blob: 1a8bf76bfd038a7fe84fdd6b9443c4673c8ddd9f [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
97 * if not we either need to find an adjacent mapping that we
98 * 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 Riel5c341ee2010-08-09 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);
Rik van Riel5beb4932010-03-05 13:42:07 -0800180 list_add_tail(&avc->same_anon_vma, &anon_vma->head);
Rik van Rielcba48b92010-08-09 17:18:38 -0700181 anon_vma_unlock(anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700182}
183
Rik van Riel5beb4932010-03-05 13:42:07 -0800184/*
185 * Attach the anon_vmas from src to dst.
186 * Returns 0 on success, -ENOMEM on failure.
187 */
188int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700189{
Rik van Riel5beb4932010-03-05 13:42:07 -0800190 struct anon_vma_chain *avc, *pavc;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700191
Linus Torvalds646d87b2010-04-11 17:15:03 -0700192 list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
Rik van Riel5beb4932010-03-05 13:42:07 -0800193 avc = anon_vma_chain_alloc();
194 if (!avc)
195 goto enomem_failure;
196 anon_vma_chain_link(dst, avc, pavc->anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700197 }
Rik van Riel5beb4932010-03-05 13:42:07 -0800198 return 0;
199
200 enomem_failure:
201 unlink_anon_vmas(dst);
202 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700203}
204
Rik van Riel5beb4932010-03-05 13:42:07 -0800205/*
206 * Attach vma to its own anon_vma, as well as to the anon_vmas that
207 * the corresponding VMA in the parent process is attached to.
208 * Returns 0 on success, non-zero on failure.
209 */
210int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700211{
Rik van Riel5beb4932010-03-05 13:42:07 -0800212 struct anon_vma_chain *avc;
213 struct anon_vma *anon_vma;
214
215 /* Don't bother if the parent process has no anon_vma here. */
216 if (!pvma->anon_vma)
217 return 0;
218
219 /*
220 * First, attach the new VMA to the parent VMA's anon_vmas,
221 * so rmap can find non-COWed pages in child processes.
222 */
223 if (anon_vma_clone(vma, pvma))
224 return -ENOMEM;
225
226 /* Then add our own anon_vma. */
227 anon_vma = anon_vma_alloc();
228 if (!anon_vma)
229 goto out_error;
230 avc = anon_vma_chain_alloc();
231 if (!avc)
232 goto out_error_free_anon_vma;
Rik van Riel5c341ee2010-08-09 17:18:39 -0700233
234 /*
235 * The root anon_vma's spinlock is the lock actually used when we
236 * lock any of the anon_vmas in this anon_vma tree.
237 */
238 anon_vma->root = pvma->anon_vma->root;
Rik van Riel76545062010-08-09 17:18:41 -0700239 /*
240 * With KSM refcounts, an anon_vma can stay around longer than the
241 * process it belongs to. The root anon_vma needs to be pinned
242 * until this anon_vma is freed, because the lock lives in the root.
243 */
244 get_anon_vma(anon_vma->root);
Rik van Riel5beb4932010-03-05 13:42:07 -0800245 /* Mark this anon_vma as the one where our new (COWed) pages go. */
246 vma->anon_vma = anon_vma;
Rik van Riel5c341ee2010-08-09 17:18:39 -0700247 anon_vma_chain_link(vma, avc, anon_vma);
Rik van Riel5beb4932010-03-05 13:42:07 -0800248
249 return 0;
250
251 out_error_free_anon_vma:
252 anon_vma_free(anon_vma);
253 out_error:
Rik van Riel4946d542010-04-05 12:13:33 -0400254 unlink_anon_vmas(vma);
Rik van Riel5beb4932010-03-05 13:42:07 -0800255 return -ENOMEM;
256}
257
258static void anon_vma_unlink(struct anon_vma_chain *anon_vma_chain)
259{
260 struct anon_vma *anon_vma = anon_vma_chain->anon_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700261 int empty;
262
Rik van Riel5beb4932010-03-05 13:42:07 -0800263 /* If anon_vma_fork fails, we can get an empty anon_vma_chain. */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700264 if (!anon_vma)
265 return;
266
Rik van Rielcba48b92010-08-09 17:18:38 -0700267 anon_vma_lock(anon_vma);
Rik van Riel5beb4932010-03-05 13:42:07 -0800268 list_del(&anon_vma_chain->same_anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700269
270 /* We must garbage collect the anon_vma if it's empty */
Mel Gorman7f60c212010-05-24 14:32:18 -0700271 empty = list_empty(&anon_vma->head) && !anonvma_external_refcount(anon_vma);
Rik van Rielcba48b92010-08-09 17:18:38 -0700272 anon_vma_unlock(anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700273
Rik van Riel76545062010-08-09 17:18:41 -0700274 if (empty) {
275 /* We no longer need the root anon_vma */
276 if (anon_vma->root != anon_vma)
277 drop_anon_vma(anon_vma->root);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700278 anon_vma_free(anon_vma);
Rik van Riel76545062010-08-09 17:18:41 -0700279 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700280}
281
Rik van Riel5beb4932010-03-05 13:42:07 -0800282void unlink_anon_vmas(struct vm_area_struct *vma)
283{
284 struct anon_vma_chain *avc, *next;
285
Rik van Riel5c341ee2010-08-09 17:18:39 -0700286 /*
287 * Unlink each anon_vma chained to the VMA. This list is ordered
288 * from newest to oldest, ensuring the root anon_vma gets freed last.
289 */
Rik van Riel5beb4932010-03-05 13:42:07 -0800290 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
291 anon_vma_unlink(avc);
292 list_del(&avc->same_vma);
293 anon_vma_chain_free(avc);
294 }
295}
296
Alexey Dobriyan51cc5062008-07-25 19:45:34 -0700297static void anon_vma_ctor(void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700298{
Christoph Lametera35afb82007-05-16 22:10:57 -0700299 struct anon_vma *anon_vma = data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700300
Christoph Lametera35afb82007-05-16 22:10:57 -0700301 spin_lock_init(&anon_vma->lock);
Mel Gorman7f60c212010-05-24 14:32:18 -0700302 anonvma_external_refcount_init(anon_vma);
Christoph Lametera35afb82007-05-16 22:10:57 -0700303 INIT_LIST_HEAD(&anon_vma->head);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700304}
305
306void __init anon_vma_init(void)
307{
308 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
Paul Mundt20c2df82007-07-20 10:11:58 +0900309 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor);
Rik van Riel5beb4932010-03-05 13:42:07 -0800310 anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain, SLAB_PANIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700311}
312
313/*
314 * Getting a lock on a stable anon_vma from a page off the LRU is
315 * tricky: page_lock_anon_vma rely on RCU to guard against the races.
316 */
Namhyung Kimea4525b2010-10-26 14:22:01 -0700317struct anon_vma *__page_lock_anon_vma(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700318{
Hugh Dickinsf1819422010-08-25 23:12:54 -0700319 struct anon_vma *anon_vma, *root_anon_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700320 unsigned long anon_mapping;
321
322 rcu_read_lock();
Hugh Dickins80e14822009-12-14 17:59:29 -0800323 anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping);
Hugh Dickins3ca7b3c2009-12-14 17:58:57 -0800324 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700325 goto out;
326 if (!page_mapped(page))
327 goto out;
328
329 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
Hugh Dickinsf1819422010-08-25 23:12:54 -0700330 root_anon_vma = ACCESS_ONCE(anon_vma->root);
331 spin_lock(&root_anon_vma->lock);
332
333 /*
334 * If this page is still mapped, then its anon_vma cannot have been
335 * freed. But if it has been unmapped, we have no security against
336 * the anon_vma structure being freed and reused (for another anon_vma:
337 * SLAB_DESTROY_BY_RCU guarantees that - so the spin_lock above cannot
338 * corrupt): with anon_vma_prepare() or anon_vma_fork() redirecting
339 * anon_vma->root before page_unlock_anon_vma() is called to unlock.
