1 /*
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
35 * mm->page_table_lock
36 * zone->lru_lock (in mark_page_accessed)
37 * swap_list_lock (in swap_free etc's swap_info_get)
38 * mmlist_lock (in mmput, drain_mmlist and others)
39 * swap_device_lock (in swap_duplicate, swap_info_get)
40 * mapping->private_lock (in __set_page_dirty_buffers)
41 * inode_lock (in set_page_dirty's __mark_inode_dirty)
42 * sb_lock (within inode_lock in fs/fs-writeback.c)
43 * mapping->tree_lock (widely used, in set_page_dirty,
44 * in arch-dependent flush_dcache_mmap_lock,
45 * within inode_lock in __sync_single_inode)
46 */
47
48 #include <linux/mm.h>
49 #include <linux/pagemap.h>
50 #include <linux/swap.h>
51 #include <linux/swapops.h>
52 #include <linux/slab.h>
53 #include <linux/init.h>
54 #include <linux/acct.h>
55 #include <linux/rmap.h>
56 #include <linux/rcupdate.h>
57
58 #include <asm/tlbflush.h>
59
60 //#define RMAP_DEBUG /* can be enabled only for debugging */
61
62 kmem_cache_t *anon_vma_cachep;
63
64 static inline void validate_anon_vma(struct vm_area_struct *find_vma)
65 {
66 #ifdef RMAP_DEBUG
67 struct anon_vma *anon_vma = find_vma->anon_vma;
68 struct vm_area_struct *vma;
69 unsigned int mapcount = 0;
70 int found = 0;
71
72 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
73 mapcount++;
74 BUG_ON(mapcount > 100000);
75 if (vma == find_vma)
76 found = 1;
77 }
78 BUG_ON(!found);
79 #endif
80 }
81
82 /* This must be called under the mmap_sem. */
83 int anon_vma_prepare(struct vm_area_struct *vma)
84 {
85 struct anon_vma *anon_vma = vma->anon_vma;
86
87 might_sleep();
88 if (unlikely(!anon_vma)) {
89 struct mm_struct *mm = vma->vm_mm;
90 struct anon_vma *allocated, *locked;
91
92 anon_vma = find_mergeable_anon_vma(vma);
93 if (anon_vma) {
94 allocated = NULL;
95 locked = anon_vma;
96 spin_lock(&locked->lock);
97 } else {
98 anon_vma = anon_vma_alloc();
99 if (unlikely(!anon_vma))
100 return -ENOMEM;
101 allocated = anon_vma;
102 locked = NULL;
103 }
104
105 /* page_table_lock to protect against threads */
106 spin_lock(&mm->page_table_lock);
107 if (likely(!vma->anon_vma)) {
108 vma->anon_vma = anon_vma;
109 list_add(&vma->anon_vma_node, &anon_vma->head);
110 allocated = NULL;
111 }
112 spin_unlock(&mm->page_table_lock);
113
114 if (locked)
115 spin_unlock(&locked->lock);
116 if (unlikely(allocated))
117 anon_vma_free(allocated);
118 }
119 return 0;
120 }
121
122 void __anon_vma_merge(struct vm_area_struct *vma, struct vm_area_struct *next)
123 {
124 BUG_ON(vma->anon_vma != next->anon_vma);
125 list_del(&next->anon_vma_node);
126 }
127
128 void __anon_vma_link(struct vm_area_struct *vma)
129 {
130 struct anon_vma *anon_vma = vma->anon_vma;
131
132 if (anon_vma) {
133 list_add(&vma->anon_vma_node, &anon_vma->head);
134 validate_anon_vma(vma);
135 }
136 }
137
138 void anon_vma_link(struct vm_area_struct *vma)
139 {
140 struct anon_vma *anon_vma = vma->anon_vma;
141
142 if (anon_vma) {
143 spin_lock(&anon_vma->lock);
