blob: f6f0d2dda2eae8480860cf57f5a9cfce69820716 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * mm/rmap.c - physical to virtual reverse mappings
3 *
4 * Copyright 2001, Rik van Riel <riel@conectiva.com.br>
5 * Released under the General Public License (GPL).
6 *
7 * Simple, low overhead reverse mapping scheme.
8 * Please try to keep this thing as modular as possible.
9 *
10 * Provides methods for unmapping each kind of mapped page:
11 * the anon methods track anonymous pages, and
12 * the file methods track pages belonging to an inode.
13 *
14 * Original design by Rik van Riel <riel@conectiva.com.br> 2001
15 * File methods by Dave McCracken <dmccr@us.ibm.com> 2003, 2004
16 * Anonymous methods by Andrea Arcangeli <andrea@suse.de> 2004
Hugh Dickins98f32602009-05-21 20:33:58 +010017 * Contributions by Hugh Dickins 2003, 2004
Linus Torvalds1da177e2005-04-16 15:20:36 -070018 */
19
20/*
21 * Lock ordering in mm:
22 *
Jes Sorensen1b1dcc12006-01-09 15:59:24 -080023 * inode->i_mutex (while writing or truncating, not reading or faulting)
Nick Piggin82591e62006-10-19 23:29:10 -070024 * inode->i_alloc_sem (vmtruncate_range)
25 * mm->mmap_sem
26 * page->flags PG_locked (lock_page)
27 * mapping->i_mmap_lock
28 * anon_vma->lock
29 * mm->page_table_lock or pte_lock
30 * zone->lru_lock (in mark_page_accessed, isolate_lru_page)
31 * swap_lock (in swap_duplicate, swap_info_get)
32 * mmlist_lock (in mmput, drain_mmlist and others)
33 * mapping->private_lock (in __set_page_dirty_buffers)
34 * inode_lock (in set_page_dirty's __mark_inode_dirty)
35 * sb_lock (within inode_lock in fs/fs-writeback.c)
36 * mapping->tree_lock (widely used, in set_page_dirty,
37 * in arch-dependent flush_dcache_mmap_lock,
38 * within inode_lock in __sync_single_inode)
Andi Kleen6a460792009-09-16 11:50:15 +020039 *
40 * (code doesn't rely on that order so it could be switched around)
41 * ->tasklist_lock
42 * anon_vma->lock (memory_failure, collect_procs_anon)
43 * pte map lock
Linus Torvalds1da177e2005-04-16 15:20:36 -070044 */
45
46#include <linux/mm.h>
47#include <linux/pagemap.h>
48#include <linux/swap.h>
49#include <linux/swapops.h>
50#include <linux/slab.h>
51#include <linux/init.h>
Hugh Dickins5ad64682009-12-14 17:59:24 -080052#include <linux/ksm.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070053#include <linux/rmap.h>
54#include <linux/rcupdate.h>
Christoph Lametera48d07a2006-02-01 03:05:38 -080055#include <linux/module.h>
Balbir Singh8a9f3cc2008-02-07 00:13:53 -080056#include <linux/memcontrol.h>
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070057#include <linux/mmu_notifier.h>
KOSAKI Motohiro64cdd542009-01-06 14:39:16 -080058#include <linux/migrate.h>
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +090059#include <linux/hugetlb.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070060
61#include <asm/tlbflush.h>
62
Nick Pigginb291f002008-10-18 20:26:44 -070063#include "internal.h"
64
Adrian Bunkfdd2e5f2008-10-18 20:28:38 -070065static struct kmem_cache *anon_vma_cachep;
Rik van Riel5beb4932010-03-05 13:42:07 -080066static struct kmem_cache *anon_vma_chain_cachep;
Adrian Bunkfdd2e5f2008-10-18 20:28:38 -070067
68static inline struct anon_vma *anon_vma_alloc(void)
69{
70 return kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
71}
72
Hugh Dickinsdb114b82009-12-14 17:59:25 -080073void anon_vma_free(struct anon_vma *anon_vma)
Adrian Bunkfdd2e5f2008-10-18 20:28:38 -070074{
75 kmem_cache_free(anon_vma_cachep, anon_vma);
76}
Linus Torvalds1da177e2005-04-16 15:20:36 -070077
Rik van Riel5beb4932010-03-05 13:42:07 -080078static inline struct anon_vma_chain *anon_vma_chain_alloc(void)
79{
80 return kmem_cache_alloc(anon_vma_chain_cachep, GFP_KERNEL);
81}
82
83void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
84{
85 kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
86}
87
Linus Torvaldsd9d332e2008-10-19 10:32:20 -070088/**
89 * anon_vma_prepare - attach an anon_vma to a memory region
90 * @vma: the memory region in question
91 *
92 * This makes sure the memory mapping described by 'vma' has
93 * an 'anon_vma' attached to it, so that we can associate the
94 * anonymous pages mapped into it with that anon_vma.
95 *
96 * The common case will be that we already have one, but if
97 * if not we either need to find an adjacent mapping that we
98 * can re-use the anon_vma from (very common when the only
99 * reason for splitting a vma has been mprotect()), or we
100 * allocate a new one.
101 *
102 * Anon-vma allocations are very subtle, because we may have
103 * optimistically looked up an anon_vma in page_lock_anon_vma()
104 * and that may actually touch the spinlock even in the newly
105 * allocated vma (it depends on RCU to make sure that the
106 * anon_vma isn't actually destroyed).
107 *
108 * As a result, we need to do proper anon_vma locking even
109 * for the new allocation. At the same time, we do not want
110 * to do any locking for the common case of already having
111 * an anon_vma.
112 *
113 * This must be called with the mmap_sem held for reading.
114 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700115int anon_vma_prepare(struct vm_area_struct *vma)
116{
117 struct anon_vma *anon_vma = vma->anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -0800118 struct anon_vma_chain *avc;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700119
120 might_sleep();
121 if (unlikely(!anon_vma)) {
122 struct mm_struct *mm = vma->vm_mm;
Linus Torvaldsd9d332e2008-10-19 10:32:20 -0700123 struct anon_vma *allocated;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700124
Rik van Riel5beb4932010-03-05 13:42:07 -0800125 avc = anon_vma_chain_alloc();
126 if (!avc)
127 goto out_enomem;
128
Linus Torvalds1da177e2005-04-16 15:20:36 -0700129 anon_vma = find_mergeable_anon_vma(vma);
Linus Torvaldsd9d332e2008-10-19 10:32:20 -0700130 allocated = NULL;
131 if (!anon_vma) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700132 anon_vma = anon_vma_alloc();
133 if (unlikely(!anon_vma))
Rik van Riel5beb4932010-03-05 13:42:07 -0800134 goto out_enomem_free_avc;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700135 allocated = anon_vma;
Rik van Riel5c341ee12010-08-09 17:18:39 -0700136 /*
137 * This VMA had no anon_vma yet. This anon_vma is
138 * the root of any anon_vma tree that might form.
139 */
140 anon_vma->root = anon_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700141 }
142
Rik van Rielcba48b92010-08-09 17:18:38 -0700143 anon_vma_lock(anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700144 /* page_table_lock to protect against threads */
145 spin_lock(&mm->page_table_lock);
146 if (likely(!vma->anon_vma)) {
147 vma->anon_vma = anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -0800148 avc->anon_vma = anon_vma;
149 avc->vma = vma;
150 list_add(&avc->same_vma, &vma->anon_vma_chain);
Andrea Arcangeli26ba0cb2010-08-09 17:19:08 -0700151 list_add_tail(&avc->same_anon_vma, &anon_vma->head);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700152 allocated = NULL;
Oleg Nesterov31f2b0e2010-04-23 13:18:01 -0400153 avc = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700154 }
155 spin_unlock(&mm->page_table_lock);
Rik van Rielcba48b92010-08-09 17:18:38 -0700156 anon_vma_unlock(anon_vma);
Oleg Nesterov31f2b0e2010-04-23 13:18:01 -0400157
158 if (unlikely(allocated))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700159 anon_vma_free(allocated);
Oleg Nesterov31f2b0e2010-04-23 13:18:01 -0400160 if (unlikely(avc))
Rik van Riel5beb4932010-03-05 13:42:07 -0800161 anon_vma_chain_free(avc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700162 }
163 return 0;
Rik van Riel5beb4932010-03-05 13:42:07 -0800164
165 out_enomem_free_avc:
166 anon_vma_chain_free(avc);
167 out_enomem:
168 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700169}
170
Rik van Riel5beb4932010-03-05 13:42:07 -0800171static void anon_vma_chain_link(struct vm_area_struct *vma,
172 struct anon_vma_chain *avc,
173 struct anon_vma *anon_vma)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700174{
Rik van Riel5beb4932010-03-05 13:42:07 -0800175 avc->vma = vma;
176 avc->anon_vma = anon_vma;
177 list_add(&avc->same_vma, &vma->anon_vma_chain);
178
Rik van Rielcba48b92010-08-09 17:18:38 -0700179 anon_vma_lock(anon_vma);
Rik van Riel5beb4932010-03-05 13:42:07 -0800180 list_add_tail(&avc->same_anon_vma, &anon_vma->head);
Rik van Rielcba48b92010-08-09 17:18:38 -0700181 anon_vma_unlock(anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700182}
183
Rik van Riel5beb4932010-03-05 13:42:07 -0800184/*
185 * Attach the anon_vmas from src to dst.
