blob: d85a99d28c0387a28d92538a88b92ddfef0efbb2 [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>
55
56#include <asm/tlbflush.h>
57
58//#define RMAP_DEBUG /* can be enabled only for debugging */
59
60kmem_cache_t *anon_vma_cachep;
61
62static inline void validate_anon_vma(struct vm_area_struct *find_vma)
63{
64#ifdef RMAP_DEBUG
65 struct anon_vma *anon_vma = find_vma->anon_vma;
66 struct vm_area_struct *vma;
67 unsigned int mapcount = 0;
68 int found = 0;
69
70 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
71 mapcount++;
72 BUG_ON(mapcount > 100000);
73 if (vma == find_vma)
74 found = 1;
75 }
76 BUG_ON(!found);
77#endif
78}
79
80/* This must be called under the mmap_sem. */
81int anon_vma_prepare(struct vm_area_struct *vma)
82{
83 struct anon_vma *anon_vma = vma->anon_vma;
84
85 might_sleep();
86 if (unlikely(!anon_vma)) {
87 struct mm_struct *mm = vma->vm_mm;
88 struct anon_vma *allocated, *locked;
89
90 anon_vma = find_mergeable_anon_vma(vma);
91 if (anon_vma) {
92 allocated = NULL;
93 locked = anon_vma;
94 spin_lock(&locked->lock);
95 } else {
96 anon_vma = anon_vma_alloc();
97 if (unlikely(!anon_vma))
98 return -ENOMEM;
99 allocated = anon_vma;
100 locked = NULL;
101 }
102
103 /* page_table_lock to protect against threads */
104 spin_lock(&mm->page_table_lock);
105 if (likely(!vma->anon_vma)) {
106 vma->anon_vma = anon_vma;
107 list_add(&vma->anon_vma_node, &anon_vma->head);
108 allocated = NULL;
109 }
110 spin_unlock(&mm->page_table_lock);
111
112 if (locked)
113 spin_unlock(&locked->lock);
114 if (unlikely(allocated))
115 anon_vma_free(allocated);
116 }
117 return 0;
118}
119
120void __anon_vma_merge(struct vm_area_struct *vma, struct vm_area_struct *next)
121{
122 BUG_ON(vma->anon_vma != next->anon_vma);
123 list_del(&next->anon_vma_node);
124}
125
126void __anon_vma_link(struct vm_area_struct *vma)
127{
128 struct anon_vma *anon_vma = vma->anon_vma;
129
130 if (anon_vma) {
131 list_add(&vma->anon_vma_node, &anon_vma->head);
132 validate_anon_vma(vma);
133 }
134}
135
136void anon_vma_link(struct vm_area_struct *vma)
137{
138 struct anon_vma *anon_vma = vma->anon_vma;
139
140 if (anon_vma) {
141 spin_lock(&anon_vma->lock);
142 list_add(&vma->anon_vma_node, &anon_vma->head);
143 validate_anon_vma(vma);
144 spin_unlock(&anon_vma->lock);
145 }
146}
147
148void anon_vma_unlink(struct vm_area_struct *vma)
149{
150 struct anon_vma *anon_vma = vma->anon_vma;
151 int empty;
152
153 if (!anon_vma)
154 return;
155
156 spin_lock(&anon_vma->lock);
157 validate_anon_vma(vma);
158 list_del(&vma->anon_vma_node);
159
160 /* We must garbage collect the anon_vma if it's empty */
161 empty = list_empty(&anon_vma->head);
162 spin_unlock(&anon_vma->lock);
163
164 if (empty)
165 anon_vma_free(anon_vma);
166}
167
168static void anon_vma_ctor(void *data, kmem_cache_t *cachep, unsigned long flags)
169{
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 */
458 __inc_page_state(nr_mapped);
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{
501 BUG_ON(PageAnon(page));
Nick Pigginb5810032005-10-29 18:16:12 -0700502 BUG_ON(!pfn_valid(page_to_pfn(page)));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700503
504 if (atomic_inc_and_test(&page->_mapcount))
Nick Piggina74609f2006-01-06 00:11:20 -0800505 __inc_page_state(nr_mapped);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700506}
507
508/**
509 * page_remove_rmap - take down pte mapping from a page
510 * @page: page to remove mapping from
511 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -0700512 * The caller needs to hold the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700513 */
514void page_remove_rmap(struct page *page)
