blob: f21f4a1d6a1ce144d2ce45c30123eb2010f93bb8 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * mm/rmap.c - physical to virtual reverse mappings
3 *
4 * Copyright 2001, Rik van Riel <riel@conectiva.com.br>
5 * Released under the General Public License (GPL).
6 *
7 * Simple, low overhead reverse mapping scheme.
8 * Please try to keep this thing as modular as possible.
9 *
10 * Provides methods for unmapping each kind of mapped page:
11 * the anon methods track anonymous pages, and
12 * the file methods track pages belonging to an inode.
13 *
14 * Original design by Rik van Riel <riel@conectiva.com.br> 2001
15 * File methods by Dave McCracken <dmccr@us.ibm.com> 2003, 2004
16 * Anonymous methods by Andrea Arcangeli <andrea@suse.de> 2004
Hugh Dickins98f32602009-05-21 20:33:58 +010017 * Contributions by Hugh Dickins 2003, 2004
Linus Torvalds1da177e2005-04-16 15:20:36 -070018 */
19
20/*
21 * Lock ordering in mm:
22 *
Jes Sorensen1b1dcc12006-01-09 15:59:24 -080023 * inode->i_mutex (while writing or truncating, not reading or faulting)
Nick Piggin82591e62006-10-19 23:29:10 -070024 * inode->i_alloc_sem (vmtruncate_range)
25 * mm->mmap_sem
26 * page->flags PG_locked (lock_page)
27 * mapping->i_mmap_lock
28 * anon_vma->lock
29 * mm->page_table_lock or pte_lock
30 * zone->lru_lock (in mark_page_accessed, isolate_lru_page)
31 * swap_lock (in swap_duplicate, swap_info_get)
32 * mmlist_lock (in mmput, drain_mmlist and others)
33 * mapping->private_lock (in __set_page_dirty_buffers)
34 * inode_lock (in set_page_dirty's __mark_inode_dirty)
35 * sb_lock (within inode_lock in fs/fs-writeback.c)
36 * mapping->tree_lock (widely used, in set_page_dirty,
37 * in arch-dependent flush_dcache_mmap_lock,
38 * within inode_lock in __sync_single_inode)
Andi Kleen6a460792009-09-16 11:50:15 +020039 *
40 * (code doesn't rely on that order so it could be switched around)
41 * ->tasklist_lock
42 * anon_vma->lock (memory_failure, collect_procs_anon)
43 * pte map lock
Linus Torvalds1da177e2005-04-16 15:20:36 -070044 */
45
46#include <linux/mm.h>
47#include <linux/pagemap.h>
48#include <linux/swap.h>
49#include <linux/swapops.h>
50#include <linux/slab.h>
51#include <linux/init.h>
Hugh Dickins5ad64682009-12-14 17:59:24 -080052#include <linux/ksm.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070053#include <linux/rmap.h>
54#include <linux/rcupdate.h>
Christoph Lametera48d07a2006-02-01 03:05:38 -080055#include <linux/module.h>
Balbir Singh8a9f3cc2008-02-07 00:13:53 -080056#include <linux/memcontrol.h>
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070057#include <linux/mmu_notifier.h>
KOSAKI Motohiro64cdd542009-01-06 14:39:16 -080058#include <linux/migrate.h>
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +090059#include <linux/hugetlb.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070060
61#include <asm/tlbflush.h>
62
Nick Pigginb291f002008-10-18 20:26:44 -070063#include "internal.h"
64
Adrian Bunkfdd2e5f2008-10-18 20:28:38 -070065static struct kmem_cache *anon_vma_cachep;
Rik van Riel5beb4932010-03-05 13:42:07 -080066static struct kmem_cache *anon_vma_chain_cachep;
Adrian Bunkfdd2e5f2008-10-18 20:28:38 -070067
68static inline struct anon_vma *anon_vma_alloc(void)
69{
70 return kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
71}
72
Hugh Dickinsdb114b82009-12-14 17:59:25 -080073void anon_vma_free(struct anon_vma *anon_vma)
Adrian Bunkfdd2e5f2008-10-18 20:28:38 -070074{
75 kmem_cache_free(anon_vma_cachep, anon_vma);
76}
Linus Torvalds1da177e2005-04-16 15:20:36 -070077
Rik van Riel5beb4932010-03-05 13:42:07 -080078static inline struct anon_vma_chain *anon_vma_chain_alloc(void)
79{
80 return kmem_cache_alloc(anon_vma_chain_cachep, GFP_KERNEL);
81}
82
Namhyung Kime574b5f2010-10-26 14:22:02 -070083static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
Rik van Riel5beb4932010-03-05 13:42:07 -080084{
85 kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
86}
87
Linus Torvaldsd9d332e2008-10-19 10:32:20 -070088/**
89 * anon_vma_prepare - attach an anon_vma to a memory region
90 * @vma: the memory region in question
91 *
92 * This makes sure the memory mapping described by 'vma' has
93 * an 'anon_vma' attached to it, so that we can associate the
94 * anonymous pages mapped into it with that anon_vma.
95 *
96 * The common case will be that we already have one, but if
Figo.zhang23a07902010-12-27 15:14:06 +010097 * not we either need to find an adjacent mapping that we
Linus Torvaldsd9d332e2008-10-19 10:32:20 -070098 * can re-use the anon_vma from (very common when the only
99 * reason for splitting a vma has been mprotect()), or we
100 * allocate a new one.
101 *
102 * Anon-vma allocations are very subtle, because we may have
103 * optimistically looked up an anon_vma in page_lock_anon_vma()
104 * and that may actually touch the spinlock even in the newly
105 * allocated vma (it depends on RCU to make sure that the
106 * anon_vma isn't actually destroyed).
107 *
108 * As a result, we need to do proper anon_vma locking even
109 * for the new allocation. At the same time, we do not want
110 * to do any locking for the common case of already having
111 * an anon_vma.
112 *
113 * This must be called with the mmap_sem held for reading.
114 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700115int anon_vma_prepare(struct vm_area_struct *vma)
116{
117 struct anon_vma *anon_vma = vma->anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -0800118 struct anon_vma_chain *avc;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700119
120 might_sleep();
121 if (unlikely(!anon_vma)) {
122 struct mm_struct *mm = vma->vm_mm;
Linus Torvaldsd9d332e2008-10-19 10:32:20 -0700123 struct anon_vma *allocated;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700124
Rik van Riel5beb4932010-03-05 13:42:07 -0800125 avc = anon_vma_chain_alloc();
126 if (!avc)
127 goto out_enomem;
128
Linus Torvalds1da177e2005-04-16 15:20:36 -0700129 anon_vma = find_mergeable_anon_vma(vma);
Linus Torvaldsd9d332e2008-10-19 10:32:20 -0700130 allocated = NULL;
131 if (!anon_vma) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700132 anon_vma = anon_vma_alloc();
133 if (unlikely(!anon_vma))
Rik van Riel5beb4932010-03-05 13:42:07 -0800134 goto out_enomem_free_avc;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700135 allocated = anon_vma;
Rik van Riel5c341ee12010-08-09 17:18:39 -0700136 /*
137 * This VMA had no anon_vma yet. This anon_vma is
138 * the root of any anon_vma tree that might form.
139 */
140 anon_vma->root = anon_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700141 }
142
Rik van Rielcba48b92010-08-09 17:18:38 -0700143 anon_vma_lock(anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700144 /* page_table_lock to protect against threads */
145 spin_lock(&mm->page_table_lock);
146 if (likely(!vma->anon_vma)) {
147 vma->anon_vma = anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -0800148 avc->anon_vma = anon_vma;
149 avc->vma = vma;
150 list_add(&avc->same_vma, &vma->anon_vma_chain);
Andrea Arcangeli26ba0cb2010-08-09 17:19:08 -0700151 list_add_tail(&avc->same_anon_vma, &anon_vma->head);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700152 allocated = NULL;
Oleg Nesterov31f2b0e2010-04-23 13:18:01 -0400153 avc = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700154 }
155 spin_unlock(&mm->page_table_lock);
Rik van Rielcba48b92010-08-09 17:18:38 -0700156 anon_vma_unlock(anon_vma);
Oleg Nesterov31f2b0e2010-04-23 13:18:01 -0400157
158 if (unlikely(allocated))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700159 anon_vma_free(allocated);
Oleg Nesterov31f2b0e2010-04-23 13:18:01 -0400160 if (unlikely(avc))
Rik van Riel5beb4932010-03-05 13:42:07 -0800161 anon_vma_chain_free(avc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700162 }
163 return 0;
Rik van Riel5beb4932010-03-05 13:42:07 -0800164
165 out_enomem_free_avc:
166 anon_vma_chain_free(avc);
167 out_enomem:
168 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700169}
170
Rik van Riel5beb4932010-03-05 13:42:07 -0800171static void anon_vma_chain_link(struct vm_area_struct *vma,
172 struct anon_vma_chain *avc,
173 struct anon_vma *anon_vma)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700174{
Rik van Riel5beb4932010-03-05 13:42:07 -0800175 avc->vma = vma;
176 avc->anon_vma = anon_vma;
177 list_add(&avc->same_vma, &vma->anon_vma_chain);
178
Rik van Rielcba48b92010-08-09 17:18:38 -0700179 anon_vma_lock(anon_vma);
Andrea Arcangeli05759d32011-01-13 15:46:53 -0800180 /*
181 * It's critical to add new vmas to the tail of the anon_vma,
182 * see comment in huge_memory.c:__split_huge_page().
