blob: 23295f65ae43386574f9058e30e28cdca9a8a708 [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)
Peter Zijlstra3d48ae42011-05-24 17:12:06 -070027 * mapping->i_mmap_mutex
Peter Zijlstra2b575eb2011-05-24 17:12:11 -070028 * anon_vma->mutex
Nick Piggin82591e62006-10-19 23:29:10 -070029 * mm->page_table_lock or pte_lock
30 * zone->lru_lock (in mark_page_accessed, isolate_lru_page)
31 * swap_lock (in swap_duplicate, swap_info_get)
32 * mmlist_lock (in mmput, drain_mmlist and others)
33 * mapping->private_lock (in __set_page_dirty_buffers)
Dave Chinner250df6e2011-03-22 22:23:36 +110034 * inode->i_lock (in set_page_dirty's __mark_inode_dirty)
Dave Chinnera66979a2011-03-22 22:23:41 +110035 * inode_wb_list_lock (in set_page_dirty's __mark_inode_dirty)
Nick Piggin82591e62006-10-19 23:29:10 -070036 * sb_lock (within inode_lock in fs/fs-writeback.c)
37 * mapping->tree_lock (widely used, in set_page_dirty,
38 * in arch-dependent flush_dcache_mmap_lock,
Dave Chinnera66979a2011-03-22 22:23:41 +110039 * within inode_wb_list_lock in __sync_single_inode)
Andi Kleen6a460792009-09-16 11:50:15 +020040 *
Peter Zijlstra9b679322011-06-27 16:18:09 -070041 * anon_vma->mutex,mapping->i_mutex (memory_failure, collect_procs_anon)
42 * ->tasklist_lock
Andi Kleen6a460792009-09-16 11:50:15 +020043 * 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{
Peter Zijlstra01d8b202011-03-22 16:32:49 -070070 struct anon_vma *anon_vma;
71
72 anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
73 if (anon_vma) {
74 atomic_set(&anon_vma->refcount, 1);
75 /*
76 * Initialise the anon_vma root to point to itself. If called
77 * from fork, the root will be reset to the parents anon_vma.
78 */
79 anon_vma->root = anon_vma;
80 }
81
82 return anon_vma;
Adrian Bunkfdd2e5f2008-10-18 20:28:38 -070083}
84
Peter Zijlstra01d8b202011-03-22 16:32:49 -070085static inline void anon_vma_free(struct anon_vma *anon_vma)
Adrian Bunkfdd2e5f2008-10-18 20:28:38 -070086{
Peter Zijlstra01d8b202011-03-22 16:32:49 -070087 VM_BUG_ON(atomic_read(&anon_vma->refcount));
Peter Zijlstra88c22082011-05-24 17:12:13 -070088
89 /*
90 * Synchronize against page_lock_anon_vma() such that
91 * we can safely hold the lock without the anon_vma getting
92 * freed.
93 *
94 * Relies on the full mb implied by the atomic_dec_and_test() from
95 * put_anon_vma() against the acquire barrier implied by
96 * mutex_trylock() from page_lock_anon_vma(). This orders:
97 *
98 * page_lock_anon_vma() VS put_anon_vma()
99 * mutex_trylock() atomic_dec_and_test()
100 * LOCK MB
101 * atomic_read() mutex_is_locked()
102 *
103 * LOCK should suffice since the actual taking of the lock must
104 * happen _before_ what follows.
105 */
106 if (mutex_is_locked(&anon_vma->root->mutex)) {
107 anon_vma_lock(anon_vma);
108 anon_vma_unlock(anon_vma);
109 }
110
Adrian Bunkfdd2e5f2008-10-18 20:28:38 -0700111 kmem_cache_free(anon_vma_cachep, anon_vma);
112}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700113
Linus Torvaldsdd347392011-06-17 19:05:36 -0700114static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp)
Rik van Riel5beb4932010-03-05 13:42:07 -0800115{
Linus Torvaldsdd347392011-06-17 19:05:36 -0700116 return kmem_cache_alloc(anon_vma_chain_cachep, gfp);
Rik van Riel5beb4932010-03-05 13:42:07 -0800117}
118
Namhyung Kime574b5f2010-10-26 14:22:02 -0700119static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
Rik van Riel5beb4932010-03-05 13:42:07 -0800120{
121 kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
122}
123
Linus Torvaldsd9d332e2008-10-19 10:32:20 -0700124/**
125 * anon_vma_prepare - attach an anon_vma to a memory region
126 * @vma: the memory region in question
127 *
128 * This makes sure the memory mapping described by 'vma' has
129 * an 'anon_vma' attached to it, so that we can associate the
130 * anonymous pages mapped into it with that anon_vma.
131 *
132 * The common case will be that we already have one, but if
Figo.zhang23a07902010-12-27 15:14:06 +0100133 * not we either need to find an adjacent mapping that we
Linus Torvaldsd9d332e2008-10-19 10:32:20 -0700134 * can re-use the anon_vma from (very common when the only
135 * reason for splitting a vma has been mprotect()), or we
136 * allocate a new one.
137 *
138 * Anon-vma allocations are very subtle, because we may have
139 * optimistically looked up an anon_vma in page_lock_anon_vma()
140 * and that may actually touch the spinlock even in the newly
141 * allocated vma (it depends on RCU to make sure that the
142 * anon_vma isn't actually destroyed).
143 *
144 * As a result, we need to do proper anon_vma locking even
145 * for the new allocation. At the same time, we do not want
146 * to do any locking for the common case of already having
147 * an anon_vma.
148 *
149 * This must be called with the mmap_sem held for reading.
150 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700151int anon_vma_prepare(struct vm_area_struct *vma)
152{
153 struct anon_vma *anon_vma = vma->anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -0800154 struct anon_vma_chain *avc;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700155
156 might_sleep();
157 if (unlikely(!anon_vma)) {
158 struct mm_struct *mm = vma->vm_mm;
Linus Torvaldsd9d332e2008-10-19 10:32:20 -0700159 struct anon_vma *allocated;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700160
Linus Torvaldsdd347392011-06-17 19:05:36 -0700161 avc = anon_vma_chain_alloc(GFP_KERNEL);
Rik van Riel5beb4932010-03-05 13:42:07 -0800162 if (!avc)
163 goto out_enomem;
164
Linus Torvalds1da177e2005-04-16 15:20:36 -0700165 anon_vma = find_mergeable_anon_vma(vma);
Linus Torvaldsd9d332e2008-10-19 10:32:20 -0700166 allocated = NULL;
167 if (!anon_vma) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700168 anon_vma = anon_vma_alloc();
169 if (unlikely(!anon_vma))
Rik van Riel5beb4932010-03-05 13:42:07 -0800170 goto out_enomem_free_avc;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700171 allocated = anon_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700172 }
173
Rik van Rielcba48b92010-08-09 17:18:38 -0700174 anon_vma_lock(anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700175 /* page_table_lock to protect against threads */
176 spin_lock(&mm->page_table_lock);
177 if (likely(!vma->anon_vma)) {
178 vma->anon_vma = anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -0800179 avc->anon_vma = anon_vma;
180 avc->vma = vma;
181 list_add(&avc->same_vma, &vma->anon_vma_chain);
Andrea Arcangeli26ba0cb2010-08-09 17:19:08 -0700182 list_add_tail(&avc->same_anon_vma, &anon_vma->head);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700183 allocated = NULL;
Oleg Nesterov31f2b0e2010-04-23 13:18:01 -0400184 avc = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700185 }
186 spin_unlock(&mm->page_table_lock);
Rik van Rielcba48b92010-08-09 17:18:38 -0700187 anon_vma_unlock(anon_vma);
Oleg Nesterov31f2b0e2010-04-23 13:18:01 -0400188
189 if (unlikely(allocated))
Peter Zijlstra01d8b202011-03-22 16:32:49 -0700190 put_anon_vma(allocated);
Oleg Nesterov31f2b0e2010-04-23 13:18:01 -0400191 if (unlikely(avc))
Rik van Riel5beb4932010-03-05 13:42:07 -0800192 anon_vma_chain_free(avc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700193 }
194 return 0;
Rik van Riel5beb4932010-03-05 13:42:07 -0800195
196 out_enomem_free_avc:
197 anon_vma_chain_free(avc);
198 out_enomem:
199 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700200}
201
Linus Torvaldsbb4aa392011-06-16 20:44:51 -0700202/*
203 * This is a useful helper function for locking the anon_vma root as
204 * we traverse the vma->anon_vma_chain, looping over anon_vma's that
205 * have the same vma.
206 *
207 * Such anon_vma's should have the same root, so you'd expect to see
208 * just a single mutex_lock for the whole traversal.
209 */
210static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
211{
212 struct anon_vma *new_root = anon_vma->root;
213 if (new_root != root) {
214 if (WARN_ON_ONCE(root))
215 mutex_unlock(&root->mutex);
216 root = new_root;
217 mutex_lock(&root->mutex);
218 }
219 return root;
220}
221
222static inline void unlock_anon_vma_root(struct anon_vma *root)
223{
224 if (root)
225 mutex_unlock(&root->mutex);
226}
227
Rik van Riel5beb4932010-03-05 13:42:07 -0800228static void anon_vma_chain_link(struct vm_area_struct *vma,
229 struct anon_vma_chain *avc,
230 struct anon_vma *anon_vma)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700231{
Rik van Riel5beb4932010-03-05 13:42:07 -0800232 avc->vma = vma;
233 avc->anon_vma = anon_vma;
234 list_add(&avc->same_vma, &vma->anon_vma_chain);
235
Andrea Arcangeli05759d32011-01-13 15:46:53 -0800236 /*
237 * It's critical to add new vmas to the tail of the anon_vma,
238 * see comment in huge_memory.c:__split_huge_page().
239 */
Rik van Riel5beb4932010-03-05 13:42:07 -0800240 list_add_tail(&avc->same_anon_vma, &anon_vma->head);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700241}
242
Rik van Riel5beb4932010-03-05 13:42:07 -0800243/*
244 * Attach the anon_vmas from src to dst.
245 * Returns 0 on success, -ENOMEM on failure.
