blob: 669acb22b572190ff2dcc6544c968bbd7516859e [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>
Nick Piggin7de6b802006-12-22 01:09:33 -080050#include <linux/kallsyms.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070051
52#include <asm/tlbflush.h>
53
Pekka Enbergfcc234f2006-03-22 00:08:13 -080054struct kmem_cache *anon_vma_cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -070055
56static inline void validate_anon_vma(struct vm_area_struct *find_vma)
57{
Nick Pigginb7ab7952006-03-22 00:08:42 -080058#ifdef CONFIG_DEBUG_VM
Linus Torvalds1da177e2005-04-16 15:20:36 -070059 struct anon_vma *anon_vma = find_vma->anon_vma;
60 struct vm_area_struct *vma;
61 unsigned int mapcount = 0;
62 int found = 0;
63
64 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
65 mapcount++;
66 BUG_ON(mapcount > 100000);
67 if (vma == find_vma)
68 found = 1;
69 }
70 BUG_ON(!found);
71#endif
72}
73
74/* This must be called under the mmap_sem. */
75int anon_vma_prepare(struct vm_area_struct *vma)
76{
77 struct anon_vma *anon_vma = vma->anon_vma;
78
79 might_sleep();
80 if (unlikely(!anon_vma)) {
81 struct mm_struct *mm = vma->vm_mm;
82 struct anon_vma *allocated, *locked;
83
84 anon_vma = find_mergeable_anon_vma(vma);
85 if (anon_vma) {
86 allocated = NULL;
87 locked = anon_vma;
88 spin_lock(&locked->lock);
89 } else {
90 anon_vma = anon_vma_alloc();
91 if (unlikely(!anon_vma))
92 return -ENOMEM;
93 allocated = anon_vma;
94 locked = NULL;
95 }
96
97 /* page_table_lock to protect against threads */
98 spin_lock(&mm->page_table_lock);
99 if (likely(!vma->anon_vma)) {
100 vma->anon_vma = anon_vma;
Christoph Lameter06972122006-06-23 02:03:35 -0700101 list_add_tail(&vma->anon_vma_node, &anon_vma->head);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700102 allocated = NULL;
103 }
104 spin_unlock(&mm->page_table_lock);
105
106 if (locked)
107 spin_unlock(&locked->lock);
108 if (unlikely(allocated))
109 anon_vma_free(allocated);
110 }
111 return 0;
112}
113
114void __anon_vma_merge(struct vm_area_struct *vma, struct vm_area_struct *next)
115{
116 BUG_ON(vma->anon_vma != next->anon_vma);
117 list_del(&next->anon_vma_node);
118}
119
120void __anon_vma_link(struct vm_area_struct *vma)
121{
122 struct anon_vma *anon_vma = vma->anon_vma;
123
124 if (anon_vma) {
Christoph Lameter06972122006-06-23 02:03:35 -0700125 list_add_tail(&vma->anon_vma_node, &anon_vma->head);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700126 validate_anon_vma(vma);
127 }
128}
129
130void anon_vma_link(struct vm_area_struct *vma)
131{
132 struct anon_vma *anon_vma = vma->anon_vma;
133
134 if (anon_vma) {
135 spin_lock(&anon_vma->lock);
Christoph Lameter06972122006-06-23 02:03:35 -0700136 list_add_tail(&vma->anon_vma_node, &anon_vma->head);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700137 validate_anon_vma(vma);
138 spin_unlock(&anon_vma->lock);
139 }
140}
141
142void anon_vma_unlink(struct vm_area_struct *vma)
143{
144 struct anon_vma *anon_vma = vma->anon_vma;
145 int empty;
146
147 if (!anon_vma)
148 return;
149
150 spin_lock(&anon_vma->lock);
151 validate_anon_vma(vma);
152 list_del(&vma->anon_vma_node);
153
154 /* We must garbage collect the anon_vma if it's empty */
155 empty = list_empty(&anon_vma->head);
156 spin_unlock(&anon_vma->lock);
157
158 if (empty)
159 anon_vma_free(anon_vma);
160}
161
Pekka Enbergfcc234f2006-03-22 00:08:13 -0800162static void anon_vma_ctor(void *data, struct kmem_cache *cachep,
163 unsigned long flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700164{
165 if ((flags & (SLAB_CTOR_VERIFY|SLAB_CTOR_CONSTRUCTOR)) ==
166 SLAB_CTOR_CONSTRUCTOR) {
167 struct anon_vma *anon_vma = data;
168
169 spin_lock_init(&anon_vma->lock);
170 INIT_LIST_HEAD(&anon_vma->head);
171 }
172}
173
174void __init anon_vma_init(void)
175{
176 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
177 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor, NULL);
178}
179
180/*
181 * Getting a lock on a stable anon_vma from a page off the LRU is
182 * tricky: page_lock_anon_vma rely on RCU to guard against the races.
