blob: 278cd277bdec75cfb3de7bb01157ad2e092f622a [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;
65
66static inline struct anon_vma *anon_vma_alloc(void)
67{
68 return kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
69}
70
Hugh Dickinsdb114b82009-12-14 17:59:25 -080071void anon_vma_free(struct anon_vma *anon_vma)
Adrian Bunkfdd2e5f2008-10-18 20:28:38 -070072{
73 kmem_cache_free(anon_vma_cachep, anon_vma);
74}
Linus Torvalds1da177e2005-04-16 15:20:36 -070075
Linus Torvaldsd9d332e2008-10-19 10:32:20 -070076/**
77 * anon_vma_prepare - attach an anon_vma to a memory region
78 * @vma: the memory region in question
79 *
80 * This makes sure the memory mapping described by 'vma' has
81 * an 'anon_vma' attached to it, so that we can associate the
82 * anonymous pages mapped into it with that anon_vma.
83 *
84 * The common case will be that we already have one, but if
85 * if not we either need to find an adjacent mapping that we
86 * can re-use the anon_vma from (very common when the only
87 * reason for splitting a vma has been mprotect()), or we
88 * allocate a new one.
89 *
90 * Anon-vma allocations are very subtle, because we may have
91 * optimistically looked up an anon_vma in page_lock_anon_vma()
92 * and that may actually touch the spinlock even in the newly
93 * allocated vma (it depends on RCU to make sure that the
94 * anon_vma isn't actually destroyed).
95 *
96 * As a result, we need to do proper anon_vma locking even
97 * for the new allocation. At the same time, we do not want
98 * to do any locking for the common case of already having
99 * an anon_vma.
100 *
101 * This must be called with the mmap_sem held for reading.
102 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700103int anon_vma_prepare(struct vm_area_struct *vma)
104{
105 struct anon_vma *anon_vma = vma->anon_vma;
106
107 might_sleep();
108 if (unlikely(!anon_vma)) {
109 struct mm_struct *mm = vma->vm_mm;
Linus Torvaldsd9d332e2008-10-19 10:32:20 -0700110 struct anon_vma *allocated;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700111
112 anon_vma = find_mergeable_anon_vma(vma);
Linus Torvaldsd9d332e2008-10-19 10:32:20 -0700113 allocated = NULL;
114 if (!anon_vma) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700115 anon_vma = anon_vma_alloc();
116 if (unlikely(!anon_vma))
117 return -ENOMEM;
118 allocated = anon_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700119 }
Linus Torvaldsd9d332e2008-10-19 10:32:20 -0700120 spin_lock(&anon_vma->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700121
122 /* page_table_lock to protect against threads */
123 spin_lock(&mm->page_table_lock);
124 if (likely(!vma->anon_vma)) {
125 vma->anon_vma = anon_vma;
Christoph Lameter06972122006-06-23 02:03:35 -0700126 list_add_tail(&vma->anon_vma_node, &anon_vma->head);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700127 allocated = NULL;
128 }
129 spin_unlock(&mm->page_table_lock);
130
Linus Torvaldsd9d332e2008-10-19 10:32:20 -0700131 spin_unlock(&anon_vma->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700132 if (unlikely(allocated))
133 anon_vma_free(allocated);
134 }
135 return 0;
136}
137
138void __anon_vma_merge(struct vm_area_struct *vma, struct vm_area_struct *next)
139{
140 BUG_ON(vma->anon_vma != next->anon_vma);
141 list_del(&next->anon_vma_node);
142}
143
144void __anon_vma_link(struct vm_area_struct *vma)
145{
146 struct anon_vma *anon_vma = vma->anon_vma;
147
Hugh Dickins30acbab2007-06-27 14:09:53 -0700148 if (anon_vma)
Christoph Lameter06972122006-06-23 02:03:35 -0700149 list_add_tail(&vma->anon_vma_node, &anon_vma->head);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700150}
151
152void anon_vma_link(struct vm_area_struct *vma)
153{
154 struct anon_vma *anon_vma = vma->anon_vma;
155
156 if (anon_vma) {
157 spin_lock(&anon_vma->lock);
Christoph Lameter06972122006-06-23 02:03:35 -0700158 list_add_tail(&vma->anon_vma_node, &anon_vma->head);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700159 spin_unlock(&anon_vma->lock);
160 }
161}
162
163void anon_vma_unlink(struct vm_area_struct *vma)
164{
165 struct anon_vma *anon_vma = vma->anon_vma;
166 int empty;
167
168 if (!anon_vma)
169 return;
170
171 spin_lock(&anon_vma->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700172 list_del(&vma->anon_vma_node);
173
174 /* We must garbage collect the anon_vma if it's empty */
Hugh Dickinsdb114b82009-12-14 17:59:25 -0800175 empty = list_empty(&anon_vma->head) && !ksm_refcount(anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700176 spin_unlock(&anon_vma->lock);
177
178 if (empty)
179 anon_vma_free(anon_vma);
180}
181
Alexey Dobriyan51cc5062008-07-25 19:45:34 -0700182static void anon_vma_ctor(void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700183{
Christoph Lametera35afb82007-05-16 22:10:57 -0700184 struct anon_vma *anon_vma = data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700185
Christoph Lametera35afb82007-05-16 22:10:57 -0700186 spin_lock_init(&anon_vma->lock);
Hugh Dickinsdb114b82009-12-14 17:59:25 -0800187 ksm_refcount_init(anon_vma);
Christoph Lametera35afb82007-05-16 22:10:57 -0700188 INIT_LIST_HEAD(&anon_vma->head);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700189}
190
191void __init anon_vma_init(void)
192{
193 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
Paul Mundt20c2df82007-07-20 10:11:58 +0900194 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700195}
196
197/*
198 * Getting a lock on a stable anon_vma from a page off the LRU is
199 * tricky: page_lock_anon_vma rely on RCU to guard against the races.
200 */
Andi Kleen10be22d2009-09-16 11:50:04 +0200201struct anon_vma *page_lock_anon_vma(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700202{
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800203 struct anon_vma *anon_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700204 unsigned long anon_mapping;
205
206 rcu_read_lock();
Hugh Dickins80e14822009-12-14 17:59:29 -0800207 anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping);
Hugh Dickins3ca7b3c2009-12-14 17:58:57 -0800208 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700209 goto out;
210 if (!page_mapped(page))
211 goto out;
212
213 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
214 spin_lock(&anon_vma->lock);
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800215 return anon_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700216out:
217 rcu_read_unlock();
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800218 return NULL;
219}
220
Andi Kleen10be22d2009-09-16 11:50:04 +0200221void page_unlock_anon_vma(struct anon_vma *anon_vma)
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800222{
223 spin_unlock(&anon_vma->lock);
224 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700225}
226
227/*
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -0800228 * At what user virtual address is page expected in @vma?
