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_sem (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_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.
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)
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>
56 #include <asm/tlbflush.h>
58 //#define RMAP_DEBUG /* can be enabled only for debugging */
60 kmem_cache_t
*anon_vma_cachep
;
62 static inline void validate_anon_vma(struct vm_area_struct
*find_vma
)
65 struct anon_vma
*anon_vma
= find_vma
->anon_vma
;
66 struct vm_area_struct
*vma
;
67 unsigned int mapcount
= 0;
70 list_for_each_entry(vma
, &anon_vma
->head
, anon_vma_node
) {
72 BUG_ON(mapcount
> 100000);
80 /* This must be called under the mmap_sem. */
81 int anon_vma_prepare(struct vm_area_struct
*vma
)
83 struct anon_vma
*anon_vma
= vma
->anon_vma
;
86 if (unlikely(!anon_vma
)) {
87 struct mm_struct
*mm
= vma
->vm_mm
;
88 struct anon_vma
*allocated
, *locked
;
90 anon_vma
= find_mergeable_anon_vma(vma
);
94 spin_lock(&locked
->lock
);
96 anon_vma
= anon_vma_alloc();
97 if (unlikely(!anon_vma
))
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
);
110 spin_unlock(&mm
->page_table_lock
);
113 spin_unlock(&locked
->lock
);
114 if (unlikely(allocated
))
115 anon_vma_free(allocated
);
120 void __anon_vma_merge(struct vm_area_struct
*vma
, struct vm_area_struct
*next
)
122 BUG_ON(vma
->anon_vma
!= next
->anon_vma
);
123 list_del(&next
->anon_vma_node
);
126 void __anon_vma_link(struct vm_area_struct
*vma
)
128 struct anon_vma
*anon_vma
= vma
->anon_vma
;
131 list_add(&vma
->anon_vma_node
, &anon_vma
->head
);
132 validate_anon_vma(vma
);
136 void anon_vma_link(struct vm_area_struct
*vma
)
138 struct anon_vma
*anon_vma
= vma
->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
);
148 void anon_vma_unlink(struct vm_area_struct
*vma
)
150 struct anon_vma
*anon_vma
= vma
->anon_vma
;
156 spin_lock(&anon_vma
->lock
);
157 validate_anon_vma(vma
);
158 list_del(&vma
->anon_vma_node
);
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
);
165 anon_vma_free(anon_vma
);
168 static void anon_vma_ctor(void *data
, kmem_cache_t
*cachep
, unsigned long flags
)
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
);
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
443 * The caller needs to hold the pte lock.
445 void page_add_anon_rmap(struct page
*page
,
446 struct vm_area_struct
*vma
, unsigned long address
)
448 if (atomic_inc_and_test(&page
->_mapcount
)) {
449 struct anon_vma
*anon_vma
= vma
->anon_vma
;
452 anon_vma
= (void *) anon_vma
+ PAGE_MAPPING_ANON
;
453 page
->mapping
= (struct address_space
*) anon_vma
;
455 page
->index
= linear_page_index(vma
, address
);
457 inc_page_state(nr_mapped
);
459 /* else checking page index and mapping is racy */
463 * page_add_file_rmap - add pte mapping to a file page
464 * @page: the page to add the mapping to
466 * The caller needs to hold the pte lock.
468 void page_add_file_rmap(struct page
*page
)
470 BUG_ON(PageAnon(page
));
471 BUG_ON(!pfn_valid(page_to_pfn(page
)));
473 if (atomic_inc_and_test(&page
->_mapcount
))
474 inc_page_state(nr_mapped
);
478 * page_remove_rmap - take down pte mapping from a page
479 * @page: page to remove mapping from
481 * The caller needs to hold the pte lock.
483 void page_remove_rmap(struct page
*page
)
485 if (atomic_add_negative(-1, &page
->_mapcount
)) {
486 BUG_ON(page_mapcount(page
) < 0);
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.
496 if (page_test_and_clear_dirty(page
))
497 set_page_dirty(page
);
498 dec_page_state(nr_mapped
);
503 * Subfunctions of try_to_unmap: try_to_unmap_one called
504 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
506 static int try_to_unmap_one(struct page
*page
, struct vm_area_struct
*vma
)
508 struct mm_struct
*mm
= vma
->vm_mm
;
509 unsigned long address
;
513 int ret
= SWAP_AGAIN
;
515 address
= vma_address(page
, vma
);
516 if (address
== -EFAULT
)
519 pte
= page_check_address(page
, mm
, address
, &ptl
);
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.
