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_mutex
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->i_lock (in set_page_dirty's __mark_inode_dirty)
35 * inode_wb_list_lock (in set_page_dirty's __mark_inode_dirty)
36 * sb_lock (within inode_lock in fs/fs-writeback.c)
37 * mapping->tree_lock (widely used, in set_page_dirty,
38 * in arch-dependent flush_dcache_mmap_lock,
39 * within inode_wb_list_lock in __sync_single_inode)
41 * (code doesn't rely on that order so it could be switched around)
43 * anon_vma->mutex (memory_failure, collect_procs_anon)
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/ksm.h>
54 #include <linux/rmap.h>
55 #include <linux/rcupdate.h>
56 #include <linux/module.h>
57 #include <linux/memcontrol.h>
58 #include <linux/mmu_notifier.h>
59 #include <linux/migrate.h>
60 #include <linux/hugetlb.h>
62 #include <asm/tlbflush.h>
66 static struct kmem_cache
*anon_vma_cachep
;
67 static struct kmem_cache
*anon_vma_chain_cachep
;
69 static inline struct anon_vma
*anon_vma_alloc(void)
71 struct anon_vma
*anon_vma
;
73 anon_vma
= kmem_cache_alloc(anon_vma_cachep
, GFP_KERNEL
);
75 atomic_set(&anon_vma
->refcount
, 1);
77 * Initialise the anon_vma root to point to itself. If called
78 * from fork, the root will be reset to the parents anon_vma.
80 anon_vma
->root
= anon_vma
;
86 static inline void anon_vma_free(struct anon_vma
*anon_vma
)
88 VM_BUG_ON(atomic_read(&anon_vma
->refcount
));
91 * Synchronize against page_lock_anon_vma() such that
92 * we can safely hold the lock without the anon_vma getting
95 * Relies on the full mb implied by the atomic_dec_and_test() from
96 * put_anon_vma() against the acquire barrier implied by
97 * mutex_trylock() from page_lock_anon_vma(). This orders:
99 * page_lock_anon_vma() VS put_anon_vma()
100 * mutex_trylock() atomic_dec_and_test()
102 * atomic_read() mutex_is_locked()
104 * LOCK should suffice since the actual taking of the lock must
105 * happen _before_ what follows.
107 if (mutex_is_locked(&anon_vma
->root
->mutex
)) {
108 anon_vma_lock(anon_vma
);
109 anon_vma_unlock(anon_vma
);
112 kmem_cache_free(anon_vma_cachep
, anon_vma
);
115 static inline struct anon_vma_chain
*anon_vma_chain_alloc(void)
117 return kmem_cache_alloc(anon_vma_chain_cachep
, GFP_KERNEL
);
120 static void anon_vma_chain_free(struct anon_vma_chain
*anon_vma_chain
)
122 kmem_cache_free(anon_vma_chain_cachep
, anon_vma_chain
);
126 * anon_vma_prepare - attach an anon_vma to a memory region
127 * @vma: the memory region in question
129 * This makes sure the memory mapping described by 'vma' has
130 * an 'anon_vma' attached to it, so that we can associate the
131 * anonymous pages mapped into it with that anon_vma.
133 * The common case will be that we already have one, but if
134 * not we either need to find an adjacent mapping that we
135 * can re-use the anon_vma from (very common when the only
136 * reason for splitting a vma has been mprotect()), or we
137 * allocate a new one.
139 * Anon-vma allocations are very subtle, because we may have
140 * optimistically looked up an anon_vma in page_lock_anon_vma()
141 * and that may actually touch the spinlock even in the newly
142 * allocated vma (it depends on RCU to make sure that the
143 * anon_vma isn't actually destroyed).
145 * As a result, we need to do proper anon_vma locking even
146 * for the new allocation. At the same time, we do not want
147 * to do any locking for the common case of already having
150 * This must be called with the mmap_sem held for reading.
152 int anon_vma_prepare(struct vm_area_struct
*vma
)
154 struct anon_vma
*anon_vma
= vma
->anon_vma
;
155 struct anon_vma_chain
*avc
;
158 if (unlikely(!anon_vma
)) {
159 struct mm_struct
*mm
= vma
->vm_mm
;
160 struct anon_vma
*allocated
;
162 avc
= anon_vma_chain_alloc();
166 anon_vma
= find_mergeable_anon_vma(vma
);
169 anon_vma
= anon_vma_alloc();
170 if (unlikely(!anon_vma
))
171 goto out_enomem_free_avc
;
172 allocated
= anon_vma
;
175 anon_vma_lock(anon_vma
);
176 /* page_table_lock to protect against threads */
177 spin_lock(&mm
->page_table_lock
);
178 if (likely(!vma
->anon_vma
)) {
179 vma
->anon_vma
= anon_vma
;
180 avc
->anon_vma
= anon_vma
;
182 list_add(&avc
->same_vma
, &vma
->anon_vma_chain
);
183 list_add_tail(&avc
->same_anon_vma
, &anon_vma
->head
);
187 spin_unlock(&mm
->page_table_lock
);
188 anon_vma_unlock(anon_vma
);
190 if (unlikely(allocated
))
191 put_anon_vma(allocated
);
193 anon_vma_chain_free(avc
);
198 anon_vma_chain_free(avc
);
203 static void anon_vma_chain_link(struct vm_area_struct
*vma
,
204 struct anon_vma_chain
*avc
,
205 struct anon_vma
*anon_vma
)
208 avc
->anon_vma
= anon_vma
;
209 list_add(&avc
->same_vma
, &vma
->anon_vma_chain
);
211 anon_vma_lock(anon_vma
);
213 * It's critical to add new vmas to the tail of the anon_vma,
214 * see comment in huge_memory.c:__split_huge_page().
216 list_add_tail(&avc
->same_anon_vma
, &anon_vma
->head
);
217 anon_vma_unlock(anon_vma
);
221 * Attach the anon_vmas from src to dst.
222 * Returns 0 on success, -ENOMEM on failure.
224 int anon_vma_clone(struct vm_area_struct
*dst
, struct vm_area_struct
*src
)
226 struct anon_vma_chain
*avc
, *pavc
;
228 list_for_each_entry_reverse(pavc
, &src
->anon_vma_chain
, same_vma
) {
229 avc
= anon_vma_chain_alloc();
232 anon_vma_chain_link(dst
, avc
, pavc
->anon_vma
);
237 unlink_anon_vmas(dst
);
242 * Attach vma to its own anon_vma, as well as to the anon_vmas that
243 * the corresponding VMA in the parent process is attached to.
244 * Returns 0 on success, non-zero on failure.
246 int anon_vma_fork(struct vm_area_struct
*vma
, struct vm_area_struct
*pvma
)
248 struct anon_vma_chain
*avc
;
249 struct anon_vma
*anon_vma
;
251 /* Don't bother if the parent process has no anon_vma here. */
256 * First, attach the new VMA to the parent VMA's anon_vmas,
257 * so rmap can find non-COWed pages in child processes.
