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)
40 * (code doesn't rely on that order so it could be switched around)
42 * anon_vma->lock (memory_failure, collect_procs_anon)
47 #include <linux/pagemap.h>
48 #include <linux/swap.h>
49 #include <linux/swapops.h>
50 #include <linux/slab.h>
51 #include <linux/init.h>
52 #include <linux/ksm.h>
53 #include <linux/rmap.h>
54 #include <linux/rcupdate.h>
55 #include <linux/module.h>
56 #include <linux/memcontrol.h>
57 #include <linux/mmu_notifier.h>
58 #include <linux/migrate.h>
60 #include <asm/tlbflush.h>
64 static struct kmem_cache
*anon_vma_cachep
;
65 static struct kmem_cache
*anon_vma_chain_cachep
;
67 static inline struct anon_vma
*anon_vma_alloc(void)
69 return kmem_cache_alloc(anon_vma_cachep
, GFP_KERNEL
);
72 void anon_vma_free(struct anon_vma
*anon_vma
)
74 kmem_cache_free(anon_vma_cachep
, anon_vma
);
77 static inline struct anon_vma_chain
*anon_vma_chain_alloc(void)
79 return kmem_cache_alloc(anon_vma_chain_cachep
, GFP_KERNEL
);
82 void anon_vma_chain_free(struct anon_vma_chain
*anon_vma_chain
)
84 kmem_cache_free(anon_vma_chain_cachep
, anon_vma_chain
);
88 * anon_vma_prepare - attach an anon_vma to a memory region
89 * @vma: the memory region in question
91 * This makes sure the memory mapping described by 'vma' has
92 * an 'anon_vma' attached to it, so that we can associate the
93 * anonymous pages mapped into it with that anon_vma.
95 * The common case will be that we already have one, but if
96 * if not we either need to find an adjacent mapping that we
97 * can re-use the anon_vma from (very common when the only
98 * reason for splitting a vma has been mprotect()), or we
101 * Anon-vma allocations are very subtle, because we may have
102 * optimistically looked up an anon_vma in page_lock_anon_vma()
103 * and that may actually touch the spinlock even in the newly
104 * allocated vma (it depends on RCU to make sure that the
105 * anon_vma isn't actually destroyed).
107 * As a result, we need to do proper anon_vma locking even
108 * for the new allocation. At the same time, we do not want
109 * to do any locking for the common case of already having
112 * This must be called with the mmap_sem held for reading.
114 int anon_vma_prepare(struct vm_area_struct
*vma
)
116 struct anon_vma
*anon_vma
= vma
->anon_vma
;
117 struct anon_vma_chain
*avc
;
120 if (unlikely(!anon_vma
)) {
121 struct mm_struct
*mm
= vma
->vm_mm
;
122 struct anon_vma
*allocated
;
124 avc
= anon_vma_chain_alloc();
128 anon_vma
= find_mergeable_anon_vma(vma
);
131 anon_vma
= anon_vma_alloc();
132 if (unlikely(!anon_vma
))
133 goto out_enomem_free_avc
;
134 allocated
= anon_vma
;
136 spin_lock(&anon_vma
->lock
);
138 /* page_table_lock to protect against threads */
139 spin_lock(&mm
->page_table_lock
);
140 if (likely(!vma
->anon_vma
)) {
141 vma
->anon_vma
= anon_vma
;
142 avc
->anon_vma
= anon_vma
;
144 list_add(&avc
->same_vma
, &vma
->anon_vma_chain
);
145 list_add(&avc
->same_anon_vma
, &anon_vma
->head
);
148 spin_unlock(&mm
->page_table_lock
);
150 spin_unlock(&anon_vma
->lock
);
151 if (unlikely(allocated
)) {
152 anon_vma_free(allocated
);
153 anon_vma_chain_free(avc
);
159 anon_vma_chain_free(avc
);
164 static void anon_vma_chain_link(struct vm_area_struct
*vma
,
165 struct anon_vma_chain
*avc
,
166 struct anon_vma
*anon_vma
)
169 avc
->anon_vma
= anon_vma
;
170 list_add(&avc
->same_vma
, &vma
->anon_vma_chain
);
172 spin_lock(&anon_vma
->lock
);
173 list_add_tail(&avc
->same_anon_vma
, &anon_vma
->head
);
174 spin_unlock(&anon_vma
->lock
);
178 * Attach the anon_vmas from src to dst.
179 * Returns 0 on success, -ENOMEM on failure.
181 int anon_vma_clone(struct vm_area_struct
*dst
, struct vm_area_struct
*src
)
183 struct anon_vma_chain
*avc
, *pavc
;
185 list_for_each_entry_reverse(pavc
, &src
->anon_vma_chain
, same_vma
) {
186 avc
= anon_vma_chain_alloc();
189 anon_vma_chain_link(dst
, avc
, pavc
->anon_vma
);
194 unlink_anon_vmas(dst
);
199 * Attach vma to its own anon_vma, as well as to the anon_vmas that
200 * the corresponding VMA in the parent process is attached to.
201 * Returns 0 on success, non-zero on failure.
203 int anon_vma_fork(struct vm_area_struct
*vma
, struct vm_area_struct
*pvma
)
205 struct anon_vma_chain
*avc
;
206 struct anon_vma
*anon_vma
;
208 /* Don't bother if the parent process has no anon_vma here. */
213 * First, attach the new VMA to the parent VMA's anon_vmas,
214 * so rmap can find non-COWed pages in child processes.