528 if ((vma
->vm_flags
& VM_LOCKED
) ||
529 ptep_clear_flush_young(vma
, address
, pte
)) {
534 /* Nuke the page table entry. */
535 flush_cache_page(vma
, address
, page_to_pfn(page
));
536 pteval
= ptep_clear_flush(vma
, address
, pte
);
538 /* Move the dirty bit to the physical page now the pte is gone. */
539 if (pte_dirty(pteval
))
540 set_page_dirty(page
);
542 /* Update high watermark before we lower rss */
543 update_hiwater_rss(mm
);
545 if (PageAnon(page
)) {
546 swp_entry_t entry
= { .val
= page_private(page
) };
548 * Store the swap location in the pte.
549 * See handle_pte_fault() ...
551 BUG_ON(!PageSwapCache(page
));
552 swap_duplicate(entry
);
553 if (list_empty(&mm
->mmlist
)) {
554 spin_lock(&mmlist_lock
);
555 if (list_empty(&mm
->mmlist
))
556 list_add(&mm
->mmlist
, &init_mm
.mmlist
);
557 spin_unlock(&mmlist_lock
);
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
);
563 dec_mm_counter(mm
, file_rss
);
565 page_remove_rmap(page
);
566 page_cache_release(page
);
569 pte_unmap_unlock(pte
, ptl
);
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.
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.
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.
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.
593 #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
594 #define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
596 static void try_to_unmap_cluster(unsigned long cursor
,
597 unsigned int *mapcount
, struct vm_area_struct
*vma
)
599 struct mm_struct
*mm
= vma
->vm_mm
;
607 unsigned long address
;
610 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
)
617 pgd
= pgd_offset(mm
, address
);
618 if (!pgd_present(*pgd
))
621 pud
= pud_offset(pgd
, address
);
622 if (!pud_present(*pud
))
625 pmd
= pmd_offset(pud
, address
);
626 if (!pmd_present(*pmd
))
629 pte
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
631 /* Update high watermark before we lower rss */
632 update_hiwater_rss(mm
);
634 for (; address
< end
; pte
++, address
+= PAGE_SIZE
) {
635 if (!pte_present(*pte
))
637 page
= vm_normal_page(vma
, address
, *pte
);
638 BUG_ON(!page
|| PageAnon(page
));
640 if (ptep_clear_flush_young(vma
, address
, pte
))
643 /* Nuke the page table entry. */
644 flush_cache_page(vma
, address
, pte_pfn(*pte
));
645 pteval
= ptep_clear_flush(vma
, address
, pte
);
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
));
651 /* Move the dirty bit to the physical page now the pte is gone. */
652 if (pte_dirty(pteval
))
653 set_page_dirty(page
);
655 page_remove_rmap(page
);
656 page_cache_release(page
);
657 dec_mm_counter(mm
, file_rss
);
660 pte_unmap_unlock(pte
- 1, ptl
);
663 static int try_to_unmap_anon(struct page
*page
)
665 struct anon_vma
*anon_vma
;
666 struct vm_area_struct
*vma
;
667 int ret
= SWAP_AGAIN
;
669 anon_vma
= page_lock_anon_vma(page
);
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
))
678 spin_unlock(&anon_vma
->lock
);
683 * try_to_unmap_file - unmap file page using the object-based rmap method
684 * @page: the page to unmap
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.
689 * This function is only called from try_to_unmap for object-based pages.
691 static int try_to_unmap_file(struct page
*page
)
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
;
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
))
710 if (list_empty(&mapping
->i_mmap_nonlinear
))
713 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
,
714 shared
.vm_set
.list
) {
715 if (vma
->vm_flags
& VM_LOCKED
)
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
;
725 if (max_nl_size
== 0) { /* any nonlinears locked or reserved */
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?
737 mapcount
= page_mapcount(page
);
740 cond_resched_lock(&mapping
->i_mmap_lock
);
742 max_nl_size
= (max_nl_size
+ CLUSTER_SIZE
- 1) & CLUSTER_MASK
;
743 if (max_nl_cursor
== 0)
744 max_nl_cursor
= CLUSTER_SIZE
;
747 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
,
748 shared
.vm_set
.list
) {
749 if (vma
->vm_flags
& VM_LOCKED
)
751 cursor
= (unsigned long) vma
->vm_private_data
;
752 while ( cursor
< max_nl_cursor
&&
753 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)
760 vma
->vm_private_data
= (void *) max_nl_cursor
;
762 cond_resched_lock(&mapping
->i_mmap_lock
);
763 max_nl_cursor
+= CLUSTER_SIZE
;
764 } while (max_nl_cursor
<= max_nl_size
);
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.
771 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
, shared
.vm_set
.list
)
772 vma
->vm_private_data
= NULL
;
774 spin_unlock(&mapping
->i_mmap_lock
);
779 * try_to_unmap - try to remove all page table mappings to a page
780 * @page: the page to get unmapped
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.
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
790 int try_to_unmap(struct page
*page
)
794 BUG_ON(!PageLocked(page
));
797 ret
= try_to_unmap_anon(page
);
799 ret
= try_to_unmap_file(page
);
801 if (!page_mapped(page
))