blob: 8da044a1db0f4db1524f2450cd733e41eb8690ca [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)
Dave Chinner250df6e2011-03-22 22:23:36 +110034 * inode->i_lock (in set_page_dirty's __mark_inode_dirty)
Dave Chinnera66979a2011-03-22 22:23:41 +110035 * inode_wb_list_lock (in set_page_dirty's __mark_inode_dirty)
Nick Piggin82591e62006-10-19 23:29:10 -070036 * sb_lock (within inode_lock in fs/fs-writeback.c)
37 * mapping->tree_lock (widely used, in set_page_dirty,
38 * in arch-dependent flush_dcache_mmap_lock,
Dave Chinnera66979a2011-03-22 22:23:41 +110039 * within inode_wb_list_lock in __sync_single_inode)
Andi Kleen6a460792009-09-16 11:50:15 +020040 *
41 * (code doesn't rely on that order so it could be switched around)
42 * ->tasklist_lock
43 * anon_vma->lock (memory_failure, collect_procs_anon)
44 * pte map lock
Linus Torvalds1da177e2005-04-16 15:20:36 -070045 */
46
47#include <linux/mm.h>
48#include <linux/pagemap.h>
49#include <linux/swap.h>
50#include <linux/swapops.h>
51#include <linux/slab.h>
52#include <linux/init.h>
Hugh Dickins5ad64682009-12-14 17:59:24 -080053#include <linux/ksm.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070054#include <linux/rmap.h>
55#include <linux/rcupdate.h>
Christoph Lametera48d07a2006-02-01 03:05:38 -080056#include <linux/module.h>
Balbir Singh8a9f3cc2008-02-07 00:13:53 -080057#include <linux/memcontrol.h>
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070058#include <linux/mmu_notifier.h>
KOSAKI Motohiro64cdd542009-01-06 14:39:16 -080059#include <linux/migrate.h>
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +090060#include <linux/hugetlb.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070061
62#include <asm/tlbflush.h>
63
Nick Pigginb291f002008-10-18 20:26:44 -070064#include "internal.h"
65
Adrian Bunkfdd2e5f2008-10-18 20:28:38 -070066static struct kmem_cache *anon_vma_cachep;
Rik van Riel5beb4932010-03-05 13:42:07 -080067static struct kmem_cache *anon_vma_chain_cachep;
Adrian Bunkfdd2e5f2008-10-18 20:28:38 -070068
69static inline struct anon_vma *anon_vma_alloc(void)
70{
Peter Zijlstra01d8b202011-03-22 16:32:49 -070071 struct anon_vma *anon_vma;
72
73 anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
74 if (anon_vma) {
75 atomic_set(&anon_vma->refcount, 1);
76 /*
77 * Initialise the anon_vma root to point to itself. If called
78 * from fork, the root will be reset to the parents anon_vma.
79 */
80 anon_vma->root = anon_vma;
81 }
82
83 return anon_vma;
Adrian Bunkfdd2e5f2008-10-18 20:28:38 -070084}
85
Peter Zijlstra01d8b202011-03-22 16:32:49 -070086static inline void anon_vma_free(struct anon_vma *anon_vma)
Adrian Bunkfdd2e5f2008-10-18 20:28:38 -070087{
Peter Zijlstra01d8b202011-03-22 16:32:49 -070088 VM_BUG_ON(atomic_read(&anon_vma->refcount));
Adrian Bunkfdd2e5f2008-10-18 20:28:38 -070089 kmem_cache_free(anon_vma_cachep, anon_vma);
90}
Linus Torvalds1da177e2005-04-16 15:20:36 -070091
Rik van Riel5beb4932010-03-05 13:42:07 -080092static inline struct anon_vma_chain *anon_vma_chain_alloc(void)
93{
94 return kmem_cache_alloc(anon_vma_chain_cachep, GFP_KERNEL);
95}
96
Namhyung Kime574b5f2010-10-26 14:22:02 -070097static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
Rik van Riel5beb4932010-03-05 13:42:07 -080098{
99 kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
100}
101
Linus Torvaldsd9d332e2008-10-19 10:32:20 -0700102/**
103 * anon_vma_prepare - attach an anon_vma to a memory region
104 * @vma: the memory region in question
105 *
106 * This makes sure the memory mapping described by 'vma' has
107 * an 'anon_vma' attached to it, so that we can associate the
108 * anonymous pages mapped into it with that anon_vma.
109 *
110 * The common case will be that we already have one, but if
Figo.zhang23a07902010-12-27 15:14:06 +0100111 * not we either need to find an adjacent mapping that we
Linus Torvaldsd9d332e2008-10-19 10:32:20 -0700112 * can re-use the anon_vma from (very common when the only
113 * reason for splitting a vma has been mprotect()), or we
114 * allocate a new one.
115 *
116 * Anon-vma allocations are very subtle, because we may have
117 * optimistically looked up an anon_vma in page_lock_anon_vma()
118 * and that may actually touch the spinlock even in the newly
119 * allocated vma (it depends on RCU to make sure that the
120 * anon_vma isn't actually destroyed).
121 *
122 * As a result, we need to do proper anon_vma locking even
123 * for the new allocation. At the same time, we do not want
124 * to do any locking for the common case of already having
125 * an anon_vma.
126 *
127 * This must be called with the mmap_sem held for reading.
128 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700129int anon_vma_prepare(struct vm_area_struct *vma)
130{
131 struct anon_vma *anon_vma = vma->anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -0800132 struct anon_vma_chain *avc;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700133
134 might_sleep();
135 if (unlikely(!anon_vma)) {
136 struct mm_struct *mm = vma->vm_mm;
Linus Torvaldsd9d332e2008-10-19 10:32:20 -0700137 struct anon_vma *allocated;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700138
Rik van Riel5beb4932010-03-05 13:42:07 -0800139 avc = anon_vma_chain_alloc();
140 if (!avc)
141 goto out_enomem;
142
Linus Torvalds1da177e2005-04-16 15:20:36 -0700143 anon_vma = find_mergeable_anon_vma(vma);
Linus Torvaldsd9d332e2008-10-19 10:32:20 -0700144 allocated = NULL;
145 if (!anon_vma) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700146 anon_vma = anon_vma_alloc();
147 if (unlikely(!anon_vma))
Rik van Riel5beb4932010-03-05 13:42:07 -0800148 goto out_enomem_free_avc;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700149 allocated = anon_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700150 }
151
Rik van Rielcba48b92010-08-09 17:18:38 -0700152 anon_vma_lock(anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700153 /* page_table_lock to protect against threads */
154 spin_lock(&mm->page_table_lock);
155 if (likely(!vma->anon_vma)) {
156 vma->anon_vma = anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -0800157 avc->anon_vma = anon_vma;
158 avc->vma = vma;
159 list_add(&avc->same_vma, &vma->anon_vma_chain);
Andrea Arcangeli26ba0cb2010-08-09 17:19:08 -0700160 list_add_tail(&avc->same_anon_vma, &anon_vma->head);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700161 allocated = NULL;
Oleg Nesterov31f2b0e2010-04-23 13:18:01 -0400162 avc = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700163 }
164 spin_unlock(&mm->page_table_lock);
Rik van Rielcba48b92010-08-09 17:18:38 -0700165 anon_vma_unlock(anon_vma);
Oleg Nesterov31f2b0e2010-04-23 13:18:01 -0400166
167 if (unlikely(allocated))
Peter Zijlstra01d8b202011-03-22 16:32:49 -0700168 put_anon_vma(allocated);
Oleg Nesterov31f2b0e2010-04-23 13:18:01 -0400169 if (unlikely(avc))
Rik van Riel5beb4932010-03-05 13:42:07 -0800170 anon_vma_chain_free(avc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700171 }
172 return 0;
Rik van Riel5beb4932010-03-05 13:42:07 -0800173
174 out_enomem_free_avc:
175 anon_vma_chain_free(avc);
176 out_enomem:
177 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700178}
179
Rik van Riel5beb4932010-03-05 13:42:07 -0800180static void anon_vma_chain_link(struct vm_area_struct *vma,
181 struct anon_vma_chain *avc,
182 struct anon_vma *anon_vma)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700183{
Rik van Riel5beb4932010-03-05 13:42:07 -0800184 avc->vma = vma;
185 avc->anon_vma = anon_vma;
186 list_add(&avc->same_vma, &vma->anon_vma_chain);
187
Rik van Rielcba48b92010-08-09 17:18:38 -0700188 anon_vma_lock(anon_vma);
Andrea Arcangeli05759d32011-01-13 15:46:53 -0800189 /*
190 * It's critical to add new vmas to the tail of the anon_vma,
191 * see comment in huge_memory.c:__split_huge_page().
192 */
Rik van Riel5beb4932010-03-05 13:42:07 -0800193 list_add_tail(&avc->same_anon_vma, &anon_vma->head);
Rik van Rielcba48b92010-08-09 17:18:38 -0700194 anon_vma_unlock(anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700195}
196
Rik van Riel5beb4932010-03-05 13:42:07 -0800197/*
198 * Attach the anon_vmas from src to dst.
199 * Returns 0 on success, -ENOMEM on failure.