259 if (anon_vma_clone(vma
, pvma
))
262 /* Then add our own anon_vma. */
263 anon_vma
= anon_vma_alloc();
266 avc
= anon_vma_chain_alloc();
268 goto out_error_free_anon_vma
;
271 * The root anon_vma's spinlock is the lock actually used when we
272 * lock any of the anon_vmas in this anon_vma tree.
274 anon_vma
->root
= pvma
->anon_vma
->root
;
276 * With refcounts, an anon_vma can stay around longer than the
277 * process it belongs to. The root anon_vma needs to be pinned until
278 * this anon_vma is freed, because the lock lives in the root.
280 get_anon_vma(anon_vma
->root
);
281 /* Mark this anon_vma as the one where our new (COWed) pages go. */
282 vma
->anon_vma
= anon_vma
;
283 anon_vma_chain_link(vma
, avc
, anon_vma
);
287 out_error_free_anon_vma
:
288 put_anon_vma(anon_vma
);
290 unlink_anon_vmas(vma
);
294 static void anon_vma_unlink(struct anon_vma_chain
*anon_vma_chain
)
296 struct anon_vma
*anon_vma
= anon_vma_chain
->anon_vma
;
299 /* If anon_vma_fork fails, we can get an empty anon_vma_chain. */
303 anon_vma_lock(anon_vma
);
304 list_del(&anon_vma_chain
->same_anon_vma
);
306 /* We must garbage collect the anon_vma if it's empty */
307 empty
= list_empty(&anon_vma
->head
);
308 anon_vma_unlock(anon_vma
);
311 put_anon_vma(anon_vma
);
314 void unlink_anon_vmas(struct vm_area_struct
*vma
)
316 struct anon_vma_chain
*avc
, *next
;
319 * Unlink each anon_vma chained to the VMA. This list is ordered
320 * from newest to oldest, ensuring the root anon_vma gets freed last.
322 list_for_each_entry_safe(avc
, next
, &vma
->anon_vma_chain
, same_vma
) {
323 anon_vma_unlink(avc
);
324 list_del(&avc
->same_vma
);
325 anon_vma_chain_free(avc
);
329 static void anon_vma_ctor(void *data
)
331 struct anon_vma
*anon_vma
= data
;
333 mutex_init(&anon_vma
->mutex
);
334 atomic_set(&anon_vma
->refcount
, 0);
335 INIT_LIST_HEAD(&anon_vma
->head
);
338 void __init
anon_vma_init(void)
340 anon_vma_cachep
= kmem_cache_create("anon_vma", sizeof(struct anon_vma
),
341 0, SLAB_DESTROY_BY_RCU
|SLAB_PANIC
, anon_vma_ctor
);
342 anon_vma_chain_cachep
= KMEM_CACHE(anon_vma_chain
, SLAB_PANIC
);
346 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
348 * Since there is no serialization what so ever against page_remove_rmap()
349 * the best this function can do is return a locked anon_vma that might
350 * have been relevant to this page.
352 * The page might have been remapped to a different anon_vma or the anon_vma
353 * returned may already be freed (and even reused).
355 * All users of this function must be very careful when walking the anon_vma
356 * chain and verify that the page in question is indeed mapped in it
357 * [ something equivalent to page_mapped_in_vma() ].
359 * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap()
360 * that the anon_vma pointer from page->mapping is valid if there is a
361 * mapcount, we can dereference the anon_vma after observing those.
363 struct anon_vma
*page_get_anon_vma(struct page
*page
)
365 struct anon_vma
*anon_vma
= NULL
;
366 unsigned long anon_mapping
;
369 anon_mapping
= (unsigned long) ACCESS_ONCE(page
->mapping
);
370 if ((anon_mapping
& PAGE_MAPPING_FLAGS
) != PAGE_MAPPING_ANON
)
372 if (!page_mapped(page
))
375 anon_vma
= (struct anon_vma
*) (anon_mapping
- PAGE_MAPPING_ANON
);
376 if (!atomic_inc_not_zero(&anon_vma
->refcount
)) {
382 * If this page is still mapped, then its anon_vma cannot have been
383 * freed. But if it has been unmapped, we have no security against the
384 * anon_vma structure being freed and reused (for another anon_vma:
385 * SLAB_DESTROY_BY_RCU guarantees that - so the atomic_inc_not_zero()
386 * above cannot corrupt).
388 if (!page_mapped(page
)) {
389 put_anon_vma(anon_vma
);
399 * Similar to page_get_anon_vma() except it locks the anon_vma.
401 * Its a little more complex as it tries to keep the fast path to a single
402 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
403 * reference like with page_get_anon_vma() and then block on the mutex.
405 struct anon_vma
*page_lock_anon_vma(struct page
*page
)
407 struct anon_vma
*anon_vma
= NULL
;
408 struct anon_vma
*root_anon_vma
;
409 unsigned long anon_mapping
;
412 anon_mapping
= (unsigned long) ACCESS_ONCE(page
->mapping
);
413 if ((anon_mapping
& PAGE_MAPPING_FLAGS
) != PAGE_MAPPING_ANON
)
415 if (!page_mapped(page
))
418 anon_vma
= (struct anon_vma
*) (anon_mapping
- PAGE_MAPPING_ANON
);
419 root_anon_vma
= ACCESS_ONCE(anon_vma
->root
);
420 if (mutex_trylock(&root_anon_vma
->mutex
)) {
422 * If the page is still mapped, then this anon_vma is still
423 * its anon_vma, and holding the mutex ensures that it will
424 * not go away, see __put_anon_vma().
426 if (!page_mapped(page
)) {
427 mutex_unlock(&root_anon_vma
->mutex
);
433 /* trylock failed, we got to sleep */
434 if (!atomic_inc_not_zero(&anon_vma
->refcount
)) {
439 if (!page_mapped(page
)) {
440 put_anon_vma(anon_vma
);
445 /* we pinned the anon_vma, its safe to sleep */
447 anon_vma_lock(anon_vma
);
449 if (atomic_dec_and_test(&anon_vma
->refcount
)) {
451 * Oops, we held the last refcount, release the lock
452 * and bail -- can't simply use put_anon_vma() because
453 * we'll deadlock on the anon_vma_lock() recursion.
455 anon_vma_unlock(anon_vma
);
456 __put_anon_vma(anon_vma
);
467 void page_unlock_anon_vma(struct anon_vma
*anon_vma
)
469 anon_vma_unlock(anon_vma
);
473 * At what user virtual address is page expected in @vma?
474 * Returns virtual address or -EFAULT if page's index/offset is not
475 * within the range mapped the @vma.
