blob: 07fc94758799b778e39ac8796216e84d8c2548e0 [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>
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{
69 return kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
70}
71
Hugh Dickinsdb114b82009-12-14 17:59:25 -080072void anon_vma_free(struct anon_vma *anon_vma)
Adrian Bunkfdd2e5f2008-10-18 20:28:38 -070073{
74 kmem_cache_free(anon_vma_cachep, anon_vma);
75}
Linus Torvalds1da177e2005-04-16 15:20:36 -070076
Rik van Riel5beb4932010-03-05 13:42:07 -080077static inline struct anon_vma_chain *anon_vma_chain_alloc(void)
78{
79 return kmem_cache_alloc(anon_vma_chain_cachep, GFP_KERNEL);
80}
81
82void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
83{
84 kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
85}
86
Linus Torvaldsd9d332e2008-10-19 10:32:20 -070087/**
88 * anon_vma_prepare - attach an anon_vma to a memory region
89 * @vma: the memory region in question
90 *
91 * This makes sure the memory mapping described by 'vma' has
92 * an 'anon_vma' attached to it, so that we can associate the
93 * anonymous pages mapped into it with that anon_vma.
94 *
95 * The common case will be that we already have one, but if
96 * if not we either need to find an adjacent mapping that we
97 * can re-use the anon_vma from (very common when the only
98 * reason for splitting a vma has been mprotect()), or we
99 * allocate a new one.
100 *
101 * Anon-vma allocations are very subtle, because we may have
102 * optimistically looked up an anon_vma in page_lock_anon_vma()
103 * and that may actually touch the spinlock even in the newly
104 * allocated vma (it depends on RCU to make sure that the
105 * anon_vma isn't actually destroyed).
106 *
107 * As a result, we need to do proper anon_vma locking even
108 * for the new allocation. At the same time, we do not want
109 * to do any locking for the common case of already having
110 * an anon_vma.
111 *
112 * This must be called with the mmap_sem held for reading.
113 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700114int anon_vma_prepare(struct vm_area_struct *vma)
115{
116 struct anon_vma *anon_vma = vma->anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -0800117 struct anon_vma_chain *avc;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700118
119 might_sleep();
120 if (unlikely(!anon_vma)) {
121 struct mm_struct *mm = vma->vm_mm;
Linus Torvaldsd9d332e2008-10-19 10:32:20 -0700122 struct anon_vma *allocated;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700123
Rik van Riel5beb4932010-03-05 13:42:07 -0800124 avc = anon_vma_chain_alloc();
125 if (!avc)
126 goto out_enomem;
127
Linus Torvalds1da177e2005-04-16 15:20:36 -0700128 anon_vma = find_mergeable_anon_vma(vma);
Linus Torvaldsd9d332e2008-10-19 10:32:20 -0700129 allocated = NULL;
130 if (!anon_vma) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700131 anon_vma = anon_vma_alloc();
132 if (unlikely(!anon_vma))
Rik van Riel5beb4932010-03-05 13:42:07 -0800133 goto out_enomem_free_avc;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700134 allocated = anon_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700135 }
136
Oleg Nesterov31f2b0e2010-04-23 13:18:01 -0400137 spin_lock(&anon_vma->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700138 /* page_table_lock to protect against threads */
139 spin_lock(&mm->page_table_lock);
140 if (likely(!vma->anon_vma)) {
141 vma->anon_vma = anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -0800142 avc->anon_vma = anon_vma;
143 avc->vma = vma;
144 list_add(&avc->same_vma, &vma->anon_vma_chain);
145 list_add(&avc->same_anon_vma, &anon_vma->head);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700146 allocated = NULL;
Oleg Nesterov31f2b0e2010-04-23 13:18:01 -0400147 avc = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700148 }
149 spin_unlock(&mm->page_table_lock);
Linus Torvaldsd9d332e2008-10-19 10:32:20 -0700150 spin_unlock(&anon_vma->lock);
Oleg Nesterov31f2b0e2010-04-23 13:18:01 -0400151
152 if (unlikely(allocated))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700153 anon_vma_free(allocated);
Oleg Nesterov31f2b0e2010-04-23 13:18:01 -0400154 if (unlikely(avc))
Rik van Riel5beb4932010-03-05 13:42:07 -0800155 anon_vma_chain_free(avc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700156 }
157 return 0;
Rik van Riel5beb4932010-03-05 13:42:07 -0800158
159 out_enomem_free_avc:
160 anon_vma_chain_free(avc);
161 out_enomem:
162 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700163}
164
Rik van Riel5beb4932010-03-05 13:42:07 -0800165static void anon_vma_chain_link(struct vm_area_struct *vma,
166 struct anon_vma_chain *avc,
167 struct anon_vma *anon_vma)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700168{
Rik van Riel5beb4932010-03-05 13:42:07 -0800169 avc->vma = vma;
170 avc->anon_vma = anon_vma;
171 list_add(&avc->same_vma, &vma->anon_vma_chain);
172
173 spin_lock(&anon_vma->lock);
174 list_add_tail(&avc->same_anon_vma, &anon_vma->head);
175 spin_unlock(&anon_vma->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700176}
177
Rik van Riel5beb4932010-03-05 13:42:07 -0800178/*
179 * Attach the anon_vmas from src to dst.
180 * Returns 0 on success, -ENOMEM on failure.
181 */
182int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700183{
Rik van Riel5beb4932010-03-05 13:42:07 -0800184 struct anon_vma_chain *avc, *pavc;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700185
Linus Torvalds646d87b2010-04-11 17:15:03 -0700186 list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
Rik van Riel5beb4932010-03-05 13:42:07 -0800187 avc = anon_vma_chain_alloc();
188 if (!avc)
189 goto enomem_failure;
190 anon_vma_chain_link(dst, avc, pavc->anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700191 }
Rik van Riel5beb4932010-03-05 13:42:07 -0800192 return 0;
193
194 enomem_failure:
195 unlink_anon_vmas(dst);
196 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700197}
198
Rik van Riel5beb4932010-03-05 13:42:07 -0800199/*
200 * Attach vma to its own anon_vma, as well as to the anon_vmas that
201 * the corresponding VMA in the parent process is attached to.
202 * Returns 0 on success, non-zero on failure.
203 */
204int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700205{
Rik van Riel5beb4932010-03-05 13:42:07 -0800206 struct anon_vma_chain *avc;
207 struct anon_vma *anon_vma;
208
209 /* Don't bother if the parent process has no anon_vma here. */
210 if (!pvma->anon_vma)
211 return 0;
212
213 /*
214 * First, attach the new VMA to the parent VMA's anon_vmas,
215 * so rmap can find non-COWed pages in child processes.
