blob: ae539f0b8aa1115ce2ba86ae8b94ec48a79b8333 [file] [log] [blame]
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001/*
Izik Eidus31dbd012009-09-21 17:02:03 -07002 * Memory merging support.
3 *
4 * This code enables dynamic sharing of identical pages found in different
5 * memory areas, even if they are not shared by fork()
6 *
Izik Eidus36b25282009-09-21 17:02:06 -07007 * Copyright (C) 2008-2009 Red Hat, Inc.
Izik Eidus31dbd012009-09-21 17:02:03 -07008 * Authors:
9 * Izik Eidus
10 * Andrea Arcangeli
11 * Chris Wright
Izik Eidus36b25282009-09-21 17:02:06 -070012 * Hugh Dickins
Izik Eidus31dbd012009-09-21 17:02:03 -070013 *
14 * This work is licensed under the terms of the GNU GPL, version 2.
Hugh Dickinsf8af4da2009-09-21 17:01:57 -070015 */
16
17#include <linux/errno.h>
Izik Eidus31dbd012009-09-21 17:02:03 -070018#include <linux/mm.h>
19#include <linux/fs.h>
Hugh Dickinsf8af4da2009-09-21 17:01:57 -070020#include <linux/mman.h>
Izik Eidus31dbd012009-09-21 17:02:03 -070021#include <linux/sched.h>
22#include <linux/rwsem.h>
23#include <linux/pagemap.h>
24#include <linux/rmap.h>
25#include <linux/spinlock.h>
26#include <linux/jhash.h>
27#include <linux/delay.h>
28#include <linux/kthread.h>
29#include <linux/wait.h>
30#include <linux/slab.h>
31#include <linux/rbtree.h>
Hugh Dickins62b61f62009-12-14 17:59:33 -080032#include <linux/memory.h>
Izik Eidus31dbd012009-09-21 17:02:03 -070033#include <linux/mmu_notifier.h>
Izik Eidus2c6854f2009-09-23 15:56:04 -070034#include <linux/swap.h>
Hugh Dickinsf8af4da2009-09-21 17:01:57 -070035#include <linux/ksm.h>
Lai Jiangshand9f89842010-08-09 17:20:02 -070036#include <linux/hash.h>
Andrea Arcangeli878aee72011-01-13 15:47:10 -080037#include <linux/freezer.h>
David Rientjes72788c32011-05-24 17:11:40 -070038#include <linux/oom.h>
Hugh Dickinsf8af4da2009-09-21 17:01:57 -070039
Izik Eidus31dbd012009-09-21 17:02:03 -070040#include <asm/tlbflush.h>
Hugh Dickins73848b42009-12-14 17:59:22 -080041#include "internal.h"
Izik Eidus31dbd012009-09-21 17:02:03 -070042
43/*
44 * A few notes about the KSM scanning process,
45 * to make it easier to understand the data structures below:
46 *
47 * In order to reduce excessive scanning, KSM sorts the memory pages by their
48 * contents into a data structure that holds pointers to the pages' locations.
49 *
50 * Since the contents of the pages may change at any moment, KSM cannot just
51 * insert the pages into a normal sorted tree and expect it to find anything.
52 * Therefore KSM uses two data structures - the stable and the unstable tree.
53 *
54 * The stable tree holds pointers to all the merged pages (ksm pages), sorted
55 * by their contents. Because each such page is write-protected, searching on
56 * this tree is fully assured to be working (except when pages are unmapped),
57 * and therefore this tree is called the stable tree.
58 *
59 * In addition to the stable tree, KSM uses a second data structure called the
60 * unstable tree: this tree holds pointers to pages which have been found to
61 * be "unchanged for a period of time". The unstable tree sorts these pages
62 * by their contents, but since they are not write-protected, KSM cannot rely
63 * upon the unstable tree to work correctly - the unstable tree is liable to
64 * be corrupted as its contents are modified, and so it is called unstable.
65 *
66 * KSM solves this problem by several techniques:
67 *
68 * 1) The unstable tree is flushed every time KSM completes scanning all
69 * memory areas, and then the tree is rebuilt again from the beginning.
70 * 2) KSM will only insert into the unstable tree, pages whose hash value
71 * has not changed since the previous scan of all memory areas.
72 * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
73 * colors of the nodes and not on their contents, assuring that even when
74 * the tree gets "corrupted" it won't get out of balance, so scanning time
75 * remains the same (also, searching and inserting nodes in an rbtree uses
76 * the same algorithm, so we have no overhead when we flush and rebuild).
77 * 4) KSM never flushes the stable tree, which means that even if it were to
78 * take 10 attempts to find a page in the unstable tree, once it is found,
79 * it is secured in the stable tree. (When we scan a new page, we first
80 * compare it against the stable tree, and then against the unstable tree.)
81 */
82
83/**
84 * struct mm_slot - ksm information per mm that is being scanned
85 * @link: link to the mm_slots hash list
86 * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
Hugh Dickins6514d512009-12-14 17:59:19 -080087 * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
Izik Eidus31dbd012009-09-21 17:02:03 -070088 * @mm: the mm that this information is valid for
89 */
90struct mm_slot {
91 struct hlist_node link;
92 struct list_head mm_list;
Hugh Dickins6514d512009-12-14 17:59:19 -080093 struct rmap_item *rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -070094 struct mm_struct *mm;
95};
96
97/**
98 * struct ksm_scan - cursor for scanning
99 * @mm_slot: the current mm_slot we are scanning
100 * @address: the next address inside that to be scanned
Hugh Dickins6514d512009-12-14 17:59:19 -0800101 * @rmap_list: link to the next rmap to be scanned in the rmap_list
Izik Eidus31dbd012009-09-21 17:02:03 -0700102 * @seqnr: count of completed full scans (needed when removing unstable node)
103 *
104 * There is only the one ksm_scan instance of this cursor structure.
105 */
106struct ksm_scan {
107 struct mm_slot *mm_slot;
108 unsigned long address;
Hugh Dickins6514d512009-12-14 17:59:19 -0800109 struct rmap_item **rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -0700110 unsigned long seqnr;
111};
112
113/**
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800114 * struct stable_node - node of the stable rbtree
115 * @node: rb node of this ksm page in the stable tree
116 * @hlist: hlist head of rmap_items using this ksm page
Hugh Dickins62b61f62009-12-14 17:59:33 -0800117 * @kpfn: page frame number of this ksm page
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800118 */
119struct stable_node {
120 struct rb_node node;
121 struct hlist_head hlist;
Hugh Dickins62b61f62009-12-14 17:59:33 -0800122 unsigned long kpfn;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800123};
124
125/**
Izik Eidus31dbd012009-09-21 17:02:03 -0700126 * struct rmap_item - reverse mapping item for virtual addresses
Hugh Dickins6514d512009-12-14 17:59:19 -0800127 * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
Hugh Dickinsdb114b82009-12-14 17:59:25 -0800128 * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
Izik Eidus31dbd012009-09-21 17:02:03 -0700129 * @mm: the memory structure this rmap_item is pointing into
130 * @address: the virtual address this rmap_item tracks (+ flags in low bits)
131 * @oldchecksum: previous checksum of the page at that virtual address
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800132 * @node: rb node of this rmap_item in the unstable tree
133 * @head: pointer to stable_node heading this list in the stable tree
134 * @hlist: link into hlist of rmap_items hanging off that stable_node
Izik Eidus31dbd012009-09-21 17:02:03 -0700135 */
136struct rmap_item {
Hugh Dickins6514d512009-12-14 17:59:19 -0800137 struct rmap_item *rmap_list;
Hugh Dickinsdb114b82009-12-14 17:59:25 -0800138 struct anon_vma *anon_vma; /* when stable */
Izik Eidus31dbd012009-09-21 17:02:03 -0700139 struct mm_struct *mm;
140 unsigned long address; /* + low bits used for flags below */
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800141 unsigned int oldchecksum; /* when unstable */
Izik Eidus31dbd012009-09-21 17:02:03 -0700142 union {
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800143 struct rb_node node; /* when node of unstable tree */
144 struct { /* when listed from stable tree */
145 struct stable_node *head;
146 struct hlist_node hlist;
147 };
Izik Eidus31dbd012009-09-21 17:02:03 -0700148 };
149};
150
151#define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800152#define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */
153#define STABLE_FLAG 0x200 /* is listed from the stable tree */
Izik Eidus31dbd012009-09-21 17:02:03 -0700154
155/* The stable and unstable tree heads */
156static struct rb_root root_stable_tree = RB_ROOT;
157static struct rb_root root_unstable_tree = RB_ROOT;
158
Lai Jiangshand9f89842010-08-09 17:20:02 -0700159#define MM_SLOTS_HASH_SHIFT 10
160#define MM_SLOTS_HASH_HEADS (1 << MM_SLOTS_HASH_SHIFT)
161static struct hlist_head mm_slots_hash[MM_SLOTS_HASH_HEADS];
Izik Eidus31dbd012009-09-21 17:02:03 -0700162
163static struct mm_slot ksm_mm_head = {
164 .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
165};
166static struct ksm_scan ksm_scan = {
167 .mm_slot = &ksm_mm_head,
168};
169
170static struct kmem_cache *rmap_item_cache;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800171static struct kmem_cache *stable_node_cache;
Izik Eidus31dbd012009-09-21 17:02:03 -0700172static struct kmem_cache *mm_slot_cache;
173
174/* The number of nodes in the stable tree */
Hugh Dickinsb4028262009-09-21 17:02:09 -0700175static unsigned long ksm_pages_shared;
Izik Eidus31dbd012009-09-21 17:02:03 -0700176
Hugh Dickinse178dfd2009-09-21 17:02:10 -0700177/* The number of page slots additionally sharing those nodes */
Hugh Dickinsb4028262009-09-21 17:02:09 -0700178static unsigned long ksm_pages_sharing;
Izik Eidus31dbd012009-09-21 17:02:03 -0700179
Hugh Dickins473b0ce2009-09-21 17:02:11 -0700180/* The number of nodes in the unstable tree */
181static unsigned long ksm_pages_unshared;
182
183/* The number of rmap_items in use: to calculate pages_volatile */
184static unsigned long ksm_rmap_items;
185
Izik Eidus31dbd012009-09-21 17:02:03 -0700186/* Number of pages ksmd should scan in one batch */
Izik Eidus2c6854f2009-09-23 15:56:04 -0700187static unsigned int ksm_thread_pages_to_scan = 100;
Izik Eidus31dbd012009-09-21 17:02:03 -0700188
189/* Milliseconds ksmd should sleep between batches */
Hugh Dickins2ffd8672009-09-21 17:02:23 -0700190static unsigned int ksm_thread_sleep_millisecs = 20;
Izik Eidus31dbd012009-09-21 17:02:03 -0700191
192#define KSM_RUN_STOP 0
193#define KSM_RUN_MERGE 1
194#define KSM_RUN_UNMERGE 2
Izik Eidus2c6854f2009-09-23 15:56:04 -0700195static unsigned int ksm_run = KSM_RUN_STOP;
Izik Eidus31dbd012009-09-21 17:02:03 -0700196
197static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
198static DEFINE_MUTEX(ksm_thread_mutex);
199static DEFINE_SPINLOCK(ksm_mmlist_lock);
200
201#define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
202 sizeof(struct __struct), __alignof__(struct __struct),\
203 (__flags), NULL)
204
205static int __init ksm_slab_init(void)
206{
207 rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
208 if (!rmap_item_cache)
209 goto out;
210
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800211 stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
212 if (!stable_node_cache)
213 goto out_free1;
214
Izik Eidus31dbd012009-09-21 17:02:03 -0700215 mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
216 if (!mm_slot_cache)
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800217 goto out_free2;
Izik Eidus31dbd012009-09-21 17:02:03 -0700218
219 return 0;
220
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800221out_free2:
222 kmem_cache_destroy(stable_node_cache);
223out_free1:
Izik Eidus31dbd012009-09-21 17:02:03 -0700224 kmem_cache_destroy(rmap_item_cache);
225out:
226 return -ENOMEM;
227}
228
229static void __init ksm_slab_free(void)
230{
231 kmem_cache_destroy(mm_slot_cache);
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800232 kmem_cache_destroy(stable_node_cache);
Izik Eidus31dbd012009-09-21 17:02:03 -0700233 kmem_cache_destroy(rmap_item_cache);
234 mm_slot_cache = NULL;
235}
236
237static inline struct rmap_item *alloc_rmap_item(void)
238{
Hugh Dickins473b0ce2009-09-21 17:02:11 -0700239 struct rmap_item *rmap_item;
240
241 rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL);
242 if (rmap_item)
243 ksm_rmap_items++;
244 return rmap_item;
Izik Eidus31dbd012009-09-21 17:02:03 -0700245}
246
247static inline void free_rmap_item(struct rmap_item *rmap_item)
248{
Hugh Dickins473b0ce2009-09-21 17:02:11 -0700249 ksm_rmap_items--;
Izik Eidus31dbd012009-09-21 17:02:03 -0700250 rmap_item->mm = NULL; /* debug safety */
251 kmem_cache_free(rmap_item_cache, rmap_item);
252}
253
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800254static inline struct stable_node *alloc_stable_node(void)
255{
256 return kmem_cache_alloc(stable_node_cache, GFP_KERNEL);
257}
258
259static inline void free_stable_node(struct stable_node *stable_node)
260{
261 kmem_cache_free(stable_node_cache, stable_node);
262}
263
Izik Eidus31dbd012009-09-21 17:02:03 -0700264static inline struct mm_slot *alloc_mm_slot(void)
265{
266 if (!mm_slot_cache) /* initialization failed */
267 return NULL;
268 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
269}
270
271static inline void free_mm_slot(struct mm_slot *mm_slot)
272{
273 kmem_cache_free(mm_slot_cache, mm_slot);
274}
275
Izik Eidus31dbd012009-09-21 17:02:03 -0700276static struct mm_slot *get_mm_slot(struct mm_struct *mm)
277{
278 struct mm_slot *mm_slot;
279 struct hlist_head *bucket;
280 struct hlist_node *node;
281
Lai Jiangshand9f89842010-08-09 17:20:02 -0700282 bucket = &mm_slots_hash[hash_ptr(mm, MM_SLOTS_HASH_SHIFT)];
Izik Eidus31dbd012009-09-21 17:02:03 -0700283 hlist_for_each_entry(mm_slot, node, bucket, link) {
284 if (mm == mm_slot->mm)
285 return mm_slot;
286 }
287 return NULL;
288}
289
290static void insert_to_mm_slots_hash(struct mm_struct *mm,
291 struct mm_slot *mm_slot)
292{
293 struct hlist_head *bucket;
294
Lai Jiangshand9f89842010-08-09 17:20:02 -0700295 bucket = &mm_slots_hash[hash_ptr(mm, MM_SLOTS_HASH_SHIFT)];
Izik Eidus31dbd012009-09-21 17:02:03 -0700296 mm_slot->mm = mm;
Izik Eidus31dbd012009-09-21 17:02:03 -0700297 hlist_add_head(&mm_slot->link, bucket);
298}
299
300static inline int in_stable_tree(struct rmap_item *rmap_item)
301{
302 return rmap_item->address & STABLE_FLAG;
303}
304
305/*
Hugh Dickinsa913e182009-09-21 17:02:26 -0700306 * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
307 * page tables after it has passed through ksm_exit() - which, if necessary,
308 * takes mmap_sem briefly to serialize against them. ksm_exit() does not set
309 * a special flag: they can just back out as soon as mm_users goes to zero.
