blob: 87b9e8ad450962afa1159b763f0aa0a977ab9a88 [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
83void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
84{
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 Riel5c341ee12010-08-09 17:18:39 -0700136 /*
137 * This VMA had no anon_vma yet. This anon_vma is
138 * the root of any anon_vma tree that might form.
139 */
140 anon_vma->root = anon_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700141 }
142
Rik van Rielcba48b92010-08-09 17:18:38 -0700143 anon_vma_lock(anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700144 /* page_table_lock to protect against threads */
145 spin_lock(&mm->page_table_lock);
146 if (likely(!vma->anon_vma)) {
147 vma->anon_vma = anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -0800148 avc->anon_vma = anon_vma;
149 avc->vma = vma;
150 list_add(&avc->same_vma, &vma->anon_vma_chain);
Andrea Arcangeli26ba0cb2010-08-09 17:19:08 -0700151 list_add_tail(&avc->same_anon_vma, &anon_vma->head);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700152 allocated = NULL;
Oleg Nesterov31f2b0e2010-04-23 13:18:01 -0400153 avc = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700154 }
155 spin_unlock(&mm->page_table_lock);
Rik van Rielcba48b92010-08-09 17:18:38 -0700156 anon_vma_unlock(anon_vma);
Oleg Nesterov31f2b0e2010-04-23 13:18:01 -0400157
158 if (unlikely(allocated))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700159 anon_vma_free(allocated);
Oleg Nesterov31f2b0e2010-04-23 13:18:01 -0400160 if (unlikely(avc))
Rik van Riel5beb4932010-03-05 13:42:07 -0800161 anon_vma_chain_free(avc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700162 }
163 return 0;
Rik van Riel5beb4932010-03-05 13:42:07 -0800164
165 out_enomem_free_avc:
166 anon_vma_chain_free(avc);
167 out_enomem:
168 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700169}
170
Rik van Riel5beb4932010-03-05 13:42:07 -0800171static void anon_vma_chain_link(struct vm_area_struct *vma,
172 struct anon_vma_chain *avc,
173 struct anon_vma *anon_vma)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700174{
Rik van Riel5beb4932010-03-05 13:42:07 -0800175 avc->vma = vma;
176 avc->anon_vma = anon_vma;
177 list_add(&avc->same_vma, &vma->anon_vma_chain);
178
Rik van Rielcba48b92010-08-09 17:18:38 -0700179 anon_vma_lock(anon_vma);
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 Riel5c341ee12010-08-09 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 Riel5c341ee12010-08-09 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 Riel5c341ee12010-08-09 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 */
Andi Kleen10be22d2009-09-16 11:50:04 +0200317struct anon_vma *page_lock_anon_vma(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700318{
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800319 struct anon_vma *anon_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700320 unsigned long anon_mapping;
321
322 rcu_read_lock();
Hugh Dickins80e148222009-12-14 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);
Rik van Rielcba48b92010-08-09 17:18:38 -0700330 anon_vma_lock(anon_vma);
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800331 return anon_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700332out:
333 rcu_read_unlock();
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800334 return NULL;
335}
336
Andi Kleen10be22d2009-09-16 11:50:04 +0200337void page_unlock_anon_vma(struct anon_vma *anon_vma)
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800338{
Rik van Rielcba48b92010-08-09 17:18:38 -0700339 anon_vma_unlock(anon_vma);
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800340 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700341}
342
343/*
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -0800344 * At what user virtual address is page expected in @vma?
345 * Returns virtual address or -EFAULT if page's index/offset is not
346 * within the range mapped the @vma.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700347 */
348static inline unsigned long
349vma_address(struct page *page, struct vm_area_struct *vma)
350{
351 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
352 unsigned long address;
353
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900354 if (unlikely(is_vm_hugetlb_page(vma)))
355 pgoff = page->index << huge_page_order(page_hstate(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700356 address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
357 if (unlikely(address < vma->vm_start || address >= vma->vm_end)) {
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -0800358 /* page should be within @vma mapping range */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700359 return -EFAULT;
360 }
361 return address;
362}
363
364/*
Huang Shijiebf89c8c2009-10-01 15:44:04 -0700365 * At what user virtual address is page expected in vma?
Naoya Horiguchiab941e02010-05-11 14:06:55 -0700366 * Caller should check the page is actually part of the vma.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700367 */
368unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
369{
Andrea Arcangeli21d0d442010-08-09 17:19:10 -0700370 if (PageAnon(page)) {
371 if (vma->anon_vma->root != page_anon_vma(page)->root)
372 return -EFAULT;
373 } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) {
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800374 if (!vma->vm_file ||
375 vma->vm_file->f_mapping != page->mapping)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700376 return -EFAULT;
377 } else
378 return -EFAULT;
379 return vma_address(page, vma);
380}
381
382/*
Nikita Danilov81b40822005-05-01 08:58:36 -0700383 * Check that @page is mapped at @address into @mm.
384 *
Nick Piggin479db0b2008-08-20 14:09:18 -0700385 * If @sync is false, page_check_address may perform a racy check to avoid
386 * the page table lock when the pte is not present (helpful when reclaiming
387 * highly shared pages).
388 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -0700389 * On success returns with pte mapped and locked.
Nikita Danilov81b40822005-05-01 08:58:36 -0700390 */
Carsten Otteceffc072005-06-23 22:05:25 -0700391pte_t *page_check_address(struct page *page, struct mm_struct *mm,
Nick Piggin479db0b2008-08-20 14:09:18 -0700392 unsigned long address, spinlock_t **ptlp, int sync)
Nikita Danilov81b40822005-05-01 08:58:36 -0700393{
394 pgd_t *pgd;
395 pud_t *pud;
396 pmd_t *pmd;
397 pte_t *pte;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700398 spinlock_t *ptl;
Nikita Danilov81b40822005-05-01 08:58:36 -0700399
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900400 if (unlikely(PageHuge(page))) {
401 pte = huge_pte_offset(mm, address);
402 ptl = &mm->page_table_lock;
403 goto check;
404 }
405
Nikita Danilov81b40822005-05-01 08:58:36 -0700406 pgd = pgd_offset(mm, address);
Hugh Dickinsc0718802005-10-29 18:16:31 -0700407 if (!pgd_present(*pgd))
408 return NULL;
409
410 pud = pud_offset(pgd, address);
411 if (!pud_present(*pud))
412 return NULL;
413
414 pmd = pmd_offset(pud, address);
415 if (!pmd_present(*pmd))
416 return NULL;
417
418 pte = pte_offset_map(pmd, address);
419 /* Make a quick check before getting the lock */
Nick Piggin479db0b2008-08-20 14:09:18 -0700420 if (!sync && !pte_present(*pte)) {
Hugh Dickinsc0718802005-10-29 18:16:31 -0700421 pte_unmap(pte);
422 return NULL;
Nikita Danilov81b40822005-05-01 08:58:36 -0700423 }
Hugh Dickinsc0718802005-10-29 18:16:31 -0700424
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700425 ptl = pte_lockptr(mm, pmd);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900426check:
Hugh Dickinsc0718802005-10-29 18:16:31 -0700427 spin_lock(ptl);
428 if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
429 *ptlp = ptl;
430 return pte;
431 }
432 pte_unmap_unlock(pte, ptl);
433 return NULL;
Nikita Danilov81b40822005-05-01 08:58:36 -0700434}
435
Nick Pigginb291f002008-10-18 20:26:44 -0700436/**
437 * page_mapped_in_vma - check whether a page is really mapped in a VMA
438 * @page: the page to test
439 * @vma: the VMA to test
440 *
441 * Returns 1 if the page is mapped into the page tables of the VMA, 0
442 * if the page is not mapped into the page tables of this VMA. Only
443 * valid for normal file or anonymous VMAs.
