]>
git.proxmox.com Git - mirror_ubuntu-zesty-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)
24 * inode->i_alloc_sem (vmtruncate_range)
26 * page->flags PG_locked (lock_page)
27 * mapping->i_mmap_lock
29 * mm->page_table_lock or pte_lock
30 * zone->lru_lock (in mark_page_accessed, isolate_lru_page)
31 * swap_lock (in swap_duplicate, swap_info_get)
32 * mmlist_lock (in mmput, drain_mmlist and others)
33 * mapping->private_lock (in __set_page_dirty_buffers)
34 * inode_lock (in set_page_dirty's __mark_inode_dirty)
35 * sb_lock (within inode_lock in fs/fs-writeback.c)
36 * mapping->tree_lock (widely used, in set_page_dirty,
37 * in arch-dependent flush_dcache_mmap_lock,
38 * within inode_lock in __sync_single_inode)
42 #include <linux/pagemap.h>
43 #include <linux/swap.h>
44 #include <linux/swapops.h>
45 #include <linux/slab.h>
46 #include <linux/init.h>
47 #include <linux/rmap.h>
48 #include <linux/rcupdate.h>
49 #include <linux/module.h>
50 #include <linux/kallsyms.h>
51 #include <linux/memcontrol.h>
52 #include <linux/mmu_notifier.h>
54 #include <asm/tlbflush.h>
56 struct kmem_cache
*anon_vma_cachep
;
58 /* This must be called under the mmap_sem. */
59 int anon_vma_prepare(struct vm_area_struct
*vma
)
61 struct anon_vma
*anon_vma
= vma
->anon_vma
;
64 if (unlikely(!anon_vma
)) {
65 struct mm_struct
*mm
= vma
->vm_mm
;
66 struct anon_vma
*allocated
, *locked
;
68 anon_vma
= find_mergeable_anon_vma(vma
);
72 spin_lock(&locked
->lock
);
74 anon_vma
= anon_vma_alloc();
75 if (unlikely(!anon_vma
))
81 /* page_table_lock to protect against threads */
82 spin_lock(&mm
->page_table_lock
);
83 if (likely(!vma
->anon_vma
)) {
84 vma
->anon_vma
= anon_vma
;
85 list_add_tail(&vma
->anon_vma_node
, &anon_vma
->head
);
88 spin_unlock(&mm
->page_table_lock
);
91 spin_unlock(&locked
->lock
);
92 if (unlikely(allocated
))
93 anon_vma_free(allocated
);
98 void __anon_vma_merge(struct vm_area_struct
*vma
, struct vm_area_struct
*next
)
100 BUG_ON(vma
->anon_vma
!= next
->anon_vma
);
101 list_del(&next
->anon_vma_node
);
104 void __anon_vma_link(struct vm_area_struct
*vma
)
106 struct anon_vma
*anon_vma
= vma
->anon_vma
;
109 list_add_tail(&vma
->anon_vma_node
, &anon_vma
->head
);
112 void anon_vma_link(struct vm_area_struct
*vma
)
114 struct anon_vma
*anon_vma
= vma
->anon_vma
;
117 spin_lock(&anon_vma
->lock
);
118 list_add_tail(&vma
->anon_vma_node
, &anon_vma
->head
);
119 spin_unlock(&anon_vma
->lock
);
123 void anon_vma_unlink(struct vm_area_struct
*vma
)
125 struct anon_vma
*anon_vma
= vma
->anon_vma
;
131 spin_lock(&anon_vma
->lock
);
132 list_del(&vma
->anon_vma_node
);
134 /* We must garbage collect the anon_vma if it's empty */
135 empty
= list_empty(&anon_vma
->head
);
136 spin_unlock(&anon_vma
->lock
);
139 anon_vma_free(anon_vma
);
142 static void anon_vma_ctor(void *data
)
144 struct anon_vma
*anon_vma
= data
;
146 spin_lock_init(&anon_vma
->lock
);
147 INIT_LIST_HEAD(&anon_vma
->head
);
150 void __init
anon_vma_init(void)
152 anon_vma_cachep
= kmem_cache_create("anon_vma", sizeof(struct anon_vma
),
153 0, SLAB_DESTROY_BY_RCU
|SLAB_PANIC
, anon_vma_ctor
);
157 * Getting a lock on a stable anon_vma from a page off the LRU is
158 * tricky: page_lock_anon_vma rely on RCU to guard against the races.