310 * ksm_test_exit() is used throughout to make this test for exit: in some
311 * places for correctness, in some places just to avoid unnecessary work.
312 */
313static inline bool ksm_test_exit(struct mm_struct *mm)
314{
315 return atomic_read(&mm->mm_users) == 0;
316}
317
318/*
Izik Eidus31dbd012009-09-21 17:02:03 -0700319 * We use break_ksm to break COW on a ksm page: it's a stripped down
320 *
321 * if (get_user_pages(current, mm, addr, 1, 1, 1, &page, NULL) == 1)
322 * put_page(page);
323 *
324 * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
325 * in case the application has unmapped and remapped mm,addr meanwhile.
326 * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP
327 * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
328 */
Hugh Dickinsd952b792009-09-21 17:02:16 -0700329static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
Izik Eidus31dbd012009-09-21 17:02:03 -0700330{
331 struct page *page;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700332 int ret = 0;
Izik Eidus31dbd012009-09-21 17:02:03 -0700333
334 do {
335 cond_resched();
336 page = follow_page(vma, addr, FOLL_GET);
Dan Carpenter22eccdd2010-04-23 13:18:10 -0400337 if (IS_ERR_OR_NULL(page))
Izik Eidus31dbd012009-09-21 17:02:03 -0700338 break;
339 if (PageKsm(page))
340 ret = handle_mm_fault(vma->vm_mm, vma, addr,
341 FAULT_FLAG_WRITE);
342 else
343 ret = VM_FAULT_WRITE;
344 put_page(page);
Hugh Dickinsd952b792009-09-21 17:02:16 -0700345 } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_OOM)));
346 /*
347 * We must loop because handle_mm_fault() may back out if there's
348 * any difficulty e.g. if pte accessed bit gets updated concurrently.
349 *
350 * VM_FAULT_WRITE is what we have been hoping for: it indicates that
351 * COW has been broken, even if the vma does not permit VM_WRITE;
352 * but note that a concurrent fault might break PageKsm for us.
353 *
354 * VM_FAULT_SIGBUS could occur if we race with truncation of the
355 * backing file, which also invalidates anonymous pages: that's
356 * okay, that truncation will have unmapped the PageKsm for us.
357 *
358 * VM_FAULT_OOM: at the time of writing (late July 2009), setting
359 * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
360 * current task has TIF_MEMDIE set, and will be OOM killed on return
361 * to user; and ksmd, having no mm, would never be chosen for that.
362 *
363 * But if the mm is in a limited mem_cgroup, then the fault may fail
364 * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
365 * even ksmd can fail in this way - though it's usually breaking ksm
366 * just to undo a merge it made a moment before, so unlikely to oom.
367 *
368 * That's a pity: we might therefore have more kernel pages allocated
369 * than we're counting as nodes in the stable tree; but ksm_do_scan
370 * will retry to break_cow on each pass, so should recover the page
371 * in due course. The important thing is to not let VM_MERGEABLE
372 * be cleared while any such pages might remain in the area.
373 */
374 return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
Izik Eidus31dbd012009-09-21 17:02:03 -0700375}
376
Bob Liuef694222012-03-21 16:34:11 -0700377static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm,
378 unsigned long addr)
379{
380 struct vm_area_struct *vma;
381 if (ksm_test_exit(mm))
382 return NULL;
383 vma = find_vma(mm, addr);
384 if (!vma || vma->vm_start > addr)
385 return NULL;
386 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
387 return NULL;
388 return vma;
389}
390
Hugh Dickins8dd35572009-12-14 17:59:18 -0800391static void break_cow(struct rmap_item *rmap_item)
Izik Eidus31dbd012009-09-21 17:02:03 -0700392{
Hugh Dickins8dd35572009-12-14 17:59:18 -0800393 struct mm_struct *mm = rmap_item->mm;
394 unsigned long addr = rmap_item->address;
Izik Eidus31dbd012009-09-21 17:02:03 -0700395 struct vm_area_struct *vma;
396
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800397 /*
398 * It is not an accident that whenever we want to break COW
399 * to undo, we also need to drop a reference to the anon_vma.
400 */
Peter Zijlstra9e601092011-03-22 16:32:46 -0700401 put_anon_vma(rmap_item->anon_vma);
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800402
Hugh Dickins81464e302009-09-21 17:02:15 -0700403 down_read(&mm->mmap_sem);
Bob Liuef694222012-03-21 16:34:11 -0700404 vma = find_mergeable_vma(mm, addr);
405 if (vma)
406 break_ksm(vma, addr);
Izik Eidus31dbd012009-09-21 17:02:03 -0700407 up_read(&mm->mmap_sem);
408}
409
Andrea Arcangeli29ad7682011-01-13 15:47:19 -0800410static struct page *page_trans_compound_anon(struct page *page)
411{
412 if (PageTransCompound(page)) {
Andrea Arcangeli22e5c472011-01-13 15:47:20 -0800413 struct page *head = compound_trans_head(page);
Andrea Arcangeli29ad7682011-01-13 15:47:19 -0800414 /*
Andrea Arcangeli22e5c472011-01-13 15:47:20 -0800415 * head may actually be splitted and freed from under
416 * us but it's ok here.
Andrea Arcangeli29ad7682011-01-13 15:47:19 -0800417 */
Andrea Arcangeli29ad7682011-01-13 15:47:19 -0800418 if (PageAnon(head))
419 return head;
420 }
421 return NULL;
422}
423
Izik Eidus31dbd012009-09-21 17:02:03 -0700424static struct page *get_mergeable_page(struct rmap_item *rmap_item)
425{
426 struct mm_struct *mm = rmap_item->mm;
427 unsigned long addr = rmap_item->address;
428 struct vm_area_struct *vma;
429 struct page *page;
430
431 down_read(&mm->mmap_sem);
Bob Liuef694222012-03-21 16:34:11 -0700432 vma = find_mergeable_vma(mm, addr);
433 if (!vma)
Izik Eidus31dbd012009-09-21 17:02:03 -0700434 goto out;
435
436 page = follow_page(vma, addr, FOLL_GET);
Dan Carpenter22eccdd2010-04-23 13:18:10 -0400437 if (IS_ERR_OR_NULL(page))
Izik Eidus31dbd012009-09-21 17:02:03 -0700438 goto out;
Andrea Arcangeli29ad7682011-01-13 15:47:19 -0800439 if (PageAnon(page) || page_trans_compound_anon(page)) {
Izik Eidus31dbd012009-09-21 17:02:03 -0700440 flush_anon_page(vma, page, addr);
441 flush_dcache_page(page);
442 } else {
443 put_page(page);
444out: page = NULL;
445 }
446 up_read(&mm->mmap_sem);
447 return page;
448}
449
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800450static void remove_node_from_stable_tree(struct stable_node *stable_node)
451{
452 struct rmap_item *rmap_item;
453 struct hlist_node *hlist;
454
455 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
456 if (rmap_item->hlist.next)
457 ksm_pages_sharing--;
458 else
459 ksm_pages_shared--;
Peter Zijlstra9e601092011-03-22 16:32:46 -0700460 put_anon_vma(rmap_item->anon_vma);
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800461 rmap_item->address &= PAGE_MASK;
462 cond_resched();
463 }
464
465 rb_erase(&stable_node->node, &root_stable_tree);
466 free_stable_node(stable_node);
467}
468
469/*
470 * get_ksm_page: checks if the page indicated by the stable node
471 * is still its ksm page, despite having held no reference to it.
472 * In which case we can trust the content of the page, and it
473 * returns the gotten page; but if the page has now been zapped,
474 * remove the stale node from the stable tree and return NULL.
475 *
476 * You would expect the stable_node to hold a reference to the ksm page.
477 * But if it increments the page's count, swapping out has to wait for
478 * ksmd to come around again before it can free the page, which may take
479 * seconds or even minutes: much too unresponsive. So instead we use a
480 * "keyhole reference": access to the ksm page from the stable node peeps
481 * out through its keyhole to see if that page still holds the right key,
482 * pointing back to this stable node. This relies on freeing a PageAnon
483 * page to reset its page->mapping to NULL, and relies on no other use of
484 * a page to put something that might look like our key in page->mapping.
485 *
486 * include/linux/pagemap.h page_cache_get_speculative() is a good reference,
487 * but this is different - made simpler by ksm_thread_mutex being held, but
488 * interesting for assuming that no other use of the struct page could ever
489 * put our expected_mapping into page->mapping (or a field of the union which
490 * coincides with page->mapping). The RCU calls are not for KSM at all, but
491 * to keep the page_count protocol described with page_cache_get_speculative.
492 *
493 * Note: it is possible that get_ksm_page() will return NULL one moment,
494 * then page the next, if the page is in between page_freeze_refs() and
495 * page_unfreeze_refs(): this shouldn't be a problem anywhere, the page
496 * is on its way to being freed; but it is an anomaly to bear in mind.
497 */
498static struct page *get_ksm_page(struct stable_node *stable_node)
499{
500 struct page *page;
501 void *expected_mapping;
502
Hugh Dickins62b61f62009-12-14 17:59:33 -0800503 page = pfn_to_page(stable_node->kpfn);
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800504 expected_mapping = (void *)stable_node +
505 (PAGE_MAPPING_ANON | PAGE_MAPPING_KSM);
506 rcu_read_lock();
507 if (page->mapping != expected_mapping)
508 goto stale;
509 if (!get_page_unless_zero(page))
510 goto stale;
511 if (page->mapping != expected_mapping) {
512 put_page(page);
513 goto stale;
514 }
515 rcu_read_unlock();
516 return page;
517stale:
518 rcu_read_unlock();
519 remove_node_from_stable_tree(stable_node);
520 return NULL;
521}
522
Izik Eidus31dbd012009-09-21 17:02:03 -0700523/*
Izik Eidus31dbd012009-09-21 17:02:03 -0700524 * Removing rmap_item from stable or unstable tree.
