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 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/memcontrol.h>
51 #include <linux/mmu_notifier.h>
52 #include <linux/migrate.h>
54 #include <asm/tlbflush.h>
58 static struct kmem_cache
*anon_vma_cachep
;
60 static inline struct anon_vma
*anon_vma_alloc(void)
62 return kmem_cache_alloc(anon_vma_cachep
, GFP_KERNEL
);
65 static inline void anon_vma_free(struct anon_vma
*anon_vma
)
67 kmem_cache_free(anon_vma_cachep
, anon_vma
);
71 * anon_vma_prepare - attach an anon_vma to a memory region
72 * @vma: the memory region in question
74 * This makes sure the memory mapping described by 'vma' has
75 * an 'anon_vma' attached to it, so that we can associate the
76 * anonymous pages mapped into it with that anon_vma.
78 * The common case will be that we already have one, but if
79 * if not we either need to find an adjacent mapping that we
80 * can re-use the anon_vma from (very common when the only
81 * reason for splitting a vma has been mprotect()), or we
84 * Anon-vma allocations are very subtle, because we may have
85 * optimistically looked up an anon_vma in page_lock_anon_vma()
86 * and that may actually touch the spinlock even in the newly
87 * allocated vma (it depends on RCU to make sure that the
88 * anon_vma isn't actually destroyed).
90 * As a result, we need to do proper anon_vma locking even
91 * for the new allocation. At the same time, we do not want
92 * to do any locking for the common case of already having
95 * This must be called with the mmap_sem held for reading.
97 int anon_vma_prepare(struct vm_area_struct
*vma
)
99 struct anon_vma
*anon_vma
= vma
->anon_vma
;
102 if (unlikely(!anon_vma
)) {
103 struct mm_struct
*mm
= vma
->vm_mm
;
104 struct anon_vma
*allocated
;
106 anon_vma
= find_mergeable_anon_vma(vma
);
109 anon_vma
= anon_vma_alloc();
110 if (unlikely(!anon_vma
))
112 allocated
= anon_vma
;
114 spin_lock(&anon_vma
->lock
);
116 /* page_table_lock to protect against threads */
117 spin_lock(&mm
->page_table_lock
);
118 if (likely(!vma
->anon_vma
)) {
119 vma
->anon_vma
= anon_vma
;
120 list_add_tail(&vma
->anon_vma_node
, &anon_vma
->head
);
123 spin_unlock(&mm
->page_table_lock
);
125 spin_unlock(&anon_vma
->lock
);
126 if (unlikely(allocated
))
127 anon_vma_free(allocated
);
132 void __anon_vma_merge(struct vm_area_struct
*vma
, struct vm_area_struct
*next
)
134 BUG_ON(vma
->anon_vma
!= next
->anon_vma
);
135 list_del(&next
->anon_vma_node
);
138 void __anon_vma_link(struct vm_area_struct
*vma
)
140 struct anon_vma
*anon_vma
= vma
->anon_vma
;
143 list_add_tail(&vma
->anon_vma_node
, &anon_vma
->head
);
146 void anon_vma_link(struct vm_area_struct
*vma
)
148 struct anon_vma
*anon_vma
= vma
->anon_vma
;
151 spin_lock(&anon_vma
->lock
);
152 list_add_tail(&vma
->anon_vma_node
, &anon_vma
->head
);
153 spin_unlock(&anon_vma
->lock
);
157 void anon_vma_unlink(struct vm_area_struct
*vma
)
159 struct anon_vma
*anon_vma
= vma
->anon_vma
;
165 spin_lock(&anon_vma
->lock
);
166 list_del(&vma
->anon_vma_node
);
168 /* We must garbage collect the anon_vma if it's empty */
169 empty
= list_empty(&anon_vma
->head
);
170 spin_unlock(&anon_vma
->lock
);
173 anon_vma_free(anon_vma
);
176 static void anon_vma_ctor(void *data
)
178 struct anon_vma
*anon_vma
= data
;
180 spin_lock_init(&anon_vma
->lock
);
181 INIT_LIST_HEAD(&anon_vma
->head
);
184 void __init
anon_vma_init(void)
186 anon_vma_cachep
= kmem_cache_create("anon_vma", sizeof(struct anon_vma
),
187 0, SLAB_DESTROY_BY_RCU
|SLAB_PANIC
, anon_vma_ctor
);
191 * Getting a lock on a stable anon_vma from a page off the LRU is
192 * tricky: page_lock_anon_vma rely on RCU to guard against the races.
