blob: 40158b59729ec3f6dac4390ecd5a57db7c719b52 [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)
Linus Torvalds1da177e2005-04-16 15:20:36 -070024 * inode->i_alloc_sem
25 *
26 * When a page fault occurs in writing from user to file, down_read
Jes Sorensen1b1dcc12006-01-09 15:59:24 -080027 * of mmap_sem nests within i_mutex; in sys_msync, i_mutex nests within
28 * down_read of mmap_sem; i_mutex and down_write of mmap_sem are never
29 * taken together; in truncation, i_mutex is taken outermost.
Linus Torvalds1da177e2005-04-16 15:20:36 -070030 *
31 * mm->mmap_sem
32 * page->flags PG_locked (lock_page)
33 * mapping->i_mmap_lock
34 * anon_vma->lock
Hugh Dickinsb8072f02005-10-29 18:16:41 -070035 * mm->page_table_lock or pte_lock
Nick Piggin053837f2006-01-18 17:42:27 -080036 * zone->lru_lock (in mark_page_accessed, isolate_lru_page)
Hugh Dickins5d337b92005-09-03 15:54:41 -070037 * swap_lock (in swap_duplicate, swap_info_get)
Linus Torvalds1da177e2005-04-16 15:20:36 -070038 * mmlist_lock (in mmput, drain_mmlist and others)
Linus Torvalds1da177e2005-04-16 15:20:36 -070039 * mapping->private_lock (in __set_page_dirty_buffers)
40 * inode_lock (in set_page_dirty's __mark_inode_dirty)
41 * sb_lock (within inode_lock in fs/fs-writeback.c)
42 * mapping->tree_lock (widely used, in set_page_dirty,
43 * in arch-dependent flush_dcache_mmap_lock,
44 * within inode_lock in __sync_single_inode)
45 */
46
47#include <linux/mm.h>
48#include <linux/pagemap.h>
49#include <linux/swap.h>
50#include <linux/swapops.h>
51#include <linux/slab.h>
52#include <linux/init.h>
53#include <linux/rmap.h>
54#include <linux/rcupdate.h>
Christoph Lametera48d07a2006-02-01 03:05:38 -080055#include <linux/module.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070056
57#include <asm/tlbflush.h>
58
Pekka Enbergfcc234f2006-03-22 00:08:13 -080059struct kmem_cache *anon_vma_cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -070060
61static inline void validate_anon_vma(struct vm_area_struct *find_vma)
62{
Nick Pigginb7ab7952006-03-22 00:08:42 -080063#ifdef CONFIG_DEBUG_VM
Linus Torvalds1da177e2005-04-16 15:20:36 -070064 struct anon_vma *anon_vma = find_vma->anon_vma;
65 struct vm_area_struct *vma;
66 unsigned int mapcount = 0;
67 int found = 0;
68
69 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
70 mapcount++;
71 BUG_ON(mapcount > 100000);
72 if (vma == find_vma)
73 found = 1;
74 }
75 BUG_ON(!found);
76#endif
77}
78
79/* This must be called under the mmap_sem. */
80int anon_vma_prepare(struct vm_area_struct *vma)
81{
82 struct anon_vma *anon_vma = vma->anon_vma;
83
84 might_sleep();
85 if (unlikely(!anon_vma)) {
86 struct mm_struct *mm = vma->vm_mm;
87 struct anon_vma *allocated, *locked;
88
89 anon_vma = find_mergeable_anon_vma(vma);
90 if (anon_vma) {
91 allocated = NULL;
92 locked = anon_vma;
93 spin_lock(&locked->lock);
94 } else {
95 anon_vma = anon_vma_alloc();
96 if (unlikely(!anon_vma))
97 return -ENOMEM;
98 allocated = anon_vma;
99 locked = NULL;
100 }
101
102 /* page_table_lock to protect against threads */
103 spin_lock(&mm->page_table_lock);
104 if (likely(!vma->anon_vma)) {
105 vma->anon_vma = anon_vma;
Christoph Lameter06972122006-06-23 02:03:35 -0700106 list_add_tail(&vma->anon_vma_node, &anon_vma->head);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700107 allocated = NULL;
108 }
109 spin_unlock(&mm->page_table_lock);
110
111 if (locked)
112 spin_unlock(&locked->lock);
113 if (unlikely(allocated))
114 anon_vma_free(allocated);
115 }
116 return 0;
117}
118
119void __anon_vma_merge(struct vm_area_struct *vma, struct vm_area_struct *next)
120{
121 BUG_ON(vma->anon_vma != next->anon_vma);
122 list_del(&next->anon_vma_node);
123}
124
125void __anon_vma_link(struct vm_area_struct *vma)
126{