186 * Returns 0 on success, -ENOMEM on failure.
187 */
188int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700189{
Rik van Riel5beb4932010-03-05 13:42:07 -0800190 struct anon_vma_chain *avc, *pavc;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700191
Linus Torvalds646d87b2010-04-11 17:15:03 -0700192 list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
Rik van Riel5beb4932010-03-05 13:42:07 -0800193 avc = anon_vma_chain_alloc();
194 if (!avc)
195 goto enomem_failure;
196 anon_vma_chain_link(dst, avc, pavc->anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700197 }
Rik van Riel5beb4932010-03-05 13:42:07 -0800198 return 0;
199
200 enomem_failure:
201 unlink_anon_vmas(dst);
202 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700203}
204
Rik van Riel5beb4932010-03-05 13:42:07 -0800205/*
206 * Attach vma to its own anon_vma, as well as to the anon_vmas that
207 * the corresponding VMA in the parent process is attached to.
208 * Returns 0 on success, non-zero on failure.
209 */
210int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700211{
Rik van Riel5beb4932010-03-05 13:42:07 -0800212 struct anon_vma_chain *avc;
213 struct anon_vma *anon_vma;
214
215 /* Don't bother if the parent process has no anon_vma here. */
216 if (!pvma->anon_vma)
217 return 0;
218
219 /*
220 * First, attach the new VMA to the parent VMA's anon_vmas,
221 * so rmap can find non-COWed pages in child processes.
222 */
223 if (anon_vma_clone(vma, pvma))
224 return -ENOMEM;
225
226 /* Then add our own anon_vma. */
227 anon_vma = anon_vma_alloc();
228 if (!anon_vma)
229 goto out_error;
230 avc = anon_vma_chain_alloc();
231 if (!avc)
232 goto out_error_free_anon_vma;
Rik van Riel5c341ee12010-08-09 17:18:39 -0700233
234 /*
235 * The root anon_vma's spinlock is the lock actually used when we
236 * lock any of the anon_vmas in this anon_vma tree.
237 */
238 anon_vma->root = pvma->anon_vma->root;
Rik van Riel76545062010-08-09 17:18:41 -0700239 /*
240 * With KSM refcounts, an anon_vma can stay around longer than the
241 * process it belongs to. The root anon_vma needs to be pinned
242 * until this anon_vma is freed, because the lock lives in the root.
243 */
244 get_anon_vma(anon_vma->root);
Rik van Riel5beb4932010-03-05 13:42:07 -0800245 /* Mark this anon_vma as the one where our new (COWed) pages go. */
246 vma->anon_vma = anon_vma;
Rik van Riel5c341ee12010-08-09 17:18:39 -0700247 anon_vma_chain_link(vma, avc, anon_vma);
Rik van Riel5beb4932010-03-05 13:42:07 -0800248
249 return 0;
250
251 out_error_free_anon_vma:
252 anon_vma_free(anon_vma);
253 out_error:
Rik van Riel4946d542010-04-05 12:13:33 -0400254 unlink_anon_vmas(vma);
Rik van Riel5beb4932010-03-05 13:42:07 -0800255 return -ENOMEM;
256}
257
258static void anon_vma_unlink(struct anon_vma_chain *anon_vma_chain)
259{
260 struct anon_vma *anon_vma = anon_vma_chain->anon_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700261 int empty;
262
Rik van Riel5beb4932010-03-05 13:42:07 -0800263 /* If anon_vma_fork fails, we can get an empty anon_vma_chain. */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700264 if (!anon_vma)
265 return;
266
Rik van Rielcba48b92010-08-09 17:18:38 -0700267 anon_vma_lock(anon_vma);
Rik van Riel5beb4932010-03-05 13:42:07 -0800268 list_del(&anon_vma_chain->same_anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700269
270 /* We must garbage collect the anon_vma if it's empty */
Mel Gorman7f60c212010-05-24 14:32:18 -0700271 empty = list_empty(&anon_vma->head) && !anonvma_external_refcount(anon_vma);
Rik van Rielcba48b92010-08-09 17:18:38 -0700272 anon_vma_unlock(anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700273
Rik van Riel76545062010-08-09 17:18:41 -0700274 if (empty) {
275 /* We no longer need the root anon_vma */
276 if (anon_vma->root != anon_vma)
277 drop_anon_vma(anon_vma->root);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700278 anon_vma_free(anon_vma);
Rik van Riel76545062010-08-09 17:18:41 -0700279 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700280}
281
Rik van Riel5beb4932010-03-05 13:42:07 -0800282void unlink_anon_vmas(struct vm_area_struct *vma)
283{
284 struct anon_vma_chain *avc, *next;
285
Rik van Riel5c341ee12010-08-09 17:18:39 -0700286 /*
287 * Unlink each anon_vma chained to the VMA. This list is ordered
288 * from newest to oldest, ensuring the root anon_vma gets freed last.
289 */
Rik van Riel5beb4932010-03-05 13:42:07 -0800290 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
291 anon_vma_unlink(avc);
292 list_del(&avc->same_vma);
293 anon_vma_chain_free(avc);
294 }
295}
296
Alexey Dobriyan51cc5062008-07-25 19:45:34 -0700297static void anon_vma_ctor(void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700298{
Christoph Lametera35afb82007-05-16 22:10:57 -0700299 struct anon_vma *anon_vma = data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700300
Christoph Lametera35afb82007-05-16 22:10:57 -0700301 spin_lock_init(&anon_vma->lock);
Mel Gorman7f60c212010-05-24 14:32:18 -0700302 anonvma_external_refcount_init(anon_vma);
Christoph Lametera35afb82007-05-16 22:10:57 -0700303 INIT_LIST_HEAD(&anon_vma->head);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700304}
305
306void __init anon_vma_init(void)
307{
308 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
Paul Mundt20c2df82007-07-20 10:11:58 +0900309 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor);
Rik van Riel5beb4932010-03-05 13:42:07 -0800310 anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain, SLAB_PANIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700311}
312
313/*
314 * Getting a lock on a stable anon_vma from a page off the LRU is
315 * tricky: page_lock_anon_vma rely on RCU to guard against the races.
316 */
Andi Kleen10be22d2009-09-16 11:50:04 +0200317struct anon_vma *page_lock_anon_vma(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700318{
Hugh Dickinsf18194272010-08-25 23:12:54 -0700319 struct anon_vma *anon_vma, *root_anon_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700320 unsigned long anon_mapping;
321
322 rcu_read_lock();
Hugh Dickins80e148222009-12-14 17:59:29 -0800323 anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping);
Hugh Dickins3ca7b3c2009-12-14 17:58:57 -0800324 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700325 goto out;
326 if (!page_mapped(page))
327 goto out;
328
329 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
Hugh Dickinsf18194272010-08-25 23:12:54 -0700330 root_anon_vma = ACCESS_ONCE(anon_vma->root);
331 spin_lock(&root_anon_vma->lock);
332
333 /*
334 * If this page is still mapped, then its anon_vma cannot have been
335 * freed. But if it has been unmapped, we have no security against
336 * the anon_vma structure being freed and reused (for another anon_vma:
337 * SLAB_DESTROY_BY_RCU guarantees that - so the spin_lock above cannot
338 * corrupt): with anon_vma_prepare() or anon_vma_fork() redirecting
339 * anon_vma->root before page_unlock_anon_vma() is called to unlock.
340 */
341 if (page_mapped(page))
342 return anon_vma;
343
344 spin_unlock(&root_anon_vma->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700345out:
346 rcu_read_unlock();
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800347 return NULL;
348}
349
Andi Kleen10be22d2009-09-16 11:50:04 +0200350void page_unlock_anon_vma(struct anon_vma *anon_vma)
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800351{
Rik van Rielcba48b92010-08-09 17:18:38 -0700352 anon_vma_unlock(anon_vma);
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800353 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700354}
355
356/*
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -0800357 * At what user virtual address is page expected in @vma?
358 * Returns virtual address or -EFAULT if page's index/offset is not
359 * within the range mapped the @vma.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700360 */
361static inline unsigned long
362vma_address(struct page *page, struct vm_area_struct *vma)
363{
364 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
365 unsigned long address;
366
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900367 if (unlikely(is_vm_hugetlb_page(vma)))
368 pgoff = page->index << huge_page_order(page_hstate(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700369 address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
370 if (unlikely(address < vma->vm_start || address >= vma->vm_end)) {
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -0800371 /* page should be within @vma mapping range */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700372 return -EFAULT;
373 }
374 return address;
375}
376
377/*
Huang Shijiebf89c8c2009-10-01 15:44:04 -0700378 * At what user virtual address is page expected in vma?
Naoya Horiguchiab941e02010-05-11 14:06:55 -0700379 * Caller should check the page is actually part of the vma.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700380 */
381unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
382{
Andrea Arcangeli21d0d442010-08-09 17:19:10 -0700383 if (PageAnon(page)) {
384 if (vma->anon_vma->root != page_anon_vma(page)->root)
385 return -EFAULT;
386 } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) {
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800387 if (!vma->vm_file ||
388 vma->vm_file->f_mapping != page->mapping)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700389 return -EFAULT;
390 } else
391 return -EFAULT;
392 return vma_address(page, vma);
393}
394
395/*
Nikita Danilov81b40822005-05-01 08:58:36 -0700396 * Check that @page is mapped at @address into @mm.
