blob: 7e72ca19d68b10cd98bd90be68699f4dbe899fd1 [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)
Linus Torvalds1da177e2005-04-16 15:20:36 -070039 */
40
41#include <linux/mm.h>
42#include <linux/pagemap.h>
43#include <linux/swap.h>
44#include <linux/swapops.h>
45#include <linux/slab.h>
46#include <linux/init.h>
47#include <linux/rmap.h>
48#include <linux/rcupdate.h>
Christoph Lametera48d07a2006-02-01 03:05:38 -080049#include <linux/module.h>
Balbir Singh8a9f3cc2008-02-07 00:13:53 -080050#include <linux/memcontrol.h>
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070051#include <linux/mmu_notifier.h>
KOSAKI Motohiro64cdd542009-01-06 14:39:16 -080052#include <linux/migrate.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070053
54#include <asm/tlbflush.h>
55
Nick Pigginb291f002008-10-18 20:26:44 -070056#include "internal.h"
57
Adrian Bunkfdd2e5f2008-10-18 20:28:38 -070058static struct kmem_cache *anon_vma_cachep;
59
60static inline struct anon_vma *anon_vma_alloc(void)
61{
62 return kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
63}
64
65static inline void anon_vma_free(struct anon_vma *anon_vma)
66{
67 kmem_cache_free(anon_vma_cachep, anon_vma);
68}
Linus Torvalds1da177e2005-04-16 15:20:36 -070069
Linus Torvaldsd9d332e2008-10-19 10:32:20 -070070/**
71 * anon_vma_prepare - attach an anon_vma to a memory region
72 * @vma: the memory region in question
73 *
74 * This makes sure the memory mapping described by 'vma' has
75 * an 'anon_vma' attached to it, so that we can associate the
76 * anonymous pages mapped into it with that anon_vma.
77 *
78 * The common case will be that we already have one, but if
79 * if not we either need to find an adjacent mapping that we
80 * can re-use the anon_vma from (very common when the only
81 * reason for splitting a vma has been mprotect()), or we
82 * allocate a new one.
83 *
84 * Anon-vma allocations are very subtle, because we may have
85 * optimistically looked up an anon_vma in page_lock_anon_vma()
86 * and that may actually touch the spinlock even in the newly
87 * allocated vma (it depends on RCU to make sure that the
88 * anon_vma isn't actually destroyed).
89 *
90 * As a result, we need to do proper anon_vma locking even
91 * for the new allocation. At the same time, we do not want
92 * to do any locking for the common case of already having
93 * an anon_vma.
94 *
95 * This must be called with the mmap_sem held for reading.
96 */
Linus Torvalds1da177e2005-04-16 15:20:36 -070097int anon_vma_prepare(struct vm_area_struct *vma)
98{
99 struct anon_vma *anon_vma = vma->anon_vma;
100
101 might_sleep();
102 if (unlikely(!anon_vma)) {
103 struct mm_struct *mm = vma->vm_mm;
Linus Torvaldsd9d332e2008-10-19 10:32:20 -0700104 struct anon_vma *allocated;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700105
106 anon_vma = find_mergeable_anon_vma(vma);
Linus Torvaldsd9d332e2008-10-19 10:32:20 -0700107 allocated = NULL;
108 if (!anon_vma) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700109 anon_vma = anon_vma_alloc();
110 if (unlikely(!anon_vma))
111 return -ENOMEM;
112 allocated = anon_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700113 }
Linus Torvaldsd9d332e2008-10-19 10:32:20 -0700114 spin_lock(&anon_vma->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700115
116 /* page_table_lock to protect against threads */
117 spin_lock(&mm->page_table_lock);
118 if (likely(!vma->anon_vma)) {
119 vma->anon_vma = anon_vma;
Christoph Lameter06972122006-06-23 02:03:35 -0700120 list_add_tail(&vma->anon_vma_node, &anon_vma->head);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700121 allocated = NULL;
122 }
123 spin_unlock(&mm->page_table_lock);
124
Linus Torvaldsd9d332e2008-10-19 10:32:20 -0700125 spin_unlock(&anon_vma->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700126 if (unlikely(allocated))
127 anon_vma_free(allocated);
128 }
129 return 0;
130}
131
132void __anon_vma_merge(struct vm_area_struct *vma, struct vm_area_struct *next)
133{
134 BUG_ON(vma->anon_vma != next->anon_vma);
135 list_del(&next->anon_vma_node);
136}
137
138void __anon_vma_link(struct vm_area_struct *vma)
139{
140 struct anon_vma *anon_vma = vma->anon_vma;
141
Hugh Dickins30acbab2007-06-27 14:09:53 -0700142 if (anon_vma)
Christoph Lameter06972122006-06-23 02:03:35 -0700143 list_add_tail(&vma->anon_vma_node, &anon_vma->head);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700144}
145
146void anon_vma_link(struct vm_area_struct *vma)
147{
148 struct anon_vma *anon_vma = vma->anon_vma;
149
150 if (anon_vma) {
151 spin_lock(&anon_vma->lock);
Christoph Lameter06972122006-06-23 02:03:35 -0700152 list_add_tail(&vma->anon_vma_node, &anon_vma->head);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700153 spin_unlock(&anon_vma->lock);
154 }
155}
156
157void anon_vma_unlink(struct vm_area_struct *vma)
158{
159 struct anon_vma *anon_vma = vma->anon_vma;
160 int empty;
161
162 if (!anon_vma)
163 return;
164
165 spin_lock(&anon_vma->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700166 list_del(&vma->anon_vma_node);
167
168 /* We must garbage collect the anon_vma if it's empty */
169 empty = list_empty(&anon_vma->head);
170 spin_unlock(&anon_vma->lock);
171
172 if (empty)
173 anon_vma_free(anon_vma);
174}
175
Alexey Dobriyan51cc5062008-07-25 19:45:34 -0700176static void anon_vma_ctor(void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700177{
Christoph Lametera35afb82007-05-16 22:10:57 -0700178 struct anon_vma *anon_vma = data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700179
Christoph Lametera35afb82007-05-16 22:10:57 -0700180 spin_lock_init(&anon_vma->lock);
181 INIT_LIST_HEAD(&anon_vma->head);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700182}
183
184void __init anon_vma_init(void)
185{
186 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
Paul Mundt20c2df82007-07-20 10:11:58 +0900187 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700188}
189
190/*
191 * Getting a lock on a stable anon_vma from a page off the LRU is
192 * tricky: page_lock_anon_vma rely on RCU to guard against the races.
