}
1432
1433int __ksm_enter(struct mm_struct *mm)
1434{
Hugh Dickins6e1583842009-09-21 17:02:14 -07001435 struct mm_slot *mm_slot;
1436 int needs_wakeup;
1437
1438 mm_slot = alloc_mm_slot();
Izik Eidus31dbd012009-09-21 17:02:03 -07001439 if (!mm_slot)
1440 return -ENOMEM;
1441
Hugh Dickins6e1583842009-09-21 17:02:14 -07001442 /* Check ksm_run too? Would need tighter locking */
1443 needs_wakeup = list_empty(&ksm_mm_head.mm_list);
1444
Izik Eidus31dbd012009-09-21 17:02:03 -07001445 spin_lock(&ksm_mmlist_lock);
1446 insert_to_mm_slots_hash(mm, mm_slot);
1447 /*
1448 * Insert just behind the scanning cursor, to let the area settle
1449 * down a little; when fork is followed by immediate exec, we don't
1450 * want ksmd to waste time setting up and tearing down an rmap_list.
1451 */
1452 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
1453 spin_unlock(&ksm_mmlist_lock);
1454
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001455 set_bit(MMF_VM_MERGEABLE, &mm->flags);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001456 atomic_inc(&mm->mm_count);
Hugh Dickins6e1583842009-09-21 17:02:14 -07001457
1458 if (needs_wakeup)
1459 wake_up_interruptible(&ksm_thread_wait);
1460
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001461 return 0;
1462}
1463
Andrea Arcangeli1c2fb7a2009-09-21 17:02:22 -07001464void __ksm_exit(struct mm_struct *mm)
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001465{
Hugh Dickinscd551f92009-09-21 17:02:17 -07001466 struct mm_slot *mm_slot;
Hugh Dickins9ba69292009-09-21 17:02:20 -07001467 int easy_to_free = 0;
Hugh Dickinscd551f92009-09-21 17:02:17 -07001468
Izik Eidus31dbd012009-09-21 17:02:03 -07001469 /*
Hugh Dickins9ba69292009-09-21 17:02:20 -07001470 * This process is exiting: if it's straightforward (as is the
1471 * case when ksmd was never running), free mm_slot immediately.
1472 * But if it's at the cursor or has rmap_items linked to it, use
1473 * mmap_sem to synchronize with any break_cows before pagetables
1474 * are freed, and leave the mm_slot on the list for ksmd to free.
1475 * Beware: ksm may already have noticed it exiting and freed the slot.
Izik Eidus31dbd012009-09-21 17:02:03 -07001476 */
Hugh Dickins9ba69292009-09-21 17:02:20 -07001477
Hugh Dickinscd551f92009-09-21 17:02:17 -07001478 spin_lock(&ksm_mmlist_lock);
1479 mm_slot = get_mm_slot(mm);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001480 if (mm_slot && ksm_scan.mm_slot != mm_slot) {
Hugh Dickins6514d512009-12-14 17:59:19 -08001481 if (!mm_slot->rmap_list) {
Hugh Dickins9ba69292009-09-21 17:02:20 -07001482 hlist_del(&mm_slot->link);
1483 list_del(&mm_slot->mm_list);
1484 easy_to_free = 1;
1485 } else {
1486 list_move(&mm_slot->mm_list,
1487 &ksm_scan.mm_slot->mm_list);
1488 }
Hugh Dickinscd551f92009-09-21 17:02:17 -07001489 }
Hugh Dickinscd551f92009-09-21 17:02:17 -07001490 spin_unlock(&ksm_mmlist_lock);
1491
Hugh Dickins9ba69292009-09-21 17:02:20 -07001492 if (easy_to_free) {
1493 free_mm_slot(mm_slot);
1494 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1495 mmdrop(mm);
1496 } else if (mm_slot) {
Hugh Dickins9ba69292009-09-21 17:02:20 -07001497 down_write(&mm->mmap_sem);
1498 up_write(&mm->mmap_sem);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001499 }
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001500}
Izik Eidus31dbd012009-09-21 17:02:03 -07001501
Hugh Dickins5ad64682009-12-14 17:59:24 -08001502struct page *ksm_does_need_to_copy(struct page *page,
1503 struct vm_area_struct *vma, unsigned long address)
1504{
1505 struct page *new_page;
1506
Hugh Dickins5ad64682009-12-14 17:59:24 -08001507 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1508 if (new_page) {
1509 copy_user_highpage(new_page, page, address, vma);
1510
1511 SetPageDirty(new_page);
1512 __SetPageUptodate(new_page);
1513 SetPageSwapBacked(new_page);
1514 __set_page_locked(new_page);
1515
1516 if (page_evictable(new_page, vma))
1517 lru_cache_add_lru(new_page, LRU_ACTIVE_ANON);
1518 else
1519 add_page_to_unevictable_list(new_page);
1520 }
1521
Hugh Dickins5ad64682009-12-14 17:59:24 -08001522 return new_page;