183 */
Rik van Riel5beb4932010-03-05 13:42:07 -0800184 list_add_tail(&avc->same_anon_vma, &anon_vma->head);
Rik van Rielcba48b92010-08-09 17:18:38 -0700185 anon_vma_unlock(anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700186}
187
Rik van Riel5beb4932010-03-05 13:42:07 -0800188/*
189 * Attach the anon_vmas from src to dst.
190 * Returns 0 on success, -ENOMEM on failure.
191 */
192int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700193{
Rik van Riel5beb4932010-03-05 13:42:07 -0800194 struct anon_vma_chain *avc, *pavc;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700195
Linus Torvalds646d87b2010-04-11 17:15:03 -0700196 list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
Rik van Riel5beb4932010-03-05 13:42:07 -0800197 avc = anon_vma_chain_alloc();
198 if (!avc)
199 goto enomem_failure;
200 anon_vma_chain_link(dst, avc, pavc->anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700201 }
Rik van Riel5beb4932010-03-05 13:42:07 -0800202 return 0;
203
204 enomem_failure:
205 unlink_anon_vmas(dst);
206 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700207}
208
Rik van Riel5beb4932010-03-05 13:42:07 -0800209/*
210 * Attach vma to its own anon_vma, as well as to the anon_vmas that
211 * the corresponding VMA in the parent process is attached to.
212 * Returns 0 on success, non-zero on failure.
213 */
214int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700215{
Rik van Riel5beb4932010-03-05 13:42:07 -0800216 struct anon_vma_chain *avc;
217 struct anon_vma *anon_vma;
218
219 /* Don't bother if the parent process has no anon_vma here. */
220 if (!pvma->anon_vma)
221 return 0;
222
223 /*
224 * First, attach the new VMA to the parent VMA's anon_vmas,
225 * so rmap can find non-COWed pages in child processes.
226 */
227 if (anon_vma_clone(vma, pvma))
228 return -ENOMEM;
229
230 /* Then add our own anon_vma. */
231 anon_vma = anon_vma_alloc();
232 if (!anon_vma)
233 goto out_error;
234 avc = anon_vma_chain_alloc();
235 if (!avc)
236 goto out_error_free_anon_vma;
Rik van Riel5c341ee12010-08-09 17:18:39 -0700237
238 /*
239 * The root anon_vma's spinlock is the lock actually used when we
240 * lock any of the anon_vmas in this anon_vma tree.
241 */
242 anon_vma->root = pvma->anon_vma->root;
Rik van Riel76545062010-08-09 17:18:41 -0700243 /*
244 * With KSM refcounts, an anon_vma can stay around longer than the
245 * process it belongs to. The root anon_vma needs to be pinned
246 * until this anon_vma is freed, because the lock lives in the root.
247 */
248 get_anon_vma(anon_vma->root);
Rik van Riel5beb4932010-03-05 13:42:07 -0800249 /* Mark this anon_vma as the one where our new (COWed) pages go. */
250 vma->anon_vma = anon_vma;
Rik van Riel5c341ee12010-08-09 17:18:39 -0700251 anon_vma_chain_link(vma, avc, anon_vma);
Rik van Riel5beb4932010-03-05 13:42:07 -0800252
253 return 0;
254
255 out_error_free_anon_vma:
256 anon_vma_free(anon_vma);
257 out_error:
Rik van Riel4946d542010-04-05 12:13:33 -0400258 unlink_anon_vmas(vma);
Rik van Riel5beb4932010-03-05 13:42:07 -0800259 return -ENOMEM;
260}
261
262static void anon_vma_unlink(struct anon_vma_chain *anon_vma_chain)
263{
264 struct anon_vma *anon_vma = anon_vma_chain->anon_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700265 int empty;
266
Rik van Riel5beb4932010-03-05 13:42:07 -0800267 /* If anon_vma_fork fails, we can get an empty anon_vma_chain. */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700268 if (!anon_vma)
269 return;
270
Rik van Rielcba48b92010-08-09 17:18:38 -0700271 anon_vma_lock(anon_vma);
Rik van Riel5beb4932010-03-05 13:42:07 -0800272 list_del(&anon_vma_chain->same_anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700273
274 /* We must garbage collect the anon_vma if it's empty */
Mel Gorman7f60c212010-05-24 14:32:18 -0700275 empty = list_empty(&anon_vma->head) && !anonvma_external_refcount(anon_vma);
Rik van Rielcba48b92010-08-09 17:18:38 -0700276 anon_vma_unlock(anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700277
Rik van Riel76545062010-08-09 17:18:41 -0700278 if (empty) {
279 /* We no longer need the root anon_vma */
280 if (anon_vma->root != anon_vma)
281 drop_anon_vma(anon_vma->root);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700282 anon_vma_free(anon_vma);
Rik van Riel76545062010-08-09 17:18:41 -0700283 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700284}
285
Rik van Riel5beb4932010-03-05 13:42:07 -0800286void unlink_anon_vmas(struct vm_area_struct *vma)
287{
288 struct anon_vma_chain *avc, *next;
289
Rik van Riel5c341ee12010-08-09 17:18:39 -0700290 /*
291 * Unlink each anon_vma chained to the VMA. This list is ordered
292 * from newest to oldest, ensuring the root anon_vma gets freed last.
293 */
Rik van Riel5beb4932010-03-05 13:42:07 -0800294 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
295 anon_vma_unlink(avc);
296 list_del(&avc->same_vma);
297 anon_vma_chain_free(avc);
298 }
299}
300
Alexey Dobriyan51cc5062008-07-25 19:45:34 -0700301static void anon_vma_ctor(void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700302{
Christoph Lametera35afb82007-05-16 22:10:57 -0700303 struct anon_vma *anon_vma = data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700304
Christoph Lametera35afb82007-05-16 22:10:57 -0700305 spin_lock_init(&anon_vma->lock);
Mel Gorman7f60c212010-05-24 14:32:18 -0700306 anonvma_external_refcount_init(anon_vma);
Christoph Lametera35afb82007-05-16 22:10:57 -0700307 INIT_LIST_HEAD(&anon_vma->head);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700308}
309
310void __init anon_vma_init(void)
311{
312 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
Paul Mundt20c2df82007-07-20 10:11:58 +0900313 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor);
Rik van Riel5beb4932010-03-05 13:42:07 -0800314 anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain, SLAB_PANIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700315}
316
317/*
318 * Getting a lock on a stable anon_vma from a page off the LRU is
319 * tricky: page_lock_anon_vma rely on RCU to guard against the races.
320 */
Namhyung Kimea4525b2010-10-26 14:22:01 -0700321struct anon_vma *__page_lock_anon_vma(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700322{
Hugh Dickinsf18194272010-08-25 23:12:54 -0700323 struct anon_vma *anon_vma, *root_anon_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700324 unsigned long anon_mapping;
325
326 rcu_read_lock();
Hugh Dickins80e148222009-12-14 17:59:29 -0800327 anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping);
Hugh Dickins3ca7b3c2009-12-14 17:58:57 -0800328 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700329 goto out;
330 if (!page_mapped(page))
331 goto out;
332
333 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
Hugh Dickinsf18194272010-08-25 23:12:54 -0700334 root_anon_vma = ACCESS_ONCE(anon_vma->root);
335 spin_lock(&root_anon_vma->lock);
336
337 /*
338 * If this page is still mapped, then its anon_vma cannot have been
339 * freed. But if it has been unmapped, we have no security against
340 * the anon_vma structure being freed and reused (for another anon_vma:
341 * SLAB_DESTROY_BY_RCU guarantees that - so the spin_lock above cannot
342 * corrupt): with anon_vma_prepare() or anon_vma_fork() redirecting
343 * anon_vma->root before page_unlock_anon_vma() is called to unlock.
344 */
345 if (page_mapped(page))
346 return anon_vma;
347
348 spin_unlock(&root_anon_vma->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700349out:
350 rcu_read_unlock();
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800351 return NULL;
352}
353
Andi Kleen10be22d2009-09-16 11:50:04 +0200354void page_unlock_anon_vma(struct anon_vma *anon_vma)
Namhyung Kimea4525b2010-10-26 14:22:01 -0700355 __releases(&anon_vma->root->lock)
356 __releases(RCU)
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800357{
Rik van Rielcba48b92010-08-09 17:18:38 -0700358 anon_vma_unlock(anon_vma);
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800359 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700360}
361
362/*
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -0800363 * At what user virtual address is page expected in @vma?
364 * Returns virtual address or -EFAULT if page's index/offset is not
365 * within the range mapped the @vma.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700366 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800367inline unsigned long
Linus Torvalds1da177e2005-04-16 15:20:36 -0700368vma_address(struct page *page, struct vm_area_struct *vma)
369{
370 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
371 unsigned long address;
372
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900373 if (unlikely(is_vm_hugetlb_page(vma)))
374 pgoff = page->index << huge_page_order(page_hstate(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700375 address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
376 if (unlikely(address < vma->vm_start || address >= vma->vm_end)) {
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -0800377 /* page should be within @vma mapping range */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700378 return -EFAULT;
379 }
380 return address;
381}
382
383/*
Huang Shijiebf89c8c2009-10-01 15:44:04 -0700384 * At what user virtual address is page expected in vma?
Naoya Horiguchiab941e02010-05-11 14:06:55 -0700385 * Caller should check the page is actually part of the vma.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700386 */
387unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
388{
Andrea Arcangeli21d0d442010-08-09 17:19:10 -0700389 if (PageAnon(page)) {
Hugh Dickins4829b902010-10-02 17:46:06 -0700390 struct anon_vma *page__anon_vma = page_anon_vma(page);
391 /*
392 * Note: swapoff's unuse_vma() is more efficient with this
393 * check, and needs it to match anon_vma when KSM is active.