193 */
Andi Kleen10be22d2009-09-16 11:50:04 +0200194struct anon_vma *page_lock_anon_vma(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700195{
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800196 struct anon_vma *anon_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700197 unsigned long anon_mapping;
198
199 rcu_read_lock();
200 anon_mapping = (unsigned long) page->mapping;
201 if (!(anon_mapping & PAGE_MAPPING_ANON))
202 goto out;
203 if (!page_mapped(page))
204 goto out;
205
206 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
207 spin_lock(&anon_vma->lock);
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800208 return anon_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700209out:
210 rcu_read_unlock();
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800211 return NULL;
212}
213
Andi Kleen10be22d2009-09-16 11:50:04 +0200214void page_unlock_anon_vma(struct anon_vma *anon_vma)
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800215{
216 spin_unlock(&anon_vma->lock);
217 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700218}
219
220/*
Lee Schermerhorn3ad33b242007-11-14 16:59:10 -0800221 * At what user virtual address is page expected in @vma?
222 * Returns virtual address or -EFAULT if page's index/offset is not
223 * within the range mapped the @vma.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700224 */
225static inline unsigned long
226vma_address(struct page *page, struct vm_area_struct *vma)
227{
228 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
229 unsigned long address;
230
231 address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
232 if (unlikely(address < vma->vm_start || address >= vma->vm_end)) {
Lee Schermerhorn3ad33b242007-11-14 16:59:10 -0800233 /* page should be within @vma mapping range */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700234 return -EFAULT;
235 }
236 return address;
237}
238
239/*
240 * At what user virtual address is page expected in vma? checking that the
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800241 * page matches the vma: currently only used on anon pages, by unuse_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700242 */
243unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
244{
245 if (PageAnon(page)) {
246 if ((void *)vma->anon_vma !=
247 (void *)page->mapping - PAGE_MAPPING_ANON)
248 return -EFAULT;
249 } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) {
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800250 if (!vma->vm_file ||
251 vma->vm_file->f_mapping != page->mapping)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700252 return -EFAULT;
253 } else
254 return -EFAULT;
255 return vma_address(page, vma);
256}
257
258/*
Nikita Danilov81b40822005-05-01 08:58:36 -0700259 * Check that @page is mapped at @address into @mm.
260 *
Nick Piggin479db0b2008-08-20 14:09:18 -0700261 * If @sync is false, page_check_address may perform a racy check to avoid
262 * the page table lock when the pte is not present (helpful when reclaiming
263 * highly shared pages).
264 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -0700265 * On success returns with pte mapped and locked.
Nikita Danilov81b40822005-05-01 08:58:36 -0700266 */
Carsten Otteceffc072005-06-23 22:05:25 -0700267pte_t *page_check_address(struct page *page, struct mm_struct *mm,
Nick Piggin479db0b2008-08-20 14:09:18 -0700268 unsigned long address, spinlock_t **ptlp, int sync)
Nikita Danilov81b40822005-05-01 08:58:36 -0700269{
270 pgd_t *pgd;
271 pud_t *pud;
272 pmd_t *pmd;
273 pte_t *pte;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700274 spinlock_t *ptl;
Nikita Danilov81b40822005-05-01 08:58:36 -0700275
Nikita Danilov81b40822005-05-01 08:58:36 -0700276 pgd = pgd_offset(mm, address);
Hugh Dickinsc0718802005-10-29 18:16:31 -0700277 if (!pgd_present(*pgd))
278 return NULL;
279
280 pud = pud_offset(pgd, address);
281 if (!pud_present(*pud))
282 return NULL;
283
284 pmd = pmd_offset(pud, address);
285 if (!pmd_present(*pmd))
286 return NULL;
287
288 pte = pte_offset_map(pmd, address);
289 /* Make a quick check before getting the lock */
Nick Piggin479db0b2008-08-20 14:09:18 -0700290 if (!sync && !pte_present(*pte)) {
Hugh Dickinsc0718802005-10-29 18:16:31 -0700291 pte_unmap(pte);
292 return NULL;
Nikita Danilov81b40822005-05-01 08:58:36 -0700293 }
Hugh Dickinsc0718802005-10-29 18:16:31 -0700294
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700295 ptl = pte_lockptr(mm, pmd);
Hugh Dickinsc0718802005-10-29 18:16:31 -0700296 spin_lock(ptl);
297 if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
298 *ptlp = ptl;
299 return pte;
300 }
301 pte_unmap_unlock(pte, ptl);
302 return NULL;
Nikita Danilov81b40822005-05-01 08:58:36 -0700303}
304
Nick Pigginb291f002008-10-18 20:26:44 -0700305/**
306 * page_mapped_in_vma - check whether a page is really mapped in a VMA
307 * @page: the page to test
308 * @vma: the VMA to test
309 *
310 * Returns 1 if the page is mapped into the page tables of the VMA, 0
311 * if the page is not mapped into the page tables of this VMA. Only
312 * valid for normal file or anonymous VMAs.
313 */
314static int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
315{
316 unsigned long address;
317 pte_t *pte;
318 spinlock_t *ptl;
319
320 address = vma_address(page, vma);
321 if (address == -EFAULT) /* out of vma range */
322 return 0;
323 pte = page_check_address(page, vma->vm_mm, address, &ptl, 1);
324 if (!pte) /* the page is not in this mm */
325 return 0;
326 pte_unmap_unlock(pte, ptl);
327
328 return 1;
329}
330
Nikita Danilov81b40822005-05-01 08:58:36 -0700331/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700332 * Subfunctions of page_referenced: page_referenced_one called
333 * repeatedly from either page_referenced_anon or page_referenced_file.
334 */
335static int page_referenced_one(struct page *page,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700336 struct vm_area_struct *vma,
337 unsigned int *mapcount,
338 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700339{
340 struct mm_struct *mm = vma->vm_mm;
341 unsigned long address;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700342 pte_t *pte;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700343 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700344 int referenced = 0;
345
Linus Torvalds1da177e2005-04-16 15:20:36 -0700346 address = vma_address(page, vma);
347 if (address == -EFAULT)
348 goto out;
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);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700386out:
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700387 if (referenced)
388 *vm_flags |= vma->vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700389 return referenced;
390}
391
Balbir Singhbed71612008-02-07 00:14:01 -0800392static int page_referenced_anon(struct page *page,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700393 struct mem_cgroup *mem_cont,
394 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700395{
396 unsigned int mapcount;
397 struct anon_vma *anon_vma;
398 struct vm_area_struct *vma;
399 int referenced = 0;
400
401 anon_vma = page_lock_anon_vma(page);
402 if (!anon_vma)
403 return referenced;
404
405 mapcount = page_mapcount(page);
406 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
Balbir Singhbed71612008-02-07 00:14:01 -0800407 /*
408 * If we are reclaiming on behalf of a cgroup, skip
409 * counting on behalf of references from different
410 * cgroups
411 */
Hugh Dickinsbd845e32008-03-04 14:29:01 -0800412 if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
Balbir Singhbed71612008-02-07 00:14:01 -0800413 continue;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700414 referenced += page_referenced_one(page, vma,
415 &mapcount, vm_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700416 if (!mapcount)
417 break;
418 }
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800419
420 page_unlock_anon_vma(anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700421 return referenced;
422}
423
424/**
425 * page_referenced_file - referenced check for object-based rmap
426 * @page: the page we're checking references on.