246 */
247int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700248{
Rik van Riel5beb4932010-03-05 13:42:07 -0800249 struct anon_vma_chain *avc, *pavc;
Linus Torvaldsbb4aa392011-06-16 20:44:51 -0700250 struct anon_vma *root = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700251
Linus Torvalds646d87b2010-04-11 17:15:03 -0700252 list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
Linus Torvaldsbb4aa392011-06-16 20:44:51 -0700253 struct anon_vma *anon_vma;
254
Linus Torvaldsdd347392011-06-17 19:05:36 -0700255 avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
256 if (unlikely(!avc)) {
257 unlock_anon_vma_root(root);
258 root = NULL;
259 avc = anon_vma_chain_alloc(GFP_KERNEL);
260 if (!avc)
261 goto enomem_failure;
262 }
Linus Torvaldsbb4aa392011-06-16 20:44:51 -0700263 anon_vma = pavc->anon_vma;
264 root = lock_anon_vma_root(root, anon_vma);
265 anon_vma_chain_link(dst, avc, anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700266 }
Linus Torvaldsbb4aa392011-06-16 20:44:51 -0700267 unlock_anon_vma_root(root);
Rik van Riel5beb4932010-03-05 13:42:07 -0800268 return 0;
269
270 enomem_failure:
271 unlink_anon_vmas(dst);
272 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700273}
274
Rik van Riel5beb4932010-03-05 13:42:07 -0800275/*
276 * Attach vma to its own anon_vma, as well as to the anon_vmas that
277 * the corresponding VMA in the parent process is attached to.
278 * Returns 0 on success, non-zero on failure.
279 */
280int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700281{
Rik van Riel5beb4932010-03-05 13:42:07 -0800282 struct anon_vma_chain *avc;
283 struct anon_vma *anon_vma;
284
285 /* Don't bother if the parent process has no anon_vma here. */
286 if (!pvma->anon_vma)
287 return 0;
288
289 /*
290 * First, attach the new VMA to the parent VMA's anon_vmas,
291 * so rmap can find non-COWed pages in child processes.
292 */
293 if (anon_vma_clone(vma, pvma))
294 return -ENOMEM;
295
296 /* Then add our own anon_vma. */
297 anon_vma = anon_vma_alloc();
298 if (!anon_vma)
299 goto out_error;
Linus Torvaldsdd347392011-06-17 19:05:36 -0700300 avc = anon_vma_chain_alloc(GFP_KERNEL);
Rik van Riel5beb4932010-03-05 13:42:07 -0800301 if (!avc)
302 goto out_error_free_anon_vma;
Rik van Riel5c341ee2010-08-09 17:18:39 -0700303
304 /*
305 * The root anon_vma's spinlock is the lock actually used when we
306 * lock any of the anon_vmas in this anon_vma tree.
307 */
308 anon_vma->root = pvma->anon_vma->root;
Rik van Riel76545062010-08-09 17:18:41 -0700309 /*
Peter Zijlstra01d8b202011-03-22 16:32:49 -0700310 * With refcounts, an anon_vma can stay around longer than the
311 * process it belongs to. The root anon_vma needs to be pinned until
312 * this anon_vma is freed, because the lock lives in the root.
Rik van Riel76545062010-08-09 17:18:41 -0700313 */
314 get_anon_vma(anon_vma->root);
Rik van Riel5beb4932010-03-05 13:42:07 -0800315 /* Mark this anon_vma as the one where our new (COWed) pages go. */
316 vma->anon_vma = anon_vma;
Linus Torvaldsbb4aa392011-06-16 20:44:51 -0700317 anon_vma_lock(anon_vma);
Rik van Riel5c341ee2010-08-09 17:18:39 -0700318 anon_vma_chain_link(vma, avc, anon_vma);
Linus Torvaldsbb4aa392011-06-16 20:44:51 -0700319 anon_vma_unlock(anon_vma);
Rik van Riel5beb4932010-03-05 13:42:07 -0800320
321 return 0;
322
323 out_error_free_anon_vma:
Peter Zijlstra01d8b202011-03-22 16:32:49 -0700324 put_anon_vma(anon_vma);
Rik van Riel5beb4932010-03-05 13:42:07 -0800325 out_error:
Rik van Riel4946d542010-04-05 12:13:33 -0400326 unlink_anon_vmas(vma);
Rik van Riel5beb4932010-03-05 13:42:07 -0800327 return -ENOMEM;
328}
329
Rik van Riel5beb4932010-03-05 13:42:07 -0800330void unlink_anon_vmas(struct vm_area_struct *vma)
331{
332 struct anon_vma_chain *avc, *next;
Peter Zijlstraeee2acb2011-06-17 13:54:23 +0200333 struct anon_vma *root = NULL;
Rik van Riel5beb4932010-03-05 13:42:07 -0800334
Rik van Riel5c341ee2010-08-09 17:18:39 -0700335 /*
336 * Unlink each anon_vma chained to the VMA. This list is ordered
337 * from newest to oldest, ensuring the root anon_vma gets freed last.
338 */
Rik van Riel5beb4932010-03-05 13:42:07 -0800339 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
Peter Zijlstraeee2acb2011-06-17 13:54:23 +0200340 struct anon_vma *anon_vma = avc->anon_vma;
341
342 root = lock_anon_vma_root(root, anon_vma);
343 list_del(&avc->same_anon_vma);
344
345 /*
346 * Leave empty anon_vmas on the list - we'll need
347 * to free them outside the lock.
348 */
349 if (list_empty(&anon_vma->head))
350 continue;
351
352 list_del(&avc->same_vma);
353 anon_vma_chain_free(avc);
354 }
355 unlock_anon_vma_root(root);
356
357 /*
358 * Iterate the list once more, it now only contains empty and unlinked
359 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
360 * needing to acquire the anon_vma->root->mutex.
361 */
362 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
363 struct anon_vma *anon_vma = avc->anon_vma;
364
365 put_anon_vma(anon_vma);
366
Rik van Riel5beb4932010-03-05 13:42:07 -0800367 list_del(&avc->same_vma);
368 anon_vma_chain_free(avc);
369 }
370}
371
Alexey Dobriyan51cc5062008-07-25 19:45:34 -0700372static void anon_vma_ctor(void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700373{
Christoph Lametera35afb82007-05-16 22:10:57 -0700374 struct anon_vma *anon_vma = data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700375
Peter Zijlstra2b575eb2011-05-24 17:12:11 -0700376 mutex_init(&anon_vma->mutex);
Peter Zijlstra83813262011-03-22 16:32:48 -0700377 atomic_set(&anon_vma->refcount, 0);
Christoph Lametera35afb82007-05-16 22:10:57 -0700378 INIT_LIST_HEAD(&anon_vma->head);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700379}
380
381void __init anon_vma_init(void)
382{
383 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
Paul Mundt20c2df82007-07-20 10:11:58 +0900384 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor);
Rik van Riel5beb4932010-03-05 13:42:07 -0800385 anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain, SLAB_PANIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700386}
387
388/*
Peter Zijlstra6111e4c2011-05-24 17:12:08 -0700389 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
390 *
391 * Since there is no serialization what so ever against page_remove_rmap()
392 * the best this function can do is return a locked anon_vma that might
393 * have been relevant to this page.
394 *
395 * The page might have been remapped to a different anon_vma or the anon_vma
396 * returned may already be freed (and even reused).
397 *
Peter Zijlstrabc658c92011-05-29 10:33:44 +0200398 * In case it was remapped to a different anon_vma, the new anon_vma will be a
399 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
400 * ensure that any anon_vma obtained from the page will still be valid for as
401 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
402 *
Peter Zijlstra6111e4c2011-05-24 17:12:08 -0700403 * All users of this function must be very careful when walking the anon_vma
404 * chain and verify that the page in question is indeed mapped in it
405 * [ something equivalent to page_mapped_in_vma() ].
406 *
407 * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap()
408 * that the anon_vma pointer from page->mapping is valid if there is a
409 * mapcount, we can dereference the anon_vma after observing those.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700410 */
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700411struct anon_vma *page_get_anon_vma(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700412{
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700413 struct anon_vma *anon_vma = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700414 unsigned long anon_mapping;
415
416 rcu_read_lock();
Hugh Dickins80e14822009-12-14 17:59:29 -0800417 anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping);
Hugh Dickins3ca7b3c2009-12-14 17:58:57 -0800418 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700419 goto out;
420 if (!page_mapped(page))
421 goto out;
422
423 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700424 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
425 anon_vma = NULL;
426 goto out;
427 }
Hugh Dickinsf1819422010-08-25 23:12:54 -0700428
429 /*
430 * If this page is still mapped, then its anon_vma cannot have been
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700431 * freed. But if it has been unmapped, we have no security against the
432 * anon_vma structure being freed and reused (for another anon_vma:
433 * SLAB_DESTROY_BY_RCU guarantees that - so the atomic_inc_not_zero()
434 * above cannot corrupt).
Hugh Dickinsf1819422010-08-25 23:12:54 -0700435 */
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700436 if (!page_mapped(page)) {
437 put_anon_vma(anon_vma);
438 anon_vma = NULL;
439 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700440out:
441 rcu_read_unlock();
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700442
443 return anon_vma;
444}
445
Peter Zijlstra88c22082011-05-24 17:12:13 -0700446/*
447 * Similar to page_get_anon_vma() except it locks the anon_vma.
448 *
449 * Its a little more complex as it tries to keep the fast path to a single
450 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
451 * reference like with page_get_anon_vma() and then block on the mutex.
452 */
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700453struct anon_vma *page_lock_anon_vma(struct page *page)
454{
Peter Zijlstra88c22082011-05-24 17:12:13 -0700455 struct anon_vma *anon_vma = NULL;
Hugh Dickinseee0f252011-05-28 13:20:21 -0700456 struct anon_vma *root_anon_vma;
Peter Zijlstra88c22082011-05-24 17:12:13 -0700457 unsigned long anon_mapping;
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700458
Peter Zijlstra88c22082011-05-24 17:12:13 -0700459 rcu_read_lock();
460 anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping);
461 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
462 goto out;
463 if (!page_mapped(page))
464 goto out;
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700465
Peter Zijlstra88c22082011-05-24 17:12:13 -0700466 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
Hugh Dickinseee0f252011-05-28 13:20:21 -0700467 root_anon_vma = ACCESS_ONCE(anon_vma->root);
468 if (mutex_trylock(&root_anon_vma->mutex)) {
Peter Zijlstra88c22082011-05-24 17:12:13 -0700469 /*
Hugh Dickinseee0f252011-05-28 13:20:21 -0700470 * If the page is still mapped, then this anon_vma is still
471 * its anon_vma, and holding the mutex ensures that it will
Peter Zijlstrabc658c92011-05-29 10:33:44 +0200472 * not go away, see anon_vma_free().