216 */
217 if (anon_vma_clone(vma, pvma))
218 return -ENOMEM;
219
220 /* Then add our own anon_vma. */
221 anon_vma = anon_vma_alloc();
222 if (!anon_vma)
223 goto out_error;
224 avc = anon_vma_chain_alloc();
225 if (!avc)
226 goto out_error_free_anon_vma;
227 anon_vma_chain_link(vma, avc, anon_vma);
228 /* Mark this anon_vma as the one where our new (COWed) pages go. */
229 vma->anon_vma = anon_vma;
230
231 return 0;
232
233 out_error_free_anon_vma:
234 anon_vma_free(anon_vma);
235 out_error:
Rik van Riel4946d542010-04-05 12:13:33 -0400236 unlink_anon_vmas(vma);
Rik van Riel5beb4932010-03-05 13:42:07 -0800237 return -ENOMEM;
238}
239
240static void anon_vma_unlink(struct anon_vma_chain *anon_vma_chain)
241{
242 struct anon_vma *anon_vma = anon_vma_chain->anon_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700243 int empty;
244
Rik van Riel5beb4932010-03-05 13:42:07 -0800245 /* If anon_vma_fork fails, we can get an empty anon_vma_chain. */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700246 if (!anon_vma)
247 return;
248
249 spin_lock(&anon_vma->lock);
Rik van Riel5beb4932010-03-05 13:42:07 -0800250 list_del(&anon_vma_chain->same_anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700251
252 /* We must garbage collect the anon_vma if it's empty */
Hugh Dickinsdb114b82009-12-14 17:59:25 -0800253 empty = list_empty(&anon_vma->head) && !ksm_refcount(anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700254 spin_unlock(&anon_vma->lock);
255
256 if (empty)
257 anon_vma_free(anon_vma);
258}
259
Rik van Riel5beb4932010-03-05 13:42:07 -0800260void unlink_anon_vmas(struct vm_area_struct *vma)
261{
262 struct anon_vma_chain *avc, *next;
263
264 /* Unlink each anon_vma chained to the VMA. */
265 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
266 anon_vma_unlink(avc);
267 list_del(&avc->same_vma);
268 anon_vma_chain_free(avc);
269 }
270}
271
Alexey Dobriyan51cc5062008-07-25 19:45:34 -0700272static void anon_vma_ctor(void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700273{
Christoph Lametera35afb82007-05-16 22:10:57 -0700274 struct anon_vma *anon_vma = data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700275
Christoph Lametera35afb82007-05-16 22:10:57 -0700276 spin_lock_init(&anon_vma->lock);
Hugh Dickinsdb114b82009-12-14 17:59:25 -0800277 ksm_refcount_init(anon_vma);
Christoph Lametera35afb82007-05-16 22:10:57 -0700278 INIT_LIST_HEAD(&anon_vma->head);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700279}
280
281void __init anon_vma_init(void)
282{
283 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
Paul Mundt20c2df82007-07-20 10:11:58 +0900284 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor);
Rik van Riel5beb4932010-03-05 13:42:07 -0800285 anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain, SLAB_PANIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700286}
287
288/*
289 * Getting a lock on a stable anon_vma from a page off the LRU is
290 * tricky: page_lock_anon_vma rely on RCU to guard against the races.
291 */
Andi Kleen10be22d2009-09-16 11:50:04 +0200292struct anon_vma *page_lock_anon_vma(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700293{
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800294 struct anon_vma *anon_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700295 unsigned long anon_mapping;
296
297 rcu_read_lock();
Hugh Dickins80e14822009-12-14 17:59:29 -0800298 anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping);
Hugh Dickins3ca7b3c2009-12-14 17:58:57 -0800299 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700300 goto out;
301 if (!page_mapped(page))
302 goto out;
303
304 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
305 spin_lock(&anon_vma->lock);
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800306 return anon_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700307out:
308 rcu_read_unlock();
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800309 return NULL;
310}
311
Andi Kleen10be22d2009-09-16 11:50:04 +0200312void page_unlock_anon_vma(struct anon_vma *anon_vma)
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800313{
314 spin_unlock(&anon_vma->lock);
315 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700316}
317
318/*
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -0800319 * At what user virtual address is page expected in @vma?
320 * Returns virtual address or -EFAULT if page's index/offset is not
321 * within the range mapped the @vma.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700322 */
323static inline unsigned long
324vma_address(struct page *page, struct vm_area_struct *vma)
325{
326 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
327 unsigned long address;
328
329 address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
330 if (unlikely(address < vma->vm_start || address >= vma->vm_end)) {
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -0800331 /* page should be within @vma mapping range */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700332 return -EFAULT;
333 }
334 return address;
335}
336
337/*
Huang Shijiebf89c8c2009-10-01 15:44:04 -0700338 * At what user virtual address is page expected in vma?
339 * checking that the page matches the vma.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700340 */
341unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
342{
343 if (PageAnon(page)) {
Hugh Dickins3ca7b3c2009-12-14 17:58:57 -0800344 if (vma->anon_vma != page_anon_vma(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700345 return -EFAULT;
346 } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) {
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800347 if (!vma->vm_file ||
348 vma->vm_file->f_mapping != page->mapping)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700349 return -EFAULT;
350 } else
351 return -EFAULT;
352 return vma_address(page, vma);
353}
354
355/*
Nikita Danilov81b40822005-05-01 08:58:36 -0700356 * Check that @page is mapped at @address into @mm.
357 *
Nick Piggin479db0b2008-08-20 14:09:18 -0700358 * If @sync is false, page_check_address may perform a racy check to avoid
359 * the page table lock when the pte is not present (helpful when reclaiming
360 * highly shared pages).
361 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -0700362 * On success returns with pte mapped and locked.
Nikita Danilov81b40822005-05-01 08:58:36 -0700363 */
Carsten Otteceffc072005-06-23 22:05:25 -0700364pte_t *page_check_address(struct page *page, struct mm_struct *mm,
Nick Piggin479db0b2008-08-20 14:09:18 -0700365 unsigned long address, spinlock_t **ptlp, int sync)
Nikita Danilov81b40822005-05-01 08:58:36 -0700366{
367 pgd_t *pgd;
368 pud_t *pud;
369 pmd_t *pmd;
370 pte_t *pte;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700371 spinlock_t *ptl;
Nikita Danilov81b40822005-05-01 08:58:36 -0700372
Nikita Danilov81b40822005-05-01 08:58:36 -0700373 pgd = pgd_offset(mm, address);
Hugh Dickinsc0718802005-10-29 18:16:31 -0700374 if (!pgd_present(*pgd))
375 return NULL;
376
377 pud = pud_offset(pgd, address);
378 if (!pud_present(*pud))
379 return NULL;
380
381 pmd = pmd_offset(pud, address);
382 if (!pmd_present(*pmd))
383 return NULL;
384
385 pte = pte_offset_map(pmd, address);
386 /* Make a quick check before getting the lock */
Nick Piggin479db0b2008-08-20 14:09:18 -0700387 if (!sync && !pte_present(*pte)) {
Hugh Dickinsc0718802005-10-29 18:16:31 -0700388 pte_unmap(pte);
389 return NULL;
Nikita Danilov81b40822005-05-01 08:58:36 -0700390 }
Hugh Dickinsc0718802005-10-29 18:16:31 -0700391
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700392 ptl = pte_lockptr(mm, pmd);
Hugh Dickinsc0718802005-10-29 18:16:31 -0700393 spin_lock(ptl);
394 if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
395 *ptlp = ptl;
396 return pte;
397 }
398 pte_unmap_unlock(pte, ptl);
399 return NULL;
Nikita Danilov81b40822005-05-01 08:58:36 -0700400}
401
Nick Pigginb291f002008-10-18 20:26:44 -0700402/**
403 * page_mapped_in_vma - check whether a page is really mapped in a VMA
404 * @page: the page to test
405 * @vma: the VMA to test
406 *
407 * Returns 1 if the page is mapped into the page tables of the VMA, 0
408 * if the page is not mapped into the page tables of this VMA. Only
409 * valid for normal file or anonymous VMAs.
410 */
Andi Kleen6a460792009-09-16 11:50:15 +0200411int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
Nick Pigginb291f002008-10-18 20:26:44 -0700412{
413 unsigned long address;
414 pte_t *pte;
415 spinlock_t *ptl;
416
417 address = vma_address(page, vma);
418 if (address == -EFAULT) /* out of vma range */
419 return 0;
420 pte = page_check_address(page, vma->vm_mm, address, &ptl, 1);
421 if (!pte) /* the page is not in this mm */
422 return 0;
423 pte_unmap_unlock(pte, ptl);
424
425 return 1;
426}
427
Nikita Danilov81b40822005-05-01 08:58:36 -0700428/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700429 * Subfunctions of page_referenced: page_referenced_one called
430 * repeatedly from either page_referenced_anon or page_referenced_file.