525 * This function will clean the information from the stable/unstable tree.
526 */
527static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
528{
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800529 if (rmap_item->address & STABLE_FLAG) {
530 struct stable_node *stable_node;
Hugh Dickins5ad64682009-12-14 17:59:24 -0800531 struct page *page;
Izik Eidus31dbd012009-09-21 17:02:03 -0700532
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800533 stable_node = rmap_item->head;
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800534 page = get_ksm_page(stable_node);
535 if (!page)
536 goto out;
537
Hugh Dickins5ad64682009-12-14 17:59:24 -0800538 lock_page(page);
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800539 hlist_del(&rmap_item->hlist);
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800540 unlock_page(page);
541 put_page(page);
Hugh Dickins08beca42009-12-14 17:59:21 -0800542
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800543 if (stable_node->hlist.first)
544 ksm_pages_sharing--;
545 else
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800546 ksm_pages_shared--;
Izik Eidus31dbd012009-09-21 17:02:03 -0700547
Peter Zijlstra9e601092011-03-22 16:32:46 -0700548 put_anon_vma(rmap_item->anon_vma);
Hugh Dickins93d17712009-12-14 17:59:16 -0800549 rmap_item->address &= PAGE_MASK;
Izik Eidus31dbd012009-09-21 17:02:03 -0700550
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800551 } else if (rmap_item->address & UNSTABLE_FLAG) {
Izik Eidus31dbd012009-09-21 17:02:03 -0700552 unsigned char age;
553 /*
Hugh Dickins9ba69292009-09-21 17:02:20 -0700554 * Usually ksmd can and must skip the rb_erase, because
Izik Eidus31dbd012009-09-21 17:02:03 -0700555 * root_unstable_tree was already reset to RB_ROOT.
Hugh Dickins9ba69292009-09-21 17:02:20 -0700556 * But be careful when an mm is exiting: do the rb_erase
557 * if this rmap_item was inserted by this scan, rather
558 * than left over from before.
Izik Eidus31dbd012009-09-21 17:02:03 -0700559 */
560 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
Hugh Dickinscd551f92009-09-21 17:02:17 -0700561 BUG_ON(age > 1);
Izik Eidus31dbd012009-09-21 17:02:03 -0700562 if (!age)
563 rb_erase(&rmap_item->node, &root_unstable_tree);
Izik Eidus31dbd012009-09-21 17:02:03 -0700564
Hugh Dickins93d17712009-12-14 17:59:16 -0800565 ksm_pages_unshared--;
566 rmap_item->address &= PAGE_MASK;
567 }
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800568out:
Izik Eidus31dbd012009-09-21 17:02:03 -0700569 cond_resched(); /* we're called from many long loops */
570}
571
Izik Eidus31dbd012009-09-21 17:02:03 -0700572static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
Hugh Dickins6514d512009-12-14 17:59:19 -0800573 struct rmap_item **rmap_list)
Izik Eidus31dbd012009-09-21 17:02:03 -0700574{
Hugh Dickins6514d512009-12-14 17:59:19 -0800575 while (*rmap_list) {
576 struct rmap_item *rmap_item = *rmap_list;
577 *rmap_list = rmap_item->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -0700578 remove_rmap_item_from_tree(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -0700579 free_rmap_item(rmap_item);
580 }
581}
582
583/*
584 * Though it's very tempting to unmerge in_stable_tree(rmap_item)s rather
585 * than check every pte of a given vma, the locking doesn't quite work for
586 * that - an rmap_item is assigned to the stable tree after inserting ksm
587 * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing
588 * rmap_items from parent to child at fork time (so as not to waste time
589 * if exit comes before the next scan reaches it).
Hugh Dickins81464e302009-09-21 17:02:15 -0700590 *
591 * Similarly, although we'd like to remove rmap_items (so updating counts
592 * and freeing memory) when unmerging an area, it's easier to leave that
593 * to the next pass of ksmd - consider, for example, how ksmd might be
594 * in cmp_and_merge_page on one of the rmap_items we would be removing.
Izik Eidus31dbd012009-09-21 17:02:03 -0700595 */
Hugh Dickinsd952b792009-09-21 17:02:16 -0700596static int unmerge_ksm_pages(struct vm_area_struct *vma,
597 unsigned long start, unsigned long end)
Izik Eidus31dbd012009-09-21 17:02:03 -0700598{
599 unsigned long addr;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700600 int err = 0;
Izik Eidus31dbd012009-09-21 17:02:03 -0700601
Hugh Dickinsd952b792009-09-21 17:02:16 -0700602 for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
Hugh Dickins9ba69292009-09-21 17:02:20 -0700603 if (ksm_test_exit(vma->vm_mm))
604 break;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700605 if (signal_pending(current))
606 err = -ERESTARTSYS;
607 else
608 err = break_ksm(vma, addr);
609 }
610 return err;
Izik Eidus31dbd012009-09-21 17:02:03 -0700611}
612
Hugh Dickins2ffd8672009-09-21 17:02:23 -0700613#ifdef CONFIG_SYSFS
614/*
615 * Only called through the sysfs control interface:
616 */
Hugh Dickinsd952b792009-09-21 17:02:16 -0700617static int unmerge_and_remove_all_rmap_items(void)
Izik Eidus31dbd012009-09-21 17:02:03 -0700618{
619 struct mm_slot *mm_slot;
620 struct mm_struct *mm;
621 struct vm_area_struct *vma;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700622 int err = 0;
Izik Eidus31dbd012009-09-21 17:02:03 -0700623
Hugh Dickinsd952b792009-09-21 17:02:16 -0700624 spin_lock(&ksm_mmlist_lock);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700625 ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
Hugh Dickinsd952b792009-09-21 17:02:16 -0700626 struct mm_slot, mm_list);
627 spin_unlock(&ksm_mmlist_lock);
628
Hugh Dickins9ba69292009-09-21 17:02:20 -0700629 for (mm_slot = ksm_scan.mm_slot;
630 mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
Izik Eidus31dbd012009-09-21 17:02:03 -0700631 mm = mm_slot->mm;
632 down_read(&mm->mmap_sem);
633 for (vma = mm->mmap; vma; vma = vma->vm_next) {
Hugh Dickins9ba69292009-09-21 17:02:20 -0700634 if (ksm_test_exit(mm))
635 break;
Izik Eidus31dbd012009-09-21 17:02:03 -0700636 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
637 continue;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700638 err = unmerge_ksm_pages(vma,
639 vma->vm_start, vma->vm_end);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700640 if (err)
641 goto error;
Izik Eidus31dbd012009-09-21 17:02:03 -0700642 }
Hugh Dickins9ba69292009-09-21 17:02:20 -0700643
Hugh Dickins6514d512009-12-14 17:59:19 -0800644 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
Hugh Dickinsd952b792009-09-21 17:02:16 -0700645
646 spin_lock(&ksm_mmlist_lock);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700647 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
Hugh Dickinsd952b792009-09-21 17:02:16 -0700648 struct mm_slot, mm_list);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700649 if (ksm_test_exit(mm)) {
650 hlist_del(&mm_slot->link);
651 list_del(&mm_slot->mm_list);
652 spin_unlock(&ksm_mmlist_lock);
653
654 free_mm_slot(mm_slot);
655 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
656 up_read(&mm->mmap_sem);
657 mmdrop(mm);
658 } else {
659 spin_unlock(&ksm_mmlist_lock);
660 up_read(&mm->mmap_sem);
661 }
Izik Eidus31dbd012009-09-21 17:02:03 -0700662 }
663
Hugh Dickinsd952b792009-09-21 17:02:16 -0700664 ksm_scan.seqnr = 0;
Hugh Dickins9ba69292009-09-21 17:02:20 -0700665 return 0;
666
667error:
668 up_read(&mm->mmap_sem);
Izik Eidus31dbd012009-09-21 17:02:03 -0700669 spin_lock(&ksm_mmlist_lock);
Hugh Dickinsd952b792009-09-21 17:02:16 -0700670 ksm_scan.mm_slot = &ksm_mm_head;
Izik Eidus31dbd012009-09-21 17:02:03 -0700671 spin_unlock(&ksm_mmlist_lock);
Hugh Dickinsd952b792009-09-21 17:02:16 -0700672 return err;
Izik Eidus31dbd012009-09-21 17:02:03 -0700673}
Hugh Dickins2ffd8672009-09-21 17:02:23 -0700674#endif /* CONFIG_SYSFS */
Izik Eidus31dbd012009-09-21 17:02:03 -0700675
Izik Eidus31dbd012009-09-21 17:02:03 -0700676static u32 calc_checksum(struct page *page)
677{
678 u32 checksum;
Cong Wang9b04c5f2011-11-25 23:14:39 +0800679 void *addr = kmap_atomic(page);
Izik Eidus31dbd012009-09-21 17:02:03 -0700680 checksum = jhash2(addr, PAGE_SIZE / 4, 17);
Cong Wang9b04c5f2011-11-25 23:14:39 +0800681 kunmap_atomic(addr);
Izik Eidus31dbd012009-09-21 17:02:03 -0700682 return checksum;
683}
684
685static int memcmp_pages(struct page *page1, struct page *page2)
686{
687 char *addr1, *addr2;
688 int ret;
689
Cong Wang9b04c5f2011-11-25 23:14:39 +0800690 addr1 = kmap_atomic(page1);
691 addr2 = kmap_atomic(page2);
Izik Eidus31dbd012009-09-21 17:02:03 -0700692 ret = memcmp(addr1, addr2, PAGE_SIZE);
Cong Wang9b04c5f2011-11-25 23:14:39 +0800693 kunmap_atomic(addr2);
694 kunmap_atomic(addr1);
Izik Eidus31dbd012009-09-21 17:02:03 -0700695 return ret;
696}
697
698static inline int pages_identical(struct page *page1, struct page *page2)
699{
700 return !memcmp_pages(page1, page2);
701}
702
703static int write_protect_page(struct vm_area_struct *vma, struct page *page,
704 pte_t *orig_pte)
705{
706 struct mm_struct *mm = vma->vm_mm;
707 unsigned long addr;
708 pte_t *ptep;
709 spinlock_t *ptl;
710 int swapped;
711 int err = -EFAULT;
Haggai Eran6bdb9132012-10-08 16:33:35 -0700712 unsigned long mmun_start; /* For mmu_notifiers */
713 unsigned long mmun_end; /* For mmu_notifiers */
Izik Eidus31dbd012009-09-21 17:02:03 -0700714
715 addr = page_address_in_vma(page, vma);
716 if (addr == -EFAULT)
717 goto out;
718
Andrea Arcangeli29ad7682011-01-13 15:47:19 -0800719 BUG_ON(PageTransCompound(page));
Haggai Eran6bdb9132012-10-08 16:33:35 -0700720
721 mmun_start = addr;
722 mmun_end = addr + PAGE_SIZE;
723 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
724
Izik Eidus31dbd012009-09-21 17:02:03 -0700725 ptep = page_check_address(page, mm, addr, &ptl, 0);
726 if (!ptep)
Haggai Eran6bdb9132012-10-08 16:33:35 -0700727 goto out_mn;
Izik Eidus31dbd012009-09-21 17:02:03 -0700728
Hugh Dickins4e316352010-10-02 17:49:08 -0700729 if (pte_write(*ptep) || pte_dirty(*ptep)) {
Izik Eidus31dbd012009-09-21 17:02:03 -0700730 pte_t entry;
731
732 swapped = PageSwapCache(page);
733 flush_cache_page(vma, addr, page_to_pfn(page));
734 /*
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300735 * Ok this is tricky, when get_user_pages_fast() run it doesn't
Izik Eidus31dbd012009-09-21 17:02:03 -0700736 * take any lock, therefore the check that we are going to make
737 * with the pagecount against the mapcount is racey and
738 * O_DIRECT can happen right after the check.
739 * So we clear the pte and flush the tlb before the check
740 * this assure us that no O_DIRECT can happen after the check
741 * or in the middle of the check.