160 static struct anon_vma
*page_lock_anon_vma(struct page
*page
)
162 struct anon_vma
*anon_vma
;
163 unsigned long anon_mapping
;
166 anon_mapping
= (unsigned long) page
->mapping
;
167 if (!(anon_mapping
& PAGE_MAPPING_ANON
))
169 if (!page_mapped(page
))
172 anon_vma
= (struct anon_vma
*) (anon_mapping
- PAGE_MAPPING_ANON
);
173 spin_lock(&anon_vma
->lock
);
180 static void page_unlock_anon_vma(struct anon_vma
*anon_vma
)
182 spin_unlock(&anon_vma
->lock
);
187 * At what user virtual address is page expected in @vma?
188 * Returns virtual address or -EFAULT if page's index/offset is not
189 * within the range mapped the @vma.
191 static inline unsigned long
192 vma_address(struct page
*page
, struct vm_area_struct
*vma
)
194 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
195 unsigned long address
;
197 address
= vma
->vm_start
+ ((pgoff
- vma
->vm_pgoff
) << PAGE_SHIFT
);
198 if (unlikely(address
< vma
->vm_start
|| address
>= vma
->vm_end
)) {
199 /* page should be within @vma mapping range */
206 * At what user virtual address is page expected in vma? checking that the
207 * page matches the vma: currently only used on anon pages, by unuse_vma;
209 unsigned long page_address_in_vma(struct page
*page
, struct vm_area_struct
*vma
)
211 if (PageAnon(page
)) {
212 if ((void *)vma
->anon_vma
!=
213 (void *)page
->mapping
- PAGE_MAPPING_ANON
)
215 } else if (page
->mapping
&& !(vma
->vm_flags
& VM_NONLINEAR
)) {
217 vma
->vm_file
->f_mapping
!= page
->mapping
)
221 return vma_address(page
, vma
);
225 * Check that @page is mapped at @address into @mm.
227 * On success returns with pte mapped and locked.
229 pte_t
*page_check_address(struct page
*page
, struct mm_struct
*mm
,
230 unsigned long address
, spinlock_t
**ptlp
)
238 pgd
= pgd_offset(mm
, address
);
239 if (!pgd_present(*pgd
))
242 pud
= pud_offset(pgd
, address
);
243 if (!pud_present(*pud
))
246 pmd
= pmd_offset(pud
, address
);
247 if (!pmd_present(*pmd
))
250 pte
= pte_offset_map(pmd
, address
);
251 /* Make a quick check before getting the lock */
252 if (!pte_present(*pte
)) {
257 ptl
= pte_lockptr(mm
, pmd
);
259 if (pte_present(*pte
) && page_to_pfn(page
) == pte_pfn(*pte
)) {
263 pte_unmap_unlock(pte
, ptl
);
268 * Subfunctions of page_referenced: page_referenced_one called
269 * repeatedly from either page_referenced_anon or page_referenced_file.
271 static int page_referenced_one(struct page
*page
,
272 struct vm_area_struct
*vma
, unsigned int *mapcount
)
274 struct mm_struct
*mm
= vma
->vm_mm
;
275 unsigned long address
;
280 address
= vma_address(page
, vma
);
281 if (address
== -EFAULT
)
284 pte
= page_check_address(page
, mm
, address
, &ptl
);
288 if (vma
->vm_flags
& VM_LOCKED
) {
290 *mapcount
= 1; /* break early from loop */
291 } else if (ptep_clear_flush_young_notify(vma
, address
, pte
))
294 /* Pretend the page is referenced if the task has the
295 swap token and is in the middle of a page fault. */
296 if (mm
!= current
->mm
&& has_swap_token(mm
) &&
297 rwsem_is_locked(&mm
->mmap_sem
))
301 pte_unmap_unlock(pte
, ptl
);
306 static int page_referenced_anon(struct page
*page
,
307 struct mem_cgroup
*mem_cont
)
309 unsigned int mapcount
;
310 struct anon_vma
*anon_vma
;
311 struct vm_area_struct
*vma
;
314 anon_vma
= page_lock_anon_vma(page
);
318 mapcount
= page_mapcount(page
);
319 list_for_each_entry(vma
, &anon_vma
->head
, anon_vma_node
) {
321 * If we are reclaiming on behalf of a cgroup, skip
322 * counting on behalf of references from different
325 if (mem_cont
&& !mm_match_cgroup(vma
->vm_mm
, mem_cont
))
327 referenced
+= page_referenced_one(page
, vma
, &mapcount
);
332 page_unlock_anon_vma(anon_vma
);
337 * page_referenced_file - referenced check for object-based rmap
338 * @page: the page we're checking references on.