183 */
184static struct anon_vma *page_lock_anon_vma(struct page *page)
185{
186 struct anon_vma *anon_vma = NULL;
187 unsigned long anon_mapping;
188
189 rcu_read_lock();
190 anon_mapping = (unsigned long) page->mapping;
191 if (!(anon_mapping & PAGE_MAPPING_ANON))
192 goto out;
193 if (!page_mapped(page))
194 goto out;
195
196 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
197 spin_lock(&anon_vma->lock);
198out:
199 rcu_read_unlock();
200 return anon_vma;
201}
202
203/*
204 * At what user virtual address is page expected in vma?
205 */
206static inline unsigned long
207vma_address(struct page *page, struct vm_area_struct *vma)
208{
209 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
210 unsigned long address;
211
212 address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
213 if (unlikely(address < vma->vm_start || address >= vma->vm_end)) {
214 /* page should be within any vma from prio_tree_next */
215 BUG_ON(!PageAnon(page));
216 return -EFAULT;
217 }
218 return address;
219}
220
221/*
222 * At what user virtual address is page expected in vma? checking that the
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800223 * page matches the vma: currently only used on anon pages, by unuse_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700224 */
225unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
226{
227 if (PageAnon(page)) {
228 if ((void *)vma->anon_vma !=
229 (void *)page->mapping - PAGE_MAPPING_ANON)
230 return -EFAULT;
231 } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) {
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800232 if (!vma->vm_file ||
233 vma->vm_file->f_mapping != page->mapping)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700234 return -EFAULT;
235 } else
236 return -EFAULT;
237 return vma_address(page, vma);
238}
239
240/*
Nikita Danilov81b40822005-05-01 08:58:36 -0700241 * Check that @page is mapped at @address into @mm.
242 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -0700243 * On success returns with pte mapped and locked.
Nikita Danilov81b40822005-05-01 08:58:36 -0700244 */
Carsten Otteceffc072005-06-23 22:05:25 -0700245pte_t *page_check_address(struct page *page, struct mm_struct *mm,
Hugh Dickinsc0718802005-10-29 18:16:31 -0700246 unsigned long address, spinlock_t **ptlp)
Nikita Danilov81b40822005-05-01 08:58:36 -0700247{
248 pgd_t *pgd;
249 pud_t *pud;
250 pmd_t *pmd;
251 pte_t *pte;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700252 spinlock_t *ptl;
Nikita Danilov81b40822005-05-01 08:58:36 -0700253
Nikita Danilov81b40822005-05-01 08:58:36 -0700254 pgd = pgd_offset(mm, address);
Hugh Dickinsc0718802005-10-29 18:16:31 -0700255 if (!pgd_present(*pgd))
256 return NULL;
257
258 pud = pud_offset(pgd, address);
259 if (!pud_present(*pud))
260 return NULL;
261
262 pmd = pmd_offset(pud, address);
263 if (!pmd_present(*pmd))
264 return NULL;
265
266 pte = pte_offset_map(pmd, address);
267 /* Make a quick check before getting the lock */
268 if (!pte_present(*pte)) {
269 pte_unmap(pte);
270 return NULL;
Nikita Danilov81b40822005-05-01 08:58:36 -0700271 }
Hugh Dickinsc0718802005-10-29 18:16:31 -0700272
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700273 ptl = pte_lockptr(mm, pmd);
Hugh Dickinsc0718802005-10-29 18:16:31 -0700274 spin_lock(ptl);
275 if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
276 *ptlp = ptl;
277 return pte;
278 }
279 pte_unmap_unlock(pte, ptl);
280 return NULL;
Nikita Danilov81b40822005-05-01 08:58:36 -0700281}
282
283/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700284 * Subfunctions of page_referenced: page_referenced_one called
285 * repeatedly from either page_referenced_anon or page_referenced_file.