194 static struct anon_vma
*page_lock_anon_vma(struct page
*page
)
196 struct anon_vma
*anon_vma
;
197 unsigned long anon_mapping
;
200 anon_mapping
= (unsigned long) page
->mapping
;
201 if (!(anon_mapping
& PAGE_MAPPING_ANON
))
203 if (!page_mapped(page
))
206 anon_vma
= (struct anon_vma
*) (anon_mapping
- PAGE_MAPPING_ANON
);
207 spin_lock(&anon_vma
->lock
);
214 static void page_unlock_anon_vma(struct anon_vma
*anon_vma
)
216 spin_unlock(&anon_vma
->lock
);
221 * At what user virtual address is page expected in @vma?
222 * Returns virtual address or -EFAULT if page's index/offset is not
223 * within the range mapped the @vma.
225 static inline unsigned long
226 vma_address(struct page
*page
, struct vm_area_struct
*vma
)
228 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
229 unsigned long address
;
231 address
= vma
->vm_start
+ ((pgoff
- vma
->vm_pgoff
) << PAGE_SHIFT
);
232 if (unlikely(address
< vma
->vm_start
|| address
>= vma
->vm_end
)) {
233 /* page should be within @vma mapping range */
240 * At what user virtual address is page expected in vma? checking that the
241 * page matches the vma: currently only used on anon pages, by unuse_vma;
243 unsigned long page_address_in_vma(struct page
*page
, struct vm_area_struct
*vma
)
245 if (PageAnon(page
)) {
246 if ((void *)vma
->anon_vma
!=
247 (void *)page
->mapping
- PAGE_MAPPING_ANON
)
249 } else if (page
->mapping
&& !(vma
->vm_flags
& VM_NONLINEAR
)) {
251 vma
->vm_file
->f_mapping
!= page
->mapping
)
255 return vma_address(page
, vma
);
259 * Check that @page is mapped at @address into @mm.
261 * If @sync is false, page_check_address may perform a racy check to avoid
262 * the page table lock when the pte is not present (helpful when reclaiming
263 * highly shared pages).
265 * On success returns with pte mapped and locked.
267 pte_t
*page_check_address(struct page
*page
, struct mm_struct
*mm
,
268 unsigned long address
, spinlock_t
**ptlp
, int sync
)
276 pgd
= pgd_offset(mm
, address
);
277 if (!pgd_present(*pgd
))
280 pud
= pud_offset(pgd
, address
);
281 if (!pud_present(*pud
))
284 pmd
= pmd_offset(pud
, address
);
285 if (!pmd_present(*pmd
))
288 pte
= pte_offset_map(pmd
, address
);
289 /* Make a quick check before getting the lock */
290 if (!sync
&& !pte_present(*pte
)) {
295 ptl
= pte_lockptr(mm
, pmd
);
297 if (pte_present(*pte
) && page_to_pfn(page
) == pte_pfn(*pte
)) {
301 pte_unmap_unlock(pte
, ptl
);
306 * page_mapped_in_vma - check whether a page is really mapped in a VMA
307 * @page: the page to test
308 * @vma: the VMA to test
310 * Returns 1 if the page is mapped into the page tables of the VMA, 0
311 * if the page is not mapped into the page tables of this VMA. Only
312 * valid for normal file or anonymous VMAs.
314 static int page_mapped_in_vma(struct page
*page
, struct vm_area_struct
*vma
)
316 unsigned long address
;
320 address
= vma_address(page
, vma
);
321 if (address
== -EFAULT
) /* out of vma range */
323 pte
= page_check_address(page
, vma
->vm_mm
, address
, &ptl
, 1);
324 if (!pte
) /* the page is not in this mm */
326 pte_unmap_unlock(pte
, ptl
);
332 * Subfunctions of page_referenced: page_referenced_one called
333 * repeatedly from either page_referenced_anon or page_referenced_file.
335 static int page_referenced_one(struct page
*page
,
336 struct vm_area_struct
*vma
, unsigned int *mapcount
)
338 struct mm_struct
*mm
= vma
->vm_mm
;
339 unsigned long address
;
344 address
= vma_address(page
, vma
);
345 if (address
== -EFAULT
)
348 pte
= page_check_address(page
, mm
, address
, &ptl
, 0);
353 * Don't want to elevate referenced for mlocked page that gets this far,
354 * in order that it progresses to try_to_unmap and is moved to the
357 if (vma
->vm_flags
& VM_LOCKED
) {
358 *mapcount
= 1; /* break early from loop */
362 if (ptep_clear_flush_young_notify(vma
, address
, pte
)) {
364 * Don't treat a reference through a sequentially read
365 * mapping as such. If the page has been used in
366 * another mapping, we will catch it; if this other
367 * mapping is already gone, the unmap path will have
368 * set PG_referenced or activated the page.
