]>
git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/blob - mm/rmap.c
2 * mm/rmap.c - physical to virtual reverse mappings
4 * Copyright 2001, Rik van Riel <riel@conectiva.com.br>
5 * Released under the General Public License (GPL).
7 * Simple, low overhead reverse mapping scheme.
8 * Please try to keep this thing as modular as possible.
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.
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
21 * Lock ordering in mm:
23 * inode->i_mutex (while writing or truncating, not reading or faulting)
26 * When a page fault occurs in writing from user to file, down_read
27 * 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.
32 * page->flags PG_locked (lock_page)
33 * mapping->i_mmap_lock
35 * mm->page_table_lock or pte_lock
36 * zone->lru_lock (in mark_page_accessed, isolate_lru_page)
37 * swap_lock (in swap_duplicate, swap_info_get)
38 * mmlist_lock (in mmput, drain_mmlist and others)
39 * 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)
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 #include <linux/module.h>
57 #include <asm/tlbflush.h>
59 struct kmem_cache
*anon_vma_cachep
;
61 static inline void validate_anon_vma(struct vm_area_struct
*find_vma
)
63 #ifdef CONFIG_DEBUG_VM
64 struct anon_vma
*anon_vma
= find_vma
->anon_vma
;
65 struct vm_area_struct
*vma
;
66 unsigned int mapcount
= 0;
69 list_for_each_entry(vma
, &anon_vma
->head
, anon_vma_node
) {
71 BUG_ON(mapcount
> 100000);
79 /* This must be called under the mmap_sem. */
80 int anon_vma_prepare(struct vm_area_struct
*vma
)
82 struct anon_vma
*anon_vma
= vma
->anon_vma
;
85 if (unlikely(!anon_vma
)) {
86 struct mm_struct
*mm
= vma
->vm_mm
;
87 struct anon_vma
*allocated
, *locked
;
89 anon_vma
= find_mergeable_anon_vma(vma
);
93 spin_lock(&locked
->lock
);
95 anon_vma
= anon_vma_alloc();
96 if (unlikely(!anon_vma
))
102 /* page_table_lock to protect against threads */
103 spin_lock(&mm
->page_table_lock
);
104 if (likely(!vma
->anon_vma
)) {
105 vma
->anon_vma
= anon_vma
;
106 list_add_tail(&vma
->anon_vma_node
, &anon_vma
->head
);
109 spin_unlock(&mm
->page_table_lock
);
112 spin_unlock(&locked
->lock
);
113 if (unlikely(allocated
))
114 anon_vma_free(allocated
);
119 void __anon_vma_merge(struct vm_area_struct
*vma
, struct vm_area_struct
*next
)
121 BUG_ON(vma
->anon_vma
!= next
->anon_vma
);
122 list_del(&next
->anon_vma_node
);
125 void __anon_vma_link(struct vm_area_struct
*vma
)
127 struct anon_vma
*anon_vma
= vma
->anon_vma
;
130 list_add_tail(&vma
->anon_vma_node
, &anon_vma
->head
);
131 validate_anon_vma(vma
);
135 void anon_vma_link(struct vm_area_struct
*vma
)
137 struct anon_vma
*anon_vma
= vma
->anon_vma
;
140 spin_lock(&anon_vma
->lock
);
141 list_add_tail(&vma
->anon_vma_node
, &anon_vma
->head
);
142 validate_anon_vma(vma
);
143 spin_unlock(&anon_vma
->lock
);
147 void anon_vma_unlink(struct vm_area_struct
*vma
)
149 struct anon_vma
*anon_vma
= vma
->anon_vma
;
155 spin_lock(&anon_vma
->lock
);
156 validate_anon_vma(vma
);
157 list_del(&vma
->anon_vma_node
);
159 /* We must garbage collect the anon_vma if it's empty */
160 empty
= list_empty(&anon_vma
->head
);
161 spin_unlock(&anon_vma
->lock
);
164 anon_vma_free(anon_vma
);
167 static void anon_vma_ctor(void *data
, struct kmem_cache
*cachep
,
170 if ((flags
& (SLAB_CTOR_VERIFY
|SLAB_CTOR_CONSTRUCTOR
)) ==
171 SLAB_CTOR_CONSTRUCTOR
) {
172 struct anon_vma
*anon_vma
= data
;
174 spin_lock_init(&anon_vma
->lock
);
175 INIT_LIST_HEAD(&anon_vma
->head
);
179 void __init
anon_vma_init(void)
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
);
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.