286 */
287static int page_referenced_one(struct page *page,
Rik van Rielf7b7fd82005-11-28 13:44:07 -0800288 struct vm_area_struct *vma, unsigned int *mapcount)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700289{
290 struct mm_struct *mm = vma->vm_mm;
291 unsigned long address;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700292 pte_t *pte;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700293 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700294 int referenced = 0;
295
Linus Torvalds1da177e2005-04-16 15:20:36 -0700296 address = vma_address(page, vma);
297 if (address == -EFAULT)
298 goto out;
299
Hugh Dickinsc0718802005-10-29 18:16:31 -0700300 pte = page_check_address(page, mm, address, &ptl);
301 if (!pte)
302 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700303
Hugh Dickinsc0718802005-10-29 18:16:31 -0700304 if (ptep_clear_flush_young(vma, address, pte))
305 referenced++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700306
Hugh Dickinsc0718802005-10-29 18:16:31 -0700307 /* Pretend the page is referenced if the task has the
308 swap token and is in the middle of a page fault. */
Rik van Rielf7b7fd82005-11-28 13:44:07 -0800309 if (mm != current->mm && has_swap_token(mm) &&
Hugh Dickinsc0718802005-10-29 18:16:31 -0700310 rwsem_is_locked(&mm->mmap_sem))
311 referenced++;
312
313 (*mapcount)--;
314 pte_unmap_unlock(pte, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700315out:
316 return referenced;
317}
318
Rik van Rielf7b7fd82005-11-28 13:44:07 -0800319static int page_referenced_anon(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700320{
321 unsigned int mapcount;
322 struct anon_vma *anon_vma;
323 struct vm_area_struct *vma;
324 int referenced = 0;
325
326 anon_vma = page_lock_anon_vma(page);
327 if (!anon_vma)
328 return referenced;
329
330 mapcount = page_mapcount(page);
331 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
Rik van Rielf7b7fd82005-11-28 13:44:07 -0800332 referenced += page_referenced_one(page, vma, &mapcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700333 if (!mapcount)
334 break;
335 }
336 spin_unlock(&anon_vma->lock);
337 return referenced;
338}
339
340/**
341 * page_referenced_file - referenced check for object-based rmap
342 * @page: the page we're checking references on.
343 *
344 * For an object-based mapped page, find all the places it is mapped and
345 * check/clear the referenced flag. This is done by following the page->mapping
346 * pointer, then walking the chain of vmas it holds. It returns the number
347 * of references it found.
348 *
349 * This function is only called from page_referenced for object-based pages.
350 */
Rik van Rielf7b7fd82005-11-28 13:44:07 -0800351static int page_referenced_file(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700352{
353 unsigned int mapcount;
354 struct address_space *mapping = page->mapping;
355 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
356 struct vm_area_struct *vma;
357 struct prio_tree_iter iter;
358 int referenced = 0;
359
360 /*
361 * The caller's checks on page->mapping and !PageAnon have made
362 * sure that this is a file page: the check for page->mapping
363 * excludes the case just before it gets set on an anon page.
364 */
365 BUG_ON(PageAnon(page));
366
367 /*
368 * The page lock not only makes sure that page->mapping cannot
369 * suddenly be NULLified by truncation, it makes sure that the
370 * structure at mapping cannot be freed and reused yet,
371 * so we can safely take mapping->i_mmap_lock.
372 */
373 BUG_ON(!PageLocked(page));
374
375 spin_lock(&mapping->i_mmap_lock);
376
377 /*
378 * i_mmap_lock does not stabilize mapcount at all, but mapcount
379 * is more likely to be accurate if we note it after spinning.
380 */
381 mapcount = page_mapcount(page);
382
383 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
384 if ((vma->vm_flags & (VM_LOCKED|VM_MAYSHARE))
385 == (VM_LOCKED|VM_MAYSHARE)) {
386 referenced++;
387 break;
388 }
Rik van Rielf7b7fd82005-11-28 13:44:07 -0800389 referenced += page_referenced_one(page, vma, &mapcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700390 if (!mapcount)
391 break;
392 }
393
394 spin_unlock(&mapping->i_mmap_lock);
395 return referenced;
396}
397
398/**
399 * page_referenced - test if the page was referenced
400 * @page: the page to test
401 * @is_locked: caller holds lock on the page
402 *
403 * Quick test_and_clear_referenced for all mappings to a page,
404 * returns the number of ptes which referenced the page.