431 */
Hugh Dickins5ad64682009-12-14 17:59:24 -0800432int page_referenced_one(struct page *page, struct vm_area_struct *vma,
433 unsigned long address, unsigned int *mapcount,
434 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700435{
436 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700437 pte_t *pte;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700438 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700439 int referenced = 0;
440
Nick Piggin479db0b2008-08-20 14:09:18 -0700441 pte = page_check_address(page, mm, address, &ptl, 0);
Hugh Dickinsc0718802005-10-29 18:16:31 -0700442 if (!pte)
443 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700444
Nick Pigginb291f002008-10-18 20:26:44 -0700445 /*
446 * Don't want to elevate referenced for mlocked page that gets this far,
447 * in order that it progresses to try_to_unmap and is moved to the
448 * unevictable list.
449 */
Hugh Dickins5a9bbdc2008-02-04 22:29:23 -0800450 if (vma->vm_flags & VM_LOCKED) {
Hugh Dickins5a9bbdc2008-02-04 22:29:23 -0800451 *mapcount = 1; /* break early from loop */
Minchan Kim03ef83a2009-08-26 14:29:23 -0700452 *vm_flags |= VM_LOCKED;
Nick Pigginb291f002008-10-18 20:26:44 -0700453 goto out_unmap;
454 }
455
Johannes Weiner4917e5d2009-01-06 14:39:17 -0800456 if (ptep_clear_flush_young_notify(vma, address, pte)) {
457 /*
458 * Don't treat a reference through a sequentially read
459 * mapping as such. If the page has been used in
460 * another mapping, we will catch it; if this other
461 * mapping is already gone, the unmap path will have
462 * set PG_referenced or activated the page.
463 */
464 if (likely(!VM_SequentialReadHint(vma)))
465 referenced++;
466 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700467
Hugh Dickinsc0718802005-10-29 18:16:31 -0700468 /* Pretend the page is referenced if the task has the
469 swap token and is in the middle of a page fault. */
Rik van Rielf7b7fd82005-11-28 13:44:07 -0800470 if (mm != current->mm && has_swap_token(mm) &&
Hugh Dickinsc0718802005-10-29 18:16:31 -0700471 rwsem_is_locked(&mm->mmap_sem))
472 referenced++;
473
Nick Pigginb291f002008-10-18 20:26:44 -0700474out_unmap:
Hugh Dickinsc0718802005-10-29 18:16:31 -0700475 (*mapcount)--;
476 pte_unmap_unlock(pte, ptl);
Huang Shijie273f0472009-12-14 17:58:51 -0800477
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700478 if (referenced)
479 *vm_flags |= vma->vm_flags;
Huang Shijie273f0472009-12-14 17:58:51 -0800480out:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700481 return referenced;
482}
483
Balbir Singhbed71612008-02-07 00:14:01 -0800484static int page_referenced_anon(struct page *page,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700485 struct mem_cgroup *mem_cont,
486 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700487{
488 unsigned int mapcount;
489 struct anon_vma *anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -0800490 struct anon_vma_chain *avc;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700491 int referenced = 0;
492
493 anon_vma = page_lock_anon_vma(page);
494 if (!anon_vma)
495 return referenced;
496
497 mapcount = page_mapcount(page);
Rik van Riel5beb4932010-03-05 13:42:07 -0800498 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
499 struct vm_area_struct *vma = avc->vma;
Hugh Dickins1cb17292009-12-14 17:59:01 -0800500 unsigned long address = vma_address(page, vma);
501 if (address == -EFAULT)
502 continue;
Balbir Singhbed71612008-02-07 00:14:01 -0800503 /*
504 * If we are reclaiming on behalf of a cgroup, skip
505 * counting on behalf of references from different
506 * cgroups
507 */
Hugh Dickinsbd845e32008-03-04 14:29:01 -0800508 if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
Balbir Singhbed71612008-02-07 00:14:01 -0800509 continue;
Hugh Dickins1cb17292009-12-14 17:59:01 -0800510 referenced += page_referenced_one(page, vma, address,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700511 &mapcount, vm_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700512 if (!mapcount)
513 break;
514 }
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800515
516 page_unlock_anon_vma(anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700517 return referenced;
518}
519
520/**
521 * page_referenced_file - referenced check for object-based rmap
522 * @page: the page we're checking references on.
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700523 * @mem_cont: target memory controller
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700524 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
Linus Torvalds1da177e2005-04-16 15:20:36 -0700525 *
526 * For an object-based mapped page, find all the places it is mapped and
527 * check/clear the referenced flag. This is done by following the page->mapping
528 * pointer, then walking the chain of vmas it holds. It returns the number
529 * of references it found.
530 *
531 * This function is only called from page_referenced for object-based pages.
532 */
Balbir Singhbed71612008-02-07 00:14:01 -0800533static int page_referenced_file(struct page *page,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700534 struct mem_cgroup *mem_cont,
535 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700536{
537 unsigned int mapcount;
538 struct address_space *mapping = page->mapping;
539 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
540 struct vm_area_struct *vma;
541 struct prio_tree_iter iter;
542 int referenced = 0;
543
544 /*
545 * The caller's checks on page->mapping and !PageAnon have made
546 * sure that this is a file page: the check for page->mapping
547 * excludes the case just before it gets set on an anon page.
548 */
549 BUG_ON(PageAnon(page));
550
551 /*
552 * The page lock not only makes sure that page->mapping cannot
553 * suddenly be NULLified by truncation, it makes sure that the
554 * structure at mapping cannot be freed and reused yet,
555 * so we can safely take mapping->i_mmap_lock.
556 */
557 BUG_ON(!PageLocked(page));
558
559 spin_lock(&mapping->i_mmap_lock);
560
561 /*
562 * i_mmap_lock does not stabilize mapcount at all, but mapcount
563 * is more likely to be accurate if we note it after spinning.
564 */
565 mapcount = page_mapcount(page);
566
567 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
Hugh Dickins1cb17292009-12-14 17:59:01 -0800568 unsigned long address = vma_address(page, vma);
569 if (address == -EFAULT)
570 continue;
Balbir Singhbed71612008-02-07 00:14:01 -0800571 /*
572 * If we are reclaiming on behalf of a cgroup, skip
573 * counting on behalf of references from different
574 * cgroups
575 */
Hugh Dickinsbd845e32008-03-04 14:29:01 -0800576 if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
Balbir Singhbed71612008-02-07 00:14:01 -0800577 continue;
Hugh Dickins1cb17292009-12-14 17:59:01 -0800578 referenced += page_referenced_one(page, vma, address,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700579 &mapcount, vm_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700580 if (!mapcount)
581 break;
582 }
583
584 spin_unlock(&mapping->i_mmap_lock);
585 return referenced;
586}
587
588/**
589 * page_referenced - test if the page was referenced
590 * @page: the page to test
591 * @is_locked: caller holds lock on the page
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700592 * @mem_cont: target memory controller
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700593 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
Linus Torvalds1da177e2005-04-16 15:20:36 -0700594 *
595 * Quick test_and_clear_referenced for all mappings to a page,
596 * returns the number of ptes which referenced the page.