127 struct anon_vma *anon_vma = vma->anon_vma;
128
129 if (anon_vma) {
Christoph Lameter06972122006-06-23 02:03:35 -0700130 list_add_tail(&vma->anon_vma_node, &anon_vma->head);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700131 validate_anon_vma(vma);
132 }
133}
134
135void anon_vma_link(struct vm_area_struct *vma)
136{
137 struct anon_vma *anon_vma = vma->anon_vma;
138
139 if (anon_vma) {
140 spin_lock(&anon_vma->lock);
Christoph Lameter06972122006-06-23 02:03:35 -0700141 list_add_tail(&vma->anon_vma_node, &anon_vma->head);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700142 validate_anon_vma(vma);
143 spin_unlock(&anon_vma->lock);
144 }
145}
146
147void anon_vma_unlink(struct vm_area_struct *vma)
148{
149 struct anon_vma *anon_vma = vma->anon_vma;
150 int empty;
151
152 if (!anon_vma)
153 return;
154
155 spin_lock(&anon_vma->lock);
156 validate_anon_vma(vma);
157 list_del(&vma->anon_vma_node);
158
159 /* We must garbage collect the anon_vma if it's empty */
160 empty = list_empty(&anon_vma->head);
161 spin_unlock(&anon_vma->lock);
162
163 if (empty)
164 anon_vma_free(anon_vma);
165}
166
Pekka Enbergfcc234f2006-03-22 00:08:13 -0800167static void anon_vma_ctor(void *data, struct kmem_cache *cachep,
168 unsigned long flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700169{
170 if ((flags & (SLAB_CTOR_VERIFY|SLAB_CTOR_CONSTRUCTOR)) ==
171 SLAB_CTOR_CONSTRUCTOR) {
172 struct anon_vma *anon_vma = data;
173
174 spin_lock_init(&anon_vma->lock);
175 INIT_LIST_HEAD(&anon_vma->head);
176 }
177}
178
179void __init anon_vma_init(void)
180{
181 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
182 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor, NULL);
183}
184
185/*
186 * Getting a lock on a stable anon_vma from a page off the LRU is
187 * tricky: page_lock_anon_vma rely on RCU to guard against the races.
188 */
189static struct anon_vma *page_lock_anon_vma(struct page *page)
190{
191 struct anon_vma *anon_vma = NULL;
192 unsigned long anon_mapping;
193
194 rcu_read_lock();
195 anon_mapping = (unsigned long) page->mapping;
196 if (!(anon_mapping & PAGE_MAPPING_ANON))
197 goto out;
198 if (!page_mapped(page))
199 goto out;
200
201 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
202 spin_lock(&anon_vma->lock);
203out:
204 rcu_read_unlock();
205 return anon_vma;
206}
207
208/*
209 * At what user virtual address is page expected in vma?
210 */
211static inline unsigned long
212vma_address(struct page *page, struct vm_area_struct *vma)
213{
214 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
215 unsigned long address;
216
217 address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
218 if (unlikely(address < vma->vm_start || address >= vma->vm_end)) {
219 /* page should be within any vma from prio_tree_next */
220 BUG_ON(!PageAnon(page));
221 return -EFAULT;
222 }
223 return address;
224}
225
226/*
227 * At what user virtual address is page expected in vma? checking that the
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800228 * page matches the vma: currently only used on anon pages, by unuse_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700229 */
230unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
231{
232 if (PageAnon(page)) {
233 if ((void *)vma->anon_vma !=
234 (void *)page->mapping - PAGE_MAPPING_ANON)
235 return -EFAULT;
236 } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) {
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800237 if (!vma->vm_file ||
238 vma->vm_file->f_mapping != page->mapping)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700239 return -EFAULT;
240 } else
241 return -EFAULT;
242 return vma_address(page, vma);
243}
244
245/*
Nikita Danilov81b40822005-05-01 08:58:36 -0700246 * Check that @page is mapped at @address into @mm.