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700427 * @mem_cont: target memory controller
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700428 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
Linus Torvalds1da177e2005-04-16 15:20:36 -0700429 *
430 * For an object-based mapped page, find all the places it is mapped and
431 * check/clear the referenced flag. This is done by following the page->mapping
432 * pointer, then walking the chain of vmas it holds. It returns the number
433 * of references it found.
434 *
435 * This function is only called from page_referenced for object-based pages.
436 */
Balbir Singhbed71612008-02-07 00:14:01 -0800437static int page_referenced_file(struct page *page,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700438 struct mem_cgroup *mem_cont,
439 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700440{
441 unsigned int mapcount;
442 struct address_space *mapping = page->mapping;
443 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
444 struct vm_area_struct *vma;
445 struct prio_tree_iter iter;
446 int referenced = 0;
447
448 /*
449 * The caller's checks on page->mapping and !PageAnon have made
450 * sure that this is a file page: the check for page->mapping
451 * excludes the case just before it gets set on an anon page.
452 */
453 BUG_ON(PageAnon(page));
454
455 /*
456 * The page lock not only makes sure that page->mapping cannot
457 * suddenly be NULLified by truncation, it makes sure that the
458 * structure at mapping cannot be freed and reused yet,
459 * so we can safely take mapping->i_mmap_lock.
460 */
461 BUG_ON(!PageLocked(page));
462
463 spin_lock(&mapping->i_mmap_lock);
464
465 /*
466 * i_mmap_lock does not stabilize mapcount at all, but mapcount
467 * is more likely to be accurate if we note it after spinning.
468 */
469 mapcount = page_mapcount(page);
470
471 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
Balbir Singhbed71612008-02-07 00:14:01 -0800472 /*
473 * If we are reclaiming on behalf of a cgroup, skip
474 * counting on behalf of references from different
475 * cgroups
476 */
Hugh Dickinsbd845e32008-03-04 14:29:01 -0800477 if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
Balbir Singhbed71612008-02-07 00:14:01 -0800478 continue;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700479 referenced += page_referenced_one(page, vma,
480 &mapcount, vm_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700481 if (!mapcount)
482 break;
483 }
484
485 spin_unlock(&mapping->i_mmap_lock);
486 return referenced;
487}
488
489/**
490 * page_referenced - test if the page was referenced
491 * @page: the page to test
492 * @is_locked: caller holds lock on the page
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700493 * @mem_cont: target memory controller
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700494 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
Linus Torvalds1da177e2005-04-16 15:20:36 -0700495 *
496 * Quick test_and_clear_referenced for all mappings to a page,
497 * returns the number of ptes which referenced the page.
498 */
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700499int page_referenced(struct page *page,
500 int is_locked,
501 struct mem_cgroup *mem_cont,
502 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700503{
504 int referenced = 0;
505
Linus Torvalds1da177e2005-04-16 15:20:36 -0700506 if (TestClearPageReferenced(page))
507 referenced++;
508
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700509 *vm_flags = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700510 if (page_mapped(page) && page->mapping) {
511 if (PageAnon(page))
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700512 referenced += page_referenced_anon(page, mem_cont,
513 vm_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700514 else if (is_locked)
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700515 referenced += page_referenced_file(page, mem_cont,
516 vm_flags);
Nick Piggin529ae9a2008-08-02 12:01:03 +0200517 else if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700518 referenced++;
519 else {
520 if (page->mapping)
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700521 referenced += page_referenced_file(page,
522 mem_cont, vm_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700523 unlock_page(page);
524 }
525 }
Christian Borntraeger5b7baf02008-03-25 18:47:12 +0100526
527 if (page_test_and_clear_young(page))
528 referenced++;
529
Linus Torvalds1da177e2005-04-16 15:20:36 -0700530 return referenced;
531}
532
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700533static int page_mkclean_one(struct page *page, struct vm_area_struct *vma)
534{
535 struct mm_struct *mm = vma->vm_mm;
536 unsigned long address;
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100537 pte_t *pte;
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700538 spinlock_t *ptl;
539 int ret = 0;
540
541 address = vma_address(page, vma);
542 if (address == -EFAULT)
543 goto out;
544
Nick Piggin479db0b2008-08-20 14:09:18 -0700545 pte = page_check_address(page, mm, address, &ptl, 1);
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700546 if (!pte)
547 goto out;
548
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100549 if (pte_dirty(*pte) || pte_write(*pte)) {
550 pte_t entry;
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700551
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100552 flush_cache_page(vma, address, pte_pfn(*pte));
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700553 entry = ptep_clear_flush_notify(vma, address, pte);
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100554 entry = pte_wrprotect(entry);
555 entry = pte_mkclean(entry);
Al Virod6e88e62006-12-29 16:48:35 -0800556 set_pte_at(mm, address, pte, entry);
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100557 ret = 1;
558 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700559
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700560 pte_unmap_unlock(pte, ptl);
561out:
562 return ret;
563}
564
565static int page_mkclean_file(struct address_space *mapping, struct page *page)
566{
567 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
568 struct vm_area_struct *vma;
569 struct prio_tree_iter iter;
570 int ret = 0;
571
572 BUG_ON(PageAnon(page));
573
574 spin_lock(&mapping->i_mmap_lock);
575 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
576 if (vma->vm_flags & VM_SHARED)
577 ret += page_mkclean_one(page, vma);
578 }
579 spin_unlock(&mapping->i_mmap_lock);
580 return ret;
581}
582
583int page_mkclean(struct page *page)
584{
585 int ret = 0;
586
587 BUG_ON(!PageLocked(page));
588
589 if (page_mapped(page)) {
590 struct address_space *mapping = page_mapping(page);
Christian Borntraegerce7e9fa2007-11-20 11:13:36 +0100591 if (mapping) {
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700592 ret = page_mkclean_file(mapping, page);
Christian Borntraegerce7e9fa2007-11-20 11:13:36 +0100593 if (page_test_dirty(page)) {
594 page_clear_dirty(page);
595 ret = 1;
596 }
Martin Schwidefsky6c210482007-04-27 16:01:57 +0200597 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700598 }
599
600 return ret;
601}
Jaya Kumar60b59be2007-05-08 00:37:37 -0700602EXPORT_SYMBOL_GPL(page_mkclean);
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700603
Linus Torvalds1da177e2005-04-16 15:20:36 -0700604/**
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700605 * __page_set_anon_rmap - setup new anonymous rmap
Nick Piggin9617d952006-01-06 00:11:12 -0800606 * @page: the page to add the mapping to
607 * @vma: the vm area in which the mapping is added
608 * @address: the user virtual address mapped
609 */
610static void __page_set_anon_rmap(struct page *page,
611 struct vm_area_struct *vma, unsigned long address)
612{
613 struct anon_vma *anon_vma = vma->anon_vma;
614
615 BUG_ON(!anon_vma);
616 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
617 page->mapping = (struct address_space *) anon_vma;
618
619 page->index = linear_page_index(vma, address);
620
Nick Piggina74609f2006-01-06 00:11:20 -0800621 /*
622 * nr_mapped state can be updated without turning off
623 * interrupts because it is not modified via interrupt.