229 * Returns virtual address or -EFAULT if page's index/offset is not
230 * within the range mapped the @vma.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700231 */
232static inline unsigned long
233vma_address(struct page *page, struct vm_area_struct *vma)
234{
235 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
236 unsigned long address;
237
238 address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
239 if (unlikely(address < vma->vm_start || address >= vma->vm_end)) {
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -0800240 /* page should be within @vma mapping range */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700241 return -EFAULT;
242 }
243 return address;
244}
245
246/*
Huang Shijiebf89c8c2009-10-01 15:44:04 -0700247 * At what user virtual address is page expected in vma?
248 * checking that the page matches the vma.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700249 */
250unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
251{
252 if (PageAnon(page)) {
Hugh Dickins3ca7b3c2009-12-14 17:58:57 -0800253 if (vma->anon_vma != page_anon_vma(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700254 return -EFAULT;
255 } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) {
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800256 if (!vma->vm_file ||
257 vma->vm_file->f_mapping != page->mapping)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700258 return -EFAULT;
259 } else
260 return -EFAULT;
261 return vma_address(page, vma);
262}
263
264/*
Nikita Danilov81b40822005-05-01 08:58:36 -0700265 * Check that @page is mapped at @address into @mm.
266 *
Nick Piggin479db0b2008-08-20 14:09:18 -0700267 * If @sync is false, page_check_address may perform a racy check to avoid
268 * the page table lock when the pte is not present (helpful when reclaiming
269 * highly shared pages).
270 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -0700271 * On success returns with pte mapped and locked.
Nikita Danilov81b40822005-05-01 08:58:36 -0700272 */
Carsten Otteceffc072005-06-23 22:05:25 -0700273pte_t *page_check_address(struct page *page, struct mm_struct *mm,
Nick Piggin479db0b2008-08-20 14:09:18 -0700274 unsigned long address, spinlock_t **ptlp, int sync)
Nikita Danilov81b40822005-05-01 08:58:36 -0700275{
276 pgd_t *pgd;
277 pud_t *pud;
278 pmd_t *pmd;
279 pte_t *pte;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700280 spinlock_t *ptl;
Nikita Danilov81b40822005-05-01 08:58:36 -0700281
Nikita Danilov81b40822005-05-01 08:58:36 -0700282 pgd = pgd_offset(mm, address);
Hugh Dickinsc0718802005-10-29 18:16:31 -0700283 if (!pgd_present(*pgd))
284 return NULL;
285
286 pud = pud_offset(pgd, address);
287 if (!pud_present(*pud))
288 return NULL;
289
290 pmd = pmd_offset(pud, address);
291 if (!pmd_present(*pmd))
292 return NULL;
293
294 pte = pte_offset_map(pmd, address);
295 /* Make a quick check before getting the lock */
Nick Piggin479db0b2008-08-20 14:09:18 -0700296 if (!sync && !pte_present(*pte)) {
Hugh Dickinsc0718802005-10-29 18:16:31 -0700297 pte_unmap(pte);
298 return NULL;
Nikita Danilov81b40822005-05-01 08:58:36 -0700299 }
Hugh Dickinsc0718802005-10-29 18:16:31 -0700300
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700301 ptl = pte_lockptr(mm, pmd);
Hugh Dickinsc0718802005-10-29 18:16:31 -0700302 spin_lock(ptl);
303 if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
304 *ptlp = ptl;
305 return pte;
306 }
307 pte_unmap_unlock(pte, ptl);
308 return NULL;
Nikita Danilov81b40822005-05-01 08:58:36 -0700309}
310
Nick Pigginb291f002008-10-18 20:26:44 -0700311/**
312 * page_mapped_in_vma - check whether a page is really mapped in a VMA
313 * @page: the page to test
314 * @vma: the VMA to test
315 *
316 * Returns 1 if the page is mapped into the page tables of the VMA, 0
317 * if the page is not mapped into the page tables of this VMA. Only
318 * valid for normal file or anonymous VMAs.
319 */
Andi Kleen6a460792009-09-16 11:50:15 +0200320int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
Nick Pigginb291f002008-10-18 20:26:44 -0700321{
322 unsigned long address;
323 pte_t *pte;
324 spinlock_t *ptl;
325
326 address = vma_address(page, vma);
327 if (address == -EFAULT) /* out of vma range */
328 return 0;
329 pte = page_check_address(page, vma->vm_mm, address, &ptl, 1);
330 if (!pte) /* the page is not in this mm */
331 return 0;
332 pte_unmap_unlock(pte, ptl);
333
334 return 1;
335}
336
Nikita Danilov81b40822005-05-01 08:58:36 -0700337/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700338 * Subfunctions of page_referenced: page_referenced_one called
339 * repeatedly from either page_referenced_anon or page_referenced_file.
340 */
Hugh Dickins5ad64682009-12-14 17:59:24 -0800341int page_referenced_one(struct page *page, struct vm_area_struct *vma,
342 unsigned long address, unsigned int *mapcount,
343 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700344{
345 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700346 pte_t *pte;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700347 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700348 int referenced = 0;
349
Nick Piggin479db0b2008-08-20 14:09:18 -0700350 pte = page_check_address(page, mm, address, &ptl, 0);
Hugh Dickinsc0718802005-10-29 18:16:31 -0700351 if (!pte)
352 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700353
Nick Pigginb291f002008-10-18 20:26:44 -0700354 /*
355 * Don't want to elevate referenced for mlocked page that gets this far,
356 * in order that it progresses to try_to_unmap and is moved to the
357 * unevictable list.
358 */
Hugh Dickins5a9bbdc2008-02-04 22:29:23 -0800359 if (vma->vm_flags & VM_LOCKED) {
Hugh Dickins5a9bbdc2008-02-04 22:29:23 -0800360 *mapcount = 1; /* break early from loop */
Minchan Kim03ef83a2009-08-26 14:29:23 -0700361 *vm_flags |= VM_LOCKED;
Nick Pigginb291f002008-10-18 20:26:44 -0700362 goto out_unmap;
363 }
364
Johannes Weiner4917e5d2009-01-06 14:39:17 -0800365 if (ptep_clear_flush_young_notify(vma, address, pte)) {
366 /*
367 * Don't treat a reference through a sequentially read
368 * mapping as such. If the page has been used in
369 * another mapping, we will catch it; if this other
370 * mapping is already gone, the unmap path will have
371 * set PG_referenced or activated the page.