444 */
Andi Kleen6a460792009-09-16 11:50:15 +0200445int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
Nick Pigginb291f002008-10-18 20:26:44 -0700446{
447 unsigned long address;
448 pte_t *pte;
449 spinlock_t *ptl;
450
451 address = vma_address(page, vma);
452 if (address == -EFAULT) /* out of vma range */
453 return 0;
454 pte = page_check_address(page, vma->vm_mm, address, &ptl, 1);
455 if (!pte) /* the page is not in this mm */
456 return 0;
457 pte_unmap_unlock(pte, ptl);
458
459 return 1;
460}
461
Nikita Danilov81b40822005-05-01 08:58:36 -0700462/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700463 * Subfunctions of page_referenced: page_referenced_one called
464 * repeatedly from either page_referenced_anon or page_referenced_file.
465 */
Hugh Dickins5ad64682009-12-14 17:59:24 -0800466int page_referenced_one(struct page *page, struct vm_area_struct *vma,
467 unsigned long address, unsigned int *mapcount,
468 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700469{
470 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700471 pte_t *pte;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700472 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700473 int referenced = 0;
474
Nick Piggin479db0b2008-08-20 14:09:18 -0700475 pte = page_check_address(page, mm, address, &ptl, 0);
Hugh Dickinsc0718802005-10-29 18:16:31 -0700476 if (!pte)
477 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700478
Nick Pigginb291f002008-10-18 20:26:44 -0700479 /*
480 * Don't want to elevate referenced for mlocked page that gets this far,
481 * in order that it progresses to try_to_unmap and is moved to the
482 * unevictable list.
483 */
Hugh Dickins5a9bbdc2008-02-04 22:29:23 -0800484 if (vma->vm_flags & VM_LOCKED) {
Hugh Dickins5a9bbdc2008-02-04 22:29:23 -0800485 *mapcount = 1; /* break early from loop */
Minchan Kim03ef83a2009-08-26 14:29:23 -0700486 *vm_flags |= VM_LOCKED;
Nick Pigginb291f002008-10-18 20:26:44 -0700487 goto out_unmap;
488 }
489
Johannes Weiner4917e5d2009-01-06 14:39:17 -0800490 if (ptep_clear_flush_young_notify(vma, address, pte)) {
491 /*
492 * Don't treat a reference through a sequentially read
493 * mapping as such. If the page has been used in
494 * another mapping, we will catch it; if this other
495 * mapping is already gone, the unmap path will have
496 * set PG_referenced or activated the page.
497 */
498 if (likely(!VM_SequentialReadHint(vma)))
499 referenced++;
500 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700501
Hugh Dickinsc0718802005-10-29 18:16:31 -0700502 /* Pretend the page is referenced if the task has the
503 swap token and is in the middle of a page fault. */
Rik van Rielf7b7fd82005-11-28 13:44:07 -0800504 if (mm != current->mm && has_swap_token(mm) &&
Hugh Dickinsc0718802005-10-29 18:16:31 -0700505 rwsem_is_locked(&mm->mmap_sem))
506 referenced++;
507
Nick Pigginb291f002008-10-18 20:26:44 -0700508out_unmap:
Hugh Dickinsc0718802005-10-29 18:16:31 -0700509 (*mapcount)--;
510 pte_unmap_unlock(pte, ptl);
Huang Shijie273f0472009-12-14 17:58:51 -0800511
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700512 if (referenced)
513 *vm_flags |= vma->vm_flags;
Huang Shijie273f0472009-12-14 17:58:51 -0800514out:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700515 return referenced;
516}
517
Balbir Singhbed71612008-02-07 00:14:01 -0800518static int page_referenced_anon(struct page *page,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700519 struct mem_cgroup *mem_cont,
520 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700521{
522 unsigned int mapcount;
523 struct anon_vma *anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -0800524 struct anon_vma_chain *avc;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700525 int referenced = 0;
526
527 anon_vma = page_lock_anon_vma(page);
528 if (!anon_vma)
529 return referenced;
530
531 mapcount = page_mapcount(page);
Rik van Riel5beb4932010-03-05 13:42:07 -0800532 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
533 struct vm_area_struct *vma = avc->vma;
Hugh Dickins1cb17292009-12-14 17:59:01 -0800534 unsigned long address = vma_address(page, vma);
535 if (address == -EFAULT)
536 continue;
Balbir Singhbed71612008-02-07 00:14:01 -0800537 /*
538 * If we are reclaiming on behalf of a cgroup, skip
539 * counting on behalf of references from different
540 * cgroups
541 */
Hugh Dickinsbd845e32008-03-04 14:29:01 -0800542 if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
Balbir Singhbed71612008-02-07 00:14:01 -0800543 continue;
Hugh Dickins1cb17292009-12-14 17:59:01 -0800544 referenced += page_referenced_one(page, vma, address,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700545 &mapcount, vm_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700546 if (!mapcount)
547 break;
548 }
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800549
550 page_unlock_anon_vma(anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700551 return referenced;
552}
553
554/**
555 * page_referenced_file - referenced check for object-based rmap
556 * @page: the page we're checking references on.
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700557 * @mem_cont: target memory controller
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700558 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
Linus Torvalds1da177e2005-04-16 15:20:36 -0700559 *
560 * For an object-based mapped page, find all the places it is mapped and
561 * check/clear the referenced flag. This is done by following the page->mapping
562 * pointer, then walking the chain of vmas it holds. It returns the number
563 * of references it found.
564 *
565 * This function is only called from page_referenced for object-based pages.
566 */
Balbir Singhbed71612008-02-07 00:14:01 -0800567static int page_referenced_file(struct page *page,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700568 struct mem_cgroup *mem_cont,
569 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700570{
571 unsigned int mapcount;
572 struct address_space *mapping = page->mapping;
573 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
574 struct vm_area_struct *vma;
575 struct prio_tree_iter iter;
576 int referenced = 0;
577
578 /*
579 * The caller's checks on page->mapping and !PageAnon have made
580 * sure that this is a file page: the check for page->mapping
581 * excludes the case just before it gets set on an anon page.
582 */
583 BUG_ON(PageAnon(page));
584
585 /*
586 * The page lock not only makes sure that page->mapping cannot
587 * suddenly be NULLified by truncation, it makes sure that the
588 * structure at mapping cannot be freed and reused yet,
589 * so we can safely take mapping->i_mmap_lock.
590 */
591 BUG_ON(!PageLocked(page));
592
593 spin_lock(&mapping->i_mmap_lock);
594
595 /*
596 * i_mmap_lock does not stabilize mapcount at all, but mapcount
597 * is more likely to be accurate if we note it after spinning.
598 */
599 mapcount = page_mapcount(page);
600
601 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
Hugh Dickins1cb17292009-12-14 17:59:01 -0800602 unsigned long address = vma_address(page, vma);
603 if (address == -EFAULT)
604 continue;
Balbir Singhbed71612008-02-07 00:14:01 -0800605 /*
606 * If we are reclaiming on behalf of a cgroup, skip
607 * counting on behalf of references from different
608 * cgroups
609 */
Hugh Dickinsbd845e32008-03-04 14:29:01 -0800610 if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
Balbir Singhbed71612008-02-07 00:14:01 -0800611 continue;
Hugh Dickins1cb17292009-12-14 17:59:01 -0800612 referenced += page_referenced_one(page, vma, address,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700613 &mapcount, vm_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700614 if (!mapcount)
615 break;
616 }
617
618 spin_unlock(&mapping->i_mmap_lock);
619 return referenced;
620}
621
622/**
623 * page_referenced - test if the page was referenced
624 * @page: the page to test
625 * @is_locked: caller holds lock on the page
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700626 * @mem_cont: target memory controller
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700627 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
Linus Torvalds1da177e2005-04-16 15:20:36 -0700628 *
629 * Quick test_and_clear_referenced for all mappings to a page,
630 * returns the number of ptes which referenced the page.