339 * @mem_cont: target memory controller
341 * For an object-based mapped page, find all the places it is mapped and
342 * check/clear the referenced flag. This is done by following the page->mapping
343 * pointer, then walking the chain of vmas it holds. It returns the number
344 * of references it found.
346 * This function is only called from page_referenced for object-based pages.
348 static int page_referenced_file(struct page
*page
,
349 struct mem_cgroup
*mem_cont
)
351 unsigned int mapcount
;
352 struct address_space
*mapping
= page
->mapping
;
353 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
354 struct vm_area_struct
*vma
;
355 struct prio_tree_iter iter
;
359 * The caller's checks on page->mapping and !PageAnon have made
360 * sure that this is a file page: the check for page->mapping
361 * excludes the case just before it gets set on an anon page.
363 BUG_ON(PageAnon(page
));
366 * The page lock not only makes sure that page->mapping cannot
367 * suddenly be NULLified by truncation, it makes sure that the
368 * structure at mapping cannot be freed and reused yet,
369 * so we can safely take mapping->i_mmap_lock.
371 BUG_ON(!PageLocked(page
));
373 spin_lock(&mapping
->i_mmap_lock
);
376 * i_mmap_lock does not stabilize mapcount at all, but mapcount
377 * is more likely to be accurate if we note it after spinning.
379 mapcount
= page_mapcount(page
);
381 vma_prio_tree_foreach(vma
, &iter
, &mapping
->i_mmap
, pgoff
, pgoff
) {
383 * If we are reclaiming on behalf of a cgroup, skip
384 * counting on behalf of references from different
387 if (mem_cont
&& !mm_match_cgroup(vma
->vm_mm
, mem_cont
))
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
407 * @mem_cont: target memory controller
409 * Quick test_and_clear_referenced for all mappings to a page,
410 * returns the number of ptes which referenced the page.
412 int page_referenced(struct page
*page
, int is_locked
,
413 struct mem_cgroup
*mem_cont
)
417 if (TestClearPageReferenced(page
))
420 if (page_mapped(page
) && page
->mapping
) {
422 referenced
+= page_referenced_anon(page
, mem_cont
);
424 referenced
+= page_referenced_file(page
, mem_cont
);
425 else if (TestSetPageLocked(page
))
430 page_referenced_file(page
, mem_cont
);
435 if (page_test_and_clear_young(page
))
441 static int page_mkclean_one(struct page
*page
, struct vm_area_struct
*vma
)
443 struct mm_struct
*mm
= vma
->vm_mm
;
444 unsigned long address
;
449 address
= vma_address(page
, vma
);
450 if (address
== -EFAULT
)
453 pte
= page_check_address(page
, mm
, address
, &ptl
);
457 if (pte_dirty(*pte
) || pte_write(*pte
)) {
460 flush_cache_page(vma
, address
, pte_pfn(*pte
));
461 entry
= ptep_clear_flush_notify(vma
, address
, pte
);
462 entry
= pte_wrprotect(entry
);
463 entry
= pte_mkclean(entry
);
464 set_pte_at(mm
, address
, pte
, entry
);
468 pte_unmap_unlock(pte
, ptl
);
473 static int page_mkclean_file(struct address_space
*mapping
, struct page
*page
)
475 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
476 struct vm_area_struct
*vma
;
477 struct prio_tree_iter iter
;
480 BUG_ON(PageAnon(page
));
482 spin_lock(&mapping
->i_mmap_lock
);
483 vma_prio_tree_foreach(vma
, &iter
, &mapping
->i_mmap
, pgoff
, pgoff
) {
484 if (vma
->vm_flags
& VM_SHARED
)
485 ret
+= page_mkclean_one(page
, vma
);
487 spin_unlock(&mapping
->i_mmap_lock
);
491 int page_mkclean(struct page
*page
)
495 BUG_ON(!PageLocked(page
));
497 if (page_mapped(page
)) {
498 struct address_space
*mapping
= page_mapping(page
);
500 ret
= page_mkclean_file(mapping
, page
);
501 if (page_test_dirty(page
)) {
502 page_clear_dirty(page
);
510 EXPORT_SYMBOL_GPL(page_mkclean
);
513 * __page_set_anon_rmap - setup new anonymous rmap
514 * @page: the page to add the mapping to
515 * @vma: the vm area in which the mapping is added
516 * @address: the user virtual address mapped
518 static void __page_set_anon_rmap(struct page
*page
,
519 struct vm_area_struct
*vma
, unsigned long address
)
521 struct anon_vma
*anon_vma
= vma
->anon_vma
;
524 anon_vma
= (void *) anon_vma
+ PAGE_MAPPING_ANON
;
525 page
->mapping
= (struct address_space
*) anon_vma
;
527 page
->index
= linear_page_index(vma
, address
);
530 * nr_mapped state can be updated without turning off
531 * interrupts because it is not modified via interrupt.