200 */
201int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700202{
Rik van Riel5beb4932010-03-05 13:42:07 -0800203 struct anon_vma_chain *avc, *pavc;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700204
Linus Torvalds646d87b2010-04-11 17:15:03 -0700205 list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
Rik van Riel5beb4932010-03-05 13:42:07 -0800206 avc = anon_vma_chain_alloc();
207 if (!avc)
208 goto enomem_failure;
209 anon_vma_chain_link(dst, avc, pavc->anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700210 }
Rik van Riel5beb4932010-03-05 13:42:07 -0800211 return 0;
212
213 enomem_failure:
214 unlink_anon_vmas(dst);
215 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700216}
217
Rik van Riel5beb4932010-03-05 13:42:07 -0800218/*
219 * Attach vma to its own anon_vma, as well as to the anon_vmas that
220 * the corresponding VMA in the parent process is attached to.
221 * Returns 0 on success, non-zero on failure.
222 */
223int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700224{
Rik van Riel5beb4932010-03-05 13:42:07 -0800225 struct anon_vma_chain *avc;
226 struct anon_vma *anon_vma;
227
228 /* Don't bother if the parent process has no anon_vma here. */
229 if (!pvma->anon_vma)
230 return 0;
231
232 /*
233 * First, attach the new VMA to the parent VMA's anon_vmas,
234 * so rmap can find non-COWed pages in child processes.
235 */
236 if (anon_vma_clone(vma, pvma))
237 return -ENOMEM;
238
239 /* Then add our own anon_vma. */
240 anon_vma = anon_vma_alloc();
241 if (!anon_vma)
242 goto out_error;
243 avc = anon_vma_chain_alloc();
244 if (!avc)
245 goto out_error_free_anon_vma;
Rik van Riel5c341ee12010-08-09 17:18:39 -0700246
247 /*
248 * The root anon_vma's spinlock is the lock actually used when we
249 * lock any of the anon_vmas in this anon_vma tree.
250 */
251 anon_vma->root = pvma->anon_vma->root;
Rik van Riel76545062010-08-09 17:18:41 -0700252 /*
Peter Zijlstra01d8b202011-03-22 16:32:49 -0700253 * With refcounts, an anon_vma can stay around longer than the
254 * process it belongs to. The root anon_vma needs to be pinned until
255 * this anon_vma is freed, because the lock lives in the root.
Rik van Riel76545062010-08-09 17:18:41 -0700256 */
257 get_anon_vma(anon_vma->root);
Rik van Riel5beb4932010-03-05 13:42:07 -0800258 /* Mark this anon_vma as the one where our new (COWed) pages go. */
259 vma->anon_vma = anon_vma;
Rik van Riel5c341ee12010-08-09 17:18:39 -0700260 anon_vma_chain_link(vma, avc, anon_vma);
Rik van Riel5beb4932010-03-05 13:42:07 -0800261
262 return 0;
263
264 out_error_free_anon_vma:
Peter Zijlstra01d8b202011-03-22 16:32:49 -0700265 put_anon_vma(anon_vma);
Rik van Riel5beb4932010-03-05 13:42:07 -0800266 out_error:
Rik van Riel4946d542010-04-05 12:13:33 -0400267 unlink_anon_vmas(vma);
Rik van Riel5beb4932010-03-05 13:42:07 -0800268 return -ENOMEM;
269}
270
271static void anon_vma_unlink(struct anon_vma_chain *anon_vma_chain)
272{
273 struct anon_vma *anon_vma = anon_vma_chain->anon_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700274 int empty;
275
Rik van Riel5beb4932010-03-05 13:42:07 -0800276 /* If anon_vma_fork fails, we can get an empty anon_vma_chain. */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700277 if (!anon_vma)
278 return;
279
Rik van Rielcba48b92010-08-09 17:18:38 -0700280 anon_vma_lock(anon_vma);
Rik van Riel5beb4932010-03-05 13:42:07 -0800281 list_del(&anon_vma_chain->same_anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700282
283 /* We must garbage collect the anon_vma if it's empty */
Peter Zijlstra01d8b202011-03-22 16:32:49 -0700284 empty = list_empty(&anon_vma->head);
Rik van Rielcba48b92010-08-09 17:18:38 -0700285 anon_vma_unlock(anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700286
Peter Zijlstra01d8b202011-03-22 16:32:49 -0700287 if (empty)
288 put_anon_vma(anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700289}
290
Rik van Riel5beb4932010-03-05 13:42:07 -0800291void unlink_anon_vmas(struct vm_area_struct *vma)
292{
293 struct anon_vma_chain *avc, *next;
294
Rik van Riel5c341ee12010-08-09 17:18:39 -0700295 /*
296 * Unlink each anon_vma chained to the VMA. This list is ordered
297 * from newest to oldest, ensuring the root anon_vma gets freed last.
298 */
Rik van Riel5beb4932010-03-05 13:42:07 -0800299 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
300 anon_vma_unlink(avc);
301 list_del(&avc->same_vma);
302 anon_vma_chain_free(avc);
303 }
304}
305
Alexey Dobriyan51cc5062008-07-25 19:45:34 -0700306static void anon_vma_ctor(void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700307{
Christoph Lametera35afb82007-05-16 22:10:57 -0700308 struct anon_vma *anon_vma = data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700309
Christoph Lametera35afb82007-05-16 22:10:57 -0700310 spin_lock_init(&anon_vma->lock);
Peter Zijlstra83813262011-03-22 16:32:48 -0700311 atomic_set(&anon_vma->refcount, 0);
Christoph Lametera35afb82007-05-16 22:10:57 -0700312 INIT_LIST_HEAD(&anon_vma->head);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700313}
314
315void __init anon_vma_init(void)
316{
317 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
Paul Mundt20c2df82007-07-20 10:11:58 +0900318 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor);
Rik van Riel5beb4932010-03-05 13:42:07 -0800319 anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain, SLAB_PANIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700320}
321
322/*
323 * Getting a lock on a stable anon_vma from a page off the LRU is
324 * tricky: page_lock_anon_vma rely on RCU to guard against the races.
325 */
Namhyung Kimea4525b2010-10-26 14:22:01 -0700326struct anon_vma *__page_lock_anon_vma(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700327{
Hugh Dickinsf18194272010-08-25 23:12:54 -0700328 struct anon_vma *anon_vma, *root_anon_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700329 unsigned long anon_mapping;
330
331 rcu_read_lock();
Hugh Dickins80e148222009-12-14 17:59:29 -0800332 anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping);
Hugh Dickins3ca7b3c2009-12-14 17:58:57 -0800333 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700334 goto out;
335 if (!page_mapped(page))
336 goto out;
337
338 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
Hugh Dickinsf18194272010-08-25 23:12:54 -0700339 root_anon_vma = ACCESS_ONCE(anon_vma->root);
340 spin_lock(&root_anon_vma->lock);
341
342 /*
343 * If this page is still mapped, then its anon_vma cannot have been
344 * freed. But if it has been unmapped, we have no security against
345 * the anon_vma structure being freed and reused (for another anon_vma:
346 * SLAB_DESTROY_BY_RCU guarantees that - so the spin_lock above cannot
347 * corrupt): with anon_vma_prepare() or anon_vma_fork() redirecting
348 * anon_vma->root before page_unlock_anon_vma() is called to unlock.
349 */
350 if (page_mapped(page))
351 return anon_vma;
352
353 spin_unlock(&root_anon_vma->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700354out:
355 rcu_read_unlock();
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800356 return NULL;
357}
358
Andi Kleen10be22d2009-09-16 11:50:04 +0200359void page_unlock_anon_vma(struct anon_vma *anon_vma)
Namhyung Kimea4525b2010-10-26 14:22:01 -0700360 __releases(&anon_vma->root->lock)
361 __releases(RCU)
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800362{
Rik van Rielcba48b92010-08-09 17:18:38 -0700363 anon_vma_unlock(anon_vma);
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800364 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700365}
366
367/*
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -0800368 * At what user virtual address is page expected in @vma?
369 * Returns virtual address or -EFAULT if page's index/offset is not
370 * within the range mapped the @vma.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700371 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800372inline unsigned long
Linus Torvalds1da177e2005-04-16 15:20:36 -0700373vma_address(struct page *page, struct vm_area_struct *vma)
374{
375 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
376 unsigned long address;
377
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900378 if (unlikely(is_vm_hugetlb_page(vma)))
379 pgoff = page->index << huge_page_order(page_hstate(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700380 address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
381 if (unlikely(address < vma->vm_start || address >= vma->vm_end)) {
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -0800382 /* page should be within @vma mapping range */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700383 return -EFAULT;
384 }
385 return address;
386}
387
388/*
Huang Shijiebf89c8c2009-10-01 15:44:04 -0700389 * At what user virtual address is page expected in vma?
Naoya Horiguchiab941e02010-05-11 14:06:55 -0700390 * Caller should check the page is actually part of the vma.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700391 */
392unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
393{
Andrea Arcangeli21d0d442010-08-09 17:19:10 -0700394 if (PageAnon(page)) {
Hugh Dickins4829b902010-10-02 17:46:06 -0700395 struct anon_vma *page__anon_vma = page_anon_vma(page);
396 /*
397 * Note: swapoff's unuse_vma() is more efficient with this
398 * check, and needs it to match anon_vma when KSM is active.
399 */
400 if (!vma->anon_vma || !page__anon_vma ||
401 vma->anon_vma->root != page__anon_vma->root)
Andrea Arcangeli21d0d442010-08-09 17:19:10 -0700402 return -EFAULT;
403 } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) {
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800404 if (!vma->vm_file ||
405 vma->vm_file->f_mapping != page->mapping)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700406 return -EFAULT;
407 } else
408 return -EFAULT;
409 return vma_address(page, vma);
410}
411
412/*
Nikita Danilov81b40822005-05-01 08:58:36 -0700413 * Check that @page is mapped at @address into @mm.