631 */
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700632int page_referenced(struct page *page,
633 int is_locked,
634 struct mem_cgroup *mem_cont,
635 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700636{
637 int referenced = 0;
Hugh Dickins5ad64682009-12-14 17:59:24 -0800638 int we_locked = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700639
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700640 *vm_flags = 0;
Hugh Dickins3ca7b3c2009-12-14 17:58:57 -0800641 if (page_mapped(page) && page_rmapping(page)) {
Hugh Dickins5ad64682009-12-14 17:59:24 -0800642 if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
643 we_locked = trylock_page(page);
644 if (!we_locked) {
645 referenced++;
646 goto out;
647 }
648 }
649 if (unlikely(PageKsm(page)))
650 referenced += page_referenced_ksm(page, mem_cont,
651 vm_flags);
652 else if (PageAnon(page))
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700653 referenced += page_referenced_anon(page, mem_cont,
654 vm_flags);
Hugh Dickins5ad64682009-12-14 17:59:24 -0800655 else if (page->mapping)
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700656 referenced += page_referenced_file(page, mem_cont,
657 vm_flags);
Hugh Dickins5ad64682009-12-14 17:59:24 -0800658 if (we_locked)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700659 unlock_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700660 }
Hugh Dickins5ad64682009-12-14 17:59:24 -0800661out:
Christian Borntraeger5b7baf02008-03-25 18:47:12 +0100662 if (page_test_and_clear_young(page))
663 referenced++;
664
Linus Torvalds1da177e2005-04-16 15:20:36 -0700665 return referenced;
666}
667
Hugh Dickins1cb17292009-12-14 17:59:01 -0800668static int page_mkclean_one(struct page *page, struct vm_area_struct *vma,
669 unsigned long address)
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700670{
671 struct mm_struct *mm = vma->vm_mm;
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100672 pte_t *pte;
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700673 spinlock_t *ptl;
674 int ret = 0;
675
Nick Piggin479db0b2008-08-20 14:09:18 -0700676 pte = page_check_address(page, mm, address, &ptl, 1);
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700677 if (!pte)
678 goto out;
679
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100680 if (pte_dirty(*pte) || pte_write(*pte)) {
681 pte_t entry;
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700682
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100683 flush_cache_page(vma, address, pte_pfn(*pte));
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700684 entry = ptep_clear_flush_notify(vma, address, pte);
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100685 entry = pte_wrprotect(entry);
686 entry = pte_mkclean(entry);
Al Virod6e88e62006-12-29 16:48:35 -0800687 set_pte_at(mm, address, pte, entry);
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100688 ret = 1;
689 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700690
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700691 pte_unmap_unlock(pte, ptl);
692out:
693 return ret;
694}
695
696static int page_mkclean_file(struct address_space *mapping, struct page *page)
697{
698 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
699 struct vm_area_struct *vma;
700 struct prio_tree_iter iter;
701 int ret = 0;
702
703 BUG_ON(PageAnon(page));
704
705 spin_lock(&mapping->i_mmap_lock);
706 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
Hugh Dickins1cb17292009-12-14 17:59:01 -0800707 if (vma->vm_flags & VM_SHARED) {
708 unsigned long address = vma_address(page, vma);
709 if (address == -EFAULT)
710 continue;
711 ret += page_mkclean_one(page, vma, address);
712 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700713 }
714 spin_unlock(&mapping->i_mmap_lock);
715 return ret;
716}
717
718int page_mkclean(struct page *page)
719{
720 int ret = 0;
721
722 BUG_ON(!PageLocked(page));
723
724 if (page_mapped(page)) {
725 struct address_space *mapping = page_mapping(page);
Christian Borntraegerce7e9fa2007-11-20 11:13:36 +0100726 if (mapping) {
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700727 ret = page_mkclean_file(mapping, page);
Christian Borntraegerce7e9fa2007-11-20 11:13:36 +0100728 if (page_test_dirty(page)) {
729 page_clear_dirty(page);
730 ret = 1;
731 }
Martin Schwidefsky6c210482007-04-27 16:01:57 +0200732 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700733 }
734
735 return ret;
736}
Jaya Kumar60b59be2007-05-08 00:37:37 -0700737EXPORT_SYMBOL_GPL(page_mkclean);
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700738
Linus Torvalds1da177e2005-04-16 15:20:36 -0700739/**
Rik van Rielc44b6742010-03-05 13:42:09 -0800740 * page_move_anon_rmap - move a page to our anon_vma
741 * @page: the page to move to our anon_vma
742 * @vma: the vma the page belongs to
743 * @address: the user virtual address mapped
744 *
745 * When a page belongs exclusively to one process after a COW event,
746 * that page can be moved into the anon_vma that belongs to just that
747 * process, so the rmap code will not search the parent or sibling
748 * processes.
749 */
750void page_move_anon_rmap(struct page *page,
751 struct vm_area_struct *vma, unsigned long address)
752{
753 struct anon_vma *anon_vma = vma->anon_vma;
754
755 VM_BUG_ON(!PageLocked(page));
756 VM_BUG_ON(!anon_vma);
757 VM_BUG_ON(page->index != linear_page_index(vma, address));
758
759 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
760 page->mapping = (struct address_space *) anon_vma;
761}
762
763/**
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700764 * __page_set_anon_rmap - setup new anonymous rmap
Nick Piggin9617d952006-01-06 00:11:12 -0800765 * @page: the page to add the mapping to
766 * @vma: the vm area in which the mapping is added
767 * @address: the user virtual address mapped
Rik van Riele8a03fe2010-04-14 17:59:28 -0400768 * @exclusive: the page is exclusively owned by the current process
Nick Piggin9617d952006-01-06 00:11:12 -0800769 */
770static void __page_set_anon_rmap(struct page *page,
Rik van Riele8a03fe2010-04-14 17:59:28 -0400771 struct vm_area_struct *vma, unsigned long address, int exclusive)
Nick Piggin9617d952006-01-06 00:11:12 -0800772{
Rik van Riele8a03fe2010-04-14 17:59:28 -0400773 struct anon_vma *anon_vma = vma->anon_vma;
Nick Piggin9617d952006-01-06 00:11:12 -0800774
Rik van Riele8a03fe2010-04-14 17:59:28 -0400775 BUG_ON(!anon_vma);
Linus Torvaldsea900022010-04-12 12:44:29 -0700776
777 /*
Rik van Riele8a03fe2010-04-14 17:59:28 -0400778 * If the page isn't exclusively mapped into this vma,
779 * we must use the _oldest_ possible anon_vma for the
780 * page mapping!
Linus Torvaldsea900022010-04-12 12:44:29 -0700781 */
Rik van Riele8a03fe2010-04-14 17:59:28 -0400782 if (!exclusive) {
Andrea Arcangeli288468c2010-08-09 17:19:09 -0700783 if (PageAnon(page))
784 return;
785 anon_vma = anon_vma->root;
786 } else {
787 /*
788 * In this case, swapped-out-but-not-discarded swap-cache
789 * is remapped. So, no need to update page->mapping here.
790 * We convice anon_vma poitned by page->mapping is not obsolete
791 * because vma->anon_vma is necessary to be a family of it.