372 */
373 if (likely(!VM_SequentialReadHint(vma)))
374 referenced++;
375 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700376
Hugh Dickinsc0718802005-10-29 18:16:31 -0700377 /* Pretend the page is referenced if the task has the
378 swap token and is in the middle of a page fault. */
Rik van Rielf7b7fd82005-11-28 13:44:07 -0800379 if (mm != current->mm && has_swap_token(mm) &&
Hugh Dickinsc0718802005-10-29 18:16:31 -0700380 rwsem_is_locked(&mm->mmap_sem))
381 referenced++;
382
Nick Pigginb291f002008-10-18 20:26:44 -0700383out_unmap:
Hugh Dickinsc0718802005-10-29 18:16:31 -0700384 (*mapcount)--;
385 pte_unmap_unlock(pte, ptl);
Huang Shijie273f0472009-12-14 17:58:51 -0800386
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700387 if (referenced)
388 *vm_flags |= vma->vm_flags;
Huang Shijie273f0472009-12-14 17:58:51 -0800389out:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700390 return referenced;
391}
392
Balbir Singhbed71612008-02-07 00:14:01 -0800393static int page_referenced_anon(struct page *page,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700394 struct mem_cgroup *mem_cont,
395 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700396{
397 unsigned int mapcount;
398 struct anon_vma *anon_vma;
399 struct vm_area_struct *vma;
400 int referenced = 0;
401
402 anon_vma = page_lock_anon_vma(page);
403 if (!anon_vma)
404 return referenced;
405
406 mapcount = page_mapcount(page);
407 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
Hugh Dickins1cb17292009-12-14 17:59:01 -0800408 unsigned long address = vma_address(page, vma);
409 if (address == -EFAULT)
410 continue;
Balbir Singhbed71612008-02-07 00:14:01 -0800411 /*
412 * If we are reclaiming on behalf of a cgroup, skip
413 * counting on behalf of references from different
414 * cgroups
415 */
Hugh Dickinsbd845e32008-03-04 14:29:01 -0800416 if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
Balbir Singhbed71612008-02-07 00:14:01 -0800417 continue;
Hugh Dickins1cb17292009-12-14 17:59:01 -0800418 referenced += page_referenced_one(page, vma, address,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700419 &mapcount, vm_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700420 if (!mapcount)
421 break;
422 }
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800423
424 page_unlock_anon_vma(anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700425 return referenced;
426}
427
428/**
429 * page_referenced_file - referenced check for object-based rmap
430 * @page: the page we're checking references on.
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700431 * @mem_cont: target memory controller
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700432 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
Linus Torvalds1da177e2005-04-16 15:20:36 -0700433 *
434 * For an object-based mapped page, find all the places it is mapped and
435 * check/clear the referenced flag. This is done by following the page->mapping
436 * pointer, then walking the chain of vmas it holds. It returns the number
437 * of references it found.
438 *
439 * This function is only called from page_referenced for object-based pages.
440 */
Balbir Singhbed71612008-02-07 00:14:01 -0800441static int page_referenced_file(struct page *page,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700442 struct mem_cgroup *mem_cont,
443 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700444{
445 unsigned int mapcount;
446 struct address_space *mapping = page->mapping;
447 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
448 struct vm_area_struct *vma;
449 struct prio_tree_iter iter;
450 int referenced = 0;
451
452 /*
453 * The caller's checks on page->mapping and !PageAnon have made
454 * sure that this is a file page: the check for page->mapping
455 * excludes the case just before it gets set on an anon page.
456 */
457 BUG_ON(PageAnon(page));
458
459 /*
460 * The page lock not only makes sure that page->mapping cannot
461 * suddenly be NULLified by truncation, it makes sure that the
462 * structure at mapping cannot be freed and reused yet,
463 * so we can safely take mapping->i_mmap_lock.
464 */
465 BUG_ON(!PageLocked(page));
466
467 spin_lock(&mapping->i_mmap_lock);
468
469 /*
470 * i_mmap_lock does not stabilize mapcount at all, but mapcount
471 * is more likely to be accurate if we note it after spinning.
472 */
473 mapcount = page_mapcount(page);
474
475 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
Hugh Dickins1cb17292009-12-14 17:59:01 -0800476 unsigned long address = vma_address(page, vma);
477 if (address == -EFAULT)
478 continue;
Balbir Singhbed71612008-02-07 00:14:01 -0800479 /*
480 * If we are reclaiming on behalf of a cgroup, skip
481 * counting on behalf of references from different
482 * cgroups
483 */
Hugh Dickinsbd845e32008-03-04 14:29:01 -0800484 if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
Balbir Singhbed71612008-02-07 00:14:01 -0800485 continue;
Hugh Dickins1cb17292009-12-14 17:59:01 -0800486 referenced += page_referenced_one(page, vma, address,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700487 &mapcount, vm_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700488 if (!mapcount)
489 break;
490 }
491
492 spin_unlock(&mapping->i_mmap_lock);
493 return referenced;
494}
495
496/**
497 * page_referenced - test if the page was referenced
498 * @page: the page to test
499 * @is_locked: caller holds lock on the page
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700500 * @mem_cont: target memory controller
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700501 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
Linus Torvalds1da177e2005-04-16 15:20:36 -0700502 *
503 * Quick test_and_clear_referenced for all mappings to a page,
504 * returns the number of ptes which referenced the page.
505 */
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700506int page_referenced(struct page *page,
507 int is_locked,
508 struct mem_cgroup *mem_cont,
509 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700510{
511 int referenced = 0;
Hugh Dickins5ad64682009-12-14 17:59:24 -0800512 int we_locked = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700513
Linus Torvalds1da177e2005-04-16 15:20:36 -0700514 if (TestClearPageReferenced(page))
515 referenced++;
516
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700517 *vm_flags = 0;
Hugh Dickins3ca7b3c2009-12-14 17:58:57 -0800518 if (page_mapped(page) && page_rmapping(page)) {
Hugh Dickins5ad64682009-12-14 17:59:24 -0800519 if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
520 we_locked = trylock_page(page);
521 if (!we_locked) {
522 referenced++;
523 goto out;
524 }
525 }
526 if (unlikely(PageKsm(page)))
527 referenced += page_referenced_ksm(page, mem_cont,
528 vm_flags);
529 else if (PageAnon(page))
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700530 referenced += page_referenced_anon(page, mem_cont,
531 vm_flags);
Hugh Dickins5ad64682009-12-14 17:59:24 -0800532 else if (page->mapping)
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700533 referenced += page_referenced_file(page, mem_cont,
534 vm_flags);
Hugh Dickins5ad64682009-12-14 17:59:24 -0800535 if (we_locked)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700536 unlock_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700537 }