597 */
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700598int page_referenced(struct page *page,
599 int is_locked,
600 struct mem_cgroup *mem_cont,
601 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700602{
603 int referenced = 0;
Hugh Dickins5ad64682009-12-14 17:59:24 -0800604 int we_locked = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700605
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700606 *vm_flags = 0;
Hugh Dickins3ca7b3c2009-12-14 17:58:57 -0800607 if (page_mapped(page) && page_rmapping(page)) {
Hugh Dickins5ad64682009-12-14 17:59:24 -0800608 if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
609 we_locked = trylock_page(page);
610 if (!we_locked) {
611 referenced++;
612 goto out;
613 }
614 }
615 if (unlikely(PageKsm(page)))
616 referenced += page_referenced_ksm(page, mem_cont,
617 vm_flags);
618 else if (PageAnon(page))
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700619 referenced += page_referenced_anon(page, mem_cont,
620 vm_flags);
Hugh Dickins5ad64682009-12-14 17:59:24 -0800621 else if (page->mapping)
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700622 referenced += page_referenced_file(page, mem_cont,
623 vm_flags);
Hugh Dickins5ad64682009-12-14 17:59:24 -0800624 if (we_locked)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700625 unlock_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700626 }
Hugh Dickins5ad64682009-12-14 17:59:24 -0800627out:
Christian Borntraeger5b7baf02008-03-25 18:47:12 +0100628 if (page_test_and_clear_young(page))
629 referenced++;
630
Linus Torvalds1da177e2005-04-16 15:20:36 -0700631 return referenced;
632}
633
Hugh Dickins1cb17292009-12-14 17:59:01 -0800634static int page_mkclean_one(struct page *page, struct vm_area_struct *vma,
635 unsigned long address)
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700636{
637 struct mm_struct *mm = vma->vm_mm;
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100638 pte_t *pte;
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700639 spinlock_t *ptl;
640 int ret = 0;
641
Nick Piggin479db0b2008-08-20 14:09:18 -0700642 pte = page_check_address(page, mm, address, &ptl, 1);
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700643 if (!pte)
644 goto out;
645
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100646 if (pte_dirty(*pte) || pte_write(*pte)) {
647 pte_t entry;
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700648
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100649 flush_cache_page(vma, address, pte_pfn(*pte));
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700650 entry = ptep_clear_flush_notify(vma, address, pte);
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100651 entry = pte_wrprotect(entry);
652 entry = pte_mkclean(entry);
Al Virod6e88e62006-12-29 16:48:35 -0800653 set_pte_at(mm, address, pte, entry);
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100654 ret = 1;
655 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700656
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700657 pte_unmap_unlock(pte, ptl);
658out:
659 return ret;
660}
661
662static int page_mkclean_file(struct address_space *mapping, struct page *page)
663{
664 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
665 struct vm_area_struct *vma;
666 struct prio_tree_iter iter;
667 int ret = 0;
668
669 BUG_ON(PageAnon(page));
670
671 spin_lock(&mapping->i_mmap_lock);
672 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
Hugh Dickins1cb17292009-12-14 17:59:01 -0800673 if (vma->vm_flags & VM_SHARED) {
674 unsigned long address = vma_address(page, vma);
675 if (address == -EFAULT)
676 continue;
677 ret += page_mkclean_one(page, vma, address);
678 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700679 }
680 spin_unlock(&mapping->i_mmap_lock);
681 return ret;
682}
683
684int page_mkclean(struct page *page)
685{
686 int ret = 0;
687
688 BUG_ON(!PageLocked(page));
689
690 if (page_mapped(page)) {
691 struct address_space *mapping = page_mapping(page);
Christian Borntraegerce7e9fa2007-11-20 11:13:36 +0100692 if (mapping) {
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700693 ret = page_mkclean_file(mapping, page);
Christian Borntraegerce7e9fa2007-11-20 11:13:36 +0100694 if (page_test_dirty(page)) {
695 page_clear_dirty(page);
696 ret = 1;
697 }
Martin Schwidefsky6c210482007-04-27 16:01:57 +0200698 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700699 }
700
701 return ret;
702}
Jaya Kumar60b59be2007-05-08 00:37:37 -0700703EXPORT_SYMBOL_GPL(page_mkclean);
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700704
Linus Torvalds1da177e2005-04-16 15:20:36 -0700705/**
Rik van Rielc44b6742010-03-05 13:42:09 -0800706 * page_move_anon_rmap - move a page to our anon_vma
707 * @page: the page to move to our anon_vma
708 * @vma: the vma the page belongs to
709 * @address: the user virtual address mapped
710 *
711 * When a page belongs exclusively to one process after a COW event,
712 * that page can be moved into the anon_vma that belongs to just that
713 * process, so the rmap code will not search the parent or sibling
714 * processes.
715 */
716void page_move_anon_rmap(struct page *page,
717 struct vm_area_struct *vma, unsigned long address)
718{
719 struct anon_vma *anon_vma = vma->anon_vma;
720
721 VM_BUG_ON(!PageLocked(page));
722 VM_BUG_ON(!anon_vma);
723 VM_BUG_ON(page->index != linear_page_index(vma, address));
724
725 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
726 page->mapping = (struct address_space *) anon_vma;
727}
728
729/**
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700730 * __page_set_anon_rmap - setup new anonymous rmap
Nick Piggin9617d952006-01-06 00:11:12 -0800731 * @page: the page to add the mapping to
732 * @vma: the vm area in which the mapping is added
733 * @address: the user virtual address mapped
Rik van Riele8a03fe2010-04-14 17:59:28 -0400734 * @exclusive: the page is exclusively owned by the current process
Nick Piggin9617d952006-01-06 00:11:12 -0800735 */
736static void __page_set_anon_rmap(struct page *page,
Rik van Riele8a03fe2010-04-14 17:59:28 -0400737 struct vm_area_struct *vma, unsigned long address, int exclusive)
Nick Piggin9617d952006-01-06 00:11:12 -0800738{
Rik van Riele8a03fe2010-04-14 17:59:28 -0400739 struct anon_vma *anon_vma = vma->anon_vma;
Nick Piggin9617d952006-01-06 00:11:12 -0800740
Rik van Riele8a03fe2010-04-14 17:59:28 -0400741 BUG_ON(!anon_vma);
Linus Torvaldsea900022010-04-12 12:44:29 -0700742
743 /*
Rik van Riele8a03fe2010-04-14 17:59:28 -0400744 * If the page isn't exclusively mapped into this vma,
745 * we must use the _oldest_ possible anon_vma for the
746 * page mapping!
Linus Torvaldsea900022010-04-12 12:44:29 -0700747 *
Rik van Riele8a03fe2010-04-14 17:59:28 -0400748 * So take the last AVC chain entry in the vma, which is
749 * the deepest ancestor, and use the anon_vma from that.
Linus Torvaldsea900022010-04-12 12:44:29 -0700750 */
Rik van Riele8a03fe2010-04-14 17:59:28 -0400751 if (!exclusive) {
752 struct anon_vma_chain *avc;
753 avc = list_entry(vma->anon_vma_chain.prev, struct anon_vma_chain, same_vma);
754 anon_vma = avc->anon_vma;
755 }
Linus Torvaldsea900022010-04-12 12:44:29 -0700756
Nick Piggin9617d952006-01-06 00:11:12 -0800757 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
758 page->mapping = (struct address_space *) anon_vma;
Nick Piggin9617d952006-01-06 00:11:12 -0800759 page->index = linear_page_index(vma, address);
Nick Piggin9617d952006-01-06 00:11:12 -0800760}
761
762/**
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700763 * __page_check_anon_rmap - sanity check anonymous rmap addition
Nick Pigginc97a9e12007-05-16 22:11:21 -0700764 * @page: the page to add the mapping to
765 * @vma: the vm area in which the mapping is added
766 * @address: the user virtual address mapped
767 */
768static void __page_check_anon_rmap(struct page *page,
769 struct vm_area_struct *vma, unsigned long address)
770{
771#ifdef CONFIG_DEBUG_VM
772 /*
773 * The page's anon-rmap details (mapping and index) are guaranteed to
774 * be set up correctly at this point.
775 *
776 * We have exclusion against page_add_anon_rmap because the caller
777 * always holds the page locked, except if called from page_dup_rmap,
778 * in which case the page is already known to be setup.
779 *
780 * We have exclusion against page_add_new_anon_rmap because those pages
781 * are initially only visible via the pagetables, and the pte is locked
782 * over the call to page_add_new_anon_rmap.