414 *
Nick Piggin479db0b2008-08-20 14:09:18 -0700415 * If @sync is false, page_check_address may perform a racy check to avoid
416 * the page table lock when the pte is not present (helpful when reclaiming
417 * highly shared pages).
418 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -0700419 * On success returns with pte mapped and locked.
Nikita Danilov81b40822005-05-01 08:58:36 -0700420 */
Namhyung Kime9a81a82010-10-26 14:22:01 -0700421pte_t *__page_check_address(struct page *page, struct mm_struct *mm,
Nick Piggin479db0b2008-08-20 14:09:18 -0700422 unsigned long address, spinlock_t **ptlp, int sync)
Nikita Danilov81b40822005-05-01 08:58:36 -0700423{
424 pgd_t *pgd;
425 pud_t *pud;
426 pmd_t *pmd;
427 pte_t *pte;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700428 spinlock_t *ptl;
Nikita Danilov81b40822005-05-01 08:58:36 -0700429
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900430 if (unlikely(PageHuge(page))) {
431 pte = huge_pte_offset(mm, address);
432 ptl = &mm->page_table_lock;
433 goto check;
434 }
435
Nikita Danilov81b40822005-05-01 08:58:36 -0700436 pgd = pgd_offset(mm, address);
Hugh Dickinsc0718802005-10-29 18:16:31 -0700437 if (!pgd_present(*pgd))
438 return NULL;
439
440 pud = pud_offset(pgd, address);
441 if (!pud_present(*pud))
442 return NULL;
443
444 pmd = pmd_offset(pud, address);
445 if (!pmd_present(*pmd))
446 return NULL;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800447 if (pmd_trans_huge(*pmd))
448 return NULL;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700449
450 pte = pte_offset_map(pmd, address);
451 /* Make a quick check before getting the lock */
Nick Piggin479db0b2008-08-20 14:09:18 -0700452 if (!sync && !pte_present(*pte)) {
Hugh Dickinsc0718802005-10-29 18:16:31 -0700453 pte_unmap(pte);
454 return NULL;
Nikita Danilov81b40822005-05-01 08:58:36 -0700455 }
Hugh Dickinsc0718802005-10-29 18:16:31 -0700456
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700457 ptl = pte_lockptr(mm, pmd);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900458check:
Hugh Dickinsc0718802005-10-29 18:16:31 -0700459 spin_lock(ptl);
460 if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
461 *ptlp = ptl;
462 return pte;
463 }
464 pte_unmap_unlock(pte, ptl);
465 return NULL;
Nikita Danilov81b40822005-05-01 08:58:36 -0700466}
467
Nick Pigginb291f002008-10-18 20:26:44 -0700468/**
469 * page_mapped_in_vma - check whether a page is really mapped in a VMA
470 * @page: the page to test
471 * @vma: the VMA to test
472 *
473 * Returns 1 if the page is mapped into the page tables of the VMA, 0
474 * if the page is not mapped into the page tables of this VMA. Only
475 * valid for normal file or anonymous VMAs.
476 */
Andi Kleen6a460792009-09-16 11:50:15 +0200477int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
Nick Pigginb291f002008-10-18 20:26:44 -0700478{
479 unsigned long address;
480 pte_t *pte;
481 spinlock_t *ptl;
482
483 address = vma_address(page, vma);
484 if (address == -EFAULT) /* out of vma range */
485 return 0;
486 pte = page_check_address(page, vma->vm_mm, address, &ptl, 1);
487 if (!pte) /* the page is not in this mm */
488 return 0;
489 pte_unmap_unlock(pte, ptl);
490
491 return 1;
492}
493
Nikita Danilov81b40822005-05-01 08:58:36 -0700494/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700495 * Subfunctions of page_referenced: page_referenced_one called
496 * repeatedly from either page_referenced_anon or page_referenced_file.
497 */
Hugh Dickins5ad64682009-12-14 17:59:24 -0800498int page_referenced_one(struct page *page, struct vm_area_struct *vma,
499 unsigned long address, unsigned int *mapcount,
500 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700501{
502 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700503 int referenced = 0;
504
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800505 if (unlikely(PageTransHuge(page))) {
506 pmd_t *pmd;
507
508 spin_lock(&mm->page_table_lock);
Andrea Arcangeli2da28bf2011-03-11 14:58:29 -0800509 /*
510 * rmap might return false positives; we must filter
511 * these out using page_check_address_pmd().
512 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800513 pmd = page_check_address_pmd(page, mm, address,
514 PAGE_CHECK_ADDRESS_PMD_FLAG);
Andrea Arcangeli2da28bf2011-03-11 14:58:29 -0800515 if (!pmd) {
516 spin_unlock(&mm->page_table_lock);
517 goto out;
518 }
519
520 if (vma->vm_flags & VM_LOCKED) {
521 spin_unlock(&mm->page_table_lock);
522 *mapcount = 0; /* break early from loop */
523 *vm_flags |= VM_LOCKED;
524 goto out;
525 }
526
527 /* go ahead even if the pmd is pmd_trans_splitting() */
528 if (pmdp_clear_flush_young_notify(vma, address, pmd))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800529 referenced++;
530 spin_unlock(&mm->page_table_lock);
531 } else {
532 pte_t *pte;
533 spinlock_t *ptl;
534
Andrea Arcangeli2da28bf2011-03-11 14:58:29 -0800535 /*
536 * rmap might return false positives; we must filter
537 * these out using page_check_address().
538 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800539 pte = page_check_address(page, mm, address, &ptl, 0);
540 if (!pte)
541 goto out;
542
Andrea Arcangeli2da28bf2011-03-11 14:58:29 -0800543 if (vma->vm_flags & VM_LOCKED) {
544 pte_unmap_unlock(pte, ptl);
545 *mapcount = 0; /* break early from loop */
546 *vm_flags |= VM_LOCKED;
547 goto out;
548 }
549
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800550 if (ptep_clear_flush_young_notify(vma, address, pte)) {
551 /*
552 * Don't treat a reference through a sequentially read
553 * mapping as such. If the page has been used in
554 * another mapping, we will catch it; if this other
555 * mapping is already gone, the unmap path will have
556 * set PG_referenced or activated the page.
557 */
558 if (likely(!VM_SequentialReadHint(vma)))
559 referenced++;
560 }
561 pte_unmap_unlock(pte, ptl);
562 }
563
Andrea Arcangeli2da28bf2011-03-11 14:58:29 -0800564 /* Pretend the page is referenced if the task has the
565 swap token and is in the middle of a page fault. */
566 if (mm != current->mm && has_swap_token(mm) &&
567 rwsem_is_locked(&mm->mmap_sem))
568 referenced++;
569
Hugh Dickinsc0718802005-10-29 18:16:31 -0700570 (*mapcount)--;
Huang Shijie273f0472009-12-14 17:58:51 -0800571
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700572 if (referenced)
573 *vm_flags |= vma->vm_flags;
Huang Shijie273f0472009-12-14 17:58:51 -0800574out:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700575 return referenced;
576}
577
Balbir Singhbed71612008-02-07 00:14:01 -0800578static int page_referenced_anon(struct page *page,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700579 struct mem_cgroup *mem_cont,
580 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700581{
582 unsigned int mapcount;
583 struct anon_vma *anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -0800584 struct anon_vma_chain *avc;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700585 int referenced = 0;
586
587 anon_vma = page_lock_anon_vma(page);
588 if (!anon_vma)
589 return referenced;
590
591 mapcount = page_mapcount(page);
Rik van Riel5beb4932010-03-05 13:42:07 -0800592 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
593 struct vm_area_struct *vma = avc->vma;
Hugh Dickins1cb17292009-12-14 17:59:01 -0800594 unsigned long address = vma_address(page, vma);
595 if (address == -EFAULT)
596 continue;
Balbir Singhbed71612008-02-07 00:14:01 -0800597 /*
598 * If we are reclaiming on behalf of a cgroup, skip
599 * counting on behalf of references from different
600 * cgroups
601 */
Hugh Dickinsbd845e32008-03-04 14:29:01 -0800602 if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
Balbir Singhbed71612008-02-07 00:14:01 -0800603 continue;
Hugh Dickins1cb17292009-12-14 17:59:01 -0800604 referenced += page_referenced_one(page, vma, address,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700605 &mapcount, vm_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700606 if (!mapcount)
607 break;
608 }
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800609
610 page_unlock_anon_vma(anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700611 return referenced;
612}
613
614/**
615 * page_referenced_file - referenced check for object-based rmap
616 * @page: the page we're checking references on.
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700617 * @mem_cont: target memory controller
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700618 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
Linus Torvalds1da177e2005-04-16 15:20:36 -0700619 *
620 * For an object-based mapped page, find all the places it is mapped and
621 * check/clear the referenced flag. This is done by following the page->mapping
622 * pointer, then walking the chain of vmas it holds. It returns the number
623 * of references it found.
624 *
625 * This function is only called from page_referenced for object-based pages.
626 */
Balbir Singhbed71612008-02-07 00:14:01 -0800627static int page_referenced_file(struct page *page,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700628 struct mem_cgroup *mem_cont,
629 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700630{
631 unsigned int mapcount;
632 struct address_space *mapping = page->mapping;
633 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
634 struct vm_area_struct *vma;
635 struct prio_tree_iter iter;
636 int referenced = 0;
637
638 /*
639 * The caller's checks on page->mapping and !PageAnon have made
640 * sure that this is a file page: the check for page->mapping
641 * excludes the case just before it gets set on an anon page.