397 *
Nick Piggin479db0b2008-08-20 14:09:18 -0700398 * If @sync is false, page_check_address may perform a racy check to avoid
399 * the page table lock when the pte is not present (helpful when reclaiming
400 * highly shared pages).
401 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -0700402 * On success returns with pte mapped and locked.
Nikita Danilov81b40822005-05-01 08:58:36 -0700403 */
Carsten Otteceffc072005-06-23 22:05:25 -0700404pte_t *page_check_address(struct page *page, struct mm_struct *mm,
Nick Piggin479db0b2008-08-20 14:09:18 -0700405 unsigned long address, spinlock_t **ptlp, int sync)
Nikita Danilov81b40822005-05-01 08:58:36 -0700406{
407 pgd_t *pgd;
408 pud_t *pud;
409 pmd_t *pmd;
410 pte_t *pte;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700411 spinlock_t *ptl;
Nikita Danilov81b40822005-05-01 08:58:36 -0700412
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900413 if (unlikely(PageHuge(page))) {
414 pte = huge_pte_offset(mm, address);
415 ptl = &mm->page_table_lock;
416 goto check;
417 }
418
Nikita Danilov81b40822005-05-01 08:58:36 -0700419 pgd = pgd_offset(mm, address);
Hugh Dickinsc0718802005-10-29 18:16:31 -0700420 if (!pgd_present(*pgd))
421 return NULL;
422
423 pud = pud_offset(pgd, address);
424 if (!pud_present(*pud))
425 return NULL;
426
427 pmd = pmd_offset(pud, address);
428 if (!pmd_present(*pmd))
429 return NULL;
430
431 pte = pte_offset_map(pmd, address);
432 /* Make a quick check before getting the lock */
Nick Piggin479db0b2008-08-20 14:09:18 -0700433 if (!sync && !pte_present(*pte)) {
Hugh Dickinsc0718802005-10-29 18:16:31 -0700434 pte_unmap(pte);
435 return NULL;
Nikita Danilov81b40822005-05-01 08:58:36 -0700436 }
Hugh Dickinsc0718802005-10-29 18:16:31 -0700437
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700438 ptl = pte_lockptr(mm, pmd);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900439check:
Hugh Dickinsc0718802005-10-29 18:16:31 -0700440 spin_lock(ptl);
441 if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
442 *ptlp = ptl;
443 return pte;
444 }
445 pte_unmap_unlock(pte, ptl);
446 return NULL;
Nikita Danilov81b40822005-05-01 08:58:36 -0700447}
448
Nick Pigginb291f002008-10-18 20:26:44 -0700449/**
450 * page_mapped_in_vma - check whether a page is really mapped in a VMA
451 * @page: the page to test
452 * @vma: the VMA to test
453 *
454 * Returns 1 if the page is mapped into the page tables of the VMA, 0
455 * if the page is not mapped into the page tables of this VMA. Only
456 * valid for normal file or anonymous VMAs.
457 */
Andi Kleen6a460792009-09-16 11:50:15 +0200458int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
Nick Pigginb291f002008-10-18 20:26:44 -0700459{
460 unsigned long address;
461 pte_t *pte;
462 spinlock_t *ptl;
463
464 address = vma_address(page, vma);
465 if (address == -EFAULT) /* out of vma range */
466 return 0;
467 pte = page_check_address(page, vma->vm_mm, address, &ptl, 1);
468 if (!pte) /* the page is not in this mm */
469 return 0;
470 pte_unmap_unlock(pte, ptl);
471
472 return 1;
473}
474
Nikita Danilov81b40822005-05-01 08:58:36 -0700475/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700476 * Subfunctions of page_referenced: page_referenced_one called
477 * repeatedly from either page_referenced_anon or page_referenced_file.
478 */
Hugh Dickins5ad64682009-12-14 17:59:24 -0800479int page_referenced_one(struct page *page, struct vm_area_struct *vma,
480 unsigned long address, unsigned int *mapcount,
481 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700482{
483 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700484 pte_t *pte;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700485 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700486 int referenced = 0;
487
Nick Piggin479db0b2008-08-20 14:09:18 -0700488 pte = page_check_address(page, mm, address, &ptl, 0);
Hugh Dickinsc0718802005-10-29 18:16:31 -0700489 if (!pte)
490 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700491
Nick Pigginb291f002008-10-18 20:26:44 -0700492 /*
493 * Don't want to elevate referenced for mlocked page that gets this far,
494 * in order that it progresses to try_to_unmap and is moved to the
495 * unevictable list.
496 */
Hugh Dickins5a9bbdc2008-02-04 22:29:23 -0800497 if (vma->vm_flags & VM_LOCKED) {
Hugh Dickins5a9bbdc2008-02-04 22:29:23 -0800498 *mapcount = 1; /* break early from loop */
Minchan Kim03ef83a2009-08-26 14:29:23 -0700499 *vm_flags |= VM_LOCKED;
Nick Pigginb291f002008-10-18 20:26:44 -0700500 goto out_unmap;
501 }
502
Johannes Weiner4917e5d2009-01-06 14:39:17 -0800503 if (ptep_clear_flush_young_notify(vma, address, pte)) {
504 /*
505 * Don't treat a reference through a sequentially read
506 * mapping as such. If the page has been used in
507 * another mapping, we will catch it; if this other
508 * mapping is already gone, the unmap path will have
509 * set PG_referenced or activated the page.
510 */
511 if (likely(!VM_SequentialReadHint(vma)))
512 referenced++;
513 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700514
Hugh Dickinsc0718802005-10-29 18:16:31 -0700515 /* Pretend the page is referenced if the task has the
516 swap token and is in the middle of a page fault. */
Rik van Rielf7b7fd82005-11-28 13:44:07 -0800517 if (mm != current->mm && has_swap_token(mm) &&
Hugh Dickinsc0718802005-10-29 18:16:31 -0700518 rwsem_is_locked(&mm->mmap_sem))
519 referenced++;
520
Nick Pigginb291f002008-10-18 20:26:44 -0700521out_unmap:
Hugh Dickinsc0718802005-10-29 18:16:31 -0700522 (*mapcount)--;
523 pte_unmap_unlock(pte, ptl);
Huang Shijie273f0472009-12-14 17:58:51 -0800524
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700525 if (referenced)
526 *vm_flags |= vma->vm_flags;
Huang Shijie273f0472009-12-14 17:58:51 -0800527out:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700528 return referenced;
529}
530
Balbir Singhbed71612008-02-07 00:14:01 -0800531static int page_referenced_anon(struct page *page,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700532 struct mem_cgroup *mem_cont,
533 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700534{
535 unsigned int mapcount;
536 struct anon_vma *anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -0800537 struct anon_vma_chain *avc;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700538 int referenced = 0;
539
540 anon_vma = page_lock_anon_vma(page);
541 if (!anon_vma)
542 return referenced;
543
544 mapcount = page_mapcount(page);
Rik van Riel5beb4932010-03-05 13:42:07 -0800545 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
546 struct vm_area_struct *vma = avc->vma;
Hugh Dickins1cb17292009-12-14 17:59:01 -0800547 unsigned long address = vma_address(page, vma);
548 if (address == -EFAULT)
549 continue;
Balbir Singhbed71612008-02-07 00:14:01 -0800550 /*
551 * If we are reclaiming on behalf of a cgroup, skip
552 * counting on behalf of references from different
553 * cgroups
554 */
Hugh Dickinsbd845e32008-03-04 14:29:01 -0800555 if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
Balbir Singhbed71612008-02-07 00:14:01 -0800556 continue;
Hugh Dickins1cb17292009-12-14 17:59:01 -0800557 referenced += page_referenced_one(page, vma, address,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700558 &mapcount, vm_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700559 if (!mapcount)
560 break;
561 }
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800562
563 page_unlock_anon_vma(anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700564 return referenced;
565}
566
567/**
568 * page_referenced_file - referenced check for object-based rmap
569 * @page: the page we're checking references on.
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700570 * @mem_cont: target memory controller
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700571 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
Linus Torvalds1da177e2005-04-16 15:20:36 -0700572 *
573 * For an object-based mapped page, find all the places it is mapped and
574 * check/clear the referenced flag. This is done by following the page->mapping
575 * pointer, then walking the chain of vmas it holds. It returns the number
576 * of references it found.
577 *
578 * This function is only called from page_referenced for object-based pages.
579 */
Balbir Singhbed71612008-02-07 00:14:01 -0800580static int page_referenced_file(struct page *page,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700581 struct mem_cgroup *mem_cont,
582 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700583{
584 unsigned int mapcount;
585 struct address_space *mapping = page->mapping;
586 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
587 struct vm_area_struct *vma;
588 struct prio_tree_iter iter;
589 int referenced = 0;
590
591 /*
592 * The caller's checks on page->mapping and !PageAnon have made
593 * sure that this is a file page: the check for page->mapping
594 * excludes the case just before it gets set on an anon page.
595 */
596 BUG_ON(PageAnon(page));
597
598 /*
599 * The page lock not only makes sure that page->mapping cannot
600 * suddenly be NULLified by truncation, it makes sure that the
601 * structure at mapping cannot be freed and reused yet,
602 * so we can safely take mapping->i_mmap_lock.