189 static struct anon_vma
*page_lock_anon_vma(struct page
*page
)
191 struct anon_vma
*anon_vma
= NULL
;
192 unsigned long anon_mapping
;
195 anon_mapping
= (unsigned long) page
->mapping
;
196 if (!(anon_mapping
& PAGE_MAPPING_ANON
))
198 if (!page_mapped(page
))
201 anon_vma
= (struct anon_vma
*) (anon_mapping
- PAGE_MAPPING_ANON
);
202 spin_lock(&anon_vma
->lock
);
209 * At what user virtual address is page expected in vma?
211 static inline unsigned long
212 vma_address(struct page
*page
, struct vm_area_struct
*vma
)
214 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
215 unsigned long address
;
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
));
227 * At what user virtual address is page expected in vma? checking that the
228 * page matches the vma: currently only used on anon pages, by unuse_vma;
230 unsigned long page_address_in_vma(struct page
*page
, struct vm_area_struct
*vma
)
232 if (PageAnon(page
)) {
233 if ((void *)vma
->anon_vma
!=
234 (void *)page
->mapping
- PAGE_MAPPING_ANON
)
236 } else if (page
->mapping
&& !(vma
->vm_flags
& VM_NONLINEAR
)) {
238 vma
->vm_file
->f_mapping
!= page
->mapping
)
242 return vma_address(page
, vma
);
246 * Check that @page is mapped at @address into @mm.
248 * On success returns with pte mapped and locked.
250 pte_t
*page_check_address(struct page
*page
, struct mm_struct
*mm
,
251 unsigned long address
, spinlock_t
**ptlp
)
259 pgd
= pgd_offset(mm
, address
);
260 if (!pgd_present(*pgd
))
263 pud
= pud_offset(pgd
, address
);
264 if (!pud_present(*pud
))
267 pmd
= pmd_offset(pud
, address
);
268 if (!pmd_present(*pmd
))
271 pte
= pte_offset_map(pmd
, address
);
272 /* Make a quick check before getting the lock */
273 if (!pte_present(*pte
)) {
278 ptl
= pte_lockptr(mm
, pmd
);
280 if (pte_present(*pte
) && page_to_pfn(page
) == pte_pfn(*pte
)) {
284 pte_unmap_unlock(pte
, ptl
);
289 * Subfunctions of page_referenced: page_referenced_one called
290 * repeatedly from either page_referenced_anon or page_referenced_file.
292 static int page_referenced_one(struct page
*page
,
293 struct vm_area_struct
*vma
, unsigned int *mapcount
)
295 struct mm_struct
*mm
= vma
->vm_mm
;
296 unsigned long address
;
301 address
= vma_address(page
, vma
);
302 if (address
== -EFAULT
)
305 pte
= page_check_address(page
, mm
, address
, &ptl
);
309 if (ptep_clear_flush_young(vma
, address
, pte
))
312 /* Pretend the page is referenced if the task has the
313 swap token and is in the middle of a page fault. */
314 if (mm
!= current
->mm
&& has_swap_token(mm
) &&
315 rwsem_is_locked(&mm
->mmap_sem
))
319 pte_unmap_unlock(pte
, ptl
);
324 static int page_referenced_anon(struct page
*page
)
326 unsigned int mapcount
;
327 struct anon_vma
*anon_vma
;
328 struct vm_area_struct
*vma
;
331 anon_vma
= page_lock_anon_vma(page
);
335 mapcount
= page_mapcount(page
);
336 list_for_each_entry(vma
, &anon_vma
->head
, anon_vma_node
) {
337 referenced
+= page_referenced_one(page
, vma
, &mapcount
);
341 spin_unlock(&anon_vma
->lock
);
346 * page_referenced_file - referenced check for object-based rmap
347 * @page: the page we're checking references on.
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.