624 */
Christoph Lameterf3dbd342006-06-30 01:55:36 -0700625 __inc_zone_page_state(page, NR_ANON_PAGES);
Nick Piggin9617d952006-01-06 00:11:12 -0800626}
627
628/**
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700629 * __page_check_anon_rmap - sanity check anonymous rmap addition
Nick Pigginc97a9e12007-05-16 22:11:21 -0700630 * @page: the page to add the mapping to
631 * @vma: the vm area in which the mapping is added
632 * @address: the user virtual address mapped
633 */
634static void __page_check_anon_rmap(struct page *page,
635 struct vm_area_struct *vma, unsigned long address)
636{
637#ifdef CONFIG_DEBUG_VM
638 /*
639 * The page's anon-rmap details (mapping and index) are guaranteed to
640 * be set up correctly at this point.
641 *
642 * We have exclusion against page_add_anon_rmap because the caller
643 * always holds the page locked, except if called from page_dup_rmap,
644 * in which case the page is already known to be setup.
645 *
646 * We have exclusion against page_add_new_anon_rmap because those pages
647 * are initially only visible via the pagetables, and the pte is locked
648 * over the call to page_add_new_anon_rmap.
649 */
650 struct anon_vma *anon_vma = vma->anon_vma;
651 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
652 BUG_ON(page->mapping != (struct address_space *)anon_vma);
653 BUG_ON(page->index != linear_page_index(vma, address));
654#endif
655}
656
657/**
Linus Torvalds1da177e2005-04-16 15:20:36 -0700658 * page_add_anon_rmap - add pte mapping to an anonymous page
659 * @page: the page to add the mapping to
660 * @vma: the vm area in which the mapping is added
661 * @address: the user virtual address mapped
662 *
Nick Pigginc97a9e12007-05-16 22:11:21 -0700663 * The caller needs to hold the pte lock and the page must be locked.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700664 */
665void page_add_anon_rmap(struct page *page,
666 struct vm_area_struct *vma, unsigned long address)
667{
Nick Pigginc97a9e12007-05-16 22:11:21 -0700668 VM_BUG_ON(!PageLocked(page));
669 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
Nick Piggin9617d952006-01-06 00:11:12 -0800670 if (atomic_inc_and_test(&page->_mapcount))
671 __page_set_anon_rmap(page, vma, address);
KAMEZAWA Hiroyuki69029cd2008-07-25 01:47:14 -0700672 else
Nick Pigginc97a9e12007-05-16 22:11:21 -0700673 __page_check_anon_rmap(page, vma, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700674}
675
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700676/**
Nick Piggin9617d952006-01-06 00:11:12 -0800677 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
678 * @page: the page to add the mapping to
679 * @vma: the vm area in which the mapping is added
680 * @address: the user virtual address mapped
681 *
682 * Same as page_add_anon_rmap but must only be called on *new* pages.
683 * This means the inc-and-test can be bypassed.
Nick Pigginc97a9e12007-05-16 22:11:21 -0700684 * Page does not have to be locked.
Nick Piggin9617d952006-01-06 00:11:12 -0800685 */
686void page_add_new_anon_rmap(struct page *page,
687 struct vm_area_struct *vma, unsigned long address)
688{
Hugh Dickinsb5934c52009-01-06 14:39:25 -0800689 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
Hugh Dickinscbf84b72009-01-06 14:39:27 -0800690 SetPageSwapBacked(page);
691 atomic_set(&page->_mapcount, 0); /* increment count (starts at -1) */
Nick Piggin9617d952006-01-06 00:11:12 -0800692 __page_set_anon_rmap(page, vma, address);
Hugh Dickinsb5934c52009-01-06 14:39:25 -0800693 if (page_evictable(page, vma))
Hugh Dickinscbf84b72009-01-06 14:39:27 -0800694 lru_cache_add_lru(page, LRU_ACTIVE_ANON);
Hugh Dickinsb5934c52009-01-06 14:39:25 -0800695 else
696 add_page_to_unevictable_list(page);
Nick Piggin9617d952006-01-06 00:11:12 -0800697}
698
Linus Torvalds1da177e2005-04-16 15:20:36 -0700699/**
700 * page_add_file_rmap - add pte mapping to a file page
701 * @page: the page to add the mapping to
702 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -0700703 * The caller needs to hold the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700704 */
705void page_add_file_rmap(struct page *page)
706{
Balbir Singhd69b0422009-06-17 16:26:34 -0700707 if (atomic_inc_and_test(&page->_mapcount)) {
Christoph Lameter65ba55f2006-06-30 01:55:34 -0700708 __inc_zone_page_state(page, NR_FILE_MAPPED);
Balbir Singhd69b0422009-06-17 16:26:34 -0700709 mem_cgroup_update_mapped_file_stat(page, 1);
710 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700711}
712
Nick Pigginc97a9e12007-05-16 22:11:21 -0700713#ifdef CONFIG_DEBUG_VM
714/**
715 * page_dup_rmap - duplicate pte mapping to a page
716 * @page: the page to add the mapping to
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700717 * @vma: the vm area being duplicated
718 * @address: the user virtual address mapped
Nick Pigginc97a9e12007-05-16 22:11:21 -0700719 *
720 * For copy_page_range only: minimal extract from page_add_file_rmap /
721 * page_add_anon_rmap, avoiding unnecessary tests (already checked) so it's
722 * quicker.