1523}
1524
1525int page_referenced_ksm(struct page *page, struct mem_cgroup *memcg,
1526 unsigned long *vm_flags)
1527{
1528 struct stable_node *stable_node;
1529 struct rmap_item *rmap_item;
1530 struct hlist_node *hlist;
1531 unsigned int mapcount = page_mapcount(page);
1532 int referenced = 0;
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001533 int search_new_forks = 0;
Hugh Dickins5ad64682009-12-14 17:59:24 -08001534
1535 VM_BUG_ON(!PageKsm(page));
1536 VM_BUG_ON(!PageLocked(page));
1537
1538 stable_node = page_stable_node(page);
1539 if (!stable_node)
1540 return 0;
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001541again:
Hugh Dickins5ad64682009-12-14 17:59:24 -08001542 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001543 struct anon_vma *anon_vma = rmap_item->anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -08001544 struct anon_vma_chain *vmac;
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001545 struct vm_area_struct *vma;
Hugh Dickins5ad64682009-12-14 17:59:24 -08001546
Rik van Rielcba48b92010-08-09 17:18:38 -07001547 anon_vma_lock(anon_vma);
Rik van Riel5beb4932010-03-05 13:42:07 -08001548 list_for_each_entry(vmac, &anon_vma->head, same_anon_vma) {
1549 vma = vmac->vma;
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001550 if (rmap_item->address < vma->vm_start ||
1551 rmap_item->address >= vma->vm_end)
1552 continue;
1553 /*
1554 * Initially we examine only the vma which covers this
1555 * rmap_item; but later, if there is still work to do,
1556 * we examine covering vmas in other mms: in case they
1557 * were forked from the original since ksmd passed.
1558 */
1559 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1560 continue;
Hugh Dickins5ad64682009-12-14 17:59:24 -08001561
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001562 if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
1563 continue;
1564
1565 referenced += page_referenced_one(page, vma,
Hugh Dickins5ad64682009-12-14 17:59:24 -08001566 rmap_item->address, &mapcount, vm_flags);
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001567 if (!search_new_forks || !mapcount)
1568 break;
1569 }
Rik van Rielcba48b92010-08-09 17:18:38 -07001570 anon_vma_unlock(anon_vma);
Hugh Dickins5ad64682009-12-14 17:59:24 -08001571 if (!mapcount)
1572 goto out;
1573 }
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001574 if (!search_new_forks++)
1575 goto again;
Hugh Dickins5ad64682009-12-14 17:59:24 -08001576out:
Hugh Dickins5ad64682009-12-14 17:59:24 -08001577 return referenced;
1578}
1579
1580int try_to_unmap_ksm(struct page *page, enum ttu_flags flags)
1581{
1582 struct stable_node *stable_node;
1583 struct hlist_node *hlist;
1584 struct rmap_item *rmap_item;
1585 int ret = SWAP_AGAIN;
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001586 int search_new_forks = 0;
Hugh Dickins5ad64682009-12-14 17:59:24 -08001587
1588 VM_BUG_ON(!PageKsm(page));
1589 VM_BUG_ON(!PageLocked(page));
1590
1591 stable_node = page_stable_node(page);
1592 if (!stable_node)
1593 return SWAP_FAIL;
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001594again:
Hugh Dickins5ad64682009-12-14 17:59:24 -08001595 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001596 struct anon_vma *anon_vma = rmap_item->anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -08001597 struct anon_vma_chain *vmac;
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001598 struct vm_area_struct *vma;
Hugh Dickins5ad64682009-12-14 17:59:24 -08001599
Rik van Rielcba48b92010-08-09 17:18:38 -07001600 anon_vma_lock(anon_vma);
Rik van Riel5beb4932010-03-05 13:42:07 -08001601 list_for_each_entry(vmac, &anon_vma->head, same_anon_vma) {
1602 vma = vmac->vma;
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001603 if (rmap_item->address < vma->vm_start ||
1604 rmap_item->address >= vma->vm_end)
1605 continue;
1606 /*
1607 * Initially we examine only the vma which covers this
1608 * rmap_item; but later, if there is still work to do,
1609 * we examine covering vmas in other mms: in case they
1610 * were forked from the original since ksmd passed.