792 */
793 if (PageAnon(page))
794 return;
Rik van Riele8a03fe2010-04-14 17:59:28 -0400795 }
Linus Torvaldsea900022010-04-12 12:44:29 -0700796
Nick Piggin9617d952006-01-06 00:11:12 -0800797 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
798 page->mapping = (struct address_space *) anon_vma;
Nick Piggin9617d952006-01-06 00:11:12 -0800799 page->index = linear_page_index(vma, address);
Nick Piggin9617d952006-01-06 00:11:12 -0800800}
801
802/**
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700803 * __page_check_anon_rmap - sanity check anonymous rmap addition
Nick Pigginc97a9e12007-05-16 22:11:21 -0700804 * @page: the page to add the mapping to
805 * @vma: the vm area in which the mapping is added
806 * @address: the user virtual address mapped
807 */
808static void __page_check_anon_rmap(struct page *page,
809 struct vm_area_struct *vma, unsigned long address)
810{
811#ifdef CONFIG_DEBUG_VM
812 /*
813 * The page's anon-rmap details (mapping and index) are guaranteed to
814 * be set up correctly at this point.
815 *
816 * We have exclusion against page_add_anon_rmap because the caller
817 * always holds the page locked, except if called from page_dup_rmap,
818 * in which case the page is already known to be setup.
819 *
820 * We have exclusion against page_add_new_anon_rmap because those pages
821 * are initially only visible via the pagetables, and the pte is locked
822 * over the call to page_add_new_anon_rmap.
823 */
Andrea Arcangeli44ab57a2010-08-09 17:19:10 -0700824 BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root);
Nick Pigginc97a9e12007-05-16 22:11:21 -0700825 BUG_ON(page->index != linear_page_index(vma, address));
826#endif
827}
828
829/**
Linus Torvalds1da177e2005-04-16 15:20:36 -0700830 * page_add_anon_rmap - add pte mapping to an anonymous page
831 * @page: the page to add the mapping to
832 * @vma: the vm area in which the mapping is added
833 * @address: the user virtual address mapped
834 *
Hugh Dickins5ad64682009-12-14 17:59:24 -0800835 * The caller needs to hold the pte lock, and the page must be locked in
Hugh Dickins80e148222009-12-14 17:59:29 -0800836 * the anon_vma case: to serialize mapping,index checking after setting,
837 * and to ensure that PageAnon is not being upgraded racily to PageKsm
838 * (but PageKsm is never downgraded to PageAnon).
Linus Torvalds1da177e2005-04-16 15:20:36 -0700839 */
840void page_add_anon_rmap(struct page *page,
841 struct vm_area_struct *vma, unsigned long address)
842{
Rik van Rielad8c2ee2010-08-09 17:19:48 -0700843 do_page_add_anon_rmap(page, vma, address, 0);
844}
845
846/*
847 * Special version of the above for do_swap_page, which often runs
848 * into pages that are exclusively owned by the current process.
849 * Everybody else should continue to use page_add_anon_rmap above.
850 */
851void do_page_add_anon_rmap(struct page *page,
852 struct vm_area_struct *vma, unsigned long address, int exclusive)
853{
Hugh Dickins5ad64682009-12-14 17:59:24 -0800854 int first = atomic_inc_and_test(&page->_mapcount);
855 if (first)
856 __inc_zone_page_state(page, NR_ANON_PAGES);
857 if (unlikely(PageKsm(page)))
858 return;
859
Nick Pigginc97a9e12007-05-16 22:11:21 -0700860 VM_BUG_ON(!PageLocked(page));
861 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
Hugh Dickins5ad64682009-12-14 17:59:24 -0800862 if (first)
Rik van Rielad8c2ee2010-08-09 17:19:48 -0700863 __page_set_anon_rmap(page, vma, address, exclusive);
KAMEZAWA Hiroyuki69029cd2008-07-25 01:47:14 -0700864 else
Nick Pigginc97a9e12007-05-16 22:11:21 -0700865 __page_check_anon_rmap(page, vma, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700866}
867
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700868/**
Nick Piggin9617d952006-01-06 00:11:12 -0800869 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
870 * @page: the page to add the mapping to
871 * @vma: the vm area in which the mapping is added
872 * @address: the user virtual address mapped
873 *
874 * Same as page_add_anon_rmap but must only be called on *new* pages.
875 * This means the inc-and-test can be bypassed.
Nick Pigginc97a9e12007-05-16 22:11:21 -0700876 * Page does not have to be locked.
Nick Piggin9617d952006-01-06 00:11:12 -0800877 */
878void page_add_new_anon_rmap(struct page *page,
879 struct vm_area_struct *vma, unsigned long address)
880{
Hugh Dickinsb5934c52009-01-06 14:39:25 -0800881 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
Hugh Dickinscbf84b72009-01-06 14:39:27 -0800882 SetPageSwapBacked(page);
883 atomic_set(&page->_mapcount, 0); /* increment count (starts at -1) */
Hugh Dickins5ad64682009-12-14 17:59:24 -0800884 __inc_zone_page_state(page, NR_ANON_PAGES);
Rik van Riele8a03fe2010-04-14 17:59:28 -0400885 __page_set_anon_rmap(page, vma, address, 1);
Hugh Dickinsb5934c52009-01-06 14:39:25 -0800886 if (page_evictable(page, vma))
Hugh Dickinscbf84b72009-01-06 14:39:27 -0800887 lru_cache_add_lru(page, LRU_ACTIVE_ANON);
Hugh Dickinsb5934c52009-01-06 14:39:25 -0800888 else
889 add_page_to_unevictable_list(page);
Nick Piggin9617d952006-01-06 00:11:12 -0800890}
891
Linus Torvalds1da177e2005-04-16 15:20:36 -0700892/**
893 * page_add_file_rmap - add pte mapping to a file page
894 * @page: the page to add the mapping to
895 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -0700896 * The caller needs to hold the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700897 */
898void page_add_file_rmap(struct page *page)
899{
Balbir Singhd69b0422009-06-17 16:26:34 -0700900 if (atomic_inc_and_test(&page->_mapcount)) {
Christoph Lameter65ba55f2006-06-30 01:55:34 -0700901 __inc_zone_page_state(page, NR_FILE_MAPPED);
KAMEZAWA Hiroyukid8046582009-12-15 16:47:09 -0800902 mem_cgroup_update_file_mapped(page, 1);
Balbir Singhd69b0422009-06-17 16:26:34 -0700903 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700904}
905
906/**
907 * page_remove_rmap - take down pte mapping from a page
908 * @page: page to remove mapping from
909 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -0700910 * The caller needs to hold the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700911 */
Hugh Dickinsedc315f2009-01-06 14:40:11 -0800912void page_remove_rmap(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700913{
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -0700914 /* page still mapped by someone else? */
915 if (!atomic_add_negative(-1, &page->_mapcount))
916 return;
917
918 /*
919 * Now that the last pte has gone, s390 must transfer dirty
920 * flag from storage key to struct page. We can usually skip
921 * this if the page is anon, so about to be freed; but perhaps
922 * not if it's in swapcache - there might be another pte slot
923 * containing the swap entry, but page not yet written to swap.
924 */
925 if ((!PageAnon(page) || PageSwapCache(page)) && page_test_dirty(page)) {
926 page_clear_dirty(page);
927 set_page_dirty(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700928 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900929 /*
930 * Hugepages are not counted in NR_ANON_PAGES nor NR_FILE_MAPPED
931 * and not charged by memcg for now.