247 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -0700248 * On success returns with pte mapped and locked.
Nikita Danilov81b40822005-05-01 08:58:36 -0700249 */
Carsten Otteceffc072005-06-23 22:05:25 -0700250pte_t *page_check_address(struct page *page, struct mm_struct *mm,
Hugh Dickinsc0718802005-10-29 18:16:31 -0700251 unsigned long address, spinlock_t **ptlp)
Nikita Danilov81b40822005-05-01 08:58:36 -0700252{
253 pgd_t *pgd;
254 pud_t *pud;
255 pmd_t *pmd;
256 pte_t *pte;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700257 spinlock_t *ptl;
Nikita Danilov81b40822005-05-01 08:58:36 -0700258
Nikita Danilov81b40822005-05-01 08:58:36 -0700259 pgd = pgd_offset(mm, address);
Hugh Dickinsc0718802005-10-29 18:16:31 -0700260 if (!pgd_present(*pgd))
261 return NULL;
262
263 pud = pud_offset(pgd, address);
264 if (!pud_present(*pud))
265 return NULL;
266
267 pmd = pmd_offset(pud, address);
268 if (!pmd_present(*pmd))
269 return NULL;
270
271 pte = pte_offset_map(pmd, address);
272 /* Make a quick check before getting the lock */
273 if (!pte_present(*pte)) {
274 pte_unmap(pte);
275 return NULL;
Nikita Danilov81b40822005-05-01 08:58:36 -0700276 }
Hugh Dickinsc0718802005-10-29 18:16:31 -0700277
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700278 ptl = pte_lockptr(mm, pmd);
Hugh Dickinsc0718802005-10-29 18:16:31 -0700279 spin_lock(ptl);
280 if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
281 *ptlp = ptl;
282 return pte;
283 }
284 pte_unmap_unlock(pte, ptl);
285 return NULL;
Nikita Danilov81b40822005-05-01 08:58:36 -0700286}
287
288/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700289 * Subfunctions of page_referenced: page_referenced_one called
290 * repeatedly from either page_referenced_anon or page_referenced_file.
291 */
292static int page_referenced_one(struct page *page,
Rik van Rielf7b7fd82005-11-28 13:44:07 -0800293 struct vm_area_struct *vma, unsigned int *mapcount)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700294{
295 struct mm_struct *mm = vma->vm_mm;
296 unsigned long address;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700297 pte_t *pte;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700298 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700299 int referenced = 0;
300
Linus Torvalds1da177e2005-04-16 15:20:36 -0700301 address = vma_address(page, vma);
302 if (address == -EFAULT)
303 goto out;
304
Hugh Dickinsc0718802005-10-29 18:16:31 -0700305 pte = page_check_address(page, mm, address, &ptl);
306 if (!pte)
307 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700308
Hugh Dickinsc0718802005-10-29 18:16:31 -0700309 if (ptep_clear_flush_young(vma, address, pte))
310 referenced++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700311
Hugh Dickinsc0718802005-10-29 18:16:31 -0700312 /* Pretend the page is referenced if the task has the
313 swap token and is in the middle of a page fault. */
Rik van Rielf7b7fd82005-11-28 13:44:07 -0800314 if (mm != current->mm && has_swap_token(mm) &&
Hugh Dickinsc0718802005-10-29 18:16:31 -0700315 rwsem_is_locked(&mm->mmap_sem))
316 referenced++;
317
318 (*mapcount)--;
319 pte_unmap_unlock(pte, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700320out:
321 return referenced;
322}
323
Rik van Rielf7b7fd82005-11-28 13:44:07 -0800324static int page_referenced_anon(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700325{
326 unsigned int mapcount;
327 struct anon_vma *anon_vma;
328 struct vm_area_struct *vma;
329 int referenced = 0;
330
331 anon_vma = page_lock_anon_vma(page);
332 if (!anon_vma)
333 return referenced;
334
335 mapcount = page_mapcount(page);
336 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
Rik van Rielf7b7fd82005-11-28 13:44:07 -0800337 referenced += page_referenced_one(page, vma, &mapcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700338 if (!mapcount)
339 break;
340 }
341 spin_unlock(&anon_vma->lock);
342 return referenced;
343}
344
345/**
346 * page_referenced_file - referenced check for object-based rmap
347 * @page: the page we're checking references on.