642 */
643 BUG_ON(PageAnon(page));
644
645 /*
646 * The page lock not only makes sure that page->mapping cannot
647 * suddenly be NULLified by truncation, it makes sure that the
648 * structure at mapping cannot be freed and reused yet,
649 * so we can safely take mapping->i_mmap_lock.
650 */
651 BUG_ON(!PageLocked(page));
652
653 spin_lock(&mapping->i_mmap_lock);
654
655 /*
656 * i_mmap_lock does not stabilize mapcount at all, but mapcount
657 * is more likely to be accurate if we note it after spinning.
658 */
659 mapcount = page_mapcount(page);
660
661 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
Hugh Dickins1cb17292009-12-14 17:59:01 -0800662 unsigned long address = vma_address(page, vma);
663 if (address == -EFAULT)
664 continue;
Balbir Singhbed71612008-02-07 00:14:01 -0800665 /*
666 * If we are reclaiming on behalf of a cgroup, skip
667 * counting on behalf of references from different
668 * cgroups
669 */
Hugh Dickinsbd845e32008-03-04 14:29:01 -0800670 if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
Balbir Singhbed71612008-02-07 00:14:01 -0800671 continue;
Hugh Dickins1cb17292009-12-14 17:59:01 -0800672 referenced += page_referenced_one(page, vma, address,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700673 &mapcount, vm_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700674 if (!mapcount)
675 break;
676 }
677
678 spin_unlock(&mapping->i_mmap_lock);
679 return referenced;
680}
681
682/**
683 * page_referenced - test if the page was referenced
684 * @page: the page to test
685 * @is_locked: caller holds lock on the page
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700686 * @mem_cont: target memory controller
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700687 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
Linus Torvalds1da177e2005-04-16 15:20:36 -0700688 *
689 * Quick test_and_clear_referenced for all mappings to a page,
690 * returns the number of ptes which referenced the page.
691 */
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700692int page_referenced(struct page *page,
693 int is_locked,
694 struct mem_cgroup *mem_cont,
695 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700696{
697 int referenced = 0;
Hugh Dickins5ad64682009-12-14 17:59:24 -0800698 int we_locked = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700699
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700700 *vm_flags = 0;
Hugh Dickins3ca7b3c2009-12-14 17:58:57 -0800701 if (page_mapped(page) && page_rmapping(page)) {
Hugh Dickins5ad64682009-12-14 17:59:24 -0800702 if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
703 we_locked = trylock_page(page);
704 if (!we_locked) {
705 referenced++;
706 goto out;
707 }
708 }
709 if (unlikely(PageKsm(page)))
710 referenced += page_referenced_ksm(page, mem_cont,
711 vm_flags);
712 else if (PageAnon(page))
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700713 referenced += page_referenced_anon(page, mem_cont,
714 vm_flags);
Hugh Dickins5ad64682009-12-14 17:59:24 -0800715 else if (page->mapping)
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700716 referenced += page_referenced_file(page, mem_cont,
717 vm_flags);
Hugh Dickins5ad64682009-12-14 17:59:24 -0800718 if (we_locked)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700719 unlock_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700720 }
Hugh Dickins5ad64682009-12-14 17:59:24 -0800721out:
Christian Borntraeger5b7baf02008-03-25 18:47:12 +0100722 if (page_test_and_clear_young(page))
723 referenced++;
724
Linus Torvalds1da177e2005-04-16 15:20:36 -0700725 return referenced;
726}
727
Hugh Dickins1cb17292009-12-14 17:59:01 -0800728static int page_mkclean_one(struct page *page, struct vm_area_struct *vma,
729 unsigned long address)
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700730{
731 struct mm_struct *mm = vma->vm_mm;
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100732 pte_t *pte;
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700733 spinlock_t *ptl;
734 int ret = 0;
735
Nick Piggin479db0b2008-08-20 14:09:18 -0700736 pte = page_check_address(page, mm, address, &ptl, 1);
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700737 if (!pte)
738 goto out;
739
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100740 if (pte_dirty(*pte) || pte_write(*pte)) {
741 pte_t entry;
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700742
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100743 flush_cache_page(vma, address, pte_pfn(*pte));
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700744 entry = ptep_clear_flush_notify(vma, address, pte);
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100745 entry = pte_wrprotect(entry);
746 entry = pte_mkclean(entry);
Al Virod6e88e62006-12-29 16:48:35 -0800747 set_pte_at(mm, address, pte, entry);
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100748 ret = 1;
749 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700750
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700751 pte_unmap_unlock(pte, ptl);
752out:
753 return ret;
754}
755
756static int page_mkclean_file(struct address_space *mapping, struct page *page)
757{
758 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
759 struct vm_area_struct *vma;
760 struct prio_tree_iter iter;
761 int ret = 0;
762
763 BUG_ON(PageAnon(page));
764
765 spin_lock(&mapping->i_mmap_lock);
766 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
Hugh Dickins1cb17292009-12-14 17:59:01 -0800767 if (vma->vm_flags & VM_SHARED) {
768 unsigned long address = vma_address(page, vma);
769 if (address == -EFAULT)
770 continue;
771 ret += page_mkclean_one(page, vma, address);
772 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700773 }
774 spin_unlock(&mapping->i_mmap_lock);
775 return ret;
776}
777
778int page_mkclean(struct page *page)
779{
780 int ret = 0;
781
782 BUG_ON(!PageLocked(page));
783
784 if (page_mapped(page)) {
785 struct address_space *mapping = page_mapping(page);
Christian Borntraegerce7e9fa2007-11-20 11:13:36 +0100786 if (mapping) {
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700787 ret = page_mkclean_file(mapping, page);
Christian Borntraegerce7e9fa2007-11-20 11:13:36 +0100788 if (page_test_dirty(page)) {
Martin Schwidefskye2b8d7a2010-10-25 16:10:14 +0200789 page_clear_dirty(page, 1);
Christian Borntraegerce7e9fa2007-11-20 11:13:36 +0100790 ret = 1;
791 }
Martin Schwidefsky6c210482007-04-27 16:01:57 +0200792 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700793 }
794
795 return ret;
796}
Jaya Kumar60b59be2007-05-08 00:37:37 -0700797EXPORT_SYMBOL_GPL(page_mkclean);
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700798
Linus Torvalds1da177e2005-04-16 15:20:36 -0700799/**
Rik van Rielc44b6742010-03-05 13:42:09 -0800800 * page_move_anon_rmap - move a page to our anon_vma
801 * @page: the page to move to our anon_vma
802 * @vma: the vma the page belongs to
803 * @address: the user virtual address mapped
804 *
805 * When a page belongs exclusively to one process after a COW event,
806 * that page can be moved into the anon_vma that belongs to just that
807 * process, so the rmap code will not search the parent or sibling
808 * processes.
809 */
810void page_move_anon_rmap(struct page *page,
811 struct vm_area_struct *vma, unsigned long address)
812{
813 struct anon_vma *anon_vma = vma->anon_vma;
814
815 VM_BUG_ON(!PageLocked(page));
816 VM_BUG_ON(!anon_vma);
817 VM_BUG_ON(page->index != linear_page_index(vma, address));
818
819 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
820 page->mapping = (struct address_space *) anon_vma;
821}
822
823/**
Andi Kleen4e1c1972010-09-22 12:43:56 +0200824 * __page_set_anon_rmap - set up new anonymous rmap
825 * @page: Page to add to rmap
826 * @vma: VM area to add page to.
827 * @address: User virtual address of the mapping
Rik van Riele8a03fe2010-04-14 17:59:28 -0400828 * @exclusive: the page is exclusively owned by the current process
Nick Piggin9617d952006-01-06 00:11:12 -0800829 */
830static void __page_set_anon_rmap(struct page *page,
Rik van Riele8a03fe2010-04-14 17:59:28 -0400831 struct vm_area_struct *vma, unsigned long address, int exclusive)
Nick Piggin9617d952006-01-06 00:11:12 -0800832{
Rik van Riele8a03fe2010-04-14 17:59:28 -0400833 struct anon_vma *anon_vma = vma->anon_vma;
Nick Piggin9617d952006-01-06 00:11:12 -0800834
Rik van Riele8a03fe2010-04-14 17:59:28 -0400835 BUG_ON(!anon_vma);
Linus Torvaldsea900022010-04-12 12:44:29 -0700836
Andi Kleen4e1c1972010-09-22 12:43:56 +0200837 if (PageAnon(page))
838 return;
839
Linus Torvaldsea900022010-04-12 12:44:29 -0700840 /*
Rik van Riele8a03fe2010-04-14 17:59:28 -0400841 * If the page isn't exclusively mapped into this vma,
842 * we must use the _oldest_ possible anon_vma for the
843 * page mapping!