144 list_add(&vma->anon_vma_node, &anon_vma->head);
145 validate_anon_vma(vma);
146 spin_unlock(&anon_vma->lock);
147 }
148 }
149
150 void anon_vma_unlink(struct vm_area_struct *vma)
151 {
152 struct anon_vma *anon_vma = vma->anon_vma;
153 int empty;
154
155 if (!anon_vma)
156 return;
157
158 spin_lock(&anon_vma->lock);
159 validate_anon_vma(vma);
160 list_del(&vma->anon_vma_node);
161
162 /* We must garbage collect the anon_vma if it's empty */
163 empty = list_empty(&anon_vma->head);
164 spin_unlock(&anon_vma->lock);
165
166 if (empty)
167 anon_vma_free(anon_vma);
168 }
169
170 static void anon_vma_ctor(void *data, kmem_cache_t *cachep, unsigned long flags)
171 {
172 if ((flags & (SLAB_CTOR_VERIFY|SLAB_CTOR_CONSTRUCTOR)) ==
173 SLAB_CTOR_CONSTRUCTOR) {
174 struct anon_vma *anon_vma = data;
175
176 spin_lock_init(&anon_vma->lock);
177 INIT_LIST_HEAD(&anon_vma->head);
178 }
179 }
180
181 void __init anon_vma_init(void)
182 {
183 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
184 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor, NULL);
185 }
186
187 /*
188 * Getting a lock on a stable anon_vma from a page off the LRU is
189 * tricky: page_lock_anon_vma rely on RCU to guard against the races.
190 */
191 static struct anon_vma *page_lock_anon_vma(struct page *page)
192 {
193 struct anon_vma *anon_vma = NULL;
194 unsigned long anon_mapping;
195
196 rcu_read_lock();
197 anon_mapping = (unsigned long) page->mapping;
198 if (!(anon_mapping & PAGE_MAPPING_ANON))
199 goto out;
200 if (!page_mapped(page))
201 goto out;
202
203 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
204 spin_lock(&anon_vma->lock);
205 out:
206 rcu_read_unlock();
207 return anon_vma;
208 }
209
210 /*
211 * At what user virtual address is page expected in vma?
212 */
213 static inline unsigned long
214 vma_address(struct page *page, struct vm_area_struct *vma)
215 {
216 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
217 unsigned long address;
218
219 address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
220 if (unlikely(address < vma->vm_start || address >= vma->vm_end)) {
221 /* page should be within any vma from prio_tree_next */
222 BUG_ON(!PageAnon(page));
223 return -EFAULT;
224 }
225 return address;
226 }
227
228 /*
229 * At what user virtual address is page expected in vma? checking that the
230 * page matches the vma: currently only used by unuse_process, on anon pages.
231 */
232 unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
233 {
234 if (PageAnon(page)) {
235 if ((void *)vma->anon_vma !=
236 (void *)page->mapping - PAGE_MAPPING_ANON)
237 return -EFAULT;
238 } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) {
239 if (vma->vm_file->f_mapping != page->mapping)
240 return -EFAULT;
241 } else
242 return -EFAULT;
243 return vma_address(page, vma);
244 }
245
246 /*
247 * Subfunctions of page_referenced: page_referenced_one called
248 * repeatedly from either page_referenced_anon or page_referenced_file.