Hugh Dickins5ad64682009-12-14 17:59:24 -0800538out:
Christian Borntraeger5b7baf02008-03-25 18:47:12 +0100539 if (page_test_and_clear_young(page))
540 referenced++;
541
Linus Torvalds1da177e2005-04-16 15:20:36 -0700542 return referenced;
543}
544
Hugh Dickins1cb17292009-12-14 17:59:01 -0800545static int page_mkclean_one(struct page *page, struct vm_area_struct *vma,
546 unsigned long address)
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700547{
548 struct mm_struct *mm = vma->vm_mm;
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100549 pte_t *pte;
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700550 spinlock_t *ptl;
551 int ret = 0;
552
Nick Piggin479db0b2008-08-20 14:09:18 -0700553 pte = page_check_address(page, mm, address, &ptl, 1);
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700554 if (!pte)
555 goto out;
556
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100557 if (pte_dirty(*pte) || pte_write(*pte)) {
558 pte_t entry;
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700559
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100560 flush_cache_page(vma, address, pte_pfn(*pte));
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700561 entry = ptep_clear_flush_notify(vma, address, pte);
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100562 entry = pte_wrprotect(entry);
563 entry = pte_mkclean(entry);
Al Virod6e88e62006-12-29 16:48:35 -0800564 set_pte_at(mm, address, pte, entry);
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100565 ret = 1;
566 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700567
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700568 pte_unmap_unlock(pte, ptl);
569out:
570 return ret;
571}
572
573static int page_mkclean_file(struct address_space *mapping, struct page *page)
574{
575 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
576 struct vm_area_struct *vma;
577 struct prio_tree_iter iter;
578 int ret = 0;
579
580 BUG_ON(PageAnon(page));
581
582 spin_lock(&mapping->i_mmap_lock);
583 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
Hugh Dickins1cb17292009-12-14 17:59:01 -0800584 if (vma->vm_flags & VM_SHARED) {
585 unsigned long address = vma_address(page, vma);
586 if (address == -EFAULT)
587 continue;
588 ret += page_mkclean_one(page, vma, address);
589 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700590 }
591 spin_unlock(&mapping->i_mmap_lock);
592 return ret;
593}
594
595int page_mkclean(struct page *page)
596{
597 int ret = 0;
598
599 BUG_ON(!PageLocked(page));
600
601 if (page_mapped(page)) {
602 struct address_space *mapping = page_mapping(page);
Christian Borntraegerce7e9fa2007-11-20 11:13:36 +0100603 if (mapping) {
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700604 ret = page_mkclean_file(mapping, page);
Christian Borntraegerce7e9fa2007-11-20 11:13:36 +0100605 if (page_test_dirty(page)) {
606 page_clear_dirty(page);
607 ret = 1;
608 }
Martin Schwidefsky6c210482007-04-27 16:01:57 +0200609 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700610 }
611
612 return ret;
613}
Jaya Kumar60b59be2007-05-08 00:37:37 -0700614EXPORT_SYMBOL_GPL(page_mkclean);
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700615
Linus Torvalds1da177e2005-04-16 15:20:36 -0700616/**
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700617 * __page_set_anon_rmap - setup new anonymous rmap
Nick Piggin9617d952006-01-06 00:11:12 -0800618 * @page: the page to add the mapping to
619 * @vma: the vm area in which the mapping is added
620 * @address: the user virtual address mapped
621 */
622static void __page_set_anon_rmap(struct page *page,
623 struct vm_area_struct *vma, unsigned long address)
624{
625 struct anon_vma *anon_vma = vma->anon_vma;
626
627 BUG_ON(!anon_vma);
628 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
629 page->mapping = (struct address_space *) anon_vma;
Nick Piggin9617d952006-01-06 00:11:12 -0800630 page->index = linear_page_index(vma, address);
Nick Piggin9617d952006-01-06 00:11:12 -0800631}
632
633/**
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700634 * __page_check_anon_rmap - sanity check anonymous rmap addition
Nick Pigginc97a9e12007-05-16 22:11:21 -0700635 * @page: the page to add the mapping to
636 * @vma: the vm area in which the mapping is added
637 * @address: the user virtual address mapped
638 */
639static void __page_check_anon_rmap(struct page *page,
640 struct vm_area_struct *vma, unsigned long address)
641{
642#ifdef CONFIG_DEBUG_VM
643 /*
644 * The page's anon-rmap details (mapping and index) are guaranteed to
645 * be set up correctly at this point.
646 *
647 * We have exclusion against page_add_anon_rmap because the caller
648 * always holds the page locked, except if called from page_dup_rmap,
649 * in which case the page is already known to be setup.
650 *
651 * We have exclusion against page_add_new_anon_rmap because those pages
652 * are initially only visible via the pagetables, and the pte is locked
653 * over the call to page_add_new_anon_rmap.
654 */
655 struct anon_vma *anon_vma = vma->anon_vma;
656 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
657 BUG_ON(page->mapping != (struct address_space *)anon_vma);
658 BUG_ON(page->index != linear_page_index(vma, address));
659#endif
660}
661
662/**
Linus Torvalds1da177e2005-04-16 15:20:36 -0700663 * page_add_anon_rmap - add pte mapping to an anonymous page
664 * @page: the page to add the mapping to
665 * @vma: the vm area in which the mapping is added
666 * @address: the user virtual address mapped
667 *
Hugh Dickins5ad64682009-12-14 17:59:24 -0800668 * The caller needs to hold the pte lock, and the page must be locked in
Hugh Dickins80e14822009-12-14 17:59:29 -0800669 * the anon_vma case: to serialize mapping,index checking after setting,
670 * and to ensure that PageAnon is not being upgraded racily to PageKsm
671 * (but PageKsm is never downgraded to PageAnon).
Linus Torvalds1da177e2005-04-16 15:20:36 -0700672 */
673void page_add_anon_rmap(struct page *page,
674 struct vm_area_struct *vma, unsigned long address)
675{
Hugh Dickins5ad64682009-12-14 17:59:24 -0800676 int first = atomic_inc_and_test(&page->_mapcount);
677 if (first)
678 __inc_zone_page_state(page, NR_ANON_PAGES);
679 if (unlikely(PageKsm(page)))
680 return;
681
Nick Pigginc97a9e12007-05-16 22:11:21 -0700682 VM_BUG_ON(!PageLocked(page));
683 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
Hugh Dickins5ad64682009-12-14 17:59:24 -0800684 if (first)
Nick Piggin9617d952006-01-06 00:11:12 -0800685 __page_set_anon_rmap(page, vma, address);
KAMEZAWA Hiroyuki69029cd2008-07-25 01:47:14 -0700686 else
Nick Pigginc97a9e12007-05-16 22:11:21 -0700687 __page_check_anon_rmap(page, vma, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700688}
689
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700690/**
Nick Piggin9617d952006-01-06 00:11:12 -0800691 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
692 * @page: the page to add the mapping to
693 * @vma: the vm area in which the mapping is added
694 * @address: the user virtual address mapped
695 *
696 * Same as page_add_anon_rmap but must only be called on *new* pages.
697 * This means the inc-and-test can be bypassed.
Nick Pigginc97a9e12007-05-16 22:11:21 -0700698 * Page does not have to be locked.