478 vma_address(struct page
*page
, struct vm_area_struct
*vma
)
480 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
481 unsigned long address
;
483 if (unlikely(is_vm_hugetlb_page(vma
)))
484 pgoff
= page
->index
<< huge_page_order(page_hstate(page
));
485 address
= vma
->vm_start
+ ((pgoff
- vma
->vm_pgoff
) << PAGE_SHIFT
);
486 if (unlikely(address
< vma
->vm_start
|| address
>= vma
->vm_end
)) {
487 /* page should be within @vma mapping range */
494 * At what user virtual address is page expected in vma?
495 * Caller should check the page is actually part of the vma.
497 unsigned long page_address_in_vma(struct page
*page
, struct vm_area_struct
*vma
)
499 if (PageAnon(page
)) {
500 struct anon_vma
*page__anon_vma
= page_anon_vma(page
);
502 * Note: swapoff's unuse_vma() is more efficient with this
503 * check, and needs it to match anon_vma when KSM is active.
505 if (!vma
->anon_vma
|| !page__anon_vma
||
506 vma
->anon_vma
->root
!= page__anon_vma
->root
)
508 } else if (page
->mapping
&& !(vma
->vm_flags
& VM_NONLINEAR
)) {
510 vma
->vm_file
->f_mapping
!= page
->mapping
)
514 return vma_address(page
, vma
);
518 * Check that @page is mapped at @address into @mm.
520 * If @sync is false, page_check_address may perform a racy check to avoid
521 * the page table lock when the pte is not present (helpful when reclaiming
522 * highly shared pages).
524 * On success returns with pte mapped and locked.
526 pte_t
*__page_check_address(struct page
*page
, struct mm_struct
*mm
,
527 unsigned long address
, spinlock_t
**ptlp
, int sync
)
535 if (unlikely(PageHuge(page
))) {
536 pte
= huge_pte_offset(mm
, address
);
537 ptl
= &mm
->page_table_lock
;
541 pgd
= pgd_offset(mm
, address
);
542 if (!pgd_present(*pgd
))
545 pud
= pud_offset(pgd
, address
);
546 if (!pud_present(*pud
))
549 pmd
= pmd_offset(pud
, address
);
550 if (!pmd_present(*pmd
))
552 if (pmd_trans_huge(*pmd
))
555 pte
= pte_offset_map(pmd
, address
);
556 /* Make a quick check before getting the lock */
557 if (!sync
&& !pte_present(*pte
)) {
562 ptl
= pte_lockptr(mm
, pmd
);
565 if (pte_present(*pte
) && page_to_pfn(page
) == pte_pfn(*pte
)) {
569 pte_unmap_unlock(pte
, ptl
);
574 * page_mapped_in_vma - check whether a page is really mapped in a VMA
575 * @page: the page to test
576 * @vma: the VMA to test
578 * Returns 1 if the page is mapped into the page tables of the VMA, 0
579 * if the page is not mapped into the page tables of this VMA. Only
580 * valid for normal file or anonymous VMAs.
582 int page_mapped_in_vma(struct page
*page
, struct vm_area_struct
*vma
)
584 unsigned long address
;
588 address
= vma_address(page
, vma
);
589 if (address
== -EFAULT
) /* out of vma range */
591 pte
= page_check_address(page
, vma
->vm_mm
, address
, &ptl
, 1);
592 if (!pte
) /* the page is not in this mm */
594 pte_unmap_unlock(pte
, ptl
);
600 * Subfunctions of page_referenced: page_referenced_one called
601 * repeatedly from either page_referenced_anon or page_referenced_file.
603 int page_referenced_one(struct page
*page
, struct vm_area_struct
*vma
,
604 unsigned long address
, unsigned int *mapcount
,
605 unsigned long *vm_flags
)
607 struct mm_struct
*mm
= vma
->vm_mm
;
610 if (unlikely(PageTransHuge(page
))) {
613 spin_lock(&mm
->page_table_lock
);
615 * rmap might return false positives; we must filter
616 * these out using page_check_address_pmd().
618 pmd
= page_check_address_pmd(page
, mm
, address
,
619 PAGE_CHECK_ADDRESS_PMD_FLAG
);
621 spin_unlock(&mm
->page_table_lock
);
625 if (vma
->vm_flags
& VM_LOCKED
) {
626 spin_unlock(&mm
->page_table_lock
);
627 *mapcount
= 0; /* break early from loop */
628 *vm_flags
|= VM_LOCKED
;
632 /* go ahead even if the pmd is pmd_trans_splitting() */
633 if (pmdp_clear_flush_young_notify(vma
, address
, pmd
))
635 spin_unlock(&mm
->page_table_lock
);
641 * rmap might return false positives; we must filter
642 * these out using page_check_address().
644 pte
= page_check_address(page
, mm
, address
, &ptl
, 0);
648 if (vma
->vm_flags
& VM_LOCKED
) {
649 pte_unmap_unlock(pte
, ptl
);
650 *mapcount
= 0; /* break early from loop */
651 *vm_flags
|= VM_LOCKED
;
655 if (ptep_clear_flush_young_notify(vma
, address
, pte
)) {
657 * Don't treat a reference through a sequentially read
658 * mapping as such. If the page has been used in
659 * another mapping, we will catch it; if this other
660 * mapping is already gone, the unmap path will have
661 * set PG_referenced or activated the page.
663 if (likely(!VM_SequentialReadHint(vma
)))
666 pte_unmap_unlock(pte
, ptl
);
669 /* Pretend the page is referenced if the task has the
670 swap token and is in the middle of a page fault. */
671 if (mm
!= current
->mm
&& has_swap_token(mm
) &&
672 rwsem_is_locked(&mm
->mmap_sem
))
678 *vm_flags
|= vma
->vm_flags
;
683 static int page_referenced_anon(struct page
*page
,
684 struct mem_cgroup
*mem_cont
,
685 unsigned long *vm_flags
)
687 unsigned int mapcount
;
688 struct anon_vma
*anon_vma
;
689 struct anon_vma_chain
*avc
;
692 anon_vma
= page_lock_anon_vma(page
);
696 mapcount
= page_mapcount(page
);
697 list_for_each_entry(avc
, &anon_vma
->head
, same_anon_vma
) {
698 struct vm_area_struct
*vma
= avc
->vma
;
699 unsigned long address
= vma_address(page
, vma
);
700 if (address
== -EFAULT
)
703 * If we are reclaiming on behalf of a cgroup, skip
704 * counting on behalf of references from different
707 if (mem_cont
&& !mm_match_cgroup(vma
->vm_mm
, mem_cont
))
709 referenced
+= page_referenced_one(page
, vma
, address
,
710 &mapcount
, vm_flags
);
715 page_unlock_anon_vma(anon_vma
);
720 * page_referenced_file - referenced check for object-based rmap
721 * @page: the page we're checking references on.