348 *
349 * For an object-based mapped page, find all the places it is mapped and
350 * check/clear the referenced flag. This is done by following the page->mapping
351 * pointer, then walking the chain of vmas it holds. It returns the number
352 * of references it found.
353 *
354 * This function is only called from page_referenced for object-based pages.
355 */
Rik van Rielf7b7fd82005-11-28 13:44:07 -0800356static int page_referenced_file(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700357{
358 unsigned int mapcount;
359 struct address_space *mapping = page->mapping;
360 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
361 struct vm_area_struct *vma;
362 struct prio_tree_iter iter;
363 int referenced = 0;
364
365 /*
366 * The caller's checks on page->mapping and !PageAnon have made
367 * sure that this is a file page: the check for page->mapping
368 * excludes the case just before it gets set on an anon page.
369 */
370 BUG_ON(PageAnon(page));
371
372 /*
373 * The page lock not only makes sure that page->mapping cannot
374 * suddenly be NULLified by truncation, it makes sure that the
375 * structure at mapping cannot be freed and reused yet,
376 * so we can safely take mapping->i_mmap_lock.
377 */
378 BUG_ON(!PageLocked(page));
379
380 spin_lock(&mapping->i_mmap_lock);
381
382 /*
383 * i_mmap_lock does not stabilize mapcount at all, but mapcount
384 * is more likely to be accurate if we note it after spinning.
385 */
386 mapcount = page_mapcount(page);
387
388 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
389 if ((vma->vm_flags & (VM_LOCKED|VM_MAYSHARE))
390 == (VM_LOCKED|VM_MAYSHARE)) {
391 referenced++;
392 break;
393 }
Rik van Rielf7b7fd82005-11-28 13:44:07 -0800394 referenced += page_referenced_one(page, vma, &mapcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700395 if (!mapcount)
396 break;
397 }
398
399 spin_unlock(&mapping->i_mmap_lock);
400 return referenced;
401}
402
403/**
404 * page_referenced - test if the page was referenced
405 * @page: the page to test
406 * @is_locked: caller holds lock on the page
407 *
408 * Quick test_and_clear_referenced for all mappings to a page,
409 * returns the number of ptes which referenced the page.
410 */
Rik van Rielf7b7fd82005-11-28 13:44:07 -0800411int page_referenced(struct page *page, int is_locked)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700412{
413 int referenced = 0;
414
Linus Torvalds1da177e2005-04-16 15:20:36 -0700415 if (page_test_and_clear_young(page))
416 referenced++;
417
418 if (TestClearPageReferenced(page))
419 referenced++;
420
421 if (page_mapped(page) && page->mapping) {
422 if (PageAnon(page))
Rik van Rielf7b7fd82005-11-28 13:44:07 -0800423 referenced += page_referenced_anon(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700424 else if (is_locked)
Rik van Rielf7b7fd82005-11-28 13:44:07 -0800425 referenced += page_referenced_file(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700426 else if (TestSetPageLocked(page))
427 referenced++;
428 else {
429 if (page->mapping)
Rik van Rielf7b7fd82005-11-28 13:44:07 -0800430 referenced += page_referenced_file(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700431 unlock_page(page);
432 }
433 }
434 return referenced;
435}
436
437/**
Nick Piggin9617d952006-01-06 00:11:12 -0800438 * page_set_anon_rmap - setup new anonymous rmap
439 * @page: the page to add the mapping to
440 * @vma: the vm area in which the mapping is added
441 * @address: the user virtual address mapped
442 */
443static void __page_set_anon_rmap(struct page *page,
444 struct vm_area_struct *vma, unsigned long address)
445{
446 struct anon_vma *anon_vma = vma->anon_vma;
447
448 BUG_ON(!anon_vma);
449 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
450 page->mapping = (struct address_space *) anon_vma;
451
452 page->index = linear_page_index(vma, address);
453
Nick Piggina74609f2006-01-06 00:11:20 -0800454 /*
455 * nr_mapped state can be updated without turning off
456 * interrupts because it is not modified via interrupt.