1611 */
1612 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1613 continue;
1614
1615 ret = try_to_unmap_one(page, vma,
1616 rmap_item->address, flags);
1617 if (ret != SWAP_AGAIN || !page_mapped(page)) {
Rik van Rielcba48b92010-08-09 17:18:38 -07001618 anon_vma_unlock(anon_vma);
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001619 goto out;
1620 }
1621 }
Rik van Rielcba48b92010-08-09 17:18:38 -07001622 anon_vma_unlock(anon_vma);
Hugh Dickins5ad64682009-12-14 17:59:24 -08001623 }
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001624 if (!search_new_forks++)
1625 goto again;
Hugh Dickins5ad64682009-12-14 17:59:24 -08001626out:
Hugh Dickins5ad64682009-12-14 17:59:24 -08001627 return ret;
1628}
1629
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001630#ifdef CONFIG_MIGRATION
1631int rmap_walk_ksm(struct page *page, int (*rmap_one)(struct page *,
1632 struct vm_area_struct *, unsigned long, void *), void *arg)
1633{
1634 struct stable_node *stable_node;
1635 struct hlist_node *hlist;
1636 struct rmap_item *rmap_item;
1637 int ret = SWAP_AGAIN;
1638 int search_new_forks = 0;
1639
1640 VM_BUG_ON(!PageKsm(page));
1641 VM_BUG_ON(!PageLocked(page));
1642
1643 stable_node = page_stable_node(page);
1644 if (!stable_node)
1645 return ret;
1646again:
1647 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1648 struct anon_vma *anon_vma = rmap_item->anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -08001649 struct anon_vma_chain *vmac;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001650 struct vm_area_struct *vma;
1651
Rik van Rielcba48b92010-08-09 17:18:38 -07001652 anon_vma_lock(anon_vma);
Rik van Riel5beb4932010-03-05 13:42:07 -08001653 list_for_each_entry(vmac, &anon_vma->head, same_anon_vma) {
1654 vma = vmac->vma;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001655 if (rmap_item->address < vma->vm_start ||
1656 rmap_item->address >= vma->vm_end)
1657 continue;
1658 /*
1659 * Initially we examine only the vma which covers this
1660 * rmap_item; but later, if there is still work to do,
1661 * we examine covering vmas in other mms: in case they
1662 * were forked from the original since ksmd passed.