Peter Zijlstra88c22082011-05-24 17:12:13 -0700473 */
Hugh Dickinseee0f252011-05-28 13:20:21 -0700474 if (!page_mapped(page)) {
475 mutex_unlock(&root_anon_vma->mutex);
Peter Zijlstra88c22082011-05-24 17:12:13 -0700476 anon_vma = NULL;
477 }
478 goto out;
479 }
480
481 /* trylock failed, we got to sleep */
482 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
483 anon_vma = NULL;
484 goto out;
485 }
486
487 if (!page_mapped(page)) {
488 put_anon_vma(anon_vma);
489 anon_vma = NULL;
490 goto out;
491 }
492
493 /* we pinned the anon_vma, its safe to sleep */
494 rcu_read_unlock();
495 anon_vma_lock(anon_vma);
496
497 if (atomic_dec_and_test(&anon_vma->refcount)) {
498 /*
499 * Oops, we held the last refcount, release the lock
500 * and bail -- can't simply use put_anon_vma() because
501 * we'll deadlock on the anon_vma_lock() recursion.
502 */
503 anon_vma_unlock(anon_vma);
504 __put_anon_vma(anon_vma);
505 anon_vma = NULL;
506 }
507
508 return anon_vma;
509
510out:
511 rcu_read_unlock();
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700512 return anon_vma;
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800513}
514
Andi Kleen10be22d2009-09-16 11:50:04 +0200515void page_unlock_anon_vma(struct anon_vma *anon_vma)
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800516{
Rik van Rielcba48b92010-08-09 17:18:38 -0700517 anon_vma_unlock(anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700518}
519
520/*
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -0800521 * At what user virtual address is page expected in @vma?
522 * Returns virtual address or -EFAULT if page's index/offset is not
523 * within the range mapped the @vma.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700524 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800525inline unsigned long
Linus Torvalds1da177e2005-04-16 15:20:36 -0700526vma_address(struct page *page, struct vm_area_struct *vma)
527{
528 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
529 unsigned long address;
530
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900531 if (unlikely(is_vm_hugetlb_page(vma)))
532 pgoff = page->index << huge_page_order(page_hstate(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700533 address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
534 if (unlikely(address < vma->vm_start || address >= vma->vm_end)) {
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -0800535 /* page should be within @vma mapping range */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700536 return -EFAULT;
537 }
538 return address;
539}
540
541/*
Huang Shijiebf89c8c2009-10-01 15:44:04 -0700542 * At what user virtual address is page expected in vma?
Naoya Horiguchiab941e02010-05-11 14:06:55 -0700543 * Caller should check the page is actually part of the vma.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700544 */
545unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
546{
Andrea Arcangeli21d0d442010-08-09 17:19:10 -0700547 if (PageAnon(page)) {
Hugh Dickins4829b902010-10-02 17:46:06 -0700548 struct anon_vma *page__anon_vma = page_anon_vma(page);
549 /*
550 * Note: swapoff's unuse_vma() is more efficient with this
551 * check, and needs it to match anon_vma when KSM is active.
552 */
553 if (!vma->anon_vma || !page__anon_vma ||
554 vma->anon_vma->root != page__anon_vma->root)
Andrea Arcangeli21d0d442010-08-09 17:19:10 -0700555 return -EFAULT;
556 } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) {
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800557 if (!vma->vm_file ||
558 vma->vm_file->f_mapping != page->mapping)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700559 return -EFAULT;
560 } else
561 return -EFAULT;
562 return vma_address(page, vma);
563}
564
565/*
Nikita Danilov81b40822005-05-01 08:58:36 -0700566 * Check that @page is mapped at @address into @mm.
567 *
Nick Piggin479db0b2008-08-20 14:09:18 -0700568 * If @sync is false, page_check_address may perform a racy check to avoid
569 * the page table lock when the pte is not present (helpful when reclaiming
570 * highly shared pages).
571 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -0700572 * On success returns with pte mapped and locked.
Nikita Danilov81b40822005-05-01 08:58:36 -0700573 */
Namhyung Kime9a81a82010-10-26 14:22:01 -0700574pte_t *__page_check_address(struct page *page, struct mm_struct *mm,
Nick Piggin479db0b2008-08-20 14:09:18 -0700575 unsigned long address, spinlock_t **ptlp, int sync)
Nikita Danilov81b40822005-05-01 08:58:36 -0700576{
577 pgd_t *pgd;
578 pud_t *pud;
579 pmd_t *pmd;
580 pte_t *pte;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700581 spinlock_t *ptl;
Nikita Danilov81b40822005-05-01 08:58:36 -0700582
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900583 if (unlikely(PageHuge(page))) {
584 pte = huge_pte_offset(mm, address);
585 ptl = &mm->page_table_lock;
586 goto check;
587 }
588
Nikita Danilov81b40822005-05-01 08:58:36 -0700589 pgd = pgd_offset(mm, address);
Hugh Dickinsc0718802005-10-29 18:16:31 -0700590 if (!pgd_present(*pgd))
591 return NULL;
592
593 pud = pud_offset(pgd, address);
594 if (!pud_present(*pud))
595 return NULL;
596
597 pmd = pmd_offset(pud, address);
598 if (!pmd_present(*pmd))
599 return NULL;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800600 if (pmd_trans_huge(*pmd))
601 return NULL;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700602
603 pte = pte_offset_map(pmd, address);
604 /* Make a quick check before getting the lock */
Nick Piggin479db0b2008-08-20 14:09:18 -0700605 if (!sync && !pte_present(*pte)) {
Hugh Dickinsc0718802005-10-29 18:16:31 -0700606 pte_unmap(pte);
607 return NULL;
Nikita Danilov81b40822005-05-01 08:58:36 -0700608 }
Hugh Dickinsc0718802005-10-29 18:16:31 -0700609
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700610 ptl = pte_lockptr(mm, pmd);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900611check:
Hugh Dickinsc0718802005-10-29 18:16:31 -0700612 spin_lock(ptl);
613 if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
614 *ptlp = ptl;
615 return pte;
616 }
617 pte_unmap_unlock(pte, ptl);
618 return NULL;
Nikita Danilov81b40822005-05-01 08:58:36 -0700619}
620
Nick Pigginb291f002008-10-18 20:26:44 -0700621/**
622 * page_mapped_in_vma - check whether a page is really mapped in a VMA
623 * @page: the page to test
624 * @vma: the VMA to test
625 *
626 * Returns 1 if the page is mapped into the page tables of the VMA, 0
627 * if the page is not mapped into the page tables of this VMA. Only
628 * valid for normal file or anonymous VMAs.
629 */
Andi Kleen6a460792009-09-16 11:50:15 +0200630int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
Nick Pigginb291f002008-10-18 20:26:44 -0700631{
632 unsigned long address;
633 pte_t *pte;
634 spinlock_t *ptl;
635
636 address = vma_address(page, vma);
637 if (address == -EFAULT) /* out of vma range */
638 return 0;
639 pte = page_check_address(page, vma->vm_mm, address, &ptl, 1);
640 if (!pte) /* the page is not in this mm */
641 return 0;
642 pte_unmap_unlock(pte, ptl);
643
644 return 1;
645}
646
Nikita Danilov81b40822005-05-01 08:58:36 -0700647/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700648 * Subfunctions of page_referenced: page_referenced_one called
649 * repeatedly from either page_referenced_anon or page_referenced_file.
650 */
Hugh Dickins5ad64682009-12-14 17:59:24 -0800651int page_referenced_one(struct page *page, struct vm_area_struct *vma,
652 unsigned long address, unsigned int *mapcount,
653 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700654{
655 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700656 int referenced = 0;
657
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800658 if (unlikely(PageTransHuge(page))) {
659 pmd_t *pmd;
660
661 spin_lock(&mm->page_table_lock);
Andrea Arcangeli2da28bf2011-03-11 14:58:29 -0800662 /*
663 * rmap might return false positives; we must filter
664 * these out using page_check_address_pmd().
665 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800666 pmd = page_check_address_pmd(page, mm, address,
667 PAGE_CHECK_ADDRESS_PMD_FLAG);
Andrea Arcangeli2da28bf2011-03-11 14:58:29 -0800668 if (!pmd) {
669 spin_unlock(&mm->page_table_lock);
670 goto out;
671 }
672
673 if (vma->vm_flags & VM_LOCKED) {
674 spin_unlock(&mm->page_table_lock);
675 *mapcount = 0; /* break early from loop */
676 *vm_flags |= VM_LOCKED;
677 goto out;
678 }
679
680 /* go ahead even if the pmd is pmd_trans_splitting() */
681 if (pmdp_clear_flush_young_notify(vma, address, pmd))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800682 referenced++;
683 spin_unlock(&mm->page_table_lock);
684 } else {
685 pte_t *pte;
686 spinlock_t *ptl;
687
Andrea Arcangeli2da28bf2011-03-11 14:58:29 -0800688 /*
689 * rmap might return false positives; we must filter
690 * these out using page_check_address().
691 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800692 pte = page_check_address(page, mm, address, &ptl, 0);
693 if (!pte)
694 goto out;
695
Andrea Arcangeli2da28bf2011-03-11 14:58:29 -0800696 if (vma->vm_flags & VM_LOCKED) {
697 pte_unmap_unlock(pte, ptl);
698 *mapcount = 0; /* break early from loop */
699 *vm_flags |= VM_LOCKED;
700 goto out;
701 }
702
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800703 if (ptep_clear_flush_young_notify(vma, address, pte)) {
704 /*
705 * Don't treat a reference through a sequentially read
706 * mapping as such. If the page has been used in
707 * another mapping, we will catch it; if this other
708 * mapping is already gone, the unmap path will have
709 * set PG_referenced or activated the page.