Linus Torvaldsea900022010-04-12 12:44:29 -0700844 */
Andi Kleen4e1c1972010-09-22 12:43:56 +0200845 if (!exclusive)
Andrea Arcangeli288468c2010-08-09 17:19:09 -0700846 anon_vma = anon_vma->root;
Linus Torvaldsea900022010-04-12 12:44:29 -0700847
Nick Piggin9617d952006-01-06 00:11:12 -0800848 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
849 page->mapping = (struct address_space *) anon_vma;
Nick Piggin9617d952006-01-06 00:11:12 -0800850 page->index = linear_page_index(vma, address);
Nick Piggin9617d952006-01-06 00:11:12 -0800851}
852
853/**
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700854 * __page_check_anon_rmap - sanity check anonymous rmap addition
Nick Pigginc97a9e12007-05-16 22:11:21 -0700855 * @page: the page to add the mapping to
856 * @vma: the vm area in which the mapping is added
857 * @address: the user virtual address mapped
858 */
859static void __page_check_anon_rmap(struct page *page,
860 struct vm_area_struct *vma, unsigned long address)
861{
862#ifdef CONFIG_DEBUG_VM
863 /*
864 * The page's anon-rmap details (mapping and index) are guaranteed to
865 * be set up correctly at this point.
866 *
867 * We have exclusion against page_add_anon_rmap because the caller
868 * always holds the page locked, except if called from page_dup_rmap,
869 * in which case the page is already known to be setup.
870 *
871 * We have exclusion against page_add_new_anon_rmap because those pages
872 * are initially only visible via the pagetables, and the pte is locked
873 * over the call to page_add_new_anon_rmap.
874 */
Andrea Arcangeli44ab57a2010-08-09 17:19:10 -0700875 BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root);
Nick Pigginc97a9e12007-05-16 22:11:21 -0700876 BUG_ON(page->index != linear_page_index(vma, address));
877#endif
878}
879
880/**
Linus Torvalds1da177e2005-04-16 15:20:36 -0700881 * page_add_anon_rmap - add pte mapping to an anonymous page
882 * @page: the page to add the mapping to
883 * @vma: the vm area in which the mapping is added
884 * @address: the user virtual address mapped
885 *
Hugh Dickins5ad64682009-12-14 17:59:24 -0800886 * The caller needs to hold the pte lock, and the page must be locked in
Hugh Dickins80e148222009-12-14 17:59:29 -0800887 * the anon_vma case: to serialize mapping,index checking after setting,
888 * and to ensure that PageAnon is not being upgraded racily to PageKsm
889 * (but PageKsm is never downgraded to PageAnon).
Linus Torvalds1da177e2005-04-16 15:20:36 -0700890 */
891void page_add_anon_rmap(struct page *page,
892 struct vm_area_struct *vma, unsigned long address)
893{
Rik van Rielad8c2ee2010-08-09 17:19:48 -0700894 do_page_add_anon_rmap(page, vma, address, 0);
895}
896
897/*
898 * Special version of the above for do_swap_page, which often runs
899 * into pages that are exclusively owned by the current process.
900 * Everybody else should continue to use page_add_anon_rmap above.
901 */
902void do_page_add_anon_rmap(struct page *page,
903 struct vm_area_struct *vma, unsigned long address, int exclusive)
904{
Hugh Dickins5ad64682009-12-14 17:59:24 -0800905 int first = atomic_inc_and_test(&page->_mapcount);
Andrea Arcangeli79134172011-01-13 15:46:58 -0800906 if (first) {
907 if (!PageTransHuge(page))
908 __inc_zone_page_state(page, NR_ANON_PAGES);
909 else
910 __inc_zone_page_state(page,
911 NR_ANON_TRANSPARENT_HUGEPAGES);
912 }
Hugh Dickins5ad64682009-12-14 17:59:24 -0800913 if (unlikely(PageKsm(page)))
914 return;
915
Nick Pigginc97a9e12007-05-16 22:11:21 -0700916 VM_BUG_ON(!PageLocked(page));
917 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
Hugh Dickins5ad64682009-12-14 17:59:24 -0800918 if (first)
Rik van Rielad8c2ee2010-08-09 17:19:48 -0700919 __page_set_anon_rmap(page, vma, address, exclusive);
KAMEZAWA Hiroyuki69029cd2008-07-25 01:47:14 -0700920 else
Nick Pigginc97a9e12007-05-16 22:11:21 -0700921 __page_check_anon_rmap(page, vma, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700922}
923
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700924/**
Nick Piggin9617d952006-01-06 00:11:12 -0800925 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
926 * @page: the page to add the mapping to
927 * @vma: the vm area in which the mapping is added
928 * @address: the user virtual address mapped
929 *
930 * Same as page_add_anon_rmap but must only be called on *new* pages.
931 * This means the inc-and-test can be bypassed.
Nick Pigginc97a9e12007-05-16 22:11:21 -0700932 * Page does not have to be locked.
Nick Piggin9617d952006-01-06 00:11:12 -0800933 */
934void page_add_new_anon_rmap(struct page *page,
935 struct vm_area_struct *vma, unsigned long address)
936{
Hugh Dickinsb5934c52009-01-06 14:39:25 -0800937 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
Hugh Dickinscbf84b72009-01-06 14:39:27 -0800938 SetPageSwapBacked(page);
939 atomic_set(&page->_mapcount, 0); /* increment count (starts at -1) */
Andrea Arcangeli79134172011-01-13 15:46:58 -0800940 if (!PageTransHuge(page))
941 __inc_zone_page_state(page, NR_ANON_PAGES);
942 else
943 __inc_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
Rik van Riele8a03fe2010-04-14 17:59:28 -0400944 __page_set_anon_rmap(page, vma, address, 1);
Hugh Dickinsb5934c52009-01-06 14:39:25 -0800945 if (page_evictable(page, vma))
Hugh Dickinscbf84b72009-01-06 14:39:27 -0800946 lru_cache_add_lru(page, LRU_ACTIVE_ANON);
Hugh Dickinsb5934c52009-01-06 14:39:25 -0800947 else
948 add_page_to_unevictable_list(page);
Nick Piggin9617d952006-01-06 00:11:12 -0800949}
950
Linus Torvalds1da177e2005-04-16 15:20:36 -0700951/**
952 * page_add_file_rmap - add pte mapping to a file page
953 * @page: the page to add the mapping to
954 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -0700955 * The caller needs to hold the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700956 */
957void page_add_file_rmap(struct page *page)
958{
Balbir Singhd69b0422009-06-17 16:26:34 -0700959 if (atomic_inc_and_test(&page->_mapcount)) {
Christoph Lameter65ba55f2006-06-30 01:55:34 -0700960 __inc_zone_page_state(page, NR_FILE_MAPPED);
Greg Thelen2a7106f2011-01-13 15:47:37 -0800961 mem_cgroup_inc_page_stat(page, MEMCG_NR_FILE_MAPPED);
Balbir Singhd69b0422009-06-17 16:26:34 -0700962 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700963}
964
965/**
966 * page_remove_rmap - take down pte mapping from a page
967 * @page: page to remove mapping from
968 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -0700969 * The caller needs to hold the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700970 */
Hugh Dickinsedc315f2009-01-06 14:40:11 -0800971void page_remove_rmap(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700972{
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -0700973 /* page still mapped by someone else? */
974 if (!atomic_add_negative(-1, &page->_mapcount))
975 return;
976
977 /*
978 * Now that the last pte has gone, s390 must transfer dirty
979 * flag from storage key to struct page. We can usually skip
980 * this if the page is anon, so about to be freed; but perhaps
981 * not if it's in swapcache - there might be another pte slot
982 * containing the swap entry, but page not yet written to swap.
983 */
984 if ((!PageAnon(page) || PageSwapCache(page)) && page_test_dirty(page)) {
Martin Schwidefskye2b8d7a2010-10-25 16:10:14 +0200985 page_clear_dirty(page, 1);
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -0700986 set_page_dirty(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700987 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900988 /*
989 * Hugepages are not counted in NR_ANON_PAGES nor NR_FILE_MAPPED
990 * and not charged by memcg for now.
991 */
992 if (unlikely(PageHuge(page)))
993 return;
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -0700994 if (PageAnon(page)) {
995 mem_cgroup_uncharge_page(page);
Andrea Arcangeli79134172011-01-13 15:46:58 -0800996 if (!PageTransHuge(page))
997 __dec_zone_page_state(page, NR_ANON_PAGES);
998 else
999 __dec_zone_page_state(page,
1000 NR_ANON_TRANSPARENT_HUGEPAGES);
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -07001001 } else {
1002 __dec_zone_page_state(page, NR_FILE_MAPPED);
Greg Thelen2a7106f2011-01-13 15:47:37 -08001003 mem_cgroup_dec_page_stat(page, MEMCG_NR_FILE_MAPPED);
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -07001004 }
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -07001005 /*
1006 * It would be tidy to reset the PageAnon mapping here,
1007 * but that might overwrite a racing page_add_anon_rmap
1008 * which increments mapcount after us but sets mapping
1009 * before us: so leave the reset to free_hot_cold_page,
1010 * and remember that it's only reliable while mapped.
1011 * Leaving it set also helps swapoff to reinstate ptes
1012 * faster for those pages still in swapcache.
1013 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001014}
1015
1016/*
1017 * Subfunctions of try_to_unmap: try_to_unmap_one called
1018 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
1019 */
Hugh Dickins5ad64682009-12-14 17:59:24 -08001020int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
1021 unsigned long address, enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001022{
1023 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001024 pte_t *pte;
1025 pte_t pteval;
Hugh Dickinsc0718802005-10-29 18:16:31 -07001026 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001027 int ret = SWAP_AGAIN;
1028
Nick Piggin479db0b2008-08-20 14:09:18 -07001029 pte = page_check_address(page, mm, address, &ptl, 0);
Hugh Dickinsc0718802005-10-29 18:16:31 -07001030 if (!pte)
Nikita Danilov81b40822005-05-01 08:58:36 -07001031 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001032
1033 /*
1034 * If the page is mlock()d, we cannot swap it out.
