blob: 16521664010ddc06b5dfb25303413bedb6a58041 [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
17 * Contributions by Hugh Dickins <hugh@veritas.com> 2003, 2004
18 */
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 */
Hugh Dickins2afd1c92009-01-06 14:39:26 -0800194static struct 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
Hugh Dickins2afd1c92009-01-06 14:39:26 -0800214static void 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 Schermerhorn3ad33b22007-11-14 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 Schermerhorn3ad33b22007-11-14 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,
Rik van Rielf7b7fd82005-11-28 13:44:07 -0800336 struct vm_area_struct *vma, unsigned int *mapcount)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700337{
338 struct mm_struct *mm = vma->vm_mm;
339 unsigned long address;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700340 pte_t *pte;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700341 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700342 int referenced = 0;
343
Linus Torvalds1da177e2005-04-16 15:20:36 -0700344 address = vma_address(page, vma);
345 if (address == -EFAULT)
346 goto out;
347
Nick Piggin479db0b2008-08-20 14:09:18 -0700348 pte = page_check_address(page, mm, address, &ptl, 0);
Hugh Dickinsc0718802005-10-29 18:16:31 -0700349 if (!pte)
350 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700351
Nick Pigginb291f002008-10-18 20:26:44 -0700352 /*
353 * Don't want to elevate referenced for mlocked page that gets this far,
354 * in order that it progresses to try_to_unmap and is moved to the
355 * unevictable list.
356 */
Hugh Dickins5a9bbdc2008-02-04 22:29:23 -0800357 if (vma->vm_flags & VM_LOCKED) {
Hugh Dickins5a9bbdc2008-02-04 22:29:23 -0800358 *mapcount = 1; /* break early from loop */
Nick Pigginb291f002008-10-18 20:26:44 -0700359 goto out_unmap;
360 }
361
Johannes Weiner4917e5d2009-01-06 14:39:17 -0800362 if (ptep_clear_flush_young_notify(vma, address, pte)) {
363 /*
364 * Don't treat a reference through a sequentially read
365 * mapping as such. If the page has been used in
366 * another mapping, we will catch it; if this other
367 * mapping is already gone, the unmap path will have
368 * set PG_referenced or activated the page.
369 */
370 if (likely(!VM_SequentialReadHint(vma)))
371 referenced++;
372 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700373
Hugh Dickinsc0718802005-10-29 18:16:31 -0700374 /* Pretend the page is referenced if the task has the
375 swap token and is in the middle of a page fault. */
Rik van Rielf7b7fd82005-11-28 13:44:07 -0800376 if (mm != current->mm && has_swap_token(mm) &&
Hugh Dickinsc0718802005-10-29 18:16:31 -0700377 rwsem_is_locked(&mm->mmap_sem))
378 referenced++;
379
Nick Pigginb291f002008-10-18 20:26:44 -0700380out_unmap:
Hugh Dickinsc0718802005-10-29 18:16:31 -0700381 (*mapcount)--;
382 pte_unmap_unlock(pte, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700383out:
384 return referenced;
385}
386
Balbir Singhbed71612008-02-07 00:14:01 -0800387static int page_referenced_anon(struct page *page,
388 struct mem_cgroup *mem_cont)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700389{
390 unsigned int mapcount;
391 struct anon_vma *anon_vma;
392 struct vm_area_struct *vma;
393 int referenced = 0;
394
395 anon_vma = page_lock_anon_vma(page);
396 if (!anon_vma)
397 return referenced;
398
399 mapcount = page_mapcount(page);
400 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
Balbir Singhbed71612008-02-07 00:14:01 -0800401 /*
402 * If we are reclaiming on behalf of a cgroup, skip
403 * counting on behalf of references from different
404 * cgroups
405 */
Hugh Dickinsbd845e32008-03-04 14:29:01 -0800406 if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
Balbir Singhbed71612008-02-07 00:14:01 -0800407 continue;
Rik van Rielf7b7fd82005-11-28 13:44:07 -0800408 referenced += page_referenced_one(page, vma, &mapcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700409 if (!mapcount)
410 break;
411 }
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800412
413 page_unlock_anon_vma(anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700414 return referenced;
415}
416
417/**
418 * page_referenced_file - referenced check for object-based rmap
419 * @page: the page we're checking references on.
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700420 * @mem_cont: target memory controller
Linus Torvalds1da177e2005-04-16 15:20:36 -0700421 *
422 * For an object-based mapped page, find all the places it is mapped and
423 * check/clear the referenced flag. This is done by following the page->mapping
424 * pointer, then walking the chain of vmas it holds. It returns the number
425 * of references it found.
426 *
427 * This function is only called from page_referenced for object-based pages.
428 */
Balbir Singhbed71612008-02-07 00:14:01 -0800429static int page_referenced_file(struct page *page,
430 struct mem_cgroup *mem_cont)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700431{
432 unsigned int mapcount;
433 struct address_space *mapping = page->mapping;
434 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
435 struct vm_area_struct *vma;
436 struct prio_tree_iter iter;
437 int referenced = 0;
438
439 /*
440 * The caller's checks on page->mapping and !PageAnon have made
441 * sure that this is a file page: the check for page->mapping
442 * excludes the case just before it gets set on an anon page.