354 * This function is only called from page_referenced for object-based pages.
356 static int page_referenced_file(struct page
*page
)
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
;
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.
370 BUG_ON(PageAnon(page
));
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.
378 BUG_ON(!PageLocked(page
));
380 spin_lock(&mapping
->i_mmap_lock
);
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.
386 mapcount
= page_mapcount(page
);
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
)) {
394 referenced
+= page_referenced_one(page
, vma
, &mapcount
);
399 spin_unlock(&mapping
->i_mmap_lock
);
404 * page_referenced - test if the page was referenced
405 * @page: the page to test
406 * @is_locked: caller holds lock on the page
408 * Quick test_and_clear_referenced for all mappings to a page,
409 * returns the number of ptes which referenced the page.
411 int page_referenced(struct page
*page
, int is_locked
)
415 if (page_test_and_clear_young(page
))
418 if (TestClearPageReferenced(page
))
421 if (page_mapped(page
) && page
->mapping
) {
423 referenced
+= page_referenced_anon(page
);
425 referenced
+= page_referenced_file(page
);
426 else if (TestSetPageLocked(page
))
430 referenced
+= page_referenced_file(page
);
437 static int page_mkclean_one(struct page
*page
, struct vm_area_struct
*vma
)
439 struct mm_struct
*mm
= vma
->vm_mm
;
440 unsigned long address
;
445 address
= vma_address(page
, vma
);
446 if (address
== -EFAULT
)
449 pte
= page_check_address(page
, mm
, address
, &ptl
);
453 if (!pte_dirty(*pte
) && !pte_write(*pte
))
456 entry
= ptep_get_and_clear(mm
, address
, pte
);
457 entry
= pte_mkclean(entry
);
458 entry
= pte_wrprotect(entry
);
459 ptep_establish(vma
, address
, pte
, entry
);
460 lazy_mmu_prot_update(entry
);
464 pte_unmap_unlock(pte
, ptl
);
469 static int page_mkclean_file(struct address_space
*mapping
, struct page
*page
)
471 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
472 struct vm_area_struct
*vma
;
473 struct prio_tree_iter iter
;
476 BUG_ON(PageAnon(page
));
478 spin_lock(&mapping
->i_mmap_lock
);
479 vma_prio_tree_foreach(vma
, &iter
, &mapping
->i_mmap
, pgoff
, pgoff
) {
480 if (vma
->vm_flags
& VM_SHARED
)
481 ret
+= page_mkclean_one(page
, vma
);
483 spin_unlock(&mapping
->i_mmap_lock
);
487 int page_mkclean(struct page
*page
)
491 BUG_ON(!PageLocked(page
));
493 if (page_mapped(page
)) {
494 struct address_space
*mapping
= page_mapping(page
);
496 ret
= page_mkclean_file(mapping
, page
);
503 * page_set_anon_rmap - setup new anonymous rmap
504 * @page: the page to add the mapping to
505 * @vma: the vm area in which the mapping is added
506 * @address: the user virtual address mapped
508 static void __page_set_anon_rmap(struct page
*page
,
509 struct vm_area_struct
*vma
, unsigned long address
)
511 struct anon_vma
*anon_vma
= vma
->anon_vma
;
514 anon_vma
= (void *) anon_vma
+ PAGE_MAPPING_ANON
;
515 page
->mapping
= (struct address_space
*) anon_vma
;
517 page
->index
= linear_page_index(vma
, address
);
520 * nr_mapped state can be updated without turning off
521 * interrupts because it is not modified via interrupt.
523 __inc_zone_page_state(page
, NR_ANON_PAGES
);
527 * page_add_anon_rmap - add pte mapping to an anonymous page
528 * @page: the page to add the mapping to
529 * @vma: the vm area in which the mapping is added
530 * @address: the user virtual address mapped
532 * The caller needs to hold the pte lock.
534 void page_add_anon_rmap(struct page
*page
,
535 struct vm_area_struct
*vma
, unsigned long address
)
537 if (atomic_inc_and_test(&page
->_mapcount
))
538 __page_set_anon_rmap(page
, vma
, address
);
539 /* else checking page index and mapping is racy */
543 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
544 * @page: the page to add the mapping to
545 * @vma: the vm area in which the mapping is added
546 * @address: the user virtual address mapped
548 * Same as page_add_anon_rmap but must only be called on *new* pages.