533 __inc_zone_page_state(page
, NR_ANON_PAGES
);
537 * __page_check_anon_rmap - sanity check anonymous rmap addition
538 * @page: the page to add the mapping to
539 * @vma: the vm area in which the mapping is added
540 * @address: the user virtual address mapped
542 static void __page_check_anon_rmap(struct page
*page
,
543 struct vm_area_struct
*vma
, unsigned long address
)
545 #ifdef CONFIG_DEBUG_VM
547 * The page's anon-rmap details (mapping and index) are guaranteed to
548 * be set up correctly at this point.
550 * We have exclusion against page_add_anon_rmap because the caller
551 * always holds the page locked, except if called from page_dup_rmap,
552 * in which case the page is already known to be setup.
554 * We have exclusion against page_add_new_anon_rmap because those pages
555 * are initially only visible via the pagetables, and the pte is locked
556 * over the call to page_add_new_anon_rmap.
558 struct anon_vma
*anon_vma
= vma
->anon_vma
;
559 anon_vma
= (void *) anon_vma
+ PAGE_MAPPING_ANON
;
560 BUG_ON(page
->mapping
!= (struct address_space
*)anon_vma
);
561 BUG_ON(page
->index
!= linear_page_index(vma
, address
));
566 * page_add_anon_rmap - add pte mapping to an anonymous page
567 * @page: the page to add the mapping to
568 * @vma: the vm area in which the mapping is added
569 * @address: the user virtual address mapped
571 * The caller needs to hold the pte lock and the page must be locked.
573 void page_add_anon_rmap(struct page
*page
,
574 struct vm_area_struct
*vma
, unsigned long address
)
576 VM_BUG_ON(!PageLocked(page
));
577 VM_BUG_ON(address
< vma
->vm_start
|| address
>= vma
->vm_end
);
578 if (atomic_inc_and_test(&page
->_mapcount
))
579 __page_set_anon_rmap(page
, vma
, address
);
581 __page_check_anon_rmap(page
, vma
, address
);
585 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
586 * @page: the page to add the mapping to
587 * @vma: the vm area in which the mapping is added
588 * @address: the user virtual address mapped
590 * Same as page_add_anon_rmap but must only be called on *new* pages.
591 * This means the inc-and-test can be bypassed.
592 * Page does not have to be locked.
594 void page_add_new_anon_rmap(struct page
*page
,
595 struct vm_area_struct
*vma
, unsigned long address
)
597 BUG_ON(address
< vma
->vm_start
|| address
>= vma
->vm_end
);
598 atomic_set(&page
->_mapcount
, 0); /* elevate count by 1 (starts at -1) */
599 __page_set_anon_rmap(page
, vma
, address
);
603 * page_add_file_rmap - add pte mapping to a file page
604 * @page: the page to add the mapping to
606 * The caller needs to hold the pte lock.
608 void page_add_file_rmap(struct page
*page
)
610 if (atomic_inc_and_test(&page
->_mapcount
))
611 __inc_zone_page_state(page
, NR_FILE_MAPPED
);
614 #ifdef CONFIG_DEBUG_VM
616 * page_dup_rmap - duplicate pte mapping to a page
617 * @page: the page to add the mapping to
618 * @vma: the vm area being duplicated
619 * @address: the user virtual address mapped
621 * For copy_page_range only: minimal extract from page_add_file_rmap /
622 * page_add_anon_rmap, avoiding unnecessary tests (already checked) so it's
625 * The caller needs to hold the pte lock.