405 */
Rik van Rielf7b7fd82005-11-28 13:44:07 -0800406int page_referenced(struct page *page, int is_locked)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700407{
408 int referenced = 0;
409
Linus Torvalds1da177e2005-04-16 15:20:36 -0700410 if (page_test_and_clear_young(page))
411 referenced++;
412
413 if (TestClearPageReferenced(page))
414 referenced++;
415
416 if (page_mapped(page) && page->mapping) {
417 if (PageAnon(page))
Rik van Rielf7b7fd82005-11-28 13:44:07 -0800418 referenced += page_referenced_anon(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700419 else if (is_locked)
Rik van Rielf7b7fd82005-11-28 13:44:07 -0800420 referenced += page_referenced_file(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700421 else if (TestSetPageLocked(page))
422 referenced++;
423 else {
424 if (page->mapping)
Rik van Rielf7b7fd82005-11-28 13:44:07 -0800425 referenced += page_referenced_file(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700426 unlock_page(page);
427 }
428 }
429 return referenced;
430}
431
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700432static int page_mkclean_one(struct page *page, struct vm_area_struct *vma)
433{
434 struct mm_struct *mm = vma->vm_mm;
435 unsigned long address;
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100436 pte_t *pte;
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700437 spinlock_t *ptl;
438 int ret = 0;
439
440 address = vma_address(page, vma);
441 if (address == -EFAULT)
442 goto out;
443
444 pte = page_check_address(page, mm, address, &ptl);
445 if (!pte)
446 goto out;
447
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100448 if (pte_dirty(*pte) || pte_write(*pte)) {
449 pte_t entry;
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700450
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100451 flush_cache_page(vma, address, pte_pfn(*pte));
452 entry = ptep_clear_flush(vma, address, pte);
453 entry = pte_wrprotect(entry);
454 entry = pte_mkclean(entry);
Al Virod6e88e62006-12-29 16:48:35 -0800455 set_pte_at(mm, address, pte, entry);
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100456 lazy_mmu_prot_update(entry);
457 ret = 1;
458 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700459
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700460 pte_unmap_unlock(pte, ptl);
461out:
462 return ret;
463}
464
465static int page_mkclean_file(struct address_space *mapping, struct page *page)
466{
467 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
468 struct vm_area_struct *vma;
469 struct prio_tree_iter iter;
470 int ret = 0;
471
472 BUG_ON(PageAnon(page));
473
474 spin_lock(&mapping->i_mmap_lock);
475 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
476 if (vma->vm_flags & VM_SHARED)
477 ret += page_mkclean_one(page, vma);
478 }
479 spin_unlock(&mapping->i_mmap_lock);
480 return ret;
481}
482
483int page_mkclean(struct page *page)
484{
485 int ret = 0;
486
487 BUG_ON(!PageLocked(page));
488
489 if (page_mapped(page)) {
490 struct address_space *mapping = page_mapping(page);
491 if (mapping)
492 ret = page_mkclean_file(mapping, page);
493 }
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100494 if (page_test_and_clear_dirty(page))
495 ret = 1;
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700496
497 return ret;
498}
499
Linus Torvalds1da177e2005-04-16 15:20:36 -0700500/**
Nick Piggin9617d952006-01-06 00:11:12 -0800501 * page_set_anon_rmap - setup new anonymous rmap
502 * @page: the page to add the mapping to
503 * @vma: the vm area in which the mapping is added
504 * @address: the user virtual address mapped
505 */
506static void __page_set_anon_rmap(struct page *page,
507 struct vm_area_struct *vma, unsigned long address)
508{
509 struct anon_vma *anon_vma = vma->anon_vma;
510
511 BUG_ON(!anon_vma);
512 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
513 page->mapping = (struct address_space *) anon_vma;
514
515 page->index = linear_page_index(vma, address);
516
Nick Piggina74609f2006-01-06 00:11:20 -0800517 /*
518 * nr_mapped state can be updated without turning off
519 * interrupts because it is not modified via interrupt.
520 */
Christoph Lameterf3dbd342006-06-30 01:55:36 -0700521 __inc_zone_page_state(page, NR_ANON_PAGES);
Nick Piggin9617d952006-01-06 00:11:12 -0800522}
523
524/**
Linus Torvalds1da177e2005-04-16 15:20:36 -0700525 * page_add_anon_rmap - add pte mapping to an anonymous page
526 * @page: the page to add the mapping to
527 * @vma: the vm area in which the mapping is added
528 * @address: the user virtual address mapped
529 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -0700530 * The caller needs to hold the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700531 */
532void page_add_anon_rmap(struct page *page,
533 struct vm_area_struct *vma, unsigned long address)
534{
Nick Piggin9617d952006-01-06 00:11:12 -0800535 if (atomic_inc_and_test(&page->_mapcount))
536 __page_set_anon_rmap(page, vma, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700537 /* else checking page index and mapping is racy */
538}
539
Nick Piggin9617d952006-01-06 00:11:12 -0800540/*
541 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
542 * @page: the page to add the mapping to
543 * @vma: the vm area in which the mapping is added
544 * @address: the user virtual address mapped
545 *
546 * Same as page_add_anon_rmap but must only be called on *new* pages.