549 * This means the inc-and-test can be bypassed.
551 void page_add_new_anon_rmap(struct page
*page
,
552 struct vm_area_struct
*vma
, unsigned long address
)
554 atomic_set(&page
->_mapcount
, 0); /* elevate count by 1 (starts at -1) */
555 __page_set_anon_rmap(page
, vma
, address
);
559 * page_add_file_rmap - add pte mapping to a file page
560 * @page: the page to add the mapping to
562 * The caller needs to hold the pte lock.
564 void page_add_file_rmap(struct page
*page
)
566 if (atomic_inc_and_test(&page
->_mapcount
))
567 __inc_zone_page_state(page
, NR_FILE_MAPPED
);
571 * page_remove_rmap - take down pte mapping from a page
572 * @page: page to remove mapping from
574 * The caller needs to hold the pte lock.
576 void page_remove_rmap(struct page
*page
)
578 if (atomic_add_negative(-1, &page
->_mapcount
)) {
579 if (unlikely(page_mapcount(page
) < 0)) {
580 printk (KERN_EMERG
"Eeek! page_mapcount(page) went negative! (%d)\n", page_mapcount(page
));
581 printk (KERN_EMERG
" page->flags = %lx\n", page
->flags
);
582 printk (KERN_EMERG
" page->count = %x\n", page_count(page
));
583 printk (KERN_EMERG
" page->mapping = %p\n", page
->mapping
);
588 * It would be tidy to reset the PageAnon mapping here,
589 * but that might overwrite a racing page_add_anon_rmap
590 * which increments mapcount after us but sets mapping
591 * before us: so leave the reset to free_hot_cold_page,
592 * and remember that it's only reliable while mapped.
593 * Leaving it set also helps swapoff to reinstate ptes
594 * faster for those pages still in swapcache.
596 if (page_test_and_clear_dirty(page
))
597 set_page_dirty(page
);
598 __dec_zone_page_state(page
,
599 PageAnon(page
) ? NR_ANON_PAGES
: NR_FILE_MAPPED
);
604 * Subfunctions of try_to_unmap: try_to_unmap_one called
605 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
607 static int try_to_unmap_one(struct page
*page
, struct vm_area_struct
*vma
,
610 struct mm_struct
*mm
= vma
->vm_mm
;
611 unsigned long address
;
615 int ret
= SWAP_AGAIN
;
617 address
= vma_address(page
, vma
);
618 if (address
== -EFAULT
)
621 pte
= page_check_address(page
, mm
, address
, &ptl
);
626 * If the page is mlock()d, we cannot swap it out.
627 * If it's recently referenced (perhaps page_referenced
628 * skipped over this mm) then we should reactivate it.
630 if (!migration
&& ((vma
->vm_flags
& VM_LOCKED
) ||
631 (ptep_clear_flush_young(vma
, address
, pte
)))) {
636 /* Nuke the page table entry. */
637 flush_cache_page(vma
, address
, page_to_pfn(page
));
638 pteval
= ptep_clear_flush(vma
, address
, pte
);
640 /* Move the dirty bit to the physical page now the pte is gone. */
641 if (pte_dirty(pteval
))
642 set_page_dirty(page
);
644 /* Update high watermark before we lower rss */
645 update_hiwater_rss(mm
);
647 if (PageAnon(page
)) {
648 swp_entry_t entry
= { .val
= page_private(page
) };
650 if (PageSwapCache(page
)) {
652 * Store the swap location in the pte.
653 * See handle_pte_fault() ...
655 swap_duplicate(entry
);
656 if (list_empty(&mm
->mmlist
)) {
657 spin_lock(&mmlist_lock
);
658 if (list_empty(&mm
->mmlist
))
659 list_add(&mm
->mmlist
, &init_mm
.mmlist
);
660 spin_unlock(&mmlist_lock
);
662 dec_mm_counter(mm
, anon_rss
);
663 #ifdef CONFIG_MIGRATION
666 * Store the pfn of the page in a special migration
667 * pte. do_swap_page() will wait until the migration
668 * pte is removed and then restart fault handling.
