blob: f853c6def15915602e484afdd3a741099d543887 [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 *
23 * inode->i_sem (while writing or truncating, not reading or faulting)
24 * inode->i_alloc_sem
25 *
26 * When a page fault occurs in writing from user to file, down_read
27 * of mmap_sem nests within i_sem; in sys_msync, i_sem nests within
28 * down_read of mmap_sem; i_sem and down_write of mmap_sem are never
29 * taken together; in truncation, i_sem is taken outermost.
30 *
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
Linus Torvalds1da177e2005-04-16 15:20:36 -070036 * zone->lru_lock (in mark_page_accessed)
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/**
438 * page_add_anon_rmap - add pte mapping to an anonymous page
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 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -0700443 * The caller needs to hold the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700444 */
445void page_add_anon_rmap(struct page *page,
446 struct vm_area_struct *vma, unsigned long address)
447{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700448 if (atomic_inc_and_test(&page->_mapcount)) {
Nick Piggin2822c1a2005-09-03 15:54:47 -0700449 struct anon_vma *anon_vma = vma->anon_vma;
Nick Piggin2822c1a2005-09-03 15:54:47 -0700450
451 BUG_ON(!anon_vma);
452 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700453 page->mapping = (struct address_space *) anon_vma;
Nick Piggin2822c1a2005-09-03 15:54:47 -0700454
Nick Piggin4d7670e2005-09-03 15:54:48 -0700455 page->index = linear_page_index(vma, address);
Nick Piggin2822c1a2005-09-03 15:54:47 -0700456
Linus Torvalds1da177e2005-04-16 15:20:36 -0700457 inc_page_state(nr_mapped);
458 }
459 /* else checking page index and mapping is racy */
460}
461
462/**
463 * page_add_file_rmap - add pte mapping to a file page
464 * @page: the page to add the mapping to
465 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -0700466 * The caller needs to hold the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700467 */
468void page_add_file_rmap(struct page *page)
469{
470 BUG_ON(PageAnon(page));
Nick Pigginb5810032005-10-29 18:16:12 -0700471 BUG_ON(!pfn_valid(page_to_pfn(page)));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700472
473 if (atomic_inc_and_test(&page->_mapcount))
474 inc_page_state(nr_mapped);
475}
476
477/**
478 * page_remove_rmap - take down pte mapping from a page
479 * @page: page to remove mapping from
480 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -0700481 * The caller needs to hold the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700482 */
483void page_remove_rmap(struct page *page)
484{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700485 if (atomic_add_negative(-1, &page->_mapcount)) {
486 BUG_ON(page_mapcount(page) < 0);
487 /*
488 * It would be tidy to reset the PageAnon mapping here,
489 * but that might overwrite a racing page_add_anon_rmap
490 * which increments mapcount after us but sets mapping
491 * before us: so leave the reset to free_hot_cold_page,
492 * and remember that it's only reliable while mapped.
493 * Leaving it set also helps swapoff to reinstate ptes
494 * faster for those pages still in swapcache.
495 */
496 if (page_test_and_clear_dirty(page))
497 set_page_dirty(page);
498 dec_page_state(nr_mapped);
499 }
500}
501
502/*
503 * Subfunctions of try_to_unmap: try_to_unmap_one called
504 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
505 */
506static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma)
507{
508 struct mm_struct *mm = vma->vm_mm;
509 unsigned long address;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700510 pte_t *pte;
511 pte_t pteval;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700512 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700513 int ret = SWAP_AGAIN;
514
Linus Torvalds1da177e2005-04-16 15:20:36 -0700515 address = vma_address(page, vma);
516 if (address == -EFAULT)
517 goto out;
518
Hugh Dickinsc0718802005-10-29 18:16:31 -0700519 pte = page_check_address(page, mm, address, &ptl);
520 if (!pte)
Nikita Danilov81b40822005-05-01 08:58:36 -0700521 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700522
523 /*
524 * If the page is mlock()d, we cannot swap it out.
525 * If it's recently referenced (perhaps page_referenced
526 * skipped over this mm) then we should reactivate it.
527 */
Hugh Dickins101d2be2005-11-21 21:32:16 -0800528 if ((vma->vm_flags & VM_LOCKED) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -0700529 ptep_clear_flush_young(vma, address, pte)) {
530 ret = SWAP_FAIL;
531 goto out_unmap;
532 }
533
Linus Torvalds1da177e2005-04-16 15:20:36 -0700534 /* Nuke the page table entry. */
535 flush_cache_page(vma, address, page_to_pfn(page));
536 pteval = ptep_clear_flush(vma, address, pte);
537
538 /* Move the dirty bit to the physical page now the pte is gone. */
539 if (pte_dirty(pteval))
540 set_page_dirty(page);
541
Hugh Dickins365e9c872005-10-29 18:16:18 -0700542 /* Update high watermark before we lower rss */
543 update_hiwater_rss(mm);
544
Linus Torvalds1da177e2005-04-16 15:20:36 -0700545 if (PageAnon(page)) {
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700546 swp_entry_t entry = { .val = page_private(page) };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700547 /*
548 * Store the swap location in the pte.