340 */
341 if (page_mapped(page))
342 return anon_vma;
343
344 spin_unlock(&root_anon_vma->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700345out:
346 rcu_read_unlock();
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800347 return NULL;
348}
349
Andi Kleen10be22d2009-09-16 11:50:04 +0200350void page_unlock_anon_vma(struct anon_vma *anon_vma)
Namhyung Kimea4525b2010-10-26 14:22:01 -0700351 __releases(&anon_vma->root->lock)
352 __releases(RCU)
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800353{
Rik van Rielcba48b92010-08-09 17:18:38 -0700354 anon_vma_unlock(anon_vma);
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800355 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700356}
357
358/*
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -0800359 * At what user virtual address is page expected in @vma?
360 * Returns virtual address or -EFAULT if page's index/offset is not
361 * within the range mapped the @vma.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700362 */
363static inline unsigned long
364vma_address(struct page *page, struct vm_area_struct *vma)
365{
366 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
367 unsigned long address;
368
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900369 if (unlikely(is_vm_hugetlb_page(vma)))
370 pgoff = page->index << huge_page_order(page_hstate(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700371 address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
372 if (unlikely(address < vma->vm_start || address >= vma->vm_end)) {
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -0800373 /* page should be within @vma mapping range */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700374 return -EFAULT;
375 }
376 return address;
377}
378
379/*
Huang Shijiebf89c8c2009-10-01 15:44:04 -0700380 * At what user virtual address is page expected in vma?
Naoya Horiguchiab941e02010-05-11 14:06:55 -0700381 * Caller should check the page is actually part of the vma.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700382 */
383unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
384{
Andrea Arcangeli21d0d442010-08-09 17:19:10 -0700385 if (PageAnon(page)) {
Hugh Dickins4829b902010-10-02 17:46:06 -0700386 struct anon_vma *page__anon_vma = page_anon_vma(page);
387 /*
388 * Note: swapoff's unuse_vma() is more efficient with this
389 * check, and needs it to match anon_vma when KSM is active.
390 */
391 if (!vma->anon_vma || !page__anon_vma ||
392 vma->anon_vma->root != page__anon_vma->root)
Andrea Arcangeli21d0d442010-08-09 17:19:10 -0700393 return -EFAULT;
394 } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) {
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800395 if (!vma->vm_file ||
396 vma->vm_file->f_mapping != page->mapping)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700397 return -EFAULT;
398 } else
399 return -EFAULT;
400 return vma_address(page, vma);
401}
402
403/*
Nikita Danilov81b40822005-05-01 08:58:36 -0700404 * Check that @page is mapped at @address into @mm.
405 *
Nick Piggin479db0b2008-08-20 14:09:18 -0700406 * If @sync is false, page_check_address may perform a racy check to avoid
407 * the page table lock when the pte is not present (helpful when reclaiming
408 * highly shared pages).
409 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -0700410 * On success returns with pte mapped and locked.
Nikita Danilov81b40822005-05-01 08:58:36 -0700411 */
Namhyung Kime9a81a82010-10-26 14:22:01 -0700412pte_t *__page_check_address(struct page *page, struct mm_struct *mm,
Nick Piggin479db0b2008-08-20 14:09:18 -0700413 unsigned long address, spinlock_t **ptlp, int sync)
Nikita Danilov81b40822005-05-01 08:58:36 -0700414{
415 pgd_t *pgd;
416 pud_t *pud;
417 pmd_t *pmd;
418 pte_t *pte;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700419 spinlock_t *ptl;
Nikita Danilov81b40822005-05-01 08:58:36 -0700420
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900421 if (unlikely(PageHuge(page))) {
422 pte = huge_pte_offset(mm, address);
423 ptl = &mm->page_table_lock;
424 goto check;
425 }
426
Nikita Danilov81b40822005-05-01 08:58:36 -0700427 pgd = pgd_offset(mm, address);
Hugh Dickinsc0718802005-10-29 18:16:31 -0700428 if (!pgd_present(*pgd))
429 return NULL;
430
431 pud = pud_offset(pgd, address);
432 if (!pud_present(*pud))
433 return NULL;
434
435 pmd = pmd_offset(pud, address);
436 if (!pmd_present(*pmd))
437 return NULL;
438
439 pte = pte_offset_map(pmd, address);
440 /* Make a quick check before getting the lock */
Nick Piggin479db0b2008-08-20 14:09:18 -0700441 if (!sync && !pte_present(*pte)) {
Hugh Dickinsc0718802005-10-29 18:16:31 -0700442 pte_unmap(pte);
443 return NULL;
Nikita Danilov81b40822005-05-01 08:58:36 -0700444 }
Hugh Dickinsc0718802005-10-29 18:16:31 -0700445
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700446 ptl = pte_lockptr(mm, pmd);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900447check:
Hugh Dickinsc0718802005-10-29 18:16:31 -0700448 spin_lock(ptl);
449 if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
450 *ptlp = ptl;
451 return pte;
452 }
453 pte_unmap_unlock(pte, ptl);
454 return NULL;
Nikita Danilov81b40822005-05-01 08:58:36 -0700455}
456
Nick Pigginb291f002008-10-18 20:26:44 -0700457/**
458 * page_mapped_in_vma - check whether a page is really mapped in a VMA
459 * @page: the page to test
460 * @vma: the VMA to test
461 *
462 * Returns 1 if the page is mapped into the page tables of the VMA, 0
463 * if the page is not mapped into the page tables of this VMA. Only
464 * valid for normal file or anonymous VMAs.
465 */
Andi Kleen6a460792009-09-16 11:50:15 +0200466int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
Nick Pigginb291f002008-10-18 20:26:44 -0700467{
468 unsigned long address;
469 pte_t *pte;
470 spinlock_t *ptl;
471
472 address = vma_address(page, vma);
473 if (address == -EFAULT) /* out of vma range */
474 return 0;
475 pte = page_check_address(page, vma->vm_mm, address, &ptl, 1);
476 if (!pte) /* the page is not in this mm */
477 return 0;
478 pte_unmap_unlock(pte, ptl);
479
480 return 1;
481}
482
Nikita Danilov81b40822005-05-01 08:58:36 -0700483/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700484 * Subfunctions of page_referenced: page_referenced_one called
485 * repeatedly from either page_referenced_anon or page_referenced_file.
486 */
Hugh Dickins5ad64682009-12-14 17:59:24 -0800487int page_referenced_one(struct page *page, struct vm_area_struct *vma,
488 unsigned long address, unsigned int *mapcount,
489 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700490{
491 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700492 pte_t *pte;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700493 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700494 int referenced = 0;
495
Nick Piggin479db0b2008-08-20 14:09:18 -0700496 pte = page_check_address(page, mm, address, &ptl, 0);
Hugh Dickinsc0718802005-10-29 18:16:31 -0700497 if (!pte)
498 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700499
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) {
Hugh Dickins5a9bbdc2008-02-04 22:29:23 -0800506 *mapcount = 1; /* break early from loop */
Minchan Kim03ef83a2009-08-26 14:29:23 -0700507 *vm_flags |= VM_LOCKED;
Nick Pigginb291f002008-10-18 20:26:44 -0700508 goto out_unmap;
509 }
510
Johannes Weiner4917e5d2009-01-06 14:39:17 -0800511 if (ptep_clear_flush_young_notify(vma, address, pte)) {
512 /*
513 * Don't treat a reference through a sequentially read
514 * mapping as such. If the page has been used in
515 * another mapping, we will catch it; if this other
516 * mapping is already gone, the unmap path will have
517 * set PG_referenced or activated the page.
