blob: 7df7984d476c8661b900725c0e2c3f33c883a32d [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 * mm->mmap_sem
25 * page->flags PG_locked (lock_page)
Peter Zijlstra3d48ae42011-05-24 17:12:06 -070026 * mapping->i_mmap_mutex
Peter Zijlstra2b575eb2011-05-24 17:12:11 -070027 * anon_vma->mutex
Nick Piggin82591e62006-10-19 23:29:10 -070028 * mm->page_table_lock or pte_lock
29 * zone->lru_lock (in mark_page_accessed, isolate_lru_page)
30 * swap_lock (in swap_duplicate, swap_info_get)
31 * mmlist_lock (in mmput, drain_mmlist and others)
32 * mapping->private_lock (in __set_page_dirty_buffers)
Dave Chinner250df6e2011-03-22 22:23:36 +110033 * inode->i_lock (in set_page_dirty's __mark_inode_dirty)
Christoph Hellwigf758eea2011-04-21 18:19:44 -060034 * bdi.wb->list_lock (in set_page_dirty's __mark_inode_dirty)
Nick Piggin82591e62006-10-19 23:29:10 -070035 * 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,
Christoph Hellwigf758eea2011-04-21 18:19:44 -060038 * within bdi.wb->list_lock in __sync_single_inode)
Andi Kleen6a460792009-09-16 11:50:15 +020039 *
Peter Zijlstra9b679322011-06-27 16:18:09 -070040 * anon_vma->mutex,mapping->i_mutex (memory_failure, collect_procs_anon)
41 * ->tasklist_lock
Andi Kleen6a460792009-09-16 11:50:15 +020042 * pte map lock
Linus Torvalds1da177e2005-04-16 15:20:36 -070043 */
44
45#include <linux/mm.h>
46#include <linux/pagemap.h>
47#include <linux/swap.h>
48#include <linux/swapops.h>
49#include <linux/slab.h>
50#include <linux/init.h>
Hugh Dickins5ad64682009-12-14 17:59:24 -080051#include <linux/ksm.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070052#include <linux/rmap.h>
53#include <linux/rcupdate.h>
Paul Gortmakerb95f1b312011-10-16 02:01:52 -040054#include <linux/export.h>
Balbir Singh8a9f3cc2008-02-07 00:13:53 -080055#include <linux/memcontrol.h>
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070056#include <linux/mmu_notifier.h>
KOSAKI Motohiro64cdd542009-01-06 14:39:16 -080057#include <linux/migrate.h>
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +090058#include <linux/hugetlb.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070059
60#include <asm/tlbflush.h>
61
Nick Pigginb291f002008-10-18 20:26:44 -070062#include "internal.h"
63
Adrian Bunkfdd2e5f2008-10-18 20:28:38 -070064static struct kmem_cache *anon_vma_cachep;
Rik van Riel5beb4932010-03-05 13:42:07 -080065static struct kmem_cache *anon_vma_chain_cachep;
Adrian Bunkfdd2e5f2008-10-18 20:28:38 -070066
67static inline struct anon_vma *anon_vma_alloc(void)
68{
Peter Zijlstra01d8b202011-03-22 16:32:49 -070069 struct anon_vma *anon_vma;
70
71 anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
72 if (anon_vma) {
73 atomic_set(&anon_vma->refcount, 1);
74 /*
75 * Initialise the anon_vma root to point to itself. If called
76 * from fork, the root will be reset to the parents anon_vma.
77 */
78 anon_vma->root = anon_vma;
79 }
80
81 return anon_vma;
Adrian Bunkfdd2e5f2008-10-18 20:28:38 -070082}
83
Peter Zijlstra01d8b202011-03-22 16:32:49 -070084static inline void anon_vma_free(struct anon_vma *anon_vma)
Adrian Bunkfdd2e5f2008-10-18 20:28:38 -070085{
Peter Zijlstra01d8b202011-03-22 16:32:49 -070086 VM_BUG_ON(atomic_read(&anon_vma->refcount));
Peter Zijlstra88c22082011-05-24 17:12:13 -070087
88 /*
89 * Synchronize against page_lock_anon_vma() such that
90 * we can safely hold the lock without the anon_vma getting
91 * freed.
92 *
93 * Relies on the full mb implied by the atomic_dec_and_test() from
94 * put_anon_vma() against the acquire barrier implied by
95 * mutex_trylock() from page_lock_anon_vma(). This orders:
96 *
97 * page_lock_anon_vma() VS put_anon_vma()
98 * mutex_trylock() atomic_dec_and_test()
99 * LOCK MB
100 * atomic_read() mutex_is_locked()
101 *
102 * LOCK should suffice since the actual taking of the lock must
103 * happen _before_ what follows.
104 */
105 if (mutex_is_locked(&anon_vma->root->mutex)) {
106 anon_vma_lock(anon_vma);
107 anon_vma_unlock(anon_vma);
108 }
109
Adrian Bunkfdd2e5f2008-10-18 20:28:38 -0700110 kmem_cache_free(anon_vma_cachep, anon_vma);
111}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700112
Linus Torvaldsdd347392011-06-17 19:05:36 -0700113static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp)
Rik van Riel5beb4932010-03-05 13:42:07 -0800114{
Linus Torvaldsdd347392011-06-17 19:05:36 -0700115 return kmem_cache_alloc(anon_vma_chain_cachep, gfp);
Rik van Riel5beb4932010-03-05 13:42:07 -0800116}
117
Namhyung Kime574b5f2010-10-26 14:22:02 -0700118static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
Rik van Riel5beb4932010-03-05 13:42:07 -0800119{
120 kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
121}
122
Kautuk Consul6583a842012-03-21 16:34:01 -0700123static void anon_vma_chain_link(struct vm_area_struct *vma,
124 struct anon_vma_chain *avc,
125 struct anon_vma *anon_vma)
126{
127 avc->vma = vma;
128 avc->anon_vma = anon_vma;
129 list_add(&avc->same_vma, &vma->anon_vma_chain);
Michel Lespinassebf181b92012-10-08 16:31:39 -0700130 anon_vma_interval_tree_insert(avc, &anon_vma->rb_root);
Kautuk Consul6583a842012-03-21 16:34:01 -0700131}
132
Linus Torvaldsd9d332e2008-10-19 10:32:20 -0700133/**
134 * anon_vma_prepare - attach an anon_vma to a memory region
135 * @vma: the memory region in question
136 *
137 * This makes sure the memory mapping described by 'vma' has
138 * an 'anon_vma' attached to it, so that we can associate the
139 * anonymous pages mapped into it with that anon_vma.
140 *
141 * The common case will be that we already have one, but if
Figo.zhang23a07902010-12-27 15:14:06 +0100142 * not we either need to find an adjacent mapping that we
Linus Torvaldsd9d332e2008-10-19 10:32:20 -0700143 * can re-use the anon_vma from (very common when the only
144 * reason for splitting a vma has been mprotect()), or we
145 * allocate a new one.
146 *
147 * Anon-vma allocations are very subtle, because we may have
148 * optimistically looked up an anon_vma in page_lock_anon_vma()
149 * and that may actually touch the spinlock even in the newly
150 * allocated vma (it depends on RCU to make sure that the
151 * anon_vma isn't actually destroyed).
152 *
153 * As a result, we need to do proper anon_vma locking even
154 * for the new allocation. At the same time, we do not want
155 * to do any locking for the common case of already having
156 * an anon_vma.
157 *
158 * This must be called with the mmap_sem held for reading.
159 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700160int anon_vma_prepare(struct vm_area_struct *vma)
161{
162 struct anon_vma *anon_vma = vma->anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -0800163 struct anon_vma_chain *avc;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700164
165 might_sleep();
166 if (unlikely(!anon_vma)) {
167 struct mm_struct *mm = vma->vm_mm;
Linus Torvaldsd9d332e2008-10-19 10:32:20 -0700168 struct anon_vma *allocated;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700169
Linus Torvaldsdd347392011-06-17 19:05:36 -0700170 avc = anon_vma_chain_alloc(GFP_KERNEL);
Rik van Riel5beb4932010-03-05 13:42:07 -0800171 if (!avc)
172 goto out_enomem;
173
Linus Torvalds1da177e2005-04-16 15:20:36 -0700174 anon_vma = find_mergeable_anon_vma(vma);
Linus Torvaldsd9d332e2008-10-19 10:32:20 -0700175 allocated = NULL;
176 if (!anon_vma) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700177 anon_vma = anon_vma_alloc();
178 if (unlikely(!anon_vma))
Rik van Riel5beb4932010-03-05 13:42:07 -0800179 goto out_enomem_free_avc;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700180 allocated = anon_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700181 }
182
Rik van Rielcba48b92010-08-09 17:18:38 -0700183 anon_vma_lock(anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700184 /* page_table_lock to protect against threads */
185 spin_lock(&mm->page_table_lock);
186 if (likely(!vma->anon_vma)) {
187 vma->anon_vma = anon_vma;
Kautuk Consul6583a842012-03-21 16:34:01 -0700188 anon_vma_chain_link(vma, avc, anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700189 allocated = NULL;
Oleg Nesterov31f2b0e2010-04-23 13:18:01 -0400190 avc = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700191 }
192 spin_unlock(&mm->page_table_lock);
Rik van Rielcba48b92010-08-09 17:18:38 -0700193 anon_vma_unlock(anon_vma);
Oleg Nesterov31f2b0e2010-04-23 13:18:01 -0400194
195 if (unlikely(allocated))
Peter Zijlstra01d8b202011-03-22 16:32:49 -0700196 put_anon_vma(allocated);
Oleg Nesterov31f2b0e2010-04-23 13:18:01 -0400197 if (unlikely(avc))
Rik van Riel5beb4932010-03-05 13:42:07 -0800198 anon_vma_chain_free(avc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700199 }
200 return 0;
Rik van Riel5beb4932010-03-05 13:42:07 -0800201
202 out_enomem_free_avc:
203 anon_vma_chain_free(avc);
204 out_enomem:
205 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700206}
207
Linus Torvaldsbb4aa392011-06-16 20:44:51 -0700208/*
209 * This is a useful helper function for locking the anon_vma root as
210 * we traverse the vma->anon_vma_chain, looping over anon_vma's that
211 * have the same vma.
212 *
213 * Such anon_vma's should have the same root, so you'd expect to see
214 * just a single mutex_lock for the whole traversal.
215 */
216static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
217{
218 struct anon_vma *new_root = anon_vma->root;
219 if (new_root != root) {
220 if (WARN_ON_ONCE(root))
221 mutex_unlock(&root->mutex);
222 root = new_root;
223 mutex_lock(&root->mutex);
224 }
225 return root;
226}
227
228static inline void unlock_anon_vma_root(struct anon_vma *root)
229{
230 if (root)
231 mutex_unlock(&root->mutex);
232}
233
Rik van Riel5beb4932010-03-05 13:42:07 -0800234/*
235 * Attach the anon_vmas from src to dst.