443 */
444 BUG_ON(PageAnon(page));
445
446 /*
447 * The page lock not only makes sure that page->mapping cannot
448 * suddenly be NULLified by truncation, it makes sure that the
449 * structure at mapping cannot be freed and reused yet,
450 * so we can safely take mapping->i_mmap_lock.
451 */
452 BUG_ON(!PageLocked(page));
453
454 spin_lock(&mapping->i_mmap_lock);
455
456 /*
457 * i_mmap_lock does not stabilize mapcount at all, but mapcount
458 * is more likely to be accurate if we note it after spinning.
459 */
460 mapcount = page_mapcount(page);
461
462 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
Balbir Singhbed71612008-02-07 00:14:01 -0800463 /*
464 * If we are reclaiming on behalf of a cgroup, skip
465 * counting on behalf of references from different
466 * cgroups
467 */
Hugh Dickinsbd845e32008-03-04 14:29:01 -0800468 if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
Balbir Singhbed71612008-02-07 00:14:01 -0800469 continue;
Rik van Rielf7b7fd82005-11-28 13:44:07 -0800470 referenced += page_referenced_one(page, vma, &mapcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700471 if (!mapcount)
472 break;
473 }
474
475 spin_unlock(&mapping->i_mmap_lock);
476 return referenced;
477}
478
479/**
480 * page_referenced - test if the page was referenced
481 * @page: the page to test
482 * @is_locked: caller holds lock on the page
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700483 * @mem_cont: target memory controller
Linus Torvalds1da177e2005-04-16 15:20:36 -0700484 *
485 * Quick test_and_clear_referenced for all mappings to a page,
486 * returns the number of ptes which referenced the page.
487 */
Balbir Singhbed71612008-02-07 00:14:01 -0800488int page_referenced(struct page *page, int is_locked,
489 struct mem_cgroup *mem_cont)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700490{
491 int referenced = 0;
492
Linus Torvalds1da177e2005-04-16 15:20:36 -0700493 if (TestClearPageReferenced(page))
494 referenced++;
495
496 if (page_mapped(page) && page->mapping) {
497 if (PageAnon(page))
Balbir Singhbed71612008-02-07 00:14:01 -0800498 referenced += page_referenced_anon(page, mem_cont);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700499 else if (is_locked)
Balbir Singhbed71612008-02-07 00:14:01 -0800500 referenced += page_referenced_file(page, mem_cont);
Nick Piggin529ae9a2008-08-02 12:01:03 +0200501 else if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700502 referenced++;
503 else {
504 if (page->mapping)
Balbir Singhbed71612008-02-07 00:14:01 -0800505 referenced +=
506 page_referenced_file(page, mem_cont);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700507 unlock_page(page);
508 }
509 }
Christian Borntraeger5b7baf02008-03-25 18:47:12 +0100510
511 if (page_test_and_clear_young(page))
512 referenced++;
513
Linus Torvalds1da177e2005-04-16 15:20:36 -0700514 return referenced;
515}
516
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700517static int page_mkclean_one(struct page *page, struct vm_area_struct *vma)
518{
519 struct mm_struct *mm = vma->vm_mm;
520 unsigned long address;
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100521 pte_t *pte;
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700522 spinlock_t *ptl;
523 int ret = 0;
524
525 address = vma_address(page, vma);
526 if (address == -EFAULT)
527 goto out;
528
Nick Piggin479db0b2008-08-20 14:09:18 -0700529 pte = page_check_address(page, mm, address, &ptl, 1);
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700530 if (!pte)
531 goto out;
532
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100533 if (pte_dirty(*pte) || pte_write(*pte)) {
534 pte_t entry;
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700535
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100536 flush_cache_page(vma, address, pte_pfn(*pte));
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700537 entry = ptep_clear_flush_notify(vma, address, pte);
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100538 entry = pte_wrprotect(entry);
539 entry = pte_mkclean(entry);
Al Virod6e88e62006-12-29 16:48:35 -0800540 set_pte_at(mm, address, pte, entry);
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100541 ret = 1;
542 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700543
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700544 pte_unmap_unlock(pte, ptl);
545out:
546 return ret;
547}
548
549static int page_mkclean_file(struct address_space *mapping, struct page *page)
550{
551 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
552 struct vm_area_struct *vma;
553 struct prio_tree_iter iter;
554 int ret = 0;
555
556 BUG_ON(PageAnon(page));
557
558 spin_lock(&mapping->i_mmap_lock);
559 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
560 if (vma->vm_flags & VM_SHARED)
561 ret += page_mkclean_one(page, vma);
562 }
563 spin_unlock(&mapping->i_mmap_lock);
564 return ret;
565}
566
567int page_mkclean(struct page *page)
568{
569 int ret = 0;
570
571 BUG_ON(!PageLocked(page));
572
573 if (page_mapped(page)) {
574 struct address_space *mapping = page_mapping(page);
Christian Borntraegerce7e9fa2007-11-20 11:13:36 +0100575 if (mapping) {
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700576 ret = page_mkclean_file(mapping, page);
Christian Borntraegerce7e9fa2007-11-20 11:13:36 +0100577 if (page_test_dirty(page)) {
578 page_clear_dirty(page);
579 ret = 1;
580 }
Martin Schwidefsky6c210482007-04-27 16:01:57 +0200581 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700582 }
583
584 return ret;
585}
Jaya Kumar60b59be2007-05-08 00:37:37 -0700586EXPORT_SYMBOL_GPL(page_mkclean);
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700587
Linus Torvalds1da177e2005-04-16 15:20:36 -0700588/**
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700589 * __page_set_anon_rmap - setup new anonymous rmap
Nick Piggin9617d952006-01-06 00:11:12 -0800590 * @page: the page to add the mapping to
591 * @vma: the vm area in which the mapping is added
592 * @address: the user virtual address mapped
593 */
594static void __page_set_anon_rmap(struct page *page,
595 struct vm_area_struct *vma, unsigned long address)
596{
597 struct anon_vma *anon_vma = vma->anon_vma;
598
599 BUG_ON(!anon_vma);
600 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
601 page->mapping = (struct address_space *) anon_vma;
602
603 page->index = linear_page_index(vma, address);
604
Nick Piggina74609f2006-01-06 00:11:20 -0800605 /*
606 * nr_mapped state can be updated without turning off
607 * interrupts because it is not modified via interrupt.