547 * This means the inc-and-test can be bypassed.
548 */
549void page_add_new_anon_rmap(struct page *page,
550 struct vm_area_struct *vma, unsigned long address)
551{
552 atomic_set(&page->_mapcount, 0); /* elevate count by 1 (starts at -1) */
553 __page_set_anon_rmap(page, vma, address);
554}
555
Linus Torvalds1da177e2005-04-16 15:20:36 -0700556/**
557 * page_add_file_rmap - add pte mapping to a file page
558 * @page: the page to add the mapping to
559 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -0700560 * The caller needs to hold the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700561 */
562void page_add_file_rmap(struct page *page)
563{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700564 if (atomic_inc_and_test(&page->_mapcount))
Christoph Lameter65ba55f2006-06-30 01:55:34 -0700565 __inc_zone_page_state(page, NR_FILE_MAPPED);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700566}
567
568/**
569 * page_remove_rmap - take down pte mapping from a page
570 * @page: page to remove mapping from
571 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -0700572 * The caller needs to hold the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700573 */
Nick Piggin7de6b802006-12-22 01:09:33 -0800574void page_remove_rmap(struct page *page, struct vm_area_struct *vma)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700575{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700576 if (atomic_add_negative(-1, &page->_mapcount)) {
Nick Pigginb7ab7952006-03-22 00:08:42 -0800577 if (unlikely(page_mapcount(page) < 0)) {
Dave Jonesef2bf0d2006-01-08 01:01:00 -0800578 printk (KERN_EMERG "Eeek! page_mapcount(page) went negative! (%d)\n", page_mapcount(page));
Nick Piggin7de6b802006-12-22 01:09:33 -0800579 printk (KERN_EMERG " page pfn = %lx\n", page_to_pfn(page));
Dave Jonesef2bf0d2006-01-08 01:01:00 -0800580 printk (KERN_EMERG " page->flags = %lx\n", page->flags);
581 printk (KERN_EMERG " page->count = %x\n", page_count(page));
582 printk (KERN_EMERG " page->mapping = %p\n", page->mapping);
Nick Piggin7de6b802006-12-22 01:09:33 -0800583 print_symbol (KERN_EMERG " vma->vm_ops = %s\n", (unsigned long)vma->vm_ops);
584 if (vma->vm_ops)
585 print_symbol (KERN_EMERG " vma->vm_ops->nopage = %s\n", (unsigned long)vma->vm_ops->nopage);
586 if (vma->vm_file && vma->vm_file->f_op)
587 print_symbol (KERN_EMERG " vma->vm_file->f_op->mmap = %s\n", (unsigned long)vma->vm_file->f_op->mmap);
Dave Jonesb16bc642006-10-11 01:21:27 -0700588 BUG();
Dave Jonesef2bf0d2006-01-08 01:01:00 -0800589 }
Dave Jonesb16bc642006-10-11 01:21:27 -0700590
Linus Torvalds1da177e2005-04-16 15:20:36 -0700591 /*
592 * It would be tidy to reset the PageAnon mapping here,
593 * but that might overwrite a racing page_add_anon_rmap
594 * which increments mapcount after us but sets mapping
595 * before us: so leave the reset to free_hot_cold_page,
596 * and remember that it's only reliable while mapped.
597 * Leaving it set also helps swapoff to reinstate ptes
598 * faster for those pages still in swapcache.