722 * @mem_cont: target memory controller
723 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
725 * For an object-based mapped page, find all the places it is mapped and
726 * check/clear the referenced flag. This is done by following the page->mapping
727 * pointer, then walking the chain of vmas it holds. It returns the number
728 * of references it found.
730 * This function is only called from page_referenced for object-based pages.
732 static int page_referenced_file(struct page
*page
,
733 struct mem_cgroup
*mem_cont
,
734 unsigned long *vm_flags
)
736 unsigned int mapcount
;
737 struct address_space
*mapping
= page
->mapping
;
738 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
739 struct vm_area_struct
*vma
;
740 struct prio_tree_iter iter
;
744 * The caller's checks on page->mapping and !PageAnon have made
745 * sure that this is a file page: the check for page->mapping
746 * excludes the case just before it gets set on an anon page.
748 BUG_ON(PageAnon(page
));
751 * The page lock not only makes sure that page->mapping cannot
752 * suddenly be NULLified by truncation, it makes sure that the
753 * structure at mapping cannot be freed and reused yet,
754 * so we can safely take mapping->i_mmap_mutex.
756 BUG_ON(!PageLocked(page
));
758 mutex_lock(&mapping
->i_mmap_mutex
);
761 * i_mmap_mutex does not stabilize mapcount at all, but mapcount
762 * is more likely to be accurate if we note it after spinning.
764 mapcount
= page_mapcount(page
);
766 vma_prio_tree_foreach(vma
, &iter
, &mapping
->i_mmap
, pgoff
, pgoff
) {
767 unsigned long address
= vma_address(page
, vma
);
768 if (address
== -EFAULT
)
771 * If we are reclaiming on behalf of a cgroup, skip
772 * counting on behalf of references from different
775 if (mem_cont
&& !mm_match_cgroup(vma
->vm_mm
, mem_cont
))
777 referenced
+= page_referenced_one(page
, vma
, address
,
778 &mapcount
, vm_flags
);
783 mutex_unlock(&mapping
->i_mmap_mutex
);
788 * page_referenced - test if the page was referenced
789 * @page: the page to test
790 * @is_locked: caller holds lock on the page
791 * @mem_cont: target memory controller
792 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
794 * Quick test_and_clear_referenced for all mappings to a page,
795 * returns the number of ptes which referenced the page.
797 int page_referenced(struct page
*page
,
799 struct mem_cgroup
*mem_cont
,
800 unsigned long *vm_flags
)
806 if (page_mapped(page
) && page_rmapping(page
)) {
807 if (!is_locked
&& (!PageAnon(page
) || PageKsm(page
))) {
808 we_locked
= trylock_page(page
);
814 if (unlikely(PageKsm(page
)))
815 referenced
+= page_referenced_ksm(page
, mem_cont
,
817 else if (PageAnon(page
))
818 referenced
+= page_referenced_anon(page
, mem_cont
,
820 else if (page
->mapping
)
821 referenced
+= page_referenced_file(page
, mem_cont
,
827 if (page_test_and_clear_young(page_to_pfn(page
)))
833 static int page_mkclean_one(struct page
*page
, struct vm_area_struct
*vma
,
834 unsigned long address
)
836 struct mm_struct
*mm
= vma
->vm_mm
;
841 pte
= page_check_address(page
, mm
, address
, &ptl
, 1);
845 if (pte_dirty(*pte
) || pte_write(*pte
)) {
848 flush_cache_page(vma
, address
, pte_pfn(*pte
));
849 entry
= ptep_clear_flush_notify(vma
, address
, pte
);
850 entry
= pte_wrprotect(entry
);
851 entry
= pte_mkclean(entry
);
852 set_pte_at(mm
, address
, pte
, entry
);
856 pte_unmap_unlock(pte
, ptl
);
861 static int page_mkclean_file(struct address_space
*mapping
, struct page
*page
)
863 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
864 struct vm_area_struct
*vma
;
865 struct prio_tree_iter iter
;
868 BUG_ON(PageAnon(page
));
870 mutex_lock(&mapping
->i_mmap_mutex
);
871 vma_prio_tree_foreach(vma
, &iter
, &mapping
->i_mmap
, pgoff
, pgoff
) {
872 if (vma
->vm_flags
& VM_SHARED
) {
873 unsigned long address
= vma_address(page
, vma
);
874 if (address
== -EFAULT
)
876 ret
+= page_mkclean_one(page
, vma
, address
);
879 mutex_unlock(&mapping
->i_mmap_mutex
);
883 int page_mkclean(struct page
*page
)
887 BUG_ON(!PageLocked(page
));
889 if (page_mapped(page
)) {
890 struct address_space
*mapping
= page_mapping(page
);
892 ret
= page_mkclean_file(mapping
, page
);
893 if (page_test_and_clear_dirty(page_to_pfn(page
), 1))
900 EXPORT_SYMBOL_GPL(page_mkclean
);
903 * page_move_anon_rmap - move a page to our anon_vma
904 * @page: the page to move to our anon_vma
905 * @vma: the vma the page belongs to
906 * @address: the user virtual address mapped
908 * When a page belongs exclusively to one process after a COW event,
909 * that page can be moved into the anon_vma that belongs to just that
910 * process, so the rmap code will not search the parent or sibling
913 void page_move_anon_rmap(struct page
*page
,
914 struct vm_area_struct
*vma
, unsigned long address
)
916 struct anon_vma
*anon_vma
= vma
->anon_vma
;
918 VM_BUG_ON(!PageLocked(page
));
919 VM_BUG_ON(!anon_vma
);
920 VM_BUG_ON(page
->index
!= linear_page_index(vma
, address
));
922 anon_vma
= (void *) anon_vma
+ PAGE_MAPPING_ANON
;
923 page
->mapping
= (struct address_space
*) anon_vma
;
927 * __page_set_anon_rmap - set up new anonymous rmap
928 * @page: Page to add to rmap
929 * @vma: VM area to add page to.
930 * @address: User virtual address of the mapping
931 * @exclusive: the page is exclusively owned by the current process
933 static void __page_set_anon_rmap(struct page
*page
,
934 struct vm_area_struct
*vma
, unsigned long address
, int exclusive
)
936 struct anon_vma
*anon_vma
= vma
->anon_vma
;
944 * If the page isn't exclusively mapped into this vma,
945 * we must use the _oldest_ possible anon_vma for the
949 anon_vma
= anon_vma
->root
;
951 anon_vma
= (void *) anon_vma
+ PAGE_MAPPING_ANON
;
952 page
->mapping
= (struct address_space
*) anon_vma
;
953 page
->index
= linear_page_index(vma
, address
);
957 * __page_check_anon_rmap - sanity check anonymous rmap addition
958 * @page: the page to add the mapping to
959 * @vma: the vm area in which the mapping is added
960 * @address: the user virtual address mapped
962 static void __page_check_anon_rmap(struct page
*page
,
963 struct vm_area_struct
*vma
, unsigned long address
)
965 #ifdef CONFIG_DEBUG_VM
967 * The page's anon-rmap details (mapping and index) are guaranteed to
968 * be set up correctly at this point.