710 */
711 if (likely(!VM_SequentialReadHint(vma)))
712 referenced++;
713 }
714 pte_unmap_unlock(pte, ptl);
715 }
716
Andrea Arcangeli2da28bf2011-03-11 14:58:29 -0800717 /* Pretend the page is referenced if the task has the
718 swap token and is in the middle of a page fault. */
719 if (mm != current->mm && has_swap_token(mm) &&
720 rwsem_is_locked(&mm->mmap_sem))
721 referenced++;
722
Hugh Dickinsc0718802005-10-29 18:16:31 -0700723 (*mapcount)--;
Huang Shijie273f0472009-12-14 17:58:51 -0800724
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700725 if (referenced)
726 *vm_flags |= vma->vm_flags;
Huang Shijie273f0472009-12-14 17:58:51 -0800727out:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700728 return referenced;
729}
730
Balbir Singhbed71612008-02-07 00:14:01 -0800731static int page_referenced_anon(struct page *page,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700732 struct mem_cgroup *mem_cont,
733 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700734{
735 unsigned int mapcount;
736 struct anon_vma *anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -0800737 struct anon_vma_chain *avc;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700738 int referenced = 0;
739
740 anon_vma = page_lock_anon_vma(page);
741 if (!anon_vma)
742 return referenced;
743
744 mapcount = page_mapcount(page);
Rik van Riel5beb4932010-03-05 13:42:07 -0800745 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
746 struct vm_area_struct *vma = avc->vma;
Hugh Dickins1cb17292009-12-14 17:59:01 -0800747 unsigned long address = vma_address(page, vma);
748 if (address == -EFAULT)
749 continue;
Balbir Singhbed71612008-02-07 00:14:01 -0800750 /*
751 * If we are reclaiming on behalf of a cgroup, skip
752 * counting on behalf of references from different
753 * cgroups
754 */
Hugh Dickinsbd845e32008-03-04 14:29:01 -0800755 if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
Balbir Singhbed71612008-02-07 00:14:01 -0800756 continue;
Hugh Dickins1cb17292009-12-14 17:59:01 -0800757 referenced += page_referenced_one(page, vma, address,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700758 &mapcount, vm_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700759 if (!mapcount)
760 break;
761 }
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800762
763 page_unlock_anon_vma(anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700764 return referenced;
765}
766
767/**
768 * page_referenced_file - referenced check for object-based rmap
769 * @page: the page we're checking references on.
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700770 * @mem_cont: target memory controller
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700771 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
Linus Torvalds1da177e2005-04-16 15:20:36 -0700772 *
773 * For an object-based mapped page, find all the places it is mapped and
774 * check/clear the referenced flag. This is done by following the page->mapping
775 * pointer, then walking the chain of vmas it holds. It returns the number
776 * of references it found.
777 *
778 * This function is only called from page_referenced for object-based pages.
779 */
Balbir Singhbed71612008-02-07 00:14:01 -0800780static int page_referenced_file(struct page *page,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700781 struct mem_cgroup *mem_cont,
782 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700783{
784 unsigned int mapcount;
785 struct address_space *mapping = page->mapping;
786 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
787 struct vm_area_struct *vma;
788 struct prio_tree_iter iter;
789 int referenced = 0;
790
791 /*
792 * The caller's checks on page->mapping and !PageAnon have made
793 * sure that this is a file page: the check for page->mapping
794 * excludes the case just before it gets set on an anon page.
795 */
796 BUG_ON(PageAnon(page));
797
798 /*
799 * The page lock not only makes sure that page->mapping cannot
800 * suddenly be NULLified by truncation, it makes sure that the
801 * structure at mapping cannot be freed and reused yet,
Peter Zijlstra3d48ae42011-05-24 17:12:06 -0700802 * so we can safely take mapping->i_mmap_mutex.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700803 */
804 BUG_ON(!PageLocked(page));
805
Peter Zijlstra3d48ae42011-05-24 17:12:06 -0700806 mutex_lock(&mapping->i_mmap_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700807
808 /*
Peter Zijlstra3d48ae42011-05-24 17:12:06 -0700809 * i_mmap_mutex does not stabilize mapcount at all, but mapcount
Linus Torvalds1da177e2005-04-16 15:20:36 -0700810 * is more likely to be accurate if we note it after spinning.
811 */
812 mapcount = page_mapcount(page);
813
814 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
Hugh Dickins1cb17292009-12-14 17:59:01 -0800815 unsigned long address = vma_address(page, vma);
816 if (address == -EFAULT)
817 continue;
Balbir Singhbed71612008-02-07 00:14:01 -0800818 /*
819 * If we are reclaiming on behalf of a cgroup, skip
820 * counting on behalf of references from different
821 * cgroups
822 */
Hugh Dickinsbd845e32008-03-04 14:29:01 -0800823 if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
Balbir Singhbed71612008-02-07 00:14:01 -0800824 continue;
Hugh Dickins1cb17292009-12-14 17:59:01 -0800825 referenced += page_referenced_one(page, vma, address,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700826 &mapcount, vm_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700827 if (!mapcount)
828 break;
829 }
830
Peter Zijlstra3d48ae42011-05-24 17:12:06 -0700831 mutex_unlock(&mapping->i_mmap_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700832 return referenced;
833}
834
835/**
836 * page_referenced - test if the page was referenced
837 * @page: the page to test
838 * @is_locked: caller holds lock on the page
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700839 * @mem_cont: target memory controller
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700840 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
Linus Torvalds1da177e2005-04-16 15:20:36 -0700841 *
842 * Quick test_and_clear_referenced for all mappings to a page,
843 * returns the number of ptes which referenced the page.
844 */
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700845int page_referenced(struct page *page,
846 int is_locked,
847 struct mem_cgroup *mem_cont,
848 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700849{
850 int referenced = 0;
Hugh Dickins5ad64682009-12-14 17:59:24 -0800851 int we_locked = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700852
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700853 *vm_flags = 0;
Hugh Dickins3ca7b3c2009-12-14 17:58:57 -0800854 if (page_mapped(page) && page_rmapping(page)) {
Hugh Dickins5ad64682009-12-14 17:59:24 -0800855 if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
856 we_locked = trylock_page(page);
857 if (!we_locked) {
858 referenced++;
859 goto out;
860 }
861 }
862 if (unlikely(PageKsm(page)))
863 referenced += page_referenced_ksm(page, mem_cont,
864 vm_flags);
865 else if (PageAnon(page))
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700866 referenced += page_referenced_anon(page, mem_cont,
867 vm_flags);
Hugh Dickins5ad64682009-12-14 17:59:24 -0800868 else if (page->mapping)
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700869 referenced += page_referenced_file(page, mem_cont,
870 vm_flags);
Hugh Dickins5ad64682009-12-14 17:59:24 -0800871 if (we_locked)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700872 unlock_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700873 }
Hugh Dickins5ad64682009-12-14 17:59:24 -0800874out:
Martin Schwidefsky2d425522011-05-23 10:24:39 +0200875 if (page_test_and_clear_young(page_to_pfn(page)))
Christian Borntraeger5b7baf02008-03-25 18:47:12 +0100876 referenced++;
877
Linus Torvalds1da177e2005-04-16 15:20:36 -0700878 return referenced;
879}
880
Hugh Dickins1cb17292009-12-14 17:59:01 -0800881static int page_mkclean_one(struct page *page, struct vm_area_struct *vma,
882 unsigned long address)
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700883{
884 struct mm_struct *mm = vma->vm_mm;
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100885 pte_t *pte;
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700886 spinlock_t *ptl;
887 int ret = 0;
888
Nick Piggin479db0b2008-08-20 14:09:18 -0700889 pte = page_check_address(page, mm, address, &ptl, 1);
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700890 if (!pte)
891 goto out;
892
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100893 if (pte_dirty(*pte) || pte_write(*pte)) {
894 pte_t entry;
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700895
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100896 flush_cache_page(vma, address, pte_pfn(*pte));
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700897 entry = ptep_clear_flush_notify(vma, address, pte);
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100898 entry = pte_wrprotect(entry);
899 entry = pte_mkclean(entry);
Al Virod6e88e62006-12-29 16:48:35 -0800900 set_pte_at(mm, address, pte, entry);
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100901 ret = 1;
902 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700903
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700904 pte_unmap_unlock(pte, ptl);
905out:
906 return ret;
907}
908
909static int page_mkclean_file(struct address_space *mapping, struct page *page)
910{
911 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
912 struct vm_area_struct *vma;
913 struct prio_tree_iter iter;
914 int ret = 0;
915
916 BUG_ON(PageAnon(page));
917
Peter Zijlstra3d48ae42011-05-24 17:12:06 -0700918 mutex_lock(&mapping->i_mmap_mutex);
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700919 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
Hugh Dickins1cb17292009-12-14 17:59:01 -0800920 if (vma->vm_flags & VM_SHARED) {
921 unsigned long address = vma_address(page, vma);
922 if (address == -EFAULT)
923 continue;
924 ret += page_mkclean_one(page, vma, address);
925 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700926 }
Peter Zijlstra3d48ae42011-05-24 17:12:06 -0700927 mutex_unlock(&mapping->i_mmap_mutex);
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700928 return ret;
929}
930
931int page_mkclean(struct page *page)
932{
933 int ret = 0;
934
935 BUG_ON(!PageLocked(page));
936
937 if (page_mapped(page)) {
938 struct address_space *mapping = page_mapping(page);
Christian Borntraegerce7e9fa2007-11-20 11:13:36 +0100939 if (mapping) {
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700940 ret = page_mkclean_file(mapping, page);
Martin Schwidefsky2d425522011-05-23 10:24:39 +0200941 if (page_test_and_clear_dirty(page_to_pfn(page), 1))
Christian Borntraegerce7e9fa2007-11-20 11:13:36 +0100942 ret = 1;
Martin Schwidefsky6c210482007-04-27 16:01:57 +0200943 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700944 }
945
946 return ret;
947}
Jaya Kumar60b59be2007-05-08 00:37:37 -0700948EXPORT_SYMBOL_GPL(page_mkclean);
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700949
Linus Torvalds1da177e2005-04-16 15:20:36 -0700950/**
Rik van Rielc44b6742010-03-05 13:42:09 -0800951 * page_move_anon_rmap - move a page to our anon_vma
952 * @page: the page to move to our anon_vma
953 * @vma: the vma the page belongs to
954 * @address: the user virtual address mapped
955 *
956 * When a page belongs exclusively to one process after a COW event,
957 * that page can be moved into the anon_vma that belongs to just that
958 * process, so the rmap code will not search the parent or sibling
959 * processes.