599 */
600 if (page_test_and_clear_dirty(page))
601 set_page_dirty(page);
Christoph Lameterf3dbd342006-06-30 01:55:36 -0700602 __dec_zone_page_state(page,
603 PageAnon(page) ? NR_ANON_PAGES : NR_FILE_MAPPED);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700604 }
605}
606
607/*
608 * Subfunctions of try_to_unmap: try_to_unmap_one called
609 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
610 */
Christoph Lametera48d07a2006-02-01 03:05:38 -0800611static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
Christoph Lameter73523492006-06-23 02:03:27 -0700612 int migration)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700613{
614 struct mm_struct *mm = vma->vm_mm;
615 unsigned long address;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700616 pte_t *pte;
617 pte_t pteval;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700618 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700619 int ret = SWAP_AGAIN;
620
Linus Torvalds1da177e2005-04-16 15:20:36 -0700621 address = vma_address(page, vma);
622 if (address == -EFAULT)
623 goto out;
624
Hugh Dickinsc0718802005-10-29 18:16:31 -0700625 pte = page_check_address(page, mm, address, &ptl);
626 if (!pte)
Nikita Danilov81b40822005-05-01 08:58:36 -0700627 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700628
629 /*
630 * If the page is mlock()d, we cannot swap it out.
631 * If it's recently referenced (perhaps page_referenced
632 * skipped over this mm) then we should reactivate it.
633 */
Christoph Lametere6a15302006-06-25 05:46:49 -0700634 if (!migration && ((vma->vm_flags & VM_LOCKED) ||
635 (ptep_clear_flush_young(vma, address, pte)))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700636 ret = SWAP_FAIL;
637 goto out_unmap;
638 }
639
Linus Torvalds1da177e2005-04-16 15:20:36 -0700640 /* Nuke the page table entry. */
641 flush_cache_page(vma, address, page_to_pfn(page));
642 pteval = ptep_clear_flush(vma, address, pte);
643
644 /* Move the dirty bit to the physical page now the pte is gone. */
645 if (pte_dirty(pteval))
646 set_page_dirty(page);
647
Hugh Dickins365e9c872005-10-29 18:16:18 -0700648 /* Update high watermark before we lower rss */
649 update_hiwater_rss(mm);
650
Linus Torvalds1da177e2005-04-16 15:20:36 -0700651 if (PageAnon(page)) {
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700652 swp_entry_t entry = { .val = page_private(page) };
Christoph Lameter06972122006-06-23 02:03:35 -0700653
654 if (PageSwapCache(page)) {
655 /*
656 * Store the swap location in the pte.
657 * See handle_pte_fault() ...
658 */
659 swap_duplicate(entry);
660 if (list_empty(&mm->mmlist)) {
661 spin_lock(&mmlist_lock);
662 if (list_empty(&mm->mmlist))
663 list_add(&mm->mmlist, &init_mm.mmlist);
664 spin_unlock(&mmlist_lock);
665 }
Christoph Lameter442c9132006-06-23 02:03:38 -0700666 dec_mm_counter(mm, anon_rss);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700667#ifdef CONFIG_MIGRATION
Christoph Lameter06972122006-06-23 02:03:35 -0700668 } else {
669 /*
670 * Store the pfn of the page in a special migration
671 * pte. do_swap_page() will wait until the migration
672 * pte is removed and then restart fault handling.
673 */
674 BUG_ON(!migration);
675 entry = make_migration_entry(page, pte_write(pteval));
Christoph Lameter04e62a22006-06-23 02:03:38 -0700676#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700677 }
678 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
679 BUG_ON(pte_file(*pte));
Hugh Dickins42946212005-10-29 18:16:05 -0700680 } else
Christoph Lameter04e62a22006-06-23 02:03:38 -0700681#ifdef CONFIG_MIGRATION
682 if (migration) {
683 /* Establish migration entry for a file page */
684 swp_entry_t entry;
685 entry = make_migration_entry(page, pte_write(pteval));
686 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
687 } else
688#endif
Hugh Dickins42946212005-10-29 18:16:05 -0700689 dec_mm_counter(mm, file_rss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700690
Christoph Lameter04e62a22006-06-23 02:03:38 -0700691
Nick Piggin7de6b802006-12-22 01:09:33 -0800692 page_remove_rmap(page, vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700693 page_cache_release(page);
694
695out_unmap:
Hugh Dickinsc0718802005-10-29 18:16:31 -0700696 pte_unmap_unlock(pte, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700697out:
698 return ret;
699}
700
701/*
702 * objrmap doesn't work for nonlinear VMAs because the assumption that
703 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
704 * Consequently, given a particular page and its ->index, we cannot locate the
705 * ptes which are mapping that page without an exhaustive linear search.
706 *
707 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
708 * maps the file to which the target page belongs. The ->vm_private_data field
709 * holds the current cursor into that scan. Successive searches will circulate
710 * around the vma's virtual address space.
711 *
712 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
713 * more scanning pressure is placed against them as well. Eventually pages
714 * will become fully unmapped and are eligible for eviction.