723 *
724 * The caller needs to hold the pte lock.
725 */
726void page_dup_rmap(struct page *page, struct vm_area_struct *vma, unsigned long address)
727{
Nick Pigginc97a9e12007-05-16 22:11:21 -0700728 if (PageAnon(page))
729 __page_check_anon_rmap(page, vma, address);
730 atomic_inc(&page->_mapcount);
731}
732#endif
733
Linus Torvalds1da177e2005-04-16 15:20:36 -0700734/**
735 * page_remove_rmap - take down pte mapping from a page
736 * @page: page to remove mapping from
737 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -0700738 * The caller needs to hold the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700739 */
Hugh Dickinsedc315f2009-01-06 14:40:11 -0800740void page_remove_rmap(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700741{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700742 if (atomic_add_negative(-1, &page->_mapcount)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700743 /*
Hugh Dickins16f8c5b2008-08-20 14:09:04 -0700744 * Now that the last pte has gone, s390 must transfer dirty
745 * flag from storage key to struct page. We can usually skip
746 * this if the page is anon, so about to be freed; but perhaps
747 * not if it's in swapcache - there might be another pte slot
748 * containing the swap entry, but page not yet written to swap.
749 */
750 if ((!PageAnon(page) || PageSwapCache(page)) &&
751 page_test_dirty(page)) {
752 page_clear_dirty(page);
753 set_page_dirty(page);
754 }
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -0700755 if (PageAnon(page))
756 mem_cgroup_uncharge_page(page);
Hugh Dickins16f8c5b2008-08-20 14:09:04 -0700757 __dec_zone_page_state(page,
758 PageAnon(page) ? NR_ANON_PAGES : NR_FILE_MAPPED);
Balbir Singhd69b0422009-06-17 16:26:34 -0700759 mem_cgroup_update_mapped_file_stat(page, -1);
Hugh Dickins16f8c5b2008-08-20 14:09:04 -0700760 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700761 * It would be tidy to reset the PageAnon mapping here,
762 * but that might overwrite a racing page_add_anon_rmap
763 * which increments mapcount after us but sets mapping
764 * before us: so leave the reset to free_hot_cold_page,
765 * and remember that it's only reliable while mapped.
766 * Leaving it set also helps swapoff to reinstate ptes
767 * faster for those pages still in swapcache.
768 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700769 }
770}
771
772/*
773 * Subfunctions of try_to_unmap: try_to_unmap_one called
774 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
775 */
Christoph Lametera48d07a2006-02-01 03:05:38 -0800776static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
Andi Kleen14fa31b2009-09-16 11:50:10 +0200777 enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700778{
779 struct mm_struct *mm = vma->vm_mm;
780 unsigned long address;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700781 pte_t *pte;
782 pte_t pteval;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700783 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700784 int ret = SWAP_AGAIN;
785
Linus Torvalds1da177e2005-04-16 15:20:36 -0700786 address = vma_address(page, vma);
787 if (address == -EFAULT)
788 goto out;
789
Nick Piggin479db0b2008-08-20 14:09:18 -0700790 pte = page_check_address(page, mm, address, &ptl, 0);
Hugh Dickinsc0718802005-10-29 18:16:31 -0700791 if (!pte)
Nikita Danilov81b40822005-05-01 08:58:36 -0700792 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700793
794 /*
795 * If the page is mlock()d, we cannot swap it out.
796 * If it's recently referenced (perhaps page_referenced
797 * skipped over this mm) then we should reactivate it.
798 */
Andi Kleen14fa31b2009-09-16 11:50:10 +0200799 if (!(flags & TTU_IGNORE_MLOCK)) {
Nick Pigginb291f002008-10-18 20:26:44 -0700800 if (vma->vm_flags & VM_LOCKED) {
801 ret = SWAP_MLOCK;
802 goto out_unmap;
803 }
Andi Kleen14fa31b2009-09-16 11:50:10 +0200804 }
805 if (!(flags & TTU_IGNORE_ACCESS)) {
Nick Pigginb291f002008-10-18 20:26:44 -0700806 if (ptep_clear_flush_young_notify(vma, address, pte)) {
807 ret = SWAP_FAIL;
808 goto out_unmap;
809 }
810 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700811
Linus Torvalds1da177e2005-04-16 15:20:36 -0700812 /* Nuke the page table entry. */
813 flush_cache_page(vma, address, page_to_pfn(page));
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700814 pteval = ptep_clear_flush_notify(vma, address, pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700815
816 /* Move the dirty bit to the physical page now the pte is gone. */
817 if (pte_dirty(pteval))
818 set_page_dirty(page);
819
Hugh Dickins365e9c872005-10-29 18:16:18 -0700820 /* Update high watermark before we lower rss */
821 update_hiwater_rss(mm);
822
Andi Kleen888b9f72009-09-16 11:50:11 +0200823 if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
824 if (PageAnon(page))
825 dec_mm_counter(mm, anon_rss);
826 else
827 dec_mm_counter(mm, file_rss);
828 set_pte_at(mm, address, pte,
829 swp_entry_to_pte(make_hwpoison_entry(page)));
830 } else if (PageAnon(page)) {
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700831 swp_entry_t entry = { .val = page_private(page) };
Christoph Lameter06972122006-06-23 02:03:35 -0700832
833 if (PageSwapCache(page)) {
834 /*
835 * Store the swap location in the pte.
836 * See handle_pte_fault() ...
837 */
838 swap_duplicate(entry);
839 if (list_empty(&mm->mmlist)) {
840 spin_lock(&mmlist_lock);
841 if (list_empty(&mm->mmlist))
842 list_add(&mm->mmlist, &init_mm.mmlist);
843 spin_unlock(&mmlist_lock);
844 }
Christoph Lameter442c9132006-06-23 02:03:38 -0700845 dec_mm_counter(mm, anon_rss);
KOSAKI Motohiro64cdd542009-01-06 14:39:16 -0800846 } else if (PAGE_MIGRATION) {
Christoph Lameter06972122006-06-23 02:03:35 -0700847 /*
848 * Store the pfn of the page in a special migration
849 * pte. do_swap_page() will wait until the migration
850 * pte is removed and then restart fault handling.