549 * See handle_pte_fault() ...
550 */
551 BUG_ON(!PageSwapCache(page));
552 swap_duplicate(entry);
553 if (list_empty(&mm->mmlist)) {
554 spin_lock(&mmlist_lock);
Hugh Dickinsf412ac02005-10-29 18:16:41 -0700555 if (list_empty(&mm->mmlist))
556 list_add(&mm->mmlist, &init_mm.mmlist);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700557 spin_unlock(&mmlist_lock);
558 }
559 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
560 BUG_ON(pte_file(*pte));
561 dec_mm_counter(mm, anon_rss);
Hugh Dickins42946212005-10-29 18:16:05 -0700562 } else
563 dec_mm_counter(mm, file_rss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700564
Linus Torvalds1da177e2005-04-16 15:20:36 -0700565 page_remove_rmap(page);
566 page_cache_release(page);
567
568out_unmap:
Hugh Dickinsc0718802005-10-29 18:16:31 -0700569 pte_unmap_unlock(pte, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700570out:
571 return ret;
572}
573
574/*
575 * objrmap doesn't work for nonlinear VMAs because the assumption that
576 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
577 * Consequently, given a particular page and its ->index, we cannot locate the
578 * ptes which are mapping that page without an exhaustive linear search.
579 *
580 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
581 * maps the file to which the target page belongs. The ->vm_private_data field
582 * holds the current cursor into that scan. Successive searches will circulate
583 * around the vma's virtual address space.
584 *
585 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
586 * more scanning pressure is placed against them as well. Eventually pages
587 * will become fully unmapped and are eligible for eviction.
588 *
589 * For very sparsely populated VMAs this is a little inefficient - chances are
590 * there there won't be many ptes located within the scan cluster. In this case
591 * maybe we could scan further - to the end of the pte page, perhaps.
592 */
593#define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
594#define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
595
596static void try_to_unmap_cluster(unsigned long cursor,
597 unsigned int *mapcount, struct vm_area_struct *vma)
598{
599 struct mm_struct *mm = vma->vm_mm;
600 pgd_t *pgd;
601 pud_t *pud;
602 pmd_t *pmd;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700603 pte_t *pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700604 pte_t pteval;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700605 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700606 struct page *page;
607 unsigned long address;
608 unsigned long end;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700609
Linus Torvalds1da177e2005-04-16 15:20:36 -0700610 address = (vma->vm_start + cursor) & CLUSTER_MASK;
611 end = address + CLUSTER_SIZE;
612 if (address < vma->vm_start)
613 address = vma->vm_start;
614 if (end > vma->vm_end)
615 end = vma->vm_end;
616
617 pgd = pgd_offset(mm, address);
618 if (!pgd_present(*pgd))
Hugh Dickinsc0718802005-10-29 18:16:31 -0700619 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700620
621 pud = pud_offset(pgd, address);
622 if (!pud_present(*pud))
Hugh Dickinsc0718802005-10-29 18:16:31 -0700623 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700624
625 pmd = pmd_offset(pud, address);
626 if (!pmd_present(*pmd))
Hugh Dickinsc0718802005-10-29 18:16:31 -0700627 return;
628
629 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700630
Hugh Dickins365e9c872005-10-29 18:16:18 -0700631 /* Update high watermark before we lower rss */
632 update_hiwater_rss(mm);
633
Hugh Dickinsc0718802005-10-29 18:16:31 -0700634 for (; address < end; pte++, address += PAGE_SIZE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700635 if (!pte_present(*pte))
636 continue;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800637 page = vm_normal_page(vma, address, *pte);
638 BUG_ON(!page || PageAnon(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700639
640 if (ptep_clear_flush_young(vma, address, pte))
641 continue;
642
643 /* Nuke the page table entry. */
Ben Collinseca35132005-11-29 11:45:26 -0800644 flush_cache_page(vma, address, pte_pfn(*pte));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700645 pteval = ptep_clear_flush(vma, address, pte);
646
647 /* If nonlinear, store the file page offset in the pte. */
648 if (page->index != linear_page_index(vma, address))
649 set_pte_at(mm, address, pte, pgoff_to_pte(page->index));
650
651 /* Move the dirty bit to the physical page now the pte is gone. */
652 if (pte_dirty(pteval))
653 set_page_dirty(page);
654
655 page_remove_rmap(page);
656 page_cache_release(page);
Hugh Dickins42946212005-10-29 18:16:05 -0700657 dec_mm_counter(mm, file_rss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700658 (*mapcount)--;
659 }
Hugh Dickinsc0718802005-10-29 18:16:31 -0700660 pte_unmap_unlock(pte - 1, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700661}
662
663static int try_to_unmap_anon(struct page *page)
664{
665 struct anon_vma *anon_vma;
666 struct vm_area_struct *vma;
667 int ret = SWAP_AGAIN;
668
669 anon_vma = page_lock_anon_vma(page);
670 if (!anon_vma)
671 return ret;
672
673 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
674 ret = try_to_unmap_one(page, vma);
675 if (ret == SWAP_FAIL || !page_mapped(page))
676 break;
677 }
678 spin_unlock(&anon_vma->lock);
679 return ret;
680}
681
682/**
683 * try_to_unmap_file - unmap file page using the object-based rmap method
684 * @page: the page to unmap
685 *
686 * Find all the mappings of a page using the mapping pointer and the vma chains
687 * contained in the address_space struct it points to.