370 if (likely(!VM_SequentialReadHint(vma
)))
374 /* Pretend the page is referenced if the task has the
375 swap token and is in the middle of a page fault. */
376 if (mm
!= current
->mm
&& has_swap_token(mm
) &&
377 rwsem_is_locked(&mm
->mmap_sem
))
382 pte_unmap_unlock(pte
, ptl
);
387 static int page_referenced_anon(struct page
*page
,
388 struct mem_cgroup
*mem_cont
)
390 unsigned int mapcount
;
391 struct anon_vma
*anon_vma
;
392 struct vm_area_struct
*vma
;
395 anon_vma
= page_lock_anon_vma(page
);
399 mapcount
= page_mapcount(page
);
400 list_for_each_entry(vma
, &anon_vma
->head
, anon_vma_node
) {
402 * If we are reclaiming on behalf of a cgroup, skip
403 * counting on behalf of references from different
406 if (mem_cont
&& !mm_match_cgroup(vma
->vm_mm
, mem_cont
))
408 referenced
+= page_referenced_one(page
, vma
, &mapcount
);
413 page_unlock_anon_vma(anon_vma
);
418 * page_referenced_file - referenced check for object-based rmap
419 * @page: the page we're checking references on.
420 * @mem_cont: target memory controller
422 * For an object-based mapped page, find all the places it is mapped and
423 * check/clear the referenced flag. This is done by following the page->mapping
424 * pointer, then walking the chain of vmas it holds. It returns the number
425 * of references it found.
427 * This function is only called from page_referenced for object-based pages.
429 static int page_referenced_file(struct page
*page
,
430 struct mem_cgroup
*mem_cont
)
432 unsigned int mapcount
;
433 struct address_space
*mapping
= page
->mapping
;
434 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
435 struct vm_area_struct
*vma
;
436 struct prio_tree_iter iter
;
440 * The caller's checks on page->mapping and !PageAnon have made
441 * sure that this is a file page: the check for page->mapping
442 * excludes the case just before it gets set on an anon page.
444 BUG_ON(PageAnon(page
));
447 * The page lock not only makes sure that page->mapping cannot
448 * suddenly be NULLified by truncation, it makes sure that the
449 * structure at mapping cannot be freed and reused yet,
450 * so we can safely take mapping->i_mmap_lock.
452 BUG_ON(!PageLocked(page
));
454 spin_lock(&mapping
->i_mmap_lock
);
457 * i_mmap_lock does not stabilize mapcount at all, but mapcount
458 * is more likely to be accurate if we note it after spinning.
460 mapcount
= page_mapcount(page
);
462 vma_prio_tree_foreach(vma
, &iter
, &mapping
->i_mmap
, pgoff
, pgoff
) {
464 * If we are reclaiming on behalf of a cgroup, skip
465 * counting on behalf of references from different
468 if (mem_cont
&& !mm_match_cgroup(vma
->vm_mm
, mem_cont
))
470 referenced
+= page_referenced_one(page
, vma
, &mapcount
);
475 spin_unlock(&mapping
->i_mmap_lock
);
480 * page_referenced - test if the page was referenced
481 * @page: the page to test
482 * @is_locked: caller holds lock on the page
483 * @mem_cont: target memory controller
485 * Quick test_and_clear_referenced for all mappings to a page,
486 * returns the number of ptes which referenced the page.