932 */
933 if (unlikely(PageHuge(page)))
934 return;
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -0700935 if (PageAnon(page)) {
936 mem_cgroup_uncharge_page(page);
937 __dec_zone_page_state(page, NR_ANON_PAGES);
938 } else {
939 __dec_zone_page_state(page, NR_FILE_MAPPED);
KAMEZAWA Hiroyukid8046582009-12-15 16:47:09 -0800940 mem_cgroup_update_file_mapped(page, -1);
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -0700941 }
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -0700942 /*
943 * It would be tidy to reset the PageAnon mapping here,
944 * but that might overwrite a racing page_add_anon_rmap
945 * which increments mapcount after us but sets mapping
946 * before us: so leave the reset to free_hot_cold_page,
947 * and remember that it's only reliable while mapped.
948 * Leaving it set also helps swapoff to reinstate ptes
949 * faster for those pages still in swapcache.
950 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700951}
952
953/*
954 * Subfunctions of try_to_unmap: try_to_unmap_one called
955 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
956 */
Hugh Dickins5ad64682009-12-14 17:59:24 -0800957int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
958 unsigned long address, enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700959{
960 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700961 pte_t *pte;
962 pte_t pteval;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700963 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700964 int ret = SWAP_AGAIN;
965
Nick Piggin479db0b2008-08-20 14:09:18 -0700966 pte = page_check_address(page, mm, address, &ptl, 0);
Hugh Dickinsc0718802005-10-29 18:16:31 -0700967 if (!pte)
Nikita Danilov81b40822005-05-01 08:58:36 -0700968 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700969
970 /*
971 * If the page is mlock()d, we cannot swap it out.
972 * If it's recently referenced (perhaps page_referenced
973 * skipped over this mm) then we should reactivate it.
974 */
Andi Kleen14fa31b2009-09-16 11:50:10 +0200975 if (!(flags & TTU_IGNORE_MLOCK)) {
KOSAKI Motohirocaed0f482009-12-14 17:59:45 -0800976 if (vma->vm_flags & VM_LOCKED)
977 goto out_mlock;
978
Hugh Dickinsaf8e3352009-12-14 17:58:59 -0800979 if (TTU_ACTION(flags) == TTU_MUNLOCK)
Hugh Dickins53f79ac2009-12-14 17:58:58 -0800980 goto out_unmap;
Andi Kleen14fa31b2009-09-16 11:50:10 +0200981 }
982 if (!(flags & TTU_IGNORE_ACCESS)) {
Nick Pigginb291f002008-10-18 20:26:44 -0700983 if (ptep_clear_flush_young_notify(vma, address, pte)) {
984 ret = SWAP_FAIL;
985 goto out_unmap;
986 }
987 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700988
Linus Torvalds1da177e2005-04-16 15:20:36 -0700989 /* Nuke the page table entry. */
990 flush_cache_page(vma, address, page_to_pfn(page));
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700991 pteval = ptep_clear_flush_notify(vma, address, pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700992
993 /* Move the dirty bit to the physical page now the pte is gone. */
994 if (pte_dirty(pteval))
995 set_page_dirty(page);
996
Hugh Dickins365e9c872005-10-29 18:16:18 -0700997 /* Update high watermark before we lower rss */
998 update_hiwater_rss(mm);
999
Andi Kleen888b9f72009-09-16 11:50:11 +02001000 if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
1001 if (PageAnon(page))
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001002 dec_mm_counter(mm, MM_ANONPAGES);
Andi Kleen888b9f72009-09-16 11:50:11 +02001003 else
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001004 dec_mm_counter(mm, MM_FILEPAGES);
Andi Kleen888b9f72009-09-16 11:50:11 +02001005 set_pte_at(mm, address, pte,
1006 swp_entry_to_pte(make_hwpoison_entry(page)));
1007 } else if (PageAnon(page)) {
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07001008 swp_entry_t entry = { .val = page_private(page) };
Christoph Lameter06972122006-06-23 02:03:35 -07001009
1010 if (PageSwapCache(page)) {
1011 /*
1012 * Store the swap location in the pte.
1013 * See handle_pte_fault() ...
1014 */
Hugh Dickins570a335b2009-12-14 17:58:46 -08001015 if (swap_duplicate(entry) < 0) {
1016 set_pte_at(mm, address, pte, pteval);
1017 ret = SWAP_FAIL;
1018 goto out_unmap;
1019 }
Christoph Lameter06972122006-06-23 02:03:35 -07001020 if (list_empty(&mm->mmlist)) {
1021 spin_lock(&mmlist_lock);
1022 if (list_empty(&mm->mmlist))
1023 list_add(&mm->mmlist, &init_mm.mmlist);
1024 spin_unlock(&mmlist_lock);
1025 }
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001026 dec_mm_counter(mm, MM_ANONPAGES);
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08001027 inc_mm_counter(mm, MM_SWAPENTS);
KOSAKI Motohiro64cdd542009-01-06 14:39:16 -08001028 } else if (PAGE_MIGRATION) {
Christoph Lameter06972122006-06-23 02:03:35 -07001029 /*
1030 * Store the pfn of the page in a special migration
1031 * pte. do_swap_page() will wait until the migration
1032 * pte is removed and then restart fault handling.
1033 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001034 BUG_ON(TTU_ACTION(flags) != TTU_MIGRATION);
Christoph Lameter06972122006-06-23 02:03:35 -07001035 entry = make_migration_entry(page, pte_write(pteval));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001036 }
1037 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
1038 BUG_ON(pte_file(*pte));
Andi Kleen14fa31b2009-09-16 11:50:10 +02001039 } else if (PAGE_MIGRATION && (TTU_ACTION(flags) == TTU_MIGRATION)) {
Christoph Lameter04e62a22006-06-23 02:03:38 -07001040 /* Establish migration entry for a file page */
1041 swp_entry_t entry;
1042 entry = make_migration_entry(page, pte_write(pteval));
1043 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
1044 } else
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001045 dec_mm_counter(mm, MM_FILEPAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001046
Hugh Dickinsedc315f2009-01-06 14:40:11 -08001047 page_remove_rmap(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001048 page_cache_release(page);
1049
1050out_unmap:
Hugh Dickinsc0718802005-10-29 18:16:31 -07001051 pte_unmap_unlock(pte, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001052out:
1053 return ret;
KOSAKI Motohirocaed0f482009-12-14 17:59:45 -08001054
1055out_mlock:
1056 pte_unmap_unlock(pte, ptl);
1057
1058
1059 /*
1060 * We need mmap_sem locking, Otherwise VM_LOCKED check makes
1061 * unstable result and race. Plus, We can't wait here because
1062 * we now hold anon_vma->lock or mapping->i_mmap_lock.
1063 * if trylock failed, the page remain in evictable lru and later
1064 * vmscan could retry to move the page to unevictable lru if the
1065 * page is actually mlocked.
1066 */
1067 if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
1068 if (vma->vm_flags & VM_LOCKED) {
1069 mlock_vma_page(page);
1070 ret = SWAP_MLOCK;
1071 }
1072 up_read(&vma->vm_mm->mmap_sem);
1073 }
1074 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001075}
1076
1077/*
1078 * objrmap doesn't work for nonlinear VMAs because the assumption that
1079 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
1080 * Consequently, given a particular page and its ->index, we cannot locate the
1081 * ptes which are mapping that page without an exhaustive linear search.
1082 *
1083 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
1084 * maps the file to which the target page belongs. The ->vm_private_data field
1085 * holds the current cursor into that scan. Successive searches will circulate
1086 * around the vma's virtual address space.
1087 *
1088 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
1089 * more scanning pressure is placed against them as well. Eventually pages
1090 * will become fully unmapped and are eligible for eviction.