603 */
604 BUG_ON(!PageLocked(page));
605
606 spin_lock(&mapping->i_mmap_lock);
607
608 /*
609 * i_mmap_lock does not stabilize mapcount at all, but mapcount
610 * is more likely to be accurate if we note it after spinning.
611 */
612 mapcount = page_mapcount(page);
613
614 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
Hugh Dickins1cb17292009-12-14 17:59:01 -0800615 unsigned long address = vma_address(page, vma);
616 if (address == -EFAULT)
617 continue;
Balbir Singhbed71612008-02-07 00:14:01 -0800618 /*
619 * If we are reclaiming on behalf of a cgroup, skip
620 * counting on behalf of references from different
621 * cgroups
622 */
Hugh Dickinsbd845e32008-03-04 14:29:01 -0800623 if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
Balbir Singhbed71612008-02-07 00:14:01 -0800624 continue;
Hugh Dickins1cb17292009-12-14 17:59:01 -0800625 referenced += page_referenced_one(page, vma, address,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700626 &mapcount, vm_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700627 if (!mapcount)
628 break;
629 }
630
631 spin_unlock(&mapping->i_mmap_lock);
632 return referenced;
633}
634
635/**
636 * page_referenced - test if the page was referenced
637 * @page: the page to test
638 * @is_locked: caller holds lock on the page
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700639 * @mem_cont: target memory controller
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700640 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
Linus Torvalds1da177e2005-04-16 15:20:36 -0700641 *
642 * Quick test_and_clear_referenced for all mappings to a page,
643 * returns the number of ptes which referenced the page.
644 */
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700645int page_referenced(struct page *page,
646 int is_locked,
647 struct mem_cgroup *mem_cont,
648 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700649{
650 int referenced = 0;
Hugh Dickins5ad64682009-12-14 17:59:24 -0800651 int we_locked = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700652
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700653 *vm_flags = 0;
Hugh Dickins3ca7b3c2009-12-14 17:58:57 -0800654 if (page_mapped(page) && page_rmapping(page)) {
Hugh Dickins5ad64682009-12-14 17:59:24 -0800655 if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
656 we_locked = trylock_page(page);
657 if (!we_locked) {
658 referenced++;
659 goto out;
660 }
661 }
662 if (unlikely(PageKsm(page)))
663 referenced += page_referenced_ksm(page, mem_cont,
664 vm_flags);
665 else if (PageAnon(page))
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700666 referenced += page_referenced_anon(page, mem_cont,
667 vm_flags);
Hugh Dickins5ad64682009-12-14 17:59:24 -0800668 else if (page->mapping)
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700669 referenced += page_referenced_file(page, mem_cont,
670 vm_flags);
Hugh Dickins5ad64682009-12-14 17:59:24 -0800671 if (we_locked)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700672 unlock_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700673 }
Hugh Dickins5ad64682009-12-14 17:59:24 -0800674out:
Christian Borntraeger5b7baf02008-03-25 18:47:12 +0100675 if (page_test_and_clear_young(page))
676 referenced++;
677
Linus Torvalds1da177e2005-04-16 15:20:36 -0700678 return referenced;
679}
680
Hugh Dickins1cb17292009-12-14 17:59:01 -0800681static int page_mkclean_one(struct page *page, struct vm_area_struct *vma,
682 unsigned long address)
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700683{
684 struct mm_struct *mm = vma->vm_mm;
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100685 pte_t *pte;
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700686 spinlock_t *ptl;
687 int ret = 0;
688
Nick Piggin479db0b2008-08-20 14:09:18 -0700689 pte = page_check_address(page, mm, address, &ptl, 1);
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700690 if (!pte)
691 goto out;
692
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100693 if (pte_dirty(*pte) || pte_write(*pte)) {
694 pte_t entry;
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700695
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100696 flush_cache_page(vma, address, pte_pfn(*pte));
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700697 entry = ptep_clear_flush_notify(vma, address, pte);
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100698 entry = pte_wrprotect(entry);
699 entry = pte_mkclean(entry);
Al Virod6e88e62006-12-29 16:48:35 -0800700 set_pte_at(mm, address, pte, entry);
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100701 ret = 1;
702 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700703
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700704 pte_unmap_unlock(pte, ptl);
705out:
706 return ret;
707}
708
709static int page_mkclean_file(struct address_space *mapping, struct page *page)
710{
711 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
712 struct vm_area_struct *vma;
713 struct prio_tree_iter iter;
714 int ret = 0;
715
716 BUG_ON(PageAnon(page));
717
718 spin_lock(&mapping->i_mmap_lock);
719 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
Hugh Dickins1cb17292009-12-14 17:59:01 -0800720 if (vma->vm_flags & VM_SHARED) {
721 unsigned long address = vma_address(page, vma);
722 if (address == -EFAULT)
723 continue;
724 ret += page_mkclean_one(page, vma, address);
725 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700726 }
727 spin_unlock(&mapping->i_mmap_lock);
728 return ret;
729}
730
731int page_mkclean(struct page *page)
732{
733 int ret = 0;
734
735 BUG_ON(!PageLocked(page));
736
737 if (page_mapped(page)) {
738 struct address_space *mapping = page_mapping(page);
Christian Borntraegerce7e9fa2007-11-20 11:13:36 +0100739 if (mapping) {
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700740 ret = page_mkclean_file(mapping, page);
Christian Borntraegerce7e9fa2007-11-20 11:13:36 +0100741 if (page_test_dirty(page)) {
742 page_clear_dirty(page);
743 ret = 1;
744 }
Martin Schwidefsky6c210482007-04-27 16:01:57 +0200745 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700746 }
747
748 return ret;
749}
Jaya Kumar60b59be2007-05-08 00:37:37 -0700750EXPORT_SYMBOL_GPL(page_mkclean);
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700751
Linus Torvalds1da177e2005-04-16 15:20:36 -0700752/**
Rik van Rielc44b6742010-03-05 13:42:09 -0800753 * page_move_anon_rmap - move a page to our anon_vma
754 * @page: the page to move to our anon_vma
755 * @vma: the vma the page belongs to
756 * @address: the user virtual address mapped
757 *
758 * When a page belongs exclusively to one process after a COW event,
759 * that page can be moved into the anon_vma that belongs to just that
760 * process, so the rmap code will not search the parent or sibling
761 * processes.
762 */
763void page_move_anon_rmap(struct page *page,
764 struct vm_area_struct *vma, unsigned long address)
765{
766 struct anon_vma *anon_vma = vma->anon_vma;
767
768 VM_BUG_ON(!PageLocked(page));
769 VM_BUG_ON(!anon_vma);
770 VM_BUG_ON(page->index != linear_page_index(vma, address));
771
772 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
773 page->mapping = (struct address_space *) anon_vma;
774}
775
776/**
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700777 * __page_set_anon_rmap - setup new anonymous rmap
Nick Piggin9617d952006-01-06 00:11:12 -0800778 * @page: the page to add the mapping to
779 * @vma: the vm area in which the mapping is added
780 * @address: the user virtual address mapped
Rik van Riele8a03fe2010-04-14 17:59:28 -0400781 * @exclusive: the page is exclusively owned by the current process
Nick Piggin9617d952006-01-06 00:11:12 -0800782 */
783static void __page_set_anon_rmap(struct page *page,
Rik van Riele8a03fe2010-04-14 17:59:28 -0400784 struct vm_area_struct *vma, unsigned long address, int exclusive)
Nick Piggin9617d952006-01-06 00:11:12 -0800785{
Rik van Riele8a03fe2010-04-14 17:59:28 -0400786 struct anon_vma *anon_vma = vma->anon_vma;
Nick Piggin9617d952006-01-06 00:11:12 -0800787
Rik van Riele8a03fe2010-04-14 17:59:28 -0400788 BUG_ON(!anon_vma);
Linus Torvaldsea900022010-04-12 12:44:29 -0700789
790 /*
Rik van Riele8a03fe2010-04-14 17:59:28 -0400791 * If the page isn't exclusively mapped into this vma,
792 * we must use the _oldest_ possible anon_vma for the
793 * page mapping!
Linus Torvaldsea900022010-04-12 12:44:29 -0700794 */
Rik van Riele8a03fe2010-04-14 17:59:28 -0400795 if (!exclusive) {
Andrea Arcangeli288468c2010-08-09 17:19:09 -0700796 if (PageAnon(page))
797 return;
798 anon_vma = anon_vma->root;
799 } else {
800 /*
801 * In this case, swapped-out-but-not-discarded swap-cache
802 * is remapped. So, no need to update page->mapping here.
803 * We convice anon_vma poitned by page->mapping is not obsolete
804 * because vma->anon_vma is necessary to be a family of it.
805 */
806 if (PageAnon(page))
807 return;
Rik van Riele8a03fe2010-04-14 17:59:28 -0400808 }
Linus Torvaldsea900022010-04-12 12:44:29 -0700809
Nick Piggin9617d952006-01-06 00:11:12 -0800810 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
811 page->mapping = (struct address_space *) anon_vma;
Nick Piggin9617d952006-01-06 00:11:12 -0800812 page->index = linear_page_index(vma, address);
Nick Piggin9617d952006-01-06 00:11:12 -0800813}
814
815/**
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700816 * __page_check_anon_rmap - sanity check anonymous rmap addition
Nick Pigginc97a9e12007-05-16 22:11:21 -0700817 * @page: the page to add the mapping to
818 * @vma: the vm area in which the mapping is added
819 * @address: the user virtual address mapped
820 */
821static void __page_check_anon_rmap(struct page *page,
822 struct vm_area_struct *vma, unsigned long address)
823{
824#ifdef CONFIG_DEBUG_VM
825 /*
826 * The page's anon-rmap details (mapping and index) are guaranteed to
827 * be set up correctly at this point.