488 int page_referenced(struct page
*page
, int is_locked
,
489 struct mem_cgroup
*mem_cont
)
493 if (TestClearPageReferenced(page
))
496 if (page_mapped(page
) && page
->mapping
) {
498 referenced
+= page_referenced_anon(page
, mem_cont
);
500 referenced
+= page_referenced_file(page
, mem_cont
);
501 else if (!trylock_page(page
))
506 page_referenced_file(page
, mem_cont
);
511 if (page_test_and_clear_young(page
))
517 static int page_mkclean_one(struct page
*page
, struct vm_area_struct
*vma
)
519 struct mm_struct
*mm
= vma
->vm_mm
;
520 unsigned long address
;
525 address
= vma_address(page
, vma
);
526 if (address
== -EFAULT
)
529 pte
= page_check_address(page
, mm
, address
, &ptl
, 1);
533 if (pte_dirty(*pte
) || pte_write(*pte
)) {
536 flush_cache_page(vma
, address
, pte_pfn(*pte
));
537 entry
= ptep_clear_flush_notify(vma
, address
, pte
);
538 entry
= pte_wrprotect(entry
);
539 entry
= pte_mkclean(entry
);
540 set_pte_at(mm
, address
, pte
, entry
);
544 pte_unmap_unlock(pte
, ptl
);
549 static int page_mkclean_file(struct address_space
*mapping
, struct page
*page
)
551 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
552 struct vm_area_struct
*vma
;
553 struct prio_tree_iter iter
;
556 BUG_ON(PageAnon(page
));
558 spin_lock(&mapping
->i_mmap_lock
);
559 vma_prio_tree_foreach(vma
, &iter
, &mapping
->i_mmap
, pgoff
, pgoff
) {
560 if (vma
->vm_flags
& VM_SHARED
)
561 ret
+= page_mkclean_one(page
, vma
);
563 spin_unlock(&mapping
->i_mmap_lock
);
567 int page_mkclean(struct page
*page
)
571 BUG_ON(!PageLocked(page
));
573 if (page_mapped(page
)) {
574 struct address_space
*mapping
= page_mapping(page
);
576 ret
= page_mkclean_file(mapping
, page
);
577 if (page_test_dirty(page
)) {
578 page_clear_dirty(page
);
586 EXPORT_SYMBOL_GPL(page_mkclean
);
589 * __page_set_anon_rmap - setup new anonymous rmap
590 * @page: the page to add the mapping to
591 * @vma: the vm area in which the mapping is added
592 * @address: the user virtual address mapped
594 static void __page_set_anon_rmap(struct page
*page
,
595 struct vm_area_struct
*vma
, unsigned long address
)
597 struct anon_vma
*anon_vma
= vma
->anon_vma
;
600 anon_vma
= (void *) anon_vma
+ PAGE_MAPPING_ANON
;
601 page
->mapping
= (struct address_space
*) anon_vma
;
603 page
->index
= linear_page_index(vma
, address
);
606 * nr_mapped state can be updated without turning off
607 * interrupts because it is not modified via interrupt.
609 __inc_zone_page_state(page
, NR_ANON_PAGES
);
613 * __page_check_anon_rmap - sanity check anonymous rmap addition
614 * @page: the page to add the mapping to
615 * @vma: the vm area in which the mapping is added
616 * @address: the user virtual address mapped
618 static void __page_check_anon_rmap(struct page
*page
,
619 struct vm_area_struct
*vma
, unsigned long address
)
621 #ifdef CONFIG_DEBUG_VM
623 * The page's anon-rmap details (mapping and index) are guaranteed to
624 * be set up correctly at this point.
626 * We have exclusion against page_add_anon_rmap because the caller
627 * always holds the page locked, except if called from page_dup_rmap,
628 * in which case the page is already known to be setup.
630 * We have exclusion against page_add_new_anon_rmap because those pages
631 * are initially only visible via the pagetables, and the pte is locked
632 * over the call to page_add_new_anon_rmap.
634 struct anon_vma
*anon_vma
= vma
->anon_vma
;
635 anon_vma
= (void *) anon_vma
+ PAGE_MAPPING_ANON
;
636 BUG_ON(page
->mapping
!= (struct address_space
*)anon_vma
);
637 BUG_ON(page
->index
!= linear_page_index(vma
, address
));
642 * page_add_anon_rmap - add pte mapping to an anonymous page
643 * @page: the page to add the mapping to
644 * @vma: the vm area in which the mapping is added
645 * @address: the user virtual address mapped
647 * The caller needs to hold the pte lock and the page must be locked.
649 void page_add_anon_rmap(struct page
*page
,
650 struct vm_area_struct
*vma
, unsigned long address
)
652 VM_BUG_ON(!PageLocked(page
));
653 VM_BUG_ON(address
< vma
->vm_start
|| address
>= vma
->vm_end
);
654 if (atomic_inc_and_test(&page
->_mapcount
))
655 __page_set_anon_rmap(page
, vma
, address
);
657 __page_check_anon_rmap(page
, vma
, address
);
661 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
662 * @page: the page to add the mapping to
663 * @vma: the vm area in which the mapping is added
664 * @address: the user virtual address mapped
666 * Same as page_add_anon_rmap but must only be called on *new* pages.
667 * This means the inc-and-test can be bypassed.
668 * Page does not have to be locked.
670 void page_add_new_anon_rmap(struct page
*page
,
671 struct vm_area_struct
*vma
, unsigned long address
)
673 VM_BUG_ON(address
< vma
->vm_start
|| address
>= vma
->vm_end
);
674 SetPageSwapBacked(page
);
675 atomic_set(&page
->_mapcount
, 0); /* increment count (starts at -1) */
676 __page_set_anon_rmap(page
, vma
, address
);
677 if (page_evictable(page
, vma
))
678 lru_cache_add_lru(page
, LRU_ACTIVE_ANON
);
680 add_page_to_unevictable_list(page
);
684 * page_add_file_rmap - add pte mapping to a file page
685 * @page: the page to add the mapping to
687 * The caller needs to hold the pte lock.