608 */
Christoph Lameterf3dbd342006-06-30 01:55:36 -0700609 __inc_zone_page_state(page, NR_ANON_PAGES);
Nick Piggin9617d952006-01-06 00:11:12 -0800610}
611
612/**
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700613 * __page_check_anon_rmap - sanity check anonymous rmap addition
Nick Pigginc97a9e12007-05-16 22:11:21 -0700614 * @page: the page to add the mapping to
615 * @vma: the vm area in which the mapping is added
616 * @address: the user virtual address mapped
617 */
618static void __page_check_anon_rmap(struct page *page,
619 struct vm_area_struct *vma, unsigned long address)
620{
621#ifdef CONFIG_DEBUG_VM
622 /*
623 * The page's anon-rmap details (mapping and index) are guaranteed to
624 * be set up correctly at this point.
625 *
626 * We have exclusion against page_add_anon_rmap because the caller
627 * always holds the page locked, except if called from page_dup_rmap,
628 * in which case the page is already known to be setup.
629 *
630 * We have exclusion against page_add_new_anon_rmap because those pages
631 * are initially only visible via the pagetables, and the pte is locked
632 * over the call to page_add_new_anon_rmap.
633 */
634 struct anon_vma *anon_vma = vma->anon_vma;
635 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
636 BUG_ON(page->mapping != (struct address_space *)anon_vma);
637 BUG_ON(page->index != linear_page_index(vma, address));
638#endif
639}
640
641/**
Linus Torvalds1da177e2005-04-16 15:20:36 -0700642 * page_add_anon_rmap - add pte mapping to an anonymous page
643 * @page: the page to add the mapping to
644 * @vma: the vm area in which the mapping is added
645 * @address: the user virtual address mapped
646 *
Nick Pigginc97a9e12007-05-16 22:11:21 -0700647 * The caller needs to hold the pte lock and the page must be locked.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700648 */
649void page_add_anon_rmap(struct page *page,
650 struct vm_area_struct *vma, unsigned long address)
651{
Nick Pigginc97a9e12007-05-16 22:11:21 -0700652 VM_BUG_ON(!PageLocked(page));
653 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
Nick Piggin9617d952006-01-06 00:11:12 -0800654 if (atomic_inc_and_test(&page->_mapcount))
655 __page_set_anon_rmap(page, vma, address);
KAMEZAWA Hiroyuki69029cd2008-07-25 01:47:14 -0700656 else
Nick Pigginc97a9e12007-05-16 22:11:21 -0700657 __page_check_anon_rmap(page, vma, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700658}
659
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700660/**
Nick Piggin9617d952006-01-06 00:11:12 -0800661 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
662 * @page: the page to add the mapping to
663 * @vma: the vm area in which the mapping is added
664 * @address: the user virtual address mapped
665 *
666 * Same as page_add_anon_rmap but must only be called on *new* pages.
667 * This means the inc-and-test can be bypassed.
Nick Pigginc97a9e12007-05-16 22:11:21 -0700668 * Page does not have to be locked.