457 */
Christoph Lameterf3dbd342006-06-30 01:55:36 -0700458 __inc_zone_page_state(page, NR_ANON_PAGES);
Nick Piggin9617d952006-01-06 00:11:12 -0800459}
460
461/**
Linus Torvalds1da177e2005-04-16 15:20:36 -0700462 * page_add_anon_rmap - add pte mapping to an anonymous page
463 * @page: the page to add the mapping to
464 * @vma: the vm area in which the mapping is added
465 * @address: the user virtual address mapped
466 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -0700467 * The caller needs to hold the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700468 */
469void page_add_anon_rmap(struct page *page,
470 struct vm_area_struct *vma, unsigned long address)
471{
Nick Piggin9617d952006-01-06 00:11:12 -0800472 if (atomic_inc_and_test(&page->_mapcount))
473 __page_set_anon_rmap(page, vma, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700474 /* else checking page index and mapping is racy */
475}
476
Nick Piggin9617d952006-01-06 00:11:12 -0800477/*
478 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
479 * @page: the page to add the mapping to
480 * @vma: the vm area in which the mapping is added
481 * @address: the user virtual address mapped
482 *
483 * Same as page_add_anon_rmap but must only be called on *new* pages.
484 * This means the inc-and-test can be bypassed.
485 */
486void page_add_new_anon_rmap(struct page *page,
487 struct vm_area_struct *vma, unsigned long address)
488{
489 atomic_set(&page->_mapcount, 0); /* elevate count by 1 (starts at -1) */
490 __page_set_anon_rmap(page, vma, address);
491}
492
Linus Torvalds1da177e2005-04-16 15:20:36 -0700493/**
494 * page_add_file_rmap - add pte mapping to a file page
495 * @page: the page to add the mapping to
496 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -0700497 * The caller needs to hold the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700498 */
499void page_add_file_rmap(struct page *page)
500{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700501 if (atomic_inc_and_test(&page->_mapcount))
Christoph Lameter65ba55f2006-06-30 01:55:34 -0700502 __inc_zone_page_state(page, NR_FILE_MAPPED);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700503}
504
505/**
506 * page_remove_rmap - take down pte mapping from a page
507 * @page: page to remove mapping from
508 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -0700509 * The caller needs to hold the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700510 */
511void page_remove_rmap(struct page *page)
512{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700513 if (atomic_add_negative(-1, &page->_mapcount)) {
Nick Pigginb7ab7952006-03-22 00:08:42 -0800514#ifdef CONFIG_DEBUG_VM
515 if (unlikely(page_mapcount(page) < 0)) {
Dave Jonesef2bf0d2006-01-08 01:01:00 -0800516 printk (KERN_EMERG "Eeek! page_mapcount(page) went negative! (%d)\n", page_mapcount(page));
517 printk (KERN_EMERG " page->flags = %lx\n", page->flags);
518 printk (KERN_EMERG " page->count = %x\n", page_count(page));
519 printk (KERN_EMERG " page->mapping = %p\n", page->mapping);
520 }
Nick Pigginb7ab7952006-03-22 00:08:42 -0800521#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700522 BUG_ON(page_mapcount(page) < 0);
523 /*
524 * It would be tidy to reset the PageAnon mapping here,
525 * but that might overwrite a racing page_add_anon_rmap
526 * which increments mapcount after us but sets mapping
527 * before us: so leave the reset to free_hot_cold_page,
528 * and remember that it's only reliable while mapped.
529 * Leaving it set also helps swapoff to reinstate ptes
530 * faster for those pages still in swapcache.
531 */
532 if (page_test_and_clear_dirty(page))
533 set_page_dirty(page);
Christoph Lameterf3dbd342006-06-30 01:55:36 -0700534 __dec_zone_page_state(page,
535 PageAnon(page) ? NR_ANON_PAGES : NR_FILE_MAPPED);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700536 }
537}
538
539/*
540 * Subfunctions of try_to_unmap: try_to_unmap_one called
541 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
542 */
Christoph Lametera48d07a2006-02-01 03:05:38 -0800543static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
Christoph Lameter73523492006-06-23 02:03:27 -0700544 int migration)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700545{
546 struct mm_struct *mm = vma->vm_mm;
547 unsigned long address;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700548 pte_t *pte;
549 pte_t pteval;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700550 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700551 int ret = SWAP_AGAIN;
552
Linus Torvalds1da177e2005-04-16 15:20:36 -0700553 address = vma_address(page, vma);
554 if (address == -EFAULT)
555 goto out;
556
Hugh Dickinsc0718802005-10-29 18:16:31 -0700557 pte = page_check_address(page, mm, address, &ptl);
558 if (!pte)
Nikita Danilov81b40822005-05-01 08:58:36 -0700559 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700560
561 /*
562 * If the page is mlock()d, we cannot swap it out.