828 *
829 * We have exclusion against page_add_anon_rmap because the caller
830 * always holds the page locked, except if called from page_dup_rmap,
831 * in which case the page is already known to be setup.
832 *
833 * We have exclusion against page_add_new_anon_rmap because those pages
834 * are initially only visible via the pagetables, and the pte is locked
835 * over the call to page_add_new_anon_rmap.
836 */
Andrea Arcangeli44ab57a2010-08-09 17:19:10 -0700837 BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root);
Nick Pigginc97a9e12007-05-16 22:11:21 -0700838 BUG_ON(page->index != linear_page_index(vma, address));
839#endif
840}
841
842/**
Linus Torvalds1da177e2005-04-16 15:20:36 -0700843 * page_add_anon_rmap - add pte mapping to an anonymous page
844 * @page: the page to add the mapping to
845 * @vma: the vm area in which the mapping is added
846 * @address: the user virtual address mapped
847 *
Hugh Dickins5ad64682009-12-14 17:59:24 -0800848 * The caller needs to hold the pte lock, and the page must be locked in
Hugh Dickins80e148222009-12-14 17:59:29 -0800849 * the anon_vma case: to serialize mapping,index checking after setting,
850 * and to ensure that PageAnon is not being upgraded racily to PageKsm
851 * (but PageKsm is never downgraded to PageAnon).
Linus Torvalds1da177e2005-04-16 15:20:36 -0700852 */
853void page_add_anon_rmap(struct page *page,
854 struct vm_area_struct *vma, unsigned long address)
855{
Rik van Rielad8c2ee2010-08-09 17:19:48 -0700856 do_page_add_anon_rmap(page, vma, address, 0);
857}
858
859/*
860 * Special version of the above for do_swap_page, which often runs
861 * into pages that are exclusively owned by the current process.
862 * Everybody else should continue to use page_add_anon_rmap above.
863 */
864void do_page_add_anon_rmap(struct page *page,
865 struct vm_area_struct *vma, unsigned long address, int exclusive)
866{
Hugh Dickins5ad64682009-12-14 17:59:24 -0800867 int first = atomic_inc_and_test(&page->_mapcount);
868 if (first)
869 __inc_zone_page_state(page, NR_ANON_PAGES);
870 if (unlikely(PageKsm(page)))
871 return;
872
Nick Pigginc97a9e12007-05-16 22:11:21 -0700873 VM_BUG_ON(!PageLocked(page));
874 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
Hugh Dickins5ad64682009-12-14 17:59:24 -0800875 if (first)
Rik van Rielad8c2ee2010-08-09 17:19:48 -0700876 __page_set_anon_rmap(page, vma, address, exclusive);
KAMEZAWA Hiroyuki69029cd2008-07-25 01:47:14 -0700877 else
Nick Pigginc97a9e12007-05-16 22:11:21 -0700878 __page_check_anon_rmap(page, vma, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700879}
880
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700881/**
Nick Piggin9617d952006-01-06 00:11:12 -0800882 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
883 * @page: the page to add the mapping to
884 * @vma: the vm area in which the mapping is added
885 * @address: the user virtual address mapped
886 *
887 * Same as page_add_anon_rmap but must only be called on *new* pages.
888 * This means the inc-and-test can be bypassed.
Nick Pigginc97a9e12007-05-16 22:11:21 -0700889 * Page does not have to be locked.
Nick Piggin9617d952006-01-06 00:11:12 -0800890 */
891void page_add_new_anon_rmap(struct page *page,
892 struct vm_area_struct *vma, unsigned long address)
893{
Hugh Dickinsb5934c52009-01-06 14:39:25 -0800894 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
Hugh Dickinscbf84b72009-01-06 14:39:27 -0800895 SetPageSwapBacked(page);
896 atomic_set(&page->_mapcount, 0); /* increment count (starts at -1) */
Hugh Dickins5ad64682009-12-14 17:59:24 -0800897 __inc_zone_page_state(page, NR_ANON_PAGES);
Rik van Riele8a03fe2010-04-14 17:59:28 -0400898 __page_set_anon_rmap(page, vma, address, 1);
Hugh Dickinsb5934c52009-01-06 14:39:25 -0800899 if (page_evictable(page, vma))
Hugh Dickinscbf84b72009-01-06 14:39:27 -0800900 lru_cache_add_lru(page, LRU_ACTIVE_ANON);
Hugh Dickinsb5934c52009-01-06 14:39:25 -0800901 else
902 add_page_to_unevictable_list(page);
Nick Piggin9617d952006-01-06 00:11:12 -0800903}
904
Linus Torvalds1da177e2005-04-16 15:20:36 -0700905/**
906 * page_add_file_rmap - add pte mapping to a file page
907 * @page: the page to add the mapping to
908 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -0700909 * The caller needs to hold the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700910 */
911void page_add_file_rmap(struct page *page)
912{
Balbir Singhd69b0422009-06-17 16:26:34 -0700913 if (atomic_inc_and_test(&page->_mapcount)) {
Christoph Lameter65ba55f2006-06-30 01:55:34 -0700914 __inc_zone_page_state(page, NR_FILE_MAPPED);
KAMEZAWA Hiroyukid8046582009-12-15 16:47:09 -0800915 mem_cgroup_update_file_mapped(page, 1);
Balbir Singhd69b0422009-06-17 16:26:34 -0700916 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700917}
918
919/**
920 * page_remove_rmap - take down pte mapping from a page
921 * @page: page to remove mapping from
922 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -0700923 * The caller needs to hold the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700924 */
Hugh Dickinsedc315f2009-01-06 14:40:11 -0800925void page_remove_rmap(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700926{
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -0700927 /* page still mapped by someone else? */
928 if (!atomic_add_negative(-1, &page->_mapcount))
929 return;
930
931 /*
932 * Now that the last pte has gone, s390 must transfer dirty
933 * flag from storage key to struct page. We can usually skip
934 * this if the page is anon, so about to be freed; but perhaps
935 * not if it's in swapcache - there might be another pte slot
936 * containing the swap entry, but page not yet written to swap.
937 */
938 if ((!PageAnon(page) || PageSwapCache(page)) && page_test_dirty(page)) {
939 page_clear_dirty(page);
940 set_page_dirty(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700941 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900942 /*
943 * Hugepages are not counted in NR_ANON_PAGES nor NR_FILE_MAPPED
944 * and not charged by memcg for now.
945 */
946 if (unlikely(PageHuge(page)))
947 return;
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -0700948 if (PageAnon(page)) {
949 mem_cgroup_uncharge_page(page);
950 __dec_zone_page_state(page, NR_ANON_PAGES);
951 } else {
952 __dec_zone_page_state(page, NR_FILE_MAPPED);
KAMEZAWA Hiroyukid8046582009-12-15 16:47:09 -0800953 mem_cgroup_update_file_mapped(page, -1);
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -0700954 }
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -0700955 /*
956 * It would be tidy to reset the PageAnon mapping here,
957 * but that might overwrite a racing page_add_anon_rmap
958 * which increments mapcount after us but sets mapping
959 * before us: so leave the reset to free_hot_cold_page,
960 * and remember that it's only reliable while mapped.
961 * Leaving it set also helps swapoff to reinstate ptes
962 * faster for those pages still in swapcache.
963 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700964}
965
966/*
967 * Subfunctions of try_to_unmap: try_to_unmap_one called
968 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
969 */
Hugh Dickins5ad64682009-12-14 17:59:24 -0800970int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
971 unsigned long address, enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700972{
973 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700974 pte_t *pte;
975 pte_t pteval;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700976 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700977 int ret = SWAP_AGAIN;
978
Nick Piggin479db0b2008-08-20 14:09:18 -0700979 pte = page_check_address(page, mm, address, &ptl, 0);
Hugh Dickinsc0718802005-10-29 18:16:31 -0700980 if (!pte)
Nikita Danilov81b40822005-05-01 08:58:36 -0700981 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700982
983 /*
984 * If the page is mlock()d, we cannot swap it out.
985 * If it's recently referenced (perhaps page_referenced
986 * skipped over this mm) then we should reactivate it.