1035 * If it's recently referenced (perhaps page_referenced
1036 * skipped over this mm) then we should reactivate it.
1037 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001038 if (!(flags & TTU_IGNORE_MLOCK)) {
KOSAKI Motohirocaed0f482009-12-14 17:59:45 -08001039 if (vma->vm_flags & VM_LOCKED)
1040 goto out_mlock;
1041
Hugh Dickinsaf8e3352009-12-14 17:58:59 -08001042 if (TTU_ACTION(flags) == TTU_MUNLOCK)
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001043 goto out_unmap;
Andi Kleen14fa31b2009-09-16 11:50:10 +02001044 }
1045 if (!(flags & TTU_IGNORE_ACCESS)) {
Nick Pigginb291f002008-10-18 20:26:44 -07001046 if (ptep_clear_flush_young_notify(vma, address, pte)) {
1047 ret = SWAP_FAIL;
1048 goto out_unmap;
1049 }
1050 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001051
Linus Torvalds1da177e2005-04-16 15:20:36 -07001052 /* Nuke the page table entry. */
1053 flush_cache_page(vma, address, page_to_pfn(page));
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001054 pteval = ptep_clear_flush_notify(vma, address, pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001055
1056 /* Move the dirty bit to the physical page now the pte is gone. */
1057 if (pte_dirty(pteval))
1058 set_page_dirty(page);
1059
Hugh Dickins365e9c872005-10-29 18:16:18 -07001060 /* Update high watermark before we lower rss */
1061 update_hiwater_rss(mm);
1062
Andi Kleen888b9f72009-09-16 11:50:11 +02001063 if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
1064 if (PageAnon(page))
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001065 dec_mm_counter(mm, MM_ANONPAGES);
Andi Kleen888b9f72009-09-16 11:50:11 +02001066 else
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001067 dec_mm_counter(mm, MM_FILEPAGES);
Andi Kleen888b9f72009-09-16 11:50:11 +02001068 set_pte_at(mm, address, pte,
1069 swp_entry_to_pte(make_hwpoison_entry(page)));
1070 } else if (PageAnon(page)) {
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07001071 swp_entry_t entry = { .val = page_private(page) };
Christoph Lameter06972122006-06-23 02:03:35 -07001072
1073 if (PageSwapCache(page)) {
1074 /*
1075 * Store the swap location in the pte.
1076 * See handle_pte_fault() ...
1077 */
Hugh Dickins570a335b2009-12-14 17:58:46 -08001078 if (swap_duplicate(entry) < 0) {
1079 set_pte_at(mm, address, pte, pteval);
1080 ret = SWAP_FAIL;
1081 goto out_unmap;
1082 }
Christoph Lameter06972122006-06-23 02:03:35 -07001083 if (list_empty(&mm->mmlist)) {
1084 spin_lock(&mmlist_lock);
1085 if (list_empty(&mm->mmlist))
1086 list_add(&mm->mmlist, &init_mm.mmlist);
1087 spin_unlock(&mmlist_lock);
1088 }
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001089 dec_mm_counter(mm, MM_ANONPAGES);
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08001090 inc_mm_counter(mm, MM_SWAPENTS);
KOSAKI Motohiro64cdd542009-01-06 14:39:16 -08001091 } else if (PAGE_MIGRATION) {
Christoph Lameter06972122006-06-23 02:03:35 -07001092 /*
1093 * Store the pfn of the page in a special migration
1094 * pte. do_swap_page() will wait until the migration
1095 * pte is removed and then restart fault handling.
1096 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001097 BUG_ON(TTU_ACTION(flags) != TTU_MIGRATION);
Christoph Lameter06972122006-06-23 02:03:35 -07001098 entry = make_migration_entry(page, pte_write(pteval));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001099 }
1100 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
1101 BUG_ON(pte_file(*pte));
Andi Kleen14fa31b2009-09-16 11:50:10 +02001102 } else if (PAGE_MIGRATION && (TTU_ACTION(flags) == TTU_MIGRATION)) {
Christoph Lameter04e62a22006-06-23 02:03:38 -07001103 /* Establish migration entry for a file page */
1104 swp_entry_t entry;
1105 entry = make_migration_entry(page, pte_write(pteval));
1106 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
1107 } else
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001108 dec_mm_counter(mm, MM_FILEPAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001109
Hugh Dickinsedc315f2009-01-06 14:40:11 -08001110 page_remove_rmap(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001111 page_cache_release(page);
1112
1113out_unmap:
Hugh Dickinsc0718802005-10-29 18:16:31 -07001114 pte_unmap_unlock(pte, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001115out:
1116 return ret;
KOSAKI Motohirocaed0f482009-12-14 17:59:45 -08001117
1118out_mlock:
1119 pte_unmap_unlock(pte, ptl);
1120
1121
1122 /*
1123 * We need mmap_sem locking, Otherwise VM_LOCKED check makes
1124 * unstable result and race. Plus, We can't wait here because
1125 * we now hold anon_vma->lock or mapping->i_mmap_lock.
1126 * if trylock failed, the page remain in evictable lru and later
1127 * vmscan could retry to move the page to unevictable lru if the
1128 * page is actually mlocked.
1129 */
1130 if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
1131 if (vma->vm_flags & VM_LOCKED) {
1132 mlock_vma_page(page);
1133 ret = SWAP_MLOCK;
1134 }
1135 up_read(&vma->vm_mm->mmap_sem);
1136 }
1137 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001138}
1139
1140/*
1141 * objrmap doesn't work for nonlinear VMAs because the assumption that
1142 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
1143 * Consequently, given a particular page and its ->index, we cannot locate the
1144 * ptes which are mapping that page without an exhaustive linear search.
1145 *
1146 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
1147 * maps the file to which the target page belongs. The ->vm_private_data field
1148 * holds the current cursor into that scan. Successive searches will circulate
1149 * around the vma's virtual address space.
1150 *
1151 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
1152 * more scanning pressure is placed against them as well. Eventually pages
1153 * will become fully unmapped and are eligible for eviction.
1154 *
1155 * For very sparsely populated VMAs this is a little inefficient - chances are
1156 * there there won't be many ptes located within the scan cluster. In this case
1157 * maybe we could scan further - to the end of the pte page, perhaps.
Nick Pigginb291f002008-10-18 20:26:44 -07001158 *
1159 * Mlocked pages: check VM_LOCKED under mmap_sem held for read, if we can
1160 * acquire it without blocking. If vma locked, mlock the pages in the cluster,
1161 * rather than unmapping them. If we encounter the "check_page" that vmscan is
1162 * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001163 */
1164#define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
1165#define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
1166
Nick Pigginb291f002008-10-18 20:26:44 -07001167static int try_to_unmap_cluster(unsigned long cursor, unsigned int *mapcount,
1168 struct vm_area_struct *vma, struct page *check_page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001169{
1170 struct mm_struct *mm = vma->vm_mm;
1171 pgd_t *pgd;
1172 pud_t *pud;
1173 pmd_t *pmd;
Hugh Dickinsc0718802005-10-29 18:16:31 -07001174 pte_t *pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001175 pte_t pteval;
Hugh Dickinsc0718802005-10-29 18:16:31 -07001176 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001177 struct page *page;
1178 unsigned long address;
1179 unsigned long end;
Nick Pigginb291f002008-10-18 20:26:44 -07001180 int ret = SWAP_AGAIN;
1181 int locked_vma = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001182
Linus Torvalds1da177e2005-04-16 15:20:36 -07001183 address = (vma->vm_start + cursor) & CLUSTER_MASK;
1184 end = address + CLUSTER_SIZE;
1185 if (address < vma->vm_start)
1186 address = vma->vm_start;
1187 if (end > vma->vm_end)
1188 end = vma->vm_end;
1189
1190 pgd = pgd_offset(mm, address);
1191 if (!pgd_present(*pgd))
Nick Pigginb291f002008-10-18 20:26:44 -07001192 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001193
1194 pud = pud_offset(pgd, address);
1195 if (!pud_present(*pud))
Nick Pigginb291f002008-10-18 20:26:44 -07001196 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001197
1198 pmd = pmd_offset(pud, address);
1199 if (!pmd_present(*pmd))
Nick Pigginb291f002008-10-18 20:26:44 -07001200 return ret;
1201
1202 /*
Hugh Dickinsaf8e3352009-12-14 17:58:59 -08001203 * If we can acquire the mmap_sem for read, and vma is VM_LOCKED,
Nick Pigginb291f002008-10-18 20:26:44 -07001204 * keep the sem while scanning the cluster for mlocking pages.