960 */
961void page_move_anon_rmap(struct page *page,
962 struct vm_area_struct *vma, unsigned long address)
963{
964 struct anon_vma *anon_vma = vma->anon_vma;
965
966 VM_BUG_ON(!PageLocked(page));
967 VM_BUG_ON(!anon_vma);
968 VM_BUG_ON(page->index != linear_page_index(vma, address));
969
970 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
971 page->mapping = (struct address_space *) anon_vma;
972}
973
974/**
Andi Kleen4e1c1972010-09-22 12:43:56 +0200975 * __page_set_anon_rmap - set up new anonymous rmap
976 * @page: Page to add to rmap
977 * @vma: VM area to add page to.
978 * @address: User virtual address of the mapping
Rik van Riele8a03fe2010-04-14 17:59:28 -0400979 * @exclusive: the page is exclusively owned by the current process
Nick Piggin9617d952006-01-06 00:11:12 -0800980 */
981static void __page_set_anon_rmap(struct page *page,
Rik van Riele8a03fe2010-04-14 17:59:28 -0400982 struct vm_area_struct *vma, unsigned long address, int exclusive)
Nick Piggin9617d952006-01-06 00:11:12 -0800983{
Rik van Riele8a03fe2010-04-14 17:59:28 -0400984 struct anon_vma *anon_vma = vma->anon_vma;
Nick Piggin9617d952006-01-06 00:11:12 -0800985
Rik van Riele8a03fe2010-04-14 17:59:28 -0400986 BUG_ON(!anon_vma);
Linus Torvaldsea900022010-04-12 12:44:29 -0700987
Andi Kleen4e1c1972010-09-22 12:43:56 +0200988 if (PageAnon(page))
989 return;
990
Linus Torvaldsea900022010-04-12 12:44:29 -0700991 /*
Rik van Riele8a03fe2010-04-14 17:59:28 -0400992 * If the page isn't exclusively mapped into this vma,
993 * we must use the _oldest_ possible anon_vma for the
994 * page mapping!
Linus Torvaldsea900022010-04-12 12:44:29 -0700995 */
Andi Kleen4e1c1972010-09-22 12:43:56 +0200996 if (!exclusive)
Andrea Arcangeli288468c2010-08-09 17:19:09 -0700997 anon_vma = anon_vma->root;
Linus Torvaldsea900022010-04-12 12:44:29 -0700998
Nick Piggin9617d952006-01-06 00:11:12 -0800999 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1000 page->mapping = (struct address_space *) anon_vma;
Nick Piggin9617d952006-01-06 00:11:12 -08001001 page->index = linear_page_index(vma, address);
Nick Piggin9617d952006-01-06 00:11:12 -08001002}
1003
1004/**
Randy Dunlap43d8eac2008-03-19 17:00:43 -07001005 * __page_check_anon_rmap - sanity check anonymous rmap addition
Nick Pigginc97a9e12007-05-16 22:11:21 -07001006 * @page: the page to add the mapping to
1007 * @vma: the vm area in which the mapping is added
1008 * @address: the user virtual address mapped
1009 */
1010static void __page_check_anon_rmap(struct page *page,
1011 struct vm_area_struct *vma, unsigned long address)
1012{
1013#ifdef CONFIG_DEBUG_VM
1014 /*
1015 * The page's anon-rmap details (mapping and index) are guaranteed to
1016 * be set up correctly at this point.
1017 *
1018 * We have exclusion against page_add_anon_rmap because the caller
1019 * always holds the page locked, except if called from page_dup_rmap,
1020 * in which case the page is already known to be setup.
1021 *
1022 * We have exclusion against page_add_new_anon_rmap because those pages
1023 * are initially only visible via the pagetables, and the pte is locked
1024 * over the call to page_add_new_anon_rmap.
1025 */
Andrea Arcangeli44ab57a2010-08-09 17:19:10 -07001026 BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root);
Nick Pigginc97a9e12007-05-16 22:11:21 -07001027 BUG_ON(page->index != linear_page_index(vma, address));
1028#endif
1029}
1030
1031/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07001032 * page_add_anon_rmap - add pte mapping to an anonymous page
1033 * @page: the page to add the mapping to
1034 * @vma: the vm area in which the mapping is added
1035 * @address: the user virtual address mapped
1036 *
Hugh Dickins5ad64682009-12-14 17:59:24 -08001037 * The caller needs to hold the pte lock, and the page must be locked in
Hugh Dickins80e14822009-12-14 17:59:29 -08001038 * the anon_vma case: to serialize mapping,index checking after setting,
1039 * and to ensure that PageAnon is not being upgraded racily to PageKsm
1040 * (but PageKsm is never downgraded to PageAnon).
Linus Torvalds1da177e2005-04-16 15:20:36 -07001041 */
1042void page_add_anon_rmap(struct page *page,
1043 struct vm_area_struct *vma, unsigned long address)
1044{
Rik van Rielad8c2ee2010-08-09 17:19:48 -07001045 do_page_add_anon_rmap(page, vma, address, 0);
1046}
1047
1048/*
1049 * Special version of the above for do_swap_page, which often runs
1050 * into pages that are exclusively owned by the current process.
1051 * Everybody else should continue to use page_add_anon_rmap above.
1052 */
1053void do_page_add_anon_rmap(struct page *page,
1054 struct vm_area_struct *vma, unsigned long address, int exclusive)
1055{
Hugh Dickins5ad64682009-12-14 17:59:24 -08001056 int first = atomic_inc_and_test(&page->_mapcount);
Andrea Arcangeli79134172011-01-13 15:46:58 -08001057 if (first) {
1058 if (!PageTransHuge(page))
1059 __inc_zone_page_state(page, NR_ANON_PAGES);
1060 else
1061 __inc_zone_page_state(page,
1062 NR_ANON_TRANSPARENT_HUGEPAGES);
1063 }
Hugh Dickins5ad64682009-12-14 17:59:24 -08001064 if (unlikely(PageKsm(page)))
1065 return;
1066
Nick Pigginc97a9e12007-05-16 22:11:21 -07001067 VM_BUG_ON(!PageLocked(page));
Hugh Dickins5dbe0af2011-05-28 13:17:04 -07001068 /* address might be in next vma when migration races vma_adjust */
Hugh Dickins5ad64682009-12-14 17:59:24 -08001069 if (first)
Rik van Rielad8c2ee2010-08-09 17:19:48 -07001070 __page_set_anon_rmap(page, vma, address, exclusive);
KAMEZAWA Hiroyuki69029cd2008-07-25 01:47:14 -07001071 else
Nick Pigginc97a9e12007-05-16 22:11:21 -07001072 __page_check_anon_rmap(page, vma, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001073}
1074
Randy Dunlap43d8eac2008-03-19 17:00:43 -07001075/**
Nick Piggin9617d952006-01-06 00:11:12 -08001076 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1077 * @page: the page to add the mapping to
1078 * @vma: the vm area in which the mapping is added
1079 * @address: the user virtual address mapped
1080 *
1081 * Same as page_add_anon_rmap but must only be called on *new* pages.
1082 * This means the inc-and-test can be bypassed.
Nick Pigginc97a9e12007-05-16 22:11:21 -07001083 * Page does not have to be locked.
Nick Piggin9617d952006-01-06 00:11:12 -08001084 */
1085void page_add_new_anon_rmap(struct page *page,
1086 struct vm_area_struct *vma, unsigned long address)
1087{
Hugh Dickinsb5934c52009-01-06 14:39:25 -08001088 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
Hugh Dickinscbf84b72009-01-06 14:39:27 -08001089 SetPageSwapBacked(page);
1090 atomic_set(&page->_mapcount, 0); /* increment count (starts at -1) */
Andrea Arcangeli79134172011-01-13 15:46:58 -08001091 if (!PageTransHuge(page))
1092 __inc_zone_page_state(page, NR_ANON_PAGES);
1093 else
1094 __inc_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
Rik van Riele8a03fe2010-04-14 17:59:28 -04001095 __page_set_anon_rmap(page, vma, address, 1);
Hugh Dickinsb5934c52009-01-06 14:39:25 -08001096 if (page_evictable(page, vma))
Hugh Dickinscbf84b72009-01-06 14:39:27 -08001097 lru_cache_add_lru(page, LRU_ACTIVE_ANON);
Hugh Dickinsb5934c52009-01-06 14:39:25 -08001098 else
1099 add_page_to_unevictable_list(page);
Nick Piggin9617d952006-01-06 00:11:12 -08001100}
1101
Linus Torvalds1da177e2005-04-16 15:20:36 -07001102/**
1103 * page_add_file_rmap - add pte mapping to a file page
1104 * @page: the page to add the mapping to
1105 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -07001106 * The caller needs to hold the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001107 */
1108void page_add_file_rmap(struct page *page)
1109{
Balbir Singhd69b0422009-06-17 16:26:34 -07001110 if (atomic_inc_and_test(&page->_mapcount)) {
Christoph Lameter65ba55f2006-06-30 01:55:34 -07001111 __inc_zone_page_state(page, NR_FILE_MAPPED);
Greg Thelen2a7106f2011-01-13 15:47:37 -08001112 mem_cgroup_inc_page_stat(page, MEMCG_NR_FILE_MAPPED);
Balbir Singhd69b0422009-06-17 16:26:34 -07001113 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001114}
1115
1116/**
1117 * page_remove_rmap - take down pte mapping from a page
1118 * @page: page to remove mapping from
1119 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -07001120 * The caller needs to hold the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001121 */
Hugh Dickinsedc315f2009-01-06 14:40:11 -08001122void page_remove_rmap(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001123{
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -07001124 /* page still mapped by someone else? */
1125 if (!atomic_add_negative(-1, &page->_mapcount))
1126 return;
1127
1128 /*
1129 * Now that the last pte has gone, s390 must transfer dirty
1130 * flag from storage key to struct page. We can usually skip
1131 * this if the page is anon, so about to be freed; but perhaps
1132 * not if it's in swapcache - there might be another pte slot
1133 * containing the swap entry, but page not yet written to swap.