987 */
Andi Kleen14fa31b2009-09-16 11:50:10 +0200988 if (!(flags & TTU_IGNORE_MLOCK)) {
KOSAKI Motohirocaed0f482009-12-14 17:59:45 -0800989 if (vma->vm_flags & VM_LOCKED)
990 goto out_mlock;
991
Hugh Dickinsaf8e3352009-12-14 17:58:59 -0800992 if (TTU_ACTION(flags) == TTU_MUNLOCK)
Hugh Dickins53f79ac2009-12-14 17:58:58 -0800993 goto out_unmap;
Andi Kleen14fa31b2009-09-16 11:50:10 +0200994 }
995 if (!(flags & TTU_IGNORE_ACCESS)) {
Nick Pigginb291f002008-10-18 20:26:44 -0700996 if (ptep_clear_flush_young_notify(vma, address, pte)) {
997 ret = SWAP_FAIL;
998 goto out_unmap;
999 }
1000 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001001
Linus Torvalds1da177e2005-04-16 15:20:36 -07001002 /* Nuke the page table entry. */
1003 flush_cache_page(vma, address, page_to_pfn(page));
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001004 pteval = ptep_clear_flush_notify(vma, address, pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001005
1006 /* Move the dirty bit to the physical page now the pte is gone. */
1007 if (pte_dirty(pteval))
1008 set_page_dirty(page);
1009
Hugh Dickins365e9c872005-10-29 18:16:18 -07001010 /* Update high watermark before we lower rss */
1011 update_hiwater_rss(mm);
1012
Andi Kleen888b9f72009-09-16 11:50:11 +02001013 if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
1014 if (PageAnon(page))
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001015 dec_mm_counter(mm, MM_ANONPAGES);
Andi Kleen888b9f72009-09-16 11:50:11 +02001016 else
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001017 dec_mm_counter(mm, MM_FILEPAGES);
Andi Kleen888b9f72009-09-16 11:50:11 +02001018 set_pte_at(mm, address, pte,
1019 swp_entry_to_pte(make_hwpoison_entry(page)));
1020 } else if (PageAnon(page)) {
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07001021 swp_entry_t entry = { .val = page_private(page) };
Christoph Lameter06972122006-06-23 02:03:35 -07001022
1023 if (PageSwapCache(page)) {
1024 /*
1025 * Store the swap location in the pte.
1026 * See handle_pte_fault() ...
1027 */
Hugh Dickins570a335b2009-12-14 17:58:46 -08001028 if (swap_duplicate(entry) < 0) {
1029 set_pte_at(mm, address, pte, pteval);
1030 ret = SWAP_FAIL;
1031 goto out_unmap;
1032 }
Christoph Lameter06972122006-06-23 02:03:35 -07001033 if (list_empty(&mm->mmlist)) {
1034 spin_lock(&mmlist_lock);
1035 if (list_empty(&mm->mmlist))
1036 list_add(&mm->mmlist, &init_mm.mmlist);
1037 spin_unlock(&mmlist_lock);
1038 }
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001039 dec_mm_counter(mm, MM_ANONPAGES);
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08001040 inc_mm_counter(mm, MM_SWAPENTS);
KOSAKI Motohiro64cdd542009-01-06 14:39:16 -08001041 } else if (PAGE_MIGRATION) {
Christoph Lameter06972122006-06-23 02:03:35 -07001042 /*
1043 * Store the pfn of the page in a special migration
1044 * pte. do_swap_page() will wait until the migration
1045 * pte is removed and then restart fault handling.
1046 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001047 BUG_ON(TTU_ACTION(flags) != TTU_MIGRATION);
Christoph Lameter06972122006-06-23 02:03:35 -07001048 entry = make_migration_entry(page, pte_write(pteval));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001049 }
1050 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
1051 BUG_ON(pte_file(*pte));
Andi Kleen14fa31b2009-09-16 11:50:10 +02001052 } else if (PAGE_MIGRATION && (TTU_ACTION(flags) == TTU_MIGRATION)) {
Christoph Lameter04e62a22006-06-23 02:03:38 -07001053 /* Establish migration entry for a file page */
1054 swp_entry_t entry;
1055 entry = make_migration_entry(page, pte_write(pteval));
1056 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
1057 } else
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001058 dec_mm_counter(mm, MM_FILEPAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001059
Hugh Dickinsedc315f2009-01-06 14:40:11 -08001060 page_remove_rmap(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001061 page_cache_release(page);
1062
1063out_unmap:
Hugh Dickinsc0718802005-10-29 18:16:31 -07001064 pte_unmap_unlock(pte, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001065out:
1066 return ret;
KOSAKI Motohirocaed0f482009-12-14 17:59:45 -08001067
1068out_mlock:
1069 pte_unmap_unlock(pte, ptl);
1070
1071
1072 /*
1073 * We need mmap_sem locking, Otherwise VM_LOCKED check makes
1074 * unstable result and race. Plus, We can't wait here because
1075 * we now hold anon_vma->lock or mapping->i_mmap_lock.
1076 * if trylock failed, the page remain in evictable lru and later
1077 * vmscan could retry to move the page to unevictable lru if the
1078 * page is actually mlocked.
1079 */
1080 if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
1081 if (vma->vm_flags & VM_LOCKED) {
1082 mlock_vma_page(page);
1083 ret = SWAP_MLOCK;
1084 }
1085 up_read(&vma->vm_mm->mmap_sem);
1086 }
1087 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001088}
1089
1090/*
1091 * objrmap doesn't work for nonlinear VMAs because the assumption that
1092 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
1093 * Consequently, given a particular page and its ->index, we cannot locate the
1094 * ptes which are mapping that page without an exhaustive linear search.
1095 *
1096 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
1097 * maps the file to which the target page belongs. The ->vm_private_data field
1098 * holds the current cursor into that scan. Successive searches will circulate
1099 * around the vma's virtual address space.
1100 *
1101 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
1102 * more scanning pressure is placed against them as well. Eventually pages
1103 * will become fully unmapped and are eligible for eviction.
1104 *
1105 * For very sparsely populated VMAs this is a little inefficient - chances are
1106 * there there won't be many ptes located within the scan cluster. In this case
1107 * maybe we could scan further - to the end of the pte page, perhaps.
Nick Pigginb291f002008-10-18 20:26:44 -07001108 *
1109 * Mlocked pages: check VM_LOCKED under mmap_sem held for read, if we can
1110 * acquire it without blocking. If vma locked, mlock the pages in the cluster,
1111 * rather than unmapping them. If we encounter the "check_page" that vmscan is
1112 * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001113 */
1114#define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
1115#define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
1116
Nick Pigginb291f002008-10-18 20:26:44 -07001117static int try_to_unmap_cluster(unsigned long cursor, unsigned int *mapcount,
1118 struct vm_area_struct *vma, struct page *check_page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001119{
1120 struct mm_struct *mm = vma->vm_mm;
1121 pgd_t *pgd;
1122 pud_t *pud;
1123 pmd_t *pmd;
Hugh Dickinsc0718802005-10-29 18:16:31 -07001124 pte_t *pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001125 pte_t pteval;
Hugh Dickinsc0718802005-10-29 18:16:31 -07001126 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001127 struct page *page;
1128 unsigned long address;
1129 unsigned long end;
Nick Pigginb291f002008-10-18 20:26:44 -07001130 int ret = SWAP_AGAIN;
1131 int locked_vma = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001132
Linus Torvalds1da177e2005-04-16 15:20:36 -07001133 address = (vma->vm_start + cursor) & CLUSTER_MASK;
1134 end = address + CLUSTER_SIZE;
1135 if (address < vma->vm_start)
1136 address = vma->vm_start;
1137 if (end > vma->vm_end)
1138 end = vma->vm_end;
1139
1140 pgd = pgd_offset(mm, address);
1141 if (!pgd_present(*pgd))
Nick Pigginb291f002008-10-18 20:26:44 -07001142 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001143
1144 pud = pud_offset(pgd, address);
1145 if (!pud_present(*pud))
Nick Pigginb291f002008-10-18 20:26:44 -07001146 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001147
1148 pmd = pmd_offset(pud, address);
1149 if (!pmd_present(*pmd))
Nick Pigginb291f002008-10-18 20:26:44 -07001150 return ret;
1151
1152 /*
Hugh Dickinsaf8e3352009-12-14 17:58:59 -08001153 * If we can acquire the mmap_sem for read, and vma is VM_LOCKED,
Nick Pigginb291f002008-10-18 20:26:44 -07001154 * keep the sem while scanning the cluster for mlocking pages.