249 */
250 static int page_referenced_one(struct page *page,
251 struct vm_area_struct *vma, unsigned int *mapcount, int ignore_token)
252 {
253 struct mm_struct *mm = vma->vm_mm;
254 unsigned long address;
255 pgd_t *pgd;
256 pud_t *pud;
257 pmd_t *pmd;
258 pte_t *pte;
259 int referenced = 0;
260
261 if (!mm->rss)
262 goto out;
263 address = vma_address(page, vma);
264 if (address == -EFAULT)
265 goto out;
266
267 spin_lock(&mm->page_table_lock);
268
269 pgd = pgd_offset(mm, address);
270 if (!pgd_present(*pgd))
271 goto out_unlock;
272
273 pud = pud_offset(pgd, address);
274 if (!pud_present(*pud))
275 goto out_unlock;
276
277 pmd = pmd_offset(pud, address);
278 if (!pmd_present(*pmd))
279 goto out_unlock;
280
281 pte = pte_offset_map(pmd, address);
282 if (!pte_present(*pte))
283 goto out_unmap;
284
285 if (page_to_pfn(page) != pte_pfn(*pte))
286 goto out_unmap;
287
288 if (ptep_clear_flush_young(vma, address, pte))
289 referenced++;
290
291 if (mm != current->mm && !ignore_token && has_swap_token(mm))
292 referenced++;
293
294 (*mapcount)--;
295
296 out_unmap:
297 pte_unmap(pte);
298 out_unlock:
299 spin_unlock(&mm->page_table_lock);
300 out:
301 return referenced;
302 }
303
304 static int page_referenced_anon(struct page *page, int ignore_token)
305 {
306 unsigned int mapcount;
307 struct anon_vma *anon_vma;
308 struct vm_area_struct *vma;
309 int referenced = 0;
310
311 anon_vma = page_lock_anon_vma(page);
312 if (!anon_vma)
313 return referenced;
314
315 mapcount = page_mapcount(page);
316 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
317 referenced += page_referenced_one(page, vma, &mapcount,
318 ignore_token);
319 if (!mapcount)
320 break;
321 }
322 spin_unlock(&anon_vma->lock);
323 return referenced;
324 }
325
326 /**
327 * page_referenced_file - referenced check for object-based rmap
328 * @page: the page we're checking references on.
329 *
330 * For an object-based mapped page, find all the places it is mapped and
331 * check/clear the referenced flag. This is done by following the page->mapping
332 * pointer, then walking the chain of vmas it holds. It returns the number
333 * of references it found.
334 *
335 * This function is only called from page_referenced for object-based pages.
336 */
337 static int page_referenced_file(struct page *page, int ignore_token)
338 {
339 unsigned int mapcount;
340 struct address_space *mapping = page->mapping;
341 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
342 struct vm_area_struct *vma;
343 struct prio_tree_iter iter;
344 int referenced = 0;
345
346 /*
347 * The caller's checks on page->mapping and !PageAnon have made
348 * sure that this is a file page: the check for page->mapping
349 * excludes the case just before it gets set on an anon page.
350 */
351 BUG_ON(PageAnon(page));
352
353 /*
354 * The page lock not only makes sure that page->mapping cannot
355 * suddenly be NULLified by truncation, it makes sure that the
356 * structure at mapping cannot be freed and reused yet,
357 * so we can safely take mapping->i_mmap_lock.
358 */
359 BUG_ON(!PageLocked(page));
360
361 spin_lock(&mapping->i_mmap_lock);
362
363 /*
364 * i_mmap_lock does not stabilize mapcount at all, but mapcount
365 * is more likely to be accurate if we note it after spinning.
366 */
367 mapcount = page_mapcount(page);
368
369 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
370 if ((vma->vm_flags & (VM_LOCKED|VM_MAYSHARE))
371 == (VM_LOCKED|VM_MAYSHARE)) {
372 referenced++;
373 break;
374 }
375 referenced += page_referenced_one(page, vma, &mapcount,
376 ignore_token);
377 if (!mapcount)
378 break;
379 }
380
381 spin_unlock(&mapping->i_mmap_lock);
382 return referenced;
383 }
384
385 /**
386 * page_referenced - test if the page was referenced
387 * @page: the page to test
388 * @is_locked: caller holds lock on the page
389 *
390 * Quick test_and_clear_referenced for all mappings to a page,
391 * returns the number of ptes which referenced the page.