518 */
519 if (likely(!VM_SequentialReadHint(vma)))
520 referenced++;
521 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700522
Hugh Dickinsc0718802005-10-29 18:16:31 -0700523 /* Pretend the page is referenced if the task has the
524 swap token and is in the middle of a page fault. */
Rik van Rielf7b7fd82005-11-28 13:44:07 -0800525 if (mm != current->mm && has_swap_token(mm) &&
Hugh Dickinsc0718802005-10-29 18:16:31 -0700526 rwsem_is_locked(&mm->mmap_sem))
527 referenced++;
528
Nick Pigginb291f002008-10-18 20:26:44 -0700529out_unmap:
Hugh Dickinsc0718802005-10-29 18:16:31 -0700530 (*mapcount)--;
531 pte_unmap_unlock(pte, ptl);
Huang Shijie273f0472009-12-14 17:58:51 -0800532
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700533 if (referenced)
534 *vm_flags |= vma->vm_flags;
Huang Shijie273f0472009-12-14 17:58:51 -0800535out:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700536 return referenced;
537}
538
Balbir Singhbed71612008-02-07 00:14:01 -0800539static int page_referenced_anon(struct page *page,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700540 struct mem_cgroup *mem_cont,
541 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700542{
543 unsigned int mapcount;
544 struct anon_vma *anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -0800545 struct anon_vma_chain *avc;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700546 int referenced = 0;
547
548 anon_vma = page_lock_anon_vma(page);
549 if (!anon_vma)
550 return referenced;
551
552 mapcount = page_mapcount(page);
Rik van Riel5beb4932010-03-05 13:42:07 -0800553 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
554 struct vm_area_struct *vma = avc->vma;
Hugh Dickins1cb17292009-12-14 17:59:01 -0800555 unsigned long address = vma_address(page, vma);
556 if (address == -EFAULT)
557 continue;
Balbir Singhbed71612008-02-07 00:14:01 -0800558 /*
559 * If we are reclaiming on behalf of a cgroup, skip
560 * counting on behalf of references from different
561 * cgroups
562 */
Hugh Dickinsbd845e32008-03-04 14:29:01 -0800563 if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
Balbir Singhbed71612008-02-07 00:14:01 -0800564 continue;
Hugh Dickins1cb17292009-12-14 17:59:01 -0800565 referenced += page_referenced_one(page, vma, address,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700566 &mapcount, vm_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700567 if (!mapcount)
568 break;
569 }
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800570
571 page_unlock_anon_vma(anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700572 return referenced;
573}
574
575/**
576 * page_referenced_file - referenced check for object-based rmap
577 * @page: the page we're checking references on.
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700578 * @mem_cont: target memory controller
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700579 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
Linus Torvalds1da177e2005-04-16 15:20:36 -0700580 *
581 * For an object-based mapped page, find all the places it is mapped and
582 * check/clear the referenced flag. This is done by following the page->mapping
583 * pointer, then walking the chain of vmas it holds. It returns the number
584 * of references it found.
585 *
586 * This function is only called from page_referenced for object-based pages.
587 */
Balbir Singhbed71612008-02-07 00:14:01 -0800588static int page_referenced_file(struct page *page,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700589 struct mem_cgroup *mem_cont,
590 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700591{
592 unsigned int mapcount;
593 struct address_space *mapping = page->mapping;
594 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
595 struct vm_area_struct *vma;
596 struct prio_tree_iter iter;
597 int referenced = 0;
598
599 /*
600 * The caller's checks on page->mapping and !PageAnon have made
601 * sure that this is a file page: the check for page->mapping
602 * excludes the case just before it gets set on an anon page.
603 */
604 BUG_ON(PageAnon(page));
605
606 /*
607 * The page lock not only makes sure that page->mapping cannot
608 * suddenly be NULLified by truncation, it makes sure that the
609 * structure at mapping cannot be freed and reused yet,
610 * so we can safely take mapping->i_mmap_lock.
611 */
612 BUG_ON(!PageLocked(page));
613
614 spin_lock(&mapping->i_mmap_lock);
615
616 /*
617 * i_mmap_lock does not stabilize mapcount at all, but mapcount
618 * is more likely to be accurate if we note it after spinning.
619 */
620 mapcount = page_mapcount(page);
621
622 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
Hugh Dickins1cb17292009-12-14 17:59:01 -0800623 unsigned long address = vma_address(page, vma);
624 if (address == -EFAULT)
625 continue;
Balbir Singhbed71612008-02-07 00:14:01 -0800626 /*
627 * If we are reclaiming on behalf of a cgroup, skip
628 * counting on behalf of references from different
629 * cgroups
630 */
Hugh Dickinsbd845e32008-03-04 14:29:01 -0800631 if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
Balbir Singhbed71612008-02-07 00:14:01 -0800632 continue;
Hugh Dickins1cb17292009-12-14 17:59:01 -0800633 referenced += page_referenced_one(page, vma, address,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700634 &mapcount, vm_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700635 if (!mapcount)
636 break;
637 }
638
639 spin_unlock(&mapping->i_mmap_lock);
640 return referenced;
641}
642
643/**
644 * page_referenced - test if the page was referenced
645 * @page: the page to test
646 * @is_locked: caller holds lock on the page
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700647 * @mem_cont: target memory controller
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700648 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
Linus Torvalds1da177e2005-04-16 15:20:36 -0700649 *
650 * Quick test_and_clear_referenced for all mappings to a page,
651 * returns the number of ptes which referenced the page.
652 */
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700653int page_referenced(struct page *page,
654 int is_locked,
655 struct mem_cgroup *mem_cont,
656 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700657{
658 int referenced = 0;
Hugh Dickins5ad64682009-12-14 17:59:24 -0800659 int we_locked = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700660
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700661 *vm_flags = 0;
Hugh Dickins3ca7b3c2009-12-14 17:58:57 -0800662 if (page_mapped(page) && page_rmapping(page)) {
Hugh Dickins5ad64682009-12-14 17:59:24 -0800663 if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
664 we_locked = trylock_page(page);
665 if (!we_locked) {
666 referenced++;
667 goto out;
668 }
669 }
670 if (unlikely(PageKsm(page)))
671 referenced += page_referenced_ksm(page, mem_cont,
672 vm_flags);
673 else if (PageAnon(page))
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700674 referenced += page_referenced_anon(page, mem_cont,
675 vm_flags);
Hugh Dickins5ad64682009-12-14 17:59:24 -0800676 else if (page->mapping)
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700677 referenced += page_referenced_file(page, mem_cont,
678 vm_flags);
Hugh Dickins5ad64682009-12-14 17:59:24 -0800679 if (we_locked)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700680 unlock_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700681 }
Hugh Dickins5ad64682009-12-14 17:59:24 -0800682out:
Christian Borntraeger5b7baf02008-03-25 18:47:12 +0100683 if (page_test_and_clear_young(page))
684 referenced++;
685
Linus Torvalds1da177e2005-04-16 15:20:36 -0700686 return referenced;
687}
688
Hugh Dickins1cb17292009-12-14 17:59:01 -0800689static int page_mkclean_one(struct page *page, struct vm_area_struct *vma,
690 unsigned long address)
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700691{
692 struct mm_struct *mm = vma->vm_mm;
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100693 pte_t *pte;
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700694 spinlock_t *ptl;
695 int ret = 0;
696
Nick Piggin479db0b2008-08-20 14:09:18 -0700697 pte = page_check_address(page, mm, address, &ptl, 1);
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700698 if (!pte)
699 goto out;
700
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100701 if (pte_dirty(*pte) || pte_write(*pte)) {
702 pte_t entry;
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700703
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100704 flush_cache_page(vma, address, pte_pfn(*pte));
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700705 entry = ptep_clear_flush_notify(vma, address, pte);
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100706 entry = pte_wrprotect(entry);
707 entry = pte_mkclean(entry);
Al Virod6e88e62006-12-29 16:48:35 -0800708 set_pte_at(mm, address, pte, entry);
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100709 ret = 1;
710 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700711
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700712 pte_unmap_unlock(pte, ptl);
713out:
714 return ret;
715}
716
717static int page_mkclean_file(struct address_space *mapping, struct page *page)
718{
719 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
720 struct vm_area_struct *vma;
721 struct prio_tree_iter iter;
722 int ret = 0;
723
724 BUG_ON(PageAnon(page));
725
726 spin_lock(&mapping->i_mmap_lock);
727 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
Hugh Dickins1cb17292009-12-14 17:59:01 -0800728 if (vma->vm_flags & VM_SHARED) {
729 unsigned long address = vma_address(page, vma);
730 if (address == -EFAULT)
731 continue;
732 ret += page_mkclean_one(page, vma, address);
733 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700734 }
735 spin_unlock(&mapping->i_mmap_lock);
736 return ret;
737}
738
739int page_mkclean(struct page *page)
740{
741 int ret = 0;
742
743 BUG_ON(!PageLocked(page));
744
745 if (page_mapped(page)) {
746 struct address_space *mapping = page_mapping(page);
Christian Borntraegerce7e9fa2007-11-20 11:13:36 +0100747 if (mapping) {
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700748 ret = page_mkclean_file(mapping, page);
Christian Borntraegerce7e9fa2007-11-20 11:13:36 +0100749 if (page_test_dirty(page)) {
Martin Schwidefskye2b8d7a2010-10-25 16:10:14 +0200750 page_clear_dirty(page, 1);
Christian Borntraegerce7e9fa2007-11-20 11:13:36 +0100751 ret = 1;
752 }
Martin Schwidefsky6c210482007-04-27 16:01:57 +0200753 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700754 }
755
756 return ret;
757}
Jaya Kumar60b59be2007-05-08 00:37:37 -0700758EXPORT_SYMBOL_GPL(page_mkclean);
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700759
Linus Torvalds1da177e2005-04-16 15:20:36 -0700760/**
Rik van Rielc44b6742010-03-05 13:42:09 -0800761 * page_move_anon_rmap - move a page to our anon_vma
762 * @page: the page to move to our anon_vma
763 * @vma: the vma the page belongs to
764 * @address: the user virtual address mapped
765 *
766 * When a page belongs exclusively to one process after a COW event,
767 * that page can be moved into the anon_vma that belongs to just that
768 * process, so the rmap code will not search the parent or sibling
769 * processes.