394 */
395 if (!vma->anon_vma || !page__anon_vma ||
396 vma->anon_vma->root != page__anon_vma->root)
Andrea Arcangeli21d0d442010-08-09 17:19:10 -0700397 return -EFAULT;
398 } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) {
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800399 if (!vma->vm_file ||
400 vma->vm_file->f_mapping != page->mapping)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700401 return -EFAULT;
402 } else
403 return -EFAULT;
404 return vma_address(page, vma);
405}
406
407/*
Nikita Danilov81b40822005-05-01 08:58:36 -0700408 * Check that @page is mapped at @address into @mm.
409 *
Nick Piggin479db0b2008-08-20 14:09:18 -0700410 * If @sync is false, page_check_address may perform a racy check to avoid
411 * the page table lock when the pte is not present (helpful when reclaiming
412 * highly shared pages).
413 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -0700414 * On success returns with pte mapped and locked.
Nikita Danilov81b40822005-05-01 08:58:36 -0700415 */
Namhyung Kime9a81a82010-10-26 14:22:01 -0700416pte_t *__page_check_address(struct page *page, struct mm_struct *mm,
Nick Piggin479db0b2008-08-20 14:09:18 -0700417 unsigned long address, spinlock_t **ptlp, int sync)
Nikita Danilov81b40822005-05-01 08:58:36 -0700418{
419 pgd_t *pgd;
420 pud_t *pud;
421 pmd_t *pmd;
422 pte_t *pte;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700423 spinlock_t *ptl;
Nikita Danilov81b40822005-05-01 08:58:36 -0700424
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900425 if (unlikely(PageHuge(page))) {
426 pte = huge_pte_offset(mm, address);
427 ptl = &mm->page_table_lock;
428 goto check;
429 }
430
Nikita Danilov81b40822005-05-01 08:58:36 -0700431 pgd = pgd_offset(mm, address);
Hugh Dickinsc0718802005-10-29 18:16:31 -0700432 if (!pgd_present(*pgd))
433 return NULL;
434
435 pud = pud_offset(pgd, address);
436 if (!pud_present(*pud))
437 return NULL;
438
439 pmd = pmd_offset(pud, address);
440 if (!pmd_present(*pmd))
441 return NULL;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800442 if (pmd_trans_huge(*pmd))
443 return NULL;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700444
445 pte = pte_offset_map(pmd, address);
446 /* Make a quick check before getting the lock */
Nick Piggin479db0b2008-08-20 14:09:18 -0700447 if (!sync && !pte_present(*pte)) {
Hugh Dickinsc0718802005-10-29 18:16:31 -0700448 pte_unmap(pte);
449 return NULL;
Nikita Danilov81b40822005-05-01 08:58:36 -0700450 }
Hugh Dickinsc0718802005-10-29 18:16:31 -0700451
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700452 ptl = pte_lockptr(mm, pmd);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900453check:
Hugh Dickinsc0718802005-10-29 18:16:31 -0700454 spin_lock(ptl);
455 if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
456 *ptlp = ptl;
457 return pte;
458 }
459 pte_unmap_unlock(pte, ptl);
460 return NULL;
Nikita Danilov81b40822005-05-01 08:58:36 -0700461}
462
Nick Pigginb291f002008-10-18 20:26:44 -0700463/**
464 * page_mapped_in_vma - check whether a page is really mapped in a VMA
465 * @page: the page to test
466 * @vma: the VMA to test
467 *
468 * Returns 1 if the page is mapped into the page tables of the VMA, 0
469 * if the page is not mapped into the page tables of this VMA. Only
470 * valid for normal file or anonymous VMAs.
471 */
Andi Kleen6a460792009-09-16 11:50:15 +0200472int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
Nick Pigginb291f002008-10-18 20:26:44 -0700473{
474 unsigned long address;
475 pte_t *pte;
476 spinlock_t *ptl;
477
478 address = vma_address(page, vma);
479 if (address == -EFAULT) /* out of vma range */
480 return 0;
481 pte = page_check_address(page, vma->vm_mm, address, &ptl, 1);
482 if (!pte) /* the page is not in this mm */
483 return 0;
484 pte_unmap_unlock(pte, ptl);
485
486 return 1;
487}
488
Nikita Danilov81b40822005-05-01 08:58:36 -0700489/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700490 * Subfunctions of page_referenced: page_referenced_one called
491 * repeatedly from either page_referenced_anon or page_referenced_file.
492 */
Hugh Dickins5ad64682009-12-14 17:59:24 -0800493int page_referenced_one(struct page *page, struct vm_area_struct *vma,
494 unsigned long address, unsigned int *mapcount,
495 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700496{
497 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700498 int referenced = 0;
499
Nick Pigginb291f002008-10-18 20:26:44 -0700500 /*
501 * Don't want to elevate referenced for mlocked page that gets this far,
502 * in order that it progresses to try_to_unmap and is moved to the
503 * unevictable list.
504 */
Hugh Dickins5a9bbdc2008-02-04 22:29:23 -0800505 if (vma->vm_flags & VM_LOCKED) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800506 *mapcount = 0; /* break early from loop */
Minchan Kim03ef83a2009-08-26 14:29:23 -0700507 *vm_flags |= VM_LOCKED;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800508 goto out;
Johannes Weiner4917e5d2009-01-06 14:39:17 -0800509 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700510
Hugh Dickinsc0718802005-10-29 18:16:31 -0700511 /* Pretend the page is referenced if the task has the
512 swap token and is in the middle of a page fault. */
Rik van Rielf7b7fd82005-11-28 13:44:07 -0800513 if (mm != current->mm && has_swap_token(mm) &&
Hugh Dickinsc0718802005-10-29 18:16:31 -0700514 rwsem_is_locked(&mm->mmap_sem))
515 referenced++;
516
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800517 if (unlikely(PageTransHuge(page))) {
518 pmd_t *pmd;
519
520 spin_lock(&mm->page_table_lock);
521 pmd = page_check_address_pmd(page, mm, address,
522 PAGE_CHECK_ADDRESS_PMD_FLAG);
523 if (pmd && !pmd_trans_splitting(*pmd) &&
524 pmdp_clear_flush_young_notify(vma, address, pmd))
525 referenced++;
526 spin_unlock(&mm->page_table_lock);
527 } else {
528 pte_t *pte;
529 spinlock_t *ptl;
530
531 pte = page_check_address(page, mm, address, &ptl, 0);
532 if (!pte)
533 goto out;
534
535 if (ptep_clear_flush_young_notify(vma, address, pte)) {
536 /*
537 * Don't treat a reference through a sequentially read
538 * mapping as such. If the page has been used in
539 * another mapping, we will catch it; if this other
540 * mapping is already gone, the unmap path will have
541 * set PG_referenced or activated the page.
542 */
543 if (likely(!VM_SequentialReadHint(vma)))
544 referenced++;
545 }
546 pte_unmap_unlock(pte, ptl);
547 }
548
Hugh Dickinsc0718802005-10-29 18:16:31 -0700549 (*mapcount)--;
Huang Shijie273f0472009-12-14 17:58:51 -0800550
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700551 if (referenced)
552 *vm_flags |= vma->vm_flags;
Huang Shijie273f0472009-12-14 17:58:51 -0800553out:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700554 return referenced;
555}
556
Balbir Singhbed71612008-02-07 00:14:01 -0800557static int page_referenced_anon(struct page *page,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700558 struct mem_cgroup *mem_cont,
559 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700560{
561 unsigned int mapcount;
562 struct anon_vma *anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -0800563 struct anon_vma_chain *avc;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700564 int referenced = 0;
565
566 anon_vma = page_lock_anon_vma(page);
567 if (!anon_vma)
568 return referenced;
569
570 mapcount = page_mapcount(page);
Rik van Riel5beb4932010-03-05 13:42:07 -0800571 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
572 struct vm_area_struct *vma = avc->vma;
Hugh Dickins1cb17292009-12-14 17:59:01 -0800573 unsigned long address = vma_address(page, vma);
574 if (address == -EFAULT)
575 continue;
Balbir Singhbed71612008-02-07 00:14:01 -0800576 /*
577 * If we are reclaiming on behalf of a cgroup, skip
578 * counting on behalf of references from different
579 * cgroups
580 */
Hugh Dickinsbd845e32008-03-04 14:29:01 -0800581 if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
Balbir Singhbed71612008-02-07 00:14:01 -0800582 continue;
Hugh Dickins1cb17292009-12-14 17:59:01 -0800583 referenced += page_referenced_one(page, vma, address,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700584 &mapcount, vm_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700585 if (!mapcount)
586 break;
587 }
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800588
589 page_unlock_anon_vma(anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700590 return referenced;
591}
592
593/**
594 * page_referenced_file - referenced check for object-based rmap
595 * @page: the page we're checking references on.
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700596 * @mem_cont: target memory controller
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700597 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
Linus Torvalds1da177e2005-04-16 15:20:36 -0700598 *
599 * For an object-based mapped page, find all the places it is mapped and
600 * check/clear the referenced flag. This is done by following the page->mapping
601 * pointer, then walking the chain of vmas it holds. It returns the number
602 * of references it found.
603 *
604 * This function is only called from page_referenced for object-based pages.
605 */
Balbir Singhbed71612008-02-07 00:14:01 -0800606static int page_referenced_file(struct page *page,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700607 struct mem_cgroup *mem_cont,
608 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700609{
610 unsigned int mapcount;
611 struct address_space *mapping = page->mapping;
612 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
613 struct vm_area_struct *vma;
614 struct prio_tree_iter iter;
615 int referenced = 0;
616
617 /*
618 * The caller's checks on page->mapping and !PageAnon have made
619 * sure that this is a file page: the check for page->mapping
620 * excludes the case just before it gets set on an anon page.