1091 *
1092 * For very sparsely populated VMAs this is a little inefficient - chances are
1093 * there there won't be many ptes located within the scan cluster. In this case
1094 * maybe we could scan further - to the end of the pte page, perhaps.
Nick Pigginb291f002008-10-18 20:26:44 -07001095 *
1096 * Mlocked pages: check VM_LOCKED under mmap_sem held for read, if we can
1097 * acquire it without blocking. If vma locked, mlock the pages in the cluster,
1098 * rather than unmapping them. If we encounter the "check_page" that vmscan is
1099 * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001100 */
1101#define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
1102#define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
1103
Nick Pigginb291f002008-10-18 20:26:44 -07001104static int try_to_unmap_cluster(unsigned long cursor, unsigned int *mapcount,
1105 struct vm_area_struct *vma, struct page *check_page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001106{
1107 struct mm_struct *mm = vma->vm_mm;
1108 pgd_t *pgd;
1109 pud_t *pud;
1110 pmd_t *pmd;
Hugh Dickinsc0718802005-10-29 18:16:31 -07001111 pte_t *pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001112 pte_t pteval;
Hugh Dickinsc0718802005-10-29 18:16:31 -07001113 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001114 struct page *page;
1115 unsigned long address;
1116 unsigned long end;
Nick Pigginb291f002008-10-18 20:26:44 -07001117 int ret = SWAP_AGAIN;
1118 int locked_vma = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001119
Linus Torvalds1da177e2005-04-16 15:20:36 -07001120 address = (vma->vm_start + cursor) & CLUSTER_MASK;
1121 end = address + CLUSTER_SIZE;
1122 if (address < vma->vm_start)
1123 address = vma->vm_start;
1124 if (end > vma->vm_end)
1125 end = vma->vm_end;
1126
1127 pgd = pgd_offset(mm, address);
1128 if (!pgd_present(*pgd))
Nick Pigginb291f002008-10-18 20:26:44 -07001129 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001130
1131 pud = pud_offset(pgd, address);
1132 if (!pud_present(*pud))
Nick Pigginb291f002008-10-18 20:26:44 -07001133 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001134
1135 pmd = pmd_offset(pud, address);
1136 if (!pmd_present(*pmd))
Nick Pigginb291f002008-10-18 20:26:44 -07001137 return ret;
1138
1139 /*
Hugh Dickinsaf8e3352009-12-14 17:58:59 -08001140 * If we can acquire the mmap_sem for read, and vma is VM_LOCKED,
Nick Pigginb291f002008-10-18 20:26:44 -07001141 * keep the sem while scanning the cluster for mlocking pages.
1142 */
Hugh Dickinsaf8e3352009-12-14 17:58:59 -08001143 if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
Nick Pigginb291f002008-10-18 20:26:44 -07001144 locked_vma = (vma->vm_flags & VM_LOCKED);
1145 if (!locked_vma)
1146 up_read(&vma->vm_mm->mmap_sem); /* don't need it */
1147 }
Hugh Dickinsc0718802005-10-29 18:16:31 -07001148
1149 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001150
Hugh Dickins365e9c872005-10-29 18:16:18 -07001151 /* Update high watermark before we lower rss */
1152 update_hiwater_rss(mm);
1153
Hugh Dickinsc0718802005-10-29 18:16:31 -07001154 for (; address < end; pte++, address += PAGE_SIZE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001155 if (!pte_present(*pte))
1156 continue;
Linus Torvalds6aab3412005-11-28 14:34:23 -08001157 page = vm_normal_page(vma, address, *pte);
1158 BUG_ON(!page || PageAnon(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001159
Nick Pigginb291f002008-10-18 20:26:44 -07001160 if (locked_vma) {
1161 mlock_vma_page(page); /* no-op if already mlocked */
1162 if (page == check_page)
1163 ret = SWAP_MLOCK;
1164 continue; /* don't unmap */
1165 }
1166
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001167 if (ptep_clear_flush_young_notify(vma, address, pte))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001168 continue;
1169
1170 /* Nuke the page table entry. */
Ben Collinseca35132005-11-29 11:45:26 -08001171 flush_cache_page(vma, address, pte_pfn(*pte));
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001172 pteval = ptep_clear_flush_notify(vma, address, pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001173
1174 /* If nonlinear, store the file page offset in the pte. */
1175 if (page->index != linear_page_index(vma, address))
1176 set_pte_at(mm, address, pte, pgoff_to_pte(page->index));
1177
1178 /* Move the dirty bit to the physical page now the pte is gone. */
1179 if (pte_dirty(pteval))
1180 set_page_dirty(page);
1181
Hugh Dickinsedc315f2009-01-06 14:40:11 -08001182 page_remove_rmap(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001183 page_cache_release(page);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001184 dec_mm_counter(mm, MM_FILEPAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001185 (*mapcount)--;
1186 }
Hugh Dickinsc0718802005-10-29 18:16:31 -07001187 pte_unmap_unlock(pte - 1, ptl);
Nick Pigginb291f002008-10-18 20:26:44 -07001188 if (locked_vma)
1189 up_read(&vma->vm_mm->mmap_sem);
1190 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001191}
1192
Mel Gormana8bef8ff2010-05-24 14:32:24 -07001193static bool is_vma_temporary_stack(struct vm_area_struct *vma)
1194{
1195 int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP);
1196
1197 if (!maybe_stack)
1198 return false;
1199
1200 if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) ==
1201 VM_STACK_INCOMPLETE_SETUP)
1202 return true;
1203
1204 return false;
1205}
1206
Nick Pigginb291f002008-10-18 20:26:44 -07001207/**
1208 * try_to_unmap_anon - unmap or unlock anonymous page using the object-based
1209 * rmap method
1210 * @page: the page to unmap/unlock
Huang Shijie8051be52009-12-14 17:58:50 -08001211 * @flags: action and flags
Nick Pigginb291f002008-10-18 20:26:44 -07001212 *
1213 * Find all the mappings of a page using the mapping pointer and the vma chains
1214 * contained in the anon_vma struct it points to.
1215 *
1216 * This function is only called from try_to_unmap/try_to_munlock for
1217 * anonymous pages.
1218 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1219 * where the page was found will be held for write. So, we won't recheck
1220 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1221 * 'LOCKED.
1222 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001223static int try_to_unmap_anon(struct page *page, enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001224{
1225 struct anon_vma *anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -08001226 struct anon_vma_chain *avc;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001227 int ret = SWAP_AGAIN;
Nick Pigginb291f002008-10-18 20:26:44 -07001228
Linus Torvalds1da177e2005-04-16 15:20:36 -07001229 anon_vma = page_lock_anon_vma(page);
1230 if (!anon_vma)
1231 return ret;
1232
Rik van Riel5beb4932010-03-05 13:42:07 -08001233 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1234 struct vm_area_struct *vma = avc->vma;
Mel Gormana8bef8ff2010-05-24 14:32:24 -07001235 unsigned long address;
1236
1237 /*
1238 * During exec, a temporary VMA is setup and later moved.
1239 * The VMA is moved under the anon_vma lock but not the
1240 * page tables leading to a race where migration cannot
1241 * find the migration ptes. Rather than increasing the
1242 * locking requirements of exec(), migration skips
1243 * temporary VMAs until after exec() completes.