216 if (anon_vma_clone(vma
, pvma
))
219 /* Then add our own anon_vma. */
220 anon_vma
= anon_vma_alloc();
223 avc
= anon_vma_chain_alloc();
225 goto out_error_free_anon_vma
;
226 anon_vma_chain_link(vma
, avc
, anon_vma
);
227 /* Mark this anon_vma as the one where our new (COWed) pages go. */
228 vma
->anon_vma
= anon_vma
;
232 out_error_free_anon_vma
:
233 anon_vma_free(anon_vma
);
235 unlink_anon_vmas(vma
);
239 static void anon_vma_unlink(struct anon_vma_chain
*anon_vma_chain
)
241 struct anon_vma
*anon_vma
= anon_vma_chain
->anon_vma
;
244 /* If anon_vma_fork fails, we can get an empty anon_vma_chain. */
248 spin_lock(&anon_vma
->lock
);
249 list_del(&anon_vma_chain
->same_anon_vma
);
251 /* We must garbage collect the anon_vma if it's empty */
252 empty
= list_empty(&anon_vma
->head
) && !ksm_refcount(anon_vma
);
253 spin_unlock(&anon_vma
->lock
);
256 anon_vma_free(anon_vma
);
259 void unlink_anon_vmas(struct vm_area_struct
*vma
)
261 struct anon_vma_chain
*avc
, *next
;
263 /* Unlink each anon_vma chained to the VMA. */
264 list_for_each_entry_safe(avc
, next
, &vma
->anon_vma_chain
, same_vma
) {
265 anon_vma_unlink(avc
);
266 list_del(&avc
->same_vma
);
267 anon_vma_chain_free(avc
);
271 static void anon_vma_ctor(void *data
)
273 struct anon_vma
*anon_vma
= data
;
275 spin_lock_init(&anon_vma
->lock
);
276 ksm_refcount_init(anon_vma
);
277 INIT_LIST_HEAD(&anon_vma
->head
);
280 void __init
anon_vma_init(void)
282 anon_vma_cachep
= kmem_cache_create("anon_vma", sizeof(struct anon_vma
),
283 0, SLAB_DESTROY_BY_RCU
|SLAB_PANIC
, anon_vma_ctor
);
284 anon_vma_chain_cachep
= KMEM_CACHE(anon_vma_chain
, SLAB_PANIC
);
288 * Getting a lock on a stable anon_vma from a page off the LRU is
289 * tricky: page_lock_anon_vma rely on RCU to guard against the races.
291 struct anon_vma
*page_lock_anon_vma(struct page
*page
)
293 struct anon_vma
*anon_vma
;
294 unsigned long anon_mapping
;
297 anon_mapping
= (unsigned long) ACCESS_ONCE(page
->mapping
);
298 if ((anon_mapping
& PAGE_MAPPING_FLAGS
) != PAGE_MAPPING_ANON
)
300 if (!page_mapped(page
))
303 anon_vma
= (struct anon_vma
*) (anon_mapping
- PAGE_MAPPING_ANON
);
304 spin_lock(&anon_vma
->lock
);
311 void page_unlock_anon_vma(struct anon_vma
*anon_vma
)
313 spin_unlock(&anon_vma
->lock
);
318 * At what user virtual address is page expected in @vma?
319 * Returns virtual address or -EFAULT if page's index/offset is not
320 * within the range mapped the @vma.
322 static inline unsigned long
323 vma_address(struct page
*page
, struct vm_area_struct
*vma
)
325 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
326 unsigned long address
;
328 address
= vma
->vm_start
+ ((pgoff
- vma
->vm_pgoff
) << PAGE_SHIFT
);
329 if (unlikely(address
< vma
->vm_start
|| address
>= vma
->vm_end
)) {
330 /* page should be within @vma mapping range */
337 * At what user virtual address is page expected in vma?
338 * checking that the page matches the vma.
340 unsigned long page_address_in_vma(struct page
*page
, struct vm_area_struct
*vma
)
342 if (PageAnon(page
)) {
343 if (vma
->anon_vma
!= page_anon_vma(page
))
345 } else if (page
->mapping
&& !(vma
->vm_flags
& VM_NONLINEAR
)) {
347 vma
->vm_file
->f_mapping
!= page
->mapping
)
351 return vma_address(page
, vma
);
355 * Check that @page is mapped at @address into @mm.
357 * If @sync is false, page_check_address may perform a racy check to avoid
358 * the page table lock when the pte is not present (helpful when reclaiming
359 * highly shared pages).
361 * On success returns with pte mapped and locked.
363 pte_t
*page_check_address(struct page
*page
, struct mm_struct
*mm
,
364 unsigned long address
, spinlock_t
**ptlp
, int sync
)
372 pgd
= pgd_offset(mm
, address
);
373 if (!pgd_present(*pgd
))
376 pud
= pud_offset(pgd
, address
);
377 if (!pud_present(*pud
))
380 pmd
= pmd_offset(pud
, address
);
381 if (!pmd_present(*pmd
))
384 pte
= pte_offset_map(pmd
, address
);
385 /* Make a quick check before getting the lock */
386 if (!sync
&& !pte_present(*pte
)) {
391 ptl
= pte_lockptr(mm
, pmd
);
393 if (pte_present(*pte
) && page_to_pfn(page
) == pte_pfn(*pte
)) {
397 pte_unmap_unlock(pte
, ptl
);
402 * page_mapped_in_vma - check whether a page is really mapped in a VMA
403 * @page: the page to test
404 * @vma: the VMA to test
406 * Returns 1 if the page is mapped into the page tables of the VMA, 0
407 * if the page is not mapped into the page tables of this VMA. Only
408 * valid for normal file or anonymous VMAs.