627 void page_dup_rmap(struct page
*page
, struct vm_area_struct
*vma
, unsigned long address
)
629 BUG_ON(page_mapcount(page
) == 0);
631 __page_check_anon_rmap(page
, vma
, address
);
632 atomic_inc(&page
->_mapcount
);
637 * page_remove_rmap - take down pte mapping from a page
638 * @page: page to remove mapping from
639 * @vma: the vm area in which the mapping is removed
641 * The caller needs to hold the pte lock.
643 void page_remove_rmap(struct page
*page
, struct vm_area_struct
*vma
)
645 if (atomic_add_negative(-1, &page
->_mapcount
)) {
646 if (unlikely(page_mapcount(page
) < 0)) {
647 printk (KERN_EMERG
"Eeek! page_mapcount(page) went negative! (%d)\n", page_mapcount(page
));
648 printk (KERN_EMERG
" page pfn = %lx\n", page_to_pfn(page
));
649 printk (KERN_EMERG
" page->flags = %lx\n", page
->flags
);
650 printk (KERN_EMERG
" page->count = %x\n", page_count(page
));
651 printk (KERN_EMERG
" page->mapping = %p\n", page
->mapping
);
652 print_symbol (KERN_EMERG
" vma->vm_ops = %s\n", (unsigned long)vma
->vm_ops
);
654 print_symbol (KERN_EMERG
" vma->vm_ops->fault = %s\n", (unsigned long)vma
->vm_ops
->fault
);
656 if (vma
->vm_file
&& vma
->vm_file
->f_op
)
657 print_symbol (KERN_EMERG
" vma->vm_file->f_op->mmap = %s\n", (unsigned long)vma
->vm_file
->f_op
->mmap
);
662 * It would be tidy to reset the PageAnon mapping here,
663 * but that might overwrite a racing page_add_anon_rmap
664 * which increments mapcount after us but sets mapping
665 * before us: so leave the reset to free_hot_cold_page,
666 * and remember that it's only reliable while mapped.
667 * Leaving it set also helps swapoff to reinstate ptes
668 * faster for those pages still in swapcache.
670 if (page_test_dirty(page
)) {
671 page_clear_dirty(page
);
672 set_page_dirty(page
);
674 mem_cgroup_uncharge_page(page
);
676 __dec_zone_page_state(page
,
677 PageAnon(page
) ? NR_ANON_PAGES
: NR_FILE_MAPPED
);
682 * Subfunctions of try_to_unmap: try_to_unmap_one called
683 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
685 static int try_to_unmap_one(struct page
*page
, struct vm_area_struct
*vma
,
688 struct mm_struct
*mm
= vma
->vm_mm
;
689 unsigned long address
;
693 int ret
= SWAP_AGAIN
;
695 address
= vma_address(page
, vma
);
696 if (address
== -EFAULT
)
699 pte
= page_check_address(page
, mm
, address
, &ptl
);
704 * If the page is mlock()d, we cannot swap it out.
705 * If it's recently referenced (perhaps page_referenced
706 * skipped over this mm) then we should reactivate it.
708 if (!migration
&& ((vma
->vm_flags
& VM_LOCKED
) ||
709 (ptep_clear_flush_young_notify(vma
, address
, pte
)))) {
714 /* Nuke the page table entry. */
715 flush_cache_page(vma
, address
, page_to_pfn(page
));
716 pteval
= ptep_clear_flush_notify(vma
, address
, pte
);
718 /* Move the dirty bit to the physical page now the pte is gone. */
719 if (pte_dirty(pteval
))
720 set_page_dirty(page
);
722 /* Update high watermark before we lower rss */
723 update_hiwater_rss(mm
);
725 if (PageAnon(page
)) {
726 swp_entry_t entry
= { .val
= page_private(page
) };
728 if (PageSwapCache(page
)) {
730 * Store the swap location in the pte.
731 * See handle_pte_fault() ...