1244 */
1245 if (PAGE_MIGRATION && (flags & TTU_MIGRATION) &&
1246 is_vma_temporary_stack(vma))
1247 continue;
1248
1249 address = vma_address(page, vma);
Hugh Dickins1cb17292009-12-14 17:59:01 -08001250 if (address == -EFAULT)
1251 continue;
1252 ret = try_to_unmap_one(page, vma, address, flags);
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001253 if (ret != SWAP_AGAIN || !page_mapped(page))
1254 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001255 }
Oleg Nesterov34bbd702007-02-28 20:13:49 -08001256
1257 page_unlock_anon_vma(anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001258 return ret;
1259}
1260
1261/**
Nick Pigginb291f002008-10-18 20:26:44 -07001262 * try_to_unmap_file - unmap/unlock file page using the object-based rmap method
1263 * @page: the page to unmap/unlock
Andi Kleen14fa31b2009-09-16 11:50:10 +02001264 * @flags: action and flags
Linus Torvalds1da177e2005-04-16 15:20:36 -07001265 *
1266 * Find all the mappings of a page using the mapping pointer and the vma chains
1267 * contained in the address_space struct it points to.
1268 *
Nick Pigginb291f002008-10-18 20:26:44 -07001269 * This function is only called from try_to_unmap/try_to_munlock for
1270 * object-based pages.
1271 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1272 * where the page was found will be held for write. So, we won't recheck
1273 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1274 * 'LOCKED.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001275 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001276static int try_to_unmap_file(struct page *page, enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001277{
1278 struct address_space *mapping = page->mapping;
1279 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1280 struct vm_area_struct *vma;
1281 struct prio_tree_iter iter;
1282 int ret = SWAP_AGAIN;
1283 unsigned long cursor;
1284 unsigned long max_nl_cursor = 0;
1285 unsigned long max_nl_size = 0;
1286 unsigned int mapcount;
1287
1288 spin_lock(&mapping->i_mmap_lock);
1289 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
Hugh Dickins1cb17292009-12-14 17:59:01 -08001290 unsigned long address = vma_address(page, vma);
1291 if (address == -EFAULT)
1292 continue;
1293 ret = try_to_unmap_one(page, vma, address, flags);
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001294 if (ret != SWAP_AGAIN || !page_mapped(page))
1295 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001296 }
1297
1298 if (list_empty(&mapping->i_mmap_nonlinear))
1299 goto out;
1300
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001301 /*
1302 * We don't bother to try to find the munlocked page in nonlinears.
1303 * It's costly. Instead, later, page reclaim logic may call
1304 * try_to_unmap(TTU_MUNLOCK) and recover PG_mlocked lazily.
1305 */
1306 if (TTU_ACTION(flags) == TTU_MUNLOCK)
1307 goto out;
1308
Linus Torvalds1da177e2005-04-16 15:20:36 -07001309 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1310 shared.vm_set.list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001311 cursor = (unsigned long) vma->vm_private_data;
1312 if (cursor > max_nl_cursor)
1313 max_nl_cursor = cursor;
1314 cursor = vma->vm_end - vma->vm_start;
1315 if (cursor > max_nl_size)
1316 max_nl_size = cursor;
1317 }
1318
Nick Pigginb291f002008-10-18 20:26:44 -07001319 if (max_nl_size == 0) { /* all nonlinears locked or reserved ? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001320 ret = SWAP_FAIL;
1321 goto out;
1322 }
1323
1324 /*
1325 * We don't try to search for this page in the nonlinear vmas,
1326 * and page_referenced wouldn't have found it anyway. Instead
1327 * just walk the nonlinear vmas trying to age and unmap some.
1328 * The mapcount of the page we came in with is irrelevant,
1329 * but even so use it as a guide to how hard we should try?
1330 */
1331 mapcount = page_mapcount(page);
1332 if (!mapcount)
1333 goto out;
1334 cond_resched_lock(&mapping->i_mmap_lock);
1335
1336 max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
1337 if (max_nl_cursor == 0)
1338 max_nl_cursor = CLUSTER_SIZE;
1339
1340 do {
1341 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1342 shared.vm_set.list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001343 cursor = (unsigned long) vma->vm_private_data;
Hugh Dickins839b9682005-09-03 15:54:43 -07001344 while ( cursor < max_nl_cursor &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07001345 cursor < vma->vm_end - vma->vm_start) {
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001346 if (try_to_unmap_cluster(cursor, &mapcount,
1347 vma, page) == SWAP_MLOCK)
1348 ret = SWAP_MLOCK;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001349 cursor += CLUSTER_SIZE;
1350 vma->vm_private_data = (void *) cursor;
1351 if ((int)mapcount <= 0)
1352 goto out;
1353 }
1354 vma->vm_private_data = (void *) max_nl_cursor;
1355 }
1356 cond_resched_lock(&mapping->i_mmap_lock);
1357 max_nl_cursor += CLUSTER_SIZE;
1358 } while (max_nl_cursor <= max_nl_size);
1359
1360 /*
1361 * Don't loop forever (perhaps all the remaining pages are
1362 * in locked vmas). Reset cursor on all unreserved nonlinear
1363 * vmas, now forgetting on which ones it had fallen behind.
1364 */
Hugh Dickins101d2be2005-11-21 21:32:16 -08001365 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1366 vma->vm_private_data = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001367out:
1368 spin_unlock(&mapping->i_mmap_lock);
1369 return ret;
1370}
1371
1372/**
1373 * try_to_unmap - try to remove all page table mappings to a page
1374 * @page: the page to get unmapped
Andi Kleen14fa31b2009-09-16 11:50:10 +02001375 * @flags: action and flags
Linus Torvalds1da177e2005-04-16 15:20:36 -07001376 *
1377 * Tries to remove all the page table entries which are mapping this
1378 * page, used in the pageout path. Caller must hold the page lock.
1379 * Return values are:
1380 *
1381 * SWAP_SUCCESS - we succeeded in removing all mappings
1382 * SWAP_AGAIN - we missed a mapping, try again later
1383 * SWAP_FAIL - the page is unswappable
Nick Pigginb291f002008-10-18 20:26:44 -07001384 * SWAP_MLOCK - page is mlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001385 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001386int try_to_unmap(struct page *page, enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001387{
1388 int ret;
1389
Linus Torvalds1da177e2005-04-16 15:20:36 -07001390 BUG_ON(!PageLocked(page));
1391
Hugh Dickins5ad64682009-12-14 17:59:24 -08001392 if (unlikely(PageKsm(page)))
1393 ret = try_to_unmap_ksm(page, flags);
1394 else if (PageAnon(page))
Andi Kleen14fa31b2009-09-16 11:50:10 +02001395 ret = try_to_unmap_anon(page, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001396 else
Andi Kleen14fa31b2009-09-16 11:50:10 +02001397 ret = try_to_unmap_file(page, flags);
Nick Pigginb291f002008-10-18 20:26:44 -07001398 if (ret != SWAP_MLOCK && !page_mapped(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001399 ret = SWAP_SUCCESS;
1400 return ret;
1401}
Nikita Danilov81b40822005-05-01 08:58:36 -07001402
Nick Pigginb291f002008-10-18 20:26:44 -07001403/**
1404 * try_to_munlock - try to munlock a page
1405 * @page: the page to be munlocked
1406 *
1407 * Called from munlock code. Checks all of the VMAs mapping the page
1408 * to make sure nobody else has this page mlocked. The page will be
1409 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1410 *
1411 * Return values are:
1412 *
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001413 * SWAP_AGAIN - no vma is holding page mlocked, or,
Nick Pigginb291f002008-10-18 20:26:44 -07001414 * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem
Hugh Dickins5ad64682009-12-14 17:59:24 -08001415 * SWAP_FAIL - page cannot be located at present
Nick Pigginb291f002008-10-18 20:26:44 -07001416 * SWAP_MLOCK - page is now mlocked.