563 * If it's recently referenced (perhaps page_referenced
564 * skipped over this mm) then we should reactivate it.
565 */
Christoph Lametere6a15302006-06-25 05:46:49 -0700566 if (!migration && ((vma->vm_flags & VM_LOCKED) ||
567 (ptep_clear_flush_young(vma, address, pte)))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700568 ret = SWAP_FAIL;
569 goto out_unmap;
570 }
571
Linus Torvalds1da177e2005-04-16 15:20:36 -0700572 /* Nuke the page table entry. */
573 flush_cache_page(vma, address, page_to_pfn(page));
574 pteval = ptep_clear_flush(vma, address, pte);
575
576 /* Move the dirty bit to the physical page now the pte is gone. */
577 if (pte_dirty(pteval))
578 set_page_dirty(page);
579
Hugh Dickins365e9c872005-10-29 18:16:18 -0700580 /* Update high watermark before we lower rss */
581 update_hiwater_rss(mm);
582
Linus Torvalds1da177e2005-04-16 15:20:36 -0700583 if (PageAnon(page)) {
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700584 swp_entry_t entry = { .val = page_private(page) };
Christoph Lameter06972122006-06-23 02:03:35 -0700585
586 if (PageSwapCache(page)) {
587 /*
588 * Store the swap location in the pte.
589 * See handle_pte_fault() ...
590 */
591 swap_duplicate(entry);
592 if (list_empty(&mm->mmlist)) {
593 spin_lock(&mmlist_lock);
594 if (list_empty(&mm->mmlist))
595 list_add(&mm->mmlist, &init_mm.mmlist);
596 spin_unlock(&mmlist_lock);
597 }
Christoph Lameter442c9132006-06-23 02:03:38 -0700598 dec_mm_counter(mm, anon_rss);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700599#ifdef CONFIG_MIGRATION
Christoph Lameter06972122006-06-23 02:03:35 -0700600 } else {
601 /*
602 * Store the pfn of the page in a special migration
603 * pte. do_swap_page() will wait until the migration
604 * pte is removed and then restart fault handling.
605 */
606 BUG_ON(!migration);
607 entry = make_migration_entry(page, pte_write(pteval));
Christoph Lameter04e62a22006-06-23 02:03:38 -0700608#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700609 }
610 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
611 BUG_ON(pte_file(*pte));
Hugh Dickins42946212005-10-29 18:16:05 -0700612 } else
Christoph Lameter04e62a22006-06-23 02:03:38 -0700613#ifdef CONFIG_MIGRATION
614 if (migration) {
615 /* Establish migration entry for a file page */
616 swp_entry_t entry;
617 entry = make_migration_entry(page, pte_write(pteval));
618 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
619 } else
620#endif
Hugh Dickins42946212005-10-29 18:16:05 -0700621 dec_mm_counter(mm, file_rss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700622
Christoph Lameter04e62a22006-06-23 02:03:38 -0700623
Linus Torvalds1da177e2005-04-16 15:20:36 -0700624 page_remove_rmap(page);
625 page_cache_release(page);
626
627out_unmap:
Hugh Dickinsc0718802005-10-29 18:16:31 -0700628 pte_unmap_unlock(pte, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700629out:
630 return ret;
631}
632
633/*
634 * objrmap doesn't work for nonlinear VMAs because the assumption that
635 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
636 * Consequently, given a particular page and its ->index, we cannot locate the
637 * ptes which are mapping that page without an exhaustive linear search.
638 *
639 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
640 * maps the file to which the target page belongs. The ->vm_private_data field
641 * holds the current cursor into that scan. Successive searches will circulate
642 * around the vma's virtual address space.
643 *
644 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
645 * more scanning pressure is placed against them as well. Eventually pages
646 * will become fully unmapped and are eligible for eviction.
647 *
648 * For very sparsely populated VMAs this is a little inefficient - chances are
649 * there there won't be many ptes located within the scan cluster. In this case
650 * maybe we could scan further - to the end of the pte page, perhaps.