742 */
743 entry = ptep_clear_flush(vma, addr, ptep);
744 /*
745 * Check that no O_DIRECT or similar I/O is in progress on the
746 * page
747 */
Hugh Dickins31e855e2009-12-14 17:59:17 -0800748 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
Robin Holtcb532372010-03-23 13:35:26 -0700749 set_pte_at(mm, addr, ptep, entry);
Izik Eidus31dbd012009-09-21 17:02:03 -0700750 goto out_unlock;
751 }
Hugh Dickins4e316352010-10-02 17:49:08 -0700752 if (pte_dirty(entry))
753 set_page_dirty(page);
754 entry = pte_mkclean(pte_wrprotect(entry));
Izik Eidus31dbd012009-09-21 17:02:03 -0700755 set_pte_at_notify(mm, addr, ptep, entry);
756 }
757 *orig_pte = *ptep;
758 err = 0;
759
760out_unlock:
761 pte_unmap_unlock(ptep, ptl);
Haggai Eran6bdb9132012-10-08 16:33:35 -0700762out_mn:
763 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Izik Eidus31dbd012009-09-21 17:02:03 -0700764out:
765 return err;
766}
767
768/**
769 * replace_page - replace page in vma by new ksm page
Hugh Dickins8dd35572009-12-14 17:59:18 -0800770 * @vma: vma that holds the pte pointing to page
771 * @page: the page we are replacing by kpage
772 * @kpage: the ksm page we replace page by
Izik Eidus31dbd012009-09-21 17:02:03 -0700773 * @orig_pte: the original value of the pte
774 *
775 * Returns 0 on success, -EFAULT on failure.
776 */
Hugh Dickins8dd35572009-12-14 17:59:18 -0800777static int replace_page(struct vm_area_struct *vma, struct page *page,
778 struct page *kpage, pte_t orig_pte)
Izik Eidus31dbd012009-09-21 17:02:03 -0700779{
780 struct mm_struct *mm = vma->vm_mm;
781 pgd_t *pgd;
782 pud_t *pud;
783 pmd_t *pmd;
784 pte_t *ptep;
785 spinlock_t *ptl;
786 unsigned long addr;
Izik Eidus31dbd012009-09-21 17:02:03 -0700787 int err = -EFAULT;
Haggai Eran6bdb9132012-10-08 16:33:35 -0700788 unsigned long mmun_start; /* For mmu_notifiers */
789 unsigned long mmun_end; /* For mmu_notifiers */
Izik Eidus31dbd012009-09-21 17:02:03 -0700790
Hugh Dickins8dd35572009-12-14 17:59:18 -0800791 addr = page_address_in_vma(page, vma);
Izik Eidus31dbd012009-09-21 17:02:03 -0700792 if (addr == -EFAULT)
793 goto out;
794
795 pgd = pgd_offset(mm, addr);
796 if (!pgd_present(*pgd))
797 goto out;
798
799 pud = pud_offset(pgd, addr);
800 if (!pud_present(*pud))
801 goto out;
802
803 pmd = pmd_offset(pud, addr);
Andrea Arcangeli29ad7682011-01-13 15:47:19 -0800804 BUG_ON(pmd_trans_huge(*pmd));
Izik Eidus31dbd012009-09-21 17:02:03 -0700805 if (!pmd_present(*pmd))
806 goto out;
807
Haggai Eran6bdb9132012-10-08 16:33:35 -0700808 mmun_start = addr;
809 mmun_end = addr + PAGE_SIZE;
810 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
811
Izik Eidus31dbd012009-09-21 17:02:03 -0700812 ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
813 if (!pte_same(*ptep, orig_pte)) {
814 pte_unmap_unlock(ptep, ptl);
Haggai Eran6bdb9132012-10-08 16:33:35 -0700815 goto out_mn;
Izik Eidus31dbd012009-09-21 17:02:03 -0700816 }
817
Hugh Dickins8dd35572009-12-14 17:59:18 -0800818 get_page(kpage);
Hugh Dickins5ad64682009-12-14 17:59:24 -0800819 page_add_anon_rmap(kpage, vma, addr);
Izik Eidus31dbd012009-09-21 17:02:03 -0700820
821 flush_cache_page(vma, addr, pte_pfn(*ptep));
822 ptep_clear_flush(vma, addr, ptep);
Hugh Dickins8dd35572009-12-14 17:59:18 -0800823 set_pte_at_notify(mm, addr, ptep, mk_pte(kpage, vma->vm_page_prot));
Izik Eidus31dbd012009-09-21 17:02:03 -0700824
Hugh Dickins8dd35572009-12-14 17:59:18 -0800825 page_remove_rmap(page);
Hugh Dickinsae52a2a2011-01-13 15:46:28 -0800826 if (!page_mapped(page))
827 try_to_free_swap(page);
Hugh Dickins8dd35572009-12-14 17:59:18 -0800828 put_page(page);
Izik Eidus31dbd012009-09-21 17:02:03 -0700829
830 pte_unmap_unlock(ptep, ptl);
831 err = 0;
Haggai Eran6bdb9132012-10-08 16:33:35 -0700832out_mn:
833 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Izik Eidus31dbd012009-09-21 17:02:03 -0700834out:
835 return err;
836}
837
Andrea Arcangeli29ad7682011-01-13 15:47:19 -0800838static int page_trans_compound_anon_split(struct page *page)
839{
840 int ret = 0;
841 struct page *transhuge_head = page_trans_compound_anon(page);
842 if (transhuge_head) {
843 /* Get the reference on the head to split it. */
844 if (get_page_unless_zero(transhuge_head)) {
845 /*
846 * Recheck we got the reference while the head
847 * was still anonymous.
848 */
849 if (PageAnon(transhuge_head))
850 ret = split_huge_page(transhuge_head);
851 else
852 /*
853 * Retry later if split_huge_page run
854 * from under us.
855 */
856 ret = 1;
857 put_page(transhuge_head);
858 } else
859 /* Retry later if split_huge_page run from under us. */
860 ret = 1;
861 }
862 return ret;
863}
864
Izik Eidus31dbd012009-09-21 17:02:03 -0700865/*
866 * try_to_merge_one_page - take two pages and merge them into one
Hugh Dickins8dd35572009-12-14 17:59:18 -0800867 * @vma: the vma that holds the pte pointing to page
868 * @page: the PageAnon page that we want to replace with kpage
Hugh Dickins80e148222009-12-14 17:59:29 -0800869 * @kpage: the PageKsm page that we want to map instead of page,
870 * or NULL the first time when we want to use page as kpage.
Izik Eidus31dbd012009-09-21 17:02:03 -0700871 *
872 * This function returns 0 if the pages were merged, -EFAULT otherwise.
873 */
874static int try_to_merge_one_page(struct vm_area_struct *vma,
Hugh Dickins8dd35572009-12-14 17:59:18 -0800875 struct page *page, struct page *kpage)
Izik Eidus31dbd012009-09-21 17:02:03 -0700876{
877 pte_t orig_pte = __pte(0);
878 int err = -EFAULT;
879
Hugh Dickinsdb114b82009-12-14 17:59:25 -0800880 if (page == kpage) /* ksm page forked */
881 return 0;
882
Izik Eidus31dbd012009-09-21 17:02:03 -0700883 if (!(vma->vm_flags & VM_MERGEABLE))
884 goto out;
Andrea Arcangeli29ad7682011-01-13 15:47:19 -0800885 if (PageTransCompound(page) && page_trans_compound_anon_split(page))
886 goto out;
887 BUG_ON(PageTransCompound(page));
Hugh Dickins8dd35572009-12-14 17:59:18 -0800888 if (!PageAnon(page))
Izik Eidus31dbd012009-09-21 17:02:03 -0700889 goto out;
890
Izik Eidus31dbd012009-09-21 17:02:03 -0700891 /*
892 * We need the page lock to read a stable PageSwapCache in
893 * write_protect_page(). We use trylock_page() instead of
894 * lock_page() because we don't want to wait here - we
895 * prefer to continue scanning and merging different pages,
896 * then come back to this page when it is unlocked.
897 */
Hugh Dickins8dd35572009-12-14 17:59:18 -0800898 if (!trylock_page(page))
Hugh Dickins31e855e2009-12-14 17:59:17 -0800899 goto out;
Izik Eidus31dbd012009-09-21 17:02:03 -0700900 /*
901 * If this anonymous page is mapped only here, its pte may need
902 * to be write-protected. If it's mapped elsewhere, all of its
903 * ptes are necessarily already write-protected. But in either
904 * case, we need to lock and check page_count is not raised.
905 */
Hugh Dickins80e148222009-12-14 17:59:29 -0800906 if (write_protect_page(vma, page, &orig_pte) == 0) {
907 if (!kpage) {
908 /*
909 * While we hold page lock, upgrade page from
910 * PageAnon+anon_vma to PageKsm+NULL stable_node:
911 * stable_tree_insert() will update stable_node.
912 */
913 set_page_stable_node(page, NULL);
914 mark_page_accessed(page);
915 err = 0;
916 } else if (pages_identical(page, kpage))
917 err = replace_page(vma, page, kpage, orig_pte);
918 }
Izik Eidus31dbd012009-09-21 17:02:03 -0700919
Hugh Dickins80e148222009-12-14 17:59:29 -0800920 if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
Hugh Dickins73848b42009-12-14 17:59:22 -0800921 munlock_vma_page(page);
Hugh Dickins5ad64682009-12-14 17:59:24 -0800922 if (!PageMlocked(kpage)) {
923 unlock_page(page);
Hugh Dickins5ad64682009-12-14 17:59:24 -0800924 lock_page(kpage);
925 mlock_vma_page(kpage);
926 page = kpage; /* for final unlock */
927 }
928 }
Hugh Dickins73848b42009-12-14 17:59:22 -0800929
Hugh Dickins8dd35572009-12-14 17:59:18 -0800930 unlock_page(page);
Izik Eidus31dbd012009-09-21 17:02:03 -0700931out:
932 return err;
933}
934
935/*
Hugh Dickins81464e302009-09-21 17:02:15 -0700936 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
937 * but no new kernel page is allocated: kpage must already be a ksm page.
Hugh Dickins8dd35572009-12-14 17:59:18 -0800938 *
939 * This function returns 0 if the pages were merged, -EFAULT otherwise.
Hugh Dickins81464e302009-09-21 17:02:15 -0700940 */
Hugh Dickins8dd35572009-12-14 17:59:18 -0800941static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
942 struct page *page, struct page *kpage)
Hugh Dickins81464e302009-09-21 17:02:15 -0700943{
Hugh Dickins8dd35572009-12-14 17:59:18 -0800944 struct mm_struct *mm = rmap_item->mm;
Hugh Dickins81464e302009-09-21 17:02:15 -0700945 struct vm_area_struct *vma;
946 int err = -EFAULT;
947
Hugh Dickins8dd35572009-12-14 17:59:18 -0800948 down_read(&mm->mmap_sem);
949 if (ksm_test_exit(mm))
950 goto out;
951 vma = find_vma(mm, rmap_item->address);
952 if (!vma || vma->vm_start > rmap_item->address)
Hugh Dickins9ba69292009-09-21 17:02:20 -0700953 goto out;
954
Hugh Dickins8dd35572009-12-14 17:59:18 -0800955 err = try_to_merge_one_page(vma, page, kpage);
Hugh Dickinsdb114b82009-12-14 17:59:25 -0800956 if (err)
957 goto out;
958
959 /* Must get reference to anon_vma while still holding mmap_sem */
Peter Zijlstra9e601092011-03-22 16:32:46 -0700960 rmap_item->anon_vma = vma->anon_vma;
961 get_anon_vma(vma->anon_vma);
Hugh Dickins81464e302009-09-21 17:02:15 -0700962out:
Hugh Dickins8dd35572009-12-14 17:59:18 -0800963 up_read(&mm->mmap_sem);
Hugh Dickins81464e302009-09-21 17:02:15 -0700964 return err;
965}
966
967/*
Izik Eidus31dbd012009-09-21 17:02:03 -0700968 * try_to_merge_two_pages - take two identical pages and prepare them
969 * to be merged into one page.
970 *
Hugh Dickins8dd35572009-12-14 17:59:18 -0800971 * This function returns the kpage if we successfully merged two identical
972 * pages into one ksm page, NULL otherwise.
Izik Eidus31dbd012009-09-21 17:02:03 -0700973 *
Hugh Dickins80e148222009-12-14 17:59:29 -0800974 * Note that this function upgrades page to ksm page: if one of the pages
Izik Eidus31dbd012009-09-21 17:02:03 -0700975 * is already a ksm page, try_to_merge_with_ksm_page should be used.
976 */
Hugh Dickins8dd35572009-12-14 17:59:18 -0800977static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
978 struct page *page,
979 struct rmap_item *tree_rmap_item,
980 struct page *tree_page)
Izik Eidus31dbd012009-09-21 17:02:03 -0700981{
Hugh Dickins80e148222009-12-14 17:59:29 -0800982 int err;
Izik Eidus31dbd012009-09-21 17:02:03 -0700983
Hugh Dickins80e148222009-12-14 17:59:29 -0800984 err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
Izik Eidus31dbd012009-09-21 17:02:03 -0700985 if (!err) {
Hugh Dickins8dd35572009-12-14 17:59:18 -0800986 err = try_to_merge_with_ksm_page(tree_rmap_item,
Hugh Dickins80e148222009-12-14 17:59:29 -0800987 tree_page, page);
Izik Eidus31dbd012009-09-21 17:02:03 -0700988 /*
Hugh Dickins81464e302009-09-21 17:02:15 -0700989 * If that fails, we have a ksm page with only one pte
990 * pointing to it: so break it.