621 */
622 BUG_ON(PageAnon(page));
623
624 /*
625 * The page lock not only makes sure that page->mapping cannot
626 * suddenly be NULLified by truncation, it makes sure that the
627 * structure at mapping cannot be freed and reused yet,
628 * so we can safely take mapping->i_mmap_lock.
629 */
630 BUG_ON(!PageLocked(page));
631
632 spin_lock(&mapping->i_mmap_lock);
633
634 /*
635 * i_mmap_lock does not stabilize mapcount at all, but mapcount
636 * is more likely to be accurate if we note it after spinning.
637 */
638 mapcount = page_mapcount(page);
639
640 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
Hugh Dickins1cb17292009-12-14 17:59:01 -0800641 unsigned long address = vma_address(page, vma);
642 if (address == -EFAULT)
643 continue;
Balbir Singhbed71612008-02-07 00:14:01 -0800644 /*
645 * If we are reclaiming on behalf of a cgroup, skip
646 * counting on behalf of references from different
647 * cgroups
648 */
Hugh Dickinsbd845e32008-03-04 14:29:01 -0800649 if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
Balbir Singhbed71612008-02-07 00:14:01 -0800650 continue;
Hugh Dickins1cb17292009-12-14 17:59:01 -0800651 referenced += page_referenced_one(page, vma, address,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700652 &mapcount, vm_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700653 if (!mapcount)
654 break;
655 }
656
657 spin_unlock(&mapping->i_mmap_lock);
658 return referenced;
659}
660
661/**
662 * page_referenced - test if the page was referenced
663 * @page: the page to test
664 * @is_locked: caller holds lock on the page
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700665 * @mem_cont: target memory controller
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700666 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
Linus Torvalds1da177e2005-04-16 15:20:36 -0700667 *
668 * Quick test_and_clear_referenced for all mappings to a page,
669 * returns the number of ptes which referenced the page.
670 */
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700671int page_referenced(struct page *page,
672 int is_locked,
673 struct mem_cgroup *mem_cont,
674 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700675{
676 int referenced = 0;
Hugh Dickins5ad64682009-12-14 17:59:24 -0800677 int we_locked = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700678
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700679 *vm_flags = 0;
Hugh Dickins3ca7b3c2009-12-14 17:58:57 -0800680 if (page_mapped(page) && page_rmapping(page)) {
Hugh Dickins5ad64682009-12-14 17:59:24 -0800681 if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
682 we_locked = trylock_page(page);
683 if (!we_locked) {
684 referenced++;
685 goto out;
686 }
687 }
688 if (unlikely(PageKsm(page)))
689 referenced += page_referenced_ksm(page, mem_cont,
690 vm_flags);
691 else if (PageAnon(page))
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700692 referenced += page_referenced_anon(page, mem_cont,
693 vm_flags);
Hugh Dickins5ad64682009-12-14 17:59:24 -0800694 else if (page->mapping)
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700695 referenced += page_referenced_file(page, mem_cont,
696 vm_flags);
Hugh Dickins5ad64682009-12-14 17:59:24 -0800697 if (we_locked)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700698 unlock_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700699 }
Hugh Dickins5ad64682009-12-14 17:59:24 -0800700out:
Christian Borntraeger5b7baf02008-03-25 18:47:12 +0100701 if (page_test_and_clear_young(page))
702 referenced++;
703
Linus Torvalds1da177e2005-04-16 15:20:36 -0700704 return referenced;
705}
706
Hugh Dickins1cb17292009-12-14 17:59:01 -0800707static int page_mkclean_one(struct page *page, struct vm_area_struct *vma,
708 unsigned long address)
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700709{
710 struct mm_struct *mm = vma->vm_mm;
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100711 pte_t *pte;
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700712 spinlock_t *ptl;
713 int ret = 0;
714
Nick Piggin479db0b2008-08-20 14:09:18 -0700715 pte = page_check_address(page, mm, address, &ptl, 1);
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700716 if (!pte)
717 goto out;
718
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100719 if (pte_dirty(*pte) || pte_write(*pte)) {
720 pte_t entry;
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700721
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100722 flush_cache_page(vma, address, pte_pfn(*pte));
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700723 entry = ptep_clear_flush_notify(vma, address, pte);
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100724 entry = pte_wrprotect(entry);
725 entry = pte_mkclean(entry);
Al Virod6e88e62006-12-29 16:48:35 -0800726 set_pte_at(mm, address, pte, entry);
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100727 ret = 1;
728 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700729
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700730 pte_unmap_unlock(pte, ptl);
731out:
732 return ret;
733}
734
735static int page_mkclean_file(struct address_space *mapping, struct page *page)
736{
737 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
738 struct vm_area_struct *vma;
739 struct prio_tree_iter iter;
740 int ret = 0;
741
742 BUG_ON(PageAnon(page));
743
744 spin_lock(&mapping->i_mmap_lock);
745 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
Hugh Dickins1cb17292009-12-14 17:59:01 -0800746 if (vma->vm_flags & VM_SHARED) {
747 unsigned long address = vma_address(page, vma);
748 if (address == -EFAULT)
749 continue;
750 ret += page_mkclean_one(page, vma, address);
751 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700752 }
753 spin_unlock(&mapping->i_mmap_lock);
754 return ret;
755}
756
757int page_mkclean(struct page *page)
758{
759 int ret = 0;
760
761 BUG_ON(!PageLocked(page));
762
763 if (page_mapped(page)) {
764 struct address_space *mapping = page_mapping(page);
Christian Borntraegerce7e9fa2007-11-20 11:13:36 +0100765 if (mapping) {
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700766 ret = page_mkclean_file(mapping, page);
Christian Borntraegerce7e9fa2007-11-20 11:13:36 +0100767 if (page_test_dirty(page)) {
Martin Schwidefskye2b8d7a2010-10-25 16:10:14 +0200768 page_clear_dirty(page, 1);
Christian Borntraegerce7e9fa2007-11-20 11:13:36 +0100769 ret = 1;
770 }
Martin Schwidefsky6c210482007-04-27 16:01:57 +0200771 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700772 }
773
774 return ret;
775}
Jaya Kumar60b59be2007-05-08 00:37:37 -0700776EXPORT_SYMBOL_GPL(page_mkclean);
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700777
Linus Torvalds1da177e2005-04-16 15:20:36 -0700778/**
Rik van Rielc44b6742010-03-05 13:42:09 -0800779 * page_move_anon_rmap - move a page to our anon_vma
780 * @page: the page to move to our anon_vma
781 * @vma: the vma the page belongs to
782 * @address: the user virtual address mapped
783 *
784 * When a page belongs exclusively to one process after a COW event,
785 * that page can be moved into the anon_vma that belongs to just that
786 * process, so the rmap code will not search the parent or sibling
787 * processes.
788 */
789void page_move_anon_rmap(struct page *page,
790 struct vm_area_struct *vma, unsigned long address)
791{
792 struct anon_vma *anon_vma = vma->anon_vma;
793
794 VM_BUG_ON(!PageLocked(page));
795 VM_BUG_ON(!anon_vma);
796 VM_BUG_ON(page->index != linear_page_index(vma, address));
797
798 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
799 page->mapping = (struct address_space *) anon_vma;
800}
801
802/**
Andi Kleen4e1c1972010-09-22 12:43:56 +0200803 * __page_set_anon_rmap - set up new anonymous rmap
804 * @page: Page to add to rmap
805 * @vma: VM area to add page to.
806 * @address: User virtual address of the mapping
Rik van Riele8a03fe2010-04-14 17:59:28 -0400807 * @exclusive: the page is exclusively owned by the current process
Nick Piggin9617d952006-01-06 00:11:12 -0800808 */
809static void __page_set_anon_rmap(struct page *page,
Rik van Riele8a03fe2010-04-14 17:59:28 -0400810 struct vm_area_struct *vma, unsigned long address, int exclusive)
Nick Piggin9617d952006-01-06 00:11:12 -0800811{
Rik van Riele8a03fe2010-04-14 17:59:28 -0400812 struct anon_vma *anon_vma = vma->anon_vma;
Nick Piggin9617d952006-01-06 00:11:12 -0800813
Rik van Riele8a03fe2010-04-14 17:59:28 -0400814 BUG_ON(!anon_vma);
Linus Torvaldsea900022010-04-12 12:44:29 -0700815
Andi Kleen4e1c1972010-09-22 12:43:56 +0200816 if (PageAnon(page))
817 return;
818
Linus Torvaldsea900022010-04-12 12:44:29 -0700819 /*
Rik van Riele8a03fe2010-04-14 17:59:28 -0400820 * If the page isn't exclusively mapped into this vma,
821 * we must use the _oldest_ possible anon_vma for the
822 * page mapping!
Linus Torvaldsea900022010-04-12 12:44:29 -0700823 */
Andi Kleen4e1c1972010-09-22 12:43:56 +0200824 if (!exclusive)
Andrea Arcangeli288468c2010-08-09 17:19:09 -0700825 anon_vma = anon_vma->root;
Linus Torvaldsea900022010-04-12 12:44:29 -0700826
Nick Piggin9617d952006-01-06 00:11:12 -0800827 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
828 page->mapping = (struct address_space *) anon_vma;
Nick Piggin9617d952006-01-06 00:11:12 -0800829 page->index = linear_page_index(vma, address);
Nick Piggin9617d952006-01-06 00:11:12 -0800830}
831
832/**
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700833 * __page_check_anon_rmap - sanity check anonymous rmap addition
Nick Pigginc97a9e12007-05-16 22:11:21 -0700834 * @page: the page to add the mapping to
835 * @vma: the vm area in which the mapping is added
836 * @address: the user virtual address mapped
837 */
838static void __page_check_anon_rmap(struct page *page,
839 struct vm_area_struct *vma, unsigned long address)
840{
841#ifdef CONFIG_DEBUG_VM
842 /*
843 * The page's anon-rmap details (mapping and index) are guaranteed to
844 * be set up correctly at this point.