671 entry
= make_migration_entry(page
, pte_write(pteval
));
674 set_pte_at(mm
, address
, pte
, swp_entry_to_pte(entry
));
675 BUG_ON(pte_file(*pte
));
677 #ifdef CONFIG_MIGRATION
679 /* Establish migration entry for a file page */
681 entry
= make_migration_entry(page
, pte_write(pteval
));
682 set_pte_at(mm
, address
, pte
, swp_entry_to_pte(entry
));
685 dec_mm_counter(mm
, file_rss
);
688 page_remove_rmap(page
);
689 page_cache_release(page
);
692 pte_unmap_unlock(pte
, ptl
);
698 * objrmap doesn't work for nonlinear VMAs because the assumption that
699 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
700 * Consequently, given a particular page and its ->index, we cannot locate the
701 * ptes which are mapping that page without an exhaustive linear search.
703 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
704 * maps the file to which the target page belongs. The ->vm_private_data field
705 * holds the current cursor into that scan. Successive searches will circulate
706 * around the vma's virtual address space.
708 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
709 * more scanning pressure is placed against them as well. Eventually pages
710 * will become fully unmapped and are eligible for eviction.
712 * For very sparsely populated VMAs this is a little inefficient - chances are
713 * there there won't be many ptes located within the scan cluster. In this case
714 * maybe we could scan further - to the end of the pte page, perhaps.
716 #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
717 #define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
719 static void try_to_unmap_cluster(unsigned long cursor
,
720 unsigned int *mapcount
, struct vm_area_struct
*vma
)
722 struct mm_struct
*mm
= vma
->vm_mm
;
730 unsigned long address
;
733 address
= (vma
->vm_start
+ cursor
) & CLUSTER_MASK
;
734 end
= address
+ CLUSTER_SIZE
;
735 if (address
< vma
->vm_start
)
736 address
= vma
->vm_start
;
737 if (end
> vma
->vm_end
)
740 pgd
= pgd_offset(mm
, address
);
741 if (!pgd_present(*pgd
))
744 pud
= pud_offset(pgd
, address
);
745 if (!pud_present(*pud
))
748 pmd
= pmd_offset(pud
, address
);
749 if (!pmd_present(*pmd
))
752 pte
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
754 /* Update high watermark before we lower rss */
755 update_hiwater_rss(mm
);
757 for (; address
< end
; pte
++, address
+= PAGE_SIZE
) {
758 if (!pte_present(*pte
))
760 page
= vm_normal_page(vma
, address
, *pte
);
761 BUG_ON(!page
|| PageAnon(page
));
763 if (ptep_clear_flush_young(vma
, address
, pte
))
766 /* Nuke the page table entry. */
767 flush_cache_page(vma
, address
, pte_pfn(*pte
));
768 pteval
= ptep_clear_flush(vma
, address
, pte
);
770 /* If nonlinear, store the file page offset in the pte. */
771 if (page
->index
!= linear_page_index(vma
, address
))
772 set_pte_at(mm
, address
, pte
, pgoff_to_pte(page
->index
));
774 /* Move the dirty bit to the physical page now the pte is gone. */
775 if (pte_dirty(pteval
))
776 set_page_dirty(page
);
778 page_remove_rmap(page
);
779 page_cache_release(page
);
780 dec_mm_counter(mm
, file_rss
);
783 pte_unmap_unlock(pte
- 1, ptl
);
786 static int try_to_unmap_anon(struct page
*page
, int migration
)
788 struct anon_vma
*anon_vma
;
789 struct vm_area_struct
*vma
;
790 int ret
= SWAP_AGAIN
;
792 anon_vma
= page_lock_anon_vma(page
);
796 list_for_each_entry(vma
, &anon_vma
->head
, anon_vma_node
) {
797 ret
= try_to_unmap_one(page
, vma
, migration
);
798 if (ret
== SWAP_FAIL
|| !page_mapped(page
))
801 spin_unlock(&anon_vma
->lock
);
806 * try_to_unmap_file - unmap file page using the object-based rmap method
807 * @page: the page to unmap
809 * Find all the mappings of a page using the mapping pointer and the vma chains
810 * contained in the address_space struct it points to.