392 */
393 int page_referenced(struct page *page, int is_locked, int ignore_token)
394 {
395 int referenced = 0;
396
397 if (!swap_token_default_timeout)
398 ignore_token = 1;
399
400 if (page_test_and_clear_young(page))
401 referenced++;
402
403 if (TestClearPageReferenced(page))
404 referenced++;
405
406 if (page_mapped(page) && page->mapping) {
407 if (PageAnon(page))
408 referenced += page_referenced_anon(page, ignore_token);
409 else if (is_locked)
410 referenced += page_referenced_file(page, ignore_token);
411 else if (TestSetPageLocked(page))
412 referenced++;
413 else {
414 if (page->mapping)
415 referenced += page_referenced_file(page,
416 ignore_token);
417 unlock_page(page);
418 }
419 }
420 return referenced;
421 }
422
423 /**
424 * page_add_anon_rmap - add pte mapping to an anonymous page
425 * @page: the page to add the mapping to
426 * @vma: the vm area in which the mapping is added
427 * @address: the user virtual address mapped
428 *
429 * The caller needs to hold the mm->page_table_lock.
430 */
431 void page_add_anon_rmap(struct page *page,
432 struct vm_area_struct *vma, unsigned long address)
433 {
434 struct anon_vma *anon_vma = vma->anon_vma;
435 pgoff_t index;
436
437 BUG_ON(PageReserved(page));
438 BUG_ON(!anon_vma);
439
440 vma->vm_mm->anon_rss++;
441
442 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
443 index = (address - vma->vm_start) >> PAGE_SHIFT;
444 index += vma->vm_pgoff;
445 index >>= PAGE_CACHE_SHIFT - PAGE_SHIFT;
446
447 if (atomic_inc_and_test(&page->_mapcount)) {
448 page->index = index;
449 page->mapping = (struct address_space *) anon_vma;
450 inc_page_state(nr_mapped);
451 }
452 /* else checking page index and mapping is racy */
453 }
454
455 /**
456 * page_add_file_rmap - add pte mapping to a file page
457 * @page: the page to add the mapping to
458 *
459 * The caller needs to hold the mm->page_table_lock.
460 */
461 void page_add_file_rmap(struct page *page)
462 {
463 BUG_ON(PageAnon(page));
464 if (!pfn_valid(page_to_pfn(page)) || PageReserved(page))
465 return;
466
467 if (atomic_inc_and_test(&page->_mapcount))
468 inc_page_state(nr_mapped);
469 }
470
471 /**
472 * page_remove_rmap - take down pte mapping from a page
473 * @page: page to remove mapping from
474 *
475 * Caller needs to hold the mm->page_table_lock.
476 */
477 void page_remove_rmap(struct page *page)
478 {
479 BUG_ON(PageReserved(page));
480
481 if (atomic_add_negative(-1, &page->_mapcount)) {
482 BUG_ON(page_mapcount(page) < 0);
483 /*
484 * It would be tidy to reset the PageAnon mapping here,
485 * but that might overwrite a racing page_add_anon_rmap
486 * which increments mapcount after us but sets mapping
487 * before us: so leave the reset to free_hot_cold_page,
488 * and remember that it's only reliable while mapped.
489 * Leaving it set also helps swapoff to reinstate ptes
490 * faster for those pages still in swapcache.
491 */
492 if (page_test_and_clear_dirty(page))
493 set_page_dirty(page);
494 dec_page_state(nr_mapped);
495 }
496 }
497
498 /*
499 * Subfunctions of try_to_unmap: try_to_unmap_one called
500 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
501 */
502 static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma)
503 {
504 struct mm_struct *mm = vma->vm_mm;
505 unsigned long address;
506 pgd_t *pgd;
507 pud_t *pud;
508 pmd_t *pmd;
509 pte_t *pte;
510 pte_t pteval;
511 int ret = SWAP_AGAIN;
512
513 if (!mm->rss)
514 goto out;
515 address = vma_address(page, vma);
516 if (address == -EFAULT)
517 goto out;
518
519 /*
520 * We need the page_table_lock to protect us from page faults,
521 * munmap, fork, etc...