Nick Piggin9617d952006-01-06 00:11:12 -0800699 */
700void page_add_new_anon_rmap(struct page *page,
701 struct vm_area_struct *vma, unsigned long address)
702{
Hugh Dickinsb5934c52009-01-06 14:39:25 -0800703 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
Hugh Dickinscbf84b72009-01-06 14:39:27 -0800704 SetPageSwapBacked(page);
705 atomic_set(&page->_mapcount, 0); /* increment count (starts at -1) */
Hugh Dickins5ad64682009-12-14 17:59:24 -0800706 __inc_zone_page_state(page, NR_ANON_PAGES);
Nick Piggin9617d952006-01-06 00:11:12 -0800707 __page_set_anon_rmap(page, vma, address);
Hugh Dickinsb5934c52009-01-06 14:39:25 -0800708 if (page_evictable(page, vma))
Hugh Dickinscbf84b72009-01-06 14:39:27 -0800709 lru_cache_add_lru(page, LRU_ACTIVE_ANON);
Hugh Dickinsb5934c52009-01-06 14:39:25 -0800710 else
711 add_page_to_unevictable_list(page);
Nick Piggin9617d952006-01-06 00:11:12 -0800712}
713
Linus Torvalds1da177e2005-04-16 15:20:36 -0700714/**
715 * page_add_file_rmap - add pte mapping to a file page
716 * @page: the page to add the mapping to
717 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -0700718 * The caller needs to hold the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700719 */
720void page_add_file_rmap(struct page *page)
721{
Balbir Singhd69b0422009-06-17 16:26:34 -0700722 if (atomic_inc_and_test(&page->_mapcount)) {
Christoph Lameter65ba55f2006-06-30 01:55:34 -0700723 __inc_zone_page_state(page, NR_FILE_MAPPED);
KAMEZAWA Hiroyukid8046582009-12-15 16:47:09 -0800724 mem_cgroup_update_file_mapped(page, 1);
Balbir Singhd69b0422009-06-17 16:26:34 -0700725 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700726}
727
728/**
729 * page_remove_rmap - take down pte mapping from a page
730 * @page: page to remove mapping from
731 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -0700732 * The caller needs to hold the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700733 */
Hugh Dickinsedc315f2009-01-06 14:40:11 -0800734void page_remove_rmap(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700735{
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -0700736 /* page still mapped by someone else? */
737 if (!atomic_add_negative(-1, &page->_mapcount))
738 return;
739
740 /*
741 * Now that the last pte has gone, s390 must transfer dirty
742 * flag from storage key to struct page. We can usually skip
743 * this if the page is anon, so about to be freed; but perhaps
744 * not if it's in swapcache - there might be another pte slot
745 * containing the swap entry, but page not yet written to swap.
746 */
747 if ((!PageAnon(page) || PageSwapCache(page)) && page_test_dirty(page)) {
748 page_clear_dirty(page);
749 set_page_dirty(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700750 }
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -0700751 if (PageAnon(page)) {
752 mem_cgroup_uncharge_page(page);
753 __dec_zone_page_state(page, NR_ANON_PAGES);
754 } else {
755 __dec_zone_page_state(page, NR_FILE_MAPPED);
KAMEZAWA Hiroyukid8046582009-12-15 16:47:09 -0800756 mem_cgroup_update_file_mapped(page, -1);
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -0700757 }
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -0700758 /*
759 * It would be tidy to reset the PageAnon mapping here,
760 * but that might overwrite a racing page_add_anon_rmap
761 * which increments mapcount after us but sets mapping
762 * before us: so leave the reset to free_hot_cold_page,
763 * and remember that it's only reliable while mapped.
764 * Leaving it set also helps swapoff to reinstate ptes
765 * faster for those pages still in swapcache.
766 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700767}
768
769/*
770 * Subfunctions of try_to_unmap: try_to_unmap_one called
771 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
772 */
Hugh Dickins5ad64682009-12-14 17:59:24 -0800773int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
774 unsigned long address, enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700775{
776 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700777 pte_t *pte;
778 pte_t pteval;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700779 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700780 int ret = SWAP_AGAIN;
781
Nick Piggin479db0b2008-08-20 14:09:18 -0700782 pte = page_check_address(page, mm, address, &ptl, 0);
Hugh Dickinsc0718802005-10-29 18:16:31 -0700783 if (!pte)
Nikita Danilov81b40822005-05-01 08:58:36 -0700784 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700785
786 /*
787 * If the page is mlock()d, we cannot swap it out.
788 * If it's recently referenced (perhaps page_referenced
789 * skipped over this mm) then we should reactivate it.
790 */
Andi Kleen14fa31b2009-09-16 11:50:10 +0200791 if (!(flags & TTU_IGNORE_MLOCK)) {
KOSAKI Motohirocaed0f42009-12-14 17:59:45 -0800792 if (vma->vm_flags & VM_LOCKED)
793 goto out_mlock;
794
Hugh Dickinsaf8e3352009-12-14 17:58:59 -0800795 if (TTU_ACTION(flags) == TTU_MUNLOCK)
Hugh Dickins53f79ac2009-12-14 17:58:58 -0800796 goto out_unmap;
Andi Kleen14fa31b2009-09-16 11:50:10 +0200797 }
798 if (!(flags & TTU_IGNORE_ACCESS)) {
Nick Pigginb291f002008-10-18 20:26:44 -0700799 if (ptep_clear_flush_young_notify(vma, address, pte)) {
800 ret = SWAP_FAIL;
801 goto out_unmap;
802 }
803 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700804
Linus Torvalds1da177e2005-04-16 15:20:36 -0700805 /* Nuke the page table entry. */
806 flush_cache_page(vma, address, page_to_pfn(page));
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700807 pteval = ptep_clear_flush_notify(vma, address, pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700808
809 /* Move the dirty bit to the physical page now the pte is gone. */
810 if (pte_dirty(pteval))
811 set_page_dirty(page);
812
Hugh Dickins365e9c872005-10-29 18:16:18 -0700813 /* Update high watermark before we lower rss */
814 update_hiwater_rss(mm);
815
Andi Kleen888b9f72009-09-16 11:50:11 +0200816 if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
817 if (PageAnon(page))
818 dec_mm_counter(mm, anon_rss);
819 else
820 dec_mm_counter(mm, file_rss);
821 set_pte_at(mm, address, pte,
822 swp_entry_to_pte(make_hwpoison_entry(page)));
823 } else if (PageAnon(page)) {
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700824 swp_entry_t entry = { .val = page_private(page) };
Christoph Lameter06972122006-06-23 02:03:35 -0700825
826 if (PageSwapCache(page)) {
827 /*
828 * Store the swap location in the pte.
829 * See handle_pte_fault() ...
830 */
Hugh Dickins570a3352009-12-14 17:58:46 -0800831 if (swap_duplicate(entry) < 0) {
832 set_pte_at(mm, address, pte, pteval);
833 ret = SWAP_FAIL;
834 goto out_unmap;
835 }
Christoph Lameter06972122006-06-23 02:03:35 -0700836 if (list_empty(&mm->mmlist)) {
837 spin_lock(&mmlist_lock);
838 if (list_empty(&mm->mmlist))
839 list_add(&mm->mmlist, &init_mm.mmlist);
840 spin_unlock(&mmlist_lock);
841 }
Christoph Lameter442c9132006-06-23 02:03:38 -0700842 dec_mm_counter(mm, anon_rss);
KOSAKI Motohiro64cdd542009-01-06 14:39:16 -0800843 } else if (PAGE_MIGRATION) {
Christoph Lameter06972122006-06-23 02:03:35 -0700844 /*
845 * Store the pfn of the page in a special migration
846 * pte. do_swap_page() will wait until the migration
847 * pte is removed and then restart fault handling.