812 * This function is only called from try_to_unmap for object-based pages.
814 static int try_to_unmap_file(struct page
*page
, int migration
)
816 struct address_space
*mapping
= page
->mapping
;
817 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
818 struct vm_area_struct
*vma
;
819 struct prio_tree_iter iter
;
820 int ret
= SWAP_AGAIN
;
821 unsigned long cursor
;
822 unsigned long max_nl_cursor
= 0;
823 unsigned long max_nl_size
= 0;
824 unsigned int mapcount
;
826 spin_lock(&mapping
->i_mmap_lock
);
827 vma_prio_tree_foreach(vma
, &iter
, &mapping
->i_mmap
, pgoff
, pgoff
) {
828 ret
= try_to_unmap_one(page
, vma
, migration
);
829 if (ret
== SWAP_FAIL
|| !page_mapped(page
))
833 if (list_empty(&mapping
->i_mmap_nonlinear
))
836 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
,
837 shared
.vm_set
.list
) {
838 if ((vma
->vm_flags
& VM_LOCKED
) && !migration
)
840 cursor
= (unsigned long) vma
->vm_private_data
;
841 if (cursor
> max_nl_cursor
)
842 max_nl_cursor
= cursor
;
843 cursor
= vma
->vm_end
- vma
->vm_start
;
844 if (cursor
> max_nl_size
)
845 max_nl_size
= cursor
;
848 if (max_nl_size
== 0) { /* any nonlinears locked or reserved */
854 * We don't try to search for this page in the nonlinear vmas,
855 * and page_referenced wouldn't have found it anyway. Instead
856 * just walk the nonlinear vmas trying to age and unmap some.
857 * The mapcount of the page we came in with is irrelevant,
858 * but even so use it as a guide to how hard we should try?
860 mapcount
= page_mapcount(page
);
863 cond_resched_lock(&mapping
->i_mmap_lock
);
865 max_nl_size
= (max_nl_size
+ CLUSTER_SIZE
- 1) & CLUSTER_MASK
;
866 if (max_nl_cursor
== 0)
867 max_nl_cursor
= CLUSTER_SIZE
;
870 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
,
871 shared
.vm_set
.list
) {
872 if ((vma
->vm_flags
& VM_LOCKED
) && !migration
)
874 cursor
= (unsigned long) vma
->vm_private_data
;
875 while ( cursor
< max_nl_cursor
&&
876 cursor
< vma
->vm_end
- vma
->vm_start
) {
877 try_to_unmap_cluster(cursor
, &mapcount
, vma
);
878 cursor
+= CLUSTER_SIZE
;
879 vma
->vm_private_data
= (void *) cursor
;
880 if ((int)mapcount
<= 0)
883 vma
->vm_private_data
= (void *) max_nl_cursor
;
885 cond_resched_lock(&mapping
->i_mmap_lock
);
886 max_nl_cursor
+= CLUSTER_SIZE
;
887 } while (max_nl_cursor
<= max_nl_size
);
890 * Don't loop forever (perhaps all the remaining pages are
891 * in locked vmas). Reset cursor on all unreserved nonlinear
892 * vmas, now forgetting on which ones it had fallen behind.
894 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
, shared
.vm_set
.list
)
895 vma
->vm_private_data
= NULL
;
897 spin_unlock(&mapping
->i_mmap_lock
);
902 * try_to_unmap - try to remove all page table mappings to a page
903 * @page: the page to get unmapped
905 * Tries to remove all the page table entries which are mapping this
906 * page, used in the pageout path. Caller must hold the page lock.
909 * SWAP_SUCCESS - we succeeded in removing all mappings
910 * SWAP_AGAIN - we missed a mapping, try again later
911 * SWAP_FAIL - the page is unswappable
913 int try_to_unmap(struct page
*page
, int migration
)
917 BUG_ON(!PageLocked(page
));
920 ret
= try_to_unmap_anon(page
, migration
);
922 ret
= try_to_unmap_file(page
, migration
);
924 if (!page_mapped(page
))