770 */
771void page_move_anon_rmap(struct page *page,
772 struct vm_area_struct *vma, unsigned long address)
773{
774 struct anon_vma *anon_vma = vma->anon_vma;
775
776 VM_BUG_ON(!PageLocked(page));
777 VM_BUG_ON(!anon_vma);
778 VM_BUG_ON(page->index != linear_page_index(vma, address));
779
780 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
781 page->mapping = (struct address_space *) anon_vma;
782}
783
784/**
Andi Kleen4e1c1972010-09-22 12:43:56 +0200785 * __page_set_anon_rmap - set up new anonymous rmap
786 * @page: Page to add to rmap
787 * @vma: VM area to add page to.
788 * @address: User virtual address of the mapping
Rik van Riele8a03fe2010-04-14 17:59:28 -0400789 * @exclusive: the page is exclusively owned by the current process
Nick Piggin9617d952006-01-06 00:11:12 -0800790 */
791static void __page_set_anon_rmap(struct page *page,
Rik van Riele8a03fe2010-04-14 17:59:28 -0400792 struct vm_area_struct *vma, unsigned long address, int exclusive)
Nick Piggin9617d952006-01-06 00:11:12 -0800793{
Rik van Riele8a03fe2010-04-14 17:59:28 -0400794 struct anon_vma *anon_vma = vma->anon_vma;
Nick Piggin9617d952006-01-06 00:11:12 -0800795
Rik van Riele8a03fe2010-04-14 17:59:28 -0400796 BUG_ON(!anon_vma);
Linus Torvaldsea900022010-04-12 12:44:29 -0700797
Andi Kleen4e1c1972010-09-22 12:43:56 +0200798 if (PageAnon(page))
799 return;
800
Linus Torvaldsea900022010-04-12 12:44:29 -0700801 /*
Rik van Riele8a03fe2010-04-14 17:59:28 -0400802 * If the page isn't exclusively mapped into this vma,
803 * we must use the _oldest_ possible anon_vma for the
804 * page mapping!
Linus Torvaldsea900022010-04-12 12:44:29 -0700805 */
Andi Kleen4e1c1972010-09-22 12:43:56 +0200806 if (!exclusive)
Andrea Arcangeli288468c2010-08-09 17:19:09 -0700807 anon_vma = anon_vma->root;
Linus Torvaldsea900022010-04-12 12:44:29 -0700808
Nick Piggin9617d952006-01-06 00:11:12 -0800809 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
810 page->mapping = (struct address_space *) anon_vma;
Nick Piggin9617d952006-01-06 00:11:12 -0800811 page->index = linear_page_index(vma, address);
Nick Piggin9617d952006-01-06 00:11:12 -0800812}
813
814/**
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700815 * __page_check_anon_rmap - sanity check anonymous rmap addition
Nick Pigginc97a9e12007-05-16 22:11:21 -0700816 * @page: the page to add the mapping to
817 * @vma: the vm area in which the mapping is added
818 * @address: the user virtual address mapped
819 */
820static void __page_check_anon_rmap(struct page *page,
821 struct vm_area_struct *vma, unsigned long address)
822{
823#ifdef CONFIG_DEBUG_VM
824 /*
825 * The page's anon-rmap details (mapping and index) are guaranteed to
826 * be set up correctly at this point.
827 *
828 * We have exclusion against page_add_anon_rmap because the caller
829 * always holds the page locked, except if called from page_dup_rmap,
830 * in which case the page is already known to be setup.
831 *
832 * We have exclusion against page_add_new_anon_rmap because those pages
833 * are initially only visible via the pagetables, and the pte is locked
834 * over the call to page_add_new_anon_rmap.
835 */
Andrea Arcangeli44ab57a2010-08-09 17:19:10 -0700836 BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root);
Nick Pigginc97a9e12007-05-16 22:11:21 -0700837 BUG_ON(page->index != linear_page_index(vma, address));
838#endif
839}
840
841/**
Linus Torvalds1da177e2005-04-16 15:20:36 -0700842 * page_add_anon_rmap - add pte mapping to an anonymous page
843 * @page: the page to add the mapping to
844 * @vma: the vm area in which the mapping is added
845 * @address: the user virtual address mapped
846 *
Hugh Dickins5ad64682009-12-14 17:59:24 -0800847 * The caller needs to hold the pte lock, and the page must be locked in
Hugh Dickins80e14822009-12-14 17:59:29 -0800848 * the anon_vma case: to serialize mapping,index checking after setting,
849 * and to ensure that PageAnon is not being upgraded racily to PageKsm
850 * (but PageKsm is never downgraded to PageAnon).
Linus Torvalds1da177e2005-04-16 15:20:36 -0700851 */
852void page_add_anon_rmap(struct page *page,
853 struct vm_area_struct *vma, unsigned long address)
854{
Rik van Rielad8c2ee2010-08-09 17:19:48 -0700855 do_page_add_anon_rmap(page, vma, address, 0);
856}
857
858/*
859 * Special version of the above for do_swap_page, which often runs
860 * into pages that are exclusively owned by the current process.
861 * Everybody else should continue to use page_add_anon_rmap above.
862 */
863void do_page_add_anon_rmap(struct page *page,
864 struct vm_area_struct *vma, unsigned long address, int exclusive)
865{
Hugh Dickins5ad64682009-12-14 17:59:24 -0800866 int first = atomic_inc_and_test(&page->_mapcount);
867 if (first)
868 __inc_zone_page_state(page, NR_ANON_PAGES);
869 if (unlikely(PageKsm(page)))
870 return;
871
Nick Pigginc97a9e12007-05-16 22:11:21 -0700872 VM_BUG_ON(!PageLocked(page));
873 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
Hugh Dickins5ad64682009-12-14 17:59:24 -0800874 if (first)
Rik van Rielad8c2ee2010-08-09 17:19:48 -0700875 __page_set_anon_rmap(page, vma, address, exclusive);
KAMEZAWA Hiroyuki69029cd2008-07-25 01:47:14 -0700876 else
Nick Pigginc97a9e12007-05-16 22:11:21 -0700877 __page_check_anon_rmap(page, vma, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700878}
879
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700880/**
Nick Piggin9617d952006-01-06 00:11:12 -0800881 * page_add_new_anon_rmap - add pte mapping to a new 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 *
886 * Same as page_add_anon_rmap but must only be called on *new* pages.
887 * This means the inc-and-test can be bypassed.
Nick Pigginc97a9e12007-05-16 22:11:21 -0700888 * Page does not have to be locked.