236 * Returns 0 on success, -ENOMEM on failure.
237 */
238int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700239{
Rik van Riel5beb4932010-03-05 13:42:07 -0800240 struct anon_vma_chain *avc, *pavc;
Linus Torvaldsbb4aa392011-06-16 20:44:51 -0700241 struct anon_vma *root = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700242
Linus Torvalds646d87b2010-04-11 17:15:03 -0700243 list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
Linus Torvaldsbb4aa392011-06-16 20:44:51 -0700244 struct anon_vma *anon_vma;
245
Linus Torvaldsdd347392011-06-17 19:05:36 -0700246 avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
247 if (unlikely(!avc)) {
248 unlock_anon_vma_root(root);
249 root = NULL;
250 avc = anon_vma_chain_alloc(GFP_KERNEL);
251 if (!avc)
252 goto enomem_failure;
253 }
Linus Torvaldsbb4aa392011-06-16 20:44:51 -0700254 anon_vma = pavc->anon_vma;
255 root = lock_anon_vma_root(root, anon_vma);
256 anon_vma_chain_link(dst, avc, anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700257 }
Linus Torvaldsbb4aa392011-06-16 20:44:51 -0700258 unlock_anon_vma_root(root);
Rik van Riel5beb4932010-03-05 13:42:07 -0800259 return 0;
260
261 enomem_failure:
262 unlink_anon_vmas(dst);
263 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700264}
265
Rik van Riel5beb4932010-03-05 13:42:07 -0800266/*
267 * Attach vma to its own anon_vma, as well as to the anon_vmas that
268 * the corresponding VMA in the parent process is attached to.
269 * Returns 0 on success, non-zero on failure.
270 */
271int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700272{
Rik van Riel5beb4932010-03-05 13:42:07 -0800273 struct anon_vma_chain *avc;
274 struct anon_vma *anon_vma;
275
276 /* Don't bother if the parent process has no anon_vma here. */
277 if (!pvma->anon_vma)
278 return 0;
279
280 /*
281 * First, attach the new VMA to the parent VMA's anon_vmas,
282 * so rmap can find non-COWed pages in child processes.
283 */
284 if (anon_vma_clone(vma, pvma))
285 return -ENOMEM;
286
287 /* Then add our own anon_vma. */
288 anon_vma = anon_vma_alloc();
289 if (!anon_vma)
290 goto out_error;
Linus Torvaldsdd347392011-06-17 19:05:36 -0700291 avc = anon_vma_chain_alloc(GFP_KERNEL);
Rik van Riel5beb4932010-03-05 13:42:07 -0800292 if (!avc)
293 goto out_error_free_anon_vma;
Rik van Riel5c341ee12010-08-09 17:18:39 -0700294
295 /*
296 * The root anon_vma's spinlock is the lock actually used when we
297 * lock any of the anon_vmas in this anon_vma tree.
298 */
299 anon_vma->root = pvma->anon_vma->root;
Rik van Riel76545062010-08-09 17:18:41 -0700300 /*
Peter Zijlstra01d8b202011-03-22 16:32:49 -0700301 * With refcounts, an anon_vma can stay around longer than the
302 * process it belongs to. The root anon_vma needs to be pinned until
303 * this anon_vma is freed, because the lock lives in the root.
Rik van Riel76545062010-08-09 17:18:41 -0700304 */
305 get_anon_vma(anon_vma->root);
Rik van Riel5beb4932010-03-05 13:42:07 -0800306 /* Mark this anon_vma as the one where our new (COWed) pages go. */
307 vma->anon_vma = anon_vma;
Linus Torvaldsbb4aa392011-06-16 20:44:51 -0700308 anon_vma_lock(anon_vma);
Rik van Riel5c341ee12010-08-09 17:18:39 -0700309 anon_vma_chain_link(vma, avc, anon_vma);
Linus Torvaldsbb4aa392011-06-16 20:44:51 -0700310 anon_vma_unlock(anon_vma);
Rik van Riel5beb4932010-03-05 13:42:07 -0800311
312 return 0;
313
314 out_error_free_anon_vma:
Peter Zijlstra01d8b202011-03-22 16:32:49 -0700315 put_anon_vma(anon_vma);
Rik van Riel5beb4932010-03-05 13:42:07 -0800316 out_error:
Rik van Riel4946d542010-04-05 12:13:33 -0400317 unlink_anon_vmas(vma);
Rik van Riel5beb4932010-03-05 13:42:07 -0800318 return -ENOMEM;
319}
320
Rik van Riel5beb4932010-03-05 13:42:07 -0800321void unlink_anon_vmas(struct vm_area_struct *vma)
322{
323 struct anon_vma_chain *avc, *next;
Peter Zijlstraeee2acb2011-06-17 13:54:23 +0200324 struct anon_vma *root = NULL;
Rik van Riel5beb4932010-03-05 13:42:07 -0800325
Rik van Riel5c341ee12010-08-09 17:18:39 -0700326 /*
327 * Unlink each anon_vma chained to the VMA. This list is ordered
328 * from newest to oldest, ensuring the root anon_vma gets freed last.
329 */
Rik van Riel5beb4932010-03-05 13:42:07 -0800330 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
Peter Zijlstraeee2acb2011-06-17 13:54:23 +0200331 struct anon_vma *anon_vma = avc->anon_vma;
332
333 root = lock_anon_vma_root(root, anon_vma);
Michel Lespinassebf181b92012-10-08 16:31:39 -0700334 anon_vma_interval_tree_remove(avc, &anon_vma->rb_root);
Peter Zijlstraeee2acb2011-06-17 13:54:23 +0200335
336 /*
337 * Leave empty anon_vmas on the list - we'll need
338 * to free them outside the lock.
339 */
Michel Lespinassebf181b92012-10-08 16:31:39 -0700340 if (RB_EMPTY_ROOT(&anon_vma->rb_root))
Peter Zijlstraeee2acb2011-06-17 13:54:23 +0200341 continue;
342
343 list_del(&avc->same_vma);
344 anon_vma_chain_free(avc);
345 }
346 unlock_anon_vma_root(root);
347
348 /*
349 * Iterate the list once more, it now only contains empty and unlinked
350 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
351 * needing to acquire the anon_vma->root->mutex.
352 */
353 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
354 struct anon_vma *anon_vma = avc->anon_vma;
355
356 put_anon_vma(anon_vma);
357
Rik van Riel5beb4932010-03-05 13:42:07 -0800358 list_del(&avc->same_vma);
359 anon_vma_chain_free(avc);
360 }
361}
362
Alexey Dobriyan51cc5062008-07-25 19:45:34 -0700363static void anon_vma_ctor(void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700364{
Christoph Lametera35afb82007-05-16 22:10:57 -0700365 struct anon_vma *anon_vma = data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700366
Peter Zijlstra2b575eb2011-05-24 17:12:11 -0700367 mutex_init(&anon_vma->mutex);
Peter Zijlstra83813262011-03-22 16:32:48 -0700368 atomic_set(&anon_vma->refcount, 0);
Michel Lespinassebf181b92012-10-08 16:31:39 -0700369 anon_vma->rb_root = RB_ROOT;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700370}
371
372void __init anon_vma_init(void)
373{
374 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
Paul Mundt20c2df82007-07-20 10:11:58 +0900375 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor);
Rik van Riel5beb4932010-03-05 13:42:07 -0800376 anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain, SLAB_PANIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700377}
378
379/*
Peter Zijlstra6111e4c2011-05-24 17:12:08 -0700380 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
381 *
382 * Since there is no serialization what so ever against page_remove_rmap()
383 * the best this function can do is return a locked anon_vma that might
384 * have been relevant to this page.
385 *
386 * The page might have been remapped to a different anon_vma or the anon_vma
387 * returned may already be freed (and even reused).
388 *
Peter Zijlstrabc658c92011-05-29 10:33:44 +0200389 * In case it was remapped to a different anon_vma, the new anon_vma will be a
390 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
391 * ensure that any anon_vma obtained from the page will still be valid for as
392 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
393 *
Peter Zijlstra6111e4c2011-05-24 17:12:08 -0700394 * All users of this function must be very careful when walking the anon_vma
395 * chain and verify that the page in question is indeed mapped in it
396 * [ something equivalent to page_mapped_in_vma() ].
397 *
398 * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap()
399 * that the anon_vma pointer from page->mapping is valid if there is a
400 * mapcount, we can dereference the anon_vma after observing those.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700401 */
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700402struct anon_vma *page_get_anon_vma(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700403{
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700404 struct anon_vma *anon_vma = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700405 unsigned long anon_mapping;
406
407 rcu_read_lock();
Hugh Dickins80e148222009-12-14 17:59:29 -0800408 anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping);
Hugh Dickins3ca7b3c2009-12-14 17:58:57 -0800409 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700410 goto out;
411 if (!page_mapped(page))
412 goto out;
413
414 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700415 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
416 anon_vma = NULL;
417 goto out;
418 }
Hugh Dickinsf18194272010-08-25 23:12:54 -0700419
420 /*
421 * If this page is still mapped, then its anon_vma cannot have been
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700422 * freed. But if it has been unmapped, we have no security against the
423 * anon_vma structure being freed and reused (for another anon_vma:
424 * SLAB_DESTROY_BY_RCU guarantees that - so the atomic_inc_not_zero()
425 * above cannot corrupt).
Hugh Dickinsf18194272010-08-25 23:12:54 -0700426 */
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700427 if (!page_mapped(page)) {
428 put_anon_vma(anon_vma);
429 anon_vma = NULL;
430 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700431out:
432 rcu_read_unlock();
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700433
434 return anon_vma;
435}
436
Peter Zijlstra88c22082011-05-24 17:12:13 -0700437/*
438 * Similar to page_get_anon_vma() except it locks the anon_vma.
439 *
440 * Its a little more complex as it tries to keep the fast path to a single
441 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
442 * reference like with page_get_anon_vma() and then block on the mutex.
443 */
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700444struct anon_vma *page_lock_anon_vma(struct page *page)
445{
Peter Zijlstra88c22082011-05-24 17:12:13 -0700446 struct anon_vma *anon_vma = NULL;
Hugh Dickinseee0f252011-05-28 13:20:21 -0700447 struct anon_vma *root_anon_vma;
Peter Zijlstra88c22082011-05-24 17:12:13 -0700448 unsigned long anon_mapping;
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700449
Peter Zijlstra88c22082011-05-24 17:12:13 -0700450 rcu_read_lock();
451 anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping);
452 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
453 goto out;
454 if (!page_mapped(page))
455 goto out;
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700456
Peter Zijlstra88c22082011-05-24 17:12:13 -0700457 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
Hugh Dickinseee0f252011-05-28 13:20:21 -0700458 root_anon_vma = ACCESS_ONCE(anon_vma->root);
459 if (mutex_trylock(&root_anon_vma->mutex)) {
Peter Zijlstra88c22082011-05-24 17:12:13 -0700460 /*
Hugh Dickinseee0f252011-05-28 13:20:21 -0700461 * If the page is still mapped, then this anon_vma is still
462 * its anon_vma, and holding the mutex ensures that it will
Peter Zijlstrabc658c92011-05-29 10:33:44 +0200463 * not go away, see anon_vma_free().