1205 */
Hugh Dickinsaf8e3352009-12-14 17:58:59 -08001206 if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
Nick Pigginb291f002008-10-18 20:26:44 -07001207 locked_vma = (vma->vm_flags & VM_LOCKED);
1208 if (!locked_vma)
1209 up_read(&vma->vm_mm->mmap_sem); /* don't need it */
1210 }
Hugh Dickinsc0718802005-10-29 18:16:31 -07001211
1212 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001213
Hugh Dickins365e9c872005-10-29 18:16:18 -07001214 /* Update high watermark before we lower rss */
1215 update_hiwater_rss(mm);
1216
Hugh Dickinsc0718802005-10-29 18:16:31 -07001217 for (; address < end; pte++, address += PAGE_SIZE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001218 if (!pte_present(*pte))
1219 continue;
Linus Torvalds6aab3412005-11-28 14:34:23 -08001220 page = vm_normal_page(vma, address, *pte);
1221 BUG_ON(!page || PageAnon(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001222
Nick Pigginb291f002008-10-18 20:26:44 -07001223 if (locked_vma) {
1224 mlock_vma_page(page); /* no-op if already mlocked */
1225 if (page == check_page)
1226 ret = SWAP_MLOCK;
1227 continue; /* don't unmap */
1228 }
1229
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001230 if (ptep_clear_flush_young_notify(vma, address, pte))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001231 continue;
1232
1233 /* Nuke the page table entry. */
Ben Collinseca35132005-11-29 11:45:26 -08001234 flush_cache_page(vma, address, pte_pfn(*pte));
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001235 pteval = ptep_clear_flush_notify(vma, address, pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001236
1237 /* If nonlinear, store the file page offset in the pte. */
1238 if (page->index != linear_page_index(vma, address))
1239 set_pte_at(mm, address, pte, pgoff_to_pte(page->index));
1240
1241 /* Move the dirty bit to the physical page now the pte is gone. */
1242 if (pte_dirty(pteval))
1243 set_page_dirty(page);
1244
Hugh Dickinsedc315f2009-01-06 14:40:11 -08001245 page_remove_rmap(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001246 page_cache_release(page);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001247 dec_mm_counter(mm, MM_FILEPAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001248 (*mapcount)--;
1249 }
Hugh Dickinsc0718802005-10-29 18:16:31 -07001250 pte_unmap_unlock(pte - 1, ptl);
Nick Pigginb291f002008-10-18 20:26:44 -07001251 if (locked_vma)
1252 up_read(&vma->vm_mm->mmap_sem);
1253 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001254}
1255
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001256bool is_vma_temporary_stack(struct vm_area_struct *vma)
Mel Gormana8bef8ff2010-05-24 14:32:24 -07001257{
1258 int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP);
1259
1260 if (!maybe_stack)
1261 return false;
1262
1263 if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) ==
1264 VM_STACK_INCOMPLETE_SETUP)
1265 return true;
1266
1267 return false;
1268}
1269
Nick Pigginb291f002008-10-18 20:26:44 -07001270/**
1271 * try_to_unmap_anon - unmap or unlock anonymous page using the object-based
1272 * rmap method
1273 * @page: the page to unmap/unlock
Huang Shijie8051be52009-12-14 17:58:50 -08001274 * @flags: action and flags
Nick Pigginb291f002008-10-18 20:26:44 -07001275 *
1276 * Find all the mappings of a page using the mapping pointer and the vma chains
1277 * contained in the anon_vma struct it points to.
1278 *
1279 * This function is only called from try_to_unmap/try_to_munlock for
1280 * anonymous pages.
1281 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1282 * where the page was found will be held for write. So, we won't recheck
1283 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1284 * 'LOCKED.
1285 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001286static int try_to_unmap_anon(struct page *page, enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001287{
1288 struct anon_vma *anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -08001289 struct anon_vma_chain *avc;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001290 int ret = SWAP_AGAIN;
Nick Pigginb291f002008-10-18 20:26:44 -07001291
Linus Torvalds1da177e2005-04-16 15:20:36 -07001292 anon_vma = page_lock_anon_vma(page);
1293 if (!anon_vma)
1294 return ret;
1295
Rik van Riel5beb4932010-03-05 13:42:07 -08001296 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1297 struct vm_area_struct *vma = avc->vma;
Mel Gormana8bef8ff2010-05-24 14:32:24 -07001298 unsigned long address;
1299
1300 /*
1301 * During exec, a temporary VMA is setup and later moved.
1302 * The VMA is moved under the anon_vma lock but not the
1303 * page tables leading to a race where migration cannot
1304 * find the migration ptes. Rather than increasing the
1305 * locking requirements of exec(), migration skips
1306 * temporary VMAs until after exec() completes.
1307 */
1308 if (PAGE_MIGRATION && (flags & TTU_MIGRATION) &&
1309 is_vma_temporary_stack(vma))
1310 continue;
1311
1312 address = vma_address(page, vma);
Hugh Dickins1cb17292009-12-14 17:59:01 -08001313 if (address == -EFAULT)
1314 continue;
1315 ret = try_to_unmap_one(page, vma, address, flags);
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001316 if (ret != SWAP_AGAIN || !page_mapped(page))
1317 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001318 }
Oleg Nesterov34bbd702007-02-28 20:13:49 -08001319
1320 page_unlock_anon_vma(anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001321 return ret;
1322}
1323
1324/**
Nick Pigginb291f002008-10-18 20:26:44 -07001325 * try_to_unmap_file - unmap/unlock file page using the object-based rmap method
1326 * @page: the page to unmap/unlock
Andi Kleen14fa31b2009-09-16 11:50:10 +02001327 * @flags: action and flags
Linus Torvalds1da177e2005-04-16 15:20:36 -07001328 *
1329 * Find all the mappings of a page using the mapping pointer and the vma chains
1330 * contained in the address_space struct it points to.
1331 *
Nick Pigginb291f002008-10-18 20:26:44 -07001332 * This function is only called from try_to_unmap/try_to_munlock for
1333 * object-based pages.
1334 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1335 * where the page was found will be held for write. So, we won't recheck
1336 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1337 * 'LOCKED.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001338 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001339static int try_to_unmap_file(struct page *page, enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001340{
1341 struct address_space *mapping = page->mapping;
1342 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1343 struct vm_area_struct *vma;
1344 struct prio_tree_iter iter;
1345 int ret = SWAP_AGAIN;
1346 unsigned long cursor;
1347 unsigned long max_nl_cursor = 0;
1348 unsigned long max_nl_size = 0;
1349 unsigned int mapcount;
1350
1351 spin_lock(&mapping->i_mmap_lock);
1352 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
Hugh Dickins1cb17292009-12-14 17:59:01 -08001353 unsigned long address = vma_address(page, vma);
1354 if (address == -EFAULT)
1355 continue;
1356 ret = try_to_unmap_one(page, vma, address, flags);
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001357 if (ret != SWAP_AGAIN || !page_mapped(page))
1358 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001359 }
1360
1361 if (list_empty(&mapping->i_mmap_nonlinear))
1362 goto out;
1363
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001364 /*
1365 * We don't bother to try to find the munlocked page in nonlinears.
1366 * It's costly. Instead, later, page reclaim logic may call
1367 * try_to_unmap(TTU_MUNLOCK) and recover PG_mlocked lazily.
1368 */
1369 if (TTU_ACTION(flags) == TTU_MUNLOCK)
1370 goto out;
1371
Linus Torvalds1da177e2005-04-16 15:20:36 -07001372 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1373 shared.vm_set.list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001374 cursor = (unsigned long) vma->vm_private_data;
1375 if (cursor > max_nl_cursor)
1376 max_nl_cursor = cursor;
1377 cursor = vma->vm_end - vma->vm_start;
1378 if (cursor > max_nl_size)
1379 max_nl_size = cursor;
1380 }
1381
Nick Pigginb291f002008-10-18 20:26:44 -07001382 if (max_nl_size == 0) { /* all nonlinears locked or reserved ? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001383 ret = SWAP_FAIL;
1384 goto out;
1385 }
1386
1387 /*
1388 * We don't try to search for this page in the nonlinear vmas,
1389 * and page_referenced wouldn't have found it anyway. Instead
1390 * just walk the nonlinear vmas trying to age and unmap some.
1391 * The mapcount of the page we came in with is irrelevant,
1392 * but even so use it as a guide to how hard we should try?
1393 */
1394 mapcount = page_mapcount(page);
1395 if (!mapcount)
1396 goto out;
1397 cond_resched_lock(&mapping->i_mmap_lock);
1398
1399 max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
1400 if (max_nl_cursor == 0)
1401 max_nl_cursor = CLUSTER_SIZE;
1402
1403 do {
1404 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1405 shared.vm_set.list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001406 cursor = (unsigned long) vma->vm_private_data;
Hugh Dickins839b9682005-09-03 15:54:43 -07001407 while ( cursor < max_nl_cursor &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07001408 cursor < vma->vm_end - vma->vm_start) {
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001409 if (try_to_unmap_cluster(cursor, &mapcount,
1410 vma, page) == SWAP_MLOCK)
1411 ret = SWAP_MLOCK;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001412 cursor += CLUSTER_SIZE;
1413 vma->vm_private_data = (void *) cursor;
1414 if ((int)mapcount <= 0)
1415 goto out;
1416 }
1417 vma->vm_private_data = (void *) max_nl_cursor;
1418 }
1419 cond_resched_lock(&mapping->i_mmap_lock);
1420 max_nl_cursor += CLUSTER_SIZE;
1421 } while (max_nl_cursor <= max_nl_size);
1422
1423 /*
1424 * Don't loop forever (perhaps all the remaining pages are
1425 * in locked vmas). Reset cursor on all unreserved nonlinear
1426 * vmas, now forgetting on which ones it had fallen behind.
1427 */
Hugh Dickins101d2be2005-11-21 21:32:16 -08001428 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1429 vma->vm_private_data = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001430out:
1431 spin_unlock(&mapping->i_mmap_lock);
1432 return ret;
1433}
1434
1435/**
1436 * try_to_unmap - try to remove all page table mappings to a page
1437 * @page: the page to get unmapped
Andi Kleen14fa31b2009-09-16 11:50:10 +02001438 * @flags: action and flags
Linus Torvalds1da177e2005-04-16 15:20:36 -07001439 *
1440 * Tries to remove all the page table entries which are mapping this
1441 * page, used in the pageout path. Caller must hold the page lock.