1663 */
1664 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1665 continue;
1666
1667 ret = rmap_one(page, vma, rmap_item->address, arg);
1668 if (ret != SWAP_AGAIN) {
Rik van Rielcba48b92010-08-09 17:18:38 -07001669 anon_vma_unlock(anon_vma);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001670 goto out;
1671 }
1672 }
Rik van Rielcba48b92010-08-09 17:18:38 -07001673 anon_vma_unlock(anon_vma);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001674 }
1675 if (!search_new_forks++)
1676 goto again;
1677out:
1678 return ret;
1679}
1680
1681void ksm_migrate_page(struct page *newpage, struct page *oldpage)
1682{
1683 struct stable_node *stable_node;
1684
1685 VM_BUG_ON(!PageLocked(oldpage));
1686 VM_BUG_ON(!PageLocked(newpage));
1687 VM_BUG_ON(newpage->mapping != oldpage->mapping);
1688
1689 stable_node = page_stable_node(newpage);
1690 if (stable_node) {
Hugh Dickins62b61f62009-12-14 17:59:33 -08001691 VM_BUG_ON(stable_node->kpfn != page_to_pfn(oldpage));
1692 stable_node->kpfn = page_to_pfn(newpage);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001693 }
1694}
1695#endif /* CONFIG_MIGRATION */
1696
Hugh Dickins62b61f62009-12-14 17:59:33 -08001697#ifdef CONFIG_MEMORY_HOTREMOVE
1698static struct stable_node *ksm_check_stable_tree(unsigned long start_pfn,
1699 unsigned long end_pfn)
1700{
1701 struct rb_node *node;
1702
1703 for (node = rb_first(&root_stable_tree); node; node = rb_next(node)) {
1704 struct stable_node *stable_node;
1705
1706 stable_node = rb_entry(node, struct stable_node, node);
1707 if (stable_node->kpfn >= start_pfn &&
1708 stable_node->kpfn < end_pfn)
1709 return stable_node;
1710 }
1711 return NULL;
1712}
1713
1714static int ksm_memory_callback(struct notifier_block *self,
1715 unsigned long action, void *arg)
1716{
1717 struct memory_notify *mn = arg;
1718 struct stable_node *stable_node;
1719
1720 switch (action) {
1721 case MEM_GOING_OFFLINE:
1722 /*
1723 * Keep it very simple for now: just lock out ksmd and
1724 * MADV_UNMERGEABLE while any memory is going offline.
1725 */
1726 mutex_lock(&ksm_thread_mutex);
1727 break;
1728
1729 case MEM_OFFLINE:
1730 /*
1731 * Most of the work is done by page migration; but there might
1732 * be a few stable_nodes left over, still pointing to struct
1733 * pages which have been offlined: prune those from the tree.
1734 */
1735 while ((stable_node = ksm_check_stable_tree(mn->start_pfn,
1736 mn->start_pfn + mn->nr_pages)) != NULL)
1737 remove_node_from_stable_tree(stable_node);
1738 /* fallthrough */
1739
1740 case MEM_CANCEL_OFFLINE:
1741 mutex_unlock(&ksm_thread_mutex);
1742 break;
1743 }
1744 return NOTIFY_OK;
1745}
1746#endif /* CONFIG_MEMORY_HOTREMOVE */
1747
Hugh Dickins2ffd8672009-09-21 17:02:23 -07001748#ifdef CONFIG_SYSFS
1749/*
1750 * This all compiles without CONFIG_SYSFS, but is a waste of space.