848 */
Andi Kleen14fa31b2009-09-16 11:50:10 +0200849 BUG_ON(TTU_ACTION(flags) != TTU_MIGRATION);
Christoph Lameter06972122006-06-23 02:03:35 -0700850 entry = make_migration_entry(page, pte_write(pteval));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700851 }
852 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
853 BUG_ON(pte_file(*pte));
Andi Kleen14fa31b2009-09-16 11:50:10 +0200854 } else if (PAGE_MIGRATION && (TTU_ACTION(flags) == TTU_MIGRATION)) {
Christoph Lameter04e62a22006-06-23 02:03:38 -0700855 /* Establish migration entry for a file page */
856 swp_entry_t entry;
857 entry = make_migration_entry(page, pte_write(pteval));
858 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
859 } else
Hugh Dickins42946212005-10-29 18:16:05 -0700860 dec_mm_counter(mm, file_rss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700861
Hugh Dickinsedc315f2009-01-06 14:40:11 -0800862 page_remove_rmap(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700863 page_cache_release(page);
864
865out_unmap:
Hugh Dickinsc0718802005-10-29 18:16:31 -0700866 pte_unmap_unlock(pte, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700867out:
868 return ret;
KOSAKI Motohirocaed0f42009-12-14 17:59:45 -0800869
870out_mlock:
871 pte_unmap_unlock(pte, ptl);
872
873
874 /*
875 * We need mmap_sem locking, Otherwise VM_LOCKED check makes
876 * unstable result and race. Plus, We can't wait here because
877 * we now hold anon_vma->lock or mapping->i_mmap_lock.
878 * if trylock failed, the page remain in evictable lru and later
879 * vmscan could retry to move the page to unevictable lru if the
880 * page is actually mlocked.
881 */
882 if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
883 if (vma->vm_flags & VM_LOCKED) {
884 mlock_vma_page(page);
885 ret = SWAP_MLOCK;
886 }
887 up_read(&vma->vm_mm->mmap_sem);
888 }
889 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700890}
891
892/*
893 * objrmap doesn't work for nonlinear VMAs because the assumption that
894 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
895 * Consequently, given a particular page and its ->index, we cannot locate the
896 * ptes which are mapping that page without an exhaustive linear search.
897 *
898 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
899 * maps the file to which the target page belongs. The ->vm_private_data field
900 * holds the current cursor into that scan. Successive searches will circulate
901 * around the vma's virtual address space.
902 *
903 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
904 * more scanning pressure is placed against them as well. Eventually pages
905 * will become fully unmapped and are eligible for eviction.
906 *
907 * For very sparsely populated VMAs this is a little inefficient - chances are
908 * there there won't be many ptes located within the scan cluster. In this case
909 * maybe we could scan further - to the end of the pte page, perhaps.
Nick Pigginb291f002008-10-18 20:26:44 -0700910 *
911 * Mlocked pages: check VM_LOCKED under mmap_sem held for read, if we can
912 * acquire it without blocking. If vma locked, mlock the pages in the cluster,
913 * rather than unmapping them. If we encounter the "check_page" that vmscan is
914 * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700915 */
916#define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
917#define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
918
Nick Pigginb291f002008-10-18 20:26:44 -0700919static int try_to_unmap_cluster(unsigned long cursor, unsigned int *mapcount,
920 struct vm_area_struct *vma, struct page *check_page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700921{
922 struct mm_struct *mm = vma->vm_mm;
923 pgd_t *pgd;
924 pud_t *pud;
925 pmd_t *pmd;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700926 pte_t *pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700927 pte_t pteval;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700928 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700929 struct page *page;
930 unsigned long address;
931 unsigned long end;
Nick Pigginb291f002008-10-18 20:26:44 -0700932 int ret = SWAP_AGAIN;
933 int locked_vma = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700934
Linus Torvalds1da177e2005-04-16 15:20:36 -0700935 address = (vma->vm_start + cursor) & CLUSTER_MASK;
936 end = address + CLUSTER_SIZE;
937 if (address < vma->vm_start)
938 address = vma->vm_start;
939 if (end > vma->vm_end)
940 end = vma->vm_end;
941
942 pgd = pgd_offset(mm, address);
943 if (!pgd_present(*pgd))
Nick Pigginb291f002008-10-18 20:26:44 -0700944 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700945
946 pud = pud_offset(pgd, address);
947 if (!pud_present(*pud))
Nick Pigginb291f002008-10-18 20:26:44 -0700948 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700949
950 pmd = pmd_offset(pud, address);
951 if (!pmd_present(*pmd))
Nick Pigginb291f002008-10-18 20:26:44 -0700952 return ret;
953
954 /*
Hugh Dickinsaf8e3352009-12-14 17:58:59 -0800955 * If we can acquire the mmap_sem for read, and vma is VM_LOCKED,
Nick Pigginb291f002008-10-18 20:26:44 -0700956 * keep the sem while scanning the cluster for mlocking pages.
957 */
Hugh Dickinsaf8e3352009-12-14 17:58:59 -0800958 if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
Nick Pigginb291f002008-10-18 20:26:44 -0700959 locked_vma = (vma->vm_flags & VM_LOCKED);
960 if (!locked_vma)
961 up_read(&vma->vm_mm->mmap_sem); /* don't need it */
962 }
Hugh Dickinsc0718802005-10-29 18:16:31 -0700963
964 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700965
Hugh Dickins365e9c872005-10-29 18:16:18 -0700966 /* Update high watermark before we lower rss */
967 update_hiwater_rss(mm);
968
Hugh Dickinsc0718802005-10-29 18:16:31 -0700969 for (; address < end; pte++, address += PAGE_SIZE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700970 if (!pte_present(*pte))
971 continue;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800972 page = vm_normal_page(vma, address, *pte);
973 BUG_ON(!page || PageAnon(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700974
Nick Pigginb291f002008-10-18 20:26:44 -0700975 if (locked_vma) {
976 mlock_vma_page(page); /* no-op if already mlocked */
977 if (page == check_page)
978 ret = SWAP_MLOCK;
979 continue; /* don't unmap */
980 }
981
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700982 if (ptep_clear_flush_young_notify(vma, address, pte))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700983 continue;
984
985 /* Nuke the page table entry. */
Ben Collinseca35132005-11-29 11:45:26 -0800986 flush_cache_page(vma, address, pte_pfn(*pte));
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700987 pteval = ptep_clear_flush_notify(vma, address, pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700988
989 /* If nonlinear, store the file page offset in the pte. */
990 if (page->index != linear_page_index(vma, address))
991 set_pte_at(mm, address, pte, pgoff_to_pte(page->index));
992
993 /* Move the dirty bit to the physical page now the pte is gone. */
994 if (pte_dirty(pteval))
995 set_page_dirty(page);
996
Hugh Dickinsedc315f2009-01-06 14:40:11 -0800997 page_remove_rmap(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700998 page_cache_release(page);
Hugh Dickins42946212005-10-29 18:16:05 -0700999 dec_mm_counter(mm, file_rss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001000 (*mapcount)--;
1001 }
Hugh Dickinsc0718802005-10-29 18:16:31 -07001002 pte_unmap_unlock(pte - 1, ptl);
Nick Pigginb291f002008-10-18 20:26:44 -07001003 if (locked_vma)
1004 up_read(&vma->vm_mm->mmap_sem);
1005 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001006}
1007
Nick Pigginb291f002008-10-18 20:26:44 -07001008/**
1009 * try_to_unmap_anon - unmap or unlock anonymous page using the object-based
1010 * rmap method
1011 * @page: the page to unmap/unlock
Huang Shijie8051be52009-12-14 17:58:50 -08001012 * @flags: action and flags
Nick Pigginb291f002008-10-18 20:26:44 -07001013 *
1014 * Find all the mappings of a page using the mapping pointer and the vma chains
1015 * contained in the anon_vma struct it points to.