Nick Piggin9617d952006-01-06 00:11:12 -0800889 */
890void page_add_new_anon_rmap(struct page *page,
891 struct vm_area_struct *vma, unsigned long address)
892{
Hugh Dickinsb5934c52009-01-06 14:39:25 -0800893 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
Hugh Dickinscbf84b72009-01-06 14:39:27 -0800894 SetPageSwapBacked(page);
895 atomic_set(&page->_mapcount, 0); /* increment count (starts at -1) */
Hugh Dickins5ad64682009-12-14 17:59:24 -0800896 __inc_zone_page_state(page, NR_ANON_PAGES);
Rik van Riele8a03fe2010-04-14 17:59:28 -0400897 __page_set_anon_rmap(page, vma, address, 1);
Hugh Dickinsb5934c52009-01-06 14:39:25 -0800898 if (page_evictable(page, vma))
Hugh Dickinscbf84b72009-01-06 14:39:27 -0800899 lru_cache_add_lru(page, LRU_ACTIVE_ANON);
Hugh Dickinsb5934c52009-01-06 14:39:25 -0800900 else
901 add_page_to_unevictable_list(page);
Nick Piggin9617d952006-01-06 00:11:12 -0800902}
903
Linus Torvalds1da177e2005-04-16 15:20:36 -0700904/**
905 * page_add_file_rmap - add pte mapping to a file page
906 * @page: the page to add the mapping to
907 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -0700908 * The caller needs to hold the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700909 */
910void page_add_file_rmap(struct page *page)
911{
Balbir Singhd69b0422009-06-17 16:26:34 -0700912 if (atomic_inc_and_test(&page->_mapcount)) {
Christoph Lameter65ba55f2006-06-30 01:55:34 -0700913 __inc_zone_page_state(page, NR_FILE_MAPPED);
KAMEZAWA Hiroyukid8046582009-12-15 16:47:09 -0800914 mem_cgroup_update_file_mapped(page, 1);
Balbir Singhd69b0422009-06-17 16:26:34 -0700915 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700916}
917
918/**
919 * page_remove_rmap - take down pte mapping from a page
920 * @page: page to remove mapping from
921 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -0700922 * The caller needs to hold the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700923 */
Hugh Dickinsedc315f2009-01-06 14:40:11 -0800924void page_remove_rmap(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700925{
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -0700926 /* page still mapped by someone else? */
927 if (!atomic_add_negative(-1, &page->_mapcount))
928 return;
929
930 /*
931 * Now that the last pte has gone, s390 must transfer dirty
932 * flag from storage key to struct page. We can usually skip
933 * this if the page is anon, so about to be freed; but perhaps
934 * not if it's in swapcache - there might be another pte slot
935 * containing the swap entry, but page not yet written to swap.
936 */
937 if ((!PageAnon(page) || PageSwapCache(page)) && page_test_dirty(page)) {
Martin Schwidefskye2b8d7a2010-10-25 16:10:14 +0200938 page_clear_dirty(page, 1);
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -0700939 set_page_dirty(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700940 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900941 /*
942 * Hugepages are not counted in NR_ANON_PAGES nor NR_FILE_MAPPED
943 * and not charged by memcg for now.
944 */
945 if (unlikely(PageHuge(page)))
946 return;
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -0700947 if (PageAnon(page)) {
948 mem_cgroup_uncharge_page(page);
949 __dec_zone_page_state(page, NR_ANON_PAGES);
950 } else {
951 __dec_zone_page_state(page, NR_FILE_MAPPED);
KAMEZAWA Hiroyukid8046582009-12-15 16:47:09 -0800952 mem_cgroup_update_file_mapped(page, -1);
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -0700953 }
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -0700954 /*
955 * It would be tidy to reset the PageAnon mapping here,
956 * but that might overwrite a racing page_add_anon_rmap
957 * which increments mapcount after us but sets mapping
958 * before us: so leave the reset to free_hot_cold_page,
959 * and remember that it's only reliable while mapped.
960 * Leaving it set also helps swapoff to reinstate ptes
961 * faster for those pages still in swapcache.
962 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700963}
964
965/*
966 * Subfunctions of try_to_unmap: try_to_unmap_one called
967 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
968 */
Hugh Dickins5ad64682009-12-14 17:59:24 -0800969int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
970 unsigned long address, enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700971{
972 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700973 pte_t *pte;
974 pte_t pteval;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700975 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700976 int ret = SWAP_AGAIN;
977
Nick Piggin479db0b2008-08-20 14:09:18 -0700978 pte = page_check_address(page, mm, address, &ptl, 0);
Hugh Dickinsc0718802005-10-29 18:16:31 -0700979 if (!pte)
Nikita Danilov81b40822005-05-01 08:58:36 -0700980 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700981
982 /*
983 * If the page is mlock()d, we cannot swap it out.
984 * If it's recently referenced (perhaps page_referenced
985 * skipped over this mm) then we should reactivate it.
986 */
Andi Kleen14fa31b2009-09-16 11:50:10 +0200987 if (!(flags & TTU_IGNORE_MLOCK)) {
KOSAKI Motohirocaed0f42009-12-14 17:59:45 -0800988 if (vma->vm_flags & VM_LOCKED)
989 goto out_mlock;
990
Hugh Dickinsaf8e3352009-12-14 17:58:59 -0800991 if (TTU_ACTION(flags) == TTU_MUNLOCK)
Hugh Dickins53f79ac2009-12-14 17:58:58 -0800992 goto out_unmap;
Andi Kleen14fa31b2009-09-16 11:50:10 +0200993 }
994 if (!(flags & TTU_IGNORE_ACCESS)) {
Nick Pigginb291f002008-10-18 20:26:44 -0700995 if (ptep_clear_flush_young_notify(vma, address, pte)) {
996 ret = SWAP_FAIL;
997 goto out_unmap;
998 }
999 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001000
Linus Torvalds1da177e2005-04-16 15:20:36 -07001001 /* Nuke the page table entry. */
1002 flush_cache_page(vma, address, page_to_pfn(page));
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001003 pteval = ptep_clear_flush_notify(vma, address, pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001004
1005 /* Move the dirty bit to the physical page now the pte is gone. */
1006 if (pte_dirty(pteval))
1007 set_page_dirty(page);
1008
Hugh Dickins365e9c872005-10-29 18:16:18 -07001009 /* Update high watermark before we lower rss */
1010 update_hiwater_rss(mm);
1011
Andi Kleen888b9f72009-09-16 11:50:11 +02001012 if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
1013 if (PageAnon(page))
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001014 dec_mm_counter(mm, MM_ANONPAGES);
Andi Kleen888b9f72009-09-16 11:50:11 +02001015 else
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001016 dec_mm_counter(mm, MM_FILEPAGES);
Andi Kleen888b9f72009-09-16 11:50:11 +02001017 set_pte_at(mm, address, pte,
1018 swp_entry_to_pte(make_hwpoison_entry(page)));
1019 } else if (PageAnon(page)) {
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07001020 swp_entry_t entry = { .val = page_private(page) };
Christoph Lameter06972122006-06-23 02:03:35 -07001021
1022 if (PageSwapCache(page)) {
1023 /*
1024 * Store the swap location in the pte.
1025 * See handle_pte_fault() ...
1026 */
Hugh Dickins570a3352009-12-14 17:58:46 -08001027 if (swap_duplicate(entry) < 0) {
1028 set_pte_at(mm, address, pte, pteval);
1029 ret = SWAP_FAIL;
1030 goto out_unmap;
1031 }
Christoph Lameter06972122006-06-23 02:03:35 -07001032 if (list_empty(&mm->mmlist)) {
1033 spin_lock(&mmlist_lock);
1034 if (list_empty(&mm->mmlist))
1035 list_add(&mm->mmlist, &init_mm.mmlist);
1036 spin_unlock(&mmlist_lock);
1037 }
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001038 dec_mm_counter(mm, MM_ANONPAGES);
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08001039 inc_mm_counter(mm, MM_SWAPENTS);
KOSAKI Motohiro64cdd542009-01-06 14:39:16 -08001040 } else if (PAGE_MIGRATION) {
Christoph Lameter06972122006-06-23 02:03:35 -07001041 /*
1042 * Store the pfn of the page in a special migration
1043 * pte. do_swap_page() will wait until the migration
1044 * pte is removed and then restart fault handling.
1045 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001046 BUG_ON(TTU_ACTION(flags) != TTU_MIGRATION);
Christoph Lameter06972122006-06-23 02:03:35 -07001047 entry = make_migration_entry(page, pte_write(pteval));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001048 }
1049 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
1050 BUG_ON(pte_file(*pte));
Andi Kleen14fa31b2009-09-16 11:50:10 +02001051 } else if (PAGE_MIGRATION && (TTU_ACTION(flags) == TTU_MIGRATION)) {
Christoph Lameter04e62a22006-06-23 02:03:38 -07001052 /* Establish migration entry for a file page */
1053 swp_entry_t entry;
1054 entry = make_migration_entry(page, pte_write(pteval));
1055 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
1056 } else
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001057 dec_mm_counter(mm, MM_FILEPAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001058
Hugh Dickinsedc315f2009-01-06 14:40:11 -08001059 page_remove_rmap(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001060 page_cache_release(page);
1061
1062out_unmap:
Hugh Dickinsc0718802005-10-29 18:16:31 -07001063 pte_unmap_unlock(pte, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001064out:
1065 return ret;
KOSAKI Motohirocaed0f42009-12-14 17:59:45 -08001066
1067out_mlock:
1068 pte_unmap_unlock(pte, ptl);
1069
1070
1071 /*
1072 * We need mmap_sem locking, Otherwise VM_LOCKED check makes
1073 * unstable result and race. Plus, We can't wait here because
1074 * we now hold anon_vma->lock or mapping->i_mmap_lock.