515{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700516 if (atomic_add_negative(-1, &page->_mapcount)) {
Dave Jonesef2bf0d2006-01-08 01:01:00 -0800517 if (page_mapcount(page) < 0) {
518 printk (KERN_EMERG "Eeek! page_mapcount(page) went negative! (%d)\n", page_mapcount(page));
519 printk (KERN_EMERG " page->flags = %lx\n", page->flags);
520 printk (KERN_EMERG " page->count = %x\n", page_count(page));
521 printk (KERN_EMERG " page->mapping = %p\n", page->mapping);
522 }
523
Linus Torvalds1da177e2005-04-16 15:20:36 -0700524 BUG_ON(page_mapcount(page) < 0);
525 /*
526 * It would be tidy to reset the PageAnon mapping here,
527 * but that might overwrite a racing page_add_anon_rmap
528 * which increments mapcount after us but sets mapping
529 * before us: so leave the reset to free_hot_cold_page,
530 * and remember that it's only reliable while mapped.
531 * Leaving it set also helps swapoff to reinstate ptes
532 * faster for those pages still in swapcache.
533 */
534 if (page_test_and_clear_dirty(page))
535 set_page_dirty(page);
Nick Piggina74609f2006-01-06 00:11:20 -0800536 __dec_page_state(nr_mapped);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700537 }
538}
539
540/*
541 * Subfunctions of try_to_unmap: try_to_unmap_one called
542 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
543 */
544static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma)
545{
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 */
Hugh Dickins101d2be2005-11-21 21:32:16 -0800566 if ((vma->vm_flags & VM_LOCKED) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -0700567 ptep_clear_flush_young(vma, address, pte)) {
568 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) };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700585 /*
586 * Store the swap location in the pte.
587 * See handle_pte_fault() ...
588 */
589 BUG_ON(!PageSwapCache(page));
590 swap_duplicate(entry);
591 if (list_empty(&mm->mmlist)) {
592 spin_lock(&mmlist_lock);
Hugh Dickinsf412ac02005-10-29 18:16:41 -0700593 if (list_empty(&mm->mmlist))
594 list_add(&mm->mmlist, &init_mm.mmlist);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700595 spin_unlock(&mmlist_lock);
596 }
597 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
598 BUG_ON(pte_file(*pte));
599 dec_mm_counter(mm, anon_rss);
Hugh Dickins42946212005-10-29 18:16:05 -0700600 } else
601 dec_mm_counter(mm, file_rss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700602
Linus Torvalds1da177e2005-04-16 15:20:36 -0700603 page_remove_rmap(page);
604 page_cache_release(page);
605
606out_unmap:
Hugh Dickinsc0718802005-10-29 18:16:31 -0700607 pte_unmap_unlock(pte, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700608out:
609 return ret;
610}
611
612/*
613 * objrmap doesn't work for nonlinear VMAs because the assumption that
614 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
615 * Consequently, given a particular page and its ->index, we cannot locate the
616 * ptes which are mapping that page without an exhaustive linear search.
617 *
618 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
619 * maps the file to which the target page belongs. The ->vm_private_data field
620 * holds the current cursor into that scan. Successive searches will circulate
621 * around the vma's virtual address space.
622 *
623 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
624 * more scanning pressure is placed against them as well. Eventually pages
625 * will become fully unmapped and are eligible for eviction.
626 *
627 * For very sparsely populated VMAs this is a little inefficient - chances are
628 * there there won't be many ptes located within the scan cluster. In this case
629 * maybe we could scan further - to the end of the pte page, perhaps.