Nick Piggin9617d952006-01-06 00:11:12 -0800669 */
670void page_add_new_anon_rmap(struct page *page,
671 struct vm_area_struct *vma, unsigned long address)
672{
Hugh Dickinsb5934c52009-01-06 14:39:25 -0800673 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
Hugh Dickinscbf84b72009-01-06 14:39:27 -0800674 SetPageSwapBacked(page);
675 atomic_set(&page->_mapcount, 0); /* increment count (starts at -1) */
Nick Piggin9617d952006-01-06 00:11:12 -0800676 __page_set_anon_rmap(page, vma, address);
Hugh Dickinsb5934c52009-01-06 14:39:25 -0800677 if (page_evictable(page, vma))
Hugh Dickinscbf84b72009-01-06 14:39:27 -0800678 lru_cache_add_lru(page, LRU_ACTIVE_ANON);
Hugh Dickinsb5934c52009-01-06 14:39:25 -0800679 else
680 add_page_to_unevictable_list(page);
Nick Piggin9617d952006-01-06 00:11:12 -0800681}
682
Linus Torvalds1da177e2005-04-16 15:20:36 -0700683/**
684 * page_add_file_rmap - add pte mapping to a file page
685 * @page: the page to add the mapping to
686 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -0700687 * The caller needs to hold the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700688 */
689void page_add_file_rmap(struct page *page)
690{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700691 if (atomic_inc_and_test(&page->_mapcount))
Christoph Lameter65ba55f2006-06-30 01:55:34 -0700692 __inc_zone_page_state(page, NR_FILE_MAPPED);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700693}
694
Nick Pigginc97a9e12007-05-16 22:11:21 -0700695#ifdef CONFIG_DEBUG_VM
696/**
697 * page_dup_rmap - duplicate pte mapping to a page
698 * @page: the page to add the mapping to
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700699 * @vma: the vm area being duplicated
700 * @address: the user virtual address mapped
Nick Pigginc97a9e12007-05-16 22:11:21 -0700701 *
702 * For copy_page_range only: minimal extract from page_add_file_rmap /
703 * page_add_anon_rmap, avoiding unnecessary tests (already checked) so it's
704 * quicker.
705 *
706 * The caller needs to hold the pte lock.
707 */
708void page_dup_rmap(struct page *page, struct vm_area_struct *vma, unsigned long address)
709{
Nick Pigginc97a9e12007-05-16 22:11:21 -0700710 if (PageAnon(page))
711 __page_check_anon_rmap(page, vma, address);
712 atomic_inc(&page->_mapcount);
713}
714#endif
715
Linus Torvalds1da177e2005-04-16 15:20:36 -0700716/**
717 * page_remove_rmap - take down pte mapping from a page
718 * @page: page to remove mapping from
719 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -0700720 * The caller needs to hold the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700721 */
Hugh Dickinsedc315f2009-01-06 14:40:11 -0800722void page_remove_rmap(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700723{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700724 if (atomic_add_negative(-1, &page->_mapcount)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700725 /*
Hugh Dickins16f8c5b2008-08-20 14:09:04 -0700726 * Now that the last pte has gone, s390 must transfer dirty
727 * flag from storage key to struct page. We can usually skip
728 * this if the page is anon, so about to be freed; but perhaps
729 * not if it's in swapcache - there might be another pte slot
730 * containing the swap entry, but page not yet written to swap.
731 */
732 if ((!PageAnon(page) || PageSwapCache(page)) &&
733 page_test_dirty(page)) {
734 page_clear_dirty(page);
735 set_page_dirty(page);
736 }
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -0700737 if (PageAnon(page))
738 mem_cgroup_uncharge_page(page);
Hugh Dickins16f8c5b2008-08-20 14:09:04 -0700739 __dec_zone_page_state(page,
740 PageAnon(page) ? NR_ANON_PAGES : NR_FILE_MAPPED);
741 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700742 * It would be tidy to reset the PageAnon mapping here,
743 * but that might overwrite a racing page_add_anon_rmap
744 * which increments mapcount after us but sets mapping
745 * before us: so leave the reset to free_hot_cold_page,
746 * and remember that it's only reliable while mapped.
747 * Leaving it set also helps swapoff to reinstate ptes
748 * faster for those pages still in swapcache.
749 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700750 }
751}
752
753/*
754 * Subfunctions of try_to_unmap: try_to_unmap_one called
755 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
756 */
Christoph Lametera48d07a2006-02-01 03:05:38 -0800757static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
Christoph Lameter73523492006-06-23 02:03:27 -0700758 int migration)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700759{
760 struct mm_struct *mm = vma->vm_mm;
761 unsigned long address;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700762 pte_t *pte;
763 pte_t pteval;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700764 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700765 int ret = SWAP_AGAIN;
766
Linus Torvalds1da177e2005-04-16 15:20:36 -0700767 address = vma_address(page, vma);
768 if (address == -EFAULT)
769 goto out;
770
Nick Piggin479db0b2008-08-20 14:09:18 -0700771 pte = page_check_address(page, mm, address, &ptl, 0);
Hugh Dickinsc0718802005-10-29 18:16:31 -0700772 if (!pte)
Nikita Danilov81b40822005-05-01 08:58:36 -0700773 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700774
775 /*
776 * If the page is mlock()d, we cannot swap it out.
777 * If it's recently referenced (perhaps page_referenced
778 * skipped over this mm) then we should reactivate it.