851 */
Andi Kleen14fa31b2009-09-16 11:50:10 +0200852 BUG_ON(TTU_ACTION(flags) != TTU_MIGRATION);
Christoph Lameter06972122006-06-23 02:03:35 -0700853 entry = make_migration_entry(page, pte_write(pteval));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700854 }
855 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
856 BUG_ON(pte_file(*pte));
Andi Kleen14fa31b2009-09-16 11:50:10 +0200857 } else if (PAGE_MIGRATION && (TTU_ACTION(flags) == TTU_MIGRATION)) {
Christoph Lameter04e62a22006-06-23 02:03:38 -0700858 /* Establish migration entry for a file page */
859 swp_entry_t entry;
860 entry = make_migration_entry(page, pte_write(pteval));
861 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
862 } else
Hugh Dickins42946212005-10-29 18:16:05 -0700863 dec_mm_counter(mm, file_rss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700864
Christoph Lameter04e62a22006-06-23 02:03:38 -0700865
Hugh Dickinsedc315f2009-01-06 14:40:11 -0800866 page_remove_rmap(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700867 page_cache_release(page);
868
869out_unmap:
Hugh Dickinsc0718802005-10-29 18:16:31 -0700870 pte_unmap_unlock(pte, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700871out:
872 return ret;
873}
874
875/*
876 * objrmap doesn't work for nonlinear VMAs because the assumption that
877 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
878 * Consequently, given a particular page and its ->index, we cannot locate the
879 * ptes which are mapping that page without an exhaustive linear search.
880 *
881 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
882 * maps the file to which the target page belongs. The ->vm_private_data field
883 * holds the current cursor into that scan. Successive searches will circulate
884 * around the vma's virtual address space.
885 *
886 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
887 * more scanning pressure is placed against them as well. Eventually pages
888 * will become fully unmapped and are eligible for eviction.
889 *
890 * For very sparsely populated VMAs this is a little inefficient - chances are
891 * there there won't be many ptes located within the scan cluster. In this case
892 * maybe we could scan further - to the end of the pte page, perhaps.
Nick Pigginb291f002008-10-18 20:26:44 -0700893 *
894 * Mlocked pages: check VM_LOCKED under mmap_sem held for read, if we can
895 * acquire it without blocking. If vma locked, mlock the pages in the cluster,
896 * rather than unmapping them. If we encounter the "check_page" that vmscan is
897 * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700898 */
899#define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
900#define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
901
Nick Pigginb291f002008-10-18 20:26:44 -0700902static int try_to_unmap_cluster(unsigned long cursor, unsigned int *mapcount,
903 struct vm_area_struct *vma, struct page *check_page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700904{
905 struct mm_struct *mm = vma->vm_mm;
906 pgd_t *pgd;
907 pud_t *pud;
908 pmd_t *pmd;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700909 pte_t *pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700910 pte_t pteval;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700911 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700912 struct page *page;
913 unsigned long address;
914 unsigned long end;
Nick Pigginb291f002008-10-18 20:26:44 -0700915 int ret = SWAP_AGAIN;
916 int locked_vma = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700917
Linus Torvalds1da177e2005-04-16 15:20:36 -0700918 address = (vma->vm_start + cursor) & CLUSTER_MASK;
919 end = address + CLUSTER_SIZE;
920 if (address < vma->vm_start)
921 address = vma->vm_start;
922 if (end > vma->vm_end)
923 end = vma->vm_end;
924
925 pgd = pgd_offset(mm, address);
926 if (!pgd_present(*pgd))
Nick Pigginb291f002008-10-18 20:26:44 -0700927 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700928
929 pud = pud_offset(pgd, address);
930 if (!pud_present(*pud))
Nick Pigginb291f002008-10-18 20:26:44 -0700931 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700932
933 pmd = pmd_offset(pud, address);
934 if (!pmd_present(*pmd))
Nick Pigginb291f002008-10-18 20:26:44 -0700935 return ret;
936
937 /*
938 * MLOCK_PAGES => feature is configured.
939 * if we can acquire the mmap_sem for read, and vma is VM_LOCKED,
940 * keep the sem while scanning the cluster for mlocking pages.
941 */
942 if (MLOCK_PAGES && down_read_trylock(&vma->vm_mm->mmap_sem)) {
943 locked_vma = (vma->vm_flags & VM_LOCKED);
944 if (!locked_vma)
945 up_read(&vma->vm_mm->mmap_sem); /* don't need it */
946 }
Hugh Dickinsc0718802005-10-29 18:16:31 -0700947
948 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700949
Hugh Dickins365e9c872005-10-29 18:16:18 -0700950 /* Update high watermark before we lower rss */
951 update_hiwater_rss(mm);
952
Hugh Dickinsc0718802005-10-29 18:16:31 -0700953 for (; address < end; pte++, address += PAGE_SIZE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700954 if (!pte_present(*pte))
955 continue;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800956 page = vm_normal_page(vma, address, *pte);
957 BUG_ON(!page || PageAnon(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700958
Nick Pigginb291f002008-10-18 20:26:44 -0700959 if (locked_vma) {
960 mlock_vma_page(page); /* no-op if already mlocked */
961 if (page == check_page)
962 ret = SWAP_MLOCK;
963 continue; /* don't unmap */
964 }
965
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700966 if (ptep_clear_flush_young_notify(vma, address, pte))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700967 continue;
968
969 /* Nuke the page table entry. */
Ben Collinseca35132005-11-29 11:45:26 -0800970 flush_cache_page(vma, address, pte_pfn(*pte));
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700971 pteval = ptep_clear_flush_notify(vma, address, pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700972
973 /* If nonlinear, store the file page offset in the pte. */
974 if (page->index != linear_page_index(vma, address))
975 set_pte_at(mm, address, pte, pgoff_to_pte(page->index));
976
977 /* Move the dirty bit to the physical page now the pte is gone. */
978 if (pte_dirty(pteval))
979 set_page_dirty(page);
980
Hugh Dickinsedc315f2009-01-06 14:40:11 -0800981 page_remove_rmap(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700982 page_cache_release(page);
Hugh Dickins42946212005-10-29 18:16:05 -0700983 dec_mm_counter(mm, file_rss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700984 (*mapcount)--;
985 }
Hugh Dickinsc0718802005-10-29 18:16:31 -0700986 pte_unmap_unlock(pte - 1, ptl);
Nick Pigginb291f002008-10-18 20:26:44 -0700987 if (locked_vma)
988 up_read(&vma->vm_mm->mmap_sem);
989 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700990}
991
Nick Pigginb291f002008-10-18 20:26:44 -0700992/*
993 * common handling for pages mapped in VM_LOCKED vmas
994 */
995static int try_to_mlock_page(struct page *page, struct vm_area_struct *vma)
996{
997 int mlocked = 0;
998
999 if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
1000 if (vma->vm_flags & VM_LOCKED) {
1001 mlock_vma_page(page);
1002 mlocked++; /* really mlocked the page */
1003 }
1004 up_read(&vma->vm_mm->mmap_sem);
1005 }
1006 return mlocked;
1007}
1008
1009/**
1010 * try_to_unmap_anon - unmap or unlock anonymous page using the object-based
1011 * rmap method
1012 * @page: the page to unmap/unlock
1013 * @unlock: request for unlock rather than unmap [unlikely]
1014 * @migration: unmapping for migration - ignored if @unlock
1015 *
1016 * Find all the mappings of a page using the mapping pointer and the vma chains
1017 * contained in the anon_vma struct it points to.