1155 */
Hugh Dickinsaf8e3352009-12-14 17:58:59 -08001156 if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
Nick Pigginb291f002008-10-18 20:26:44 -07001157 locked_vma = (vma->vm_flags & VM_LOCKED);
1158 if (!locked_vma)
1159 up_read(&vma->vm_mm->mmap_sem); /* don't need it */
1160 }
Hugh Dickinsc0718802005-10-29 18:16:31 -07001161
1162 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001163
Hugh Dickins365e9c872005-10-29 18:16:18 -07001164 /* Update high watermark before we lower rss */
1165 update_hiwater_rss(mm);
1166
Hugh Dickinsc0718802005-10-29 18:16:31 -07001167 for (; address < end; pte++, address += PAGE_SIZE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001168 if (!pte_present(*pte))
1169 continue;
Linus Torvalds6aab3412005-11-28 14:34:23 -08001170 page = vm_normal_page(vma, address, *pte);
1171 BUG_ON(!page || PageAnon(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001172
Nick Pigginb291f002008-10-18 20:26:44 -07001173 if (locked_vma) {
1174 mlock_vma_page(page); /* no-op if already mlocked */
1175 if (page == check_page)
1176 ret = SWAP_MLOCK;
1177 continue; /* don't unmap */
1178 }
1179
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001180 if (ptep_clear_flush_young_notify(vma, address, pte))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001181 continue;
1182
1183 /* Nuke the page table entry. */
Ben Collinseca35132005-11-29 11:45:26 -08001184 flush_cache_page(vma, address, pte_pfn(*pte));
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001185 pteval = ptep_clear_flush_notify(vma, address, pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001186
1187 /* If nonlinear, store the file page offset in the pte. */
1188 if (page->index != linear_page_index(vma, address))
1189 set_pte_at(mm, address, pte, pgoff_to_pte(page->index));
1190
1191 /* Move the dirty bit to the physical page now the pte is gone. */
1192 if (pte_dirty(pteval))
1193 set_page_dirty(page);
1194
Hugh Dickinsedc315f2009-01-06 14:40:11 -08001195 page_remove_rmap(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001196 page_cache_release(page);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001197 dec_mm_counter(mm, MM_FILEPAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001198 (*mapcount)--;
1199 }
Hugh Dickinsc0718802005-10-29 18:16:31 -07001200 pte_unmap_unlock(pte - 1, ptl);
Nick Pigginb291f002008-10-18 20:26:44 -07001201 if (locked_vma)
1202 up_read(&vma->vm_mm->mmap_sem);
1203 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001204}
1205
Mel Gormana8bef8ff2010-05-24 14:32:24 -07001206static bool is_vma_temporary_stack(struct vm_area_struct *vma)
1207{
1208 int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP);
1209
1210 if (!maybe_stack)
1211 return false;
1212
1213 if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) ==
1214 VM_STACK_INCOMPLETE_SETUP)
1215 return true;
1216
1217 return false;
1218}
1219
Nick Pigginb291f002008-10-18 20:26:44 -07001220/**
1221 * try_to_unmap_anon - unmap or unlock anonymous page using the object-based
1222 * rmap method
1223 * @page: the page to unmap/unlock
Huang Shijie8051be52009-12-14 17:58:50 -08001224 * @flags: action and flags
Nick Pigginb291f002008-10-18 20:26:44 -07001225 *
1226 * Find all the mappings of a page using the mapping pointer and the vma chains
1227 * contained in the anon_vma struct it points to.
1228 *
1229 * This function is only called from try_to_unmap/try_to_munlock for
1230 * anonymous pages.
1231 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1232 * where the page was found will be held for write. So, we won't recheck
1233 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1234 * 'LOCKED.
1235 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001236static int try_to_unmap_anon(struct page *page, enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001237{
1238 struct anon_vma *anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -08001239 struct anon_vma_chain *avc;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001240 int ret = SWAP_AGAIN;
Nick Pigginb291f002008-10-18 20:26:44 -07001241
Linus Torvalds1da177e2005-04-16 15:20:36 -07001242 anon_vma = page_lock_anon_vma(page);
1243 if (!anon_vma)
1244 return ret;
1245
Rik van Riel5beb4932010-03-05 13:42:07 -08001246 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1247 struct vm_area_struct *vma = avc->vma;
Mel Gormana8bef8ff2010-05-24 14:32:24 -07001248 unsigned long address;
1249
1250 /*
1251 * During exec, a temporary VMA is setup and later moved.
1252 * The VMA is moved under the anon_vma lock but not the
1253 * page tables leading to a race where migration cannot
1254 * find the migration ptes. Rather than increasing the
1255 * locking requirements of exec(), migration skips
1256 * temporary VMAs until after exec() completes.
1257 */
1258 if (PAGE_MIGRATION && (flags & TTU_MIGRATION) &&
1259 is_vma_temporary_stack(vma))
1260 continue;
1261
1262 address = vma_address(page, vma);
Hugh Dickins1cb17292009-12-14 17:59:01 -08001263 if (address == -EFAULT)
1264 continue;
1265 ret = try_to_unmap_one(page, vma, address, flags);
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001266 if (ret != SWAP_AGAIN || !page_mapped(page))
1267 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001268 }
Oleg Nesterov34bbd702007-02-28 20:13:49 -08001269
1270 page_unlock_anon_vma(anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001271 return ret;
1272}
1273
1274/**
Nick Pigginb291f002008-10-18 20:26:44 -07001275 * try_to_unmap_file - unmap/unlock file page using the object-based rmap method
1276 * @page: the page to unmap/unlock
Andi Kleen14fa31b2009-09-16 11:50:10 +02001277 * @flags: action and flags
Linus Torvalds1da177e2005-04-16 15:20:36 -07001278 *
1279 * Find all the mappings of a page using the mapping pointer and the vma chains
1280 * contained in the address_space struct it points to.
1281 *
Nick Pigginb291f002008-10-18 20:26:44 -07001282 * This function is only called from try_to_unmap/try_to_munlock for
1283 * object-based pages.
1284 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1285 * where the page was found will be held for write. So, we won't recheck
1286 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1287 * 'LOCKED.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001288 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001289static int try_to_unmap_file(struct page *page, enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001290{
1291 struct address_space *mapping = page->mapping;
1292 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1293 struct vm_area_struct *vma;
1294 struct prio_tree_iter iter;
1295 int ret = SWAP_AGAIN;
1296 unsigned long cursor;
1297 unsigned long max_nl_cursor = 0;
1298 unsigned long max_nl_size = 0;
1299 unsigned int mapcount;
1300
1301 spin_lock(&mapping->i_mmap_lock);
1302 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
Hugh Dickins1cb17292009-12-14 17:59:01 -08001303 unsigned long address = vma_address(page, vma);
1304 if (address == -EFAULT)
1305 continue;
1306 ret = try_to_unmap_one(page, vma, address, flags);
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001307 if (ret != SWAP_AGAIN || !page_mapped(page))
1308 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001309 }
1310
1311 if (list_empty(&mapping->i_mmap_nonlinear))
1312 goto out;
1313
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001314 /*
1315 * We don't bother to try to find the munlocked page in nonlinears.
1316 * It's costly. Instead, later, page reclaim logic may call
1317 * try_to_unmap(TTU_MUNLOCK) and recover PG_mlocked lazily.
1318 */
1319 if (TTU_ACTION(flags) == TTU_MUNLOCK)
1320 goto out;
1321
Linus Torvalds1da177e2005-04-16 15:20:36 -07001322 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1323 shared.vm_set.list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001324 cursor = (unsigned long) vma->vm_private_data;
1325 if (cursor > max_nl_cursor)
1326 max_nl_cursor = cursor;
1327 cursor = vma->vm_end - vma->vm_start;
1328 if (cursor > max_nl_size)
1329 max_nl_size = cursor;
1330 }
1331
Nick Pigginb291f002008-10-18 20:26:44 -07001332 if (max_nl_size == 0) { /* all nonlinears locked or reserved ? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001333 ret = SWAP_FAIL;
1334 goto out;
1335 }
1336
1337 /*
1338 * We don't try to search for this page in the nonlinear vmas,
1339 * and page_referenced wouldn't have found it anyway. Instead
1340 * just walk the nonlinear vmas trying to age and unmap some.
1341 * The mapcount of the page we came in with is irrelevant,
1342 * but even so use it as a guide to how hard we should try?
1343 */
1344 mapcount = page_mapcount(page);
1345 if (!mapcount)
1346 goto out;
1347 cond_resched_lock(&mapping->i_mmap_lock);
1348
1349 max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
1350 if (max_nl_cursor == 0)
1351 max_nl_cursor = CLUSTER_SIZE;
1352
1353 do {
1354 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1355 shared.vm_set.list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001356 cursor = (unsigned long) vma->vm_private_data;
Hugh Dickins839b9682005-09-03 15:54:43 -07001357 while ( cursor < max_nl_cursor &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07001358 cursor < vma->vm_end - vma->vm_start) {
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001359 if (try_to_unmap_cluster(cursor, &mapcount,
1360 vma, page) == SWAP_MLOCK)
1361 ret = SWAP_MLOCK;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001362 cursor += CLUSTER_SIZE;
1363 vma->vm_private_data = (void *) cursor;
1364 if ((int)mapcount <= 0)
1365 goto out;
1366 }
1367 vma->vm_private_data = (void *) max_nl_cursor;
1368 }
1369 cond_resched_lock(&mapping->i_mmap_lock);
1370 max_nl_cursor += CLUSTER_SIZE;
1371 } while (max_nl_cursor <= max_nl_size);
1372
1373 /*
1374 * Don't loop forever (perhaps all the remaining pages are
1375 * in locked vmas). Reset cursor on all unreserved nonlinear
1376 * vmas, now forgetting on which ones it had fallen behind.
1377 */
Hugh Dickins101d2be2005-11-21 21:32:16 -08001378 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1379 vma->vm_private_data = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001380out:
1381 spin_unlock(&mapping->i_mmap_lock);
1382 return ret;
1383}
1384
1385/**
1386 * try_to_unmap - try to remove all page table mappings to a page
1387 * @page: the page to get unmapped
Andi Kleen14fa31b2009-09-16 11:50:10 +02001388 * @flags: action and flags
Linus Torvalds1da177e2005-04-16 15:20:36 -07001389 *
1390 * Tries to remove all the page table entries which are mapping this
1391 * page, used in the pageout path. Caller must hold the page lock.
