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
);
204 * This is a useful helper function for locking the anon_vma root as
205 * we traverse the vma->anon_vma_chain, looping over anon_vma's that
208 * Such anon_vma's should have the same root, so you'd expect to see
209 * just a single mutex_lock for the whole traversal.
211 static inline struct anon_vma
*lock_anon_vma_root(struct anon_vma
*root
, struct anon_vma
*anon_vma
)
213 struct anon_vma
*new_root
= anon_vma
->root
;
214 if (new_root
!= root
) {
215 if (WARN_ON_ONCE(root
))
216 mutex_unlock(&root
->mutex
);
218 mutex_lock(&root
->mutex
);
223 static inline void unlock_anon_vma_root(struct anon_vma
*root
)
226 mutex_unlock(&root
->mutex
);
229 static void anon_vma_chain_link(struct vm_area_struct
*vma
,
230 struct anon_vma_chain
*avc
,
231 struct anon_vma
*anon_vma
)
234 avc
->anon_vma
= anon_vma
;
235 list_add(&avc
->same_vma
, &vma
->anon_vma_chain
);
238 * It's critical to add new vmas to the tail of the anon_vma,
239 * see comment in huge_memory.c:__split_huge_page().
241 list_add_tail(&avc
->same_anon_vma
, &anon_vma
->head
);
245 * Attach the anon_vmas from src to dst.
246 * Returns 0 on success, -ENOMEM on failure.
248 int anon_vma_clone(struct vm_area_struct
*dst
, struct vm_area_struct
*src
)
250 struct anon_vma_chain
*avc
, *pavc
;
251 struct anon_vma
*root
= NULL
;
253 list_for_each_entry_reverse(pavc
, &src
->anon_vma_chain
, same_vma
) {
254 struct anon_vma
*anon_vma
;
256 avc
= anon_vma_chain_alloc();
259 anon_vma
= pavc
->anon_vma
;
260 root
= lock_anon_vma_root(root
, anon_vma
);
261 anon_vma_chain_link(dst
, avc
, anon_vma
);
263 unlock_anon_vma_root(root
);
267 unlock_anon_vma_root(root
);
268 unlink_anon_vmas(dst
);
273 * Attach vma to its own anon_vma, as well as to the anon_vmas that
274 * the corresponding VMA in the parent process is attached to.
275 * Returns 0 on success, non-zero on failure.
277 int anon_vma_fork(struct vm_area_struct
*vma
, struct vm_area_struct
*pvma
)
279 struct anon_vma_chain
*avc
;
280 struct anon_vma
*anon_vma
;
282 /* Don't bother if the parent process has no anon_vma here. */
287 * First, attach the new VMA to the parent VMA's anon_vmas,
288 * so rmap can find non-COWed pages in child processes.
290 if (anon_vma_clone(vma
, pvma
))
293 /* Then add our own anon_vma. */
294 anon_vma
= anon_vma_alloc();
297 avc
= anon_vma_chain_alloc();
299 goto out_error_free_anon_vma
;
302 * The root anon_vma's spinlock is the lock actually used when we
303 * lock any of the anon_vmas in this anon_vma tree.
305 anon_vma
->root
= pvma
->anon_vma
->root
;
307 * With refcounts, an anon_vma can stay around longer than the
308 * process it belongs to. The root anon_vma needs to be pinned until
309 * this anon_vma is freed, because the lock lives in the root.
311 get_anon_vma(anon_vma
->root
);
312 /* Mark this anon_vma as the one where our new (COWed) pages go. */
313 vma
->anon_vma
= anon_vma
;
314 anon_vma_lock(anon_vma
);
315 anon_vma_chain_link(vma
, avc
, anon_vma
);
316 anon_vma_unlock(anon_vma
);
320 out_error_free_anon_vma
:
321 put_anon_vma(anon_vma
);
323 unlink_anon_vmas(vma
);
327 static void anon_vma_unlink(struct anon_vma_chain
*anon_vma_chain
)
329 struct anon_vma
*anon_vma
= anon_vma_chain
->anon_vma
;
332 /* If anon_vma_fork fails, we can get an empty anon_vma_chain. */
336 anon_vma_lock(anon_vma
);
337 list_del(&anon_vma_chain
->same_anon_vma
);
339 /* We must garbage collect the anon_vma if it's empty */
340 empty
= list_empty(&anon_vma
->head
);
341 anon_vma_unlock(anon_vma
);
344 put_anon_vma(anon_vma
);
347 void unlink_anon_vmas(struct vm_area_struct
*vma
)
349 struct anon_vma_chain
*avc
, *next
;
352 * Unlink each anon_vma chained to the VMA. This list is ordered
353 * from newest to oldest, ensuring the root anon_vma gets freed last.