688 *
689 * This function is only called from try_to_unmap for object-based pages.
690 */
691static int try_to_unmap_file(struct page *page)
692{
693 struct address_space *mapping = page->mapping;
694 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
695 struct vm_area_struct *vma;
696 struct prio_tree_iter iter;
697 int ret = SWAP_AGAIN;
698 unsigned long cursor;
699 unsigned long max_nl_cursor = 0;
700 unsigned long max_nl_size = 0;
701 unsigned int mapcount;
702
703 spin_lock(&mapping->i_mmap_lock);
704 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
705 ret = try_to_unmap_one(page, vma);
706 if (ret == SWAP_FAIL || !page_mapped(page))
707 goto out;
708 }
709
710 if (list_empty(&mapping->i_mmap_nonlinear))
711 goto out;
712
713 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
714 shared.vm_set.list) {
Hugh Dickins101d2be2005-11-21 21:32:16 -0800715 if (vma->vm_flags & VM_LOCKED)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700716 continue;
717 cursor = (unsigned long) vma->vm_private_data;
718 if (cursor > max_nl_cursor)
719 max_nl_cursor = cursor;
720 cursor = vma->vm_end - vma->vm_start;
721 if (cursor > max_nl_size)
722 max_nl_size = cursor;
723 }
724
725 if (max_nl_size == 0) { /* any nonlinears locked or reserved */
726 ret = SWAP_FAIL;
727 goto out;
728 }
729
730 /*
731 * We don't try to search for this page in the nonlinear vmas,
732 * and page_referenced wouldn't have found it anyway. Instead
733 * just walk the nonlinear vmas trying to age and unmap some.
734 * The mapcount of the page we came in with is irrelevant,
735 * but even so use it as a guide to how hard we should try?
736 */
737 mapcount = page_mapcount(page);
738 if (!mapcount)
739 goto out;
740 cond_resched_lock(&mapping->i_mmap_lock);
741
742 max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
743 if (max_nl_cursor == 0)
744 max_nl_cursor = CLUSTER_SIZE;
745
746 do {
747 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
748 shared.vm_set.list) {
Hugh Dickins101d2be2005-11-21 21:32:16 -0800749 if (vma->vm_flags & VM_LOCKED)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700750 continue;
751 cursor = (unsigned long) vma->vm_private_data;
Hugh Dickins839b9682005-09-03 15:54:43 -0700752 while ( cursor < max_nl_cursor &&
Linus Torvalds1da177e2005-04-16 15:20:36 -0700753 cursor < vma->vm_end - vma->vm_start) {
754 try_to_unmap_cluster(cursor, &mapcount, vma);
755 cursor += CLUSTER_SIZE;
756 vma->vm_private_data = (void *) cursor;
757 if ((int)mapcount <= 0)
758 goto out;
759 }
760 vma->vm_private_data = (void *) max_nl_cursor;
761 }
762 cond_resched_lock(&mapping->i_mmap_lock);
763 max_nl_cursor += CLUSTER_SIZE;
764 } while (max_nl_cursor <= max_nl_size);
765
766 /*
767 * Don't loop forever (perhaps all the remaining pages are
768 * in locked vmas). Reset cursor on all unreserved nonlinear
769 * vmas, now forgetting on which ones it had fallen behind.
770 */
Hugh Dickins101d2be2005-11-21 21:32:16 -0800771 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
772 vma->vm_private_data = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700773out:
774 spin_unlock(&mapping->i_mmap_lock);
775 return ret;
776}
777
778/**
779 * try_to_unmap - try to remove all page table mappings to a page
780 * @page: the page to get unmapped
781 *
782 * Tries to remove all the page table entries which are mapping this
783 * page, used in the pageout path. Caller must hold the page lock.
784 * Return values are:
785 *
786 * SWAP_SUCCESS - we succeeded in removing all mappings
787 * SWAP_AGAIN - we missed a mapping, try again later
788 * SWAP_FAIL - the page is unswappable
789 */
790int try_to_unmap(struct page *page)
791{
792 int ret;
793
Linus Torvalds1da177e2005-04-16 15:20:36 -0700794 BUG_ON(!PageLocked(page));
795
796 if (PageAnon(page))
797 ret = try_to_unmap_anon(page);
798 else
799 ret = try_to_unmap_file(page);
800
801 if (!page_mapped(page))
802 ret = SWAP_SUCCESS;
803 return ret;
804}
Nikita Danilov81b40822005-05-01 08:58:36 -0700805