1016 *
1017 * This function is only called from try_to_unmap/try_to_munlock for
1018 * anonymous pages.
1019 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1020 * where the page was found will be held for write. So, we won't recheck
1021 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1022 * 'LOCKED.
1023 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001024static int try_to_unmap_anon(struct page *page, enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001025{
1026 struct anon_vma *anon_vma;
1027 struct vm_area_struct *vma;
1028 int ret = SWAP_AGAIN;
Nick Pigginb291f002008-10-18 20:26:44 -07001029
Linus Torvalds1da177e2005-04-16 15:20:36 -07001030 anon_vma = page_lock_anon_vma(page);
1031 if (!anon_vma)
1032 return ret;
1033
1034 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
Hugh Dickins1cb17292009-12-14 17:59:01 -08001035 unsigned long address = vma_address(page, vma);
1036 if (address == -EFAULT)
1037 continue;
1038 ret = try_to_unmap_one(page, vma, address, flags);
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001039 if (ret != SWAP_AGAIN || !page_mapped(page))
1040 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001041 }
Oleg Nesterov34bbd702007-02-28 20:13:49 -08001042
1043 page_unlock_anon_vma(anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001044 return ret;
1045}
1046
1047/**
Nick Pigginb291f002008-10-18 20:26:44 -07001048 * try_to_unmap_file - unmap/unlock file page using the object-based rmap method
1049 * @page: the page to unmap/unlock
Andi Kleen14fa31b2009-09-16 11:50:10 +02001050 * @flags: action and flags
Linus Torvalds1da177e2005-04-16 15:20:36 -07001051 *
1052 * Find all the mappings of a page using the mapping pointer and the vma chains
1053 * contained in the address_space struct it points to.
1054 *
Nick Pigginb291f002008-10-18 20:26:44 -07001055 * This function is only called from try_to_unmap/try_to_munlock for
1056 * object-based pages.
1057 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1058 * where the page was found will be held for write. So, we won't recheck
1059 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1060 * 'LOCKED.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001061 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001062static int try_to_unmap_file(struct page *page, enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001063{
1064 struct address_space *mapping = page->mapping;
1065 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1066 struct vm_area_struct *vma;
1067 struct prio_tree_iter iter;
1068 int ret = SWAP_AGAIN;
1069 unsigned long cursor;
1070 unsigned long max_nl_cursor = 0;
1071 unsigned long max_nl_size = 0;
1072 unsigned int mapcount;
1073
1074 spin_lock(&mapping->i_mmap_lock);
1075 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
Hugh Dickins1cb17292009-12-14 17:59:01 -08001076 unsigned long address = vma_address(page, vma);
1077 if (address == -EFAULT)
1078 continue;
1079 ret = try_to_unmap_one(page, vma, address, flags);
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001080 if (ret != SWAP_AGAIN || !page_mapped(page))
1081 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001082 }
1083
1084 if (list_empty(&mapping->i_mmap_nonlinear))
1085 goto out;
1086
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001087 /*
1088 * We don't bother to try to find the munlocked page in nonlinears.
1089 * It's costly. Instead, later, page reclaim logic may call
1090 * try_to_unmap(TTU_MUNLOCK) and recover PG_mlocked lazily.
1091 */
1092 if (TTU_ACTION(flags) == TTU_MUNLOCK)
1093 goto out;
1094
Linus Torvalds1da177e2005-04-16 15:20:36 -07001095 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1096 shared.vm_set.list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001097 cursor = (unsigned long) vma->vm_private_data;
1098 if (cursor > max_nl_cursor)
1099 max_nl_cursor = cursor;
1100 cursor = vma->vm_end - vma->vm_start;
1101 if (cursor > max_nl_size)
1102 max_nl_size = cursor;
1103 }
1104
Nick Pigginb291f002008-10-18 20:26:44 -07001105 if (max_nl_size == 0) { /* all nonlinears locked or reserved ? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001106 ret = SWAP_FAIL;
1107 goto out;
1108 }
1109
1110 /*
1111 * We don't try to search for this page in the nonlinear vmas,
1112 * and page_referenced wouldn't have found it anyway. Instead
1113 * just walk the nonlinear vmas trying to age and unmap some.
1114 * The mapcount of the page we came in with is irrelevant,
1115 * but even so use it as a guide to how hard we should try?
1116 */
1117 mapcount = page_mapcount(page);
1118 if (!mapcount)
1119 goto out;
1120 cond_resched_lock(&mapping->i_mmap_lock);
1121
1122 max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
1123 if (max_nl_cursor == 0)
1124 max_nl_cursor = CLUSTER_SIZE;
1125
1126 do {
1127 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1128 shared.vm_set.list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001129 cursor = (unsigned long) vma->vm_private_data;
Hugh Dickins839b9682005-09-03 15:54:43 -07001130 while ( cursor < max_nl_cursor &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07001131 cursor < vma->vm_end - vma->vm_start) {
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001132 if (try_to_unmap_cluster(cursor, &mapcount,
1133 vma, page) == SWAP_MLOCK)
1134 ret = SWAP_MLOCK;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001135 cursor += CLUSTER_SIZE;
1136 vma->vm_private_data = (void *) cursor;
1137 if ((int)mapcount <= 0)
1138 goto out;
1139 }
1140 vma->vm_private_data = (void *) max_nl_cursor;
1141 }
1142 cond_resched_lock(&mapping->i_mmap_lock);
1143 max_nl_cursor += CLUSTER_SIZE;
1144 } while (max_nl_cursor <= max_nl_size);
1145
1146 /*
1147 * Don't loop forever (perhaps all the remaining pages are
1148 * in locked vmas). Reset cursor on all unreserved nonlinear
1149 * vmas, now forgetting on which ones it had fallen behind.