845 *
846 * We have exclusion against page_add_anon_rmap because the caller
847 * always holds the page locked, except if called from page_dup_rmap,
848 * in which case the page is already known to be setup.
849 *
850 * We have exclusion against page_add_new_anon_rmap because those pages
851 * are initially only visible via the pagetables, and the pte is locked
852 * over the call to page_add_new_anon_rmap.
853 */
Andrea Arcangeli44ab57a2010-08-09 17:19:10 -0700854 BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root);
Nick Pigginc97a9e12007-05-16 22:11:21 -0700855 BUG_ON(page->index != linear_page_index(vma, address));
856#endif
857}
858
859/**
Linus Torvalds1da177e2005-04-16 15:20:36 -0700860 * page_add_anon_rmap - add pte mapping to an anonymous page
861 * @page: the page to add the mapping to
862 * @vma: the vm area in which the mapping is added
863 * @address: the user virtual address mapped
864 *
Hugh Dickins5ad64682009-12-14 17:59:24 -0800865 * The caller needs to hold the pte lock, and the page must be locked in
Hugh Dickins80e148222009-12-14 17:59:29 -0800866 * the anon_vma case: to serialize mapping,index checking after setting,
867 * and to ensure that PageAnon is not being upgraded racily to PageKsm
868 * (but PageKsm is never downgraded to PageAnon).
Linus Torvalds1da177e2005-04-16 15:20:36 -0700869 */
870void page_add_anon_rmap(struct page *page,
871 struct vm_area_struct *vma, unsigned long address)
872{
Rik van Rielad8c2ee2010-08-09 17:19:48 -0700873 do_page_add_anon_rmap(page, vma, address, 0);
874}
875
876/*
877 * Special version of the above for do_swap_page, which often runs
878 * into pages that are exclusively owned by the current process.
879 * Everybody else should continue to use page_add_anon_rmap above.
880 */
881void do_page_add_anon_rmap(struct page *page,
882 struct vm_area_struct *vma, unsigned long address, int exclusive)
883{
Hugh Dickins5ad64682009-12-14 17:59:24 -0800884 int first = atomic_inc_and_test(&page->_mapcount);
Andrea Arcangeli79134172011-01-13 15:46:58 -0800885 if (first) {
886 if (!PageTransHuge(page))
887 __inc_zone_page_state(page, NR_ANON_PAGES);
888 else
889 __inc_zone_page_state(page,
890 NR_ANON_TRANSPARENT_HUGEPAGES);
891 }
Hugh Dickins5ad64682009-12-14 17:59:24 -0800892 if (unlikely(PageKsm(page)))
893 return;
894
Nick Pigginc97a9e12007-05-16 22:11:21 -0700895 VM_BUG_ON(!PageLocked(page));
896 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
Hugh Dickins5ad64682009-12-14 17:59:24 -0800897 if (first)
Rik van Rielad8c2ee2010-08-09 17:19:48 -0700898 __page_set_anon_rmap(page, vma, address, exclusive);
KAMEZAWA Hiroyuki69029cd2008-07-25 01:47:14 -0700899 else
Nick Pigginc97a9e12007-05-16 22:11:21 -0700900 __page_check_anon_rmap(page, vma, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700901}
902
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700903/**
Nick Piggin9617d952006-01-06 00:11:12 -0800904 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
905 * @page: the page to add the mapping to
906 * @vma: the vm area in which the mapping is added
907 * @address: the user virtual address mapped
908 *
909 * Same as page_add_anon_rmap but must only be called on *new* pages.
910 * This means the inc-and-test can be bypassed.
Nick Pigginc97a9e12007-05-16 22:11:21 -0700911 * Page does not have to be locked.
Nick Piggin9617d952006-01-06 00:11:12 -0800912 */
913void page_add_new_anon_rmap(struct page *page,
914 struct vm_area_struct *vma, unsigned long address)
915{
Hugh Dickinsb5934c52009-01-06 14:39:25 -0800916 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
Hugh Dickinscbf84b72009-01-06 14:39:27 -0800917 SetPageSwapBacked(page);
918 atomic_set(&page->_mapcount, 0); /* increment count (starts at -1) */
Andrea Arcangeli79134172011-01-13 15:46:58 -0800919 if (!PageTransHuge(page))
920 __inc_zone_page_state(page, NR_ANON_PAGES);
921 else
922 __inc_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
Rik van Riele8a03fe2010-04-14 17:59:28 -0400923 __page_set_anon_rmap(page, vma, address, 1);
Hugh Dickinsb5934c52009-01-06 14:39:25 -0800924 if (page_evictable(page, vma))
Hugh Dickinscbf84b72009-01-06 14:39:27 -0800925 lru_cache_add_lru(page, LRU_ACTIVE_ANON);
Hugh Dickinsb5934c52009-01-06 14:39:25 -0800926 else
927 add_page_to_unevictable_list(page);
Nick Piggin9617d952006-01-06 00:11:12 -0800928}
929
Linus Torvalds1da177e2005-04-16 15:20:36 -0700930/**
931 * page_add_file_rmap - add pte mapping to a file page
932 * @page: the page to add the mapping to
933 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -0700934 * The caller needs to hold the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700935 */
936void page_add_file_rmap(struct page *page)
937{
Balbir Singhd69b0422009-06-17 16:26:34 -0700938 if (atomic_inc_and_test(&page->_mapcount)) {
Christoph Lameter65ba55f2006-06-30 01:55:34 -0700939 __inc_zone_page_state(page, NR_FILE_MAPPED);
Greg Thelen2a7106f2011-01-13 15:47:37 -0800940 mem_cgroup_inc_page_stat(page, MEMCG_NR_FILE_MAPPED);
Balbir Singhd69b0422009-06-17 16:26:34 -0700941 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700942}
943
944/**
945 * page_remove_rmap - take down pte mapping from a page
946 * @page: page to remove mapping from
947 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -0700948 * The caller needs to hold the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700949 */
Hugh Dickinsedc315f2009-01-06 14:40:11 -0800950void page_remove_rmap(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700951{
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -0700952 /* page still mapped by someone else? */
953 if (!atomic_add_negative(-1, &page->_mapcount))
954 return;
955
956 /*
957 * Now that the last pte has gone, s390 must transfer dirty
958 * flag from storage key to struct page. We can usually skip
959 * this if the page is anon, so about to be freed; but perhaps
960 * not if it's in swapcache - there might be another pte slot
961 * containing the swap entry, but page not yet written to swap.
962 */
963 if ((!PageAnon(page) || PageSwapCache(page)) && page_test_dirty(page)) {
Martin Schwidefskye2b8d7a2010-10-25 16:10:14 +0200964 page_clear_dirty(page, 1);
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -0700965 set_page_dirty(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700966 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900967 /*
968 * Hugepages are not counted in NR_ANON_PAGES nor NR_FILE_MAPPED
969 * and not charged by memcg for now.
970 */
971 if (unlikely(PageHuge(page)))
972 return;
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -0700973 if (PageAnon(page)) {
974 mem_cgroup_uncharge_page(page);
Andrea Arcangeli79134172011-01-13 15:46:58 -0800975 if (!PageTransHuge(page))
976 __dec_zone_page_state(page, NR_ANON_PAGES);
977 else
978 __dec_zone_page_state(page,
979 NR_ANON_TRANSPARENT_HUGEPAGES);
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -0700980 } else {
981 __dec_zone_page_state(page, NR_FILE_MAPPED);
Greg Thelen2a7106f2011-01-13 15:47:37 -0800982 mem_cgroup_dec_page_stat(page, MEMCG_NR_FILE_MAPPED);
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -0700983 }
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -0700984 /*
985 * It would be tidy to reset the PageAnon mapping here,
986 * but that might overwrite a racing page_add_anon_rmap
987 * which increments mapcount after us but sets mapping
988 * before us: so leave the reset to free_hot_cold_page,
989 * and remember that it's only reliable while mapped.
990 * Leaving it set also helps swapoff to reinstate ptes
991 * faster for those pages still in swapcache.
992 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700993}
994
995/*
996 * Subfunctions of try_to_unmap: try_to_unmap_one called
997 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
998 */
Hugh Dickins5ad64682009-12-14 17:59:24 -0800999int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
1000 unsigned long address, enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001001{
1002 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001003 pte_t *pte;
1004 pte_t pteval;
Hugh Dickinsc0718802005-10-29 18:16:31 -07001005 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001006 int ret = SWAP_AGAIN;
1007
Nick Piggin479db0b2008-08-20 14:09:18 -07001008 pte = page_check_address(page, mm, address, &ptl, 0);
Hugh Dickinsc0718802005-10-29 18:16:31 -07001009 if (!pte)
Nikita Danilov81b40822005-05-01 08:58:36 -07001010 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001011
1012 /*
1013 * If the page is mlock()d, we cannot swap it out.
1014 * If it's recently referenced (perhaps page_referenced
1015 * skipped over this mm) then we should reactivate it.