689 void page_add_file_rmap(struct page
*page
)
691 if (atomic_inc_and_test(&page
->_mapcount
))
692 __inc_zone_page_state(page
, NR_FILE_MAPPED
);
695 #ifdef CONFIG_DEBUG_VM
697 * page_dup_rmap - duplicate pte mapping to a page
698 * @page: the page to add the mapping to
699 * @vma: the vm area being duplicated
700 * @address: the user virtual address mapped
702 * For copy_page_range only: minimal extract from page_add_file_rmap /
703 * page_add_anon_rmap, avoiding unnecessary tests (already checked) so it's
706 * The caller needs to hold the pte lock.
708 void page_dup_rmap(struct page
*page
, struct vm_area_struct
*vma
, unsigned long address
)
711 __page_check_anon_rmap(page
, vma
, address
);
712 atomic_inc(&page
->_mapcount
);
717 * page_remove_rmap - take down pte mapping from a page
718 * @page: page to remove mapping from
720 * The caller needs to hold the pte lock.
722 void page_remove_rmap(struct page
*page
)
724 if (atomic_add_negative(-1, &page
->_mapcount
)) {
726 * Now that the last pte has gone, s390 must transfer dirty
727 * flag from storage key to struct page. We can usually skip
728 * this if the page is anon, so about to be freed; but perhaps
729 * not if it's in swapcache - there might be another pte slot
730 * containing the swap entry, but page not yet written to swap.
732 if ((!PageAnon(page
) || PageSwapCache(page
)) &&
733 page_test_dirty(page
)) {
734 page_clear_dirty(page
);
735 set_page_dirty(page
);
738 mem_cgroup_uncharge_page(page
);
739 __dec_zone_page_state(page
,
740 PageAnon(page
) ? NR_ANON_PAGES
: NR_FILE_MAPPED
);
742 * It would be tidy to reset the PageAnon mapping here,
743 * but that might overwrite a racing page_add_anon_rmap
744 * which increments mapcount after us but sets mapping
745 * before us: so leave the reset to free_hot_cold_page,
746 * and remember that it's only reliable while mapped.
747 * Leaving it set also helps swapoff to reinstate ptes
748 * faster for those pages still in swapcache.
754 * Subfunctions of try_to_unmap: try_to_unmap_one called
755 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
757 static int try_to_unmap_one(struct page
*page
, struct vm_area_struct
*vma
,
760 struct mm_struct
*mm
= vma
->vm_mm
;
761 unsigned long address
;
765 int ret
= SWAP_AGAIN
;
767 address
= vma_address(page
, vma
);
768 if (address
== -EFAULT
)
771 pte
= page_check_address(page
, mm
, address
, &ptl
, 0);
776 * If the page is mlock()d, we cannot swap it out.
777 * If it's recently referenced (perhaps page_referenced
778 * skipped over this mm) then we should reactivate it.
781 if (vma
->vm_flags
& VM_LOCKED
) {
785 if (ptep_clear_flush_young_notify(vma
, address
, pte
)) {
791 /* Nuke the page table entry. */
792 flush_cache_page(vma
, address
, page_to_pfn(page
));
793 pteval
= ptep_clear_flush_notify(vma
, address
, pte
);
795 /* Move the dirty bit to the physical page now the pte is gone. */
796 if (pte_dirty(pteval
))
797 set_page_dirty(page
);
799 /* Update high watermark before we lower rss */
800 update_hiwater_rss(mm
);
802 if (PageAnon(page
)) {
803 swp_entry_t entry
= { .val
= page_private(page
) };
805 if (PageSwapCache(page
)) {
807 * Store the swap location in the pte.
808 * See handle_pte_fault() ...
810 swap_duplicate(entry
);
811 if (list_empty(&mm
->mmlist
)) {
812 spin_lock(&mmlist_lock
);
813 if (list_empty(&mm
->mmlist
))
814 list_add(&mm
->mmlist
, &init_mm
.mmlist
);
815 spin_unlock(&mmlist_lock
);
817 dec_mm_counter(mm
, anon_rss
);
818 } else if (PAGE_MIGRATION
) {
820 * Store the pfn of the page in a special migration
821 * pte. do_swap_page() will wait until the migration
822 * pte is removed and then restart fault handling.