522 */
523 spin_lock(&mm->page_table_lock);
524
525 pgd = pgd_offset(mm, address);
526 if (!pgd_present(*pgd))
527 goto out_unlock;
528
529 pud = pud_offset(pgd, address);
530 if (!pud_present(*pud))
531 goto out_unlock;
532
533 pmd = pmd_offset(pud, address);
534 if (!pmd_present(*pmd))
535 goto out_unlock;
536
537 pte = pte_offset_map(pmd, address);
538 if (!pte_present(*pte))
539 goto out_unmap;
540
541 if (page_to_pfn(page) != pte_pfn(*pte))
542 goto out_unmap;
543
544 /*
545 * If the page is mlock()d, we cannot swap it out.
546 * If it's recently referenced (perhaps page_referenced
547 * skipped over this mm) then we should reactivate it.
548 */
549 if ((vma->vm_flags & (VM_LOCKED|VM_RESERVED)) ||
550 ptep_clear_flush_young(vma, address, pte)) {
551 ret = SWAP_FAIL;
552 goto out_unmap;
553 }
554
555 /*
556 * Don't pull an anonymous page out from under get_user_pages.
557 * GUP carefully breaks COW and raises page count (while holding
558 * page_table_lock, as we have here) to make sure that the page
559 * cannot be freed. If we unmap that page here, a user write
560 * access to the virtual address will bring back the page, but
561 * its raised count will (ironically) be taken to mean it's not
562 * an exclusive swap page, do_wp_page will replace it by a copy
563 * page, and the user never get to see the data GUP was holding
564 * the original page for.
565 *
566 * This test is also useful for when swapoff (unuse_process) has
567 * to drop page lock: its reference to the page stops existing
568 * ptes from being unmapped, so swapoff can make progress.
569 */
570 if (PageSwapCache(page) &&
571 page_count(page) != page_mapcount(page) + 2) {
572 ret = SWAP_FAIL;
573 goto out_unmap;
574 }
575
576 /* Nuke the page table entry. */
577 flush_cache_page(vma, address);
578 pteval = ptep_clear_flush(vma, address, pte);
579
580 /* Move the dirty bit to the physical page now the pte is gone. */
581 if (pte_dirty(pteval))
582 set_page_dirty(page);
583
584 if (PageAnon(page)) {
585 swp_entry_t entry = { .val = page->private };
586 /*
587 * Store the swap location in the pte.
588 * See handle_pte_fault() ...
589 */
590 BUG_ON(!PageSwapCache(page));
591 swap_duplicate(entry);
592 if (list_empty(&mm->mmlist)) {
593 spin_lock(&mmlist_lock);
594 list_add(&mm->mmlist, &init_mm.mmlist);
595 spin_unlock(&mmlist_lock);
596 }
597 set_pte(pte, swp_entry_to_pte(entry));
598 BUG_ON(pte_file(*pte));
599 mm->anon_rss--;
600 }
601
602 mm->rss--;
603 acct_update_integrals();
604 page_remove_rmap(page);
605 page_cache_release(page);
606
607 out_unmap:
608 pte_unmap(pte);
609 out_unlock:
610 spin_unlock(&mm->page_table_lock);
611 out:
612 return ret;
613 }
614
615 /*
616 * objrmap doesn't work for nonlinear VMAs because the assumption that
617 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
618 * Consequently, given a particular page and its ->index, we cannot locate the
619 * ptes which are mapping that page without an exhaustive linear search.
620 *
621 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
622 * maps the file to which the target page belongs. The ->vm_private_data field
623 * holds the current cursor into that scan. Successive searches will circulate
624 * around the vma's virtual address space.
625 *
626 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
627 * more scanning pressure is placed against them as well. Eventually pages
628 * will become fully unmapped and are eligible for eviction.