Peter Zijlstra88c22082011-05-24 17:12:13 -0700464 */
Hugh Dickinseee0f252011-05-28 13:20:21 -0700465 if (!page_mapped(page)) {
466 mutex_unlock(&root_anon_vma->mutex);
Peter Zijlstra88c22082011-05-24 17:12:13 -0700467 anon_vma = NULL;
468 }
469 goto out;
470 }
471
472 /* trylock failed, we got to sleep */
473 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
474 anon_vma = NULL;
475 goto out;
476 }
477
478 if (!page_mapped(page)) {
479 put_anon_vma(anon_vma);
480 anon_vma = NULL;
481 goto out;
482 }
483
484 /* we pinned the anon_vma, its safe to sleep */
485 rcu_read_unlock();
486 anon_vma_lock(anon_vma);
487
488 if (atomic_dec_and_test(&anon_vma->refcount)) {
489 /*
490 * Oops, we held the last refcount, release the lock
491 * and bail -- can't simply use put_anon_vma() because
492 * we'll deadlock on the anon_vma_lock() recursion.
493 */
494 anon_vma_unlock(anon_vma);
495 __put_anon_vma(anon_vma);
496 anon_vma = NULL;
497 }
498
499 return anon_vma;
500
501out:
502 rcu_read_unlock();
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700503 return anon_vma;
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800504}
505
Andi Kleen10be22d2009-09-16 11:50:04 +0200506void page_unlock_anon_vma(struct anon_vma *anon_vma)
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800507{
Rik van Rielcba48b92010-08-09 17:18:38 -0700508 anon_vma_unlock(anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700509}
510
511/*
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -0800512 * At what user virtual address is page expected in @vma?
Linus Torvalds1da177e2005-04-16 15:20:36 -0700513 */
Michel Lespinasse86c2ad12012-10-08 16:31:42 -0700514static inline unsigned long
515__vma_address(struct page *page, struct vm_area_struct *vma)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700516{
517 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700518
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900519 if (unlikely(is_vm_hugetlb_page(vma)))
520 pgoff = page->index << huge_page_order(page_hstate(page));
Michel Lespinasse86c2ad12012-10-08 16:31:42 -0700521
522 return vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
523}
524
525inline unsigned long
526vma_address(struct page *page, struct vm_area_struct *vma)
527{
528 unsigned long address = __vma_address(page, vma);
529
530 /* page should be within @vma mapping range */
531 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
532
Linus Torvalds1da177e2005-04-16 15:20:36 -0700533 return address;
534}
535
536/*
Huang Shijiebf89c8c2009-10-01 15:44:04 -0700537 * At what user virtual address is page expected in vma?
Naoya Horiguchiab941e02010-05-11 14:06:55 -0700538 * Caller should check the page is actually part of the vma.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700539 */
540unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
541{
Michel Lespinasse86c2ad12012-10-08 16:31:42 -0700542 unsigned long address;
Andrea Arcangeli21d0d442010-08-09 17:19:10 -0700543 if (PageAnon(page)) {
Hugh Dickins4829b902010-10-02 17:46:06 -0700544 struct anon_vma *page__anon_vma = page_anon_vma(page);
545 /*
546 * Note: swapoff's unuse_vma() is more efficient with this
547 * check, and needs it to match anon_vma when KSM is active.
548 */
549 if (!vma->anon_vma || !page__anon_vma ||
550 vma->anon_vma->root != page__anon_vma->root)
Andrea Arcangeli21d0d442010-08-09 17:19:10 -0700551 return -EFAULT;
552 } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) {
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800553 if (!vma->vm_file ||
554 vma->vm_file->f_mapping != page->mapping)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700555 return -EFAULT;
556 } else
557 return -EFAULT;
Michel Lespinasse86c2ad12012-10-08 16:31:42 -0700558 address = __vma_address(page, vma);
559 if (unlikely(address < vma->vm_start || address >= vma->vm_end))
560 return -EFAULT;
561 return address;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700562}
563
564/*
Nikita Danilov81b40822005-05-01 08:58:36 -0700565 * Check that @page is mapped at @address into @mm.
566 *
Nick Piggin479db0b2008-08-20 14:09:18 -0700567 * If @sync is false, page_check_address may perform a racy check to avoid
568 * the page table lock when the pte is not present (helpful when reclaiming
569 * highly shared pages).
570 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -0700571 * On success returns with pte mapped and locked.
Nikita Danilov81b40822005-05-01 08:58:36 -0700572 */
Namhyung Kime9a81a82010-10-26 14:22:01 -0700573pte_t *__page_check_address(struct page *page, struct mm_struct *mm,
Nick Piggin479db0b2008-08-20 14:09:18 -0700574 unsigned long address, spinlock_t **ptlp, int sync)
Nikita Danilov81b40822005-05-01 08:58:36 -0700575{
576 pgd_t *pgd;
577 pud_t *pud;
578 pmd_t *pmd;
579 pte_t *pte;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700580 spinlock_t *ptl;
Nikita Danilov81b40822005-05-01 08:58:36 -0700581
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900582 if (unlikely(PageHuge(page))) {
583 pte = huge_pte_offset(mm, address);
584 ptl = &mm->page_table_lock;
585 goto check;
586 }
587
Nikita Danilov81b40822005-05-01 08:58:36 -0700588 pgd = pgd_offset(mm, address);
Hugh Dickinsc0718802005-10-29 18:16:31 -0700589 if (!pgd_present(*pgd))
590 return NULL;
591
592 pud = pud_offset(pgd, address);
593 if (!pud_present(*pud))
594 return NULL;
595
596 pmd = pmd_offset(pud, address);
597 if (!pmd_present(*pmd))
598 return NULL;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800599 if (pmd_trans_huge(*pmd))
600 return NULL;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700601
602 pte = pte_offset_map(pmd, address);
603 /* Make a quick check before getting the lock */
Nick Piggin479db0b2008-08-20 14:09:18 -0700604 if (!sync && !pte_present(*pte)) {
Hugh Dickinsc0718802005-10-29 18:16:31 -0700605 pte_unmap(pte);
606 return NULL;
Nikita Danilov81b40822005-05-01 08:58:36 -0700607 }
Hugh Dickinsc0718802005-10-29 18:16:31 -0700608
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700609 ptl = pte_lockptr(mm, pmd);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900610check:
Hugh Dickinsc0718802005-10-29 18:16:31 -0700611 spin_lock(ptl);
612 if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
613 *ptlp = ptl;
614 return pte;
615 }
616 pte_unmap_unlock(pte, ptl);
617 return NULL;
Nikita Danilov81b40822005-05-01 08:58:36 -0700618}
619
Nick Pigginb291f002008-10-18 20:26:44 -0700620/**
621 * page_mapped_in_vma - check whether a page is really mapped in a VMA
622 * @page: the page to test
623 * @vma: the VMA to test
624 *
625 * Returns 1 if the page is mapped into the page tables of the VMA, 0
626 * if the page is not mapped into the page tables of this VMA. Only
627 * valid for normal file or anonymous VMAs.
628 */
Andi Kleen6a460792009-09-16 11:50:15 +0200629int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
Nick Pigginb291f002008-10-18 20:26:44 -0700630{
631 unsigned long address;
632 pte_t *pte;
633 spinlock_t *ptl;
634
Michel Lespinasse86c2ad12012-10-08 16:31:42 -0700635 address = __vma_address(page, vma);
636 if (unlikely(address < vma->vm_start || address >= vma->vm_end))
Nick Pigginb291f002008-10-18 20:26:44 -0700637 return 0;
638 pte = page_check_address(page, vma->vm_mm, address, &ptl, 1);
639 if (!pte) /* the page is not in this mm */
640 return 0;
641 pte_unmap_unlock(pte, ptl);
642
643 return 1;
644}
645
Nikita Danilov81b40822005-05-01 08:58:36 -0700646/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700647 * Subfunctions of page_referenced: page_referenced_one called
648 * repeatedly from either page_referenced_anon or page_referenced_file.
649 */
Hugh Dickins5ad64682009-12-14 17:59:24 -0800650int page_referenced_one(struct page *page, struct vm_area_struct *vma,
651 unsigned long address, unsigned int *mapcount,
652 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700653{
654 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700655 int referenced = 0;
656
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800657 if (unlikely(PageTransHuge(page))) {
658 pmd_t *pmd;
659
660 spin_lock(&mm->page_table_lock);
Andrea Arcangeli2da28bf2011-03-11 14:58:29 -0800661 /*
662 * rmap might return false positives; we must filter
663 * these out using page_check_address_pmd().
664 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800665 pmd = page_check_address_pmd(page, mm, address,
666 PAGE_CHECK_ADDRESS_PMD_FLAG);
Andrea Arcangeli2da28bf2011-03-11 14:58:29 -0800667 if (!pmd) {
668 spin_unlock(&mm->page_table_lock);
669 goto out;
670 }
671
672 if (vma->vm_flags & VM_LOCKED) {
673 spin_unlock(&mm->page_table_lock);
674 *mapcount = 0; /* break early from loop */
675 *vm_flags |= VM_LOCKED;
676 goto out;
677 }
678
679 /* go ahead even if the pmd is pmd_trans_splitting() */
680 if (pmdp_clear_flush_young_notify(vma, address, pmd))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800681 referenced++;
682 spin_unlock(&mm->page_table_lock);
683 } else {
684 pte_t *pte;
685 spinlock_t *ptl;
686
Andrea Arcangeli2da28bf2011-03-11 14:58:29 -0800687 /*
688 * rmap might return false positives; we must filter
689 * these out using page_check_address().
690 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800691 pte = page_check_address(page, mm, address, &ptl, 0);
692 if (!pte)
693 goto out;
694
Andrea Arcangeli2da28bf2011-03-11 14:58:29 -0800695 if (vma->vm_flags & VM_LOCKED) {
696 pte_unmap_unlock(pte, ptl);
697 *mapcount = 0; /* break early from loop */
698 *vm_flags |= VM_LOCKED;
699 goto out;
700 }
701
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800702 if (ptep_clear_flush_young_notify(vma, address, pte)) {
703 /*
704 * Don't treat a reference through a sequentially read
705 * mapping as such. If the page has been used in
706 * another mapping, we will catch it; if this other
707 * mapping is already gone, the unmap path will have
708 * set PG_referenced or activated the page.