825 entry
= make_migration_entry(page
, pte_write(pteval
));
827 set_pte_at(mm
, address
, pte
, swp_entry_to_pte(entry
));
828 BUG_ON(pte_file(*pte
));
829 } else if (PAGE_MIGRATION
&& migration
) {
830 /* Establish migration entry for a file page */
832 entry
= make_migration_entry(page
, pte_write(pteval
));
833 set_pte_at(mm
, address
, pte
, swp_entry_to_pte(entry
));
835 dec_mm_counter(mm
, file_rss
);
838 page_remove_rmap(page
);
839 page_cache_release(page
);
842 pte_unmap_unlock(pte
, ptl
);
848 * objrmap doesn't work for nonlinear VMAs because the assumption that
849 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
850 * Consequently, given a particular page and its ->index, we cannot locate the
851 * ptes which are mapping that page without an exhaustive linear search.
853 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
854 * maps the file to which the target page belongs. The ->vm_private_data field
855 * holds the current cursor into that scan. Successive searches will circulate
856 * around the vma's virtual address space.
858 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
859 * more scanning pressure is placed against them as well. Eventually pages
860 * will become fully unmapped and are eligible for eviction.
862 * For very sparsely populated VMAs this is a little inefficient - chances are
863 * there there won't be many ptes located within the scan cluster. In this case
864 * maybe we could scan further - to the end of the pte page, perhaps.
866 * Mlocked pages: check VM_LOCKED under mmap_sem held for read, if we can
867 * acquire it without blocking. If vma locked, mlock the pages in the cluster,
868 * rather than unmapping them. If we encounter the "check_page" that vmscan is
869 * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN.
871 #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
872 #define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
874 static int try_to_unmap_cluster(unsigned long cursor
, unsigned int *mapcount
,
875 struct vm_area_struct
*vma
, struct page
*check_page
)
877 struct mm_struct
*mm
= vma
->vm_mm
;
885 unsigned long address
;
887 int ret
= SWAP_AGAIN
;
890 address
= (vma
->vm_start
+ cursor
) & CLUSTER_MASK
;
891 end
= address
+ CLUSTER_SIZE
;
892 if (address
< vma
->vm_start
)
893 address
= vma
->vm_start
;
894 if (end
> vma
->vm_end
)
897 pgd
= pgd_offset(mm
, address
);
898 if (!pgd_present(*pgd
))
901 pud
= pud_offset(pgd
, address
);
902 if (!pud_present(*pud
))
905 pmd
= pmd_offset(pud
, address
);
906 if (!pmd_present(*pmd
))
910 * MLOCK_PAGES => feature is configured.
911 * if we can acquire the mmap_sem for read, and vma is VM_LOCKED,
912 * keep the sem while scanning the cluster for mlocking pages.
914 if (MLOCK_PAGES
&& down_read_trylock(&vma
->vm_mm
->mmap_sem
)) {
915 locked_vma
= (vma
->vm_flags
& VM_LOCKED
);
917 up_read(&vma
->vm_mm
->mmap_sem
); /* don't need it */
920 pte
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
922 /* Update high watermark before we lower rss */
923 update_hiwater_rss(mm
);
925 for (; address
< end
; pte
++, address
+= PAGE_SIZE
) {
926 if (!pte_present(*pte
))
928 page
= vm_normal_page(vma
, address
, *pte
);
929 BUG_ON(!page
|| PageAnon(page
));
932 mlock_vma_page(page
); /* no-op if already mlocked */
933 if (page
== check_page
)
935 continue; /* don't unmap */
938 if (ptep_clear_flush_young_notify(vma
, address
, pte
))
941 /* Nuke the page table entry. */
942 flush_cache_page(vma
, address
, pte_pfn(*pte
));
943 pteval
= ptep_clear_flush_notify(vma
, address
, pte
);
945 /* If nonlinear, store the file page offset in the pte. */
946 if (page
->index
!= linear_page_index(vma
, address
))
947 set_pte_at(mm
, address
, pte
, pgoff_to_pte(page
->index
));
949 /* Move the dirty bit to the physical page now the pte is gone. */
950 if (pte_dirty(pteval
))
951 set_page_dirty(page
);
953 page_remove_rmap(page
);
954 page_cache_release(page
);
955 dec_mm_counter(mm
, file_rss
);
958 pte_unmap_unlock(pte
- 1, ptl
);
960 up_read(&vma
->vm_mm
->mmap_sem
);
965 * common handling for pages mapped in VM_LOCKED vmas
967 static int try_to_mlock_page(struct page
*page
, struct vm_area_struct
*vma
)
971 if (down_read_trylock(&vma
->vm_mm
->mmap_sem
)) {
972 if (vma
->vm_flags
& VM_LOCKED
) {
973 mlock_vma_page(page
);
974 mlocked
++; /* really mlocked the page */
976 up_read(&vma
->vm_mm
->mmap_sem
);
982 * try_to_unmap_anon - unmap or unlock anonymous page using the object-based
984 * @page: the page to unmap/unlock
985 * @unlock: request for unlock rather than unmap [unlikely]
986 * @migration: unmapping for migration - ignored if @unlock
988 * Find all the mappings of a page using the mapping pointer and the vma chains
989 * contained in the anon_vma struct it points to.