1075 * if trylock failed, the page remain in evictable lru and later
1076 * vmscan could retry to move the page to unevictable lru if the
1077 * page is actually mlocked.
1078 */
1079 if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
1080 if (vma->vm_flags & VM_LOCKED) {
1081 mlock_vma_page(page);
1082 ret = SWAP_MLOCK;
1083 }
1084 up_read(&vma->vm_mm->mmap_sem);
1085 }
1086 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001087}
1088
1089/*
1090 * objrmap doesn't work for nonlinear VMAs because the assumption that
1091 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
1092 * Consequently, given a particular page and its ->index, we cannot locate the
1093 * ptes which are mapping that page without an exhaustive linear search.
1094 *
1095 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
1096 * maps the file to which the target page belongs. The ->vm_private_data field
1097 * holds the current cursor into that scan. Successive searches will circulate
1098 * around the vma's virtual address space.
1099 *
1100 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
1101 * more scanning pressure is placed against them as well. Eventually pages
1102 * will become fully unmapped and are eligible for eviction.
1103 *
1104 * For very sparsely populated VMAs this is a little inefficient - chances are
1105 * there there won't be many ptes located within the scan cluster. In this case
1106 * maybe we could scan further - to the end of the pte page, perhaps.
Nick Pigginb291f002008-10-18 20:26:44 -07001107 *
1108 * Mlocked pages: check VM_LOCKED under mmap_sem held for read, if we can
1109 * acquire it without blocking. If vma locked, mlock the pages in the cluster,
1110 * rather than unmapping them. If we encounter the "check_page" that vmscan is
1111 * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001112 */
1113#define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
1114#define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
1115
Nick Pigginb291f002008-10-18 20:26:44 -07001116static int try_to_unmap_cluster(unsigned long cursor, unsigned int *mapcount,
1117 struct vm_area_struct *vma, struct page *check_page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001118{
1119 struct mm_struct *mm = vma->vm_mm;
1120 pgd_t *pgd;
1121 pud_t *pud;
1122 pmd_t *pmd;
Hugh Dickinsc0718802005-10-29 18:16:31 -07001123 pte_t *pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001124 pte_t pteval;
Hugh Dickinsc0718802005-10-29 18:16:31 -07001125 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001126 struct page *page;
1127 unsigned long address;
1128 unsigned long end;
Nick Pigginb291f002008-10-18 20:26:44 -07001129 int ret = SWAP_AGAIN;
1130 int locked_vma = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001131
Linus Torvalds1da177e2005-04-16 15:20:36 -07001132 address = (vma->vm_start + cursor) & CLUSTER_MASK;
1133 end = address + CLUSTER_SIZE;
1134 if (address < vma->vm_start)
1135 address = vma->vm_start;
1136 if (end > vma->vm_end)
1137 end = vma->vm_end;
1138
1139 pgd = pgd_offset(mm, address);
1140 if (!pgd_present(*pgd))
Nick Pigginb291f002008-10-18 20:26:44 -07001141 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001142
1143 pud = pud_offset(pgd, address);
1144 if (!pud_present(*pud))
Nick Pigginb291f002008-10-18 20:26:44 -07001145 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001146
1147 pmd = pmd_offset(pud, address);
1148 if (!pmd_present(*pmd))
Nick Pigginb291f002008-10-18 20:26:44 -07001149 return ret;
1150
1151 /*
Hugh Dickinsaf8e3352009-12-14 17:58:59 -08001152 * If we can acquire the mmap_sem for read, and vma is VM_LOCKED,
Nick Pigginb291f002008-10-18 20:26:44 -07001153 * keep the sem while scanning the cluster for mlocking pages.
1154 */
Hugh Dickinsaf8e3352009-12-14 17:58:59 -08001155 if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
Nick Pigginb291f002008-10-18 20:26:44 -07001156 locked_vma = (vma->vm_flags & VM_LOCKED);
1157 if (!locked_vma)
1158 up_read(&vma->vm_mm->mmap_sem); /* don't need it */
1159 }
Hugh Dickinsc0718802005-10-29 18:16:31 -07001160
1161 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001162
Hugh Dickins365e9c872005-10-29 18:16:18 -07001163 /* Update high watermark before we lower rss */
1164 update_hiwater_rss(mm);
1165
Hugh Dickinsc0718802005-10-29 18:16:31 -07001166 for (; address < end; pte++, address += PAGE_SIZE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001167 if (!pte_present(*pte))
1168 continue;
Linus Torvalds6aab3412005-11-28 14:34:23 -08001169 page = vm_normal_page(vma, address, *pte);
1170 BUG_ON(!page || PageAnon(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001171
Nick Pigginb291f002008-10-18 20:26:44 -07001172 if (locked_vma) {
1173 mlock_vma_page(page); /* no-op if already mlocked */
1174 if (page == check_page)
1175 ret = SWAP_MLOCK;
1176 continue; /* don't unmap */
1177 }
1178
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001179 if (ptep_clear_flush_young_notify(vma, address, pte))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001180 continue;
1181
1182 /* Nuke the page table entry. */
Ben Collinseca35132005-11-29 11:45:26 -08001183 flush_cache_page(vma, address, pte_pfn(*pte));
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001184 pteval = ptep_clear_flush_notify(vma, address, pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001185
1186 /* If nonlinear, store the file page offset in the pte. */
1187 if (page->index != linear_page_index(vma, address))
1188 set_pte_at(mm, address, pte, pgoff_to_pte(page->index));
1189
1190 /* Move the dirty bit to the physical page now the pte is gone. */
1191 if (pte_dirty(pteval))
1192 set_page_dirty(page);
1193
Hugh Dickinsedc315f2009-01-06 14:40:11 -08001194 page_remove_rmap(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001195 page_cache_release(page);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001196 dec_mm_counter(mm, MM_FILEPAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001197 (*mapcount)--;
1198 }
Hugh Dickinsc0718802005-10-29 18:16:31 -07001199 pte_unmap_unlock(pte - 1, ptl);
Nick Pigginb291f002008-10-18 20:26:44 -07001200 if (locked_vma)
1201 up_read(&vma->vm_mm->mmap_sem);
1202 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001203}
1204
Mel Gormana8bef8ff2010-05-24 14:32:24 -07001205static bool is_vma_temporary_stack(struct vm_area_struct *vma)
1206{
1207 int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP);
1208
1209 if (!maybe_stack)
1210 return false;
1211
1212 if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) ==
1213 VM_STACK_INCOMPLETE_SETUP)
1214 return true;
1215
1216 return false;
1217}
1218
Nick Pigginb291f002008-10-18 20:26:44 -07001219/**
1220 * try_to_unmap_anon - unmap or unlock anonymous page using the object-based
1221 * rmap method
1222 * @page: the page to unmap/unlock
Huang Shijie8051be52009-12-14 17:58:50 -08001223 * @flags: action and flags
Nick Pigginb291f002008-10-18 20:26:44 -07001224 *
1225 * Find all the mappings of a page using the mapping pointer and the vma chains
1226 * contained in the anon_vma struct it points to.
1227 *
1228 * This function is only called from try_to_unmap/try_to_munlock for
1229 * anonymous pages.
1230 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1231 * where the page was found will be held for write. So, we won't recheck
1232 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1233 * 'LOCKED.