715 *
716 * For very sparsely populated VMAs this is a little inefficient - chances are
717 * there there won't be many ptes located within the scan cluster. In this case
718 * maybe we could scan further - to the end of the pte page, perhaps.
719 */
720#define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
721#define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
722
723static void try_to_unmap_cluster(unsigned long cursor,
724 unsigned int *mapcount, struct vm_area_struct *vma)
725{
726 struct mm_struct *mm = vma->vm_mm;
727 pgd_t *pgd;
728 pud_t *pud;
729 pmd_t *pmd;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700730 pte_t *pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700731 pte_t pteval;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700732 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700733 struct page *page;
734 unsigned long address;
735 unsigned long end;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700736
Linus Torvalds1da177e2005-04-16 15:20:36 -0700737 address = (vma->vm_start + cursor) & CLUSTER_MASK;
738 end = address + CLUSTER_SIZE;
739 if (address < vma->vm_start)
740 address = vma->vm_start;
741 if (end > vma->vm_end)
742 end = vma->vm_end;
743
744 pgd = pgd_offset(mm, address);
745 if (!pgd_present(*pgd))
Hugh Dickinsc0718802005-10-29 18:16:31 -0700746 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700747
748 pud = pud_offset(pgd, address);
749 if (!pud_present(*pud))
Hugh Dickinsc0718802005-10-29 18:16:31 -0700750 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700751
752 pmd = pmd_offset(pud, address);
753 if (!pmd_present(*pmd))
Hugh Dickinsc0718802005-10-29 18:16:31 -0700754 return;
755
756 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700757
Hugh Dickins365e9c872005-10-29 18:16:18 -0700758 /* Update high watermark before we lower rss */
759 update_hiwater_rss(mm);
760
Hugh Dickinsc0718802005-10-29 18:16:31 -0700761 for (; address < end; pte++, address += PAGE_SIZE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700762 if (!pte_present(*pte))
763 continue;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800764 page = vm_normal_page(vma, address, *pte);
765 BUG_ON(!page || PageAnon(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700766
767 if (ptep_clear_flush_young(vma, address, pte))
768 continue;
769
770 /* Nuke the page table entry. */
Ben Collinseca35132005-11-29 11:45:26 -0800771 flush_cache_page(vma, address, pte_pfn(*pte));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700772 pteval = ptep_clear_flush(vma, address, pte);
773
774 /* If nonlinear, store the file page offset in the pte. */
775 if (page->index != linear_page_index(vma, address))
776 set_pte_at(mm, address, pte, pgoff_to_pte(page->index));
777
778 /* Move the dirty bit to the physical page now the pte is gone. */
779 if (pte_dirty(pteval))
780 set_page_dirty(page);
781
Nick Piggin7de6b802006-12-22 01:09:33 -0800782 page_remove_rmap(page, vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700783 page_cache_release(page);
Hugh Dickins42946212005-10-29 18:16:05 -0700784 dec_mm_counter(mm, file_rss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700785 (*mapcount)--;
786 }
Hugh Dickinsc0718802005-10-29 18:16:31 -0700787 pte_unmap_unlock(pte - 1, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700788}
789
Christoph Lameter73523492006-06-23 02:03:27 -0700790static int try_to_unmap_anon(struct page *page, int migration)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700791{
792 struct anon_vma *anon_vma;
793 struct vm_area_struct *vma;
794 int ret = SWAP_AGAIN;
795
796 anon_vma = page_lock_anon_vma(page);
797 if (!anon_vma)
798 return ret;
799
800 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
Christoph Lameter73523492006-06-23 02:03:27 -0700801 ret = try_to_unmap_one(page, vma, migration);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700802 if (ret == SWAP_FAIL || !page_mapped(page))
803 break;
804 }
805 spin_unlock(&anon_vma->lock);
806 return ret;
807}
808
809/**
810 * try_to_unmap_file - unmap file page using the object-based rmap method
811 * @page: the page to unmap
812 *
813 * Find all the mappings of a page using the mapping pointer and the vma chains
814 * contained in the address_space struct it points to.
815 *
816 * This function is only called from try_to_unmap for object-based pages.