1134 */
Martin Schwidefsky2d425522011-05-23 10:24:39 +02001135 if ((!PageAnon(page) || PageSwapCache(page)) &&
1136 page_test_and_clear_dirty(page_to_pfn(page), 1))
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -07001137 set_page_dirty(page);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001138 /*
1139 * Hugepages are not counted in NR_ANON_PAGES nor NR_FILE_MAPPED
1140 * and not charged by memcg for now.
1141 */
1142 if (unlikely(PageHuge(page)))
1143 return;
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -07001144 if (PageAnon(page)) {
1145 mem_cgroup_uncharge_page(page);
Andrea Arcangeli79134172011-01-13 15:46:58 -08001146 if (!PageTransHuge(page))
1147 __dec_zone_page_state(page, NR_ANON_PAGES);
1148 else
1149 __dec_zone_page_state(page,
1150 NR_ANON_TRANSPARENT_HUGEPAGES);
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -07001151 } else {
1152 __dec_zone_page_state(page, NR_FILE_MAPPED);
Greg Thelen2a7106f2011-01-13 15:47:37 -08001153 mem_cgroup_dec_page_stat(page, MEMCG_NR_FILE_MAPPED);
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -07001154 }
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -07001155 /*
1156 * It would be tidy to reset the PageAnon mapping here,
1157 * but that might overwrite a racing page_add_anon_rmap
1158 * which increments mapcount after us but sets mapping
1159 * before us: so leave the reset to free_hot_cold_page,
1160 * and remember that it's only reliable while mapped.
1161 * Leaving it set also helps swapoff to reinstate ptes
1162 * faster for those pages still in swapcache.
1163 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001164}
1165
1166/*
1167 * Subfunctions of try_to_unmap: try_to_unmap_one called
1168 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
1169 */
Hugh Dickins5ad64682009-12-14 17:59:24 -08001170int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
1171 unsigned long address, enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001172{
1173 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001174 pte_t *pte;
1175 pte_t pteval;
Hugh Dickinsc0718802005-10-29 18:16:31 -07001176 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001177 int ret = SWAP_AGAIN;
1178
Nick Piggin479db0b2008-08-20 14:09:18 -07001179 pte = page_check_address(page, mm, address, &ptl, 0);
Hugh Dickinsc0718802005-10-29 18:16:31 -07001180 if (!pte)
Nikita Danilov81b40822005-05-01 08:58:36 -07001181 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001182
1183 /*
1184 * If the page is mlock()d, we cannot swap it out.
1185 * If it's recently referenced (perhaps page_referenced
1186 * skipped over this mm) then we should reactivate it.
1187 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001188 if (!(flags & TTU_IGNORE_MLOCK)) {
KOSAKI Motohirocaed0f42009-12-14 17:59:45 -08001189 if (vma->vm_flags & VM_LOCKED)
1190 goto out_mlock;
1191
Hugh Dickinsaf8e3352009-12-14 17:58:59 -08001192 if (TTU_ACTION(flags) == TTU_MUNLOCK)
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001193 goto out_unmap;
Andi Kleen14fa31b2009-09-16 11:50:10 +02001194 }
1195 if (!(flags & TTU_IGNORE_ACCESS)) {
Nick Pigginb291f002008-10-18 20:26:44 -07001196 if (ptep_clear_flush_young_notify(vma, address, pte)) {
1197 ret = SWAP_FAIL;
1198 goto out_unmap;
1199 }
1200 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001201
Linus Torvalds1da177e2005-04-16 15:20:36 -07001202 /* Nuke the page table entry. */
1203 flush_cache_page(vma, address, page_to_pfn(page));
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001204 pteval = ptep_clear_flush_notify(vma, address, pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001205
1206 /* Move the dirty bit to the physical page now the pte is gone. */
1207 if (pte_dirty(pteval))
1208 set_page_dirty(page);
1209
Hugh Dickins365e9c872005-10-29 18:16:18 -07001210 /* Update high watermark before we lower rss */
1211 update_hiwater_rss(mm);
1212
Andi Kleen888b9f72009-09-16 11:50:11 +02001213 if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
1214 if (PageAnon(page))
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001215 dec_mm_counter(mm, MM_ANONPAGES);
Andi Kleen888b9f72009-09-16 11:50:11 +02001216 else
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001217 dec_mm_counter(mm, MM_FILEPAGES);
Andi Kleen888b9f72009-09-16 11:50:11 +02001218 set_pte_at(mm, address, pte,
1219 swp_entry_to_pte(make_hwpoison_entry(page)));
1220 } else if (PageAnon(page)) {
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07001221 swp_entry_t entry = { .val = page_private(page) };
Christoph Lameter06972122006-06-23 02:03:35 -07001222
1223 if (PageSwapCache(page)) {
1224 /*
1225 * Store the swap location in the pte.
1226 * See handle_pte_fault() ...
1227 */
Hugh Dickins570a3352009-12-14 17:58:46 -08001228 if (swap_duplicate(entry) < 0) {
1229 set_pte_at(mm, address, pte, pteval);
1230 ret = SWAP_FAIL;
1231 goto out_unmap;
1232 }
Christoph Lameter06972122006-06-23 02:03:35 -07001233 if (list_empty(&mm->mmlist)) {
1234 spin_lock(&mmlist_lock);
1235 if (list_empty(&mm->mmlist))
1236 list_add(&mm->mmlist, &init_mm.mmlist);
1237 spin_unlock(&mmlist_lock);
1238 }
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001239 dec_mm_counter(mm, MM_ANONPAGES);
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08001240 inc_mm_counter(mm, MM_SWAPENTS);
KOSAKI Motohiro64cdd542009-01-06 14:39:16 -08001241 } else if (PAGE_MIGRATION) {
Christoph Lameter06972122006-06-23 02:03:35 -07001242 /*
1243 * Store the pfn of the page in a special migration
1244 * pte. do_swap_page() will wait until the migration
1245 * pte is removed and then restart fault handling.
1246 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001247 BUG_ON(TTU_ACTION(flags) != TTU_MIGRATION);
Christoph Lameter06972122006-06-23 02:03:35 -07001248 entry = make_migration_entry(page, pte_write(pteval));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001249 }
1250 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
1251 BUG_ON(pte_file(*pte));
Andi Kleen14fa31b2009-09-16 11:50:10 +02001252 } else if (PAGE_MIGRATION && (TTU_ACTION(flags) == TTU_MIGRATION)) {
Christoph Lameter04e62a22006-06-23 02:03:38 -07001253 /* Establish migration entry for a file page */
1254 swp_entry_t entry;
1255 entry = make_migration_entry(page, pte_write(pteval));
1256 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
1257 } else
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001258 dec_mm_counter(mm, MM_FILEPAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001259
Hugh Dickinsedc315f2009-01-06 14:40:11 -08001260 page_remove_rmap(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001261 page_cache_release(page);
1262
1263out_unmap:
Hugh Dickinsc0718802005-10-29 18:16:31 -07001264 pte_unmap_unlock(pte, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001265out:
1266 return ret;
KOSAKI Motohirocaed0f42009-12-14 17:59:45 -08001267
1268out_mlock:
1269 pte_unmap_unlock(pte, ptl);
1270
1271
1272 /*
1273 * We need mmap_sem locking, Otherwise VM_LOCKED check makes
1274 * unstable result and race. Plus, We can't wait here because
Peter Zijlstra2b575eb2011-05-24 17:12:11 -07001275 * we now hold anon_vma->mutex or mapping->i_mmap_mutex.
KOSAKI Motohirocaed0f42009-12-14 17:59:45 -08001276 * if trylock failed, the page remain in evictable lru and later
1277 * vmscan could retry to move the page to unevictable lru if the
1278 * page is actually mlocked.
1279 */
1280 if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
1281 if (vma->vm_flags & VM_LOCKED) {
1282 mlock_vma_page(page);
1283 ret = SWAP_MLOCK;
1284 }
1285 up_read(&vma->vm_mm->mmap_sem);
1286 }
1287 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001288}
1289
1290/*
1291 * objrmap doesn't work for nonlinear VMAs because the assumption that
1292 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
1293 * Consequently, given a particular page and its ->index, we cannot locate the
1294 * ptes which are mapping that page without an exhaustive linear search.
1295 *
1296 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
1297 * maps the file to which the target page belongs. The ->vm_private_data field
1298 * holds the current cursor into that scan. Successive searches will circulate
1299 * around the vma's virtual address space.
1300 *
1301 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
1302 * more scanning pressure is placed against them as well. Eventually pages
1303 * will become fully unmapped and are eligible for eviction.
1304 *
1305 * For very sparsely populated VMAs this is a little inefficient - chances are
1306 * there there won't be many ptes located within the scan cluster. In this case
1307 * maybe we could scan further - to the end of the pte page, perhaps.
Nick Pigginb291f002008-10-18 20:26:44 -07001308 *
1309 * Mlocked pages: check VM_LOCKED under mmap_sem held for read, if we can
1310 * acquire it without blocking. If vma locked, mlock the pages in the cluster,
1311 * rather than unmapping them. If we encounter the "check_page" that vmscan is
1312 * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001313 */
1314#define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
1315#define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
1316
Nick Pigginb291f002008-10-18 20:26:44 -07001317static int try_to_unmap_cluster(unsigned long cursor, unsigned int *mapcount,
1318 struct vm_area_struct *vma, struct page *check_page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001319{
1320 struct mm_struct *mm = vma->vm_mm;
1321 pgd_t *pgd;
1322 pud_t *pud;
1323 pmd_t *pmd;
Hugh Dickinsc0718802005-10-29 18:16:31 -07001324 pte_t *pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001325 pte_t pteval;
Hugh Dickinsc0718802005-10-29 18:16:31 -07001326 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001327 struct page *page;
1328 unsigned long address;
1329 unsigned long end;
Nick Pigginb291f002008-10-18 20:26:44 -07001330 int ret = SWAP_AGAIN;
1331 int locked_vma = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001332
Linus Torvalds1da177e2005-04-16 15:20:36 -07001333 address = (vma->vm_start + cursor) & CLUSTER_MASK;
1334 end = address + CLUSTER_SIZE;
1335 if (address < vma->vm_start)
1336 address = vma->vm_start;
1337 if (end > vma->vm_end)
1338 end = vma->vm_end;
1339
1340 pgd = pgd_offset(mm, address);
1341 if (!pgd_present(*pgd))
Nick Pigginb291f002008-10-18 20:26:44 -07001342 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001343
1344 pud = pud_offset(pgd, address);
1345 if (!pud_present(*pud))
Nick Pigginb291f002008-10-18 20:26:44 -07001346 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001347
1348 pmd = pmd_offset(pud, address);
1349 if (!pmd_present(*pmd))
Nick Pigginb291f002008-10-18 20:26:44 -07001350 return ret;
1351
1352 /*
Hugh Dickinsaf8e3352009-12-14 17:58:59 -08001353 * If we can acquire the mmap_sem for read, and vma is VM_LOCKED,
Nick Pigginb291f002008-10-18 20:26:44 -07001354 * keep the sem while scanning the cluster for mlocking pages.