1417 */
1418int try_to_munlock(struct page *page)
1419{
1420 VM_BUG_ON(!PageLocked(page) || PageLRU(page));
1421
Hugh Dickins5ad64682009-12-14 17:59:24 -08001422 if (unlikely(PageKsm(page)))
1423 return try_to_unmap_ksm(page, TTU_MUNLOCK);
1424 else if (PageAnon(page))
Andi Kleen14fa31b2009-09-16 11:50:10 +02001425 return try_to_unmap_anon(page, TTU_MUNLOCK);
Nick Pigginb291f002008-10-18 20:26:44 -07001426 else
Andi Kleen14fa31b2009-09-16 11:50:10 +02001427 return try_to_unmap_file(page, TTU_MUNLOCK);
Nick Pigginb291f002008-10-18 20:26:44 -07001428}
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001429
Rik van Riel76545062010-08-09 17:18:41 -07001430#if defined(CONFIG_KSM) || defined(CONFIG_MIGRATION)
1431/*
1432 * Drop an anon_vma refcount, freeing the anon_vma and anon_vma->root
1433 * if necessary. Be careful to do all the tests under the lock. Once
1434 * we know we are the last user, nobody else can get a reference and we
1435 * can do the freeing without the lock.
1436 */
1437void drop_anon_vma(struct anon_vma *anon_vma)
1438{
Andrea Arcangeli44ab57a2010-08-09 17:19:10 -07001439 BUG_ON(atomic_read(&anon_vma->external_refcount) <= 0);
Rik van Riel76545062010-08-09 17:18:41 -07001440 if (atomic_dec_and_lock(&anon_vma->external_refcount, &anon_vma->root->lock)) {
1441 struct anon_vma *root = anon_vma->root;
1442 int empty = list_empty(&anon_vma->head);
1443 int last_root_user = 0;
1444 int root_empty = 0;
1445
1446 /*
1447 * The refcount on a non-root anon_vma got dropped. Drop
1448 * the refcount on the root and check if we need to free it.
1449 */
1450 if (empty && anon_vma != root) {
Andrea Arcangeli44ab57a2010-08-09 17:19:10 -07001451 BUG_ON(atomic_read(&root->external_refcount) <= 0);
Rik van Riel76545062010-08-09 17:18:41 -07001452 last_root_user = atomic_dec_and_test(&root->external_refcount);
1453 root_empty = list_empty(&root->head);
1454 }
1455 anon_vma_unlock(anon_vma);
1456
1457 if (empty) {
1458 anon_vma_free(anon_vma);
1459 if (root_empty && last_root_user)
1460 anon_vma_free(root);
1461 }
1462 }
1463}
1464#endif
1465
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001466#ifdef CONFIG_MIGRATION
1467/*
1468 * rmap_walk() and its helpers rmap_walk_anon() and rmap_walk_file():
1469 * Called by migrate.c to remove migration ptes, but might be used more later.
1470 */
1471static int rmap_walk_anon(struct page *page, int (*rmap_one)(struct page *,
1472 struct vm_area_struct *, unsigned long, void *), void *arg)
1473{
1474 struct anon_vma *anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -08001475 struct anon_vma_chain *avc;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001476 int ret = SWAP_AGAIN;
1477
1478 /*
1479 * Note: remove_migration_ptes() cannot use page_lock_anon_vma()
1480 * because that depends on page_mapped(); but not all its usages
Mel Gorman3f6c8272010-05-24 14:32:17 -07001481 * are holding mmap_sem. Users without mmap_sem are required to
1482 * take a reference count to prevent the anon_vma disappearing
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001483 */
1484 anon_vma = page_anon_vma(page);
1485 if (!anon_vma)
1486 return ret;
Rik van Rielcba48b92010-08-09 17:18:38 -07001487 anon_vma_lock(anon_vma);
Rik van Riel5beb4932010-03-05 13:42:07 -08001488 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1489 struct vm_area_struct *vma = avc->vma;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001490 unsigned long address = vma_address(page, vma);
1491 if (address == -EFAULT)
1492 continue;
1493 ret = rmap_one(page, vma, address, arg);
1494 if (ret != SWAP_AGAIN)
1495 break;
1496 }
Rik van Rielcba48b92010-08-09 17:18:38 -07001497 anon_vma_unlock(anon_vma);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001498 return ret;
1499}
1500
1501static int rmap_walk_file(struct page *page, int (*rmap_one)(struct page *,
1502 struct vm_area_struct *, unsigned long, void *), void *arg)
1503{
1504 struct address_space *mapping = page->mapping;
1505 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1506 struct vm_area_struct *vma;
1507 struct prio_tree_iter iter;
1508 int ret = SWAP_AGAIN;
1509
1510 if (!mapping)
1511 return ret;
1512 spin_lock(&mapping->i_mmap_lock);
1513 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1514 unsigned long address = vma_address(page, vma);
1515 if (address == -EFAULT)
1516 continue;
1517 ret = rmap_one(page, vma, address, arg);
1518 if (ret != SWAP_AGAIN)
1519 break;
1520 }
1521 /*
1522 * No nonlinear handling: being always shared, nonlinear vmas
1523 * never contain migration ptes. Decide what to do about this
1524 * limitation to linear when we need rmap_walk() on nonlinear.
1525 */
1526 spin_unlock(&mapping->i_mmap_lock);
1527 return ret;
1528}
1529
1530int rmap_walk(struct page *page, int (*rmap_one)(struct page *,
1531 struct vm_area_struct *, unsigned long, void *), void *arg)
1532{
1533 VM_BUG_ON(!PageLocked(page));
1534
1535 if (unlikely(PageKsm(page)))
1536 return rmap_walk_ksm(page, rmap_one, arg);
1537 else if (PageAnon(page))
1538 return rmap_walk_anon(page, rmap_one, arg);
1539 else
1540 return rmap_walk_file(page, rmap_one, arg);
1541}
1542#endif /* CONFIG_MIGRATION */
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001543
Naoya Horiguchie3390f62010-06-15 13:18:13 +09001544#ifdef CONFIG_HUGETLB_PAGE
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001545/*
1546 * The following three functions are for anonymous (private mapped) hugepages.
1547 * Unlike common anonymous pages, anonymous hugepages have no accounting code
1548 * and no lru code, because we handle hugepages differently from common pages.
1549 */
1550static void __hugepage_set_anon_rmap(struct page *page,
1551 struct vm_area_struct *vma, unsigned long address, int exclusive)
1552{
1553 struct anon_vma *anon_vma = vma->anon_vma;
1554 BUG_ON(!anon_vma);
1555 if (!exclusive) {
1556 struct anon_vma_chain *avc;
1557 avc = list_entry(vma->anon_vma_chain.prev,
1558 struct anon_vma_chain, same_vma);
1559 anon_vma = avc->anon_vma;
1560 }
1561 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1562 page->mapping = (struct address_space *) anon_vma;
1563 page->index = linear_page_index(vma, address);
1564}
1565
1566void hugepage_add_anon_rmap(struct page *page,
1567 struct vm_area_struct *vma, unsigned long address)
1568{
1569 struct anon_vma *anon_vma = vma->anon_vma;
1570 int first;
1571 BUG_ON(!anon_vma);
1572 BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1573 first = atomic_inc_and_test(&page->_mapcount);
1574 if (first)
1575 __hugepage_set_anon_rmap(page, vma, address, 0);
1576}
1577
1578void hugepage_add_new_anon_rmap(struct page *page,
1579 struct vm_area_struct *vma, unsigned long address)
1580{
1581 BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1582 atomic_set(&page->_mapcount, 0);
1583 __hugepage_set_anon_rmap(page, vma, address, 1);
1584}
Naoya Horiguchie3390f62010-06-15 13:18:13 +09001585#endif /* CONFIG_HUGETLB_PAGE */