1016 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001017 if (!(flags & TTU_IGNORE_MLOCK)) {
KOSAKI Motohirocaed0f482009-12-14 17:59:45 -08001018 if (vma->vm_flags & VM_LOCKED)
1019 goto out_mlock;
1020
Hugh Dickinsaf8e3352009-12-14 17:58:59 -08001021 if (TTU_ACTION(flags) == TTU_MUNLOCK)
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001022 goto out_unmap;
Andi Kleen14fa31b2009-09-16 11:50:10 +02001023 }
1024 if (!(flags & TTU_IGNORE_ACCESS)) {
Nick Pigginb291f002008-10-18 20:26:44 -07001025 if (ptep_clear_flush_young_notify(vma, address, pte)) {
1026 ret = SWAP_FAIL;
1027 goto out_unmap;
1028 }
1029 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001030
Linus Torvalds1da177e2005-04-16 15:20:36 -07001031 /* Nuke the page table entry. */
1032 flush_cache_page(vma, address, page_to_pfn(page));
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001033 pteval = ptep_clear_flush_notify(vma, address, pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001034
1035 /* Move the dirty bit to the physical page now the pte is gone. */
1036 if (pte_dirty(pteval))
1037 set_page_dirty(page);
1038
Hugh Dickins365e9c872005-10-29 18:16:18 -07001039 /* Update high watermark before we lower rss */
1040 update_hiwater_rss(mm);
1041
Andi Kleen888b9f72009-09-16 11:50:11 +02001042 if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
1043 if (PageAnon(page))
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001044 dec_mm_counter(mm, MM_ANONPAGES);
Andi Kleen888b9f72009-09-16 11:50:11 +02001045 else
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001046 dec_mm_counter(mm, MM_FILEPAGES);
Andi Kleen888b9f72009-09-16 11:50:11 +02001047 set_pte_at(mm, address, pte,
1048 swp_entry_to_pte(make_hwpoison_entry(page)));
1049 } else if (PageAnon(page)) {
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07001050 swp_entry_t entry = { .val = page_private(page) };
Christoph Lameter06972122006-06-23 02:03:35 -07001051
1052 if (PageSwapCache(page)) {
1053 /*
1054 * Store the swap location in the pte.
1055 * See handle_pte_fault() ...
1056 */
Hugh Dickins570a335b2009-12-14 17:58:46 -08001057 if (swap_duplicate(entry) < 0) {
1058 set_pte_at(mm, address, pte, pteval);
1059 ret = SWAP_FAIL;
1060 goto out_unmap;
1061 }
Christoph Lameter06972122006-06-23 02:03:35 -07001062 if (list_empty(&mm->mmlist)) {
1063 spin_lock(&mmlist_lock);
1064 if (list_empty(&mm->mmlist))
1065 list_add(&mm->mmlist, &init_mm.mmlist);
1066 spin_unlock(&mmlist_lock);
1067 }
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001068 dec_mm_counter(mm, MM_ANONPAGES);
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08001069 inc_mm_counter(mm, MM_SWAPENTS);
KOSAKI Motohiro64cdd542009-01-06 14:39:16 -08001070 } else if (PAGE_MIGRATION) {
Christoph Lameter06972122006-06-23 02:03:35 -07001071 /*
1072 * Store the pfn of the page in a special migration
1073 * pte. do_swap_page() will wait until the migration
1074 * pte is removed and then restart fault handling.
1075 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001076 BUG_ON(TTU_ACTION(flags) != TTU_MIGRATION);
Christoph Lameter06972122006-06-23 02:03:35 -07001077 entry = make_migration_entry(page, pte_write(pteval));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001078 }
1079 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
1080 BUG_ON(pte_file(*pte));
Andi Kleen14fa31b2009-09-16 11:50:10 +02001081 } else if (PAGE_MIGRATION && (TTU_ACTION(flags) == TTU_MIGRATION)) {
Christoph Lameter04e62a22006-06-23 02:03:38 -07001082 /* Establish migration entry for a file page */
1083 swp_entry_t entry;
1084 entry = make_migration_entry(page, pte_write(pteval));
1085 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
1086 } else
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001087 dec_mm_counter(mm, MM_FILEPAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001088
Hugh Dickinsedc315f2009-01-06 14:40:11 -08001089 page_remove_rmap(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001090 page_cache_release(page);
1091
1092out_unmap:
Hugh Dickinsc0718802005-10-29 18:16:31 -07001093 pte_unmap_unlock(pte, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001094out:
1095 return ret;
KOSAKI Motohirocaed0f482009-12-14 17:59:45 -08001096
1097out_mlock:
1098 pte_unmap_unlock(pte, ptl);
1099
1100
1101 /*
1102 * We need mmap_sem locking, Otherwise VM_LOCKED check makes
1103 * unstable result and race. Plus, We can't wait here because
1104 * we now hold anon_vma->lock or mapping->i_mmap_lock.
1105 * if trylock failed, the page remain in evictable lru and later
1106 * vmscan could retry to move the page to unevictable lru if the
1107 * page is actually mlocked.
1108 */
1109 if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
1110 if (vma->vm_flags & VM_LOCKED) {
1111 mlock_vma_page(page);
1112 ret = SWAP_MLOCK;
1113 }
1114 up_read(&vma->vm_mm->mmap_sem);
1115 }
1116 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001117}
1118
1119/*
1120 * objrmap doesn't work for nonlinear VMAs because the assumption that
1121 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
1122 * Consequently, given a particular page and its ->index, we cannot locate the
1123 * ptes which are mapping that page without an exhaustive linear search.
1124 *
1125 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
1126 * maps the file to which the target page belongs. The ->vm_private_data field
1127 * holds the current cursor into that scan. Successive searches will circulate
1128 * around the vma's virtual address space.
1129 *
1130 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
1131 * more scanning pressure is placed against them as well. Eventually pages
1132 * will become fully unmapped and are eligible for eviction.
1133 *
1134 * For very sparsely populated VMAs this is a little inefficient - chances are
1135 * there there won't be many ptes located within the scan cluster. In this case
1136 * maybe we could scan further - to the end of the pte page, perhaps.
Nick Pigginb291f002008-10-18 20:26:44 -07001137 *
1138 * Mlocked pages: check VM_LOCKED under mmap_sem held for read, if we can
1139 * acquire it without blocking. If vma locked, mlock the pages in the cluster,
1140 * rather than unmapping them. If we encounter the "check_page" that vmscan is
1141 * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001142 */
1143#define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
1144#define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
1145
Nick Pigginb291f002008-10-18 20:26:44 -07001146static int try_to_unmap_cluster(unsigned long cursor, unsigned int *mapcount,
1147 struct vm_area_struct *vma, struct page *check_page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001148{
1149 struct mm_struct *mm = vma->vm_mm;
1150 pgd_t *pgd;
1151 pud_t *pud;
1152 pmd_t *pmd;
Hugh Dickinsc0718802005-10-29 18:16:31 -07001153 pte_t *pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001154 pte_t pteval;
Hugh Dickinsc0718802005-10-29 18:16:31 -07001155 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001156 struct page *page;
1157 unsigned long address;
1158 unsigned long end;
Nick Pigginb291f002008-10-18 20:26:44 -07001159 int ret = SWAP_AGAIN;
1160 int locked_vma = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001161
Linus Torvalds1da177e2005-04-16 15:20:36 -07001162 address = (vma->vm_start + cursor) & CLUSTER_MASK;
1163 end = address + CLUSTER_SIZE;
1164 if (address < vma->vm_start)
1165 address = vma->vm_start;
1166 if (end > vma->vm_end)
1167 end = vma->vm_end;
1168
1169 pgd = pgd_offset(mm, address);
1170 if (!pgd_present(*pgd))
Nick Pigginb291f002008-10-18 20:26:44 -07001171 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001172
1173 pud = pud_offset(pgd, address);
1174 if (!pud_present(*pud))
Nick Pigginb291f002008-10-18 20:26:44 -07001175 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001176
1177 pmd = pmd_offset(pud, address);
1178 if (!pmd_present(*pmd))
Nick Pigginb291f002008-10-18 20:26:44 -07001179 return ret;
1180
1181 /*
Hugh Dickinsaf8e3352009-12-14 17:58:59 -08001182 * If we can acquire the mmap_sem for read, and vma is VM_LOCKED,
Nick Pigginb291f002008-10-18 20:26:44 -07001183 * keep the sem while scanning the cluster for mlocking pages.