410 int page_mapped_in_vma(struct page
*page
, struct vm_area_struct
*vma
)
412 unsigned long address
;
416 address
= vma_address(page
, vma
);
417 if (address
== -EFAULT
) /* out of vma range */
419 pte
= page_check_address(page
, vma
->vm_mm
, address
, &ptl
, 1);
420 if (!pte
) /* the page is not in this mm */
422 pte_unmap_unlock(pte
, ptl
);
428 * Subfunctions of page_referenced: page_referenced_one called
429 * repeatedly from either page_referenced_anon or page_referenced_file.
431 int page_referenced_one(struct page
*page
, struct vm_area_struct
*vma
,
432 unsigned long address
, unsigned int *mapcount
,
433 unsigned long *vm_flags
)
435 struct mm_struct
*mm
= vma
->vm_mm
;
440 pte
= page_check_address(page
, mm
, address
, &ptl
, 0);
445 * Don't want to elevate referenced for mlocked page that gets this far,
446 * in order that it progresses to try_to_unmap and is moved to the
449 if (vma
->vm_flags
& VM_LOCKED
) {
450 *mapcount
= 1; /* break early from loop */
451 *vm_flags
|= VM_LOCKED
;
455 if (ptep_clear_flush_young_notify(vma
, address
, pte
)) {
457 * Don't treat a reference through a sequentially read
458 * mapping as such. If the page has been used in
459 * another mapping, we will catch it; if this other
460 * mapping is already gone, the unmap path will have
461 * set PG_referenced or activated the page.
463 if (likely(!VM_SequentialReadHint(vma
)))
467 /* Pretend the page is referenced if the task has the
468 swap token and is in the middle of a page fault. */
469 if (mm
!= current
->mm
&& has_swap_token(mm
) &&
470 rwsem_is_locked(&mm
->mmap_sem
))
475 pte_unmap_unlock(pte
, ptl
);
478 *vm_flags
|= vma
->vm_flags
;
483 static int page_referenced_anon(struct page
*page
,
484 struct mem_cgroup
*mem_cont
,
485 unsigned long *vm_flags
)
487 unsigned int mapcount
;
488 struct anon_vma
*anon_vma
;
489 struct anon_vma_chain
*avc
;
492 anon_vma
= page_lock_anon_vma(page
);
496 mapcount
= page_mapcount(page
);
497 list_for_each_entry(avc
, &anon_vma
->head
, same_anon_vma
) {
498 struct vm_area_struct
*vma
= avc
->vma
;
499 unsigned long address
= vma_address(page
, vma
);
500 if (address
== -EFAULT
)
503 * If we are reclaiming on behalf of a cgroup, skip
504 * counting on behalf of references from different
507 if (mem_cont
&& !mm_match_cgroup(vma
->vm_mm
, mem_cont
))
509 referenced
+= page_referenced_one(page
, vma
, address
,
510 &mapcount
, vm_flags
);
515 page_unlock_anon_vma(anon_vma
);
520 * page_referenced_file - referenced check for object-based rmap
521 * @page: the page we're checking references on.
522 * @mem_cont: target memory controller
523 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
525 * For an object-based mapped page, find all the places it is mapped and
526 * check/clear the referenced flag. This is done by following the page->mapping
527 * pointer, then walking the chain of vmas it holds. It returns the number
528 * of references it found.
530 * This function is only called from page_referenced for object-based pages.
532 static int page_referenced_file(struct page
*page
,
533 struct mem_cgroup
*mem_cont
,
534 unsigned long *vm_flags
)
536 unsigned int mapcount
;
537 struct address_space
*mapping
= page
->mapping
;
538 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
539 struct vm_area_struct
*vma
;
540 struct prio_tree_iter iter
;
544 * The caller's checks on page->mapping and !PageAnon have made
545 * sure that this is a file page: the check for page->mapping
546 * excludes the case just before it gets set on an anon page.
548 BUG_ON(PageAnon(page
));
551 * The page lock not only makes sure that page->mapping cannot
552 * suddenly be NULLified by truncation, it makes sure that the
553 * structure at mapping cannot be freed and reused yet,
554 * so we can safely take mapping->i_mmap_lock.
556 BUG_ON(!PageLocked(page
));
558 spin_lock(&mapping
->i_mmap_lock
);
561 * i_mmap_lock does not stabilize mapcount at all, but mapcount
562 * is more likely to be accurate if we note it after spinning.
564 mapcount
= page_mapcount(page
);
566 vma_prio_tree_foreach(vma
, &iter
, &mapping
->i_mmap
, pgoff
, pgoff
) {
567 unsigned long address
= vma_address(page
, vma
);
568 if (address
== -EFAULT
)
571 * If we are reclaiming on behalf of a cgroup, skip
572 * counting on behalf of references from different
575 if (mem_cont
&& !mm_match_cgroup(vma
->vm_mm
, mem_cont
))
577 referenced
+= page_referenced_one(page
, vma
, address
,
578 &mapcount
, vm_flags
);
583 spin_unlock(&mapping
->i_mmap_lock
);
588 * page_referenced - test if the page was referenced
589 * @page: the page to test
590 * @is_locked: caller holds lock on the page
591 * @mem_cont: target memory controller
592 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
594 * Quick test_and_clear_referenced for all mappings to a page,
595 * returns the number of ptes which referenced the page.