1018 *
1019 * This function is only called from try_to_unmap/try_to_munlock for
1020 * anonymous pages.
1021 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1022 * where the page was found will be held for write. So, we won't recheck
1023 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1024 * 'LOCKED.
1025 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001026static int try_to_unmap_anon(struct page *page, enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001027{
1028 struct anon_vma *anon_vma;
1029 struct vm_area_struct *vma;
Nick Pigginb291f002008-10-18 20:26:44 -07001030 unsigned int mlocked = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001031 int ret = SWAP_AGAIN;
Andi Kleen14fa31b2009-09-16 11:50:10 +02001032 int unlock = TTU_ACTION(flags) == TTU_MUNLOCK;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001033
Nick Pigginb291f002008-10-18 20:26:44 -07001034 if (MLOCK_PAGES && unlikely(unlock))
1035 ret = SWAP_SUCCESS; /* default for try_to_munlock() */
1036
Linus Torvalds1da177e2005-04-16 15:20:36 -07001037 anon_vma = page_lock_anon_vma(page);
1038 if (!anon_vma)
1039 return ret;
1040
1041 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
Nick Pigginb291f002008-10-18 20:26:44 -07001042 if (MLOCK_PAGES && unlikely(unlock)) {
1043 if (!((vma->vm_flags & VM_LOCKED) &&
1044 page_mapped_in_vma(page, vma)))
1045 continue; /* must visit all unlocked vmas */
1046 ret = SWAP_MLOCK; /* saw at least one mlocked vma */
1047 } else {
Andi Kleen14fa31b2009-09-16 11:50:10 +02001048 ret = try_to_unmap_one(page, vma, flags);
Nick Pigginb291f002008-10-18 20:26:44 -07001049 if (ret == SWAP_FAIL || !page_mapped(page))
1050 break;
1051 }
1052 if (ret == SWAP_MLOCK) {
1053 mlocked = try_to_mlock_page(page, vma);
1054 if (mlocked)
1055 break; /* stop if actually mlocked page */
1056 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001057 }
Oleg Nesterov34bbd702007-02-28 20:13:49 -08001058
1059 page_unlock_anon_vma(anon_vma);
Nick Pigginb291f002008-10-18 20:26:44 -07001060
1061 if (mlocked)
1062 ret = SWAP_MLOCK; /* actually mlocked the page */
1063 else if (ret == SWAP_MLOCK)
1064 ret = SWAP_AGAIN; /* saw VM_LOCKED vma */
1065
Linus Torvalds1da177e2005-04-16 15:20:36 -07001066 return ret;
1067}
1068
1069/**
Nick Pigginb291f002008-10-18 20:26:44 -07001070 * try_to_unmap_file - unmap/unlock file page using the object-based rmap method
1071 * @page: the page to unmap/unlock
Andi Kleen14fa31b2009-09-16 11:50:10 +02001072 * @flags: action and flags
Linus Torvalds1da177e2005-04-16 15:20:36 -07001073 *
1074 * Find all the mappings of a page using the mapping pointer and the vma chains
1075 * contained in the address_space struct it points to.
1076 *
Nick Pigginb291f002008-10-18 20:26:44 -07001077 * This function is only called from try_to_unmap/try_to_munlock for
1078 * object-based pages.
1079 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1080 * where the page was found will be held for write. So, we won't recheck
1081 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1082 * 'LOCKED.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001083 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001084static int try_to_unmap_file(struct page *page, enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001085{
1086 struct address_space *mapping = page->mapping;
1087 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1088 struct vm_area_struct *vma;
1089 struct prio_tree_iter iter;
1090 int ret = SWAP_AGAIN;
1091 unsigned long cursor;
1092 unsigned long max_nl_cursor = 0;
1093 unsigned long max_nl_size = 0;
1094 unsigned int mapcount;
Nick Pigginb291f002008-10-18 20:26:44 -07001095 unsigned int mlocked = 0;
Andi Kleen14fa31b2009-09-16 11:50:10 +02001096 int unlock = TTU_ACTION(flags) == TTU_MUNLOCK;
Nick Pigginb291f002008-10-18 20:26:44 -07001097
1098 if (MLOCK_PAGES && unlikely(unlock))
1099 ret = SWAP_SUCCESS; /* default for try_to_munlock() */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001100
1101 spin_lock(&mapping->i_mmap_lock);
1102 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
Nick Pigginb291f002008-10-18 20:26:44 -07001103 if (MLOCK_PAGES && unlikely(unlock)) {
MinChan Kim508b9f82009-02-11 13:04:27 -08001104 if (!((vma->vm_flags & VM_LOCKED) &&
1105 page_mapped_in_vma(page, vma)))
Nick Pigginb291f002008-10-18 20:26:44 -07001106 continue; /* must visit all vmas */
1107 ret = SWAP_MLOCK;
1108 } else {
Andi Kleen14fa31b2009-09-16 11:50:10 +02001109 ret = try_to_unmap_one(page, vma, flags);
Nick Pigginb291f002008-10-18 20:26:44 -07001110 if (ret == SWAP_FAIL || !page_mapped(page))
1111 goto out;
1112 }
1113 if (ret == SWAP_MLOCK) {
1114 mlocked = try_to_mlock_page(page, vma);
1115 if (mlocked)
1116 break; /* stop if actually mlocked page */
1117 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001118 }
1119
Nick Pigginb291f002008-10-18 20:26:44 -07001120 if (mlocked)
1121 goto out;
1122
Linus Torvalds1da177e2005-04-16 15:20:36 -07001123 if (list_empty(&mapping->i_mmap_nonlinear))
1124 goto out;
1125
1126 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1127 shared.vm_set.list) {
Nick Pigginb291f002008-10-18 20:26:44 -07001128 if (MLOCK_PAGES && unlikely(unlock)) {
1129 if (!(vma->vm_flags & VM_LOCKED))
1130 continue; /* must visit all vmas */
1131 ret = SWAP_MLOCK; /* leave mlocked == 0 */
1132 goto out; /* no need to look further */
1133 }
Andi Kleen14fa31b2009-09-16 11:50:10 +02001134 if (!MLOCK_PAGES && !(flags & TTU_IGNORE_MLOCK) &&
1135 (vma->vm_flags & VM_LOCKED))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001136 continue;
1137 cursor = (unsigned long) vma->vm_private_data;
1138 if (cursor > max_nl_cursor)
1139 max_nl_cursor = cursor;
1140 cursor = vma->vm_end - vma->vm_start;
1141 if (cursor > max_nl_size)
1142 max_nl_size = cursor;
1143 }
1144
Nick Pigginb291f002008-10-18 20:26:44 -07001145 if (max_nl_size == 0) { /* all nonlinears locked or reserved ? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001146 ret = SWAP_FAIL;
1147 goto out;
1148 }
1149
1150 /*
1151 * We don't try to search for this page in the nonlinear vmas,
1152 * and page_referenced wouldn't have found it anyway. Instead
1153 * just walk the nonlinear vmas trying to age and unmap some.