779 */
Nick Pigginb291f002008-10-18 20:26:44 -0700780 if (!migration) {
781 if (vma->vm_flags & VM_LOCKED) {
782 ret = SWAP_MLOCK;
783 goto out_unmap;
784 }
785 if (ptep_clear_flush_young_notify(vma, address, pte)) {
786 ret = SWAP_FAIL;
787 goto out_unmap;
788 }
789 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700790
Linus Torvalds1da177e2005-04-16 15:20:36 -0700791 /* Nuke the page table entry. */
792 flush_cache_page(vma, address, page_to_pfn(page));
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700793 pteval = ptep_clear_flush_notify(vma, address, pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700794
795 /* Move the dirty bit to the physical page now the pte is gone. */
796 if (pte_dirty(pteval))
797 set_page_dirty(page);
798
Hugh Dickins365e9c872005-10-29 18:16:18 -0700799 /* Update high watermark before we lower rss */
800 update_hiwater_rss(mm);
801
Linus Torvalds1da177e2005-04-16 15:20:36 -0700802 if (PageAnon(page)) {
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700803 swp_entry_t entry = { .val = page_private(page) };
Christoph Lameter06972122006-06-23 02:03:35 -0700804
805 if (PageSwapCache(page)) {
806 /*
807 * Store the swap location in the pte.
808 * See handle_pte_fault() ...
809 */
810 swap_duplicate(entry);
811 if (list_empty(&mm->mmlist)) {
812 spin_lock(&mmlist_lock);
813 if (list_empty(&mm->mmlist))
814 list_add(&mm->mmlist, &init_mm.mmlist);
815 spin_unlock(&mmlist_lock);
816 }
Christoph Lameter442c9132006-06-23 02:03:38 -0700817 dec_mm_counter(mm, anon_rss);
KOSAKI Motohiro64cdd542009-01-06 14:39:16 -0800818 } else if (PAGE_MIGRATION) {
Christoph Lameter06972122006-06-23 02:03:35 -0700819 /*
820 * Store the pfn of the page in a special migration
821 * pte. do_swap_page() will wait until the migration
822 * pte is removed and then restart fault handling.
823 */
824 BUG_ON(!migration);
825 entry = make_migration_entry(page, pte_write(pteval));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700826 }
827 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
828 BUG_ON(pte_file(*pte));
KOSAKI Motohiro64cdd542009-01-06 14:39:16 -0800829 } else if (PAGE_MIGRATION && migration) {
Christoph Lameter04e62a22006-06-23 02:03:38 -0700830 /* Establish migration entry for a file page */
831 swp_entry_t entry;
832 entry = make_migration_entry(page, pte_write(pteval));
833 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
834 } else
Hugh Dickins42946212005-10-29 18:16:05 -0700835 dec_mm_counter(mm, file_rss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700836
Christoph Lameter04e62a22006-06-23 02:03:38 -0700837
Hugh Dickinsedc315f2009-01-06 14:40:11 -0800838 page_remove_rmap(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700839 page_cache_release(page);
840
841out_unmap:
Hugh Dickinsc0718802005-10-29 18:16:31 -0700842 pte_unmap_unlock(pte, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700843out:
844 return ret;
845}
846
847/*
848 * objrmap doesn't work for nonlinear VMAs because the assumption that
849 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
850 * Consequently, given a particular page and its ->index, we cannot locate the
851 * ptes which are mapping that page without an exhaustive linear search.
852 *
853 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
854 * maps the file to which the target page belongs. The ->vm_private_data field
855 * holds the current cursor into that scan. Successive searches will circulate
856 * around the vma's virtual address space.
857 *
858 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
859 * more scanning pressure is placed against them as well. Eventually pages
860 * will become fully unmapped and are eligible for eviction.
861 *
862 * For very sparsely populated VMAs this is a little inefficient - chances are
863 * there there won't be many ptes located within the scan cluster. In this case
864 * maybe we could scan further - to the end of the pte page, perhaps.
Nick Pigginb291f002008-10-18 20:26:44 -0700865 *
866 * Mlocked pages: check VM_LOCKED under mmap_sem held for read, if we can
867 * acquire it without blocking. If vma locked, mlock the pages in the cluster,
868 * rather than unmapping them. If we encounter the "check_page" that vmscan is
869 * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700870 */
871#define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
872#define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
873
Nick Pigginb291f002008-10-18 20:26:44 -0700874static int try_to_unmap_cluster(unsigned long cursor, unsigned int *mapcount,
875 struct vm_area_struct *vma, struct page *check_page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700876{
877 struct mm_struct *mm = vma->vm_mm;
878 pgd_t *pgd;
879 pud_t *pud;
880 pmd_t *pmd;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700881 pte_t *pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700882 pte_t pteval;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700883 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700884 struct page *page;
885 unsigned long address;
886 unsigned long end;
Nick Pigginb291f002008-10-18 20:26:44 -0700887 int ret = SWAP_AGAIN;
888 int locked_vma = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700889
Linus Torvalds1da177e2005-04-16 15:20:36 -0700890 address = (vma->vm_start + cursor) & CLUSTER_MASK;
891 end = address + CLUSTER_SIZE;
892 if (address < vma->vm_start)
893 address = vma->vm_start;
894 if (end > vma->vm_end)
895 end = vma->vm_end;
896
897 pgd = pgd_offset(mm, address);
898 if (!pgd_present(*pgd))
Nick Pigginb291f002008-10-18 20:26:44 -0700899 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700900
901 pud = pud_offset(pgd, address);
902 if (!pud_present(*pud))
Nick Pigginb291f002008-10-18 20:26:44 -0700903 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700904
905 pmd = pmd_offset(pud, address);
906 if (!pmd_present(*pmd))
Nick Pigginb291f002008-10-18 20:26:44 -0700907 return ret;
908
909 /*
910 * MLOCK_PAGES => feature is configured.