651 */
652#define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
653#define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
654
655static void try_to_unmap_cluster(unsigned long cursor,
656 unsigned int *mapcount, struct vm_area_struct *vma)
657{
658 struct mm_struct *mm = vma->vm_mm;
659 pgd_t *pgd;
660 pud_t *pud;
661 pmd_t *pmd;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700662 pte_t *pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700663 pte_t pteval;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700664 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700665 struct page *page;
666 unsigned long address;
667 unsigned long end;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700668
Linus Torvalds1da177e2005-04-16 15:20:36 -0700669 address = (vma->vm_start + cursor) & CLUSTER_MASK;
670 end = address + CLUSTER_SIZE;
671 if (address < vma->vm_start)
672 address = vma->vm_start;
673 if (end > vma->vm_end)
674 end = vma->vm_end;
675
676 pgd = pgd_offset(mm, address);
677 if (!pgd_present(*pgd))
Hugh Dickinsc0718802005-10-29 18:16:31 -0700678 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700679
680 pud = pud_offset(pgd, address);
681 if (!pud_present(*pud))
Hugh Dickinsc0718802005-10-29 18:16:31 -0700682 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700683
684 pmd = pmd_offset(pud, address);
685 if (!pmd_present(*pmd))
Hugh Dickinsc0718802005-10-29 18:16:31 -0700686 return;
687
688 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700689
Hugh Dickins365e9c872005-10-29 18:16:18 -0700690 /* Update high watermark before we lower rss */
691 update_hiwater_rss(mm);
692
Hugh Dickinsc0718802005-10-29 18:16:31 -0700693 for (; address < end; pte++, address += PAGE_SIZE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700694 if (!pte_present(*pte))
695 continue;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800696 page = vm_normal_page(vma, address, *pte);
697 BUG_ON(!page || PageAnon(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700698
699 if (ptep_clear_flush_young(vma, address, pte))
700 continue;
701
702 /* Nuke the page table entry. */
Ben Collinseca35132005-11-29 11:45:26 -0800703 flush_cache_page(vma, address, pte_pfn(*pte));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700704 pteval = ptep_clear_flush(vma, address, pte);
705
706 /* If nonlinear, store the file page offset in the pte. */
707 if (page->index != linear_page_index(vma, address))
708 set_pte_at(mm, address, pte, pgoff_to_pte(page->index));
709
710 /* Move the dirty bit to the physical page now the pte is gone. */
711 if (pte_dirty(pteval))
712 set_page_dirty(page);
713
714 page_remove_rmap(page);
715 page_cache_release(page);
Hugh Dickins42946212005-10-29 18:16:05 -0700716 dec_mm_counter(mm, file_rss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700717 (*mapcount)--;
718 }
Hugh Dickinsc0718802005-10-29 18:16:31 -0700719 pte_unmap_unlock(pte - 1, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700720}
721
Christoph Lameter73523492006-06-23 02:03:27 -0700722static int try_to_unmap_anon(struct page *page, int migration)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700723{
724 struct anon_vma *anon_vma;
725 struct vm_area_struct *vma;
726 int ret = SWAP_AGAIN;
727
728 anon_vma = page_lock_anon_vma(page);
729 if (!anon_vma)
730 return ret;
731
732 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
Christoph Lameter73523492006-06-23 02:03:27 -0700733 ret = try_to_unmap_one(page, vma, migration);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700734 if (ret == SWAP_FAIL || !page_mapped(page))
735 break;
736 }
737 spin_unlock(&anon_vma->lock);
738 return ret;
739}
740
741/**
742 * try_to_unmap_file - unmap file page using the object-based rmap method
743 * @page: the page to unmap
744 *
745 * Find all the mappings of a page using the mapping pointer and the vma chains
746 * contained in the address_space struct it points to.
747 *
748 * This function is only called from try_to_unmap for object-based pages.