1184 */
Hugh Dickinsaf8e3352009-12-14 17:58:59 -08001185 if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
Nick Pigginb291f002008-10-18 20:26:44 -07001186 locked_vma = (vma->vm_flags & VM_LOCKED);
1187 if (!locked_vma)
1188 up_read(&vma->vm_mm->mmap_sem); /* don't need it */
1189 }
Hugh Dickinsc0718802005-10-29 18:16:31 -07001190
1191 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001192
Hugh Dickins365e9c872005-10-29 18:16:18 -07001193 /* Update high watermark before we lower rss */
1194 update_hiwater_rss(mm);
1195
Hugh Dickinsc0718802005-10-29 18:16:31 -07001196 for (; address < end; pte++, address += PAGE_SIZE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001197 if (!pte_present(*pte))
1198 continue;
Linus Torvalds6aab3412005-11-28 14:34:23 -08001199 page = vm_normal_page(vma, address, *pte);
1200 BUG_ON(!page || PageAnon(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001201
Nick Pigginb291f002008-10-18 20:26:44 -07001202 if (locked_vma) {
1203 mlock_vma_page(page); /* no-op if already mlocked */
1204 if (page == check_page)
1205 ret = SWAP_MLOCK;
1206 continue; /* don't unmap */
1207 }
1208
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001209 if (ptep_clear_flush_young_notify(vma, address, pte))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001210 continue;
1211
1212 /* Nuke the page table entry. */
Ben Collinseca35132005-11-29 11:45:26 -08001213 flush_cache_page(vma, address, pte_pfn(*pte));
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001214 pteval = ptep_clear_flush_notify(vma, address, pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001215
1216 /* If nonlinear, store the file page offset in the pte. */
1217 if (page->index != linear_page_index(vma, address))
1218 set_pte_at(mm, address, pte, pgoff_to_pte(page->index));
1219
1220 /* Move the dirty bit to the physical page now the pte is gone. */
1221 if (pte_dirty(pteval))
1222 set_page_dirty(page);
1223
Hugh Dickinsedc315f2009-01-06 14:40:11 -08001224 page_remove_rmap(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001225 page_cache_release(page);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001226 dec_mm_counter(mm, MM_FILEPAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001227 (*mapcount)--;
1228 }
Hugh Dickinsc0718802005-10-29 18:16:31 -07001229 pte_unmap_unlock(pte - 1, ptl);
Nick Pigginb291f002008-10-18 20:26:44 -07001230 if (locked_vma)
1231 up_read(&vma->vm_mm->mmap_sem);
1232 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001233}
1234
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001235bool is_vma_temporary_stack(struct vm_area_struct *vma)
Mel Gormana8bef8ff2010-05-24 14:32:24 -07001236{
1237 int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP);
1238
1239 if (!maybe_stack)
1240 return false;
1241
1242 if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) ==
1243 VM_STACK_INCOMPLETE_SETUP)
1244 return true;
1245
1246 return false;
1247}
1248
Nick Pigginb291f002008-10-18 20:26:44 -07001249/**
1250 * try_to_unmap_anon - unmap or unlock anonymous page using the object-based
1251 * rmap method
1252 * @page: the page to unmap/unlock
Huang Shijie8051be52009-12-14 17:58:50 -08001253 * @flags: action and flags
Nick Pigginb291f002008-10-18 20:26:44 -07001254 *
1255 * Find all the mappings of a page using the mapping pointer and the vma chains
1256 * contained in the anon_vma struct it points to.
1257 *
1258 * This function is only called from try_to_unmap/try_to_munlock for
1259 * anonymous pages.
1260 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1261 * where the page was found will be held for write. So, we won't recheck
1262 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1263 * 'LOCKED.
1264 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001265static int try_to_unmap_anon(struct page *page, enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001266{
1267 struct anon_vma *anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -08001268 struct anon_vma_chain *avc;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001269 int ret = SWAP_AGAIN;
Nick Pigginb291f002008-10-18 20:26:44 -07001270
Linus Torvalds1da177e2005-04-16 15:20:36 -07001271 anon_vma = page_lock_anon_vma(page);
1272 if (!anon_vma)
1273 return ret;
1274
Rik van Riel5beb4932010-03-05 13:42:07 -08001275 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1276 struct vm_area_struct *vma = avc->vma;
Mel Gormana8bef8ff2010-05-24 14:32:24 -07001277 unsigned long address;
1278
1279 /*
1280 * During exec, a temporary VMA is setup and later moved.
1281 * The VMA is moved under the anon_vma lock but not the
1282 * page tables leading to a race where migration cannot
1283 * find the migration ptes. Rather than increasing the
1284 * locking requirements of exec(), migration skips
1285 * temporary VMAs until after exec() completes.
1286 */
1287 if (PAGE_MIGRATION && (flags & TTU_MIGRATION) &&
1288 is_vma_temporary_stack(vma))
1289 continue;
1290
1291 address = vma_address(page, vma);
Hugh Dickins1cb17292009-12-14 17:59:01 -08001292 if (address == -EFAULT)
1293 continue;
1294 ret = try_to_unmap_one(page, vma, address, flags);
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001295 if (ret != SWAP_AGAIN || !page_mapped(page))
1296 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001297 }
Oleg Nesterov34bbd702007-02-28 20:13:49 -08001298
1299 page_unlock_anon_vma(anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001300 return ret;
1301}
1302
1303/**
Nick Pigginb291f002008-10-18 20:26:44 -07001304 * try_to_unmap_file - unmap/unlock file page using the object-based rmap method
1305 * @page: the page to unmap/unlock
Andi Kleen14fa31b2009-09-16 11:50:10 +02001306 * @flags: action and flags
Linus Torvalds1da177e2005-04-16 15:20:36 -07001307 *
1308 * Find all the mappings of a page using the mapping pointer and the vma chains
1309 * contained in the address_space struct it points to.
1310 *
Nick Pigginb291f002008-10-18 20:26:44 -07001311 * This function is only called from try_to_unmap/try_to_munlock for
1312 * object-based pages.
1313 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1314 * where the page was found will be held for write. So, we won't recheck
1315 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1316 * 'LOCKED.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001317 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001318static int try_to_unmap_file(struct page *page, enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001319{
1320 struct address_space *mapping = page->mapping;
1321 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1322 struct vm_area_struct *vma;
1323 struct prio_tree_iter iter;
1324 int ret = SWAP_AGAIN;
1325 unsigned long cursor;
1326 unsigned long max_nl_cursor = 0;
1327 unsigned long max_nl_size = 0;
1328 unsigned int mapcount;
1329
1330 spin_lock(&mapping->i_mmap_lock);
1331 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
Hugh Dickins1cb17292009-12-14 17:59:01 -08001332 unsigned long address = vma_address(page, vma);
1333 if (address == -EFAULT)
1334 continue;
1335 ret = try_to_unmap_one(page, vma, address, flags);
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001336 if (ret != SWAP_AGAIN || !page_mapped(page))
1337 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001338 }
1339
1340 if (list_empty(&mapping->i_mmap_nonlinear))
1341 goto out;
1342
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001343 /*
1344 * We don't bother to try to find the munlocked page in nonlinears.
1345 * It's costly. Instead, later, page reclaim logic may call
1346 * try_to_unmap(TTU_MUNLOCK) and recover PG_mlocked lazily.
1347 */
1348 if (TTU_ACTION(flags) == TTU_MUNLOCK)
1349 goto out;
1350
Linus Torvalds1da177e2005-04-16 15:20:36 -07001351 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1352 shared.vm_set.list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001353 cursor = (unsigned long) vma->vm_private_data;
1354 if (cursor > max_nl_cursor)
1355 max_nl_cursor = cursor;
1356 cursor = vma->vm_end - vma->vm_start;
1357 if (cursor > max_nl_size)
1358 max_nl_size = cursor;
1359 }
1360
Nick Pigginb291f002008-10-18 20:26:44 -07001361 if (max_nl_size == 0) { /* all nonlinears locked or reserved ? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001362 ret = SWAP_FAIL;
1363 goto out;
1364 }
1365
1366 /*
1367 * We don't try to search for this page in the nonlinear vmas,
1368 * and page_referenced wouldn't have found it anyway. Instead
1369 * just walk the nonlinear vmas trying to age and unmap some.
1370 * The mapcount of the page we came in with is irrelevant,
1371 * but even so use it as a guide to how hard we should try?
1372 */
1373 mapcount = page_mapcount(page);
1374 if (!mapcount)
1375 goto out;
1376 cond_resched_lock(&mapping->i_mmap_lock);
1377
1378 max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
1379 if (max_nl_cursor == 0)
1380 max_nl_cursor = CLUSTER_SIZE;
1381
1382 do {
1383 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1384 shared.vm_set.list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001385 cursor = (unsigned long) vma->vm_private_data;
Hugh Dickins839b9682005-09-03 15:54:43 -07001386 while ( cursor < max_nl_cursor &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07001387 cursor < vma->vm_end - vma->vm_start) {
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001388 if (try_to_unmap_cluster(cursor, &mapcount,
1389 vma, page) == SWAP_MLOCK)
1390 ret = SWAP_MLOCK;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001391 cursor += CLUSTER_SIZE;
1392 vma->vm_private_data = (void *) cursor;
1393 if ((int)mapcount <= 0)
1394 goto out;
1395 }
1396 vma->vm_private_data = (void *) max_nl_cursor;
1397 }
1398 cond_resched_lock(&mapping->i_mmap_lock);
1399 max_nl_cursor += CLUSTER_SIZE;
1400 } while (max_nl_cursor <= max_nl_size);
1401
1402 /*
1403 * Don't loop forever (perhaps all the remaining pages are
1404 * in locked vmas). Reset cursor on all unreserved nonlinear
1405 * vmas, now forgetting on which ones it had fallen behind.
1406 */
Hugh Dickins101d2be2005-11-21 21:32:16 -08001407 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1408 vma->vm_private_data = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001409out:
1410 spin_unlock(&mapping->i_mmap_lock);
1411 return ret;
1412}
1413
1414/**
1415 * try_to_unmap - try to remove all page table mappings to a page
1416 * @page: the page to get unmapped
Andi Kleen14fa31b2009-09-16 11:50:10 +02001417 * @flags: action and flags
Linus Torvalds1da177e2005-04-16 15:20:36 -07001418 *
1419 * Tries to remove all the page table entries which are mapping this
1420 * page, used in the pageout path. Caller must hold the page lock.