783 */
Nick Pigginc97a9e12007-05-16 22:11:21 -0700784 BUG_ON(page->index != linear_page_index(vma, address));
785#endif
786}
787
788/**
Linus Torvalds1da177e2005-04-16 15:20:36 -0700789 * page_add_anon_rmap - add pte mapping to an anonymous page
790 * @page: the page to add the mapping to
791 * @vma: the vm area in which the mapping is added
792 * @address: the user virtual address mapped
793 *
Hugh Dickins5ad64682009-12-14 17:59:24 -0800794 * The caller needs to hold the pte lock, and the page must be locked in
Hugh Dickins80e14822009-12-14 17:59:29 -0800795 * the anon_vma case: to serialize mapping,index checking after setting,
796 * and to ensure that PageAnon is not being upgraded racily to PageKsm
797 * (but PageKsm is never downgraded to PageAnon).
Linus Torvalds1da177e2005-04-16 15:20:36 -0700798 */
799void page_add_anon_rmap(struct page *page,
800 struct vm_area_struct *vma, unsigned long address)
801{
Hugh Dickins5ad64682009-12-14 17:59:24 -0800802 int first = atomic_inc_and_test(&page->_mapcount);
803 if (first)
804 __inc_zone_page_state(page, NR_ANON_PAGES);
805 if (unlikely(PageKsm(page)))
806 return;
807
Nick Pigginc97a9e12007-05-16 22:11:21 -0700808 VM_BUG_ON(!PageLocked(page));
809 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
Hugh Dickins5ad64682009-12-14 17:59:24 -0800810 if (first)
Rik van Riele8a03fe2010-04-14 17:59:28 -0400811 __page_set_anon_rmap(page, vma, address, 0);
KAMEZAWA Hiroyuki69029cd2008-07-25 01:47:14 -0700812 else
Nick Pigginc97a9e12007-05-16 22:11:21 -0700813 __page_check_anon_rmap(page, vma, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700814}
815
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700816/**
Nick Piggin9617d952006-01-06 00:11:12 -0800817 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
818 * @page: the page to add the mapping to
819 * @vma: the vm area in which the mapping is added
820 * @address: the user virtual address mapped
821 *
822 * Same as page_add_anon_rmap but must only be called on *new* pages.
823 * This means the inc-and-test can be bypassed.
Nick Pigginc97a9e12007-05-16 22:11:21 -0700824 * Page does not have to be locked.
Nick Piggin9617d952006-01-06 00:11:12 -0800825 */
826void page_add_new_anon_rmap(struct page *page,
827 struct vm_area_struct *vma, unsigned long address)
828{
Hugh Dickinsb5934c52009-01-06 14:39:25 -0800829 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
Hugh Dickinscbf84b72009-01-06 14:39:27 -0800830 SetPageSwapBacked(page);
831 atomic_set(&page->_mapcount, 0); /* increment count (starts at -1) */
Hugh Dickins5ad64682009-12-14 17:59:24 -0800832 __inc_zone_page_state(page, NR_ANON_PAGES);
Rik van Riele8a03fe2010-04-14 17:59:28 -0400833 __page_set_anon_rmap(page, vma, address, 1);
Hugh Dickinsb5934c52009-01-06 14:39:25 -0800834 if (page_evictable(page, vma))
Hugh Dickinscbf84b72009-01-06 14:39:27 -0800835 lru_cache_add_lru(page, LRU_ACTIVE_ANON);
Hugh Dickinsb5934c52009-01-06 14:39:25 -0800836 else
837 add_page_to_unevictable_list(page);
Nick Piggin9617d952006-01-06 00:11:12 -0800838}
839
Linus Torvalds1da177e2005-04-16 15:20:36 -0700840/**
841 * page_add_file_rmap - add pte mapping to a file page
842 * @page: the page to add the mapping to
843 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -0700844 * The caller needs to hold the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700845 */
846void page_add_file_rmap(struct page *page)
847{
Balbir Singhd69b0422009-06-17 16:26:34 -0700848 if (atomic_inc_and_test(&page->_mapcount)) {
Christoph Lameter65ba55f2006-06-30 01:55:34 -0700849 __inc_zone_page_state(page, NR_FILE_MAPPED);
KAMEZAWA Hiroyukid8046582009-12-15 16:47:09 -0800850 mem_cgroup_update_file_mapped(page, 1);
Balbir Singhd69b0422009-06-17 16:26:34 -0700851 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700852}
853
854/**
855 * page_remove_rmap - take down pte mapping from a page
856 * @page: page to remove mapping from
857 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -0700858 * The caller needs to hold the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700859 */
Hugh Dickinsedc315f2009-01-06 14:40:11 -0800860void page_remove_rmap(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700861{
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -0700862 /* page still mapped by someone else? */
863 if (!atomic_add_negative(-1, &page->_mapcount))
864 return;
865
866 /*
867 * Now that the last pte has gone, s390 must transfer dirty
868 * flag from storage key to struct page. We can usually skip
869 * this if the page is anon, so about to be freed; but perhaps
870 * not if it's in swapcache - there might be another pte slot
871 * containing the swap entry, but page not yet written to swap.
872 */
873 if ((!PageAnon(page) || PageSwapCache(page)) && page_test_dirty(page)) {
874 page_clear_dirty(page);
875 set_page_dirty(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700876 }
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -0700877 if (PageAnon(page)) {
878 mem_cgroup_uncharge_page(page);
879 __dec_zone_page_state(page, NR_ANON_PAGES);
880 } else {
881 __dec_zone_page_state(page, NR_FILE_MAPPED);
KAMEZAWA Hiroyukid8046582009-12-15 16:47:09 -0800882 mem_cgroup_update_file_mapped(page, -1);
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -0700883 }
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -0700884 /*
885 * It would be tidy to reset the PageAnon mapping here,
886 * but that might overwrite a racing page_add_anon_rmap
887 * which increments mapcount after us but sets mapping
888 * before us: so leave the reset to free_hot_cold_page,
889 * and remember that it's only reliable while mapped.
890 * Leaving it set also helps swapoff to reinstate ptes
891 * faster for those pages still in swapcache.
892 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700893}
894
895/*
896 * Subfunctions of try_to_unmap: try_to_unmap_one called
897 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
898 */
Hugh Dickins5ad64682009-12-14 17:59:24 -0800899int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
900 unsigned long address, enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700901{
902 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700903 pte_t *pte;
904 pte_t pteval;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700905 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700906 int ret = SWAP_AGAIN;
907
Nick Piggin479db0b2008-08-20 14:09:18 -0700908 pte = page_check_address(page, mm, address, &ptl, 0);
Hugh Dickinsc0718802005-10-29 18:16:31 -0700909 if (!pte)
Nikita Danilov81b40822005-05-01 08:58:36 -0700910 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700911
912 /*
913 * If the page is mlock()d, we cannot swap it out.
914 * If it's recently referenced (perhaps page_referenced
915 * skipped over this mm) then we should reactivate it.