709 */
710 if (likely(!VM_SequentialReadHint(vma)))
711 referenced++;
712 }
713 pte_unmap_unlock(pte, ptl);
714 }
715
Hugh Dickinsc0718802005-10-29 18:16:31 -0700716 (*mapcount)--;
Huang Shijie273f0472009-12-14 17:58:51 -0800717
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700718 if (referenced)
719 *vm_flags |= vma->vm_flags;
Huang Shijie273f0472009-12-14 17:58:51 -0800720out:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700721 return referenced;
722}
723
Balbir Singhbed71612008-02-07 00:14:01 -0800724static int page_referenced_anon(struct page *page,
Johannes Weiner72835c82012-01-12 17:18:32 -0800725 struct mem_cgroup *memcg,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700726 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700727{
728 unsigned int mapcount;
729 struct anon_vma *anon_vma;
Michel Lespinassebf181b92012-10-08 16:31:39 -0700730 pgoff_t pgoff;
Rik van Riel5beb4932010-03-05 13:42:07 -0800731 struct anon_vma_chain *avc;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700732 int referenced = 0;
733
734 anon_vma = page_lock_anon_vma(page);
735 if (!anon_vma)
736 return referenced;
737
738 mapcount = page_mapcount(page);
Michel Lespinassebf181b92012-10-08 16:31:39 -0700739 pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
740 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
Rik van Riel5beb4932010-03-05 13:42:07 -0800741 struct vm_area_struct *vma = avc->vma;
Hugh Dickins1cb17292009-12-14 17:59:01 -0800742 unsigned long address = vma_address(page, vma);
Balbir Singhbed71612008-02-07 00:14:01 -0800743 /*
744 * If we are reclaiming on behalf of a cgroup, skip
745 * counting on behalf of references from different
746 * cgroups
747 */
Johannes Weiner72835c82012-01-12 17:18:32 -0800748 if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
Balbir Singhbed71612008-02-07 00:14:01 -0800749 continue;
Hugh Dickins1cb17292009-12-14 17:59:01 -0800750 referenced += page_referenced_one(page, vma, address,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700751 &mapcount, vm_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700752 if (!mapcount)
753 break;
754 }
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800755
756 page_unlock_anon_vma(anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700757 return referenced;
758}
759
760/**
761 * page_referenced_file - referenced check for object-based rmap
762 * @page: the page we're checking references on.
Johannes Weiner72835c82012-01-12 17:18:32 -0800763 * @memcg: target memory control group
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700764 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
Linus Torvalds1da177e2005-04-16 15:20:36 -0700765 *
766 * For an object-based mapped page, find all the places it is mapped and
767 * check/clear the referenced flag. This is done by following the page->mapping
768 * pointer, then walking the chain of vmas it holds. It returns the number
769 * of references it found.
770 *
771 * This function is only called from page_referenced for object-based pages.
772 */
Balbir Singhbed71612008-02-07 00:14:01 -0800773static int page_referenced_file(struct page *page,
Johannes Weiner72835c82012-01-12 17:18:32 -0800774 struct mem_cgroup *memcg,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700775 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700776{
777 unsigned int mapcount;
778 struct address_space *mapping = page->mapping;
779 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
780 struct vm_area_struct *vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700781 int referenced = 0;
782
783 /*
784 * The caller's checks on page->mapping and !PageAnon have made
785 * sure that this is a file page: the check for page->mapping
786 * excludes the case just before it gets set on an anon page.
787 */
788 BUG_ON(PageAnon(page));
789
790 /*
791 * The page lock not only makes sure that page->mapping cannot
792 * suddenly be NULLified by truncation, it makes sure that the
793 * structure at mapping cannot be freed and reused yet,
Peter Zijlstra3d48ae42011-05-24 17:12:06 -0700794 * so we can safely take mapping->i_mmap_mutex.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700795 */
796 BUG_ON(!PageLocked(page));
797
Peter Zijlstra3d48ae42011-05-24 17:12:06 -0700798 mutex_lock(&mapping->i_mmap_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700799
800 /*
Peter Zijlstra3d48ae42011-05-24 17:12:06 -0700801 * i_mmap_mutex does not stabilize mapcount at all, but mapcount
Linus Torvalds1da177e2005-04-16 15:20:36 -0700802 * is more likely to be accurate if we note it after spinning.
803 */
804 mapcount = page_mapcount(page);
805
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -0700806 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
Hugh Dickins1cb17292009-12-14 17:59:01 -0800807 unsigned long address = vma_address(page, vma);
Balbir Singhbed71612008-02-07 00:14:01 -0800808 /*
809 * If we are reclaiming on behalf of a cgroup, skip
810 * counting on behalf of references from different
811 * cgroups
812 */
Johannes Weiner72835c82012-01-12 17:18:32 -0800813 if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
Balbir Singhbed71612008-02-07 00:14:01 -0800814 continue;
Hugh Dickins1cb17292009-12-14 17:59:01 -0800815 referenced += page_referenced_one(page, vma, address,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700816 &mapcount, vm_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700817 if (!mapcount)
818 break;
819 }
820
Peter Zijlstra3d48ae42011-05-24 17:12:06 -0700821 mutex_unlock(&mapping->i_mmap_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700822 return referenced;
823}
824
825/**
826 * page_referenced - test if the page was referenced
827 * @page: the page to test
828 * @is_locked: caller holds lock on the page
Johannes Weiner72835c82012-01-12 17:18:32 -0800829 * @memcg: target memory cgroup
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700830 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
Linus Torvalds1da177e2005-04-16 15:20:36 -0700831 *
832 * Quick test_and_clear_referenced for all mappings to a page,
833 * returns the number of ptes which referenced the page.
834 */
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700835int page_referenced(struct page *page,
836 int is_locked,
Johannes Weiner72835c82012-01-12 17:18:32 -0800837 struct mem_cgroup *memcg,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700838 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700839{
840 int referenced = 0;
Hugh Dickins5ad64682009-12-14 17:59:24 -0800841 int we_locked = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700842
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700843 *vm_flags = 0;
Hugh Dickins3ca7b3c2009-12-14 17:58:57 -0800844 if (page_mapped(page) && page_rmapping(page)) {
Hugh Dickins5ad64682009-12-14 17:59:24 -0800845 if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
846 we_locked = trylock_page(page);
847 if (!we_locked) {
848 referenced++;
849 goto out;
850 }
851 }
852 if (unlikely(PageKsm(page)))
Johannes Weiner72835c82012-01-12 17:18:32 -0800853 referenced += page_referenced_ksm(page, memcg,
Hugh Dickins5ad64682009-12-14 17:59:24 -0800854 vm_flags);
855 else if (PageAnon(page))
Johannes Weiner72835c82012-01-12 17:18:32 -0800856 referenced += page_referenced_anon(page, memcg,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700857 vm_flags);
Hugh Dickins5ad64682009-12-14 17:59:24 -0800858 else if (page->mapping)
Johannes Weiner72835c82012-01-12 17:18:32 -0800859 referenced += page_referenced_file(page, memcg,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700860 vm_flags);
Hugh Dickins5ad64682009-12-14 17:59:24 -0800861 if (we_locked)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700862 unlock_page(page);
Martin Schwidefsky50a15982011-07-24 10:47:59 +0200863
864 if (page_test_and_clear_young(page_to_pfn(page)))
865 referenced++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700866 }
Hugh Dickins5ad64682009-12-14 17:59:24 -0800867out:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700868 return referenced;
869}
870
Hugh Dickins1cb17292009-12-14 17:59:01 -0800871static int page_mkclean_one(struct page *page, struct vm_area_struct *vma,
872 unsigned long address)
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700873{
874 struct mm_struct *mm = vma->vm_mm;
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100875 pte_t *pte;
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700876 spinlock_t *ptl;
877 int ret = 0;
878
Nick Piggin479db0b2008-08-20 14:09:18 -0700879 pte = page_check_address(page, mm, address, &ptl, 1);
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700880 if (!pte)
881 goto out;
882
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100883 if (pte_dirty(*pte) || pte_write(*pte)) {
884 pte_t entry;
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700885
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100886 flush_cache_page(vma, address, pte_pfn(*pte));
Sagi Grimberg2ec74c32012-10-08 16:33:33 -0700887 entry = ptep_clear_flush(vma, address, pte);
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100888 entry = pte_wrprotect(entry);
889 entry = pte_mkclean(entry);
Al Virod6e88e62006-12-29 16:48:35 -0800890 set_pte_at(mm, address, pte, entry);
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100891 ret = 1;
892 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700893
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700894 pte_unmap_unlock(pte, ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -0700895
896 if (ret)
897 mmu_notifier_invalidate_page(mm, address);
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700898out:
899 return ret;
900}
901
902static int page_mkclean_file(struct address_space *mapping, struct page *page)
903{
904 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
905 struct vm_area_struct *vma;
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700906 int ret = 0;
907
908 BUG_ON(PageAnon(page));
909
Peter Zijlstra3d48ae42011-05-24 17:12:06 -0700910 mutex_lock(&mapping->i_mmap_mutex);
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -0700911 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
Hugh Dickins1cb17292009-12-14 17:59:01 -0800912 if (vma->vm_flags & VM_SHARED) {
913 unsigned long address = vma_address(page, vma);
Hugh Dickins1cb17292009-12-14 17:59:01 -0800914 ret += page_mkclean_one(page, vma, address);
915 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700916 }
Peter Zijlstra3d48ae42011-05-24 17:12:06 -0700917 mutex_unlock(&mapping->i_mmap_mutex);
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700918 return ret;
919}
920
921int page_mkclean(struct page *page)
922{
923 int ret = 0;
924
925 BUG_ON(!PageLocked(page));
926
927 if (page_mapped(page)) {
928 struct address_space *mapping = page_mapping(page);
Christian Borntraegerce7e9fa2007-11-20 11:13:36 +0100929 if (mapping) {
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700930 ret = page_mkclean_file(mapping, page);
Martin Schwidefsky2d425522011-05-23 10:24:39 +0200931 if (page_test_and_clear_dirty(page_to_pfn(page), 1))
Christian Borntraegerce7e9fa2007-11-20 11:13:36 +0100932 ret = 1;
Martin Schwidefsky6c210482007-04-27 16:01:57 +0200933 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700934 }
935
936 return ret;
937}
Jaya Kumar60b59be2007-05-08 00:37:37 -0700938EXPORT_SYMBOL_GPL(page_mkclean);
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700939
Linus Torvalds1da177e2005-04-16 15:20:36 -0700940/**
Rik van Rielc44b6742010-03-05 13:42:09 -0800941 * page_move_anon_rmap - move a page to our anon_vma
942 * @page: the page to move to our anon_vma
943 * @vma: the vma the page belongs to
944 * @address: the user virtual address mapped
945 *
946 * When a page belongs exclusively to one process after a COW event,
947 * that page can be moved into the anon_vma that belongs to just that
948 * process, so the rmap code will not search the parent or sibling
949 * processes.