991 * This function is only called from try_to_unmap/try_to_munlock for
993 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
994 * where the page was found will be held for write. So, we won't recheck
995 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
998 static int try_to_unmap_anon(struct page
*page
, int unlock
, int migration
)
1000 struct anon_vma
*anon_vma
;
1001 struct vm_area_struct
*vma
;
1002 unsigned int mlocked
= 0;
1003 int ret
= SWAP_AGAIN
;
1005 if (MLOCK_PAGES
&& unlikely(unlock
))
1006 ret
= SWAP_SUCCESS
; /* default for try_to_munlock() */
1008 anon_vma
= page_lock_anon_vma(page
);
1012 list_for_each_entry(vma
, &anon_vma
->head
, anon_vma_node
) {
1013 if (MLOCK_PAGES
&& unlikely(unlock
)) {
1014 if (!((vma
->vm_flags
& VM_LOCKED
) &&
1015 page_mapped_in_vma(page
, vma
)))
1016 continue; /* must visit all unlocked vmas */
1017 ret
= SWAP_MLOCK
; /* saw at least one mlocked vma */
1019 ret
= try_to_unmap_one(page
, vma
, migration
);
1020 if (ret
== SWAP_FAIL
|| !page_mapped(page
))
1023 if (ret
== SWAP_MLOCK
) {
1024 mlocked
= try_to_mlock_page(page
, vma
);
1026 break; /* stop if actually mlocked page */
1030 page_unlock_anon_vma(anon_vma
);
1033 ret
= SWAP_MLOCK
; /* actually mlocked the page */
1034 else if (ret
== SWAP_MLOCK
)
1035 ret
= SWAP_AGAIN
; /* saw VM_LOCKED vma */
1041 * try_to_unmap_file - unmap/unlock file page using the object-based rmap method
1042 * @page: the page to unmap/unlock
1043 * @unlock: request for unlock rather than unmap [unlikely]
1044 * @migration: unmapping for migration - ignored if @unlock
1046 * Find all the mappings of a page using the mapping pointer and the vma chains
1047 * contained in the address_space struct it points to.
1049 * This function is only called from try_to_unmap/try_to_munlock for
1050 * object-based pages.
1051 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1052 * where the page was found will be held for write. So, we won't recheck
1053 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1056 static int try_to_unmap_file(struct page
*page
, int unlock
, int migration
)
1058 struct address_space
*mapping
= page
->mapping
;
1059 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
1060 struct vm_area_struct
*vma
;
1061 struct prio_tree_iter iter
;
1062 int ret
= SWAP_AGAIN
;
1063 unsigned long cursor
;
1064 unsigned long max_nl_cursor
= 0;
1065 unsigned long max_nl_size
= 0;
1066 unsigned int mapcount
;
1067 unsigned int mlocked
= 0;
1069 if (MLOCK_PAGES
&& unlikely(unlock
))
1070 ret
= SWAP_SUCCESS
; /* default for try_to_munlock() */
1072 spin_lock(&mapping
->i_mmap_lock
);
1073 vma_prio_tree_foreach(vma
, &iter
, &mapping
->i_mmap
, pgoff
, pgoff
) {
1074 if (MLOCK_PAGES
&& unlikely(unlock
)) {
1075 if (!((vma
->vm_flags
& VM_LOCKED
) &&
1076 page_mapped_in_vma(page
, vma
)))
1077 continue; /* must visit all vmas */
1080 ret
= try_to_unmap_one(page
, vma
, migration
);
1081 if (ret
== SWAP_FAIL
|| !page_mapped(page
))
1084 if (ret
== SWAP_MLOCK
) {
1085 mlocked
= try_to_mlock_page(page
, vma
);
1087 break; /* stop if actually mlocked page */
1094 if (list_empty(&mapping
->i_mmap_nonlinear
))
1097 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
,
1098 shared
.vm_set
.list
) {
1099 if (MLOCK_PAGES
&& unlikely(unlock
)) {
1100 if (!(vma
->vm_flags
& VM_LOCKED
))
1101 continue; /* must visit all vmas */
1102 ret
= SWAP_MLOCK
; /* leave mlocked == 0 */
1103 goto out
; /* no need to look further */
1105 if (!MLOCK_PAGES
&& !migration
&& (vma
->vm_flags
& VM_LOCKED
))
1107 cursor
= (unsigned long) vma
->vm_private_data
;
1108 if (cursor
> max_nl_cursor
)
1109 max_nl_cursor
= cursor
;
1110 cursor
= vma
->vm_end
- vma
->vm_start
;
1111 if (cursor
> max_nl_size
)
1112 max_nl_size
= cursor
;
1115 if (max_nl_size
== 0) { /* all nonlinears locked or reserved ? */
1121 * We don't try to search for this page in the nonlinear vmas,
1122 * and page_referenced wouldn't have found it anyway. Instead
1123 * just walk the nonlinear vmas trying to age and unmap some.