911 * if we can acquire the mmap_sem for read, and vma is VM_LOCKED,
912 * keep the sem while scanning the cluster for mlocking pages.
913 */
914 if (MLOCK_PAGES && down_read_trylock(&vma->vm_mm->mmap_sem)) {
915 locked_vma = (vma->vm_flags & VM_LOCKED);
916 if (!locked_vma)
917 up_read(&vma->vm_mm->mmap_sem); /* don't need it */
918 }
Hugh Dickinsc0718802005-10-29 18:16:31 -0700919
920 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700921
Hugh Dickins365e9c872005-10-29 18:16:18 -0700922 /* Update high watermark before we lower rss */
923 update_hiwater_rss(mm);
924
Hugh Dickinsc0718802005-10-29 18:16:31 -0700925 for (; address < end; pte++, address += PAGE_SIZE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700926 if (!pte_present(*pte))
927 continue;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800928 page = vm_normal_page(vma, address, *pte);
929 BUG_ON(!page || PageAnon(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700930
Nick Pigginb291f002008-10-18 20:26:44 -0700931 if (locked_vma) {
932 mlock_vma_page(page); /* no-op if already mlocked */
933 if (page == check_page)
934 ret = SWAP_MLOCK;
935 continue; /* don't unmap */
936 }
937
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700938 if (ptep_clear_flush_young_notify(vma, address, pte))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700939 continue;
940
941 /* Nuke the page table entry. */
Ben Collinseca35132005-11-29 11:45:26 -0800942 flush_cache_page(vma, address, pte_pfn(*pte));
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700943 pteval = ptep_clear_flush_notify(vma, address, pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700944
945 /* If nonlinear, store the file page offset in the pte. */
946 if (page->index != linear_page_index(vma, address))
947 set_pte_at(mm, address, pte, pgoff_to_pte(page->index));
948
949 /* Move the dirty bit to the physical page now the pte is gone. */
950 if (pte_dirty(pteval))
951 set_page_dirty(page);
952
Hugh Dickinsedc315f2009-01-06 14:40:11 -0800953 page_remove_rmap(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700954 page_cache_release(page);
Hugh Dickins42946212005-10-29 18:16:05 -0700955 dec_mm_counter(mm, file_rss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700956 (*mapcount)--;
957 }
Hugh Dickinsc0718802005-10-29 18:16:31 -0700958 pte_unmap_unlock(pte - 1, ptl);
Nick Pigginb291f002008-10-18 20:26:44 -0700959 if (locked_vma)
960 up_read(&vma->vm_mm->mmap_sem);
961 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700962}
963
Nick Pigginb291f002008-10-18 20:26:44 -0700964/*
965 * common handling for pages mapped in VM_LOCKED vmas
966 */
967static int try_to_mlock_page(struct page *page, struct vm_area_struct *vma)
968{
969 int mlocked = 0;
970
971 if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
972 if (vma->vm_flags & VM_LOCKED) {
973 mlock_vma_page(page);
974 mlocked++; /* really mlocked the page */
975 }
976 up_read(&vma->vm_mm->mmap_sem);
977 }
978 return mlocked;
979}
980
981/**
982 * try_to_unmap_anon - unmap or unlock anonymous page using the object-based
983 * rmap method
984 * @page: the page to unmap/unlock
985 * @unlock: request for unlock rather than unmap [unlikely]
986 * @migration: unmapping for migration - ignored if @unlock
987 *
988 * Find all the mappings of a page using the mapping pointer and the vma chains
989 * contained in the anon_vma struct it points to.
990 *
991 * This function is only called from try_to_unmap/try_to_munlock for
992 * anonymous pages.
993 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
994 * where the page was found will be held for write. So, we won't recheck
995 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
996 * 'LOCKED.