597 int page_referenced(struct page
*page
,
599 struct mem_cgroup
*mem_cont
,
600 unsigned long *vm_flags
)
606 if (page_mapped(page
) && page_rmapping(page
)) {
607 if (!is_locked
&& (!PageAnon(page
) || PageKsm(page
))) {
608 we_locked
= trylock_page(page
);
614 if (unlikely(PageKsm(page
)))
615 referenced
+= page_referenced_ksm(page
, mem_cont
,
617 else if (PageAnon(page
))
618 referenced
+= page_referenced_anon(page
, mem_cont
,
620 else if (page
->mapping
)
621 referenced
+= page_referenced_file(page
, mem_cont
,
627 if (page_test_and_clear_young(page
))
633 static int page_mkclean_one(struct page
*page
, struct vm_area_struct
*vma
,
634 unsigned long address
)
636 struct mm_struct
*mm
= vma
->vm_mm
;
641 pte
= page_check_address(page
, mm
, address
, &ptl
, 1);
645 if (pte_dirty(*pte
) || pte_write(*pte
)) {
648 flush_cache_page(vma
, address
, pte_pfn(*pte
));
649 entry
= ptep_clear_flush_notify(vma
, address
, pte
);
650 entry
= pte_wrprotect(entry
);
651 entry
= pte_mkclean(entry
);
652 set_pte_at(mm
, address
, pte
, entry
);
656 pte_unmap_unlock(pte
, ptl
);
661 static int page_mkclean_file(struct address_space
*mapping
, struct page
*page
)
663 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
664 struct vm_area_struct
*vma
;
665 struct prio_tree_iter iter
;
668 BUG_ON(PageAnon(page
));
670 spin_lock(&mapping
->i_mmap_lock
);
671 vma_prio_tree_foreach(vma
, &iter
, &mapping
->i_mmap
, pgoff
, pgoff
) {
672 if (vma
->vm_flags
& VM_SHARED
) {
673 unsigned long address
= vma_address(page
, vma
);
674 if (address
== -EFAULT
)
676 ret
+= page_mkclean_one(page
, vma
, address
);
679 spin_unlock(&mapping
->i_mmap_lock
);
683 int page_mkclean(struct page
*page
)
687 BUG_ON(!PageLocked(page
));
689 if (page_mapped(page
)) {
690 struct address_space
*mapping
= page_mapping(page
);
692 ret
= page_mkclean_file(mapping
, page
);
693 if (page_test_dirty(page
)) {
694 page_clear_dirty(page
);
702 EXPORT_SYMBOL_GPL(page_mkclean
);
705 * page_move_anon_rmap - move a page to our anon_vma
706 * @page: the page to move to our anon_vma
707 * @vma: the vma the page belongs to
708 * @address: the user virtual address mapped
710 * When a page belongs exclusively to one process after a COW event,
711 * that page can be moved into the anon_vma that belongs to just that
712 * process, so the rmap code will not search the parent or sibling
715 void page_move_anon_rmap(struct page
*page
,
716 struct vm_area_struct
*vma
, unsigned long address
)
718 struct anon_vma
*anon_vma
= vma
->anon_vma
;
720 VM_BUG_ON(!PageLocked(page
));
721 VM_BUG_ON(!anon_vma
);
722 VM_BUG_ON(page
->index
!= linear_page_index(vma
, address
));
724 anon_vma
= (void *) anon_vma
+ PAGE_MAPPING_ANON
;
725 page
->mapping
= (struct address_space
*) anon_vma
;
729 * __page_set_anon_rmap - setup new anonymous rmap
730 * @page: the page to add the mapping to
731 * @vma: the vm area in which the mapping is added
732 * @address: the user virtual address mapped
734 static void __page_set_anon_rmap(struct page
*page
,
735 struct vm_area_struct
*vma
, unsigned long address
)
737 struct anon_vma_chain
*avc
;
738 struct anon_vma
*anon_vma
;
740 BUG_ON(!vma
->anon_vma
);
743 * We must use the _oldest_ possible anon_vma for the page mapping!
745 * So take the last AVC chain entry in the vma, which is the deepest
746 * ancestor, and use the anon_vma from that.
748 avc
= list_entry(vma
->anon_vma_chain
.prev
, struct anon_vma_chain
, same_vma
);
749 anon_vma
= avc
->anon_vma
;
751 anon_vma
= (void *) anon_vma
+ PAGE_MAPPING_ANON
;
752 page
->mapping
= (struct address_space
*) anon_vma
;
753 page
->index
= linear_page_index(vma
, address
);
757 * __page_check_anon_rmap - sanity check anonymous rmap addition
758 * @page: the page to add the mapping to
759 * @vma: the vm area in which the mapping is added
760 * @address: the user virtual address mapped
762 static void __page_check_anon_rmap(struct page
*page
,
763 struct vm_area_struct
*vma
, unsigned long address
)
765 #ifdef CONFIG_DEBUG_VM
767 * The page's anon-rmap details (mapping and index) are guaranteed to
768 * be set up correctly at this point.
770 * We have exclusion against page_add_anon_rmap because the caller
771 * always holds the page locked, except if called from page_dup_rmap,
772 * in which case the page is already known to be setup.
774 * We have exclusion against page_add_new_anon_rmap because those pages
775 * are initially only visible via the pagetables, and the pte is locked
776 * over the call to page_add_new_anon_rmap.
778 BUG_ON(page
->index
!= linear_page_index(vma
, address
));
783 * page_add_anon_rmap - add pte mapping to an anonymous page
784 * @page: the page to add the mapping to
785 * @vma: the vm area in which the mapping is added
786 * @address: the user virtual address mapped
788 * The caller needs to hold the pte lock, and the page must be locked in
789 * the anon_vma case: to serialize mapping,index checking after setting,
790 * and to ensure that PageAnon is not being upgraded racily to PageKsm
791 * (but PageKsm is never downgraded to PageAnon).