1124 * The mapcount of the page we came in with is irrelevant,
1125 * but even so use it as a guide to how hard we should try?
1127 mapcount
= page_mapcount(page
);
1130 cond_resched_lock(&mapping
->i_mmap_lock
);
1132 max_nl_size
= (max_nl_size
+ CLUSTER_SIZE
- 1) & CLUSTER_MASK
;
1133 if (max_nl_cursor
== 0)
1134 max_nl_cursor
= CLUSTER_SIZE
;
1137 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
,
1138 shared
.vm_set
.list
) {
1139 if (!MLOCK_PAGES
&& !migration
&&
1140 (vma
->vm_flags
& VM_LOCKED
))
1142 cursor
= (unsigned long) vma
->vm_private_data
;
1143 while ( cursor
< max_nl_cursor
&&
1144 cursor
< vma
->vm_end
- vma
->vm_start
) {
1145 ret
= try_to_unmap_cluster(cursor
, &mapcount
,
1147 if (ret
== SWAP_MLOCK
)
1148 mlocked
= 2; /* to return below */
1149 cursor
+= CLUSTER_SIZE
;
1150 vma
->vm_private_data
= (void *) cursor
;
1151 if ((int)mapcount
<= 0)
1154 vma
->vm_private_data
= (void *) max_nl_cursor
;
1156 cond_resched_lock(&mapping
->i_mmap_lock
);
1157 max_nl_cursor
+= CLUSTER_SIZE
;
1158 } while (max_nl_cursor
<= max_nl_size
);
1161 * Don't loop forever (perhaps all the remaining pages are
1162 * in locked vmas). Reset cursor on all unreserved nonlinear
1163 * vmas, now forgetting on which ones it had fallen behind.
1165 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
, shared
.vm_set
.list
)
1166 vma
->vm_private_data
= NULL
;
1168 spin_unlock(&mapping
->i_mmap_lock
);
1170 ret
= SWAP_MLOCK
; /* actually mlocked the page */
1171 else if (ret
== SWAP_MLOCK
)
1172 ret
= SWAP_AGAIN
; /* saw VM_LOCKED vma */
1177 * try_to_unmap - try to remove all page table mappings to a page
1178 * @page: the page to get unmapped
1179 * @migration: migration flag
1181 * Tries to remove all the page table entries which are mapping this
1182 * page, used in the pageout path. Caller must hold the page lock.
1183 * Return values are:
1185 * SWAP_SUCCESS - we succeeded in removing all mappings
1186 * SWAP_AGAIN - we missed a mapping, try again later
1187 * SWAP_FAIL - the page is unswappable
1188 * SWAP_MLOCK - page is mlocked.
1190 int try_to_unmap(struct page
*page
, int migration
)
1194 BUG_ON(!PageLocked(page
));
1197 ret
= try_to_unmap_anon(page
, 0, migration
);
1199 ret
= try_to_unmap_file(page
, 0, migration
);
1200 if (ret
!= SWAP_MLOCK
&& !page_mapped(page
))
1205 #ifdef CONFIG_UNEVICTABLE_LRU
1207 * try_to_munlock - try to munlock a page
1208 * @page: the page to be munlocked
1210 * Called from munlock code. Checks all of the VMAs mapping the page
1211 * to make sure nobody else has this page mlocked. The page will be
1212 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1214 * Return values are:
1216 * SWAP_SUCCESS - no vma's holding page mlocked.
1217 * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem
1218 * SWAP_MLOCK - page is now mlocked.
1220 int try_to_munlock(struct page
*page
)
1222 VM_BUG_ON(!PageLocked(page
) || PageLRU(page
));
1225 return try_to_unmap_anon(page
, 1, 0);
1227 return try_to_unmap_file(page
, 1, 0);