970 * We have exclusion against page_add_anon_rmap because the caller
971 * always holds the page locked, except if called from page_dup_rmap,
972 * in which case the page is already known to be setup.
974 * We have exclusion against page_add_new_anon_rmap because those pages
975 * are initially only visible via the pagetables, and the pte is locked
976 * over the call to page_add_new_anon_rmap.
978 BUG_ON(page_anon_vma(page
)->root
!= vma
->anon_vma
->root
);
979 BUG_ON(page
->index
!= linear_page_index(vma
, address
));
984 * page_add_anon_rmap - add pte mapping to an anonymous page
985 * @page: the page to add the mapping to
986 * @vma: the vm area in which the mapping is added
987 * @address: the user virtual address mapped
989 * The caller needs to hold the pte lock, and the page must be locked in
990 * the anon_vma case: to serialize mapping,index checking after setting,
991 * and to ensure that PageAnon is not being upgraded racily to PageKsm
992 * (but PageKsm is never downgraded to PageAnon).
994 void page_add_anon_rmap(struct page
*page
,
995 struct vm_area_struct
*vma
, unsigned long address
)
997 do_page_add_anon_rmap(page
, vma
, address
, 0);
1001 * Special version of the above for do_swap_page, which often runs
1002 * into pages that are exclusively owned by the current process.
1003 * Everybody else should continue to use page_add_anon_rmap above.
1005 void do_page_add_anon_rmap(struct page
*page
,
1006 struct vm_area_struct
*vma
, unsigned long address
, int exclusive
)
1008 int first
= atomic_inc_and_test(&page
->_mapcount
);
1010 if (!PageTransHuge(page
))
1011 __inc_zone_page_state(page
, NR_ANON_PAGES
);
1013 __inc_zone_page_state(page
,
1014 NR_ANON_TRANSPARENT_HUGEPAGES
);
1016 if (unlikely(PageKsm(page
)))
1019 VM_BUG_ON(!PageLocked(page
));
1020 /* address might be in next vma when migration races vma_adjust */
1022 __page_set_anon_rmap(page
, vma
, address
, exclusive
);
1024 __page_check_anon_rmap(page
, vma
, address
);
1028 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1029 * @page: the page to add the mapping to
1030 * @vma: the vm area in which the mapping is added
1031 * @address: the user virtual address mapped
1033 * Same as page_add_anon_rmap but must only be called on *new* pages.
1034 * This means the inc-and-test can be bypassed.
1035 * Page does not have to be locked.
1037 void page_add_new_anon_rmap(struct page
*page
,
1038 struct vm_area_struct
*vma
, unsigned long address
)
1040 VM_BUG_ON(address
< vma
->vm_start
|| address
>= vma
->vm_end
);
1041 SetPageSwapBacked(page
);
1042 atomic_set(&page
->_mapcount
, 0); /* increment count (starts at -1) */
1043 if (!PageTransHuge(page
))
1044 __inc_zone_page_state(page
, NR_ANON_PAGES
);
1046 __inc_zone_page_state(page
, NR_ANON_TRANSPARENT_HUGEPAGES
);
1047 __page_set_anon_rmap(page
, vma
, address
, 1);
1048 if (page_evictable(page
, vma
))
1049 lru_cache_add_lru(page
, LRU_ACTIVE_ANON
);
1051 add_page_to_unevictable_list(page
);
1055 * page_add_file_rmap - add pte mapping to a file page
1056 * @page: the page to add the mapping to
1058 * The caller needs to hold the pte lock.
1060 void page_add_file_rmap(struct page
*page
)
1062 if (atomic_inc_and_test(&page
->_mapcount
)) {
1063 __inc_zone_page_state(page
, NR_FILE_MAPPED
);
1064 mem_cgroup_inc_page_stat(page
, MEMCG_NR_FILE_MAPPED
);
1069 * page_remove_rmap - take down pte mapping from a page
1070 * @page: page to remove mapping from
1072 * The caller needs to hold the pte lock.
1074 void page_remove_rmap(struct page
*page
)
1076 /* page still mapped by someone else? */
1077 if (!atomic_add_negative(-1, &page
->_mapcount
))
1081 * Now that the last pte has gone, s390 must transfer dirty
1082 * flag from storage key to struct page. We can usually skip
1083 * this if the page is anon, so about to be freed; but perhaps
1084 * not if it's in swapcache - there might be another pte slot
1085 * containing the swap entry, but page not yet written to swap.
1087 if ((!PageAnon(page
) || PageSwapCache(page
)) &&
1088 page_test_and_clear_dirty(page_to_pfn(page
), 1))
1089 set_page_dirty(page
);
1091 * Hugepages are not counted in NR_ANON_PAGES nor NR_FILE_MAPPED
1092 * and not charged by memcg for now.
1094 if (unlikely(PageHuge(page
)))
1096 if (PageAnon(page
)) {
1097 mem_cgroup_uncharge_page(page
);
1098 if (!PageTransHuge(page
))
1099 __dec_zone_page_state(page
, NR_ANON_PAGES
);
1101 __dec_zone_page_state(page
,
1102 NR_ANON_TRANSPARENT_HUGEPAGES
);
1104 __dec_zone_page_state(page
, NR_FILE_MAPPED
);
1105 mem_cgroup_dec_page_stat(page
, MEMCG_NR_FILE_MAPPED
);
1108 * It would be tidy to reset the PageAnon mapping here,
1109 * but that might overwrite a racing page_add_anon_rmap
1110 * which increments mapcount after us but sets mapping
1111 * before us: so leave the reset to free_hot_cold_page,
1112 * and remember that it's only reliable while mapped.
1113 * Leaving it set also helps swapoff to reinstate ptes
1114 * faster for those pages still in swapcache.
1119 * Subfunctions of try_to_unmap: try_to_unmap_one called
1120 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
1122 int try_to_unmap_one(struct page
*page
, struct vm_area_struct
*vma
,
1123 unsigned long address
, enum ttu_flags flags
)
1125 struct mm_struct
*mm
= vma
->vm_mm
;
1129 int ret
= SWAP_AGAIN
;
1131 pte
= page_check_address(page
, mm
, address
, &ptl
, 0);
1136 * If the page is mlock()d, we cannot swap it out.
1137 * If it's recently referenced (perhaps page_referenced
1138 * skipped over this mm) then we should reactivate it.