629 *
630 * For very sparsely populated VMAs this is a little inefficient - chances are
631 * there there won't be many ptes located within the scan cluster. In this case
632 * maybe we could scan further - to the end of the pte page, perhaps.
633 */
634 #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
635 #define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
636
637 static void try_to_unmap_cluster(unsigned long cursor,
638 unsigned int *mapcount, struct vm_area_struct *vma)
639 {
640 struct mm_struct *mm = vma->vm_mm;
641 pgd_t *pgd;
642 pud_t *pud;
643 pmd_t *pmd;
644 pte_t *pte;
645 pte_t pteval;
646 struct page *page;
647 unsigned long address;
648 unsigned long end;
649 unsigned long pfn;
650
651 /*
652 * We need the page_table_lock to protect us from page faults,
653 * munmap, fork, etc...
654 */
655 spin_lock(&mm->page_table_lock);
656
657 address = (vma->vm_start + cursor) & CLUSTER_MASK;
658 end = address + CLUSTER_SIZE;
659 if (address < vma->vm_start)
660 address = vma->vm_start;
661 if (end > vma->vm_end)
662 end = vma->vm_end;
663
664 pgd = pgd_offset(mm, address);
665 if (!pgd_present(*pgd))
666 goto out_unlock;
667
668 pud = pud_offset(pgd, address);
669 if (!pud_present(*pud))
670 goto out_unlock;
671
672 pmd = pmd_offset(pud, address);
673 if (!pmd_present(*pmd))
674 goto out_unlock;
675
676 for (pte = pte_offset_map(pmd, address);
677 address < end; pte++, address += PAGE_SIZE) {
678
679 if (!pte_present(*pte))
680 continue;
681
682 pfn = pte_pfn(*pte);
683 if (!pfn_valid(pfn))
684 continue;
685
686 page = pfn_to_page(pfn);
687 BUG_ON(PageAnon(page));
688 if (PageReserved(page))
689 continue;
690
691 if (ptep_clear_flush_young(vma, address, pte))
692 continue;
693
694 /* Nuke the page table entry. */
695 flush_cache_page(vma, address);
696 pteval = ptep_clear_flush(vma, address, pte);
697
698 /* If nonlinear, store the file page offset in the pte. */
699 if (page->index != linear_page_index(vma, address))
700 set_pte(pte, pgoff_to_pte(page->index));
701
702 /* Move the dirty bit to the physical page now the pte is gone. */
703 if (pte_dirty(pteval))
704 set_page_dirty(page);
705
706 page_remove_rmap(page);
707 page_cache_release(page);
708 acct_update_integrals();
709 mm->rss--;
710 (*mapcount)--;
711 }
712
713 pte_unmap(pte);
714
715 out_unlock:
716 spin_unlock(&mm->page_table_lock);
717 }
718
719 static int try_to_unmap_anon(struct page *page)
720 {
721 struct anon_vma *anon_vma;
722 struct vm_area_struct *vma;
723 int ret = SWAP_AGAIN;
724
725 anon_vma = page_lock_anon_vma(page);
726 if (!anon_vma)
727 return ret;
728
729 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
730 ret = try_to_unmap_one(page, vma);
731 if (ret == SWAP_FAIL || !page_mapped(page))
732 break;
733 }
734 spin_unlock(&anon_vma->lock);
735 return ret;
736 }
737
738 /**
739 * try_to_unmap_file - unmap file page using the object-based rmap method
740 * @page: the page to unmap
741 *
742 * Find all the mappings of a page using the mapping pointer and the vma chains
743 * contained in the address_space struct it points to.
744 *
745 * This function is only called from try_to_unmap for object-based pages.