733 swap_duplicate(entry
);
734 if (list_empty(&mm
->mmlist
)) {
735 spin_lock(&mmlist_lock
);
736 if (list_empty(&mm
->mmlist
))
737 list_add(&mm
->mmlist
, &init_mm
.mmlist
);
738 spin_unlock(&mmlist_lock
);
740 dec_mm_counter(mm
, anon_rss
);
741 #ifdef CONFIG_MIGRATION
744 * Store the pfn of the page in a special migration
745 * pte. do_swap_page() will wait until the migration
746 * pte is removed and then restart fault handling.
749 entry
= make_migration_entry(page
, pte_write(pteval
));
752 set_pte_at(mm
, address
, pte
, swp_entry_to_pte(entry
));
753 BUG_ON(pte_file(*pte
));
755 #ifdef CONFIG_MIGRATION
757 /* Establish migration entry for a file page */
759 entry
= make_migration_entry(page
, pte_write(pteval
));
760 set_pte_at(mm
, address
, pte
, swp_entry_to_pte(entry
));
763 dec_mm_counter(mm
, file_rss
);
766 page_remove_rmap(page
, vma
);
767 page_cache_release(page
);
770 pte_unmap_unlock(pte
, ptl
);
776 * objrmap doesn't work for nonlinear VMAs because the assumption that
777 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
778 * Consequently, given a particular page and its ->index, we cannot locate the
779 * ptes which are mapping that page without an exhaustive linear search.
781 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
782 * maps the file to which the target page belongs. The ->vm_private_data field
783 * holds the current cursor into that scan. Successive searches will circulate
784 * around the vma's virtual address space.
786 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
787 * more scanning pressure is placed against them as well. Eventually pages
788 * will become fully unmapped and are eligible for eviction.
790 * For very sparsely populated VMAs this is a little inefficient - chances are
791 * there there won't be many ptes located within the scan cluster. In this case
792 * maybe we could scan further - to the end of the pte page, perhaps.
794 #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
795 #define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
797 static void try_to_unmap_cluster(unsigned long cursor
,
798 unsigned int *mapcount
, struct vm_area_struct
*vma
)
800 struct mm_struct
*mm
= vma
->vm_mm
;
808 unsigned long address
;
811 address
= (vma
->vm_start
+ cursor
) & CLUSTER_MASK
;
812 end
= address
+ CLUSTER_SIZE
;
813 if (address
< vma
->vm_start
)
814 address
= vma
->vm_start
;
815 if (end
> vma
->vm_end
)
818 pgd
= pgd_offset(mm
, address
);
819 if (!pgd_present(*pgd
))
822 pud
= pud_offset(pgd
, address
);
823 if (!pud_present(*pud
))
826 pmd
= pmd_offset(pud
, address
);
827 if (!pmd_present(*pmd
))
830 pte
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
832 /* Update high watermark before we lower rss */
833 update_hiwater_rss(mm
);
835 for (; address
< end
; pte
++, address
+= PAGE_SIZE
) {
836 if (!pte_present(*pte
))
838 page
= vm_normal_page(vma
, address
, *pte
);
839 BUG_ON(!page
|| PageAnon(page
));
841 if (ptep_clear_flush_young_notify(vma
, address
, pte
))
844 /* Nuke the page table entry. */
845 flush_cache_page(vma
, address
, pte_pfn(*pte
));
846 pteval
= ptep_clear_flush_notify(vma
, address
, pte
);
848 /* If nonlinear, store the file page offset in the pte. */
849 if (page
->index
!= linear_page_index(vma
, address
))
850 set_pte_at(mm
, address
, pte
, pgoff_to_pte(page
->index
));
852 /* Move the dirty bit to the physical page now the pte is gone. */
853 if (pte_dirty(pteval
))
854 set_page_dirty(page
);
856 page_remove_rmap(page
, vma
);
857 page_cache_release(page
);
858 dec_mm_counter(mm
, file_rss
);
861 pte_unmap_unlock(pte
- 1, ptl
);
864 static int try_to_unmap_anon(struct page
*page
, int migration
)
866 struct anon_vma
*anon_vma
;
867 struct vm_area_struct
*vma
;
868 int ret
= SWAP_AGAIN
;
870 anon_vma
= page_lock_anon_vma(page
);
874 list_for_each_entry(vma
, &anon_vma
->head
, anon_vma_node
) {
875 ret
= try_to_unmap_one(page
, vma
, migration
);
876 if (ret
== SWAP_FAIL
|| !page_mapped(page
))
880 page_unlock_anon_vma(anon_vma
);
885 * try_to_unmap_file - unmap file page using the object-based rmap method
886 * @page: the page to unmap
887 * @migration: migration flag
889 * Find all the mappings of a page using the mapping pointer and the vma chains
890 * contained in the address_space struct it points to.