817 */
Christoph Lameter73523492006-06-23 02:03:27 -0700818static int try_to_unmap_file(struct page *page, int migration)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700819{
820 struct address_space *mapping = page->mapping;
821 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
822 struct vm_area_struct *vma;
823 struct prio_tree_iter iter;
824 int ret = SWAP_AGAIN;
825 unsigned long cursor;
826 unsigned long max_nl_cursor = 0;
827 unsigned long max_nl_size = 0;
828 unsigned int mapcount;
829
830 spin_lock(&mapping->i_mmap_lock);
831 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
Christoph Lameter73523492006-06-23 02:03:27 -0700832 ret = try_to_unmap_one(page, vma, migration);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700833 if (ret == SWAP_FAIL || !page_mapped(page))
834 goto out;
835 }
836
837 if (list_empty(&mapping->i_mmap_nonlinear))
838 goto out;
839
840 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
841 shared.vm_set.list) {
Christoph Lametere6a15302006-06-25 05:46:49 -0700842 if ((vma->vm_flags & VM_LOCKED) && !migration)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700843 continue;
844 cursor = (unsigned long) vma->vm_private_data;
845 if (cursor > max_nl_cursor)
846 max_nl_cursor = cursor;
847 cursor = vma->vm_end - vma->vm_start;
848 if (cursor > max_nl_size)
849 max_nl_size = cursor;
850 }
851
852 if (max_nl_size == 0) { /* any nonlinears locked or reserved */
853 ret = SWAP_FAIL;
854 goto out;
855 }
856
857 /*
858 * We don't try to search for this page in the nonlinear vmas,
859 * and page_referenced wouldn't have found it anyway. Instead
860 * just walk the nonlinear vmas trying to age and unmap some.
861 * The mapcount of the page we came in with is irrelevant,
862 * but even so use it as a guide to how hard we should try?
863 */
864 mapcount = page_mapcount(page);
865 if (!mapcount)
866 goto out;
867 cond_resched_lock(&mapping->i_mmap_lock);
868
869 max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
870 if (max_nl_cursor == 0)
871 max_nl_cursor = CLUSTER_SIZE;
872
873 do {
874 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
875 shared.vm_set.list) {
Christoph Lametere6a15302006-06-25 05:46:49 -0700876 if ((vma->vm_flags & VM_LOCKED) && !migration)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700877 continue;
878 cursor = (unsigned long) vma->vm_private_data;
Hugh Dickins839b9682005-09-03 15:54:43 -0700879 while ( cursor < max_nl_cursor &&
Linus Torvalds1da177e2005-04-16 15:20:36 -0700880 cursor < vma->vm_end - vma->vm_start) {
881 try_to_unmap_cluster(cursor, &mapcount, vma);
882 cursor += CLUSTER_SIZE;
883 vma->vm_private_data = (void *) cursor;
884 if ((int)mapcount <= 0)
885 goto out;
886 }
887 vma->vm_private_data = (void *) max_nl_cursor;
888 }
889 cond_resched_lock(&mapping->i_mmap_lock);
890 max_nl_cursor += CLUSTER_SIZE;
891 } while (max_nl_cursor <= max_nl_size);
892
893 /*
894 * Don't loop forever (perhaps all the remaining pages are
895 * in locked vmas). Reset cursor on all unreserved nonlinear
896 * vmas, now forgetting on which ones it had fallen behind.
897 */
Hugh Dickins101d2be2005-11-21 21:32:16 -0800898 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
899 vma->vm_private_data = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700900out:
901 spin_unlock(&mapping->i_mmap_lock);
902 return ret;
903}
904
905/**
906 * try_to_unmap - try to remove all page table mappings to a page
907 * @page: the page to get unmapped
908 *
909 * Tries to remove all the page table entries which are mapping this
910 * page, used in the pageout path. Caller must hold the page lock.
911 * Return values are:
912 *
913 * SWAP_SUCCESS - we succeeded in removing all mappings
914 * SWAP_AGAIN - we missed a mapping, try again later
915 * SWAP_FAIL - the page is unswappable
916 */
Christoph Lameter73523492006-06-23 02:03:27 -0700917int try_to_unmap(struct page *page, int migration)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700918{
919 int ret;
920
Linus Torvalds1da177e2005-04-16 15:20:36 -0700921 BUG_ON(!PageLocked(page));
922
923 if (PageAnon(page))
Christoph Lameter73523492006-06-23 02:03:27 -0700924 ret = try_to_unmap_anon(page, migration);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700925 else
Christoph Lameter73523492006-06-23 02:03:27 -0700926 ret = try_to_unmap_file(page, migration);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700927
928 if (!page_mapped(page))
929 ret = SWAP_SUCCESS;
930 return ret;
931}
Nikita Danilov81b40822005-05-01 08:58:36 -0700932