950 */
951void page_move_anon_rmap(struct page *page,
952 struct vm_area_struct *vma, unsigned long address)
953{
954 struct anon_vma *anon_vma = vma->anon_vma;
955
956 VM_BUG_ON(!PageLocked(page));
957 VM_BUG_ON(!anon_vma);
958 VM_BUG_ON(page->index != linear_page_index(vma, address));
959
960 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
961 page->mapping = (struct address_space *) anon_vma;
962}
963
964/**
Andi Kleen4e1c1972010-09-22 12:43:56 +0200965 * __page_set_anon_rmap - set up new anonymous rmap
966 * @page: Page to add to rmap
967 * @vma: VM area to add page to.
968 * @address: User virtual address of the mapping
Rik van Riele8a03fe2010-04-14 17:59:28 -0400969 * @exclusive: the page is exclusively owned by the current process
Nick Piggin9617d952006-01-06 00:11:12 -0800970 */
971static void __page_set_anon_rmap(struct page *page,
Rik van Riele8a03fe2010-04-14 17:59:28 -0400972 struct vm_area_struct *vma, unsigned long address, int exclusive)
Nick Piggin9617d952006-01-06 00:11:12 -0800973{
Rik van Riele8a03fe2010-04-14 17:59:28 -0400974 struct anon_vma *anon_vma = vma->anon_vma;
Nick Piggin9617d952006-01-06 00:11:12 -0800975
Rik van Riele8a03fe2010-04-14 17:59:28 -0400976 BUG_ON(!anon_vma);
Linus Torvaldsea900022010-04-12 12:44:29 -0700977
Andi Kleen4e1c1972010-09-22 12:43:56 +0200978 if (PageAnon(page))
979 return;
980
Linus Torvaldsea900022010-04-12 12:44:29 -0700981 /*
Rik van Riele8a03fe2010-04-14 17:59:28 -0400982 * If the page isn't exclusively mapped into this vma,
983 * we must use the _oldest_ possible anon_vma for the
984 * page mapping!
Linus Torvaldsea900022010-04-12 12:44:29 -0700985 */
Andi Kleen4e1c1972010-09-22 12:43:56 +0200986 if (!exclusive)
Andrea Arcangeli288468c2010-08-09 17:19:09 -0700987 anon_vma = anon_vma->root;
Linus Torvaldsea900022010-04-12 12:44:29 -0700988
Nick Piggin9617d952006-01-06 00:11:12 -0800989 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
990 page->mapping = (struct address_space *) anon_vma;
Nick Piggin9617d952006-01-06 00:11:12 -0800991 page->index = linear_page_index(vma, address);
Nick Piggin9617d952006-01-06 00:11:12 -0800992}
993
994/**
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700995 * __page_check_anon_rmap - sanity check anonymous rmap addition
Nick Pigginc97a9e12007-05-16 22:11:21 -0700996 * @page: the page to add the mapping to
997 * @vma: the vm area in which the mapping is added
998 * @address: the user virtual address mapped
999 */
1000static void __page_check_anon_rmap(struct page *page,
1001 struct vm_area_struct *vma, unsigned long address)
1002{
1003#ifdef CONFIG_DEBUG_VM
1004 /*
1005 * The page's anon-rmap details (mapping and index) are guaranteed to
1006 * be set up correctly at this point.
1007 *
1008 * We have exclusion against page_add_anon_rmap because the caller
1009 * always holds the page locked, except if called from page_dup_rmap,
1010 * in which case the page is already known to be setup.
1011 *
1012 * We have exclusion against page_add_new_anon_rmap because those pages
1013 * are initially only visible via the pagetables, and the pte is locked
1014 * over the call to page_add_new_anon_rmap.
1015 */
Andrea Arcangeli44ab57a2010-08-09 17:19:10 -07001016 BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root);
Nick Pigginc97a9e12007-05-16 22:11:21 -07001017 BUG_ON(page->index != linear_page_index(vma, address));
1018#endif
1019}
1020
1021/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07001022 * page_add_anon_rmap - add pte mapping to an anonymous page
1023 * @page: the page to add the mapping to
1024 * @vma: the vm area in which the mapping is added
1025 * @address: the user virtual address mapped
1026 *
Hugh Dickins5ad64682009-12-14 17:59:24 -08001027 * The caller needs to hold the pte lock, and the page must be locked in
Hugh Dickins80e148222009-12-14 17:59:29 -08001028 * the anon_vma case: to serialize mapping,index checking after setting,
1029 * and to ensure that PageAnon is not being upgraded racily to PageKsm
1030 * (but PageKsm is never downgraded to PageAnon).
Linus Torvalds1da177e2005-04-16 15:20:36 -07001031 */
1032void page_add_anon_rmap(struct page *page,
1033 struct vm_area_struct *vma, unsigned long address)
1034{
Rik van Rielad8c2ee2010-08-09 17:19:48 -07001035 do_page_add_anon_rmap(page, vma, address, 0);
1036}
1037
1038/*
1039 * Special version of the above for do_swap_page, which often runs
1040 * into pages that are exclusively owned by the current process.
1041 * Everybody else should continue to use page_add_anon_rmap above.
1042 */
1043void do_page_add_anon_rmap(struct page *page,
1044 struct vm_area_struct *vma, unsigned long address, int exclusive)
1045{
Hugh Dickins5ad64682009-12-14 17:59:24 -08001046 int first = atomic_inc_and_test(&page->_mapcount);
Andrea Arcangeli79134172011-01-13 15:46:58 -08001047 if (first) {
1048 if (!PageTransHuge(page))
1049 __inc_zone_page_state(page, NR_ANON_PAGES);
1050 else
1051 __inc_zone_page_state(page,
1052 NR_ANON_TRANSPARENT_HUGEPAGES);
1053 }
Hugh Dickins5ad64682009-12-14 17:59:24 -08001054 if (unlikely(PageKsm(page)))
1055 return;
1056
Nick Pigginc97a9e12007-05-16 22:11:21 -07001057 VM_BUG_ON(!PageLocked(page));
Hugh Dickins5dbe0af2011-05-28 13:17:04 -07001058 /* address might be in next vma when migration races vma_adjust */
Hugh Dickins5ad64682009-12-14 17:59:24 -08001059 if (first)
Rik van Rielad8c2ee2010-08-09 17:19:48 -07001060 __page_set_anon_rmap(page, vma, address, exclusive);
KAMEZAWA Hiroyuki69029cd2008-07-25 01:47:14 -07001061 else
Nick Pigginc97a9e12007-05-16 22:11:21 -07001062 __page_check_anon_rmap(page, vma, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001063}
1064
Randy Dunlap43d8eac2008-03-19 17:00:43 -07001065/**
Nick Piggin9617d952006-01-06 00:11:12 -08001066 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1067 * @page: the page to add the mapping to
1068 * @vma: the vm area in which the mapping is added
1069 * @address: the user virtual address mapped
1070 *
1071 * Same as page_add_anon_rmap but must only be called on *new* pages.
1072 * This means the inc-and-test can be bypassed.
Nick Pigginc97a9e12007-05-16 22:11:21 -07001073 * Page does not have to be locked.
Nick Piggin9617d952006-01-06 00:11:12 -08001074 */
1075void page_add_new_anon_rmap(struct page *page,
1076 struct vm_area_struct *vma, unsigned long address)
1077{
Hugh Dickinsb5934c52009-01-06 14:39:25 -08001078 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
Hugh Dickinscbf84b72009-01-06 14:39:27 -08001079 SetPageSwapBacked(page);
1080 atomic_set(&page->_mapcount, 0); /* increment count (starts at -1) */
Andrea Arcangeli79134172011-01-13 15:46:58 -08001081 if (!PageTransHuge(page))
1082 __inc_zone_page_state(page, NR_ANON_PAGES);
1083 else
1084 __inc_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
Rik van Riele8a03fe2010-04-14 17:59:28 -04001085 __page_set_anon_rmap(page, vma, address, 1);
Hugh Dickins39b5f292012-10-08 16:33:18 -07001086 if (!mlocked_vma_newpage(vma, page))
Hugh Dickinscbf84b72009-01-06 14:39:27 -08001087 lru_cache_add_lru(page, LRU_ACTIVE_ANON);
Hugh Dickinsb5934c52009-01-06 14:39:25 -08001088 else
1089 add_page_to_unevictable_list(page);
Nick Piggin9617d952006-01-06 00:11:12 -08001090}
1091
Linus Torvalds1da177e2005-04-16 15:20:36 -07001092/**
1093 * page_add_file_rmap - add pte mapping to a file page
1094 * @page: the page to add the mapping to
1095 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -07001096 * The caller needs to hold the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001097 */
1098void page_add_file_rmap(struct page *page)
1099{
KAMEZAWA Hiroyuki89c06bd2012-03-21 16:34:25 -07001100 bool locked;
1101 unsigned long flags;
1102
1103 mem_cgroup_begin_update_page_stat(page, &locked, &flags);
Balbir Singhd69b0422009-06-17 16:26:34 -07001104 if (atomic_inc_and_test(&page->_mapcount)) {
Christoph Lameter65ba55f2006-06-30 01:55:34 -07001105 __inc_zone_page_state(page, NR_FILE_MAPPED);
Greg Thelen2a7106f2011-01-13 15:47:37 -08001106 mem_cgroup_inc_page_stat(page, MEMCG_NR_FILE_MAPPED);
Balbir Singhd69b0422009-06-17 16:26:34 -07001107 }
KAMEZAWA Hiroyuki89c06bd2012-03-21 16:34:25 -07001108 mem_cgroup_end_update_page_stat(page, &locked, &flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001109}
1110
1111/**
1112 * page_remove_rmap - take down pte mapping from a page
1113 * @page: page to remove mapping from
1114 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -07001115 * The caller needs to hold the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001116 */
Hugh Dickinsedc315f2009-01-06 14:40:11 -08001117void page_remove_rmap(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001118{
KAMEZAWA Hiroyuki89c06bd2012-03-21 16:34:25 -07001119 bool anon = PageAnon(page);
1120 bool locked;
1121 unsigned long flags;
1122
1123 /*
1124 * The anon case has no mem_cgroup page_stat to update; but may
1125 * uncharge_page() below, where the lock ordering can deadlock if
1126 * we hold the lock against page_stat move: so avoid it on anon.
1127 */
1128 if (!anon)
1129 mem_cgroup_begin_update_page_stat(page, &locked, &flags);
1130
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -07001131 /* page still mapped by someone else? */
1132 if (!atomic_add_negative(-1, &page->_mapcount))
KAMEZAWA Hiroyuki89c06bd2012-03-21 16:34:25 -07001133 goto out;
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -07001134
1135 /*
1136 * Now that the last pte has gone, s390 must transfer dirty
1137 * flag from storage key to struct page. We can usually skip
1138 * this if the page is anon, so about to be freed; but perhaps
1139 * not if it's in swapcache - there might be another pte slot
1140 * containing the swap entry, but page not yet written to swap.