892 * This function is only called from try_to_unmap for object-based pages.
894 static int try_to_unmap_file(struct page
*page
, int migration
)
896 struct address_space
*mapping
= page
->mapping
;
897 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
898 struct vm_area_struct
*vma
;
899 struct prio_tree_iter iter
;
900 int ret
= SWAP_AGAIN
;
901 unsigned long cursor
;
902 unsigned long max_nl_cursor
= 0;
903 unsigned long max_nl_size
= 0;
904 unsigned int mapcount
;
906 spin_lock(&mapping
->i_mmap_lock
);
907 vma_prio_tree_foreach(vma
, &iter
, &mapping
->i_mmap
, pgoff
, pgoff
) {
908 ret
= try_to_unmap_one(page
, vma
, migration
);
909 if (ret
== SWAP_FAIL
|| !page_mapped(page
))
913 if (list_empty(&mapping
->i_mmap_nonlinear
))
916 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
,
917 shared
.vm_set
.list
) {
918 if ((vma
->vm_flags
& VM_LOCKED
) && !migration
)
920 cursor
= (unsigned long) vma
->vm_private_data
;
921 if (cursor
> max_nl_cursor
)
922 max_nl_cursor
= cursor
;
923 cursor
= vma
->vm_end
- vma
->vm_start
;
924 if (cursor
> max_nl_size
)
925 max_nl_size
= cursor
;
928 if (max_nl_size
== 0) { /* any nonlinears locked or reserved */
934 * We don't try to search for this page in the nonlinear vmas,
935 * and page_referenced wouldn't have found it anyway. Instead
936 * just walk the nonlinear vmas trying to age and unmap some.
937 * The mapcount of the page we came in with is irrelevant,
938 * but even so use it as a guide to how hard we should try?
940 mapcount
= page_mapcount(page
);
943 cond_resched_lock(&mapping
->i_mmap_lock
);
945 max_nl_size
= (max_nl_size
+ CLUSTER_SIZE
- 1) & CLUSTER_MASK
;
946 if (max_nl_cursor
== 0)
947 max_nl_cursor
= CLUSTER_SIZE
;
950 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
,
951 shared
.vm_set
.list
) {
952 if ((vma
->vm_flags
& VM_LOCKED
) && !migration
)
954 cursor
= (unsigned long) vma
->vm_private_data
;
955 while ( cursor
< max_nl_cursor
&&
956 cursor
< vma
->vm_end
- vma
->vm_start
) {
957 try_to_unmap_cluster(cursor
, &mapcount
, vma
);
958 cursor
+= CLUSTER_SIZE
;
959 vma
->vm_private_data
= (void *) cursor
;
960 if ((int)mapcount
<= 0)
963 vma
->vm_private_data
= (void *) max_nl_cursor
;
965 cond_resched_lock(&mapping
->i_mmap_lock
);
966 max_nl_cursor
+= CLUSTER_SIZE
;
967 } while (max_nl_cursor
<= max_nl_size
);
970 * Don't loop forever (perhaps all the remaining pages are
971 * in locked vmas). Reset cursor on all unreserved nonlinear
972 * vmas, now forgetting on which ones it had fallen behind.
974 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
, shared
.vm_set
.list
)
975 vma
->vm_private_data
= NULL
;
977 spin_unlock(&mapping
->i_mmap_lock
);
982 * try_to_unmap - try to remove all page table mappings to a page
983 * @page: the page to get unmapped
984 * @migration: migration flag
986 * Tries to remove all the page table entries which are mapping this
987 * page, used in the pageout path. Caller must hold the page lock.
990 * SWAP_SUCCESS - we succeeded in removing all mappings
991 * SWAP_AGAIN - we missed a mapping, try again later
992 * SWAP_FAIL - the page is unswappable
994 int try_to_unmap(struct page
*page
, int migration
)
998 BUG_ON(!PageLocked(page
));
1001 ret
= try_to_unmap_anon(page
, migration
);
1003 ret
= try_to_unmap_file(page
, migration
);
1005 if (!page_mapped(page
))