1234 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001235static int try_to_unmap_anon(struct page *page, enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001236{
1237 struct anon_vma *anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -08001238 struct anon_vma_chain *avc;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001239 int ret = SWAP_AGAIN;
Nick Pigginb291f002008-10-18 20:26:44 -07001240
Linus Torvalds1da177e2005-04-16 15:20:36 -07001241 anon_vma = page_lock_anon_vma(page);
1242 if (!anon_vma)
1243 return ret;
1244
Rik van Riel5beb4932010-03-05 13:42:07 -08001245 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1246 struct vm_area_struct *vma = avc->vma;
Mel Gormana8bef8ff2010-05-24 14:32:24 -07001247 unsigned long address;
1248
1249 /*
1250 * During exec, a temporary VMA is setup and later moved.
1251 * The VMA is moved under the anon_vma lock but not the
1252 * page tables leading to a race where migration cannot
1253 * find the migration ptes. Rather than increasing the
1254 * locking requirements of exec(), migration skips
1255 * temporary VMAs until after exec() completes.
1256 */
1257 if (PAGE_MIGRATION && (flags & TTU_MIGRATION) &&
1258 is_vma_temporary_stack(vma))
1259 continue;
1260
1261 address = vma_address(page, vma);
Hugh Dickins1cb17292009-12-14 17:59:01 -08001262 if (address == -EFAULT)
1263 continue;
1264 ret = try_to_unmap_one(page, vma, address, flags);
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001265 if (ret != SWAP_AGAIN || !page_mapped(page))
1266 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001267 }
Oleg Nesterov34bbd702007-02-28 20:13:49 -08001268
1269 page_unlock_anon_vma(anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001270 return ret;
1271}
1272
1273/**
Nick Pigginb291f002008-10-18 20:26:44 -07001274 * try_to_unmap_file - unmap/unlock file page using the object-based rmap method
1275 * @page: the page to unmap/unlock
Andi Kleen14fa31b2009-09-16 11:50:10 +02001276 * @flags: action and flags
Linus Torvalds1da177e2005-04-16 15:20:36 -07001277 *
1278 * Find all the mappings of a page using the mapping pointer and the vma chains
1279 * contained in the address_space struct it points to.
1280 *
Nick Pigginb291f002008-10-18 20:26:44 -07001281 * This function is only called from try_to_unmap/try_to_munlock for
1282 * object-based pages.
1283 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1284 * where the page was found will be held for write. So, we won't recheck
1285 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1286 * 'LOCKED.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001287 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001288static int try_to_unmap_file(struct page *page, enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001289{
1290 struct address_space *mapping = page->mapping;
1291 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1292 struct vm_area_struct *vma;
1293 struct prio_tree_iter iter;
1294 int ret = SWAP_AGAIN;
1295 unsigned long cursor;
1296 unsigned long max_nl_cursor = 0;
1297 unsigned long max_nl_size = 0;
1298 unsigned int mapcount;
1299
1300 spin_lock(&mapping->i_mmap_lock);
1301 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
Hugh Dickins1cb17292009-12-14 17:59:01 -08001302 unsigned long address = vma_address(page, vma);
1303 if (address == -EFAULT)
1304 continue;
1305 ret = try_to_unmap_one(page, vma, address, flags);
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001306 if (ret != SWAP_AGAIN || !page_mapped(page))
1307 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001308 }
1309
1310 if (list_empty(&mapping->i_mmap_nonlinear))
1311 goto out;
1312
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001313 /*
1314 * We don't bother to try to find the munlocked page in nonlinears.
1315 * It's costly. Instead, later, page reclaim logic may call
1316 * try_to_unmap(TTU_MUNLOCK) and recover PG_mlocked lazily.
1317 */
1318 if (TTU_ACTION(flags) == TTU_MUNLOCK)
1319 goto out;
1320
Linus Torvalds1da177e2005-04-16 15:20:36 -07001321 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1322 shared.vm_set.list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001323 cursor = (unsigned long) vma->vm_private_data;
1324 if (cursor > max_nl_cursor)
1325 max_nl_cursor = cursor;
1326 cursor = vma->vm_end - vma->vm_start;
1327 if (cursor > max_nl_size)
1328 max_nl_size = cursor;
1329 }
1330
Nick Pigginb291f002008-10-18 20:26:44 -07001331 if (max_nl_size == 0) { /* all nonlinears locked or reserved ? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001332 ret = SWAP_FAIL;
1333 goto out;
1334 }
1335
1336 /*
1337 * We don't try to search for this page in the nonlinear vmas,
1338 * and page_referenced wouldn't have found it anyway. Instead
1339 * just walk the nonlinear vmas trying to age and unmap some.
1340 * The mapcount of the page we came in with is irrelevant,
1341 * but even so use it as a guide to how hard we should try?
1342 */
1343 mapcount = page_mapcount(page);
1344 if (!mapcount)
1345 goto out;
1346 cond_resched_lock(&mapping->i_mmap_lock);
1347
1348 max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
1349 if (max_nl_cursor == 0)
1350 max_nl_cursor = CLUSTER_SIZE;
1351
1352 do {
1353 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1354 shared.vm_set.list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001355 cursor = (unsigned long) vma->vm_private_data;
Hugh Dickins839b9682005-09-03 15:54:43 -07001356 while ( cursor < max_nl_cursor &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07001357 cursor < vma->vm_end - vma->vm_start) {
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001358 if (try_to_unmap_cluster(cursor, &mapcount,
1359 vma, page) == SWAP_MLOCK)
1360 ret = SWAP_MLOCK;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001361 cursor += CLUSTER_SIZE;
1362 vma->vm_private_data = (void *) cursor;
1363 if ((int)mapcount <= 0)
1364 goto out;
1365 }
1366 vma->vm_private_data = (void *) max_nl_cursor;
1367 }
1368 cond_resched_lock(&mapping->i_mmap_lock);
1369 max_nl_cursor += CLUSTER_SIZE;
1370 } while (max_nl_cursor <= max_nl_size);
1371
1372 /*
1373 * Don't loop forever (perhaps all the remaining pages are
1374 * in locked vmas). Reset cursor on all unreserved nonlinear
1375 * vmas, now forgetting on which ones it had fallen behind.
1376 */
Hugh Dickins101d2be2005-11-21 21:32:16 -08001377 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1378 vma->vm_private_data = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001379out:
1380 spin_unlock(&mapping->i_mmap_lock);
1381 return ret;
1382}
1383
1384/**
1385 * try_to_unmap - try to remove all page table mappings to a page
1386 * @page: the page to get unmapped
Andi Kleen14fa31b2009-09-16 11:50:10 +02001387 * @flags: action and flags
Linus Torvalds1da177e2005-04-16 15:20:36 -07001388 *
1389 * Tries to remove all the page table entries which are mapping this
1390 * page, used in the pageout path. Caller must hold the page lock.
1391 * Return values are:
1392 *
1393 * SWAP_SUCCESS - we succeeded in removing all mappings
1394 * SWAP_AGAIN - we missed a mapping, try again later
1395 * SWAP_FAIL - the page is unswappable
Nick Pigginb291f002008-10-18 20:26:44 -07001396 * SWAP_MLOCK - page is mlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001397 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001398int try_to_unmap(struct page *page, enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001399{
1400 int ret;
1401
Linus Torvalds1da177e2005-04-16 15:20:36 -07001402 BUG_ON(!PageLocked(page));
1403
Hugh Dickins5ad64682009-12-14 17:59:24 -08001404 if (unlikely(PageKsm(page)))
1405 ret = try_to_unmap_ksm(page, flags);
1406 else if (PageAnon(page))
Andi Kleen14fa31b2009-09-16 11:50:10 +02001407 ret = try_to_unmap_anon(page, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001408 else
Andi Kleen14fa31b2009-09-16 11:50:10 +02001409 ret = try_to_unmap_file(page, flags);
Nick Pigginb291f002008-10-18 20:26:44 -07001410 if (ret != SWAP_MLOCK && !page_mapped(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001411 ret = SWAP_SUCCESS;
1412 return ret;
1413}
Nikita Danilov81b40822005-05-01 08:58:36 -07001414
Nick Pigginb291f002008-10-18 20:26:44 -07001415/**
1416 * try_to_munlock - try to munlock a page
1417 * @page: the page to be munlocked
1418 *
1419 * Called from munlock code. Checks all of the VMAs mapping the page
1420 * to make sure nobody else has this page mlocked. The page will be
1421 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1422 *
1423 * Return values are:
1424 *
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001425 * SWAP_AGAIN - no vma is holding page mlocked, or,
Nick Pigginb291f002008-10-18 20:26:44 -07001426 * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem
Hugh Dickins5ad64682009-12-14 17:59:24 -08001427 * SWAP_FAIL - page cannot be located at present
Nick Pigginb291f002008-10-18 20:26:44 -07001428 * SWAP_MLOCK - page is now mlocked.