630 */
631#define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
632#define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
633
634static void try_to_unmap_cluster(unsigned long cursor,
635 unsigned int *mapcount, struct vm_area_struct *vma)
636{
637 struct mm_struct *mm = vma->vm_mm;
638 pgd_t *pgd;
639 pud_t *pud;
640 pmd_t *pmd;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700641 pte_t *pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700642 pte_t pteval;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700643 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700644 struct page *page;
645 unsigned long address;
646 unsigned long end;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700647
Linus Torvalds1da177e2005-04-16 15:20:36 -0700648 address = (vma->vm_start + cursor) & CLUSTER_MASK;
649 end = address + CLUSTER_SIZE;
650 if (address < vma->vm_start)
651 address = vma->vm_start;
652 if (end > vma->vm_end)
653 end = vma->vm_end;
654
655 pgd = pgd_offset(mm, address);
656 if (!pgd_present(*pgd))
Hugh Dickinsc0718802005-10-29 18:16:31 -0700657 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700658
659 pud = pud_offset(pgd, address);
660 if (!pud_present(*pud))
Hugh Dickinsc0718802005-10-29 18:16:31 -0700661 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700662
663 pmd = pmd_offset(pud, address);
664 if (!pmd_present(*pmd))
Hugh Dickinsc0718802005-10-29 18:16:31 -0700665 return;
666
667 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700668
Hugh Dickins365e9c872005-10-29 18:16:18 -0700669 /* Update high watermark before we lower rss */
670 update_hiwater_rss(mm);
671
Hugh Dickinsc0718802005-10-29 18:16:31 -0700672 for (; address < end; pte++, address += PAGE_SIZE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700673 if (!pte_present(*pte))
674 continue;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800675 page = vm_normal_page(vma, address, *pte);
676 BUG_ON(!page || PageAnon(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700677
678 if (ptep_clear_flush_young(vma, address, pte))
679 continue;
680
681 /* Nuke the page table entry. */
Ben Collinseca35132005-11-29 11:45:26 -0800682 flush_cache_page(vma, address, pte_pfn(*pte));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700683 pteval = ptep_clear_flush(vma, address, pte);
684
685 /* If nonlinear, store the file page offset in the pte. */
686 if (page->index != linear_page_index(vma, address))
687 set_pte_at(mm, address, pte, pgoff_to_pte(page->index));
688
689 /* Move the dirty bit to the physical page now the pte is gone. */
690 if (pte_dirty(pteval))
691 set_page_dirty(page);
692
693 page_remove_rmap(page);
694 page_cache_release(page);
Hugh Dickins42946212005-10-29 18:16:05 -0700695 dec_mm_counter(mm, file_rss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700696 (*mapcount)--;
697 }
Hugh Dickinsc0718802005-10-29 18:16:31 -0700698 pte_unmap_unlock(pte - 1, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700699}
700
701static int try_to_unmap_anon(struct page *page)
702{
703 struct anon_vma *anon_vma;
704 struct vm_area_struct *vma;
705 int ret = SWAP_AGAIN;
706
707 anon_vma = page_lock_anon_vma(page);
708 if (!anon_vma)
709 return ret;
710
711 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
712 ret = try_to_unmap_one(page, vma);
713 if (ret == SWAP_FAIL || !page_mapped(page))
714 break;
715 }
716 spin_unlock(&anon_vma->lock);
717 return ret;
718}
719
720/**
721 * try_to_unmap_file - unmap file page using the object-based rmap method
722 * @page: the page to unmap
723 *
724 * Find all the mappings of a page using the mapping pointer and the vma chains
725 * contained in the address_space struct it points to.
726 *
727 * This function is only called from try_to_unmap for object-based pages.