1355 */
Hugh Dickinsaf8e3352009-12-14 17:58:59 -08001356 if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
Nick Pigginb291f002008-10-18 20:26:44 -07001357 locked_vma = (vma->vm_flags & VM_LOCKED);
1358 if (!locked_vma)
1359 up_read(&vma->vm_mm->mmap_sem); /* don't need it */
1360 }
Hugh Dickinsc0718802005-10-29 18:16:31 -07001361
1362 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001363
Hugh Dickins365e9c872005-10-29 18:16:18 -07001364 /* Update high watermark before we lower rss */
1365 update_hiwater_rss(mm);
1366
Hugh Dickinsc0718802005-10-29 18:16:31 -07001367 for (; address < end; pte++, address += PAGE_SIZE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001368 if (!pte_present(*pte))
1369 continue;
Linus Torvalds6aab3412005-11-28 14:34:23 -08001370 page = vm_normal_page(vma, address, *pte);
1371 BUG_ON(!page || PageAnon(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001372
Nick Pigginb291f002008-10-18 20:26:44 -07001373 if (locked_vma) {
1374 mlock_vma_page(page); /* no-op if already mlocked */
1375 if (page == check_page)
1376 ret = SWAP_MLOCK;
1377 continue; /* don't unmap */
1378 }
1379
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001380 if (ptep_clear_flush_young_notify(vma, address, pte))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001381 continue;
1382
1383 /* Nuke the page table entry. */
Ben Collinseca35132005-11-29 11:45:26 -08001384 flush_cache_page(vma, address, pte_pfn(*pte));
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001385 pteval = ptep_clear_flush_notify(vma, address, pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001386
1387 /* If nonlinear, store the file page offset in the pte. */
1388 if (page->index != linear_page_index(vma, address))
1389 set_pte_at(mm, address, pte, pgoff_to_pte(page->index));
1390
1391 /* Move the dirty bit to the physical page now the pte is gone. */
1392 if (pte_dirty(pteval))
1393 set_page_dirty(page);
1394
Hugh Dickinsedc315f2009-01-06 14:40:11 -08001395 page_remove_rmap(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001396 page_cache_release(page);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001397 dec_mm_counter(mm, MM_FILEPAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001398 (*mapcount)--;
1399 }
Hugh Dickinsc0718802005-10-29 18:16:31 -07001400 pte_unmap_unlock(pte - 1, ptl);
Nick Pigginb291f002008-10-18 20:26:44 -07001401 if (locked_vma)
1402 up_read(&vma->vm_mm->mmap_sem);
1403 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001404}
1405
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001406bool is_vma_temporary_stack(struct vm_area_struct *vma)
Mel Gormana8bef8ff2010-05-24 14:32:24 -07001407{
1408 int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP);
1409
1410 if (!maybe_stack)
1411 return false;
1412
1413 if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) ==
1414 VM_STACK_INCOMPLETE_SETUP)
1415 return true;
1416
1417 return false;
1418}
1419
Nick Pigginb291f002008-10-18 20:26:44 -07001420/**
1421 * try_to_unmap_anon - unmap or unlock anonymous page using the object-based
1422 * rmap method
1423 * @page: the page to unmap/unlock
Huang Shijie8051be52009-12-14 17:58:50 -08001424 * @flags: action and flags
Nick Pigginb291f002008-10-18 20:26:44 -07001425 *
1426 * Find all the mappings of a page using the mapping pointer and the vma chains
1427 * contained in the anon_vma struct it points to.
1428 *
1429 * This function is only called from try_to_unmap/try_to_munlock for
1430 * anonymous pages.
1431 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1432 * where the page was found will be held for write. So, we won't recheck
1433 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1434 * 'LOCKED.
1435 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001436static int try_to_unmap_anon(struct page *page, enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001437{
1438 struct anon_vma *anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -08001439 struct anon_vma_chain *avc;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001440 int ret = SWAP_AGAIN;
Nick Pigginb291f002008-10-18 20:26:44 -07001441
Linus Torvalds1da177e2005-04-16 15:20:36 -07001442 anon_vma = page_lock_anon_vma(page);
1443 if (!anon_vma)
1444 return ret;
1445
Rik van Riel5beb4932010-03-05 13:42:07 -08001446 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1447 struct vm_area_struct *vma = avc->vma;
Mel Gormana8bef8ff2010-05-24 14:32:24 -07001448 unsigned long address;
1449
1450 /*
1451 * During exec, a temporary VMA is setup and later moved.
1452 * The VMA is moved under the anon_vma lock but not the
1453 * page tables leading to a race where migration cannot
1454 * find the migration ptes. Rather than increasing the
1455 * locking requirements of exec(), migration skips
1456 * temporary VMAs until after exec() completes.
1457 */
1458 if (PAGE_MIGRATION && (flags & TTU_MIGRATION) &&
1459 is_vma_temporary_stack(vma))
1460 continue;
1461
1462 address = vma_address(page, vma);
Hugh Dickins1cb17292009-12-14 17:59:01 -08001463 if (address == -EFAULT)
1464 continue;
1465 ret = try_to_unmap_one(page, vma, address, flags);
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001466 if (ret != SWAP_AGAIN || !page_mapped(page))
1467 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001468 }
Oleg Nesterov34bbd702007-02-28 20:13:49 -08001469
1470 page_unlock_anon_vma(anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001471 return ret;
1472}
1473
1474/**
Nick Pigginb291f002008-10-18 20:26:44 -07001475 * try_to_unmap_file - unmap/unlock file page using the object-based rmap method
1476 * @page: the page to unmap/unlock
Andi Kleen14fa31b2009-09-16 11:50:10 +02001477 * @flags: action and flags
Linus Torvalds1da177e2005-04-16 15:20:36 -07001478 *
1479 * Find all the mappings of a page using the mapping pointer and the vma chains
1480 * contained in the address_space struct it points to.
1481 *
Nick Pigginb291f002008-10-18 20:26:44 -07001482 * This function is only called from try_to_unmap/try_to_munlock for
1483 * object-based pages.
1484 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1485 * where the page was found will be held for write. So, we won't recheck
1486 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1487 * 'LOCKED.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001488 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001489static int try_to_unmap_file(struct page *page, enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001490{
1491 struct address_space *mapping = page->mapping;
1492 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1493 struct vm_area_struct *vma;
1494 struct prio_tree_iter iter;
1495 int ret = SWAP_AGAIN;
1496 unsigned long cursor;
1497 unsigned long max_nl_cursor = 0;
1498 unsigned long max_nl_size = 0;
1499 unsigned int mapcount;
1500
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07001501 mutex_lock(&mapping->i_mmap_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001502 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
Hugh Dickins1cb17292009-12-14 17:59:01 -08001503 unsigned long address = vma_address(page, vma);
1504 if (address == -EFAULT)
1505 continue;
1506 ret = try_to_unmap_one(page, vma, address, flags);
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001507 if (ret != SWAP_AGAIN || !page_mapped(page))
1508 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001509 }
1510
1511 if (list_empty(&mapping->i_mmap_nonlinear))
1512 goto out;
1513
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001514 /*
1515 * We don't bother to try to find the munlocked page in nonlinears.
1516 * It's costly. Instead, later, page reclaim logic may call
1517 * try_to_unmap(TTU_MUNLOCK) and recover PG_mlocked lazily.
1518 */
1519 if (TTU_ACTION(flags) == TTU_MUNLOCK)
1520 goto out;
1521
Linus Torvalds1da177e2005-04-16 15:20:36 -07001522 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1523 shared.vm_set.list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001524 cursor = (unsigned long) vma->vm_private_data;
1525 if (cursor > max_nl_cursor)
1526 max_nl_cursor = cursor;
1527 cursor = vma->vm_end - vma->vm_start;
1528 if (cursor > max_nl_size)
1529 max_nl_size = cursor;
1530 }
1531
Nick Pigginb291f002008-10-18 20:26:44 -07001532 if (max_nl_size == 0) { /* all nonlinears locked or reserved ? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001533 ret = SWAP_FAIL;
1534 goto out;
1535 }
1536
1537 /*
1538 * We don't try to search for this page in the nonlinear vmas,
1539 * and page_referenced wouldn't have found it anyway. Instead
1540 * just walk the nonlinear vmas trying to age and unmap some.
1541 * The mapcount of the page we came in with is irrelevant,
1542 * but even so use it as a guide to how hard we should try?
1543 */
1544 mapcount = page_mapcount(page);
1545 if (!mapcount)
1546 goto out;
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07001547 cond_resched();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001548
1549 max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
1550 if (max_nl_cursor == 0)
1551 max_nl_cursor = CLUSTER_SIZE;
1552
1553 do {
1554 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1555 shared.vm_set.list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001556 cursor = (unsigned long) vma->vm_private_data;
Hugh Dickins839b9682005-09-03 15:54:43 -07001557 while ( cursor < max_nl_cursor &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07001558 cursor < vma->vm_end - vma->vm_start) {
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001559 if (try_to_unmap_cluster(cursor, &mapcount,
1560 vma, page) == SWAP_MLOCK)
1561 ret = SWAP_MLOCK;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001562 cursor += CLUSTER_SIZE;
1563 vma->vm_private_data = (void *) cursor;
1564 if ((int)mapcount <= 0)
1565 goto out;
1566 }
1567 vma->vm_private_data = (void *) max_nl_cursor;
1568 }
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07001569 cond_resched();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001570 max_nl_cursor += CLUSTER_SIZE;
1571 } while (max_nl_cursor <= max_nl_size);
1572
1573 /*
1574 * Don't loop forever (perhaps all the remaining pages are
1575 * in locked vmas). Reset cursor on all unreserved nonlinear
1576 * vmas, now forgetting on which ones it had fallen behind.