793 void page_add_anon_rmap(struct page
*page
,
794 struct vm_area_struct
*vma
, unsigned long address
)
796 int first
= atomic_inc_and_test(&page
->_mapcount
);
798 __inc_zone_page_state(page
, NR_ANON_PAGES
);
799 if (unlikely(PageKsm(page
)))
802 VM_BUG_ON(!PageLocked(page
));
803 VM_BUG_ON(address
< vma
->vm_start
|| address
>= vma
->vm_end
);
805 __page_set_anon_rmap(page
, vma
, address
);
807 __page_check_anon_rmap(page
, vma
, address
);
811 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
812 * @page: the page to add the mapping to
813 * @vma: the vm area in which the mapping is added
814 * @address: the user virtual address mapped
816 * Same as page_add_anon_rmap but must only be called on *new* pages.
817 * This means the inc-and-test can be bypassed.
818 * Page does not have to be locked.
820 void page_add_new_anon_rmap(struct page
*page
,
821 struct vm_area_struct
*vma
, unsigned long address
)
823 VM_BUG_ON(address
< vma
->vm_start
|| address
>= vma
->vm_end
);
824 SetPageSwapBacked(page
);
825 atomic_set(&page
->_mapcount
, 0); /* increment count (starts at -1) */
826 __inc_zone_page_state(page
, NR_ANON_PAGES
);
827 __page_set_anon_rmap(page
, vma
, address
);
828 if (page_evictable(page
, vma
))
829 lru_cache_add_lru(page
, LRU_ACTIVE_ANON
);
831 add_page_to_unevictable_list(page
);
835 * page_add_file_rmap - add pte mapping to a file page
836 * @page: the page to add the mapping to
838 * The caller needs to hold the pte lock.
840 void page_add_file_rmap(struct page
*page
)
842 if (atomic_inc_and_test(&page
->_mapcount
)) {
843 __inc_zone_page_state(page
, NR_FILE_MAPPED
);
844 mem_cgroup_update_file_mapped(page
, 1);
849 * page_remove_rmap - take down pte mapping from a page
850 * @page: page to remove mapping from
852 * The caller needs to hold the pte lock.
854 void page_remove_rmap(struct page
*page
)
856 /* page still mapped by someone else? */
857 if (!atomic_add_negative(-1, &page
->_mapcount
))
861 * Now that the last pte has gone, s390 must transfer dirty
862 * flag from storage key to struct page. We can usually skip
863 * this if the page is anon, so about to be freed; but perhaps
864 * not if it's in swapcache - there might be another pte slot
865 * containing the swap entry, but page not yet written to swap.
867 if ((!PageAnon(page
) || PageSwapCache(page
)) && page_test_dirty(page
)) {
868 page_clear_dirty(page
);
869 set_page_dirty(page
);
871 if (PageAnon(page
)) {
872 mem_cgroup_uncharge_page(page
);
873 __dec_zone_page_state(page
, NR_ANON_PAGES
);
875 __dec_zone_page_state(page
, NR_FILE_MAPPED
);
876 mem_cgroup_update_file_mapped(page
, -1);
879 * It would be tidy to reset the PageAnon mapping here,
880 * but that might overwrite a racing page_add_anon_rmap
881 * which increments mapcount after us but sets mapping
882 * before us: so leave the reset to free_hot_cold_page,
883 * and remember that it's only reliable while mapped.
884 * Leaving it set also helps swapoff to reinstate ptes
885 * faster for those pages still in swapcache.
890 * Subfunctions of try_to_unmap: try_to_unmap_one called
891 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
893 int try_to_unmap_one(struct page
*page
, struct vm_area_struct
*vma
,
894 unsigned long address
, enum ttu_flags flags
)
896 struct mm_struct
*mm
= vma
->vm_mm
;
900 int ret
= SWAP_AGAIN
;
902 pte
= page_check_address(page
, mm
, address
, &ptl
, 0);
907 * If the page is mlock()d, we cannot swap it out.
908 * If it's recently referenced (perhaps page_referenced
909 * skipped over this mm) then we should reactivate it.
911 if (!(flags
& TTU_IGNORE_MLOCK
)) {
912 if (vma
->vm_flags
& VM_LOCKED
)
915 if (TTU_ACTION(flags
) == TTU_MUNLOCK
)
918 if (!(flags
& TTU_IGNORE_ACCESS
)) {
919 if (ptep_clear_flush_young_notify(vma
, address
, pte
)) {
925 /* Nuke the page table entry. */
926 flush_cache_page(vma
, address
, page_to_pfn(page
));
927 pteval
= ptep_clear_flush_notify(vma
, address
, pte
);
929 /* Move the dirty bit to the physical page now the pte is gone. */
930 if (pte_dirty(pteval
))
931 set_page_dirty(page
);
933 /* Update high watermark before we lower rss */
934 update_hiwater_rss(mm
);
936 if (PageHWPoison(page
) && !(flags
& TTU_IGNORE_HWPOISON
)) {
938 dec_mm_counter(mm
, MM_ANONPAGES
);
940 dec_mm_counter(mm
, MM_FILEPAGES
);
941 set_pte_at(mm
, address
, pte
,
942 swp_entry_to_pte(make_hwpoison_entry(page
)));
943 } else if (PageAnon(page
)) {
944 swp_entry_t entry
= { .val
= page_private(page
) };
946 if (PageSwapCache(page
)) {
948 * Store the swap location in the pte.
949 * See handle_pte_fault() ...