1751 */
1752
Izik Eidus31dbd012009-09-21 17:02:03 -07001753#define KSM_ATTR_RO(_name) \
1754 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1755#define KSM_ATTR(_name) \
1756 static struct kobj_attribute _name##_attr = \
1757 __ATTR(_name, 0644, _name##_show, _name##_store)
1758
1759static ssize_t sleep_millisecs_show(struct kobject *kobj,
1760 struct kobj_attribute *attr, char *buf)
1761{
1762 return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
1763}
1764
1765static ssize_t sleep_millisecs_store(struct kobject *kobj,
1766 struct kobj_attribute *attr,
1767 const char *buf, size_t count)
1768{
1769 unsigned long msecs;
1770 int err;
1771
1772 err = strict_strtoul(buf, 10, &msecs);
1773 if (err || msecs > UINT_MAX)
1774 return -EINVAL;
1775
1776 ksm_thread_sleep_millisecs = msecs;
1777
1778 return count;
1779}
1780KSM_ATTR(sleep_millisecs);
1781
1782static ssize_t pages_to_scan_show(struct kobject *kobj,
1783 struct kobj_attribute *attr, char *buf)
1784{
1785 return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
1786}
1787
1788static ssize_t pages_to_scan_store(struct kobject *kobj,
1789 struct kobj_attribute *attr,
1790 const char *buf, size_t count)
1791{
1792 int err;
1793 unsigned long nr_pages;
1794
1795 err = strict_strtoul(buf, 10, &nr_pages);
1796 if (err || nr_pages > UINT_MAX)
1797 return -EINVAL;
1798
1799 ksm_thread_pages_to_scan = nr_pages;
1800
1801 return count;
1802}
1803KSM_ATTR(pages_to_scan);
1804
1805static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
1806 char *buf)
1807{
1808 return sprintf(buf, "%u\n", ksm_run);
1809}
1810
1811static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
1812 const char *buf, size_t count)
1813{
1814 int err;
1815 unsigned long flags;
1816
1817 err = strict_strtoul(buf, 10, &flags);
1818 if (err || flags > UINT_MAX)
1819 return -EINVAL;
1820 if (flags > KSM_RUN_UNMERGE)
1821 return -EINVAL;
1822
1823 /*
1824 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
1825 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
Hugh Dickinsd0f209f2009-12-14 17:59:34 -08001826 * breaking COW to free the pages_shared (but leaves mm_slots
1827 * on the list for when ksmd may be set running again).
Izik Eidus31dbd012009-09-21 17:02:03 -07001828 */
1829
1830 mutex_lock(&ksm_thread_mutex);
1831 if (ksm_run != flags) {
1832 ksm_run = flags;
Hugh Dickinsd952b792009-09-21 17:02:16 -07001833 if (flags & KSM_RUN_UNMERGE) {
Hugh Dickins35451be2009-09-21 17:02:27 -07001834 current->flags |= PF_OOM_ORIGIN;
Hugh Dickinsd952b792009-09-21 17:02:16 -07001835 err = unmerge_and_remove_all_rmap_items();
Hugh Dickins35451be2009-09-21 17:02:27 -07001836 current->flags &= ~PF_OOM_ORIGIN;
Hugh Dickinsd952b792009-09-21 17:02:16 -07001837 if (err) {
1838 ksm_run = KSM_RUN_STOP;
1839 count = err;
1840 }
1841 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001842 }
1843 mutex_unlock(&ksm_thread_mutex);
1844
1845 if (flags & KSM_RUN_MERGE)
1846 wake_up_interruptible(&ksm_thread_wait);
1847
1848 return count;
1849}
1850KSM_ATTR(run);
1851
Hugh Dickinsb4028262009-09-21 17:02:09 -07001852static ssize_t pages_shared_show(struct kobject *kobj,
1853 struct kobj_attribute *attr, char *buf)
1854{
1855 return sprintf(buf, "%lu\n", ksm_pages_shared);
1856}
1857KSM_ATTR_RO(pages_shared);
1858
1859static ssize_t pages_sharing_show(struct kobject *kobj,
1860 struct kobj_attribute *attr, char *buf)
1861{
Hugh Dickinse178dfd2009-09-21 17:02:10 -07001862 return sprintf(buf, "%lu\n", ksm_pages_sharing);
Hugh Dickinsb4028262009-09-21 17:02:09 -07001863}
1864KSM_ATTR_RO(pages_sharing);
1865
Hugh Dickins473b0ce2009-09-21 17:02:11 -07001866static ssize_t pages_unshared_show(struct kobject *kobj,
1867 struct kobj_attribute *attr, char *buf)
1868{
1869 return sprintf(buf, "%lu\n", ksm_pages_unshared);
1870}
1871KSM_ATTR_RO(pages_unshared);
1872
1873static ssize_t pages_volatile_show(struct kobject *kobj,
1874 struct kobj_attribute *attr, char *buf)
1875{
1876 long ksm_pages_volatile;
1877
1878 ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
1879 - ksm_pages_sharing - ksm_pages_unshared;
1880 /*
1881 * It was not worth any locking to calculate that statistic,
1882 * but it might therefore sometimes be negative: conceal that.