1140 if (!(flags
& TTU_IGNORE_MLOCK
)) {
1141 if (vma
->vm_flags
& VM_LOCKED
)
1144 if (TTU_ACTION(flags
) == TTU_MUNLOCK
)
1147 if (!(flags
& TTU_IGNORE_ACCESS
)) {
1148 if (ptep_clear_flush_young_notify(vma
, address
, pte
)) {
1154 /* Nuke the page table entry. */
1155 flush_cache_page(vma
, address
, page_to_pfn(page
));
1156 pteval
= ptep_clear_flush_notify(vma
, address
, pte
);
1158 /* Move the dirty bit to the physical page now the pte is gone. */
1159 if (pte_dirty(pteval
))
1160 set_page_dirty(page
);
1162 /* Update high watermark before we lower rss */
1163 update_hiwater_rss(mm
);
1165 if (PageHWPoison(page
) && !(flags
& TTU_IGNORE_HWPOISON
)) {
1167 dec_mm_counter(mm
, MM_ANONPAGES
);
1169 dec_mm_counter(mm
, MM_FILEPAGES
);
1170 set_pte_at(mm
, address
, pte
,
1171 swp_entry_to_pte(make_hwpoison_entry(page
)));
1172 } else if (PageAnon(page
)) {
1173 swp_entry_t entry
= { .val
= page_private(page
) };
1175 if (PageSwapCache(page
)) {
1177 * Store the swap location in the pte.
1178 * See handle_pte_fault() ...
1180 if (swap_duplicate(entry
) < 0) {
1181 set_pte_at(mm
, address
, pte
, pteval
);
1185 if (list_empty(&mm
->mmlist
)) {
1186 spin_lock(&mmlist_lock
);
1187 if (list_empty(&mm
->mmlist
))
1188 list_add(&mm
->mmlist
, &init_mm
.mmlist
);
1189 spin_unlock(&mmlist_lock
);
1191 dec_mm_counter(mm
, MM_ANONPAGES
);
1192 inc_mm_counter(mm
, MM_SWAPENTS
);
1193 } else if (PAGE_MIGRATION
) {
1195 * Store the pfn of the page in a special migration
1196 * pte. do_swap_page() will wait until the migration
1197 * pte is removed and then restart fault handling.
1199 BUG_ON(TTU_ACTION(flags
) != TTU_MIGRATION
);
1200 entry
= make_migration_entry(page
, pte_write(pteval
));
1202 set_pte_at(mm
, address
, pte
, swp_entry_to_pte(entry
));
1203 BUG_ON(pte_file(*pte
));
1204 } else if (PAGE_MIGRATION
&& (TTU_ACTION(flags
) == TTU_MIGRATION
)) {
1205 /* Establish migration entry for a file page */
1207 entry
= make_migration_entry(page
, pte_write(pteval
));
1208 set_pte_at(mm
, address
, pte
, swp_entry_to_pte(entry
));
1210 dec_mm_counter(mm
, MM_FILEPAGES
);
1212 page_remove_rmap(page
);
1213 page_cache_release(page
);
1216 pte_unmap_unlock(pte
, ptl
);
1221 pte_unmap_unlock(pte
, ptl
);
1225 * We need mmap_sem locking, Otherwise VM_LOCKED check makes
1226 * unstable result and race. Plus, We can't wait here because
1227 * we now hold anon_vma->mutex or mapping->i_mmap_mutex.
1228 * if trylock failed, the page remain in evictable lru and later
1229 * vmscan could retry to move the page to unevictable lru if the
1230 * page is actually mlocked.
1232 if (down_read_trylock(&vma
->vm_mm
->mmap_sem
)) {
1233 if (vma
->vm_flags
& VM_LOCKED
) {
1234 mlock_vma_page(page
);
1237 up_read(&vma
->vm_mm
->mmap_sem
);
1243 * objrmap doesn't work for nonlinear VMAs because the assumption that
1244 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
1245 * Consequently, given a particular page and its ->index, we cannot locate the
1246 * ptes which are mapping that page without an exhaustive linear search.
1248 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
1249 * maps the file to which the target page belongs. The ->vm_private_data field
1250 * holds the current cursor into that scan. Successive searches will circulate
1251 * around the vma's virtual address space.
1253 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
1254 * more scanning pressure is placed against them as well. Eventually pages
1255 * will become fully unmapped and are eligible for eviction.
1257 * For very sparsely populated VMAs this is a little inefficient - chances are
1258 * there there won't be many ptes located within the scan cluster. In this case
1259 * maybe we could scan further - to the end of the pte page, perhaps.
1261 * Mlocked pages: check VM_LOCKED under mmap_sem held for read, if we can
1262 * acquire it without blocking. If vma locked, mlock the pages in the cluster,
1263 * rather than unmapping them. If we encounter the "check_page" that vmscan is
1264 * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN.
1266 #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
1267 #define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
1269 static int try_to_unmap_cluster(unsigned long cursor
, unsigned int *mapcount
,
1270 struct vm_area_struct
*vma
, struct page
*check_page
)
1272 struct mm_struct
*mm
= vma
->vm_mm
;
1280 unsigned long address
;
1282 int ret
= SWAP_AGAIN
;
1285 address
= (vma
->vm_start
+ cursor
) & CLUSTER_MASK
;
1286 end
= address
+ CLUSTER_SIZE
;
1287 if (address
< vma
->vm_start
)
1288 address
= vma
->vm_start
;
1289 if (end
> vma
->vm_end
)
1292 pgd
= pgd_offset(mm
, address
);
1293 if (!pgd_present(*pgd
))
1296 pud
= pud_offset(pgd
, address
);
1297 if (!pud_present(*pud
))
1300 pmd
= pmd_offset(pud
, address
);
1301 if (!pmd_present(*pmd
))
1305 * If we can acquire the mmap_sem for read, and vma is VM_LOCKED,
1306 * keep the sem while scanning the cluster for mlocking pages.
1308 if (down_read_trylock(&vma
->vm_mm
->mmap_sem
)) {
1309 locked_vma
= (vma
->vm_flags
& VM_LOCKED
);
1311 up_read(&vma
->vm_mm
->mmap_sem
); /* don't need it */
1314 pte
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
1316 /* Update high watermark before we lower rss */
1317 update_hiwater_rss(mm
);
1319 for (; address
< end
; pte
++, address
+= PAGE_SIZE
) {
1320 if (!pte_present(*pte
))
1322 page
= vm_normal_page(vma
, address
, *pte
);
1323 BUG_ON(!page
|| PageAnon(page
));
1326 mlock_vma_page(page
); /* no-op if already mlocked */
1327 if (page
== check_page
)
1329 continue; /* don't unmap */
1332 if (ptep_clear_flush_young_notify(vma
, address
, pte
))
1335 /* Nuke the page table entry. */
1336 flush_cache_page(vma
, address
, pte_pfn(*pte
));
1337 pteval
= ptep_clear_flush_notify(vma
, address
, pte
);
1339 /* If nonlinear, store the file page offset in the pte. */
1340 if (page
->index
!= linear_page_index(vma
, address
))
1341 set_pte_at(mm
, address
, pte
, pgoff_to_pte(page
->index
));
1343 /* Move the dirty bit to the physical page now the pte is gone. */
1344 if (pte_dirty(pteval
))
1345 set_page_dirty(page
);
1347 page_remove_rmap(page
);
1348 page_cache_release(page
);
1349 dec_mm_counter(mm
, MM_FILEPAGES
);
1352 pte_unmap_unlock(pte
- 1, ptl
);
1354 up_read(&vma
->vm_mm
->mmap_sem
);
1358 bool is_vma_temporary_stack(struct vm_area_struct
*vma
)
1360 int maybe_stack
= vma
->vm_flags
& (VM_GROWSDOWN
| VM_GROWSUP
);
1365 if ((vma
->vm_flags
& VM_STACK_INCOMPLETE_SETUP
) ==
1366 VM_STACK_INCOMPLETE_SETUP
)
1373 * try_to_unmap_anon - unmap or unlock anonymous page using the object-based
1375 * @page: the page to unmap/unlock
1376 * @flags: action and flags
1378 * Find all the mappings of a page using the mapping pointer and the vma chains
1379 * contained in the anon_vma struct it points to.