951 if (swap_duplicate(entry
) < 0) {
952 set_pte_at(mm
, address
, pte
, pteval
);
956 if (list_empty(&mm
->mmlist
)) {
957 spin_lock(&mmlist_lock
);
958 if (list_empty(&mm
->mmlist
))
959 list_add(&mm
->mmlist
, &init_mm
.mmlist
);
960 spin_unlock(&mmlist_lock
);
962 dec_mm_counter(mm
, MM_ANONPAGES
);
963 inc_mm_counter(mm
, MM_SWAPENTS
);
964 } else if (PAGE_MIGRATION
) {
966 * Store the pfn of the page in a special migration
967 * pte. do_swap_page() will wait until the migration
968 * pte is removed and then restart fault handling.
970 BUG_ON(TTU_ACTION(flags
) != TTU_MIGRATION
);
971 entry
= make_migration_entry(page
, pte_write(pteval
));
973 set_pte_at(mm
, address
, pte
, swp_entry_to_pte(entry
));
974 BUG_ON(pte_file(*pte
));
975 } else if (PAGE_MIGRATION
&& (TTU_ACTION(flags
) == TTU_MIGRATION
)) {
976 /* Establish migration entry for a file page */
978 entry
= make_migration_entry(page
, pte_write(pteval
));
979 set_pte_at(mm
, address
, pte
, swp_entry_to_pte(entry
));
981 dec_mm_counter(mm
, MM_FILEPAGES
);
983 page_remove_rmap(page
);
984 page_cache_release(page
);
987 pte_unmap_unlock(pte
, ptl
);
992 pte_unmap_unlock(pte
, ptl
);
996 * We need mmap_sem locking, Otherwise VM_LOCKED check makes
997 * unstable result and race. Plus, We can't wait here because
998 * we now hold anon_vma->lock or mapping->i_mmap_lock.
999 * if trylock failed, the page remain in evictable lru and later
1000 * vmscan could retry to move the page to unevictable lru if the
1001 * page is actually mlocked.
1003 if (down_read_trylock(&vma
->vm_mm
->mmap_sem
)) {
1004 if (vma
->vm_flags
& VM_LOCKED
) {
1005 mlock_vma_page(page
);
1008 up_read(&vma
->vm_mm
->mmap_sem
);
1014 * objrmap doesn't work for nonlinear VMAs because the assumption that
1015 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
1016 * Consequently, given a particular page and its ->index, we cannot locate the
1017 * ptes which are mapping that page without an exhaustive linear search.
1019 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
1020 * maps the file to which the target page belongs. The ->vm_private_data field
1021 * holds the current cursor into that scan. Successive searches will circulate
1022 * around the vma's virtual address space.
1024 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
1025 * more scanning pressure is placed against them as well. Eventually pages
1026 * will become fully unmapped and are eligible for eviction.
1028 * For very sparsely populated VMAs this is a little inefficient - chances are
1029 * there there won't be many ptes located within the scan cluster. In this case
1030 * maybe we could scan further - to the end of the pte page, perhaps.
1032 * Mlocked pages: check VM_LOCKED under mmap_sem held for read, if we can
1033 * acquire it without blocking. If vma locked, mlock the pages in the cluster,
1034 * rather than unmapping them. If we encounter the "check_page" that vmscan is
1035 * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN.
1037 #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
1038 #define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
1040 static int try_to_unmap_cluster(unsigned long cursor
, unsigned int *mapcount
,
1041 struct vm_area_struct
*vma
, struct page
*check_page
)
1043 struct mm_struct
*mm
= vma
->vm_mm
;
1051 unsigned long address
;
1053 int ret
= SWAP_AGAIN
;
1056 address
= (vma
->vm_start
+ cursor
) & CLUSTER_MASK
;
1057 end
= address
+ CLUSTER_SIZE
;
1058 if (address
< vma
->vm_start
)
1059 address
= vma
->vm_start
;
1060 if (end
> vma
->vm_end
)
1063 pgd
= pgd_offset(mm
, address
);
1064 if (!pgd_present(*pgd
))
1067 pud
= pud_offset(pgd
, address
);
1068 if (!pud_present(*pud
))
1071 pmd
= pmd_offset(pud
, address
);
1072 if (!pmd_present(*pmd
))
1076 * If we can acquire the mmap_sem for read, and vma is VM_LOCKED,
1077 * keep the sem while scanning the cluster for mlocking pages.
1079 if (down_read_trylock(&vma
->vm_mm
->mmap_sem
)) {
1080 locked_vma
= (vma
->vm_flags
& VM_LOCKED
);
1082 up_read(&vma
->vm_mm
->mmap_sem
); /* don't need it */
1085 pte
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
1087 /* Update high watermark before we lower rss */
1088 update_hiwater_rss(mm
);
1090 for (; address
< end
; pte
++, address
+= PAGE_SIZE
) {
1091 if (!pte_present(*pte
))
1093 page
= vm_normal_page(vma
, address
, *pte
);
1094 BUG_ON(!page
|| PageAnon(page
));
1097 mlock_vma_page(page
); /* no-op if already mlocked */
1098 if (page
== check_page
)
1100 continue; /* don't unmap */
1103 if (ptep_clear_flush_young_notify(vma
, address
, pte
))
1106 /* Nuke the page table entry. */
1107 flush_cache_page(vma
, address
, pte_pfn(*pte
));
1108 pteval
= ptep_clear_flush_notify(vma
, address
, pte
);
1110 /* If nonlinear, store the file page offset in the pte. */
1111 if (page
->index
!= linear_page_index(vma
, address
))
1112 set_pte_at(mm
, address
, pte
, pgoff_to_pte(page
->index
));
1114 /* Move the dirty bit to the physical page now the pte is gone. */
1115 if (pte_dirty(pteval
))
1116 set_page_dirty(page
);
1118 page_remove_rmap(page
);
1119 page_cache_release(page
);
1120 dec_mm_counter(mm
, MM_FILEPAGES
);
1123 pte_unmap_unlock(pte
- 1, ptl
);
1125 up_read(&vma
->vm_mm
->mmap_sem
);
1130 * try_to_unmap_anon - unmap or unlock anonymous page using the object-based
1132 * @page: the page to unmap/unlock
1133 * @flags: action and flags
1135 * Find all the mappings of a page using the mapping pointer and the vma chains
1136 * contained in the anon_vma struct it points to.