1392 * Return values are:
1393 *
1394 * SWAP_SUCCESS - we succeeded in removing all mappings
1395 * SWAP_AGAIN - we missed a mapping, try again later
1396 * SWAP_FAIL - the page is unswappable
Nick Pigginb291f002008-10-18 20:26:44 -07001397 * SWAP_MLOCK - page is mlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001398 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001399int try_to_unmap(struct page *page, enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001400{
1401 int ret;
1402
Linus Torvalds1da177e2005-04-16 15:20:36 -07001403 BUG_ON(!PageLocked(page));
1404
Hugh Dickins5ad64682009-12-14 17:59:24 -08001405 if (unlikely(PageKsm(page)))
1406 ret = try_to_unmap_ksm(page, flags);
1407 else if (PageAnon(page))
Andi Kleen14fa31b2009-09-16 11:50:10 +02001408 ret = try_to_unmap_anon(page, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001409 else
Andi Kleen14fa31b2009-09-16 11:50:10 +02001410 ret = try_to_unmap_file(page, flags);
Nick Pigginb291f002008-10-18 20:26:44 -07001411 if (ret != SWAP_MLOCK && !page_mapped(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001412 ret = SWAP_SUCCESS;
1413 return ret;
1414}
Nikita Danilov81b40822005-05-01 08:58:36 -07001415
Nick Pigginb291f002008-10-18 20:26:44 -07001416/**
1417 * try_to_munlock - try to munlock a page
1418 * @page: the page to be munlocked
1419 *
1420 * Called from munlock code. Checks all of the VMAs mapping the page
1421 * to make sure nobody else has this page mlocked. The page will be
1422 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1423 *
1424 * Return values are:
1425 *
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001426 * SWAP_AGAIN - no vma is holding page mlocked, or,
Nick Pigginb291f002008-10-18 20:26:44 -07001427 * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem
Hugh Dickins5ad64682009-12-14 17:59:24 -08001428 * SWAP_FAIL - page cannot be located at present
Nick Pigginb291f002008-10-18 20:26:44 -07001429 * SWAP_MLOCK - page is now mlocked.
1430 */
1431int try_to_munlock(struct page *page)
1432{
1433 VM_BUG_ON(!PageLocked(page) || PageLRU(page));
1434
Hugh Dickins5ad64682009-12-14 17:59:24 -08001435 if (unlikely(PageKsm(page)))
1436 return try_to_unmap_ksm(page, TTU_MUNLOCK);
1437 else if (PageAnon(page))
Andi Kleen14fa31b2009-09-16 11:50:10 +02001438 return try_to_unmap_anon(page, TTU_MUNLOCK);
Nick Pigginb291f002008-10-18 20:26:44 -07001439 else
Andi Kleen14fa31b2009-09-16 11:50:10 +02001440 return try_to_unmap_file(page, TTU_MUNLOCK);
Nick Pigginb291f002008-10-18 20:26:44 -07001441}
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001442
Rik van Riel76545062010-08-09 17:18:41 -07001443#if defined(CONFIG_KSM) || defined(CONFIG_MIGRATION)
1444/*
1445 * Drop an anon_vma refcount, freeing the anon_vma and anon_vma->root
1446 * if necessary. Be careful to do all the tests under the lock. Once
1447 * we know we are the last user, nobody else can get a reference and we
1448 * can do the freeing without the lock.
1449 */
1450void drop_anon_vma(struct anon_vma *anon_vma)
1451{
Andrea Arcangeli44ab57a2010-08-09 17:19:10 -07001452 BUG_ON(atomic_read(&anon_vma->external_refcount) <= 0);
Rik van Riel76545062010-08-09 17:18:41 -07001453 if (atomic_dec_and_lock(&anon_vma->external_refcount, &anon_vma->root->lock)) {
1454 struct anon_vma *root = anon_vma->root;
1455 int empty = list_empty(&anon_vma->head);
1456 int last_root_user = 0;
1457 int root_empty = 0;
1458
1459 /*
1460 * The refcount on a non-root anon_vma got dropped. Drop
1461 * the refcount on the root and check if we need to free it.
1462 */
1463 if (empty && anon_vma != root) {
Andrea Arcangeli44ab57a2010-08-09 17:19:10 -07001464 BUG_ON(atomic_read(&root->external_refcount) <= 0);
Rik van Riel76545062010-08-09 17:18:41 -07001465 last_root_user = atomic_dec_and_test(&root->external_refcount);
1466 root_empty = list_empty(&root->head);
1467 }
1468 anon_vma_unlock(anon_vma);
1469
1470 if (empty) {
1471 anon_vma_free(anon_vma);
1472 if (root_empty && last_root_user)
1473 anon_vma_free(root);
1474 }
1475 }
1476}
1477#endif
1478
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001479#ifdef CONFIG_MIGRATION
1480/*
1481 * rmap_walk() and its helpers rmap_walk_anon() and rmap_walk_file():
1482 * Called by migrate.c to remove migration ptes, but might be used more later.
1483 */
1484static int rmap_walk_anon(struct page *page, int (*rmap_one)(struct page *,
1485 struct vm_area_struct *, unsigned long, void *), void *arg)
1486{
1487 struct anon_vma *anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -08001488 struct anon_vma_chain *avc;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001489 int ret = SWAP_AGAIN;
1490
1491 /*
1492 * Note: remove_migration_ptes() cannot use page_lock_anon_vma()
1493 * because that depends on page_mapped(); but not all its usages
Mel Gorman3f6c8272010-05-24 14:32:17 -07001494 * are holding mmap_sem. Users without mmap_sem are required to
1495 * take a reference count to prevent the anon_vma disappearing
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001496 */
1497 anon_vma = page_anon_vma(page);
1498 if (!anon_vma)
1499 return ret;
Rik van Rielcba48b92010-08-09 17:18:38 -07001500 anon_vma_lock(anon_vma);
Rik van Riel5beb4932010-03-05 13:42:07 -08001501 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1502 struct vm_area_struct *vma = avc->vma;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001503 unsigned long address = vma_address(page, vma);
1504 if (address == -EFAULT)
1505 continue;
1506 ret = rmap_one(page, vma, address, arg);
1507 if (ret != SWAP_AGAIN)
1508 break;
1509 }
Rik van Rielcba48b92010-08-09 17:18:38 -07001510 anon_vma_unlock(anon_vma);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001511 return ret;
1512}
1513
1514static int rmap_walk_file(struct page *page, int (*rmap_one)(struct page *,
1515 struct vm_area_struct *, unsigned long, void *), void *arg)
1516{
1517 struct address_space *mapping = page->mapping;
1518 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1519 struct vm_area_struct *vma;
1520 struct prio_tree_iter iter;
1521 int ret = SWAP_AGAIN;
1522
1523 if (!mapping)
1524 return ret;
1525 spin_lock(&mapping->i_mmap_lock);
1526 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1527 unsigned long address = vma_address(page, vma);
1528 if (address == -EFAULT)
1529 continue;
1530 ret = rmap_one(page, vma, address, arg);
1531 if (ret != SWAP_AGAIN)
1532 break;
1533 }
1534 /*
1535 * No nonlinear handling: being always shared, nonlinear vmas
1536 * never contain migration ptes. Decide what to do about this
1537 * limitation to linear when we need rmap_walk() on nonlinear.
1538 */
1539 spin_unlock(&mapping->i_mmap_lock);
1540 return ret;
1541}
1542
1543int rmap_walk(struct page *page, int (*rmap_one)(struct page *,
1544 struct vm_area_struct *, unsigned long, void *), void *arg)
1545{
1546 VM_BUG_ON(!PageLocked(page));
1547
1548 if (unlikely(PageKsm(page)))
1549 return rmap_walk_ksm(page, rmap_one, arg);
1550 else if (PageAnon(page))
1551 return rmap_walk_anon(page, rmap_one, arg);
1552 else
1553 return rmap_walk_file(page, rmap_one, arg);
1554}
1555#endif /* CONFIG_MIGRATION */
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001556
Naoya Horiguchie3390f62010-06-15 13:18:13 +09001557#ifdef CONFIG_HUGETLB_PAGE
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001558/*
1559 * The following three functions are for anonymous (private mapped) hugepages.
1560 * Unlike common anonymous pages, anonymous hugepages have no accounting code
1561 * and no lru code, because we handle hugepages differently from common pages.
1562 */
1563static void __hugepage_set_anon_rmap(struct page *page,
1564 struct vm_area_struct *vma, unsigned long address, int exclusive)
1565{
1566 struct anon_vma *anon_vma = vma->anon_vma;
1567 BUG_ON(!anon_vma);
1568 if (!exclusive) {
1569 struct anon_vma_chain *avc;
1570 avc = list_entry(vma->anon_vma_chain.prev,
1571 struct anon_vma_chain, same_vma);
1572 anon_vma = avc->anon_vma;
1573 }
1574 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1575 page->mapping = (struct address_space *) anon_vma;
1576 page->index = linear_page_index(vma, address);
1577}
1578
1579void hugepage_add_anon_rmap(struct page *page,
1580 struct vm_area_struct *vma, unsigned long address)
1581{
1582 struct anon_vma *anon_vma = vma->anon_vma;
1583 int first;
1584 BUG_ON(!anon_vma);
1585 BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1586 first = atomic_inc_and_test(&page->_mapcount);
1587 if (first)
1588 __hugepage_set_anon_rmap(page, vma, address, 0);
1589}
1590
1591void hugepage_add_new_anon_rmap(struct page *page,
1592 struct vm_area_struct *vma, unsigned long address)
1593{
1594 BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1595 atomic_set(&page->_mapcount, 0);
1596 __hugepage_set_anon_rmap(page, vma, address, 1);
1597}
Naoya Horiguchie3390f62010-06-15 13:18:13 +09001598#endif /* CONFIG_HUGETLB_PAGE */