1429 */
1430int try_to_munlock(struct page *page)
1431{
1432 VM_BUG_ON(!PageLocked(page) || PageLRU(page));
1433
Hugh Dickins5ad64682009-12-14 17:59:24 -08001434 if (unlikely(PageKsm(page)))
1435 return try_to_unmap_ksm(page, TTU_MUNLOCK);
1436 else if (PageAnon(page))
Andi Kleen14fa31b2009-09-16 11:50:10 +02001437 return try_to_unmap_anon(page, TTU_MUNLOCK);
Nick Pigginb291f002008-10-18 20:26:44 -07001438 else
Andi Kleen14fa31b2009-09-16 11:50:10 +02001439 return try_to_unmap_file(page, TTU_MUNLOCK);
Nick Pigginb291f002008-10-18 20:26:44 -07001440}
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001441
Rik van Riel76545062010-08-09 17:18:41 -07001442#if defined(CONFIG_KSM) || defined(CONFIG_MIGRATION)
1443/*
1444 * Drop an anon_vma refcount, freeing the anon_vma and anon_vma->root
1445 * if necessary. Be careful to do all the tests under the lock. Once
1446 * we know we are the last user, nobody else can get a reference and we
1447 * can do the freeing without the lock.
1448 */
1449void drop_anon_vma(struct anon_vma *anon_vma)
1450{
Andrea Arcangeli44ab57a2010-08-09 17:19:10 -07001451 BUG_ON(atomic_read(&anon_vma->external_refcount) <= 0);
Rik van Riel76545062010-08-09 17:18:41 -07001452 if (atomic_dec_and_lock(&anon_vma->external_refcount, &anon_vma->root->lock)) {
1453 struct anon_vma *root = anon_vma->root;
1454 int empty = list_empty(&anon_vma->head);
1455 int last_root_user = 0;
1456 int root_empty = 0;
1457
1458 /*
1459 * The refcount on a non-root anon_vma got dropped. Drop
1460 * the refcount on the root and check if we need to free it.
1461 */
1462 if (empty && anon_vma != root) {
Andrea Arcangeli44ab57a2010-08-09 17:19:10 -07001463 BUG_ON(atomic_read(&root->external_refcount) <= 0);
Rik van Riel76545062010-08-09 17:18:41 -07001464 last_root_user = atomic_dec_and_test(&root->external_refcount);
1465 root_empty = list_empty(&root->head);
1466 }
1467 anon_vma_unlock(anon_vma);
1468
1469 if (empty) {
1470 anon_vma_free(anon_vma);
1471 if (root_empty && last_root_user)
1472 anon_vma_free(root);
1473 }
1474 }
1475}
1476#endif
1477
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001478#ifdef CONFIG_MIGRATION
1479/*
1480 * rmap_walk() and its helpers rmap_walk_anon() and rmap_walk_file():
1481 * Called by migrate.c to remove migration ptes, but might be used more later.
1482 */
1483static int rmap_walk_anon(struct page *page, int (*rmap_one)(struct page *,
1484 struct vm_area_struct *, unsigned long, void *), void *arg)
1485{
1486 struct anon_vma *anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -08001487 struct anon_vma_chain *avc;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001488 int ret = SWAP_AGAIN;
1489
1490 /*
1491 * Note: remove_migration_ptes() cannot use page_lock_anon_vma()
1492 * because that depends on page_mapped(); but not all its usages
Mel Gorman3f6c8272010-05-24 14:32:17 -07001493 * are holding mmap_sem. Users without mmap_sem are required to
1494 * take a reference count to prevent the anon_vma disappearing
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001495 */
1496 anon_vma = page_anon_vma(page);
1497 if (!anon_vma)
1498 return ret;
Rik van Rielcba48b92010-08-09 17:18:38 -07001499 anon_vma_lock(anon_vma);
Rik van Riel5beb4932010-03-05 13:42:07 -08001500 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1501 struct vm_area_struct *vma = avc->vma;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001502 unsigned long address = vma_address(page, vma);
1503 if (address == -EFAULT)
1504 continue;
1505 ret = rmap_one(page, vma, address, arg);
1506 if (ret != SWAP_AGAIN)
1507 break;
1508 }
Rik van Rielcba48b92010-08-09 17:18:38 -07001509 anon_vma_unlock(anon_vma);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001510 return ret;
1511}
1512
1513static int rmap_walk_file(struct page *page, int (*rmap_one)(struct page *,
1514 struct vm_area_struct *, unsigned long, void *), void *arg)
1515{
1516 struct address_space *mapping = page->mapping;
1517 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1518 struct vm_area_struct *vma;
1519 struct prio_tree_iter iter;
1520 int ret = SWAP_AGAIN;
1521
1522 if (!mapping)
1523 return ret;
1524 spin_lock(&mapping->i_mmap_lock);
1525 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1526 unsigned long address = vma_address(page, vma);
1527 if (address == -EFAULT)
1528 continue;
1529 ret = rmap_one(page, vma, address, arg);
1530 if (ret != SWAP_AGAIN)
1531 break;
1532 }
1533 /*
1534 * No nonlinear handling: being always shared, nonlinear vmas
1535 * never contain migration ptes. Decide what to do about this
1536 * limitation to linear when we need rmap_walk() on nonlinear.
1537 */
1538 spin_unlock(&mapping->i_mmap_lock);
1539 return ret;
1540}
1541
1542int rmap_walk(struct page *page, int (*rmap_one)(struct page *,
1543 struct vm_area_struct *, unsigned long, void *), void *arg)
1544{
1545 VM_BUG_ON(!PageLocked(page));
1546
1547 if (unlikely(PageKsm(page)))
1548 return rmap_walk_ksm(page, rmap_one, arg);
1549 else if (PageAnon(page))
1550 return rmap_walk_anon(page, rmap_one, arg);
1551 else
1552 return rmap_walk_file(page, rmap_one, arg);
1553}
1554#endif /* CONFIG_MIGRATION */
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001555
Naoya Horiguchie3390f62010-06-15 13:18:13 +09001556#ifdef CONFIG_HUGETLB_PAGE
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001557/*
1558 * The following three functions are for anonymous (private mapped) hugepages.
1559 * Unlike common anonymous pages, anonymous hugepages have no accounting code
1560 * and no lru code, because we handle hugepages differently from common pages.
1561 */
1562static void __hugepage_set_anon_rmap(struct page *page,
1563 struct vm_area_struct *vma, unsigned long address, int exclusive)
1564{
1565 struct anon_vma *anon_vma = vma->anon_vma;
Naoya Horiguchi433abed2010-09-10 13:23:03 +09001566
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001567 BUG_ON(!anon_vma);
Naoya Horiguchi433abed2010-09-10 13:23:03 +09001568
1569 if (PageAnon(page))
1570 return;
1571 if (!exclusive)
1572 anon_vma = anon_vma->root;
1573
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001574 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1575 page->mapping = (struct address_space *) anon_vma;
1576 page->index = linear_page_index(vma, address);
1577}
1578
1579void hugepage_add_anon_rmap(struct page *page,
1580 struct vm_area_struct *vma, unsigned long address)
1581{
1582 struct anon_vma *anon_vma = vma->anon_vma;
1583 int first;
Naoya Horiguchia850ea32010-09-10 13:23:06 +09001584
1585 BUG_ON(!PageLocked(page));
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001586 BUG_ON(!anon_vma);
1587 BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1588 first = atomic_inc_and_test(&page->_mapcount);
1589 if (first)
1590 __hugepage_set_anon_rmap(page, vma, address, 0);
1591}
1592
1593void hugepage_add_new_anon_rmap(struct page *page,
1594 struct vm_area_struct *vma, unsigned long address)
1595{
1596 BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1597 atomic_set(&page->_mapcount, 0);
1598 __hugepage_set_anon_rmap(page, vma, address, 1);
1599}
Naoya Horiguchie3390f62010-06-15 13:18:13 +09001600#endif /* CONFIG_HUGETLB_PAGE */