1442 * Return values are:
1443 *
1444 * SWAP_SUCCESS - we succeeded in removing all mappings
1445 * SWAP_AGAIN - we missed a mapping, try again later
1446 * SWAP_FAIL - the page is unswappable
Nick Pigginb291f002008-10-18 20:26:44 -07001447 * SWAP_MLOCK - page is mlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001448 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001449int try_to_unmap(struct page *page, enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001450{
1451 int ret;
1452
Linus Torvalds1da177e2005-04-16 15:20:36 -07001453 BUG_ON(!PageLocked(page));
Andrea Arcangeli91600e92011-01-13 15:47:16 -08001454 VM_BUG_ON(!PageHuge(page) && PageTransHuge(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001455
Hugh Dickins5ad64682009-12-14 17:59:24 -08001456 if (unlikely(PageKsm(page)))
1457 ret = try_to_unmap_ksm(page, flags);
1458 else if (PageAnon(page))
Andi Kleen14fa31b2009-09-16 11:50:10 +02001459 ret = try_to_unmap_anon(page, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001460 else
Andi Kleen14fa31b2009-09-16 11:50:10 +02001461 ret = try_to_unmap_file(page, flags);
Nick Pigginb291f002008-10-18 20:26:44 -07001462 if (ret != SWAP_MLOCK && !page_mapped(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001463 ret = SWAP_SUCCESS;
1464 return ret;
1465}
Nikita Danilov81b40822005-05-01 08:58:36 -07001466
Nick Pigginb291f002008-10-18 20:26:44 -07001467/**
1468 * try_to_munlock - try to munlock a page
1469 * @page: the page to be munlocked
1470 *
1471 * Called from munlock code. Checks all of the VMAs mapping the page
1472 * to make sure nobody else has this page mlocked. The page will be
1473 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1474 *
1475 * Return values are:
1476 *
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001477 * SWAP_AGAIN - no vma is holding page mlocked, or,
Nick Pigginb291f002008-10-18 20:26:44 -07001478 * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem
Hugh Dickins5ad64682009-12-14 17:59:24 -08001479 * SWAP_FAIL - page cannot be located at present
Nick Pigginb291f002008-10-18 20:26:44 -07001480 * SWAP_MLOCK - page is now mlocked.
1481 */
1482int try_to_munlock(struct page *page)
1483{
1484 VM_BUG_ON(!PageLocked(page) || PageLRU(page));
1485
Hugh Dickins5ad64682009-12-14 17:59:24 -08001486 if (unlikely(PageKsm(page)))
1487 return try_to_unmap_ksm(page, TTU_MUNLOCK);
1488 else if (PageAnon(page))
Andi Kleen14fa31b2009-09-16 11:50:10 +02001489 return try_to_unmap_anon(page, TTU_MUNLOCK);
Nick Pigginb291f002008-10-18 20:26:44 -07001490 else
Andi Kleen14fa31b2009-09-16 11:50:10 +02001491 return try_to_unmap_file(page, TTU_MUNLOCK);
Nick Pigginb291f002008-10-18 20:26:44 -07001492}
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001493
Peter Zijlstra01d8b202011-03-22 16:32:49 -07001494void __put_anon_vma(struct anon_vma *anon_vma)
Rik van Riel76545062010-08-09 17:18:41 -07001495{
Peter Zijlstra01d8b202011-03-22 16:32:49 -07001496 struct anon_vma *root = anon_vma->root;
Rik van Riel76545062010-08-09 17:18:41 -07001497
Peter Zijlstra01d8b202011-03-22 16:32:49 -07001498 if (root != anon_vma && atomic_dec_and_test(&root->refcount))
1499 anon_vma_free(root);
Rik van Riel76545062010-08-09 17:18:41 -07001500
Peter Zijlstra01d8b202011-03-22 16:32:49 -07001501 anon_vma_free(anon_vma);
Rik van Riel76545062010-08-09 17:18:41 -07001502}
Rik van Riel76545062010-08-09 17:18:41 -07001503
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001504#ifdef CONFIG_MIGRATION
1505/*
1506 * rmap_walk() and its helpers rmap_walk_anon() and rmap_walk_file():
1507 * Called by migrate.c to remove migration ptes, but might be used more later.
1508 */
1509static int rmap_walk_anon(struct page *page, int (*rmap_one)(struct page *,
1510 struct vm_area_struct *, unsigned long, void *), void *arg)
1511{
1512 struct anon_vma *anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -08001513 struct anon_vma_chain *avc;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001514 int ret = SWAP_AGAIN;
1515
1516 /*
1517 * Note: remove_migration_ptes() cannot use page_lock_anon_vma()
1518 * because that depends on page_mapped(); but not all its usages
Mel Gorman3f6c8272010-05-24 14:32:17 -07001519 * are holding mmap_sem. Users without mmap_sem are required to
1520 * take a reference count to prevent the anon_vma disappearing
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001521 */
1522 anon_vma = page_anon_vma(page);
1523 if (!anon_vma)
1524 return ret;
Rik van Rielcba48b92010-08-09 17:18:38 -07001525 anon_vma_lock(anon_vma);
Rik van Riel5beb4932010-03-05 13:42:07 -08001526 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1527 struct vm_area_struct *vma = avc->vma;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001528 unsigned long address = vma_address(page, vma);
1529 if (address == -EFAULT)
1530 continue;
1531 ret = rmap_one(page, vma, address, arg);
1532 if (ret != SWAP_AGAIN)
1533 break;
1534 }
Rik van Rielcba48b92010-08-09 17:18:38 -07001535 anon_vma_unlock(anon_vma);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001536 return ret;
1537}
1538
1539static int rmap_walk_file(struct page *page, int (*rmap_one)(struct page *,
1540 struct vm_area_struct *, unsigned long, void *), void *arg)
1541{
1542 struct address_space *mapping = page->mapping;
1543 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1544 struct vm_area_struct *vma;
1545 struct prio_tree_iter iter;
1546 int ret = SWAP_AGAIN;
1547
1548 if (!mapping)
1549 return ret;
1550 spin_lock(&mapping->i_mmap_lock);
1551 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1552 unsigned long address = vma_address(page, vma);
1553 if (address == -EFAULT)
1554 continue;
1555 ret = rmap_one(page, vma, address, arg);
1556 if (ret != SWAP_AGAIN)
1557 break;
1558 }
1559 /*
1560 * No nonlinear handling: being always shared, nonlinear vmas
1561 * never contain migration ptes. Decide what to do about this
1562 * limitation to linear when we need rmap_walk() on nonlinear.
1563 */
1564 spin_unlock(&mapping->i_mmap_lock);
1565 return ret;
1566}
1567
1568int rmap_walk(struct page *page, int (*rmap_one)(struct page *,
1569 struct vm_area_struct *, unsigned long, void *), void *arg)
1570{
1571 VM_BUG_ON(!PageLocked(page));
1572
1573 if (unlikely(PageKsm(page)))
1574 return rmap_walk_ksm(page, rmap_one, arg);
1575 else if (PageAnon(page))
1576 return rmap_walk_anon(page, rmap_one, arg);
1577 else
1578 return rmap_walk_file(page, rmap_one, arg);
1579}
1580#endif /* CONFIG_MIGRATION */
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001581
Naoya Horiguchie3390f62010-06-15 13:18:13 +09001582#ifdef CONFIG_HUGETLB_PAGE
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001583/*
1584 * The following three functions are for anonymous (private mapped) hugepages.
1585 * Unlike common anonymous pages, anonymous hugepages have no accounting code
1586 * and no lru code, because we handle hugepages differently from common pages.
1587 */
1588static void __hugepage_set_anon_rmap(struct page *page,
1589 struct vm_area_struct *vma, unsigned long address, int exclusive)
1590{
1591 struct anon_vma *anon_vma = vma->anon_vma;
Naoya Horiguchi433abed2010-09-10 13:23:03 +09001592
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001593 BUG_ON(!anon_vma);
Naoya Horiguchi433abed2010-09-10 13:23:03 +09001594
1595 if (PageAnon(page))
1596 return;
1597 if (!exclusive)
1598 anon_vma = anon_vma->root;
1599
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001600 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1601 page->mapping = (struct address_space *) anon_vma;
1602 page->index = linear_page_index(vma, address);
1603}
1604
1605void hugepage_add_anon_rmap(struct page *page,
1606 struct vm_area_struct *vma, unsigned long address)
1607{
1608 struct anon_vma *anon_vma = vma->anon_vma;
1609 int first;
Naoya Horiguchia850ea32010-09-10 13:23:06 +09001610
1611 BUG_ON(!PageLocked(page));
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001612 BUG_ON(!anon_vma);
1613 BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1614 first = atomic_inc_and_test(&page->_mapcount);
1615 if (first)
1616 __hugepage_set_anon_rmap(page, vma, address, 0);
1617}
1618
1619void hugepage_add_new_anon_rmap(struct page *page,
1620 struct vm_area_struct *vma, unsigned long address)
1621{
1622 BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1623 atomic_set(&page->_mapcount, 0);
1624 __hugepage_set_anon_rmap(page, vma, address, 1);
1625}
Naoya Horiguchie3390f62010-06-15 13:18:13 +09001626#endif /* CONFIG_HUGETLB_PAGE */