Izik Eidus31dbd012009-09-21 17:02:03 -0700991 */
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800992 if (err)
Hugh Dickins8dd35572009-12-14 17:59:18 -0800993 break_cow(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -0700994 }
Hugh Dickins80e148222009-12-14 17:59:29 -0800995 return err ? NULL : page;
Izik Eidus31dbd012009-09-21 17:02:03 -0700996}
997
998/*
Hugh Dickins8dd35572009-12-14 17:59:18 -0800999 * stable_tree_search - search for page inside the stable tree
Izik Eidus31dbd012009-09-21 17:02:03 -07001000 *
1001 * This function checks if there is a page inside the stable tree
1002 * with identical content to the page that we are scanning right now.
1003 *
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001004 * This function returns the stable tree node of identical content if found,
Izik Eidus31dbd012009-09-21 17:02:03 -07001005 * NULL otherwise.
1006 */
Hugh Dickins62b61f62009-12-14 17:59:33 -08001007static struct page *stable_tree_search(struct page *page)
Izik Eidus31dbd012009-09-21 17:02:03 -07001008{
1009 struct rb_node *node = root_stable_tree.rb_node;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001010 struct stable_node *stable_node;
Izik Eidus31dbd012009-09-21 17:02:03 -07001011
Hugh Dickins08beca42009-12-14 17:59:21 -08001012 stable_node = page_stable_node(page);
1013 if (stable_node) { /* ksm page forked */
1014 get_page(page);
Hugh Dickins62b61f62009-12-14 17:59:33 -08001015 return page;
Hugh Dickins08beca42009-12-14 17:59:21 -08001016 }
1017
Izik Eidus31dbd012009-09-21 17:02:03 -07001018 while (node) {
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001019 struct page *tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001020 int ret;
1021
Hugh Dickins08beca42009-12-14 17:59:21 -08001022 cond_resched();
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001023 stable_node = rb_entry(node, struct stable_node, node);
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001024 tree_page = get_ksm_page(stable_node);
1025 if (!tree_page)
1026 return NULL;
Izik Eidus31dbd012009-09-21 17:02:03 -07001027
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001028 ret = memcmp_pages(page, tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001029
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001030 if (ret < 0) {
1031 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001032 node = node->rb_left;
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001033 } else if (ret > 0) {
1034 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001035 node = node->rb_right;
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001036 } else
Hugh Dickins62b61f62009-12-14 17:59:33 -08001037 return tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001038 }
1039
1040 return NULL;
1041}
1042
1043/*
1044 * stable_tree_insert - insert rmap_item pointing to new ksm page
1045 * into the stable tree.
1046 *
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001047 * This function returns the stable tree node just allocated on success,
1048 * NULL otherwise.
Izik Eidus31dbd012009-09-21 17:02:03 -07001049 */
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001050static struct stable_node *stable_tree_insert(struct page *kpage)
Izik Eidus31dbd012009-09-21 17:02:03 -07001051{
1052 struct rb_node **new = &root_stable_tree.rb_node;
1053 struct rb_node *parent = NULL;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001054 struct stable_node *stable_node;
Izik Eidus31dbd012009-09-21 17:02:03 -07001055
1056 while (*new) {
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001057 struct page *tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001058 int ret;
1059
Hugh Dickins08beca42009-12-14 17:59:21 -08001060 cond_resched();
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001061 stable_node = rb_entry(*new, struct stable_node, node);
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001062 tree_page = get_ksm_page(stable_node);
1063 if (!tree_page)
1064 return NULL;
Izik Eidus31dbd012009-09-21 17:02:03 -07001065
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001066 ret = memcmp_pages(kpage, tree_page);
1067 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001068
1069 parent = *new;
1070 if (ret < 0)
1071 new = &parent->rb_left;
1072 else if (ret > 0)
1073 new = &parent->rb_right;
1074 else {
1075 /*
1076 * It is not a bug that stable_tree_search() didn't
1077 * find this node: because at that time our page was
1078 * not yet write-protected, so may have changed since.
1079 */
1080 return NULL;
1081 }
1082 }
1083
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001084 stable_node = alloc_stable_node();
1085 if (!stable_node)
1086 return NULL;
Izik Eidus31dbd012009-09-21 17:02:03 -07001087
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001088 rb_link_node(&stable_node->node, parent, new);
1089 rb_insert_color(&stable_node->node, &root_stable_tree);
1090
1091 INIT_HLIST_HEAD(&stable_node->hlist);
1092
Hugh Dickins62b61f62009-12-14 17:59:33 -08001093 stable_node->kpfn = page_to_pfn(kpage);
Hugh Dickins08beca42009-12-14 17:59:21 -08001094 set_page_stable_node(kpage, stable_node);
1095
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001096 return stable_node;
Izik Eidus31dbd012009-09-21 17:02:03 -07001097}
1098
1099/*
Hugh Dickins8dd35572009-12-14 17:59:18 -08001100 * unstable_tree_search_insert - search for identical page,
1101 * else insert rmap_item into the unstable tree.
Izik Eidus31dbd012009-09-21 17:02:03 -07001102 *
1103 * This function searches for a page in the unstable tree identical to the
1104 * page currently being scanned; and if no identical page is found in the
1105 * tree, we insert rmap_item as a new object into the unstable tree.
1106 *
1107 * This function returns pointer to rmap_item found to be identical
1108 * to the currently scanned page, NULL otherwise.
1109 *
1110 * This function does both searching and inserting, because they share
1111 * the same walking algorithm in an rbtree.
1112 */
Hugh Dickins8dd35572009-12-14 17:59:18 -08001113static
1114struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1115 struct page *page,
1116 struct page **tree_pagep)
1117
Izik Eidus31dbd012009-09-21 17:02:03 -07001118{
1119 struct rb_node **new = &root_unstable_tree.rb_node;
1120 struct rb_node *parent = NULL;
1121
1122 while (*new) {
1123 struct rmap_item *tree_rmap_item;
Hugh Dickins8dd35572009-12-14 17:59:18 -08001124 struct page *tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001125 int ret;
1126
Hugh Dickinsd178f272009-11-09 15:58:23 +00001127 cond_resched();
Izik Eidus31dbd012009-09-21 17:02:03 -07001128 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
Hugh Dickins8dd35572009-12-14 17:59:18 -08001129 tree_page = get_mergeable_page(tree_rmap_item);
Dan Carpenter22eccdd2010-04-23 13:18:10 -04001130 if (IS_ERR_OR_NULL(tree_page))
Izik Eidus31dbd012009-09-21 17:02:03 -07001131 return NULL;
1132
1133 /*
Hugh Dickins8dd35572009-12-14 17:59:18 -08001134 * Don't substitute a ksm page for a forked page.
Izik Eidus31dbd012009-09-21 17:02:03 -07001135 */
Hugh Dickins8dd35572009-12-14 17:59:18 -08001136 if (page == tree_page) {
1137 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001138 return NULL;
1139 }
1140
Hugh Dickins8dd35572009-12-14 17:59:18 -08001141 ret = memcmp_pages(page, tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001142
1143 parent = *new;
1144 if (ret < 0) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001145 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001146 new = &parent->rb_left;
1147 } else if (ret > 0) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001148 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001149 new = &parent->rb_right;
1150 } else {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001151 *tree_pagep = tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001152 return tree_rmap_item;
1153 }
1154 }
1155
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001156 rmap_item->address |= UNSTABLE_FLAG;
Izik Eidus31dbd012009-09-21 17:02:03 -07001157 rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1158 rb_link_node(&rmap_item->node, parent, new);
1159 rb_insert_color(&rmap_item->node, &root_unstable_tree);
1160
Hugh Dickins473b0ce2009-09-21 17:02:11 -07001161 ksm_pages_unshared++;
Izik Eidus31dbd012009-09-21 17:02:03 -07001162 return NULL;
1163}
1164
1165/*
1166 * stable_tree_append - add another rmap_item to the linked list of
1167 * rmap_items hanging off a given node of the stable tree, all sharing
1168 * the same ksm page.
1169 */
1170static void stable_tree_append(struct rmap_item *rmap_item,
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001171 struct stable_node *stable_node)
Izik Eidus31dbd012009-09-21 17:02:03 -07001172{
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001173 rmap_item->head = stable_node;
Izik Eidus31dbd012009-09-21 17:02:03 -07001174 rmap_item->address |= STABLE_FLAG;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001175 hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
Hugh Dickinse178dfd2009-09-21 17:02:10 -07001176
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001177 if (rmap_item->hlist.next)
1178 ksm_pages_sharing++;
1179 else
1180 ksm_pages_shared++;
Izik Eidus31dbd012009-09-21 17:02:03 -07001181}
1182
1183/*
Hugh Dickins81464e302009-09-21 17:02:15 -07001184 * cmp_and_merge_page - first see if page can be merged into the stable tree;
1185 * if not, compare checksum to previous and if it's the same, see if page can
1186 * be inserted into the unstable tree, or merged with a page already there and
1187 * both transferred to the stable tree.
Izik Eidus31dbd012009-09-21 17:02:03 -07001188 *
1189 * @page: the page that we are searching identical page to.
1190 * @rmap_item: the reverse mapping into the virtual address of this page
1191 */
1192static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1193{
Izik Eidus31dbd012009-09-21 17:02:03 -07001194 struct rmap_item *tree_rmap_item;
Hugh Dickins8dd35572009-12-14 17:59:18 -08001195 struct page *tree_page = NULL;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001196 struct stable_node *stable_node;
Hugh Dickins8dd35572009-12-14 17:59:18 -08001197 struct page *kpage;
Izik Eidus31dbd012009-09-21 17:02:03 -07001198 unsigned int checksum;
1199 int err;
1200
Hugh Dickins93d17712009-12-14 17:59:16 -08001201 remove_rmap_item_from_tree(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -07001202
1203 /* We first start with searching the page inside the stable tree */
Hugh Dickins62b61f62009-12-14 17:59:33 -08001204 kpage = stable_tree_search(page);
1205 if (kpage) {
Hugh Dickins08beca42009-12-14 17:59:21 -08001206 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
Izik Eidus31dbd012009-09-21 17:02:03 -07001207 if (!err) {
1208 /*
1209 * The page was successfully merged:
1210 * add its rmap_item to the stable tree.
1211 */
Hugh Dickins5ad64682009-12-14 17:59:24 -08001212 lock_page(kpage);
Hugh Dickins62b61f62009-12-14 17:59:33 -08001213 stable_tree_append(rmap_item, page_stable_node(kpage));
Hugh Dickins5ad64682009-12-14 17:59:24 -08001214 unlock_page(kpage);
Izik Eidus31dbd012009-09-21 17:02:03 -07001215 }
Hugh Dickins8dd35572009-12-14 17:59:18 -08001216 put_page(kpage);
Izik Eidus31dbd012009-09-21 17:02:03 -07001217 return;
1218 }
1219
1220 /*
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001221 * If the hash value of the page has changed from the last time
1222 * we calculated it, this page is changing frequently: therefore we
1223 * don't want to insert it in the unstable tree, and we don't want
1224 * to waste our time searching for something identical to it there.
Izik Eidus31dbd012009-09-21 17:02:03 -07001225 */
1226 checksum = calc_checksum(page);
1227 if (rmap_item->oldchecksum != checksum) {
1228 rmap_item->oldchecksum = checksum;
1229 return;
1230 }
1231
Hugh Dickins8dd35572009-12-14 17:59:18 -08001232 tree_rmap_item =
1233 unstable_tree_search_insert(rmap_item, page, &tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001234 if (tree_rmap_item) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001235 kpage = try_to_merge_two_pages(rmap_item, page,
1236 tree_rmap_item, tree_page);
1237 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001238 /*
1239 * As soon as we merge this page, we want to remove the
1240 * rmap_item of the page we have merged with from the unstable
1241 * tree, and insert it instead as new node in the stable tree.