1577 */
Hugh Dickins101d2be2005-11-21 21:32:16 -08001578 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1579 vma->vm_private_data = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001580out:
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07001581 mutex_unlock(&mapping->i_mmap_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001582 return ret;
1583}
1584
1585/**
1586 * try_to_unmap - try to remove all page table mappings to a page
1587 * @page: the page to get unmapped
Andi Kleen14fa31b2009-09-16 11:50:10 +02001588 * @flags: action and flags
Linus Torvalds1da177e2005-04-16 15:20:36 -07001589 *
1590 * Tries to remove all the page table entries which are mapping this
1591 * page, used in the pageout path. Caller must hold the page lock.
1592 * Return values are:
1593 *
1594 * SWAP_SUCCESS - we succeeded in removing all mappings
1595 * SWAP_AGAIN - we missed a mapping, try again later
1596 * SWAP_FAIL - the page is unswappable
Nick Pigginb291f002008-10-18 20:26:44 -07001597 * SWAP_MLOCK - page is mlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001598 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001599int try_to_unmap(struct page *page, enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001600{
1601 int ret;
1602
Linus Torvalds1da177e2005-04-16 15:20:36 -07001603 BUG_ON(!PageLocked(page));
Andrea Arcangeli91600e92011-01-13 15:47:16 -08001604 VM_BUG_ON(!PageHuge(page) && PageTransHuge(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001605
Hugh Dickins5ad64682009-12-14 17:59:24 -08001606 if (unlikely(PageKsm(page)))
1607 ret = try_to_unmap_ksm(page, flags);
1608 else if (PageAnon(page))
Andi Kleen14fa31b2009-09-16 11:50:10 +02001609 ret = try_to_unmap_anon(page, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001610 else
Andi Kleen14fa31b2009-09-16 11:50:10 +02001611 ret = try_to_unmap_file(page, flags);
Nick Pigginb291f002008-10-18 20:26:44 -07001612 if (ret != SWAP_MLOCK && !page_mapped(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001613 ret = SWAP_SUCCESS;
1614 return ret;
1615}
Nikita Danilov81b40822005-05-01 08:58:36 -07001616
Nick Pigginb291f002008-10-18 20:26:44 -07001617/**
1618 * try_to_munlock - try to munlock a page
1619 * @page: the page to be munlocked
1620 *
1621 * Called from munlock code. Checks all of the VMAs mapping the page
1622 * to make sure nobody else has this page mlocked. The page will be
1623 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1624 *
1625 * Return values are:
1626 *
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001627 * SWAP_AGAIN - no vma is holding page mlocked, or,
Nick Pigginb291f002008-10-18 20:26:44 -07001628 * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem
Hugh Dickins5ad64682009-12-14 17:59:24 -08001629 * SWAP_FAIL - page cannot be located at present
Nick Pigginb291f002008-10-18 20:26:44 -07001630 * SWAP_MLOCK - page is now mlocked.
1631 */
1632int try_to_munlock(struct page *page)
1633{
1634 VM_BUG_ON(!PageLocked(page) || PageLRU(page));
1635
Hugh Dickins5ad64682009-12-14 17:59:24 -08001636 if (unlikely(PageKsm(page)))
1637 return try_to_unmap_ksm(page, TTU_MUNLOCK);
1638 else if (PageAnon(page))
Andi Kleen14fa31b2009-09-16 11:50:10 +02001639 return try_to_unmap_anon(page, TTU_MUNLOCK);
Nick Pigginb291f002008-10-18 20:26:44 -07001640 else
Andi Kleen14fa31b2009-09-16 11:50:10 +02001641 return try_to_unmap_file(page, TTU_MUNLOCK);
Nick Pigginb291f002008-10-18 20:26:44 -07001642}
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001643
Peter Zijlstra01d8b202011-03-22 16:32:49 -07001644void __put_anon_vma(struct anon_vma *anon_vma)
Rik van Riel76545062010-08-09 17:18:41 -07001645{
Peter Zijlstra01d8b202011-03-22 16:32:49 -07001646 struct anon_vma *root = anon_vma->root;
Rik van Riel76545062010-08-09 17:18:41 -07001647
Peter Zijlstra01d8b202011-03-22 16:32:49 -07001648 if (root != anon_vma && atomic_dec_and_test(&root->refcount))
1649 anon_vma_free(root);
Rik van Riel76545062010-08-09 17:18:41 -07001650
Peter Zijlstra01d8b202011-03-22 16:32:49 -07001651 anon_vma_free(anon_vma);
Rik van Riel76545062010-08-09 17:18:41 -07001652}
Rik van Riel76545062010-08-09 17:18:41 -07001653
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001654#ifdef CONFIG_MIGRATION
1655/*
1656 * rmap_walk() and its helpers rmap_walk_anon() and rmap_walk_file():
1657 * Called by migrate.c to remove migration ptes, but might be used more later.
1658 */
1659static int rmap_walk_anon(struct page *page, int (*rmap_one)(struct page *,
1660 struct vm_area_struct *, unsigned long, void *), void *arg)
1661{
1662 struct anon_vma *anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -08001663 struct anon_vma_chain *avc;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001664 int ret = SWAP_AGAIN;
1665
1666 /*
1667 * Note: remove_migration_ptes() cannot use page_lock_anon_vma()
1668 * because that depends on page_mapped(); but not all its usages
Mel Gorman3f6c8272010-05-24 14:32:17 -07001669 * are holding mmap_sem. Users without mmap_sem are required to
1670 * take a reference count to prevent the anon_vma disappearing
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001671 */
1672 anon_vma = page_anon_vma(page);
1673 if (!anon_vma)
1674 return ret;
Rik van Rielcba48b92010-08-09 17:18:38 -07001675 anon_vma_lock(anon_vma);
Rik van Riel5beb4932010-03-05 13:42:07 -08001676 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1677 struct vm_area_struct *vma = avc->vma;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001678 unsigned long address = vma_address(page, vma);
1679 if (address == -EFAULT)
1680 continue;
1681 ret = rmap_one(page, vma, address, arg);
1682 if (ret != SWAP_AGAIN)
1683 break;
1684 }
Rik van Rielcba48b92010-08-09 17:18:38 -07001685 anon_vma_unlock(anon_vma);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001686 return ret;
1687}
1688
1689static int rmap_walk_file(struct page *page, int (*rmap_one)(struct page *,
1690 struct vm_area_struct *, unsigned long, void *), void *arg)
1691{
1692 struct address_space *mapping = page->mapping;
1693 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1694 struct vm_area_struct *vma;
1695 struct prio_tree_iter iter;
1696 int ret = SWAP_AGAIN;
1697
1698 if (!mapping)
1699 return ret;
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07001700 mutex_lock(&mapping->i_mmap_mutex);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001701 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1702 unsigned long address = vma_address(page, vma);
1703 if (address == -EFAULT)
1704 continue;
1705 ret = rmap_one(page, vma, address, arg);
1706 if (ret != SWAP_AGAIN)
1707 break;
1708 }
1709 /*
1710 * No nonlinear handling: being always shared, nonlinear vmas
1711 * never contain migration ptes. Decide what to do about this
1712 * limitation to linear when we need rmap_walk() on nonlinear.
1713 */
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07001714 mutex_unlock(&mapping->i_mmap_mutex);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001715 return ret;
1716}
1717
1718int rmap_walk(struct page *page, int (*rmap_one)(struct page *,
1719 struct vm_area_struct *, unsigned long, void *), void *arg)
1720{
1721 VM_BUG_ON(!PageLocked(page));
1722
1723 if (unlikely(PageKsm(page)))
1724 return rmap_walk_ksm(page, rmap_one, arg);
1725 else if (PageAnon(page))
1726 return rmap_walk_anon(page, rmap_one, arg);
1727 else
1728 return rmap_walk_file(page, rmap_one, arg);
1729}
1730#endif /* CONFIG_MIGRATION */
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001731
Naoya Horiguchie3390f62010-06-15 13:18:13 +09001732#ifdef CONFIG_HUGETLB_PAGE
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001733/*
1734 * The following three functions are for anonymous (private mapped) hugepages.
1735 * Unlike common anonymous pages, anonymous hugepages have no accounting code
1736 * and no lru code, because we handle hugepages differently from common pages.
1737 */
1738static void __hugepage_set_anon_rmap(struct page *page,
1739 struct vm_area_struct *vma, unsigned long address, int exclusive)
1740{
1741 struct anon_vma *anon_vma = vma->anon_vma;
Naoya Horiguchi433abed2010-09-10 13:23:03 +09001742
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001743 BUG_ON(!anon_vma);
Naoya Horiguchi433abed2010-09-10 13:23:03 +09001744
1745 if (PageAnon(page))
1746 return;
1747 if (!exclusive)
1748 anon_vma = anon_vma->root;
1749
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001750 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1751 page->mapping = (struct address_space *) anon_vma;
1752 page->index = linear_page_index(vma, address);
1753}
1754
1755void hugepage_add_anon_rmap(struct page *page,
1756 struct vm_area_struct *vma, unsigned long address)
1757{
1758 struct anon_vma *anon_vma = vma->anon_vma;
1759 int first;
Naoya Horiguchia850ea32010-09-10 13:23:06 +09001760
1761 BUG_ON(!PageLocked(page));
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001762 BUG_ON(!anon_vma);
Hugh Dickins5dbe0af2011-05-28 13:17:04 -07001763 /* address might be in next vma when migration races vma_adjust */
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001764 first = atomic_inc_and_test(&page->_mapcount);
1765 if (first)
1766 __hugepage_set_anon_rmap(page, vma, address, 0);
1767}
1768
1769void hugepage_add_new_anon_rmap(struct page *page,
1770 struct vm_area_struct *vma, unsigned long address)
1771{
1772 BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1773 atomic_set(&page->_mapcount, 0);
1774 __hugepage_set_anon_rmap(page, vma, address, 1);
1775}
Naoya Horiguchie3390f62010-06-15 13:18:13 +09001776#endif /* CONFIG_HUGETLB_PAGE */