916 */
Andi Kleen14fa31b2009-09-16 11:50:10 +0200917 if (!(flags & TTU_IGNORE_MLOCK)) {
KOSAKI Motohirocaed0f42009-12-14 17:59:45 -0800918 if (vma->vm_flags & VM_LOCKED)
919 goto out_mlock;
920
Hugh Dickinsaf8e3352009-12-14 17:58:59 -0800921 if (TTU_ACTION(flags) == TTU_MUNLOCK)
Hugh Dickins53f79ac2009-12-14 17:58:58 -0800922 goto out_unmap;
Andi Kleen14fa31b2009-09-16 11:50:10 +0200923 }
924 if (!(flags & TTU_IGNORE_ACCESS)) {
Nick Pigginb291f002008-10-18 20:26:44 -0700925 if (ptep_clear_flush_young_notify(vma, address, pte)) {
926 ret = SWAP_FAIL;
927 goto out_unmap;
928 }
929 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700930
Linus Torvalds1da177e2005-04-16 15:20:36 -0700931 /* Nuke the page table entry. */
932 flush_cache_page(vma, address, page_to_pfn(page));
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700933 pteval = ptep_clear_flush_notify(vma, address, pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700934
935 /* Move the dirty bit to the physical page now the pte is gone. */
936 if (pte_dirty(pteval))
937 set_page_dirty(page);
938
Hugh Dickins365e9c872005-10-29 18:16:18 -0700939 /* Update high watermark before we lower rss */
940 update_hiwater_rss(mm);
941
Andi Kleen888b9f72009-09-16 11:50:11 +0200942 if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
943 if (PageAnon(page))
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800944 dec_mm_counter(mm, MM_ANONPAGES);
Andi Kleen888b9f72009-09-16 11:50:11 +0200945 else
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800946 dec_mm_counter(mm, MM_FILEPAGES);
Andi Kleen888b9f72009-09-16 11:50:11 +0200947 set_pte_at(mm, address, pte,
948 swp_entry_to_pte(make_hwpoison_entry(page)));
949 } else if (PageAnon(page)) {
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700950 swp_entry_t entry = { .val = page_private(page) };
Christoph Lameter06972122006-06-23 02:03:35 -0700951
952 if (PageSwapCache(page)) {
953 /*
954 * Store the swap location in the pte.
955 * See handle_pte_fault() ...
956 */
Hugh Dickins570a3352009-12-14 17:58:46 -0800957 if (swap_duplicate(entry) < 0) {
958 set_pte_at(mm, address, pte, pteval);
959 ret = SWAP_FAIL;
960 goto out_unmap;
961 }
Christoph Lameter06972122006-06-23 02:03:35 -0700962 if (list_empty(&mm->mmlist)) {
963 spin_lock(&mmlist_lock);
964 if (list_empty(&mm->mmlist))
965 list_add(&mm->mmlist, &init_mm.mmlist);
966 spin_unlock(&mmlist_lock);
967 }
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800968 dec_mm_counter(mm, MM_ANONPAGES);
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -0800969 inc_mm_counter(mm, MM_SWAPENTS);
KOSAKI Motohiro64cdd542009-01-06 14:39:16 -0800970 } else if (PAGE_MIGRATION) {
Christoph Lameter06972122006-06-23 02:03:35 -0700971 /*
972 * Store the pfn of the page in a special migration
973 * pte. do_swap_page() will wait until the migration
974 * pte is removed and then restart fault handling.
975 */
Andi Kleen14fa31b2009-09-16 11:50:10 +0200976 BUG_ON(TTU_ACTION(flags) != TTU_MIGRATION);
Christoph Lameter06972122006-06-23 02:03:35 -0700977 entry = make_migration_entry(page, pte_write(pteval));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700978 }
979 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
980 BUG_ON(pte_file(*pte));
Andi Kleen14fa31b2009-09-16 11:50:10 +0200981 } else if (PAGE_MIGRATION && (TTU_ACTION(flags) == TTU_MIGRATION)) {
Christoph Lameter04e62a22006-06-23 02:03:38 -0700982 /* Establish migration entry for a file page */
983 swp_entry_t entry;
984 entry = make_migration_entry(page, pte_write(pteval));
985 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
986 } else
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800987 dec_mm_counter(mm, MM_FILEPAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700988
Hugh Dickinsedc315f2009-01-06 14:40:11 -0800989 page_remove_rmap(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700990 page_cache_release(page);
991
992out_unmap:
Hugh Dickinsc0718802005-10-29 18:16:31 -0700993 pte_unmap_unlock(pte, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700994out:
995 return ret;
KOSAKI Motohirocaed0f42009-12-14 17:59:45 -0800996
997out_mlock:
998 pte_unmap_unlock(pte, ptl);
999
1000
1001 /*
1002 * We need mmap_sem locking, Otherwise VM_LOCKED check makes
1003 * unstable result and race. Plus, We can't wait here because
1004 * we now hold anon_vma->lock or mapping->i_mmap_lock.
1005 * if trylock failed, the page remain in evictable lru and later
1006 * vmscan could retry to move the page to unevictable lru if the
1007 * page is actually mlocked.
1008 */
1009 if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
1010 if (vma->vm_flags & VM_LOCKED) {
1011 mlock_vma_page(page);
1012 ret = SWAP_MLOCK;
1013 }
1014 up_read(&vma->vm_mm->mmap_sem);
1015 }
1016 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001017}
1018
1019/*
1020 * objrmap doesn't work for nonlinear VMAs because the assumption that
1021 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
1022 * Consequently, given a particular page and its ->index, we cannot locate the
1023 * ptes which are mapping that page without an exhaustive linear search.
1024 *
1025 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
1026 * maps the file to which the target page belongs. The ->vm_private_data field
1027 * holds the current cursor into that scan. Successive searches will circulate
1028 * around the vma's virtual address space.
1029 *
1030 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
1031 * more scanning pressure is placed against them as well. Eventually pages
1032 * will become fully unmapped and are eligible for eviction.
1033 *
1034 * For very sparsely populated VMAs this is a little inefficient - chances are
1035 * there there won't be many ptes located within the scan cluster. In this case
1036 * maybe we could scan further - to the end of the pte page, perhaps.
Nick Pigginb291f002008-10-18 20:26:44 -07001037 *
1038 * Mlocked pages: check VM_LOCKED under mmap_sem held for read, if we can
1039 * acquire it without blocking. If vma locked, mlock the pages in the cluster,
1040 * rather than unmapping them. If we encounter the "check_page" that vmscan is
1041 * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001042 */
1043#define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
1044#define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
1045
Nick Pigginb291f002008-10-18 20:26:44 -07001046static int try_to_unmap_cluster(unsigned long cursor, unsigned int *mapcount,
1047 struct vm_area_struct *vma, struct page *check_page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001048{
1049 struct mm_struct *mm = vma->vm_mm;
1050 pgd_t *pgd;
1051 pud_t *pud;
1052 pmd_t *pmd;
Hugh Dickinsc0718802005-10-29 18:16:31 -07001053 pte_t *pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001054 pte_t pteval;
Hugh Dickinsc0718802005-10-29 18:16:31 -07001055 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001056 struct page *page;
1057 unsigned long address;
1058 unsigned long end;
Nick Pigginb291f002008-10-18 20:26:44 -07001059 int ret = SWAP_AGAIN;
1060 int locked_vma = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001061
Linus Torvalds1da177e2005-04-16 15:20:36 -07001062 address = (vma->vm_start + cursor) & CLUSTER_MASK;
1063 end = address + CLUSTER_SIZE;
1064 if (address < vma->vm_start)
1065 address = vma->vm_start;
1066 if (end > vma->vm_end)
1067 end = vma->vm_end;
1068
1069 pgd = pgd_offset(mm, address);
1070 if (!pgd_present(*pgd))
Nick Pigginb291f002008-10-18 20:26:44 -07001071 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001072
1073 pud = pud_offset(pgd, address);
1074 if (!pud_present(*pud))
Nick Pigginb291f002008-10-18 20:26:44 -07001075 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001076
1077 pmd = pmd_offset(pud, address);
1078 if (!pmd_present(*pmd))
Nick Pigginb291f002008-10-18 20:26:44 -07001079 return ret;
1080
1081 /*
Hugh Dickinsaf8e3352009-12-14 17:58:59 -08001082 * If we can acquire the mmap_sem for read, and vma is VM_LOCKED,
Nick Pigginb291f002008-10-18 20:26:44 -07001083 * keep the sem while scanning the cluster for mlocking pages.