1242 */
Hugh Dickins8dd35572009-12-14 17:59:18 -08001243 if (kpage) {
Hugh Dickins93d17712009-12-14 17:59:16 -08001244 remove_rmap_item_from_tree(tree_rmap_item);
Hugh Dickins473b0ce2009-09-21 17:02:11 -07001245
Hugh Dickins5ad64682009-12-14 17:59:24 -08001246 lock_page(kpage);
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001247 stable_node = stable_tree_insert(kpage);
1248 if (stable_node) {
1249 stable_tree_append(tree_rmap_item, stable_node);
1250 stable_tree_append(rmap_item, stable_node);
1251 }
Hugh Dickins5ad64682009-12-14 17:59:24 -08001252 unlock_page(kpage);
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001253
Izik Eidus31dbd012009-09-21 17:02:03 -07001254 /*
1255 * If we fail to insert the page into the stable tree,
1256 * we will have 2 virtual addresses that are pointing
1257 * to a ksm page left outside the stable tree,
1258 * in which case we need to break_cow on both.
1259 */
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001260 if (!stable_node) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001261 break_cow(tree_rmap_item);
1262 break_cow(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -07001263 }
1264 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001265 }
1266}
1267
1268static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
Hugh Dickins6514d512009-12-14 17:59:19 -08001269 struct rmap_item **rmap_list,
Izik Eidus31dbd012009-09-21 17:02:03 -07001270 unsigned long addr)
1271{
1272 struct rmap_item *rmap_item;
1273
Hugh Dickins6514d512009-12-14 17:59:19 -08001274 while (*rmap_list) {
1275 rmap_item = *rmap_list;
Hugh Dickins93d17712009-12-14 17:59:16 -08001276 if ((rmap_item->address & PAGE_MASK) == addr)
Izik Eidus31dbd012009-09-21 17:02:03 -07001277 return rmap_item;
Izik Eidus31dbd012009-09-21 17:02:03 -07001278 if (rmap_item->address > addr)
1279 break;
Hugh Dickins6514d512009-12-14 17:59:19 -08001280 *rmap_list = rmap_item->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -07001281 remove_rmap_item_from_tree(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -07001282 free_rmap_item(rmap_item);
1283 }
1284
1285 rmap_item = alloc_rmap_item();
1286 if (rmap_item) {
1287 /* It has already been zeroed */
1288 rmap_item->mm = mm_slot->mm;
1289 rmap_item->address = addr;
Hugh Dickins6514d512009-12-14 17:59:19 -08001290 rmap_item->rmap_list = *rmap_list;
1291 *rmap_list = rmap_item;
Izik Eidus31dbd012009-09-21 17:02:03 -07001292 }
1293 return rmap_item;
1294}
1295
1296static struct rmap_item *scan_get_next_rmap_item(struct page **page)
1297{
1298 struct mm_struct *mm;
1299 struct mm_slot *slot;
1300 struct vm_area_struct *vma;
1301 struct rmap_item *rmap_item;
1302
1303 if (list_empty(&ksm_mm_head.mm_list))
1304 return NULL;
1305
1306 slot = ksm_scan.mm_slot;
1307 if (slot == &ksm_mm_head) {
Hugh Dickins2919bfd2011-01-13 15:47:29 -08001308 /*
1309 * A number of pages can hang around indefinitely on per-cpu
1310 * pagevecs, raised page count preventing write_protect_page
1311 * from merging them. Though it doesn't really matter much,
1312 * it is puzzling to see some stuck in pages_volatile until
1313 * other activity jostles them out, and they also prevented
1314 * LTP's KSM test from succeeding deterministically; so drain
1315 * them here (here rather than on entry to ksm_do_scan(),
1316 * so we don't IPI too often when pages_to_scan is set low).
1317 */
1318 lru_add_drain_all();
1319
Izik Eidus31dbd012009-09-21 17:02:03 -07001320 root_unstable_tree = RB_ROOT;
1321
1322 spin_lock(&ksm_mmlist_lock);
1323 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1324 ksm_scan.mm_slot = slot;
1325 spin_unlock(&ksm_mmlist_lock);
Hugh Dickins2b472612011-06-15 15:08:58 -07001326 /*
1327 * Although we tested list_empty() above, a racing __ksm_exit
1328 * of the last mm on the list may have removed it since then.
1329 */
1330 if (slot == &ksm_mm_head)
1331 return NULL;
Izik Eidus31dbd012009-09-21 17:02:03 -07001332next_mm:
1333 ksm_scan.address = 0;
Hugh Dickins6514d512009-12-14 17:59:19 -08001334 ksm_scan.rmap_list = &slot->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -07001335 }
1336
1337 mm = slot->mm;
1338 down_read(&mm->mmap_sem);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001339 if (ksm_test_exit(mm))
1340 vma = NULL;
1341 else
1342 vma = find_vma(mm, ksm_scan.address);
1343
1344 for (; vma; vma = vma->vm_next) {
Izik Eidus31dbd012009-09-21 17:02:03 -07001345 if (!(vma->vm_flags & VM_MERGEABLE))
1346 continue;
1347 if (ksm_scan.address < vma->vm_start)
1348 ksm_scan.address = vma->vm_start;
1349 if (!vma->anon_vma)
1350 ksm_scan.address = vma->vm_end;
1351
1352 while (ksm_scan.address < vma->vm_end) {
Hugh Dickins9ba69292009-09-21 17:02:20 -07001353 if (ksm_test_exit(mm))
1354 break;
Izik Eidus31dbd012009-09-21 17:02:03 -07001355 *page = follow_page(vma, ksm_scan.address, FOLL_GET);
Andrea Arcangeli21ae5b02011-01-13 15:47:00 -08001356 if (IS_ERR_OR_NULL(*page)) {
1357 ksm_scan.address += PAGE_SIZE;
1358 cond_resched();
1359 continue;
1360 }
Andrea Arcangeli29ad7682011-01-13 15:47:19 -08001361 if (PageAnon(*page) ||
1362 page_trans_compound_anon(*page)) {
Izik Eidus31dbd012009-09-21 17:02:03 -07001363 flush_anon_page(vma, *page, ksm_scan.address);
1364 flush_dcache_page(*page);
1365 rmap_item = get_next_rmap_item(slot,
Hugh Dickins6514d512009-12-14 17:59:19 -08001366 ksm_scan.rmap_list, ksm_scan.address);
Izik Eidus31dbd012009-09-21 17:02:03 -07001367 if (rmap_item) {
Hugh Dickins6514d512009-12-14 17:59:19 -08001368 ksm_scan.rmap_list =
1369 &rmap_item->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -07001370 ksm_scan.address += PAGE_SIZE;
1371 } else
1372 put_page(*page);
1373 up_read(&mm->mmap_sem);
1374 return rmap_item;
1375 }
Andrea Arcangeli21ae5b02011-01-13 15:47:00 -08001376 put_page(*page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001377 ksm_scan.address += PAGE_SIZE;
1378 cond_resched();
1379 }
1380 }
1381
Hugh Dickins9ba69292009-09-21 17:02:20 -07001382 if (ksm_test_exit(mm)) {
1383 ksm_scan.address = 0;
Hugh Dickins6514d512009-12-14 17:59:19 -08001384 ksm_scan.rmap_list = &slot->rmap_list;
Hugh Dickins9ba69292009-09-21 17:02:20 -07001385 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001386 /*
1387 * Nuke all the rmap_items that are above this current rmap:
1388 * because there were no VM_MERGEABLE vmas with such addresses.
1389 */
Hugh Dickins6514d512009-12-14 17:59:19 -08001390 remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
Izik Eidus31dbd012009-09-21 17:02:03 -07001391
1392 spin_lock(&ksm_mmlist_lock);
Hugh Dickinscd551f92009-09-21 17:02:17 -07001393 ksm_scan.mm_slot = list_entry(slot->mm_list.next,
1394 struct mm_slot, mm_list);
1395 if (ksm_scan.address == 0) {
1396 /*
1397 * We've completed a full scan of all vmas, holding mmap_sem
1398 * throughout, and found no VM_MERGEABLE: so do the same as
1399 * __ksm_exit does to remove this mm from all our lists now.
Hugh Dickins9ba69292009-09-21 17:02:20 -07001400 * This applies either when cleaning up after __ksm_exit
1401 * (but beware: we can reach here even before __ksm_exit),
1402 * or when all VM_MERGEABLE areas have been unmapped (and
1403 * mmap_sem then protects against race with MADV_MERGEABLE).
Hugh Dickinscd551f92009-09-21 17:02:17 -07001404 */
1405 hlist_del(&slot->link);
1406 list_del(&slot->mm_list);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001407 spin_unlock(&ksm_mmlist_lock);
1408
Hugh Dickinscd551f92009-09-21 17:02:17 -07001409 free_mm_slot(slot);
1410 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001411 up_read(&mm->mmap_sem);
1412 mmdrop(mm);
1413 } else {
1414 spin_unlock(&ksm_mmlist_lock);
1415 up_read(&mm->mmap_sem);
Hugh Dickinscd551f92009-09-21 17:02:17 -07001416 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001417
1418 /* Repeat until we've completed scanning the whole list */
Hugh Dickinscd551f92009-09-21 17:02:17 -07001419 slot = ksm_scan.mm_slot;
Izik Eidus31dbd012009-09-21 17:02:03 -07001420 if (slot != &ksm_mm_head)
1421 goto next_mm;
1422
Izik Eidus31dbd012009-09-21 17:02:03 -07001423 ksm_scan.seqnr++;
1424 return NULL;
1425}
1426
1427/**
1428 * ksm_do_scan - the ksm scanner main worker function.
1429 * @scan_npages - number of pages we want to scan before we return.
1430 */
1431static void ksm_do_scan(unsigned int scan_npages)
1432{
1433 struct rmap_item *rmap_item;
Dan Carpenter22eccdd2010-04-23 13:18:10 -04001434 struct page *uninitialized_var(page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001435
Andrea Arcangeli878aee72011-01-13 15:47:10 -08001436 while (scan_npages-- && likely(!freezing(current))) {
Izik Eidus31dbd012009-09-21 17:02:03 -07001437 cond_resched();
1438 rmap_item = scan_get_next_rmap_item(&page);
1439 if (!rmap_item)
1440 return;
1441 if (!PageKsm(page) || !in_stable_tree(rmap_item))
1442 cmp_and_merge_page(page, rmap_item);
1443 put_page(page);
1444 }
1445}
1446
Hugh Dickins6e1583842009-09-21 17:02:14 -07001447static int ksmd_should_run(void)
1448{
1449 return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
1450}
1451
Izik Eidus31dbd012009-09-21 17:02:03 -07001452static int ksm_scan_thread(void *nothing)
1453{
Andrea Arcangeli878aee72011-01-13 15:47:10 -08001454 set_freezable();
Izik Eidus339aa622009-09-21 17:02:07 -07001455 set_user_nice(current, 5);
Izik Eidus31dbd012009-09-21 17:02:03 -07001456
1457 while (!kthread_should_stop()) {
Hugh Dickins6e1583842009-09-21 17:02:14 -07001458 mutex_lock(&ksm_thread_mutex);
1459 if (ksmd_should_run())
Izik Eidus31dbd012009-09-21 17:02:03 -07001460 ksm_do_scan(ksm_thread_pages_to_scan);
Hugh Dickins6e1583842009-09-21 17:02:14 -07001461 mutex_unlock(&ksm_thread_mutex);
1462
Andrea Arcangeli878aee72011-01-13 15:47:10 -08001463 try_to_freeze();
1464
Hugh Dickins6e1583842009-09-21 17:02:14 -07001465 if (ksmd_should_run()) {
Izik Eidus31dbd012009-09-21 17:02:03 -07001466 schedule_timeout_interruptible(
1467 msecs_to_jiffies(ksm_thread_sleep_millisecs));
1468 } else {
Andrea Arcangeli878aee72011-01-13 15:47:10 -08001469 wait_event_freezable(ksm_thread_wait,
Hugh Dickins6e1583842009-09-21 17:02:14 -07001470 ksmd_should_run() || kthread_should_stop());
Izik Eidus31dbd012009-09-21 17:02:03 -07001471 }
1472 }
1473 return 0;
1474}
1475
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001476int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1477 unsigned long end, int advice, unsigned long *vm_flags)
1478{
1479 struct mm_struct *mm = vma->vm_mm;
Hugh Dickinsd952b792009-09-21 17:02:16 -07001480 int err;
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001481
1482 switch (advice) {
1483 case MADV_MERGEABLE:
1484 /*
1485 * Be somewhat over-protective for now!