728 */
729static int try_to_unmap_file(struct page *page)
730{
731 struct address_space *mapping = page->mapping;
732 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
733 struct vm_area_struct *vma;
734 struct prio_tree_iter iter;
735 int ret = SWAP_AGAIN;
736 unsigned long cursor;
737 unsigned long max_nl_cursor = 0;
738 unsigned long max_nl_size = 0;
739 unsigned int mapcount;
740
741 spin_lock(&mapping->i_mmap_lock);
742 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
743 ret = try_to_unmap_one(page, vma);
744 if (ret == SWAP_FAIL || !page_mapped(page))
745 goto out;
746 }
747
748 if (list_empty(&mapping->i_mmap_nonlinear))
749 goto out;
750
751 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
752 shared.vm_set.list) {
Hugh Dickins101d2be2005-11-21 21:32:16 -0800753 if (vma->vm_flags & VM_LOCKED)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700754 continue;
755 cursor = (unsigned long) vma->vm_private_data;
756 if (cursor > max_nl_cursor)
757 max_nl_cursor = cursor;
758 cursor = vma->vm_end - vma->vm_start;
759 if (cursor > max_nl_size)
760 max_nl_size = cursor;
761 }
762
763 if (max_nl_size == 0) { /* any nonlinears locked or reserved */
764 ret = SWAP_FAIL;
765 goto out;
766 }
767
768 /*
769 * We don't try to search for this page in the nonlinear vmas,
770 * and page_referenced wouldn't have found it anyway. Instead
771 * just walk the nonlinear vmas trying to age and unmap some.
772 * The mapcount of the page we came in with is irrelevant,
773 * but even so use it as a guide to how hard we should try?
774 */
775 mapcount = page_mapcount(page);
776 if (!mapcount)
777 goto out;
778 cond_resched_lock(&mapping->i_mmap_lock);
779
780 max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
781 if (max_nl_cursor == 0)
782 max_nl_cursor = CLUSTER_SIZE;
783
784 do {
785 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
786 shared.vm_set.list) {
Hugh Dickins101d2be2005-11-21 21:32:16 -0800787 if (vma->vm_flags & VM_LOCKED)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700788 continue;
789 cursor = (unsigned long) vma->vm_private_data;
Hugh Dickins839b9682005-09-03 15:54:43 -0700790 while ( cursor < max_nl_cursor &&
Linus Torvalds1da177e2005-04-16 15:20:36 -0700791 cursor < vma->vm_end - vma->vm_start) {
792 try_to_unmap_cluster(cursor, &mapcount, vma);
793 cursor += CLUSTER_SIZE;
794 vma->vm_private_data = (void *) cursor;
795 if ((int)mapcount <= 0)
796 goto out;
797 }
798 vma->vm_private_data = (void *) max_nl_cursor;
799 }
800 cond_resched_lock(&mapping->i_mmap_lock);
801 max_nl_cursor += CLUSTER_SIZE;
802 } while (max_nl_cursor <= max_nl_size);
803
804 /*
805 * Don't loop forever (perhaps all the remaining pages are
806 * in locked vmas). Reset cursor on all unreserved nonlinear
807 * vmas, now forgetting on which ones it had fallen behind.
808 */
Hugh Dickins101d2be2005-11-21 21:32:16 -0800809 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
810 vma->vm_private_data = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700811out:
812 spin_unlock(&mapping->i_mmap_lock);
813 return ret;
814}
815
816/**
817 * try_to_unmap - try to remove all page table mappings to a page
818 * @page: the page to get unmapped
819 *
820 * Tries to remove all the page table entries which are mapping this
821 * page, used in the pageout path. Caller must hold the page lock.
822 * Return values are:
823 *
824 * SWAP_SUCCESS - we succeeded in removing all mappings
825 * SWAP_AGAIN - we missed a mapping, try again later
826 * SWAP_FAIL - the page is unswappable
827 */
828int try_to_unmap(struct page *page)
829{
830 int ret;
831
Linus Torvalds1da177e2005-04-16 15:20:36 -0700832 BUG_ON(!PageLocked(page));
833
834 if (PageAnon(page))
835 ret = try_to_unmap_anon(page);
836 else
837 ret = try_to_unmap_file(page);
838
839 if (!page_mapped(page))
840 ret = SWAP_SUCCESS;
841 return ret;
842}
Nikita Danilov81b40822005-05-01 08:58:36 -0700843