1084 */
Hugh Dickinsaf8e3352009-12-14 17:58:59 -08001085 if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
Nick Pigginb291f002008-10-18 20:26:44 -07001086 locked_vma = (vma->vm_flags & VM_LOCKED);
1087 if (!locked_vma)
1088 up_read(&vma->vm_mm->mmap_sem); /* don't need it */
1089 }
Hugh Dickinsc0718802005-10-29 18:16:31 -07001090
1091 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001092
Hugh Dickins365e9c872005-10-29 18:16:18 -07001093 /* Update high watermark before we lower rss */
1094 update_hiwater_rss(mm);
1095
Hugh Dickinsc0718802005-10-29 18:16:31 -07001096 for (; address < end; pte++, address += PAGE_SIZE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001097 if (!pte_present(*pte))
1098 continue;
Linus Torvalds6aab3412005-11-28 14:34:23 -08001099 page = vm_normal_page(vma, address, *pte);
1100 BUG_ON(!page || PageAnon(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001101
Nick Pigginb291f002008-10-18 20:26:44 -07001102 if (locked_vma) {
1103 mlock_vma_page(page); /* no-op if already mlocked */
1104 if (page == check_page)
1105 ret = SWAP_MLOCK;
1106 continue; /* don't unmap */
1107 }
1108
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001109 if (ptep_clear_flush_young_notify(vma, address, pte))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001110 continue;
1111
1112 /* Nuke the page table entry. */
Ben Collinseca35132005-11-29 11:45:26 -08001113 flush_cache_page(vma, address, pte_pfn(*pte));
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001114 pteval = ptep_clear_flush_notify(vma, address, pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001115
1116 /* If nonlinear, store the file page offset in the pte. */
1117 if (page->index != linear_page_index(vma, address))
1118 set_pte_at(mm, address, pte, pgoff_to_pte(page->index));
1119
1120 /* Move the dirty bit to the physical page now the pte is gone. */
1121 if (pte_dirty(pteval))
1122 set_page_dirty(page);
1123
Hugh Dickinsedc315f2009-01-06 14:40:11 -08001124 page_remove_rmap(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001125 page_cache_release(page);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001126 dec_mm_counter(mm, MM_FILEPAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001127 (*mapcount)--;
1128 }
Hugh Dickinsc0718802005-10-29 18:16:31 -07001129 pte_unmap_unlock(pte - 1, ptl);
Nick Pigginb291f002008-10-18 20:26:44 -07001130 if (locked_vma)
1131 up_read(&vma->vm_mm->mmap_sem);
1132 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001133}
1134
Nick Pigginb291f002008-10-18 20:26:44 -07001135/**
1136 * try_to_unmap_anon - unmap or unlock anonymous page using the object-based
1137 * rmap method
1138 * @page: the page to unmap/unlock
Huang Shijie8051be52009-12-14 17:58:50 -08001139 * @flags: action and flags
Nick Pigginb291f002008-10-18 20:26:44 -07001140 *
1141 * Find all the mappings of a page using the mapping pointer and the vma chains
1142 * contained in the anon_vma struct it points to.
1143 *
1144 * This function is only called from try_to_unmap/try_to_munlock for
1145 * anonymous pages.
1146 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1147 * where the page was found will be held for write. So, we won't recheck
1148 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1149 * 'LOCKED.
1150 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001151static int try_to_unmap_anon(struct page *page, enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001152{
1153 struct anon_vma *anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -08001154 struct anon_vma_chain *avc;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001155 int ret = SWAP_AGAIN;
Nick Pigginb291f002008-10-18 20:26:44 -07001156
Linus Torvalds1da177e2005-04-16 15:20:36 -07001157 anon_vma = page_lock_anon_vma(page);
1158 if (!anon_vma)
1159 return ret;
1160
Rik van Riel5beb4932010-03-05 13:42:07 -08001161 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1162 struct vm_area_struct *vma = avc->vma;
Hugh Dickins1cb17292009-12-14 17:59:01 -08001163 unsigned long address = vma_address(page, vma);
1164 if (address == -EFAULT)
1165 continue;
1166 ret = try_to_unmap_one(page, vma, address, flags);
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001167 if (ret != SWAP_AGAIN || !page_mapped(page))
1168 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001169 }
Oleg Nesterov34bbd702007-02-28 20:13:49 -08001170
1171 page_unlock_anon_vma(anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001172 return ret;
1173}
1174
1175/**
Nick Pigginb291f002008-10-18 20:26:44 -07001176 * try_to_unmap_file - unmap/unlock file page using the object-based rmap method
1177 * @page: the page to unmap/unlock
Andi Kleen14fa31b2009-09-16 11:50:10 +02001178 * @flags: action and flags
Linus Torvalds1da177e2005-04-16 15:20:36 -07001179 *
1180 * Find all the mappings of a page using the mapping pointer and the vma chains
1181 * contained in the address_space struct it points to.
1182 *
Nick Pigginb291f002008-10-18 20:26:44 -07001183 * This function is only called from try_to_unmap/try_to_munlock for
1184 * object-based pages.
1185 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1186 * where the page was found will be held for write. So, we won't recheck
1187 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1188 * 'LOCKED.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001189 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001190static int try_to_unmap_file(struct page *page, enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001191{
1192 struct address_space *mapping = page->mapping;
1193 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1194 struct vm_area_struct *vma;
1195 struct prio_tree_iter iter;
1196 int ret = SWAP_AGAIN;
1197 unsigned long cursor;
1198 unsigned long max_nl_cursor = 0;
1199 unsigned long max_nl_size = 0;
1200 unsigned int mapcount;
1201
1202 spin_lock(&mapping->i_mmap_lock);
1203 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
Hugh Dickins1cb17292009-12-14 17:59:01 -08001204 unsigned long address = vma_address(page, vma);
1205 if (address == -EFAULT)
1206 continue;
1207 ret = try_to_unmap_one(page, vma, address, flags);
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001208 if (ret != SWAP_AGAIN || !page_mapped(page))
1209 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001210 }
1211
1212 if (list_empty(&mapping->i_mmap_nonlinear))
1213 goto out;
1214
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001215 /*
1216 * We don't bother to try to find the munlocked page in nonlinears.
1217 * It's costly. Instead, later, page reclaim logic may call
1218 * try_to_unmap(TTU_MUNLOCK) and recover PG_mlocked lazily.
1219 */
1220 if (TTU_ACTION(flags) == TTU_MUNLOCK)
1221 goto out;
1222
Linus Torvalds1da177e2005-04-16 15:20:36 -07001223 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1224 shared.vm_set.list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001225 cursor = (unsigned long) vma->vm_private_data;
1226 if (cursor > max_nl_cursor)
1227 max_nl_cursor = cursor;
1228 cursor = vma->vm_end - vma->vm_start;
1229 if (cursor > max_nl_size)
1230 max_nl_size = cursor;
1231 }
1232
Nick Pigginb291f002008-10-18 20:26:44 -07001233 if (max_nl_size == 0) { /* all nonlinears locked or reserved ? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001234 ret = SWAP_FAIL;
1235 goto out;
1236 }
1237
1238 /*
1239 * We don't try to search for this page in the nonlinear vmas,
1240 * and page_referenced wouldn't have found it anyway. Instead
1241 * just walk the nonlinear vmas trying to age and unmap some.
1242 * The mapcount of the page we came in with is irrelevant,
1243 * but even so use it as a guide to how hard we should try?