1486 */
1487 if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE |
1488 VM_PFNMAP | VM_IO | VM_DONTEXPAND |
Konstantin Khlebnikov314e51b2012-10-08 16:29:02 -07001489 VM_HUGETLB | VM_NONLINEAR | VM_MIXEDMAP))
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001490 return 0; /* just ignore the advice */
1491
Konstantin Khlebnikovcc2383e2012-10-08 16:28:37 -07001492#ifdef VM_SAO
1493 if (*vm_flags & VM_SAO)
1494 return 0;
1495#endif
1496
Hugh Dickinsd952b792009-09-21 17:02:16 -07001497 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
1498 err = __ksm_enter(mm);
1499 if (err)
1500 return err;
1501 }
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001502
1503 *vm_flags |= VM_MERGEABLE;
1504 break;
1505
1506 case MADV_UNMERGEABLE:
1507 if (!(*vm_flags & VM_MERGEABLE))
1508 return 0; /* just ignore the advice */
1509
Hugh Dickinsd952b792009-09-21 17:02:16 -07001510 if (vma->anon_vma) {
1511 err = unmerge_ksm_pages(vma, start, end);
1512 if (err)
1513 return err;
1514 }
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001515
1516 *vm_flags &= ~VM_MERGEABLE;
1517 break;
1518 }
1519
1520 return 0;
1521}
1522
1523int __ksm_enter(struct mm_struct *mm)
1524{
Hugh Dickins6e1583842009-09-21 17:02:14 -07001525 struct mm_slot *mm_slot;
1526 int needs_wakeup;
1527
1528 mm_slot = alloc_mm_slot();
Izik Eidus31dbd012009-09-21 17:02:03 -07001529 if (!mm_slot)
1530 return -ENOMEM;
1531
Hugh Dickins6e1583842009-09-21 17:02:14 -07001532 /* Check ksm_run too? Would need tighter locking */
1533 needs_wakeup = list_empty(&ksm_mm_head.mm_list);
1534
Izik Eidus31dbd012009-09-21 17:02:03 -07001535 spin_lock(&ksm_mmlist_lock);
1536 insert_to_mm_slots_hash(mm, mm_slot);
1537 /*
1538 * Insert just behind the scanning cursor, to let the area settle
1539 * down a little; when fork is followed by immediate exec, we don't
1540 * want ksmd to waste time setting up and tearing down an rmap_list.
1541 */
1542 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
1543 spin_unlock(&ksm_mmlist_lock);
1544
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001545 set_bit(MMF_VM_MERGEABLE, &mm->flags);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001546 atomic_inc(&mm->mm_count);
Hugh Dickins6e1583842009-09-21 17:02:14 -07001547
1548 if (needs_wakeup)
1549 wake_up_interruptible(&ksm_thread_wait);
1550
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001551 return 0;
1552}
1553
Andrea Arcangeli1c2fb7a2009-09-21 17:02:22 -07001554void __ksm_exit(struct mm_struct *mm)
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001555{
Hugh Dickinscd551f92009-09-21 17:02:17 -07001556 struct mm_slot *mm_slot;
Hugh Dickins9ba69292009-09-21 17:02:20 -07001557 int easy_to_free = 0;
Hugh Dickinscd551f92009-09-21 17:02:17 -07001558
Izik Eidus31dbd012009-09-21 17:02:03 -07001559 /*
Hugh Dickins9ba69292009-09-21 17:02:20 -07001560 * This process is exiting: if it's straightforward (as is the
1561 * case when ksmd was never running), free mm_slot immediately.
1562 * But if it's at the cursor or has rmap_items linked to it, use
1563 * mmap_sem to synchronize with any break_cows before pagetables
1564 * are freed, and leave the mm_slot on the list for ksmd to free.
1565 * Beware: ksm may already have noticed it exiting and freed the slot.
Izik Eidus31dbd012009-09-21 17:02:03 -07001566 */
Hugh Dickins9ba69292009-09-21 17:02:20 -07001567
Hugh Dickinscd551f92009-09-21 17:02:17 -07001568 spin_lock(&ksm_mmlist_lock);
1569 mm_slot = get_mm_slot(mm);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001570 if (mm_slot && ksm_scan.mm_slot != mm_slot) {
Hugh Dickins6514d512009-12-14 17:59:19 -08001571 if (!mm_slot->rmap_list) {
Hugh Dickins9ba69292009-09-21 17:02:20 -07001572 hlist_del(&mm_slot->link);
1573 list_del(&mm_slot->mm_list);
1574 easy_to_free = 1;
1575 } else {
1576 list_move(&mm_slot->mm_list,
1577 &ksm_scan.mm_slot->mm_list);
1578 }
Hugh Dickinscd551f92009-09-21 17:02:17 -07001579 }
Hugh Dickinscd551f92009-09-21 17:02:17 -07001580 spin_unlock(&ksm_mmlist_lock);
1581
Hugh Dickins9ba69292009-09-21 17:02:20 -07001582 if (easy_to_free) {
1583 free_mm_slot(mm_slot);
1584 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1585 mmdrop(mm);
1586 } else if (mm_slot) {
Hugh Dickins9ba69292009-09-21 17:02:20 -07001587 down_write(&mm->mmap_sem);
1588 up_write(&mm->mmap_sem);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001589 }
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001590}
Izik Eidus31dbd012009-09-21 17:02:03 -07001591
Hugh Dickins5ad64682009-12-14 17:59:24 -08001592struct page *ksm_does_need_to_copy(struct page *page,
1593 struct vm_area_struct *vma, unsigned long address)
1594{
1595 struct page *new_page;
1596
Hugh Dickins5ad64682009-12-14 17:59:24 -08001597 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1598 if (new_page) {
1599 copy_user_highpage(new_page, page, address, vma);
1600
1601 SetPageDirty(new_page);
1602 __SetPageUptodate(new_page);
1603 SetPageSwapBacked(new_page);
1604 __set_page_locked(new_page);
1605
Hugh Dickins39b5f292012-10-08 16:33:18 -07001606 if (!mlocked_vma_newpage(vma, new_page))
Hugh Dickins5ad64682009-12-14 17:59:24 -08001607 lru_cache_add_lru(new_page, LRU_ACTIVE_ANON);
1608 else
1609 add_page_to_unevictable_list(new_page);
1610 }
1611
Hugh Dickins5ad64682009-12-14 17:59:24 -08001612 return new_page;
1613}
1614
1615int page_referenced_ksm(struct page *page, struct mem_cgroup *memcg,
1616 unsigned long *vm_flags)
1617{
1618 struct stable_node *stable_node;
1619 struct rmap_item *rmap_item;
1620 struct hlist_node *hlist;
1621 unsigned int mapcount = page_mapcount(page);
1622 int referenced = 0;
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001623 int search_new_forks = 0;
Hugh Dickins5ad64682009-12-14 17:59:24 -08001624
1625 VM_BUG_ON(!PageKsm(page));
1626 VM_BUG_ON(!PageLocked(page));
1627
1628 stable_node = page_stable_node(page);
1629 if (!stable_node)
1630 return 0;
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001631again:
Hugh Dickins5ad64682009-12-14 17:59:24 -08001632 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001633 struct anon_vma *anon_vma = rmap_item->anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -08001634 struct anon_vma_chain *vmac;
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001635 struct vm_area_struct *vma;
Hugh Dickins5ad64682009-12-14 17:59:24 -08001636
Rik van Rielcba48b92010-08-09 17:18:38 -07001637 anon_vma_lock(anon_vma);
Michel Lespinassebf181b92012-10-08 16:31:39 -07001638 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
1639 0, ULONG_MAX) {
Rik van Riel5beb4932010-03-05 13:42:07 -08001640 vma = vmac->vma;
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001641 if (rmap_item->address < vma->vm_start ||
1642 rmap_item->address >= vma->vm_end)
1643 continue;
1644 /*
1645 * Initially we examine only the vma which covers this
1646 * rmap_item; but later, if there is still work to do,
1647 * we examine covering vmas in other mms: in case they
1648 * were forked from the original since ksmd passed.
1649 */
1650 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1651 continue;
Hugh Dickins5ad64682009-12-14 17:59:24 -08001652
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001653 if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
1654 continue;
1655
1656 referenced += page_referenced_one(page, vma,
Hugh Dickins5ad64682009-12-14 17:59:24 -08001657 rmap_item->address, &mapcount, vm_flags);
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001658 if (!search_new_forks || !mapcount)
1659 break;
1660 }
Rik van Rielcba48b92010-08-09 17:18:38 -07001661 anon_vma_unlock(anon_vma);
Hugh Dickins5ad64682009-12-14 17:59:24 -08001662 if (!mapcount)
1663 goto out;
1664 }
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001665 if (!search_new_forks++)
1666 goto again;
Hugh Dickins5ad64682009-12-14 17:59:24 -08001667out:
Hugh Dickins5ad64682009-12-14 17:59:24 -08001668 return referenced;
1669}
1670
1671int try_to_unmap_ksm(struct page *page, enum ttu_flags flags)
1672{
1673 struct stable_node *stable_node;
1674 struct hlist_node *hlist;
1675 struct rmap_item *rmap_item;
1676 int ret = SWAP_AGAIN;
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001677 int search_new_forks = 0;
Hugh Dickins5ad64682009-12-14 17:59:24 -08001678
1679 VM_BUG_ON(!PageKsm(page));
1680 VM_BUG_ON(!PageLocked(page));
1681
1682 stable_node = page_stable_node(page);
1683 if (!stable_node)
1684 return SWAP_FAIL;
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001685again:
Hugh Dickins5ad64682009-12-14 17:59:24 -08001686 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001687 struct anon_vma *anon_vma = rmap_item->anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -08001688 struct anon_vma_chain *vmac;
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001689 struct vm_area_struct *vma;
Hugh Dickins5ad64682009-12-14 17:59:24 -08001690
Rik van Rielcba48b92010-08-09 17:18:38 -07001691 anon_vma_lock(anon_vma);
Michel Lespinassebf181b92012-10-08 16:31:39 -07001692 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
1693 0, ULONG_MAX) {
Rik van Riel5beb4932010-03-05 13:42:07 -08001694 vma = vmac->vma;
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001695 if (rmap_item->address < vma->vm_start ||
1696 rmap_item->address >= vma->vm_end)
1697 continue;
1698 /*
1699 * Initially we examine only the vma which covers this
1700 * rmap_item; but later, if there is still work to do,
1701 * we examine covering vmas in other mms: in case they
1702 * were forked from the original since ksmd passed.
1703 */
1704 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1705 continue;
1706
1707 ret = try_to_unmap_one(page, vma,
1708 rmap_item->address, flags);
1709 if (ret != SWAP_AGAIN || !page_mapped(page)) {
Rik van Rielcba48b92010-08-09 17:18:38 -07001710 anon_vma_unlock(anon_vma);
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001711 goto out;
1712 }
1713 }
Rik van Rielcba48b92010-08-09 17:18:38 -07001714 anon_vma_unlock(anon_vma);
Hugh Dickins5ad64682009-12-14 17:59:24 -08001715 }
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001716 if (!search_new_forks++)
1717 goto again;
Hugh Dickins5ad64682009-12-14 17:59:24 -08001718out:
Hugh Dickins5ad64682009-12-14 17:59:24 -08001719 return ret;
1720}
1721
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001722#ifdef CONFIG_MIGRATION
1723int rmap_walk_ksm(struct page *page, int (*rmap_one)(struct page *,
1724 struct vm_area_struct *, unsigned long, void *), void *arg)
1725{
1726 struct stable_node *stable_node;
1727 struct hlist_node *hlist;
1728 struct rmap_item *rmap_item;
1729 int ret = SWAP_AGAIN;
1730 int search_new_forks = 0;
1731
1732 VM_BUG_ON(!PageKsm(page));
1733 VM_BUG_ON(!PageLocked(page));
1734
1735 stable_node = page_stable_node(page);
1736 if (!stable_node)
1737 return ret;
1738again:
1739 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1740 struct anon_vma *anon_vma = rmap_item->anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -08001741 struct anon_vma_chain *vmac;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001742 struct vm_area_struct *vma;
1743
Rik van Rielcba48b92010-08-09 17:18:38 -07001744 anon_vma_lock(anon_vma);
Michel Lespinassebf181b92012-10-08 16:31:39 -07001745 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
1746 0, ULONG_MAX) {
Rik van Riel5beb4932010-03-05 13:42:07 -08001747 vma = vmac->vma;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001748 if (rmap_item->address < vma->vm_start ||
1749 rmap_item->address >= vma->vm_end)
1750 continue;
1751 /*
1752 * Initially we examine only the vma which covers this
1753 * rmap_item; but later, if there is still work to do,
1754 * we examine covering vmas in other mms: in case they
1755 * were forked from the original since ksmd passed.