997 */
998static int try_to_unmap_anon(struct page *page, int unlock, int migration)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700999{
1000 struct anon_vma *anon_vma;
1001 struct vm_area_struct *vma;
Nick Pigginb291f002008-10-18 20:26:44 -07001002 unsigned int mlocked = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001003 int ret = SWAP_AGAIN;
1004
Nick Pigginb291f002008-10-18 20:26:44 -07001005 if (MLOCK_PAGES && unlikely(unlock))
1006 ret = SWAP_SUCCESS; /* default for try_to_munlock() */
1007
Linus Torvalds1da177e2005-04-16 15:20:36 -07001008 anon_vma = page_lock_anon_vma(page);
1009 if (!anon_vma)
1010 return ret;
1011
1012 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
Nick Pigginb291f002008-10-18 20:26:44 -07001013 if (MLOCK_PAGES && unlikely(unlock)) {
1014 if (!((vma->vm_flags & VM_LOCKED) &&
1015 page_mapped_in_vma(page, vma)))
1016 continue; /* must visit all unlocked vmas */
1017 ret = SWAP_MLOCK; /* saw at least one mlocked vma */
1018 } else {
1019 ret = try_to_unmap_one(page, vma, migration);
1020 if (ret == SWAP_FAIL || !page_mapped(page))
1021 break;
1022 }
1023 if (ret == SWAP_MLOCK) {
1024 mlocked = try_to_mlock_page(page, vma);
1025 if (mlocked)
1026 break; /* stop if actually mlocked page */
1027 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001028 }
Oleg Nesterov34bbd702007-02-28 20:13:49 -08001029
1030 page_unlock_anon_vma(anon_vma);
Nick Pigginb291f002008-10-18 20:26:44 -07001031
1032 if (mlocked)
1033 ret = SWAP_MLOCK; /* actually mlocked the page */
1034 else if (ret == SWAP_MLOCK)
1035 ret = SWAP_AGAIN; /* saw VM_LOCKED vma */
1036
Linus Torvalds1da177e2005-04-16 15:20:36 -07001037 return ret;
1038}
1039
1040/**
Nick Pigginb291f002008-10-18 20:26:44 -07001041 * try_to_unmap_file - unmap/unlock file page using the object-based rmap method
1042 * @page: the page to unmap/unlock
1043 * @unlock: request for unlock rather than unmap [unlikely]
1044 * @migration: unmapping for migration - ignored if @unlock
Linus Torvalds1da177e2005-04-16 15:20:36 -07001045 *
1046 * Find all the mappings of a page using the mapping pointer and the vma chains
1047 * contained in the address_space struct it points to.
1048 *
Nick Pigginb291f002008-10-18 20:26:44 -07001049 * This function is only called from try_to_unmap/try_to_munlock for
1050 * object-based pages.
1051 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1052 * where the page was found will be held for write. So, we won't recheck
1053 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1054 * 'LOCKED.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001055 */
Nick Pigginb291f002008-10-18 20:26:44 -07001056static int try_to_unmap_file(struct page *page, int unlock, int migration)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001057{
1058 struct address_space *mapping = page->mapping;
1059 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1060 struct vm_area_struct *vma;
1061 struct prio_tree_iter iter;
1062 int ret = SWAP_AGAIN;
1063 unsigned long cursor;
1064 unsigned long max_nl_cursor = 0;
1065 unsigned long max_nl_size = 0;
1066 unsigned int mapcount;
Nick Pigginb291f002008-10-18 20:26:44 -07001067 unsigned int mlocked = 0;
1068
1069 if (MLOCK_PAGES && unlikely(unlock))
1070 ret = SWAP_SUCCESS; /* default for try_to_munlock() */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001071
1072 spin_lock(&mapping->i_mmap_lock);
1073 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
Nick Pigginb291f002008-10-18 20:26:44 -07001074 if (MLOCK_PAGES && unlikely(unlock)) {
MinChan Kim508b9f82009-02-11 13:04:27 -08001075 if (!((vma->vm_flags & VM_LOCKED) &&
1076 page_mapped_in_vma(page, vma)))
Nick Pigginb291f002008-10-18 20:26:44 -07001077 continue; /* must visit all vmas */
1078 ret = SWAP_MLOCK;
1079 } else {
1080 ret = try_to_unmap_one(page, vma, migration);
1081 if (ret == SWAP_FAIL || !page_mapped(page))
1082 goto out;
1083 }
1084 if (ret == SWAP_MLOCK) {
1085 mlocked = try_to_mlock_page(page, vma);
1086 if (mlocked)
1087 break; /* stop if actually mlocked page */
1088 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001089 }
1090
Nick Pigginb291f002008-10-18 20:26:44 -07001091 if (mlocked)
1092 goto out;
1093
Linus Torvalds1da177e2005-04-16 15:20:36 -07001094 if (list_empty(&mapping->i_mmap_nonlinear))
1095 goto out;
1096
1097 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1098 shared.vm_set.list) {
Nick Pigginb291f002008-10-18 20:26:44 -07001099 if (MLOCK_PAGES && unlikely(unlock)) {
1100 if (!(vma->vm_flags & VM_LOCKED))
1101 continue; /* must visit all vmas */
1102 ret = SWAP_MLOCK; /* leave mlocked == 0 */
1103 goto out; /* no need to look further */
1104 }
1105 if (!MLOCK_PAGES && !migration && (vma->vm_flags & VM_LOCKED))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001106 continue;
1107 cursor = (unsigned long) vma->vm_private_data;
1108 if (cursor > max_nl_cursor)
1109 max_nl_cursor = cursor;
1110 cursor = vma->vm_end - vma->vm_start;
1111 if (cursor > max_nl_size)
1112 max_nl_size = cursor;
1113 }
1114
Nick Pigginb291f002008-10-18 20:26:44 -07001115 if (max_nl_size == 0) { /* all nonlinears locked or reserved ? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001116 ret = SWAP_FAIL;
1117 goto out;
1118 }
1119
1120 /*
1121 * We don't try to search for this page in the nonlinear vmas,
1122 * and page_referenced wouldn't have found it anyway. Instead
1123 * just walk the nonlinear vmas trying to age and unmap some.