1381 * This function is only called from try_to_unmap/try_to_munlock for
1383 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1384 * where the page was found will be held for write. So, we won't recheck
1385 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1388 static int try_to_unmap_anon(struct page
*page
, enum ttu_flags flags
)
1390 struct anon_vma
*anon_vma
;
1391 struct anon_vma_chain
*avc
;
1392 int ret
= SWAP_AGAIN
;
1394 anon_vma
= page_lock_anon_vma(page
);
1398 list_for_each_entry(avc
, &anon_vma
->head
, same_anon_vma
) {
1399 struct vm_area_struct
*vma
= avc
->vma
;
1400 unsigned long address
;
1403 * During exec, a temporary VMA is setup and later moved.
1404 * The VMA is moved under the anon_vma lock but not the
1405 * page tables leading to a race where migration cannot
1406 * find the migration ptes. Rather than increasing the
1407 * locking requirements of exec(), migration skips
1408 * temporary VMAs until after exec() completes.
1410 if (PAGE_MIGRATION
&& (flags
& TTU_MIGRATION
) &&
1411 is_vma_temporary_stack(vma
))
1414 address
= vma_address(page
, vma
);
1415 if (address
== -EFAULT
)
1417 ret
= try_to_unmap_one(page
, vma
, address
, flags
);
1418 if (ret
!= SWAP_AGAIN
|| !page_mapped(page
))
1422 page_unlock_anon_vma(anon_vma
);
1427 * try_to_unmap_file - unmap/unlock file page using the object-based rmap method
1428 * @page: the page to unmap/unlock
1429 * @flags: action and flags
1431 * Find all the mappings of a page using the mapping pointer and the vma chains
1432 * contained in the address_space struct it points to.
1434 * This function is only called from try_to_unmap/try_to_munlock for
1435 * object-based pages.
1436 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1437 * where the page was found will be held for write. So, we won't recheck
1438 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1441 static int try_to_unmap_file(struct page
*page
, enum ttu_flags flags
)
1443 struct address_space
*mapping
= page
->mapping
;
1444 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
1445 struct vm_area_struct
*vma
;
1446 struct prio_tree_iter iter
;
1447 int ret
= SWAP_AGAIN
;
1448 unsigned long cursor
;
1449 unsigned long max_nl_cursor
= 0;
1450 unsigned long max_nl_size
= 0;
1451 unsigned int mapcount
;
1453 mutex_lock(&mapping
->i_mmap_mutex
);
1454 vma_prio_tree_foreach(vma
, &iter
, &mapping
->i_mmap
, pgoff
, pgoff
) {
1455 unsigned long address
= vma_address(page
, vma
);
1456 if (address
== -EFAULT
)
1458 ret
= try_to_unmap_one(page
, vma
, address
, flags
);
1459 if (ret
!= SWAP_AGAIN
|| !page_mapped(page
))
1463 if (list_empty(&mapping
->i_mmap_nonlinear
))
1467 * We don't bother to try to find the munlocked page in nonlinears.
1468 * It's costly. Instead, later, page reclaim logic may call
1469 * try_to_unmap(TTU_MUNLOCK) and recover PG_mlocked lazily.
1471 if (TTU_ACTION(flags
) == TTU_MUNLOCK
)
1474 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
,
1475 shared
.vm_set
.list
) {
1476 cursor
= (unsigned long) vma
->vm_private_data
;
1477 if (cursor
> max_nl_cursor
)
1478 max_nl_cursor
= cursor
;
1479 cursor
= vma
->vm_end
- vma
->vm_start
;
1480 if (cursor
> max_nl_size
)
1481 max_nl_size
= cursor
;
1484 if (max_nl_size
== 0) { /* all nonlinears locked or reserved ? */
1490 * We don't try to search for this page in the nonlinear vmas,
1491 * and page_referenced wouldn't have found it anyway. Instead
1492 * just walk the nonlinear vmas trying to age and unmap some.
1493 * The mapcount of the page we came in with is irrelevant,
1494 * but even so use it as a guide to how hard we should try?
1496 mapcount
= page_mapcount(page
);
1501 max_nl_size
= (max_nl_size
+ CLUSTER_SIZE
- 1) & CLUSTER_MASK
;
1502 if (max_nl_cursor
== 0)
1503 max_nl_cursor
= CLUSTER_SIZE
;
1506 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
,
1507 shared
.vm_set
.list
) {
1508 cursor
= (unsigned long) vma
->vm_private_data
;
1509 while ( cursor
< max_nl_cursor
&&
1510 cursor
< vma
->vm_end
- vma
->vm_start
) {
1511 if (try_to_unmap_cluster(cursor
, &mapcount
,
1512 vma
, page
) == SWAP_MLOCK
)
1514 cursor
+= CLUSTER_SIZE
;
1515 vma
->vm_private_data
= (void *) cursor
;
1516 if ((int)mapcount
<= 0)
1519 vma
->vm_private_data
= (void *) max_nl_cursor
;
1522 max_nl_cursor
+= CLUSTER_SIZE
;
1523 } while (max_nl_cursor
<= max_nl_size
);
1526 * Don't loop forever (perhaps all the remaining pages are
1527 * in locked vmas). Reset cursor on all unreserved nonlinear
1528 * vmas, now forgetting on which ones it had fallen behind.
1530 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
, shared
.vm_set
.list
)
1531 vma
->vm_private_data
= NULL
;
1533 mutex_unlock(&mapping
->i_mmap_mutex
);
1538 * try_to_unmap - try to remove all page table mappings to a page
1539 * @page: the page to get unmapped
1540 * @flags: action and flags
1542 * Tries to remove all the page table entries which are mapping this
1543 * page, used in the pageout path. Caller must hold the page lock.