1150 */
Hugh Dickins101d2be2005-11-21 21:32:16 -08001151 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1152 vma->vm_private_data = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001153out:
1154 spin_unlock(&mapping->i_mmap_lock);
1155 return ret;
1156}
1157
1158/**
1159 * try_to_unmap - try to remove all page table mappings to a page
1160 * @page: the page to get unmapped
Andi Kleen14fa31b2009-09-16 11:50:10 +02001161 * @flags: action and flags
Linus Torvalds1da177e2005-04-16 15:20:36 -07001162 *
1163 * Tries to remove all the page table entries which are mapping this
1164 * page, used in the pageout path. Caller must hold the page lock.
1165 * Return values are:
1166 *
1167 * SWAP_SUCCESS - we succeeded in removing all mappings
1168 * SWAP_AGAIN - we missed a mapping, try again later
1169 * SWAP_FAIL - the page is unswappable
Nick Pigginb291f002008-10-18 20:26:44 -07001170 * SWAP_MLOCK - page is mlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001171 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001172int try_to_unmap(struct page *page, enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001173{
1174 int ret;
1175
Linus Torvalds1da177e2005-04-16 15:20:36 -07001176 BUG_ON(!PageLocked(page));
1177
Hugh Dickins5ad64682009-12-14 17:59:24 -08001178 if (unlikely(PageKsm(page)))
1179 ret = try_to_unmap_ksm(page, flags);
1180 else if (PageAnon(page))
Andi Kleen14fa31b2009-09-16 11:50:10 +02001181 ret = try_to_unmap_anon(page, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001182 else
Andi Kleen14fa31b2009-09-16 11:50:10 +02001183 ret = try_to_unmap_file(page, flags);
Nick Pigginb291f002008-10-18 20:26:44 -07001184 if (ret != SWAP_MLOCK && !page_mapped(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001185 ret = SWAP_SUCCESS;
1186 return ret;
1187}
Nikita Danilov81b40822005-05-01 08:58:36 -07001188
Nick Pigginb291f002008-10-18 20:26:44 -07001189/**
1190 * try_to_munlock - try to munlock a page
1191 * @page: the page to be munlocked
1192 *
1193 * Called from munlock code. Checks all of the VMAs mapping the page
1194 * to make sure nobody else has this page mlocked. The page will be
1195 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1196 *
1197 * Return values are:
1198 *
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001199 * SWAP_AGAIN - no vma is holding page mlocked, or,
Nick Pigginb291f002008-10-18 20:26:44 -07001200 * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem
Hugh Dickins5ad64682009-12-14 17:59:24 -08001201 * SWAP_FAIL - page cannot be located at present
Nick Pigginb291f002008-10-18 20:26:44 -07001202 * SWAP_MLOCK - page is now mlocked.
1203 */
1204int try_to_munlock(struct page *page)
1205{
1206 VM_BUG_ON(!PageLocked(page) || PageLRU(page));
1207
Hugh Dickins5ad64682009-12-14 17:59:24 -08001208 if (unlikely(PageKsm(page)))
1209 return try_to_unmap_ksm(page, TTU_MUNLOCK);
1210 else if (PageAnon(page))
Andi Kleen14fa31b2009-09-16 11:50:10 +02001211 return try_to_unmap_anon(page, TTU_MUNLOCK);
Nick Pigginb291f002008-10-18 20:26:44 -07001212 else
Andi Kleen14fa31b2009-09-16 11:50:10 +02001213 return try_to_unmap_file(page, TTU_MUNLOCK);
Nick Pigginb291f002008-10-18 20:26:44 -07001214}
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001215
1216#ifdef CONFIG_MIGRATION
1217/*
1218 * rmap_walk() and its helpers rmap_walk_anon() and rmap_walk_file():
1219 * Called by migrate.c to remove migration ptes, but might be used more later.
1220 */
1221static int rmap_walk_anon(struct page *page, int (*rmap_one)(struct page *,
1222 struct vm_area_struct *, unsigned long, void *), void *arg)
1223{
1224 struct anon_vma *anon_vma;
1225 struct vm_area_struct *vma;
1226 int ret = SWAP_AGAIN;
1227
1228 /*
1229 * Note: remove_migration_ptes() cannot use page_lock_anon_vma()
1230 * because that depends on page_mapped(); but not all its usages
1231 * are holding mmap_sem, which also gave the necessary guarantee
1232 * (that this anon_vma's slab has not already been destroyed).
1233 * This needs to be reviewed later: avoiding page_lock_anon_vma()
1234 * is risky, and currently limits the usefulness of rmap_walk().
1235 */
1236 anon_vma = page_anon_vma(page);
1237 if (!anon_vma)
1238 return ret;
1239 spin_lock(&anon_vma->lock);
1240 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
1241 unsigned long address = vma_address(page, vma);
1242 if (address == -EFAULT)
1243 continue;
1244 ret = rmap_one(page, vma, address, arg);
1245 if (ret != SWAP_AGAIN)
1246 break;
1247 }
1248 spin_unlock(&anon_vma->lock);
1249 return ret;
1250}
1251
1252static int rmap_walk_file(struct page *page, int (*rmap_one)(struct page *,
1253 struct vm_area_struct *, unsigned long, void *), void *arg)
1254{
1255 struct address_space *mapping = page->mapping;
1256 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1257 struct vm_area_struct *vma;
1258 struct prio_tree_iter iter;
1259 int ret = SWAP_AGAIN;
1260
1261 if (!mapping)
1262 return ret;
1263 spin_lock(&mapping->i_mmap_lock);
1264 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1265 unsigned long address = vma_address(page, vma);
1266 if (address == -EFAULT)
1267 continue;
1268 ret = rmap_one(page, vma, address, arg);
1269 if (ret != SWAP_AGAIN)
1270 break;
1271 }
1272 /*
1273 * No nonlinear handling: being always shared, nonlinear vmas
1274 * never contain migration ptes. Decide what to do about this
1275 * limitation to linear when we need rmap_walk() on nonlinear.
1276 */
1277 spin_unlock(&mapping->i_mmap_lock);
1278 return ret;
1279}
1280
1281int rmap_walk(struct page *page, int (*rmap_one)(struct page *,
1282 struct vm_area_struct *, unsigned long, void *), void *arg)
1283{
1284 VM_BUG_ON(!PageLocked(page));
1285
1286 if (unlikely(PageKsm(page)))
1287 return rmap_walk_ksm(page, rmap_one, arg);
1288 else if (PageAnon(page))
1289 return rmap_walk_anon(page, rmap_one, arg);
1290 else
1291 return rmap_walk_file(page, rmap_one, arg);
1292}
1293#endif /* CONFIG_MIGRATION */