749 */
Christoph Lameter73523492006-06-23 02:03:27 -0700750static int try_to_unmap_file(struct page *page, int migration)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700751{
752 struct address_space *mapping = page->mapping;
753 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
754 struct vm_area_struct *vma;
755 struct prio_tree_iter iter;
756 int ret = SWAP_AGAIN;
757 unsigned long cursor;
758 unsigned long max_nl_cursor = 0;
759 unsigned long max_nl_size = 0;
760 unsigned int mapcount;
761
762 spin_lock(&mapping->i_mmap_lock);
763 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
Christoph Lameter73523492006-06-23 02:03:27 -0700764 ret = try_to_unmap_one(page, vma, migration);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700765 if (ret == SWAP_FAIL || !page_mapped(page))
766 goto out;
767 }
768
769 if (list_empty(&mapping->i_mmap_nonlinear))
770 goto out;
771
772 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
773 shared.vm_set.list) {
Christoph Lametere6a15302006-06-25 05:46:49 -0700774 if ((vma->vm_flags & VM_LOCKED) && !migration)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700775 continue;
776 cursor = (unsigned long) vma->vm_private_data;
777 if (cursor > max_nl_cursor)
778 max_nl_cursor = cursor;
779 cursor = vma->vm_end - vma->vm_start;
780 if (cursor > max_nl_size)
781 max_nl_size = cursor;
782 }
783
784 if (max_nl_size == 0) { /* any nonlinears locked or reserved */
785 ret = SWAP_FAIL;
786 goto out;
787 }
788
789 /*
790 * We don't try to search for this page in the nonlinear vmas,
791 * and page_referenced wouldn't have found it anyway. Instead
792 * just walk the nonlinear vmas trying to age and unmap some.
793 * The mapcount of the page we came in with is irrelevant,
794 * but even so use it as a guide to how hard we should try?
795 */
796 mapcount = page_mapcount(page);
797 if (!mapcount)
798 goto out;
799 cond_resched_lock(&mapping->i_mmap_lock);
800
801 max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
802 if (max_nl_cursor == 0)
803 max_nl_cursor = CLUSTER_SIZE;
804
805 do {
806 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
807 shared.vm_set.list) {
Christoph Lametere6a15302006-06-25 05:46:49 -0700808 if ((vma->vm_flags & VM_LOCKED) && !migration)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700809 continue;
810 cursor = (unsigned long) vma->vm_private_data;
Hugh Dickins839b9682005-09-03 15:54:43 -0700811 while ( cursor < max_nl_cursor &&
Linus Torvalds1da177e2005-04-16 15:20:36 -0700812 cursor < vma->vm_end - vma->vm_start) {
813 try_to_unmap_cluster(cursor, &mapcount, vma);
814 cursor += CLUSTER_SIZE;
815 vma->vm_private_data = (void *) cursor;
816 if ((int)mapcount <= 0)
817 goto out;
818 }
819 vma->vm_private_data = (void *) max_nl_cursor;
820 }
821 cond_resched_lock(&mapping->i_mmap_lock);
822 max_nl_cursor += CLUSTER_SIZE;
823 } while (max_nl_cursor <= max_nl_size);
824
825 /*
826 * Don't loop forever (perhaps all the remaining pages are
827 * in locked vmas). Reset cursor on all unreserved nonlinear
828 * vmas, now forgetting on which ones it had fallen behind.
829 */
Hugh Dickins101d2be2005-11-21 21:32:16 -0800830 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
831 vma->vm_private_data = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700832out:
833 spin_unlock(&mapping->i_mmap_lock);
834 return ret;
835}
836
837/**
838 * try_to_unmap - try to remove all page table mappings to a page
839 * @page: the page to get unmapped
840 *
841 * Tries to remove all the page table entries which are mapping this
842 * page, used in the pageout path. Caller must hold the page lock.
843 * Return values are:
844 *
845 * SWAP_SUCCESS - we succeeded in removing all mappings
846 * SWAP_AGAIN - we missed a mapping, try again later
847 * SWAP_FAIL - the page is unswappable
848 */
Christoph Lameter73523492006-06-23 02:03:27 -0700849int try_to_unmap(struct page *page, int migration)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700850{
851 int ret;
852
Linus Torvalds1da177e2005-04-16 15:20:36 -0700853 BUG_ON(!PageLocked(page));
854
855 if (PageAnon(page))
Christoph Lameter73523492006-06-23 02:03:27 -0700856 ret = try_to_unmap_anon(page, migration);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700857 else
Christoph Lameter73523492006-06-23 02:03:27 -0700858 ret = try_to_unmap_file(page, migration);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700859
860 if (!page_mapped(page))
861 ret = SWAP_SUCCESS;
862 return ret;
863}
Nikita Danilov81b40822005-05-01 08:58:36 -0700864