1141 */
KAMEZAWA Hiroyuki89c06bd2012-03-21 16:34:25 -07001142 if ((!anon || PageSwapCache(page)) &&
Martin Schwidefsky2d425522011-05-23 10:24:39 +02001143 page_test_and_clear_dirty(page_to_pfn(page), 1))
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -07001144 set_page_dirty(page);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001145 /*
1146 * Hugepages are not counted in NR_ANON_PAGES nor NR_FILE_MAPPED
1147 * and not charged by memcg for now.
1148 */
1149 if (unlikely(PageHuge(page)))
KAMEZAWA Hiroyuki89c06bd2012-03-21 16:34:25 -07001150 goto out;
1151 if (anon) {
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -07001152 mem_cgroup_uncharge_page(page);
Andrea Arcangeli79134172011-01-13 15:46:58 -08001153 if (!PageTransHuge(page))
1154 __dec_zone_page_state(page, NR_ANON_PAGES);
1155 else
1156 __dec_zone_page_state(page,
1157 NR_ANON_TRANSPARENT_HUGEPAGES);
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -07001158 } else {
1159 __dec_zone_page_state(page, NR_FILE_MAPPED);
Greg Thelen2a7106f2011-01-13 15:47:37 -08001160 mem_cgroup_dec_page_stat(page, MEMCG_NR_FILE_MAPPED);
Hugh Dickinse6c509f2012-10-08 16:33:19 -07001161 mem_cgroup_end_update_page_stat(page, &locked, &flags);
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -07001162 }
Hugh Dickinse6c509f2012-10-08 16:33:19 -07001163 if (unlikely(PageMlocked(page)))
1164 clear_page_mlock(page);
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -07001165 /*
1166 * It would be tidy to reset the PageAnon mapping here,
1167 * but that might overwrite a racing page_add_anon_rmap
1168 * which increments mapcount after us but sets mapping
1169 * before us: so leave the reset to free_hot_cold_page,
1170 * and remember that it's only reliable while mapped.
1171 * Leaving it set also helps swapoff to reinstate ptes
1172 * faster for those pages still in swapcache.
1173 */
Hugh Dickinse6c509f2012-10-08 16:33:19 -07001174 return;
KAMEZAWA Hiroyuki89c06bd2012-03-21 16:34:25 -07001175out:
1176 if (!anon)
1177 mem_cgroup_end_update_page_stat(page, &locked, &flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001178}
1179
1180/*
1181 * Subfunctions of try_to_unmap: try_to_unmap_one called
Wanlong Gao99ef0312011-10-31 17:08:51 -07001182 * repeatedly from try_to_unmap_ksm, try_to_unmap_anon or try_to_unmap_file.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001183 */
Hugh Dickins5ad64682009-12-14 17:59:24 -08001184int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
1185 unsigned long address, enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001186{
1187 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001188 pte_t *pte;
1189 pte_t pteval;
Hugh Dickinsc0718802005-10-29 18:16:31 -07001190 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001191 int ret = SWAP_AGAIN;
1192
Nick Piggin479db0b2008-08-20 14:09:18 -07001193 pte = page_check_address(page, mm, address, &ptl, 0);
Hugh Dickinsc0718802005-10-29 18:16:31 -07001194 if (!pte)
Nikita Danilov81b40822005-05-01 08:58:36 -07001195 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001196
1197 /*
1198 * If the page is mlock()d, we cannot swap it out.
1199 * If it's recently referenced (perhaps page_referenced
1200 * skipped over this mm) then we should reactivate it.
1201 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001202 if (!(flags & TTU_IGNORE_MLOCK)) {
KOSAKI Motohirocaed0f482009-12-14 17:59:45 -08001203 if (vma->vm_flags & VM_LOCKED)
1204 goto out_mlock;
1205
Hugh Dickinsaf8e3352009-12-14 17:58:59 -08001206 if (TTU_ACTION(flags) == TTU_MUNLOCK)
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001207 goto out_unmap;
Andi Kleen14fa31b2009-09-16 11:50:10 +02001208 }
1209 if (!(flags & TTU_IGNORE_ACCESS)) {
Nick Pigginb291f002008-10-18 20:26:44 -07001210 if (ptep_clear_flush_young_notify(vma, address, pte)) {
1211 ret = SWAP_FAIL;
1212 goto out_unmap;
1213 }
1214 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001215
Linus Torvalds1da177e2005-04-16 15:20:36 -07001216 /* Nuke the page table entry. */
1217 flush_cache_page(vma, address, page_to_pfn(page));
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001218 pteval = ptep_clear_flush(vma, address, pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001219
1220 /* Move the dirty bit to the physical page now the pte is gone. */
1221 if (pte_dirty(pteval))
1222 set_page_dirty(page);
1223
Hugh Dickins365e9c872005-10-29 18:16:18 -07001224 /* Update high watermark before we lower rss */
1225 update_hiwater_rss(mm);
1226
Andi Kleen888b9f72009-09-16 11:50:11 +02001227 if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
1228 if (PageAnon(page))
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001229 dec_mm_counter(mm, MM_ANONPAGES);
Andi Kleen888b9f72009-09-16 11:50:11 +02001230 else
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001231 dec_mm_counter(mm, MM_FILEPAGES);
Andi Kleen888b9f72009-09-16 11:50:11 +02001232 set_pte_at(mm, address, pte,
1233 swp_entry_to_pte(make_hwpoison_entry(page)));
1234 } else if (PageAnon(page)) {
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07001235 swp_entry_t entry = { .val = page_private(page) };
Christoph Lameter06972122006-06-23 02:03:35 -07001236
1237 if (PageSwapCache(page)) {
1238 /*
1239 * Store the swap location in the pte.
1240 * See handle_pte_fault() ...
1241 */
Hugh Dickins570a335b2009-12-14 17:58:46 -08001242 if (swap_duplicate(entry) < 0) {
1243 set_pte_at(mm, address, pte, pteval);
1244 ret = SWAP_FAIL;
1245 goto out_unmap;
1246 }
Christoph Lameter06972122006-06-23 02:03:35 -07001247 if (list_empty(&mm->mmlist)) {
1248 spin_lock(&mmlist_lock);
1249 if (list_empty(&mm->mmlist))
1250 list_add(&mm->mmlist, &init_mm.mmlist);
1251 spin_unlock(&mmlist_lock);
1252 }
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001253 dec_mm_counter(mm, MM_ANONPAGES);
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08001254 inc_mm_counter(mm, MM_SWAPENTS);
Konstantin Khlebnikovce1744f2012-03-21 16:33:59 -07001255 } else if (IS_ENABLED(CONFIG_MIGRATION)) {
Christoph Lameter06972122006-06-23 02:03:35 -07001256 /*
1257 * Store the pfn of the page in a special migration
1258 * pte. do_swap_page() will wait until the migration
1259 * pte is removed and then restart fault handling.
1260 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001261 BUG_ON(TTU_ACTION(flags) != TTU_MIGRATION);
Christoph Lameter06972122006-06-23 02:03:35 -07001262 entry = make_migration_entry(page, pte_write(pteval));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001263 }
1264 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
1265 BUG_ON(pte_file(*pte));
Konstantin Khlebnikovce1744f2012-03-21 16:33:59 -07001266 } else if (IS_ENABLED(CONFIG_MIGRATION) &&
1267 (TTU_ACTION(flags) == TTU_MIGRATION)) {
Christoph Lameter04e62a22006-06-23 02:03:38 -07001268 /* Establish migration entry for a file page */
1269 swp_entry_t entry;
1270 entry = make_migration_entry(page, pte_write(pteval));
1271 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
1272 } else
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001273 dec_mm_counter(mm, MM_FILEPAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001274
Hugh Dickinsedc315f2009-01-06 14:40:11 -08001275 page_remove_rmap(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001276 page_cache_release(page);
1277
1278out_unmap:
Hugh Dickinsc0718802005-10-29 18:16:31 -07001279 pte_unmap_unlock(pte, ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001280 if (ret != SWAP_FAIL)
1281 mmu_notifier_invalidate_page(mm, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001282out:
1283 return ret;
KOSAKI Motohirocaed0f482009-12-14 17:59:45 -08001284
1285out_mlock:
1286 pte_unmap_unlock(pte, ptl);
1287
1288
1289 /*
1290 * We need mmap_sem locking, Otherwise VM_LOCKED check makes
1291 * unstable result and race. Plus, We can't wait here because
Peter Zijlstra2b575eb2011-05-24 17:12:11 -07001292 * we now hold anon_vma->mutex or mapping->i_mmap_mutex.
KOSAKI Motohirocaed0f482009-12-14 17:59:45 -08001293 * if trylock failed, the page remain in evictable lru and later
1294 * vmscan could retry to move the page to unevictable lru if the
1295 * page is actually mlocked.
1296 */
1297 if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
1298 if (vma->vm_flags & VM_LOCKED) {
1299 mlock_vma_page(page);
1300 ret = SWAP_MLOCK;
1301 }
1302 up_read(&vma->vm_mm->mmap_sem);
1303 }
1304 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001305}
1306
1307/*
1308 * objrmap doesn't work for nonlinear VMAs because the assumption that
1309 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
1310 * Consequently, given a particular page and its ->index, we cannot locate the
1311 * ptes which are mapping that page without an exhaustive linear search.
1312 *
1313 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
1314 * maps the file to which the target page belongs. The ->vm_private_data field
1315 * holds the current cursor into that scan. Successive searches will circulate
1316 * around the vma's virtual address space.
1317 *
1318 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
1319 * more scanning pressure is placed against them as well. Eventually pages
1320 * will become fully unmapped and are eligible for eviction.
1321 *
1322 * For very sparsely populated VMAs this is a little inefficient - chances are
1323 * there there won't be many ptes located within the scan cluster. In this case
1324 * maybe we could scan further - to the end of the pte page, perhaps.