1756 */
1757 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1758 continue;
1759
1760 ret = rmap_one(page, vma, rmap_item->address, arg);
1761 if (ret != SWAP_AGAIN) {
Rik van Rielcba48b92010-08-09 17:18:38 -07001762 anon_vma_unlock(anon_vma);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001763 goto out;
1764 }
1765 }
Rik van Rielcba48b92010-08-09 17:18:38 -07001766 anon_vma_unlock(anon_vma);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001767 }
1768 if (!search_new_forks++)
1769 goto again;
1770out:
1771 return ret;
1772}
1773
1774void ksm_migrate_page(struct page *newpage, struct page *oldpage)
1775{
1776 struct stable_node *stable_node;
1777
1778 VM_BUG_ON(!PageLocked(oldpage));
1779 VM_BUG_ON(!PageLocked(newpage));
1780 VM_BUG_ON(newpage->mapping != oldpage->mapping);
1781
1782 stable_node = page_stable_node(newpage);
1783 if (stable_node) {
Hugh Dickins62b61f62009-12-14 17:59:33 -08001784 VM_BUG_ON(stable_node->kpfn != page_to_pfn(oldpage));
1785 stable_node->kpfn = page_to_pfn(newpage);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001786 }
1787}
1788#endif /* CONFIG_MIGRATION */
1789
Hugh Dickins62b61f62009-12-14 17:59:33 -08001790#ifdef CONFIG_MEMORY_HOTREMOVE
1791static struct stable_node *ksm_check_stable_tree(unsigned long start_pfn,
1792 unsigned long end_pfn)
1793{
1794 struct rb_node *node;
1795
1796 for (node = rb_first(&root_stable_tree); node; node = rb_next(node)) {
1797 struct stable_node *stable_node;
1798
1799 stable_node = rb_entry(node, struct stable_node, node);
1800 if (stable_node->kpfn >= start_pfn &&
1801 stable_node->kpfn < end_pfn)
1802 return stable_node;
1803 }
1804 return NULL;
1805}
1806
1807static int ksm_memory_callback(struct notifier_block *self,
1808 unsigned long action, void *arg)
1809{
1810 struct memory_notify *mn = arg;
1811 struct stable_node *stable_node;
1812
1813 switch (action) {
1814 case MEM_GOING_OFFLINE:
1815 /*
1816 * Keep it very simple for now: just lock out ksmd and
1817 * MADV_UNMERGEABLE while any memory is going offline.
KOSAKI Motohiroa0b0f582010-12-02 14:31:20 -08001818 * mutex_lock_nested() is necessary because lockdep was alarmed
1819 * that here we take ksm_thread_mutex inside notifier chain
1820 * mutex, and later take notifier chain mutex inside
1821 * ksm_thread_mutex to unlock it. But that's safe because both
1822 * are inside mem_hotplug_mutex.
Hugh Dickins62b61f62009-12-14 17:59:33 -08001823 */
KOSAKI Motohiroa0b0f582010-12-02 14:31:20 -08001824 mutex_lock_nested(&ksm_thread_mutex, SINGLE_DEPTH_NESTING);
Hugh Dickins62b61f62009-12-14 17:59:33 -08001825 break;
1826
1827 case MEM_OFFLINE:
1828 /*
1829 * Most of the work is done by page migration; but there might
1830 * be a few stable_nodes left over, still pointing to struct
1831 * pages which have been offlined: prune those from the tree.
1832 */
1833 while ((stable_node = ksm_check_stable_tree(mn->start_pfn,
1834 mn->start_pfn + mn->nr_pages)) != NULL)
1835 remove_node_from_stable_tree(stable_node);
1836 /* fallthrough */
1837
1838 case MEM_CANCEL_OFFLINE:
1839 mutex_unlock(&ksm_thread_mutex);
1840 break;
1841 }
1842 return NOTIFY_OK;
1843}
1844#endif /* CONFIG_MEMORY_HOTREMOVE */
1845
Hugh Dickins2ffd8672009-09-21 17:02:23 -07001846#ifdef CONFIG_SYSFS
1847/*
1848 * This all compiles without CONFIG_SYSFS, but is a waste of space.
1849 */
1850
Izik Eidus31dbd012009-09-21 17:02:03 -07001851#define KSM_ATTR_RO(_name) \
1852 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1853#define KSM_ATTR(_name) \
1854 static struct kobj_attribute _name##_attr = \
1855 __ATTR(_name, 0644, _name##_show, _name##_store)
1856
1857static ssize_t sleep_millisecs_show(struct kobject *kobj,
1858 struct kobj_attribute *attr, char *buf)
1859{
1860 return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
1861}
1862
1863static ssize_t sleep_millisecs_store(struct kobject *kobj,
1864 struct kobj_attribute *attr,
1865 const char *buf, size_t count)
1866{
1867 unsigned long msecs;
1868 int err;
1869
1870 err = strict_strtoul(buf, 10, &msecs);
1871 if (err || msecs > UINT_MAX)
1872 return -EINVAL;
1873
1874 ksm_thread_sleep_millisecs = msecs;
1875
1876 return count;
1877}
1878KSM_ATTR(sleep_millisecs);
1879
1880static ssize_t pages_to_scan_show(struct kobject *kobj,
1881 struct kobj_attribute *attr, char *buf)
1882{
1883 return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
1884}
1885
1886static ssize_t pages_to_scan_store(struct kobject *kobj,
1887 struct kobj_attribute *attr,
1888 const char *buf, size_t count)
1889{
1890 int err;
1891 unsigned long nr_pages;
1892
1893 err = strict_strtoul(buf, 10, &nr_pages);
1894 if (err || nr_pages > UINT_MAX)
1895 return -EINVAL;
1896
1897 ksm_thread_pages_to_scan = nr_pages;
1898
1899 return count;
1900}
1901KSM_ATTR(pages_to_scan);
1902
1903static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
1904 char *buf)
1905{
1906 return sprintf(buf, "%u\n", ksm_run);
1907}
1908
1909static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
1910 const char *buf, size_t count)
1911{
1912 int err;
1913 unsigned long flags;
1914
1915 err = strict_strtoul(buf, 10, &flags);
1916 if (err || flags > UINT_MAX)
1917 return -EINVAL;
1918 if (flags > KSM_RUN_UNMERGE)
1919 return -EINVAL;
1920
1921 /*
1922 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
1923 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
Hugh Dickinsd0f209f2009-12-14 17:59:34 -08001924 * breaking COW to free the pages_shared (but leaves mm_slots
1925 * on the list for when ksmd may be set running again).
Izik Eidus31dbd012009-09-21 17:02:03 -07001926 */
1927
1928 mutex_lock(&ksm_thread_mutex);
1929 if (ksm_run != flags) {
1930 ksm_run = flags;
Hugh Dickinsd952b792009-09-21 17:02:16 -07001931 if (flags & KSM_RUN_UNMERGE) {
David Rientjes72788c32011-05-24 17:11:40 -07001932 int oom_score_adj;
1933
1934 oom_score_adj = test_set_oom_score_adj(OOM_SCORE_ADJ_MAX);
Hugh Dickinsd952b792009-09-21 17:02:16 -07001935 err = unmerge_and_remove_all_rmap_items();
David Rientjes43362a42011-10-31 17:07:18 -07001936 compare_swap_oom_score_adj(OOM_SCORE_ADJ_MAX,
1937 oom_score_adj);
Hugh Dickinsd952b792009-09-21 17:02:16 -07001938 if (err) {
1939 ksm_run = KSM_RUN_STOP;
1940 count = err;
1941 }
1942 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001943 }
1944 mutex_unlock(&ksm_thread_mutex);
1945
1946 if (flags & KSM_RUN_MERGE)
1947 wake_up_interruptible(&ksm_thread_wait);
1948
1949 return count;
1950}
1951KSM_ATTR(run);
1952
Hugh Dickinsb4028262009-09-21 17:02:09 -07001953static ssize_t pages_shared_show(struct kobject *kobj,
1954 struct kobj_attribute *attr, char *buf)
1955{
1956 return sprintf(buf, "%lu\n", ksm_pages_shared);
1957}
1958KSM_ATTR_RO(pages_shared);
1959
1960static ssize_t pages_sharing_show(struct kobject *kobj,
1961 struct kobj_attribute *attr, char *buf)
1962{
Hugh Dickinse178dfd2009-09-21 17:02:10 -07001963 return sprintf(buf, "%lu\n", ksm_pages_sharing);
Hugh Dickinsb4028262009-09-21 17:02:09 -07001964}
1965KSM_ATTR_RO(pages_sharing);
1966
Hugh Dickins473b0ce2009-09-21 17:02:11 -07001967static ssize_t pages_unshared_show(struct kobject *kobj,
1968 struct kobj_attribute *attr, char *buf)
1969{
1970 return sprintf(buf, "%lu\n", ksm_pages_unshared);
1971}
1972KSM_ATTR_RO(pages_unshared);
1973
1974static ssize_t pages_volatile_show(struct kobject *kobj,
1975 struct kobj_attribute *attr, char *buf)
1976{
1977 long ksm_pages_volatile;
1978
1979 ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
1980 - ksm_pages_sharing - ksm_pages_unshared;
1981 /*
1982 * It was not worth any locking to calculate that statistic,
1983 * but it might therefore sometimes be negative: conceal that.
1984 */
1985 if (ksm_pages_volatile < 0)
1986 ksm_pages_volatile = 0;
1987 return sprintf(buf, "%ld\n", ksm_pages_volatile);
1988}
1989KSM_ATTR_RO(pages_volatile);
1990
1991static ssize_t full_scans_show(struct kobject *kobj,
1992 struct kobj_attribute *attr, char *buf)
1993{
1994 return sprintf(buf, "%lu\n", ksm_scan.seqnr);
1995}
1996KSM_ATTR_RO(full_scans);
1997
Izik Eidus31dbd012009-09-21 17:02:03 -07001998static struct attribute *ksm_attrs[] = {
1999 &sleep_millisecs_attr.attr,
2000 &pages_to_scan_attr.attr,
2001 &run_attr.attr,
Hugh Dickinsb4028262009-09-21 17:02:09 -07002002 &pages_shared_attr.attr,
2003 &pages_sharing_attr.attr,
Hugh Dickins473b0ce2009-09-21 17:02:11 -07002004 &pages_unshared_attr.attr,
2005 &pages_volatile_attr.attr,
2006 &full_scans_attr.attr,
Izik Eidus31dbd012009-09-21 17:02:03 -07002007 NULL,
2008};
2009
2010static struct attribute_group ksm_attr_group = {
2011 .attrs = ksm_attrs,
2012 .name = "ksm",
2013};
Hugh Dickins2ffd8672009-09-21 17:02:23 -07002014#endif /* CONFIG_SYSFS */
Izik Eidus31dbd012009-09-21 17:02:03 -07002015
2016static int __init ksm_init(void)
2017{
2018 struct task_struct *ksm_thread;
2019 int err;
2020
2021 err = ksm_slab_init();
2022 if (err)
2023 goto out;
2024
Izik Eidus31dbd012009-09-21 17:02:03 -07002025 ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
2026 if (IS_ERR(ksm_thread)) {
2027 printk(KERN_ERR "ksm: creating kthread failed\n");
2028 err = PTR_ERR(ksm_thread);
Lai Jiangshand9f89842010-08-09 17:20:02 -07002029 goto out_free;
Izik Eidus31dbd012009-09-21 17:02:03 -07002030 }
2031
Hugh Dickins2ffd8672009-09-21 17:02:23 -07002032#ifdef CONFIG_SYSFS
Izik Eidus31dbd012009-09-21 17:02:03 -07002033 err = sysfs_create_group(mm_kobj, &ksm_attr_group);
2034 if (err) {
2035 printk(KERN_ERR "ksm: register sysfs failed\n");
Hugh Dickins2ffd8672009-09-21 17:02:23 -07002036 kthread_stop(ksm_thread);
Lai Jiangshand9f89842010-08-09 17:20:02 -07002037 goto out_free;
Izik Eidus31dbd012009-09-21 17:02:03 -07002038 }
Hugh Dickinsc73602a2009-10-07 16:32:22 -07002039#else
2040 ksm_run = KSM_RUN_MERGE; /* no way for user to start it */
2041
Hugh Dickins2ffd8672009-09-21 17:02:23 -07002042#endif /* CONFIG_SYSFS */
Izik Eidus31dbd012009-09-21 17:02:03 -07002043
Hugh Dickins62b61f62009-12-14 17:59:33 -08002044#ifdef CONFIG_MEMORY_HOTREMOVE
2045 /*
2046 * Choose a high priority since the callback takes ksm_thread_mutex:
2047 * later callbacks could only be taking locks which nest within that.
2048 */
2049 hotplug_memory_notifier(ksm_memory_callback, 100);
2050#endif
Izik Eidus31dbd012009-09-21 17:02:03 -07002051 return 0;
2052
Lai Jiangshand9f89842010-08-09 17:20:02 -07002053out_free:
Izik Eidus31dbd012009-09-21 17:02:03 -07002054 ksm_slab_free();
2055out:
2056 return err;
2057}
2058module_init(ksm_init)