1421 * Return values are:
1422 *
1423 * SWAP_SUCCESS - we succeeded in removing all mappings
1424 * SWAP_AGAIN - we missed a mapping, try again later
1425 * SWAP_FAIL - the page is unswappable
Nick Pigginb291f002008-10-18 20:26:44 -07001426 * SWAP_MLOCK - page is mlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001427 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001428int try_to_unmap(struct page *page, enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001429{
1430 int ret;
1431
Linus Torvalds1da177e2005-04-16 15:20:36 -07001432 BUG_ON(!PageLocked(page));
Andrea Arcangeli91600e92011-01-13 15:47:16 -08001433 VM_BUG_ON(!PageHuge(page) && PageTransHuge(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001434
Hugh Dickins5ad64682009-12-14 17:59:24 -08001435 if (unlikely(PageKsm(page)))
1436 ret = try_to_unmap_ksm(page, flags);
1437 else if (PageAnon(page))
Andi Kleen14fa31b2009-09-16 11:50:10 +02001438 ret = try_to_unmap_anon(page, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001439 else
Andi Kleen14fa31b2009-09-16 11:50:10 +02001440 ret = try_to_unmap_file(page, flags);
Nick Pigginb291f002008-10-18 20:26:44 -07001441 if (ret != SWAP_MLOCK && !page_mapped(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001442 ret = SWAP_SUCCESS;
1443 return ret;
1444}
Nikita Danilov81b40822005-05-01 08:58:36 -07001445
Nick Pigginb291f002008-10-18 20:26:44 -07001446/**
1447 * try_to_munlock - try to munlock a page
1448 * @page: the page to be munlocked
1449 *
1450 * Called from munlock code. Checks all of the VMAs mapping the page
1451 * to make sure nobody else has this page mlocked. The page will be
1452 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1453 *
1454 * Return values are:
1455 *
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001456 * SWAP_AGAIN - no vma is holding page mlocked, or,
Nick Pigginb291f002008-10-18 20:26:44 -07001457 * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem
Hugh Dickins5ad64682009-12-14 17:59:24 -08001458 * SWAP_FAIL - page cannot be located at present
Nick Pigginb291f002008-10-18 20:26:44 -07001459 * SWAP_MLOCK - page is now mlocked.
1460 */
1461int try_to_munlock(struct page *page)
1462{
1463 VM_BUG_ON(!PageLocked(page) || PageLRU(page));
1464
Hugh Dickins5ad64682009-12-14 17:59:24 -08001465 if (unlikely(PageKsm(page)))
1466 return try_to_unmap_ksm(page, TTU_MUNLOCK);
1467 else if (PageAnon(page))
Andi Kleen14fa31b2009-09-16 11:50:10 +02001468 return try_to_unmap_anon(page, TTU_MUNLOCK);
Nick Pigginb291f002008-10-18 20:26:44 -07001469 else
Andi Kleen14fa31b2009-09-16 11:50:10 +02001470 return try_to_unmap_file(page, TTU_MUNLOCK);
Nick Pigginb291f002008-10-18 20:26:44 -07001471}
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001472
Rik van Riel76545062010-08-09 17:18:41 -07001473#if defined(CONFIG_KSM) || defined(CONFIG_MIGRATION)
1474/*
1475 * Drop an anon_vma refcount, freeing the anon_vma and anon_vma->root
1476 * if necessary. Be careful to do all the tests under the lock. Once
1477 * we know we are the last user, nobody else can get a reference and we
1478 * can do the freeing without the lock.
1479 */
1480void drop_anon_vma(struct anon_vma *anon_vma)
1481{
Andrea Arcangeli44ab57a2010-08-09 17:19:10 -07001482 BUG_ON(atomic_read(&anon_vma->external_refcount) <= 0);
Rik van Riel76545062010-08-09 17:18:41 -07001483 if (atomic_dec_and_lock(&anon_vma->external_refcount, &anon_vma->root->lock)) {
1484 struct anon_vma *root = anon_vma->root;
1485 int empty = list_empty(&anon_vma->head);
1486 int last_root_user = 0;
1487 int root_empty = 0;
1488
1489 /*
1490 * The refcount on a non-root anon_vma got dropped. Drop
1491 * the refcount on the root and check if we need to free it.
1492 */
1493 if (empty && anon_vma != root) {
Andrea Arcangeli44ab57a2010-08-09 17:19:10 -07001494 BUG_ON(atomic_read(&root->external_refcount) <= 0);
Rik van Riel76545062010-08-09 17:18:41 -07001495 last_root_user = atomic_dec_and_test(&root->external_refcount);
1496 root_empty = list_empty(&root->head);
1497 }
1498 anon_vma_unlock(anon_vma);
1499
1500 if (empty) {
1501 anon_vma_free(anon_vma);
1502 if (root_empty && last_root_user)
1503 anon_vma_free(root);
1504 }
1505 }
1506}
1507#endif
1508
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001509#ifdef CONFIG_MIGRATION
1510/*
1511 * rmap_walk() and its helpers rmap_walk_anon() and rmap_walk_file():
1512 * Called by migrate.c to remove migration ptes, but might be used more later.
1513 */
1514static int rmap_walk_anon(struct page *page, int (*rmap_one)(struct page *,
1515 struct vm_area_struct *, unsigned long, void *), void *arg)
1516{
1517 struct anon_vma *anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -08001518 struct anon_vma_chain *avc;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001519 int ret = SWAP_AGAIN;
1520
1521 /*
1522 * Note: remove_migration_ptes() cannot use page_lock_anon_vma()
1523 * because that depends on page_mapped(); but not all its usages
Mel Gorman3f6c8272010-05-24 14:32:17 -07001524 * are holding mmap_sem. Users without mmap_sem are required to
1525 * take a reference count to prevent the anon_vma disappearing
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001526 */
1527 anon_vma = page_anon_vma(page);
1528 if (!anon_vma)
1529 return ret;
Rik van Rielcba48b92010-08-09 17:18:38 -07001530 anon_vma_lock(anon_vma);
Rik van Riel5beb4932010-03-05 13:42:07 -08001531 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1532 struct vm_area_struct *vma = avc->vma;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001533 unsigned long address = vma_address(page, vma);
1534 if (address == -EFAULT)
1535 continue;
1536 ret = rmap_one(page, vma, address, arg);
1537 if (ret != SWAP_AGAIN)
1538 break;
1539 }
Rik van Rielcba48b92010-08-09 17:18:38 -07001540 anon_vma_unlock(anon_vma);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001541 return ret;
1542}
1543
1544static int rmap_walk_file(struct page *page, int (*rmap_one)(struct page *,
1545 struct vm_area_struct *, unsigned long, void *), void *arg)
1546{
1547 struct address_space *mapping = page->mapping;
1548 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1549 struct vm_area_struct *vma;
1550 struct prio_tree_iter iter;
1551 int ret = SWAP_AGAIN;
1552
1553 if (!mapping)
1554 return ret;
1555 spin_lock(&mapping->i_mmap_lock);
1556 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1557 unsigned long address = vma_address(page, vma);
1558 if (address == -EFAULT)
1559 continue;
1560 ret = rmap_one(page, vma, address, arg);
1561 if (ret != SWAP_AGAIN)
1562 break;
1563 }
1564 /*
1565 * No nonlinear handling: being always shared, nonlinear vmas
1566 * never contain migration ptes. Decide what to do about this
1567 * limitation to linear when we need rmap_walk() on nonlinear.
1568 */
1569 spin_unlock(&mapping->i_mmap_lock);
1570 return ret;
1571}
1572
1573int rmap_walk(struct page *page, int (*rmap_one)(struct page *,
1574 struct vm_area_struct *, unsigned long, void *), void *arg)
1575{
1576 VM_BUG_ON(!PageLocked(page));
1577
1578 if (unlikely(PageKsm(page)))
1579 return rmap_walk_ksm(page, rmap_one, arg);
1580 else if (PageAnon(page))
1581 return rmap_walk_anon(page, rmap_one, arg);
1582 else
1583 return rmap_walk_file(page, rmap_one, arg);
1584}
1585#endif /* CONFIG_MIGRATION */
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001586
Naoya Horiguchie3390f62010-06-15 13:18:13 +09001587#ifdef CONFIG_HUGETLB_PAGE
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001588/*
1589 * The following three functions are for anonymous (private mapped) hugepages.
1590 * Unlike common anonymous pages, anonymous hugepages have no accounting code
1591 * and no lru code, because we handle hugepages differently from common pages.
1592 */
1593static void __hugepage_set_anon_rmap(struct page *page,
1594 struct vm_area_struct *vma, unsigned long address, int exclusive)
1595{
1596 struct anon_vma *anon_vma = vma->anon_vma;
Naoya Horiguchi433abed2010-09-10 13:23:03 +09001597
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001598 BUG_ON(!anon_vma);
Naoya Horiguchi433abed2010-09-10 13:23:03 +09001599
1600 if (PageAnon(page))
1601 return;
1602 if (!exclusive)
1603 anon_vma = anon_vma->root;
1604
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001605 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1606 page->mapping = (struct address_space *) anon_vma;
1607 page->index = linear_page_index(vma, address);
1608}
1609
1610void hugepage_add_anon_rmap(struct page *page,
1611 struct vm_area_struct *vma, unsigned long address)
1612{
1613 struct anon_vma *anon_vma = vma->anon_vma;
1614 int first;
Naoya Horiguchia850ea32010-09-10 13:23:06 +09001615
1616 BUG_ON(!PageLocked(page));
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001617 BUG_ON(!anon_vma);
1618 BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1619 first = atomic_inc_and_test(&page->_mapcount);
1620 if (first)
1621 __hugepage_set_anon_rmap(page, vma, address, 0);
1622}
1623
1624void hugepage_add_new_anon_rmap(struct page *page,
1625 struct vm_area_struct *vma, unsigned long address)
1626{
1627 BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1628 atomic_set(&page->_mapcount, 0);
1629 __hugepage_set_anon_rmap(page, vma, address, 1);
1630}
Naoya Horiguchie3390f62010-06-15 13:18:13 +09001631#endif /* CONFIG_HUGETLB_PAGE */