Nick Pigginb291f002008-10-18 20:26:44 -07001325 *
1326 * Mlocked pages: check VM_LOCKED under mmap_sem held for read, if we can
1327 * acquire it without blocking. If vma locked, mlock the pages in the cluster,
1328 * rather than unmapping them. If we encounter the "check_page" that vmscan is
1329 * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001330 */
1331#define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
1332#define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
1333
Nick Pigginb291f002008-10-18 20:26:44 -07001334static int try_to_unmap_cluster(unsigned long cursor, unsigned int *mapcount,
1335 struct vm_area_struct *vma, struct page *check_page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001336{
1337 struct mm_struct *mm = vma->vm_mm;
1338 pgd_t *pgd;
1339 pud_t *pud;
1340 pmd_t *pmd;
Hugh Dickinsc0718802005-10-29 18:16:31 -07001341 pte_t *pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001342 pte_t pteval;
Hugh Dickinsc0718802005-10-29 18:16:31 -07001343 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001344 struct page *page;
1345 unsigned long address;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001346 unsigned long mmun_start; /* For mmu_notifiers */
1347 unsigned long mmun_end; /* For mmu_notifiers */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001348 unsigned long end;
Nick Pigginb291f002008-10-18 20:26:44 -07001349 int ret = SWAP_AGAIN;
1350 int locked_vma = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001351
Linus Torvalds1da177e2005-04-16 15:20:36 -07001352 address = (vma->vm_start + cursor) & CLUSTER_MASK;
1353 end = address + CLUSTER_SIZE;
1354 if (address < vma->vm_start)
1355 address = vma->vm_start;
1356 if (end > vma->vm_end)
1357 end = vma->vm_end;
1358
1359 pgd = pgd_offset(mm, address);
1360 if (!pgd_present(*pgd))
Nick Pigginb291f002008-10-18 20:26:44 -07001361 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001362
1363 pud = pud_offset(pgd, address);
1364 if (!pud_present(*pud))
Nick Pigginb291f002008-10-18 20:26:44 -07001365 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001366
1367 pmd = pmd_offset(pud, address);
1368 if (!pmd_present(*pmd))
Nick Pigginb291f002008-10-18 20:26:44 -07001369 return ret;
1370
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001371 mmun_start = address;
1372 mmun_end = end;
1373 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1374
Nick Pigginb291f002008-10-18 20:26:44 -07001375 /*
Hugh Dickinsaf8e3352009-12-14 17:58:59 -08001376 * If we can acquire the mmap_sem for read, and vma is VM_LOCKED,
Nick Pigginb291f002008-10-18 20:26:44 -07001377 * keep the sem while scanning the cluster for mlocking pages.
1378 */
Hugh Dickinsaf8e3352009-12-14 17:58:59 -08001379 if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
Nick Pigginb291f002008-10-18 20:26:44 -07001380 locked_vma = (vma->vm_flags & VM_LOCKED);
1381 if (!locked_vma)
1382 up_read(&vma->vm_mm->mmap_sem); /* don't need it */
1383 }
Hugh Dickinsc0718802005-10-29 18:16:31 -07001384
1385 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001386
Hugh Dickins365e9c872005-10-29 18:16:18 -07001387 /* Update high watermark before we lower rss */
1388 update_hiwater_rss(mm);
1389
Hugh Dickinsc0718802005-10-29 18:16:31 -07001390 for (; address < end; pte++, address += PAGE_SIZE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001391 if (!pte_present(*pte))
1392 continue;
Linus Torvalds6aab3412005-11-28 14:34:23 -08001393 page = vm_normal_page(vma, address, *pte);
1394 BUG_ON(!page || PageAnon(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001395
Nick Pigginb291f002008-10-18 20:26:44 -07001396 if (locked_vma) {
1397 mlock_vma_page(page); /* no-op if already mlocked */
1398 if (page == check_page)
1399 ret = SWAP_MLOCK;
1400 continue; /* don't unmap */
1401 }
1402
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001403 if (ptep_clear_flush_young_notify(vma, address, pte))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001404 continue;
1405
1406 /* Nuke the page table entry. */
Ben Collinseca35132005-11-29 11:45:26 -08001407 flush_cache_page(vma, address, pte_pfn(*pte));
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001408 pteval = ptep_clear_flush(vma, address, pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001409
1410 /* If nonlinear, store the file page offset in the pte. */
1411 if (page->index != linear_page_index(vma, address))
1412 set_pte_at(mm, address, pte, pgoff_to_pte(page->index));
1413
1414 /* Move the dirty bit to the physical page now the pte is gone. */
1415 if (pte_dirty(pteval))
1416 set_page_dirty(page);
1417
Hugh Dickinsedc315f2009-01-06 14:40:11 -08001418 page_remove_rmap(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001419 page_cache_release(page);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001420 dec_mm_counter(mm, MM_FILEPAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001421 (*mapcount)--;
1422 }
Hugh Dickinsc0718802005-10-29 18:16:31 -07001423 pte_unmap_unlock(pte - 1, ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001424 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Nick Pigginb291f002008-10-18 20:26:44 -07001425 if (locked_vma)
1426 up_read(&vma->vm_mm->mmap_sem);
1427 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001428}
1429
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001430bool is_vma_temporary_stack(struct vm_area_struct *vma)
Mel Gormana8bef8ff2010-05-24 14:32:24 -07001431{
1432 int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP);
1433
1434 if (!maybe_stack)
1435 return false;
1436
1437 if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) ==
1438 VM_STACK_INCOMPLETE_SETUP)
1439 return true;
1440
1441 return false;
1442}
1443
Nick Pigginb291f002008-10-18 20:26:44 -07001444/**
1445 * try_to_unmap_anon - unmap or unlock anonymous page using the object-based
1446 * rmap method
1447 * @page: the page to unmap/unlock
Huang Shijie8051be52009-12-14 17:58:50 -08001448 * @flags: action and flags
Nick Pigginb291f002008-10-18 20:26:44 -07001449 *
1450 * Find all the mappings of a page using the mapping pointer and the vma chains
1451 * contained in the anon_vma struct it points to.
1452 *
1453 * This function is only called from try_to_unmap/try_to_munlock for
1454 * anonymous pages.
1455 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1456 * where the page was found will be held for write. So, we won't recheck
1457 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1458 * 'LOCKED.
1459 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001460static int try_to_unmap_anon(struct page *page, enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001461{
1462 struct anon_vma *anon_vma;
Michel Lespinassebf181b92012-10-08 16:31:39 -07001463 pgoff_t pgoff;
Rik van Riel5beb4932010-03-05 13:42:07 -08001464 struct anon_vma_chain *avc;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001465 int ret = SWAP_AGAIN;
Nick Pigginb291f002008-10-18 20:26:44 -07001466
Linus Torvalds1da177e2005-04-16 15:20:36 -07001467 anon_vma = page_lock_anon_vma(page);
1468 if (!anon_vma)
1469 return ret;
1470
Michel Lespinassebf181b92012-10-08 16:31:39 -07001471 pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1472 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
Rik van Riel5beb4932010-03-05 13:42:07 -08001473 struct vm_area_struct *vma = avc->vma;
Mel Gormana8bef8ff2010-05-24 14:32:24 -07001474 unsigned long address;
1475
1476 /*
1477 * During exec, a temporary VMA is setup and later moved.
1478 * The VMA is moved under the anon_vma lock but not the
1479 * page tables leading to a race where migration cannot
1480 * find the migration ptes. Rather than increasing the
1481 * locking requirements of exec(), migration skips
1482 * temporary VMAs until after exec() completes.
1483 */
Konstantin Khlebnikovce1744f2012-03-21 16:33:59 -07001484 if (IS_ENABLED(CONFIG_MIGRATION) && (flags & TTU_MIGRATION) &&
Mel Gormana8bef8ff2010-05-24 14:32:24 -07001485 is_vma_temporary_stack(vma))
1486 continue;
1487
1488 address = vma_address(page, vma);
Hugh Dickins1cb17292009-12-14 17:59:01 -08001489 ret = try_to_unmap_one(page, vma, address, flags);
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001490 if (ret != SWAP_AGAIN || !page_mapped(page))
1491 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001492 }
Oleg Nesterov34bbd702007-02-28 20:13:49 -08001493
1494 page_unlock_anon_vma(anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001495 return ret;
1496}
1497
1498/**
Nick Pigginb291f002008-10-18 20:26:44 -07001499 * try_to_unmap_file - unmap/unlock file page using the object-based rmap method
1500 * @page: the page to unmap/unlock
Andi Kleen14fa31b2009-09-16 11:50:10 +02001501 * @flags: action and flags
Linus Torvalds1da177e2005-04-16 15:20:36 -07001502 *
1503 * Find all the mappings of a page using the mapping pointer and the vma chains
1504 * contained in the address_space struct it points to.
1505 *
Nick Pigginb291f002008-10-18 20:26:44 -07001506 * This function is only called from try_to_unmap/try_to_munlock for
1507 * object-based pages.
1508 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1509 * where the page was found will be held for write. So, we won't recheck
1510 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1511 * 'LOCKED.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001512 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001513static int try_to_unmap_file(struct page *page, enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001514{
1515 struct address_space *mapping = page->mapping;
1516 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1517 struct vm_area_struct *vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001518 int ret = SWAP_AGAIN;
1519 unsigned long cursor;
1520 unsigned long max_nl_cursor = 0;
1521 unsigned long max_nl_size = 0;
1522 unsigned int mapcount;
1523
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07001524 mutex_lock(&mapping->i_mmap_mutex);
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07001525 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
Hugh Dickins1cb17292009-12-14 17:59:01 -08001526 unsigned long address = vma_address(page, vma);
Hugh Dickins1cb17292009-12-14 17:59:01 -08001527 ret = try_to_unmap_one(page, vma, address, flags);
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001528 if (ret != SWAP_AGAIN || !page_mapped(page))
1529 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001530 }
1531
1532 if (list_empty(&mapping->i_mmap_nonlinear))
1533 goto out;
1534
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001535 /*
1536 * We don't bother to try to find the munlocked page in nonlinears.
1537 * It's costly. Instead, later, page reclaim logic may call
1538 * try_to_unmap(TTU_MUNLOCK) and recover PG_mlocked lazily.
1539 */
1540 if (TTU_ACTION(flags) == TTU_MUNLOCK)
1541 goto out;
1542
Linus Torvalds1da177e2005-04-16 15:20:36 -07001543 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07001544 shared.nonlinear) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001545 cursor = (unsigned long) vma->vm_private_data;
1546 if (cursor > max_nl_cursor)
1547 max_nl_cursor = cursor;
1548 cursor = vma->vm_end - vma->vm_start;
1549 if (cursor > max_nl_size)
1550 max_nl_size = cursor;
1551 }
1552
Nick Pigginb291f002008-10-18 20:26:44 -07001553 if (max_nl_size == 0) { /* all nonlinears locked or reserved ? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001554 ret = SWAP_FAIL;
1555 goto out;
1556 }
1557
1558 /*
1559 * We don't try to search for this page in the nonlinear vmas,
1560 * and page_referenced wouldn't have found it anyway. Instead
1561 * just walk the nonlinear vmas trying to age and unmap some.
1562 * The mapcount of the page we came in with is irrelevant,
1563 * but even so use it as a guide to how hard we should try?