1544 * Return values are:
1546 * SWAP_SUCCESS - we succeeded in removing all mappings
1547 * SWAP_AGAIN - we missed a mapping, try again later
1548 * SWAP_FAIL - the page is unswappable
1549 * SWAP_MLOCK - page is mlocked.
1551 int try_to_unmap(struct page
*page
, enum ttu_flags flags
)
1555 BUG_ON(!PageLocked(page
));
1556 VM_BUG_ON(!PageHuge(page
) && PageTransHuge(page
));
1558 if (unlikely(PageKsm(page
)))
1559 ret
= try_to_unmap_ksm(page
, flags
);
1560 else if (PageAnon(page
))
1561 ret
= try_to_unmap_anon(page
, flags
);
1563 ret
= try_to_unmap_file(page
, flags
);
1564 if (ret
!= SWAP_MLOCK
&& !page_mapped(page
))
1570 * try_to_munlock - try to munlock a page
1571 * @page: the page to be munlocked
1573 * Called from munlock code. Checks all of the VMAs mapping the page
1574 * to make sure nobody else has this page mlocked. The page will be
1575 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1577 * Return values are:
1579 * SWAP_AGAIN - no vma is holding page mlocked, or,
1580 * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem
1581 * SWAP_FAIL - page cannot be located at present
1582 * SWAP_MLOCK - page is now mlocked.
1584 int try_to_munlock(struct page
*page
)
1586 VM_BUG_ON(!PageLocked(page
) || PageLRU(page
));
1588 if (unlikely(PageKsm(page
)))
1589 return try_to_unmap_ksm(page
, TTU_MUNLOCK
);
1590 else if (PageAnon(page
))
1591 return try_to_unmap_anon(page
, TTU_MUNLOCK
);
1593 return try_to_unmap_file(page
, TTU_MUNLOCK
);
1596 void __put_anon_vma(struct anon_vma
*anon_vma
)
1598 struct anon_vma
*root
= anon_vma
->root
;
1600 if (root
!= anon_vma
&& atomic_dec_and_test(&root
->refcount
))
1601 anon_vma_free(root
);
1603 anon_vma_free(anon_vma
);
1606 #ifdef CONFIG_MIGRATION
1608 * rmap_walk() and its helpers rmap_walk_anon() and rmap_walk_file():
1609 * Called by migrate.c to remove migration ptes, but might be used more later.
1611 static int rmap_walk_anon(struct page
*page
, int (*rmap_one
)(struct page
*,
1612 struct vm_area_struct
*, unsigned long, void *), void *arg
)
1614 struct anon_vma
*anon_vma
;
1615 struct anon_vma_chain
*avc
;
1616 int ret
= SWAP_AGAIN
;
1619 * Note: remove_migration_ptes() cannot use page_lock_anon_vma()
1620 * because that depends on page_mapped(); but not all its usages
1621 * are holding mmap_sem. Users without mmap_sem are required to
1622 * take a reference count to prevent the anon_vma disappearing
1624 anon_vma
= page_anon_vma(page
);
1627 anon_vma_lock(anon_vma
);
1628 list_for_each_entry(avc
, &anon_vma
->head
, same_anon_vma
) {
1629 struct vm_area_struct
*vma
= avc
->vma
;
1630 unsigned long address
= vma_address(page
, vma
);
1631 if (address
== -EFAULT
)
1633 ret
= rmap_one(page
, vma
, address
, arg
);
1634 if (ret
!= SWAP_AGAIN
)
1637 anon_vma_unlock(anon_vma
);
1641 static int rmap_walk_file(struct page
*page
, int (*rmap_one
)(struct page
*,
1642 struct vm_area_struct
*, unsigned long, void *), void *arg
)
1644 struct address_space
*mapping
= page
->mapping
;
1645 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
1646 struct vm_area_struct
*vma
;
1647 struct prio_tree_iter iter
;
1648 int ret
= SWAP_AGAIN
;
1652 mutex_lock(&mapping
->i_mmap_mutex
);
1653 vma_prio_tree_foreach(vma
, &iter
, &mapping
->i_mmap
, pgoff
, pgoff
) {
1654 unsigned long address
= vma_address(page
, vma
);
1655 if (address
== -EFAULT
)
1657 ret
= rmap_one(page
, vma
, address
, arg
);
1658 if (ret
!= SWAP_AGAIN
)
1662 * No nonlinear handling: being always shared, nonlinear vmas
1663 * never contain migration ptes. Decide what to do about this
1664 * limitation to linear when we need rmap_walk() on nonlinear.
1666 mutex_unlock(&mapping
->i_mmap_mutex
);
1670 int rmap_walk(struct page
*page
, int (*rmap_one
)(struct page
*,
1671 struct vm_area_struct
*, unsigned long, void *), void *arg
)
1673 VM_BUG_ON(!PageLocked(page
));
1675 if (unlikely(PageKsm(page
)))
1676 return rmap_walk_ksm(page
, rmap_one
, arg
);
1677 else if (PageAnon(page
))
1678 return rmap_walk_anon(page
, rmap_one
, arg
);
1680 return rmap_walk_file(page
, rmap_one
, arg
);
1682 #endif /* CONFIG_MIGRATION */
1684 #ifdef CONFIG_HUGETLB_PAGE
1686 * The following three functions are for anonymous (private mapped) hugepages.
1687 * Unlike common anonymous pages, anonymous hugepages have no accounting code
1688 * and no lru code, because we handle hugepages differently from common pages.
1690 static void __hugepage_set_anon_rmap(struct page
*page
,
1691 struct vm_area_struct
*vma
, unsigned long address
, int exclusive
)
1693 struct anon_vma
*anon_vma
= vma
->anon_vma
;
1700 anon_vma
= anon_vma
->root
;
1702 anon_vma
= (void *) anon_vma
+ PAGE_MAPPING_ANON
;
1703 page
->mapping
= (struct address_space
*) anon_vma
;
1704 page
->index
= linear_page_index(vma
, address
);
1707 void hugepage_add_anon_rmap(struct page
*page
,
1708 struct vm_area_struct
*vma
, unsigned long address
)
1710 struct anon_vma
*anon_vma
= vma
->anon_vma
;
1713 BUG_ON(!PageLocked(page
));
1715 /* address might be in next vma when migration races vma_adjust */
1716 first
= atomic_inc_and_test(&page
->_mapcount
);
1718 __hugepage_set_anon_rmap(page
, vma
, address
, 0);
1721 void hugepage_add_new_anon_rmap(struct page
*page
,
1722 struct vm_area_struct
*vma
, unsigned long address
)
1724 BUG_ON(address
< vma
->vm_start
|| address
>= vma
->vm_end
);
1725 atomic_set(&page
->_mapcount
, 0);
1726 __hugepage_set_anon_rmap(page
, vma
, address
, 1);
1728 #endif /* CONFIG_HUGETLB_PAGE */