1154 * The mapcount of the page we came in with is irrelevant,
1155 * but even so use it as a guide to how hard we should try?
1156 */
1157 mapcount = page_mapcount(page);
1158 if (!mapcount)
1159 goto out;
1160 cond_resched_lock(&mapping->i_mmap_lock);
1161
1162 max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
1163 if (max_nl_cursor == 0)
1164 max_nl_cursor = CLUSTER_SIZE;
1165
1166 do {
1167 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1168 shared.vm_set.list) {
Andi Kleen14fa31b2009-09-16 11:50:10 +02001169 if (!MLOCK_PAGES && !(flags & TTU_IGNORE_MLOCK) &&
Nick Pigginb291f002008-10-18 20:26:44 -07001170 (vma->vm_flags & VM_LOCKED))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001171 continue;
1172 cursor = (unsigned long) vma->vm_private_data;
Hugh Dickins839b9682005-09-03 15:54:43 -07001173 while ( cursor < max_nl_cursor &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07001174 cursor < vma->vm_end - vma->vm_start) {
Nick Pigginb291f002008-10-18 20:26:44 -07001175 ret = try_to_unmap_cluster(cursor, &mapcount,
1176 vma, page);
1177 if (ret == SWAP_MLOCK)
1178 mlocked = 2; /* to return below */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001179 cursor += CLUSTER_SIZE;
1180 vma->vm_private_data = (void *) cursor;
1181 if ((int)mapcount <= 0)
1182 goto out;
1183 }
1184 vma->vm_private_data = (void *) max_nl_cursor;
1185 }
1186 cond_resched_lock(&mapping->i_mmap_lock);
1187 max_nl_cursor += CLUSTER_SIZE;
1188 } while (max_nl_cursor <= max_nl_size);
1189
1190 /*
1191 * Don't loop forever (perhaps all the remaining pages are
1192 * in locked vmas). Reset cursor on all unreserved nonlinear
1193 * vmas, now forgetting on which ones it had fallen behind.
1194 */
Hugh Dickins101d2be2005-11-21 21:32:16 -08001195 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1196 vma->vm_private_data = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001197out:
1198 spin_unlock(&mapping->i_mmap_lock);
Nick Pigginb291f002008-10-18 20:26:44 -07001199 if (mlocked)
1200 ret = SWAP_MLOCK; /* actually mlocked the page */
1201 else if (ret == SWAP_MLOCK)
1202 ret = SWAP_AGAIN; /* saw VM_LOCKED vma */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001203 return ret;
1204}
1205
1206/**
1207 * try_to_unmap - try to remove all page table mappings to a page
1208 * @page: the page to get unmapped
Andi Kleen14fa31b2009-09-16 11:50:10 +02001209 * @flags: action and flags
Linus Torvalds1da177e2005-04-16 15:20:36 -07001210 *
1211 * Tries to remove all the page table entries which are mapping this
1212 * page, used in the pageout path. Caller must hold the page lock.
1213 * Return values are:
1214 *
1215 * SWAP_SUCCESS - we succeeded in removing all mappings
1216 * SWAP_AGAIN - we missed a mapping, try again later
1217 * SWAP_FAIL - the page is unswappable
Nick Pigginb291f002008-10-18 20:26:44 -07001218 * SWAP_MLOCK - page is mlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001219 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001220int try_to_unmap(struct page *page, enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001221{
1222 int ret;
1223
Linus Torvalds1da177e2005-04-16 15:20:36 -07001224 BUG_ON(!PageLocked(page));
1225
1226 if (PageAnon(page))
Andi Kleen14fa31b2009-09-16 11:50:10 +02001227 ret = try_to_unmap_anon(page, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001228 else
Andi Kleen14fa31b2009-09-16 11:50:10 +02001229 ret = try_to_unmap_file(page, flags);
Nick Pigginb291f002008-10-18 20:26:44 -07001230 if (ret != SWAP_MLOCK && !page_mapped(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001231 ret = SWAP_SUCCESS;
1232 return ret;
1233}
Nikita Danilov81b40822005-05-01 08:58:36 -07001234
Nick Pigginb291f002008-10-18 20:26:44 -07001235/**
1236 * try_to_munlock - try to munlock a page
1237 * @page: the page to be munlocked
1238 *
1239 * Called from munlock code. Checks all of the VMAs mapping the page
1240 * to make sure nobody else has this page mlocked. The page will be
1241 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1242 *
1243 * Return values are:
1244 *
1245 * SWAP_SUCCESS - no vma's holding page mlocked.
1246 * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem
1247 * SWAP_MLOCK - page is now mlocked.
1248 */
1249int try_to_munlock(struct page *page)
1250{
1251 VM_BUG_ON(!PageLocked(page) || PageLRU(page));
1252
1253 if (PageAnon(page))
Andi Kleen14fa31b2009-09-16 11:50:10 +02001254 return try_to_unmap_anon(page, TTU_MUNLOCK);
Nick Pigginb291f002008-10-18 20:26:44 -07001255 else
Andi Kleen14fa31b2009-09-16 11:50:10 +02001256 return try_to_unmap_file(page, TTU_MUNLOCK);
Nick Pigginb291f002008-10-18 20:26:44 -07001257}
KOSAKI Motohiro68377652009-06-16 15:32:51 -07001258