1564 */
1565 mapcount = page_mapcount(page);
1566 if (!mapcount)
1567 goto out;
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07001568 cond_resched();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001569
1570 max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
1571 if (max_nl_cursor == 0)
1572 max_nl_cursor = CLUSTER_SIZE;
1573
1574 do {
1575 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07001576 shared.nonlinear) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001577 cursor = (unsigned long) vma->vm_private_data;
Hugh Dickins839b9682005-09-03 15:54:43 -07001578 while ( cursor < max_nl_cursor &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07001579 cursor < vma->vm_end - vma->vm_start) {
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001580 if (try_to_unmap_cluster(cursor, &mapcount,
1581 vma, page) == SWAP_MLOCK)
1582 ret = SWAP_MLOCK;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001583 cursor += CLUSTER_SIZE;
1584 vma->vm_private_data = (void *) cursor;
1585 if ((int)mapcount <= 0)
1586 goto out;
1587 }
1588 vma->vm_private_data = (void *) max_nl_cursor;
1589 }
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07001590 cond_resched();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001591 max_nl_cursor += CLUSTER_SIZE;
1592 } while (max_nl_cursor <= max_nl_size);
1593
1594 /*
1595 * Don't loop forever (perhaps all the remaining pages are
1596 * in locked vmas). Reset cursor on all unreserved nonlinear
1597 * vmas, now forgetting on which ones it had fallen behind.
1598 */
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07001599 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.nonlinear)
Hugh Dickins101d2be2005-11-21 21:32:16 -08001600 vma->vm_private_data = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001601out:
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07001602 mutex_unlock(&mapping->i_mmap_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001603 return ret;
1604}
1605
1606/**
1607 * try_to_unmap - try to remove all page table mappings to a page
1608 * @page: the page to get unmapped
Andi Kleen14fa31b2009-09-16 11:50:10 +02001609 * @flags: action and flags
Linus Torvalds1da177e2005-04-16 15:20:36 -07001610 *
1611 * Tries to remove all the page table entries which are mapping this
1612 * page, used in the pageout path. Caller must hold the page lock.
1613 * Return values are:
1614 *
1615 * SWAP_SUCCESS - we succeeded in removing all mappings
1616 * SWAP_AGAIN - we missed a mapping, try again later
1617 * SWAP_FAIL - the page is unswappable
Nick Pigginb291f002008-10-18 20:26:44 -07001618 * SWAP_MLOCK - page is mlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001619 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001620int try_to_unmap(struct page *page, enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001621{
1622 int ret;
1623
Linus Torvalds1da177e2005-04-16 15:20:36 -07001624 BUG_ON(!PageLocked(page));
Andrea Arcangeli91600e92011-01-13 15:47:16 -08001625 VM_BUG_ON(!PageHuge(page) && PageTransHuge(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001626
Hugh Dickins5ad64682009-12-14 17:59:24 -08001627 if (unlikely(PageKsm(page)))
1628 ret = try_to_unmap_ksm(page, flags);
1629 else if (PageAnon(page))
Andi Kleen14fa31b2009-09-16 11:50:10 +02001630 ret = try_to_unmap_anon(page, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001631 else
Andi Kleen14fa31b2009-09-16 11:50:10 +02001632 ret = try_to_unmap_file(page, flags);
Nick Pigginb291f002008-10-18 20:26:44 -07001633 if (ret != SWAP_MLOCK && !page_mapped(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001634 ret = SWAP_SUCCESS;
1635 return ret;
1636}
Nikita Danilov81b40822005-05-01 08:58:36 -07001637
Nick Pigginb291f002008-10-18 20:26:44 -07001638/**
1639 * try_to_munlock - try to munlock a page
1640 * @page: the page to be munlocked
1641 *
1642 * Called from munlock code. Checks all of the VMAs mapping the page
1643 * to make sure nobody else has this page mlocked. The page will be
1644 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1645 *
1646 * Return values are:
1647 *
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001648 * SWAP_AGAIN - no vma is holding page mlocked, or,
Nick Pigginb291f002008-10-18 20:26:44 -07001649 * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem
Hugh Dickins5ad64682009-12-14 17:59:24 -08001650 * SWAP_FAIL - page cannot be located at present
Nick Pigginb291f002008-10-18 20:26:44 -07001651 * SWAP_MLOCK - page is now mlocked.
1652 */
1653int try_to_munlock(struct page *page)
1654{
1655 VM_BUG_ON(!PageLocked(page) || PageLRU(page));
1656
Hugh Dickins5ad64682009-12-14 17:59:24 -08001657 if (unlikely(PageKsm(page)))
1658 return try_to_unmap_ksm(page, TTU_MUNLOCK);
1659 else if (PageAnon(page))
Andi Kleen14fa31b2009-09-16 11:50:10 +02001660 return try_to_unmap_anon(page, TTU_MUNLOCK);
Nick Pigginb291f002008-10-18 20:26:44 -07001661 else
Andi Kleen14fa31b2009-09-16 11:50:10 +02001662 return try_to_unmap_file(page, TTU_MUNLOCK);
Nick Pigginb291f002008-10-18 20:26:44 -07001663}
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001664
Peter Zijlstra01d8b202011-03-22 16:32:49 -07001665void __put_anon_vma(struct anon_vma *anon_vma)
Rik van Riel76545062010-08-09 17:18:41 -07001666{
Peter Zijlstra01d8b202011-03-22 16:32:49 -07001667 struct anon_vma *root = anon_vma->root;
Rik van Riel76545062010-08-09 17:18:41 -07001668
Peter Zijlstra01d8b202011-03-22 16:32:49 -07001669 if (root != anon_vma && atomic_dec_and_test(&root->refcount))
1670 anon_vma_free(root);
Rik van Riel76545062010-08-09 17:18:41 -07001671
Peter Zijlstra01d8b202011-03-22 16:32:49 -07001672 anon_vma_free(anon_vma);
Rik van Riel76545062010-08-09 17:18:41 -07001673}
Rik van Riel76545062010-08-09 17:18:41 -07001674
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001675#ifdef CONFIG_MIGRATION
1676/*
1677 * rmap_walk() and its helpers rmap_walk_anon() and rmap_walk_file():
1678 * Called by migrate.c to remove migration ptes, but might be used more later.
1679 */
1680static int rmap_walk_anon(struct page *page, int (*rmap_one)(struct page *,
1681 struct vm_area_struct *, unsigned long, void *), void *arg)
1682{
1683 struct anon_vma *anon_vma;
Michel Lespinassebf181b92012-10-08 16:31:39 -07001684 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
Rik van Riel5beb4932010-03-05 13:42:07 -08001685 struct anon_vma_chain *avc;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001686 int ret = SWAP_AGAIN;
1687
1688 /*
1689 * Note: remove_migration_ptes() cannot use page_lock_anon_vma()
1690 * because that depends on page_mapped(); but not all its usages
Mel Gorman3f6c8272010-05-24 14:32:17 -07001691 * are holding mmap_sem. Users without mmap_sem are required to
1692 * take a reference count to prevent the anon_vma disappearing
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001693 */
1694 anon_vma = page_anon_vma(page);
1695 if (!anon_vma)
1696 return ret;
Rik van Rielcba48b92010-08-09 17:18:38 -07001697 anon_vma_lock(anon_vma);
Michel Lespinassebf181b92012-10-08 16:31:39 -07001698 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
Rik van Riel5beb4932010-03-05 13:42:07 -08001699 struct vm_area_struct *vma = avc->vma;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001700 unsigned long address = vma_address(page, vma);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001701 ret = rmap_one(page, vma, address, arg);
1702 if (ret != SWAP_AGAIN)
1703 break;
1704 }
Rik van Rielcba48b92010-08-09 17:18:38 -07001705 anon_vma_unlock(anon_vma);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001706 return ret;
1707}
1708
1709static int rmap_walk_file(struct page *page, int (*rmap_one)(struct page *,
1710 struct vm_area_struct *, unsigned long, void *), void *arg)
1711{
1712 struct address_space *mapping = page->mapping;
1713 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1714 struct vm_area_struct *vma;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001715 int ret = SWAP_AGAIN;
1716
1717 if (!mapping)
1718 return ret;
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07001719 mutex_lock(&mapping->i_mmap_mutex);
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07001720 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001721 unsigned long address = vma_address(page, vma);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001722 ret = rmap_one(page, vma, address, arg);
1723 if (ret != SWAP_AGAIN)
1724 break;
1725 }
1726 /*
1727 * No nonlinear handling: being always shared, nonlinear vmas
1728 * never contain migration ptes. Decide what to do about this
1729 * limitation to linear when we need rmap_walk() on nonlinear.
1730 */
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07001731 mutex_unlock(&mapping->i_mmap_mutex);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001732 return ret;
1733}
1734
1735int rmap_walk(struct page *page, int (*rmap_one)(struct page *,
1736 struct vm_area_struct *, unsigned long, void *), void *arg)
1737{
1738 VM_BUG_ON(!PageLocked(page));
1739
1740 if (unlikely(PageKsm(page)))
1741 return rmap_walk_ksm(page, rmap_one, arg);
1742 else if (PageAnon(page))
1743 return rmap_walk_anon(page, rmap_one, arg);
1744 else
1745 return rmap_walk_file(page, rmap_one, arg);
1746}
1747#endif /* CONFIG_MIGRATION */
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001748
Naoya Horiguchie3390f62010-06-15 13:18:13 +09001749#ifdef CONFIG_HUGETLB_PAGE
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001750/*
1751 * The following three functions are for anonymous (private mapped) hugepages.
1752 * Unlike common anonymous pages, anonymous hugepages have no accounting code
1753 * and no lru code, because we handle hugepages differently from common pages.
1754 */
1755static void __hugepage_set_anon_rmap(struct page *page,
1756 struct vm_area_struct *vma, unsigned long address, int exclusive)
1757{
1758 struct anon_vma *anon_vma = vma->anon_vma;
Naoya Horiguchi433abed2010-09-10 13:23:03 +09001759
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001760 BUG_ON(!anon_vma);
Naoya Horiguchi433abed2010-09-10 13:23:03 +09001761
1762 if (PageAnon(page))
1763 return;
1764 if (!exclusive)
1765 anon_vma = anon_vma->root;
1766
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001767 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1768 page->mapping = (struct address_space *) anon_vma;
1769 page->index = linear_page_index(vma, address);
1770}
1771
1772void hugepage_add_anon_rmap(struct page *page,
1773 struct vm_area_struct *vma, unsigned long address)
1774{
1775 struct anon_vma *anon_vma = vma->anon_vma;
1776 int first;
Naoya Horiguchia850ea32010-09-10 13:23:06 +09001777
1778 BUG_ON(!PageLocked(page));
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001779 BUG_ON(!anon_vma);
Hugh Dickins5dbe0af2011-05-28 13:17:04 -07001780 /* address might be in next vma when migration races vma_adjust */
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001781 first = atomic_inc_and_test(&page->_mapcount);
1782 if (first)
1783 __hugepage_set_anon_rmap(page, vma, address, 0);
1784}
1785
1786void hugepage_add_new_anon_rmap(struct page *page,
1787 struct vm_area_struct *vma, unsigned long address)
1788{
1789 BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1790 atomic_set(&page->_mapcount, 0);
1791 __hugepage_set_anon_rmap(page, vma, address, 1);
1792}
Naoya Horiguchie3390f62010-06-15 13:18:13 +09001793#endif /* CONFIG_HUGETLB_PAGE */