1883 */
1884 if (ksm_pages_volatile < 0)
1885 ksm_pages_volatile = 0;
1886 return sprintf(buf, "%ld\n", ksm_pages_volatile);
1887}
1888KSM_ATTR_RO(pages_volatile);
1889
1890static ssize_t full_scans_show(struct kobject *kobj,
1891 struct kobj_attribute *attr, char *buf)
1892{
1893 return sprintf(buf, "%lu\n", ksm_scan.seqnr);
1894}
1895KSM_ATTR_RO(full_scans);
1896
Izik Eidus31dbd012009-09-21 17:02:03 -07001897static struct attribute *ksm_attrs[] = {
1898 &sleep_millisecs_attr.attr,
1899 &pages_to_scan_attr.attr,
1900 &run_attr.attr,
Hugh Dickinsb4028262009-09-21 17:02:09 -07001901 &pages_shared_attr.attr,
1902 &pages_sharing_attr.attr,
Hugh Dickins473b0ce2009-09-21 17:02:11 -07001903 &pages_unshared_attr.attr,
1904 &pages_volatile_attr.attr,
1905 &full_scans_attr.attr,
Izik Eidus31dbd012009-09-21 17:02:03 -07001906 NULL,
1907};
1908
1909static struct attribute_group ksm_attr_group = {
1910 .attrs = ksm_attrs,
1911 .name = "ksm",
1912};
Hugh Dickins2ffd8672009-09-21 17:02:23 -07001913#endif /* CONFIG_SYSFS */
Izik Eidus31dbd012009-09-21 17:02:03 -07001914
1915static int __init ksm_init(void)
1916{
1917 struct task_struct *ksm_thread;
1918 int err;
1919
1920 err = ksm_slab_init();
1921 if (err)
1922 goto out;
1923
Izik Eidus31dbd012009-09-21 17:02:03 -07001924 ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
1925 if (IS_ERR(ksm_thread)) {
1926 printk(KERN_ERR "ksm: creating kthread failed\n");
1927 err = PTR_ERR(ksm_thread);
Lai Jiangshand9f89842010-08-09 17:20:02 -07001928 goto out_free;
Izik Eidus31dbd012009-09-21 17:02:03 -07001929 }
1930
Hugh Dickins2ffd8672009-09-21 17:02:23 -07001931#ifdef CONFIG_SYSFS
Izik Eidus31dbd012009-09-21 17:02:03 -07001932 err = sysfs_create_group(mm_kobj, &ksm_attr_group);
1933 if (err) {
1934 printk(KERN_ERR "ksm: register sysfs failed\n");
Hugh Dickins2ffd8672009-09-21 17:02:23 -07001935 kthread_stop(ksm_thread);
Lai Jiangshand9f89842010-08-09 17:20:02 -07001936 goto out_free;
Izik Eidus31dbd012009-09-21 17:02:03 -07001937 }
Hugh Dickinsc73602a2009-10-07 16:32:22 -07001938#else
1939 ksm_run = KSM_RUN_MERGE; /* no way for user to start it */
1940
Hugh Dickins2ffd8672009-09-21 17:02:23 -07001941#endif /* CONFIG_SYSFS */
Izik Eidus31dbd012009-09-21 17:02:03 -07001942
Hugh Dickins62b61f62009-12-14 17:59:33 -08001943#ifdef CONFIG_MEMORY_HOTREMOVE
1944 /*
1945 * Choose a high priority since the callback takes ksm_thread_mutex:
1946 * later callbacks could only be taking locks which nest within that.
1947 */
1948 hotplug_memory_notifier(ksm_memory_callback, 100);
1949#endif
Izik Eidus31dbd012009-09-21 17:02:03 -07001950 return 0;
1951
Lai Jiangshand9f89842010-08-09 17:20:02 -07001952out_free:
Izik Eidus31dbd012009-09-21 17:02:03 -07001953 ksm_slab_free();
1954out:
1955 return err;
1956}
1957module_init(ksm_init)