1138 * This function is only called from try_to_unmap/try_to_munlock for
1140 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1141 * where the page was found will be held for write. So, we won't recheck
1142 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1145 static int try_to_unmap_anon(struct page
*page
, enum ttu_flags flags
)
1147 struct anon_vma
*anon_vma
;
1148 struct anon_vma_chain
*avc
;
1149 int ret
= SWAP_AGAIN
;
1151 anon_vma
= page_lock_anon_vma(page
);
1155 list_for_each_entry(avc
, &anon_vma
->head
, same_anon_vma
) {
1156 struct vm_area_struct
*vma
= avc
->vma
;
1157 unsigned long address
= vma_address(page
, vma
);
1158 if (address
== -EFAULT
)
1160 ret
= try_to_unmap_one(page
, vma
, address
, flags
);
1161 if (ret
!= SWAP_AGAIN
|| !page_mapped(page
))
1165 page_unlock_anon_vma(anon_vma
);
1170 * try_to_unmap_file - unmap/unlock file page using the object-based rmap method
1171 * @page: the page to unmap/unlock
1172 * @flags: action and flags
1174 * Find all the mappings of a page using the mapping pointer and the vma chains
1175 * contained in the address_space struct it points to.
1177 * This function is only called from try_to_unmap/try_to_munlock for
1178 * object-based pages.
1179 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1180 * where the page was found will be held for write. So, we won't recheck
1181 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1184 static int try_to_unmap_file(struct page
*page
, enum ttu_flags flags
)
1186 struct address_space
*mapping
= page
->mapping
;
1187 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
1188 struct vm_area_struct
*vma
;
1189 struct prio_tree_iter iter
;
1190 int ret
= SWAP_AGAIN
;
1191 unsigned long cursor
;
1192 unsigned long max_nl_cursor
= 0;
1193 unsigned long max_nl_size
= 0;
1194 unsigned int mapcount
;
1196 spin_lock(&mapping
->i_mmap_lock
);
1197 vma_prio_tree_foreach(vma
, &iter
, &mapping
->i_mmap
, pgoff
, pgoff
) {
1198 unsigned long address
= vma_address(page
, vma
);
1199 if (address
== -EFAULT
)
1201 ret
= try_to_unmap_one(page
, vma
, address
, flags
);
1202 if (ret
!= SWAP_AGAIN
|| !page_mapped(page
))
1206 if (list_empty(&mapping
->i_mmap_nonlinear
))
1210 * We don't bother to try to find the munlocked page in nonlinears.
1211 * It's costly. Instead, later, page reclaim logic may call
1212 * try_to_unmap(TTU_MUNLOCK) and recover PG_mlocked lazily.
1214 if (TTU_ACTION(flags
) == TTU_MUNLOCK
)
1217 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
,
1218 shared
.vm_set
.list
) {
1219 cursor
= (unsigned long) vma
->vm_private_data
;
1220 if (cursor
> max_nl_cursor
)
1221 max_nl_cursor
= cursor
;
1222 cursor
= vma
->vm_end
- vma
->vm_start
;
1223 if (cursor
> max_nl_size
)
1224 max_nl_size
= cursor
;
1227 if (max_nl_size
== 0) { /* all nonlinears locked or reserved ? */
1233 * We don't try to search for this page in the nonlinear vmas,
1234 * and page_referenced wouldn't have found it anyway. Instead
1235 * just walk the nonlinear vmas trying to age and unmap some.
1236 * The mapcount of the page we came in with is irrelevant,
1237 * but even so use it as a guide to how hard we should try?
1239 mapcount
= page_mapcount(page
);
1242 cond_resched_lock(&mapping
->i_mmap_lock
);
1244 max_nl_size
= (max_nl_size
+ CLUSTER_SIZE
- 1) & CLUSTER_MASK
;
1245 if (max_nl_cursor
== 0)
1246 max_nl_cursor
= CLUSTER_SIZE
;
1249 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
,
1250 shared
.vm_set
.list
) {
1251 cursor
= (unsigned long) vma
->vm_private_data
;
1252 while ( cursor
< max_nl_cursor
&&
1253 cursor
< vma
->vm_end
- vma
->vm_start
) {
1254 if (try_to_unmap_cluster(cursor
, &mapcount
,
1255 vma
, page
) == SWAP_MLOCK
)
1257 cursor
+= CLUSTER_SIZE
;
1258 vma
->vm_private_data
= (void *) cursor
;
1259 if ((int)mapcount
<= 0)
1262 vma
->vm_private_data
= (void *) max_nl_cursor
;
1264 cond_resched_lock(&mapping
->i_mmap_lock
);
1265 max_nl_cursor
+= CLUSTER_SIZE
;
1266 } while (max_nl_cursor
<= max_nl_size
);
1269 * Don't loop forever (perhaps all the remaining pages are
1270 * in locked vmas). Reset cursor on all unreserved nonlinear
1271 * vmas, now forgetting on which ones it had fallen behind.