746 */
747 static int try_to_unmap_file(struct page *page)
748 {
749 struct address_space *mapping = page->mapping;
750 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
751 struct vm_area_struct *vma;
752 struct prio_tree_iter iter;
753 int ret = SWAP_AGAIN;
754 unsigned long cursor;
755 unsigned long max_nl_cursor = 0;
756 unsigned long max_nl_size = 0;
757 unsigned int mapcount;
758
759 spin_lock(&mapping->i_mmap_lock);
760 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
761 ret = try_to_unmap_one(page, vma);
762 if (ret == SWAP_FAIL || !page_mapped(page))
763 goto out;
764 }
765
766 if (list_empty(&mapping->i_mmap_nonlinear))
767 goto out;
768
769 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
770 shared.vm_set.list) {
771 if (vma->vm_flags & (VM_LOCKED|VM_RESERVED))
772 continue;
773 cursor = (unsigned long) vma->vm_private_data;
774 if (cursor > max_nl_cursor)
775 max_nl_cursor = cursor;
776 cursor = vma->vm_end - vma->vm_start;
777 if (cursor > max_nl_size)
778 max_nl_size = cursor;
779 }
780
781 if (max_nl_size == 0) { /* any nonlinears locked or reserved */
782 ret = SWAP_FAIL;
783 goto out;
784 }
785
786 /*
787 * We don't try to search for this page in the nonlinear vmas,
788 * and page_referenced wouldn't have found it anyway. Instead
789 * just walk the nonlinear vmas trying to age and unmap some.
790 * The mapcount of the page we came in with is irrelevant,
791 * but even so use it as a guide to how hard we should try?
792 */
793 mapcount = page_mapcount(page);
794 if (!mapcount)
795 goto out;
796 cond_resched_lock(&mapping->i_mmap_lock);
797
798 max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
799 if (max_nl_cursor == 0)
800 max_nl_cursor = CLUSTER_SIZE;
801
802 do {
803 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
804 shared.vm_set.list) {
805 if (vma->vm_flags & (VM_LOCKED|VM_RESERVED))
806 continue;
807 cursor = (unsigned long) vma->vm_private_data;
808 while (vma->vm_mm->rss &&
809 cursor < max_nl_cursor &&
810 cursor < vma->vm_end - vma->vm_start) {
811 try_to_unmap_cluster(cursor, &mapcount, vma);
812 cursor += CLUSTER_SIZE;
813 vma->vm_private_data = (void *) cursor;
814 if ((int)mapcount <= 0)
815 goto out;
816 }
817 vma->vm_private_data = (void *) max_nl_cursor;
818 }
819 cond_resched_lock(&mapping->i_mmap_lock);
820 max_nl_cursor += CLUSTER_SIZE;
821 } while (max_nl_cursor <= max_nl_size);
822
823 /*
824 * Don't loop forever (perhaps all the remaining pages are
825 * in locked vmas). Reset cursor on all unreserved nonlinear
826 * vmas, now forgetting on which ones it had fallen behind.
827 */
828 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
829 shared.vm_set.list) {
830 if (!(vma->vm_flags & VM_RESERVED))
831 vma->vm_private_data = NULL;
832 }
833 out:
834 spin_unlock(&mapping->i_mmap_lock);
835 return ret;
836 }
837
838 /**
839 * try_to_unmap - try to remove all page table mappings to a page
840 * @page: the page to get unmapped
841 *
842 * Tries to remove all the page table entries which are mapping this
843 * page, used in the pageout path. Caller must hold the page lock.
844 * Return values are:
845 *
846 * SWAP_SUCCESS - we succeeded in removing all mappings
847 * SWAP_AGAIN - we missed a mapping, try again later
848 * SWAP_FAIL - the page is unswappable
849 */
850 int try_to_unmap(struct page *page)
851 {
852 int ret;
853
854 BUG_ON(PageReserved(page));
855 BUG_ON(!PageLocked(page));
856
857 if (PageAnon(page))
858 ret = try_to_unmap_anon(page);
859 else
860 ret = try_to_unmap_file(page);
861
862 if (!page_mapped(page))
863 ret = SWAP_SUCCESS;
864 return ret;
865 }
866
|
This page was automatically generated by the
LXR engine.
|