1244 */
1245 mapcount = page_mapcount(page);
1246 if (!mapcount)
1247 goto out;
1248 cond_resched_lock(&mapping->i_mmap_lock);
1249
1250 max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
1251 if (max_nl_cursor == 0)
1252 max_nl_cursor = CLUSTER_SIZE;
1253
1254 do {
1255 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1256 shared.vm_set.list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001257 cursor = (unsigned long) vma->vm_private_data;
Hugh Dickins839b9682005-09-03 15:54:43 -07001258 while ( cursor < max_nl_cursor &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07001259 cursor < vma->vm_end - vma->vm_start) {
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001260 if (try_to_unmap_cluster(cursor, &mapcount,
1261 vma, page) == SWAP_MLOCK)
1262 ret = SWAP_MLOCK;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001263 cursor += CLUSTER_SIZE;
1264 vma->vm_private_data = (void *) cursor;
1265 if ((int)mapcount <= 0)
1266 goto out;
1267 }
1268 vma->vm_private_data = (void *) max_nl_cursor;
1269 }
1270 cond_resched_lock(&mapping->i_mmap_lock);
1271 max_nl_cursor += CLUSTER_SIZE;
1272 } while (max_nl_cursor <= max_nl_size);
1273
1274 /*
1275 * Don't loop forever (perhaps all the remaining pages are
1276 * in locked vmas). Reset cursor on all unreserved nonlinear
1277 * vmas, now forgetting on which ones it had fallen behind.
1278 */
Hugh Dickins101d2be2005-11-21 21:32:16 -08001279 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1280 vma->vm_private_data = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001281out:
1282 spin_unlock(&mapping->i_mmap_lock);
1283 return ret;
1284}
1285
1286/**
1287 * try_to_unmap - try to remove all page table mappings to a page
1288 * @page: the page to get unmapped
Andi Kleen14fa31b2009-09-16 11:50:10 +02001289 * @flags: action and flags
Linus Torvalds1da177e2005-04-16 15:20:36 -07001290 *
1291 * Tries to remove all the page table entries which are mapping this
1292 * page, used in the pageout path. Caller must hold the page lock.
1293 * Return values are:
1294 *
1295 * SWAP_SUCCESS - we succeeded in removing all mappings
1296 * SWAP_AGAIN - we missed a mapping, try again later
1297 * SWAP_FAIL - the page is unswappable
Nick Pigginb291f002008-10-18 20:26:44 -07001298 * SWAP_MLOCK - page is mlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001299 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001300int try_to_unmap(struct page *page, enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001301{
1302 int ret;
1303
Linus Torvalds1da177e2005-04-16 15:20:36 -07001304 BUG_ON(!PageLocked(page));
1305
Hugh Dickins5ad64682009-12-14 17:59:24 -08001306 if (unlikely(PageKsm(page)))
1307 ret = try_to_unmap_ksm(page, flags);
1308 else if (PageAnon(page))
Andi Kleen14fa31b2009-09-16 11:50:10 +02001309 ret = try_to_unmap_anon(page, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001310 else
Andi Kleen14fa31b2009-09-16 11:50:10 +02001311 ret = try_to_unmap_file(page, flags);
Nick Pigginb291f002008-10-18 20:26:44 -07001312 if (ret != SWAP_MLOCK && !page_mapped(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001313 ret = SWAP_SUCCESS;
1314 return ret;
1315}
Nikita Danilov81b40822005-05-01 08:58:36 -07001316
Nick Pigginb291f002008-10-18 20:26:44 -07001317/**
1318 * try_to_munlock - try to munlock a page
1319 * @page: the page to be munlocked
1320 *
1321 * Called from munlock code. Checks all of the VMAs mapping the page
1322 * to make sure nobody else has this page mlocked. The page will be
1323 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1324 *
1325 * Return values are:
1326 *
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001327 * SWAP_AGAIN - no vma is holding page mlocked, or,
Nick Pigginb291f002008-10-18 20:26:44 -07001328 * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem
Hugh Dickins5ad64682009-12-14 17:59:24 -08001329 * SWAP_FAIL - page cannot be located at present
Nick Pigginb291f002008-10-18 20:26:44 -07001330 * SWAP_MLOCK - page is now mlocked.
1331 */
1332int try_to_munlock(struct page *page)
1333{
1334 VM_BUG_ON(!PageLocked(page) || PageLRU(page));
1335
Hugh Dickins5ad64682009-12-14 17:59:24 -08001336 if (unlikely(PageKsm(page)))
1337 return try_to_unmap_ksm(page, TTU_MUNLOCK);
1338 else if (PageAnon(page))
Andi Kleen14fa31b2009-09-16 11:50:10 +02001339 return try_to_unmap_anon(page, TTU_MUNLOCK);
Nick Pigginb291f002008-10-18 20:26:44 -07001340 else
Andi Kleen14fa31b2009-09-16 11:50:10 +02001341 return try_to_unmap_file(page, TTU_MUNLOCK);
Nick Pigginb291f002008-10-18 20:26:44 -07001342}
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001343
1344#ifdef CONFIG_MIGRATION
1345/*
1346 * rmap_walk() and its helpers rmap_walk_anon() and rmap_walk_file():
1347 * Called by migrate.c to remove migration ptes, but might be used more later.
1348 */
1349static int rmap_walk_anon(struct page *page, int (*rmap_one)(struct page *,
1350 struct vm_area_struct *, unsigned long, void *), void *arg)
1351{
1352 struct anon_vma *anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -08001353 struct anon_vma_chain *avc;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001354 int ret = SWAP_AGAIN;
1355
1356 /*
1357 * Note: remove_migration_ptes() cannot use page_lock_anon_vma()
1358 * because that depends on page_mapped(); but not all its usages
1359 * are holding mmap_sem, which also gave the necessary guarantee
1360 * (that this anon_vma's slab has not already been destroyed).
1361 * This needs to be reviewed later: avoiding page_lock_anon_vma()
1362 * is risky, and currently limits the usefulness of rmap_walk().
1363 */
1364 anon_vma = page_anon_vma(page);
1365 if (!anon_vma)
1366 return ret;
1367 spin_lock(&anon_vma->lock);
Rik van Riel5beb4932010-03-05 13:42:07 -08001368 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1369 struct vm_area_struct *vma = avc->vma;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001370 unsigned long address = vma_address(page, vma);
1371 if (address == -EFAULT)
1372 continue;
1373 ret = rmap_one(page, vma, address, arg);
1374 if (ret != SWAP_AGAIN)
1375 break;
1376 }
1377 spin_unlock(&anon_vma->lock);
1378 return ret;
1379}
1380
1381static int rmap_walk_file(struct page *page, int (*rmap_one)(struct page *,
1382 struct vm_area_struct *, unsigned long, void *), void *arg)
1383{
1384 struct address_space *mapping = page->mapping;
1385 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1386 struct vm_area_struct *vma;
1387 struct prio_tree_iter iter;
1388 int ret = SWAP_AGAIN;
1389
1390 if (!mapping)
1391 return ret;
1392 spin_lock(&mapping->i_mmap_lock);
1393 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1394 unsigned long address = vma_address(page, vma);
1395 if (address == -EFAULT)
1396 continue;
1397 ret = rmap_one(page, vma, address, arg);
1398 if (ret != SWAP_AGAIN)
1399 break;
1400 }
1401 /*
1402 * No nonlinear handling: being always shared, nonlinear vmas
1403 * never contain migration ptes. Decide what to do about this
1404 * limitation to linear when we need rmap_walk() on nonlinear.
1405 */
1406 spin_unlock(&mapping->i_mmap_lock);
1407 return ret;
1408}
1409
1410int rmap_walk(struct page *page, int (*rmap_one)(struct page *,
1411 struct vm_area_struct *, unsigned long, void *), void *arg)
1412{
1413 VM_BUG_ON(!PageLocked(page));
1414
1415 if (unlikely(PageKsm(page)))
1416 return rmap_walk_ksm(page, rmap_one, arg);
1417 else if (PageAnon(page))
1418 return rmap_walk_anon(page, rmap_one, arg);
1419 else
1420 return rmap_walk_file(page, rmap_one, arg);
1421}
1422#endif /* CONFIG_MIGRATION */