1273 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
, shared
.vm_set
.list
)
1274 vma
->vm_private_data
= NULL
;
1276 spin_unlock(&mapping
->i_mmap_lock
);
1281 * try_to_unmap - try to remove all page table mappings to a page
1282 * @page: the page to get unmapped
1283 * @flags: action and flags
1285 * Tries to remove all the page table entries which are mapping this
1286 * page, used in the pageout path. Caller must hold the page lock.
1287 * Return values are:
1289 * SWAP_SUCCESS - we succeeded in removing all mappings
1290 * SWAP_AGAIN - we missed a mapping, try again later
1291 * SWAP_FAIL - the page is unswappable
1292 * SWAP_MLOCK - page is mlocked.
1294 int try_to_unmap(struct page
*page
, enum ttu_flags flags
)
1298 BUG_ON(!PageLocked(page
));
1300 if (unlikely(PageKsm(page
)))
1301 ret
= try_to_unmap_ksm(page
, flags
);
1302 else if (PageAnon(page
))
1303 ret
= try_to_unmap_anon(page
, flags
);
1305 ret
= try_to_unmap_file(page
, flags
);
1306 if (ret
!= SWAP_MLOCK
&& !page_mapped(page
))
1312 * try_to_munlock - try to munlock a page
1313 * @page: the page to be munlocked
1315 * Called from munlock code. Checks all of the VMAs mapping the page
1316 * to make sure nobody else has this page mlocked. The page will be
1317 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1319 * Return values are:
1321 * SWAP_AGAIN - no vma is holding page mlocked, or,
1322 * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem
1323 * SWAP_FAIL - page cannot be located at present
1324 * SWAP_MLOCK - page is now mlocked.
1326 int try_to_munlock(struct page
*page
)
1328 VM_BUG_ON(!PageLocked(page
) || PageLRU(page
));
1330 if (unlikely(PageKsm(page
)))
1331 return try_to_unmap_ksm(page
, TTU_MUNLOCK
);
1332 else if (PageAnon(page
))
1333 return try_to_unmap_anon(page
, TTU_MUNLOCK
);
1335 return try_to_unmap_file(page
, TTU_MUNLOCK
);
1338 #ifdef CONFIG_MIGRATION
1340 * rmap_walk() and its helpers rmap_walk_anon() and rmap_walk_file():
1341 * Called by migrate.c to remove migration ptes, but might be used more later.
1343 static int rmap_walk_anon(struct page
*page
, int (*rmap_one
)(struct page
*,
1344 struct vm_area_struct
*, unsigned long, void *), void *arg
)
1346 struct anon_vma
*anon_vma
;
1347 struct anon_vma_chain
*avc
;
1348 int ret
= SWAP_AGAIN
;
1351 * Note: remove_migration_ptes() cannot use page_lock_anon_vma()
1352 * because that depends on page_mapped(); but not all its usages
1353 * are holding mmap_sem, which also gave the necessary guarantee
1354 * (that this anon_vma's slab has not already been destroyed).
1355 * This needs to be reviewed later: avoiding page_lock_anon_vma()
1356 * is risky, and currently limits the usefulness of rmap_walk().
1358 anon_vma
= page_anon_vma(page
);
1361 spin_lock(&anon_vma
->lock
);
1362 list_for_each_entry(avc
, &anon_vma
->head
, same_anon_vma
) {
1363 struct vm_area_struct
*vma
= avc
->vma
;
1364 unsigned long address
= vma_address(page
, vma
);
1365 if (address
== -EFAULT
)
1367 ret
= rmap_one(page
, vma
, address
, arg
);
1368 if (ret
!= SWAP_AGAIN
)
1371 spin_unlock(&anon_vma
->lock
);
1375 static int rmap_walk_file(struct page
*page
, int (*rmap_one
)(struct page
*,
1376 struct vm_area_struct
*, unsigned long, void *), void *arg
)
1378 struct address_space
*mapping
= page
->mapping
;
1379 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
1380 struct vm_area_struct
*vma
;
1381 struct prio_tree_iter iter
;
1382 int ret
= SWAP_AGAIN
;
1386 spin_lock(&mapping
->i_mmap_lock
);
1387 vma_prio_tree_foreach(vma
, &iter
, &mapping
->i_mmap
, pgoff
, pgoff
) {
1388 unsigned long address
= vma_address(page
, vma
);
1389 if (address
== -EFAULT
)
1391 ret
= rmap_one(page
, vma
, address
, arg
);
1392 if (ret
!= SWAP_AGAIN
)
1396 * No nonlinear handling: being always shared, nonlinear vmas
1397 * never contain migration ptes. Decide what to do about this
1398 * limitation to linear when we need rmap_walk() on nonlinear.
1400 spin_unlock(&mapping
->i_mmap_lock
);
1404 int rmap_walk(struct page
*page
, int (*rmap_one
)(struct page
*,
1405 struct vm_area_struct
*, unsigned long, void *), void *arg
)
1407 VM_BUG_ON(!PageLocked(page
));
1409 if (unlikely(PageKsm(page
)))
1410 return rmap_walk_ksm(page
, rmap_one
, arg
);
1411 else if (PageAnon(page
))
1412 return rmap_walk_anon(page
, rmap_one
, arg
);
1414 return rmap_walk_file(page
, rmap_one
, arg
);
1416 #endif /* CONFIG_MIGRATION */