1124 * The mapcount of the page we came in with is irrelevant,
1125 * but even so use it as a guide to how hard we should try?
1126 */
1127 mapcount = page_mapcount(page);
1128 if (!mapcount)
1129 goto out;
1130 cond_resched_lock(&mapping->i_mmap_lock);
1131
1132 max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
1133 if (max_nl_cursor == 0)
1134 max_nl_cursor = CLUSTER_SIZE;
1135
1136 do {
1137 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1138 shared.vm_set.list) {
Nick Pigginb291f002008-10-18 20:26:44 -07001139 if (!MLOCK_PAGES && !migration &&
1140 (vma->vm_flags & VM_LOCKED))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001141 continue;
1142 cursor = (unsigned long) vma->vm_private_data;
Hugh Dickins839b9682005-09-03 15:54:43 -07001143 while ( cursor < max_nl_cursor &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07001144 cursor < vma->vm_end - vma->vm_start) {
Nick Pigginb291f002008-10-18 20:26:44 -07001145 ret = try_to_unmap_cluster(cursor, &mapcount,
1146 vma, page);
1147 if (ret == SWAP_MLOCK)
1148 mlocked = 2; /* to return below */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001149 cursor += CLUSTER_SIZE;
1150 vma->vm_private_data = (void *) cursor;
1151 if ((int)mapcount <= 0)
1152 goto out;
1153 }
1154 vma->vm_private_data = (void *) max_nl_cursor;
1155 }
1156 cond_resched_lock(&mapping->i_mmap_lock);
1157 max_nl_cursor += CLUSTER_SIZE;
1158 } while (max_nl_cursor <= max_nl_size);
1159
1160 /*
1161 * Don't loop forever (perhaps all the remaining pages are
1162 * in locked vmas). Reset cursor on all unreserved nonlinear
1163 * vmas, now forgetting on which ones it had fallen behind.
1164 */
Hugh Dickins101d2be2005-11-21 21:32:16 -08001165 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1166 vma->vm_private_data = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001167out:
1168 spin_unlock(&mapping->i_mmap_lock);
Nick Pigginb291f002008-10-18 20:26:44 -07001169 if (mlocked)
1170 ret = SWAP_MLOCK; /* actually mlocked the page */
1171 else if (ret == SWAP_MLOCK)
1172 ret = SWAP_AGAIN; /* saw VM_LOCKED vma */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001173 return ret;
1174}
1175
1176/**
1177 * try_to_unmap - try to remove all page table mappings to a page
1178 * @page: the page to get unmapped
Randy Dunlap43d8eac2008-03-19 17:00:43 -07001179 * @migration: migration flag
Linus Torvalds1da177e2005-04-16 15:20:36 -07001180 *
1181 * Tries to remove all the page table entries which are mapping this
1182 * page, used in the pageout path. Caller must hold the page lock.
1183 * Return values are:
1184 *
1185 * SWAP_SUCCESS - we succeeded in removing all mappings
1186 * SWAP_AGAIN - we missed a mapping, try again later
1187 * SWAP_FAIL - the page is unswappable
Nick Pigginb291f002008-10-18 20:26:44 -07001188 * SWAP_MLOCK - page is mlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001189 */
Christoph Lameter73523492006-06-23 02:03:27 -07001190int try_to_unmap(struct page *page, int migration)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001191{
1192 int ret;
1193
Linus Torvalds1da177e2005-04-16 15:20:36 -07001194 BUG_ON(!PageLocked(page));
1195
1196 if (PageAnon(page))
Nick Pigginb291f002008-10-18 20:26:44 -07001197 ret = try_to_unmap_anon(page, 0, migration);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001198 else
Nick Pigginb291f002008-10-18 20:26:44 -07001199 ret = try_to_unmap_file(page, 0, migration);
1200 if (ret != SWAP_MLOCK && !page_mapped(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001201 ret = SWAP_SUCCESS;
1202 return ret;
1203}
Nikita Danilov81b40822005-05-01 08:58:36 -07001204
Nick Pigginb291f002008-10-18 20:26:44 -07001205#ifdef CONFIG_UNEVICTABLE_LRU
1206/**
1207 * try_to_munlock - try to munlock a page
1208 * @page: the page to be munlocked
1209 *
1210 * Called from munlock code. Checks all of the VMAs mapping the page
1211 * to make sure nobody else has this page mlocked. The page will be
1212 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1213 *
1214 * Return values are:
1215 *
1216 * SWAP_SUCCESS - no vma's holding page mlocked.
1217 * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem
1218 * SWAP_MLOCK - page is now mlocked.
1219 */
1220int try_to_munlock(struct page *page)
1221{
1222 VM_BUG_ON(!PageLocked(page) || PageLRU(page));
1223
1224 if (PageAnon(page))
1225 return try_to_unmap_anon(page, 1, 0);
1226 else
1227 return try_to_unmap_file(page, 1, 0);
1228}
1229#endif