355 list_for_each_entry_safe(avc
, next
, &vma
->anon_vma_chain
, same_vma
) {
356 anon_vma_unlink(avc
);
357 list_del(&avc
->same_vma
);
358 anon_vma_chain_free(avc
);
362 static void anon_vma_ctor(void *data
)
364 struct anon_vma
*anon_vma
= data
;
366 mutex_init(&anon_vma
->mutex
);
367 atomic_set(&anon_vma
->refcount
, 0);
368 INIT_LIST_HEAD(&anon_vma
->head
);
371 void __init
anon_vma_init(void)
373 anon_vma_cachep
= kmem_cache_create("anon_vma", sizeof(struct anon_vma
),
374 0, SLAB_DESTROY_BY_RCU
|SLAB_PANIC
, anon_vma_ctor
);
375 anon_vma_chain_cachep
= KMEM_CACHE(anon_vma_chain
, SLAB_PANIC
);
379 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
381 * Since there is no serialization what so ever against page_remove_rmap()
382 * the best this function can do is return a locked anon_vma that might
383 * have been relevant to this page.
385 * The page might have been remapped to a different anon_vma or the anon_vma
386 * returned may already be freed (and even reused).
388 * In case it was remapped to a different anon_vma, the new anon_vma will be a
389 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
390 * ensure that any anon_vma obtained from the page will still be valid for as
391 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
393 * All users of this function must be very careful when walking the anon_vma
394 * chain and verify that the page in question is indeed mapped in it
395 * [ something equivalent to page_mapped_in_vma() ].
397 * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap()
398 * that the anon_vma pointer from page->mapping is valid if there is a
399 * mapcount, we can dereference the anon_vma after observing those.
401 struct anon_vma
*page_get_anon_vma(struct page
*page
)
403 struct anon_vma
*anon_vma
= NULL
;
404 unsigned long anon_mapping
;
407 anon_mapping
= (unsigned long) ACCESS_ONCE(page
->mapping
);
408 if ((anon_mapping
& PAGE_MAPPING_FLAGS
) != PAGE_MAPPING_ANON
)
410 if (!page_mapped(page
))
413 anon_vma
= (struct anon_vma
*) (anon_mapping
- PAGE_MAPPING_ANON
);
414 if (!atomic_inc_not_zero(&anon_vma
->refcount
)) {
420 * If this page is still mapped, then its anon_vma cannot have been
421 * freed. But if it has been unmapped, we have no security against the
422 * anon_vma structure being freed and reused (for another anon_vma:
423 * SLAB_DESTROY_BY_RCU guarantees that - so the atomic_inc_not_zero()
424 * above cannot corrupt).
426 if (!page_mapped(page
)) {
427 put_anon_vma(anon_vma
);
437 * Similar to page_get_anon_vma() except it locks the anon_vma.
439 * Its a little more complex as it tries to keep the fast path to a single
440 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
441 * reference like with page_get_anon_vma() and then block on the mutex.
443 struct anon_vma
*page_lock_anon_vma(struct page
*page
)
445 struct anon_vma
*anon_vma
= NULL
;
446 struct anon_vma
*root_anon_vma
;
447 unsigned long anon_mapping
;
450 anon_mapping
= (unsigned long) ACCESS_ONCE(page
->mapping
);
451 if ((anon_mapping
& PAGE_MAPPING_FLAGS
) != PAGE_MAPPING_ANON
)
453 if (!page_mapped(page
))
456 anon_vma
= (struct anon_vma
*) (anon_mapping
- PAGE_MAPPING_ANON
);
457 root_anon_vma
= ACCESS_ONCE(anon_vma
->root
);
458 if (mutex_trylock(&root_anon_vma
->mutex
)) {
460 * If the page is still mapped, then this anon_vma is still
461 * its anon_vma, and holding the mutex ensures that it will
462 * not go away, see anon_vma_free().
464 if (!page_mapped(page
)) {
465 mutex_unlock(&root_anon_vma
->mutex
);
471 /* trylock failed, we got to sleep */
472 if (!atomic_inc_not_zero(&anon_vma
->refcount
)) {
477 if (!page_mapped(page
)) {
478 put_anon_vma(anon_vma
);
483 /* we pinned the anon_vma, its safe to sleep */
485 anon_vma_lock(anon_vma
);
487 if (atomic_dec_and_test(&anon_vma
->refcount
)) {
489 * Oops, we held the last refcount, release the lock
490 * and bail -- can't simply use put_anon_vma() because
491 * we'll deadlock on the anon_vma_lock() recursion.
493 anon_vma_unlock(anon_vma
);
494 __put_anon_vma(anon_vma
);
505 void page_unlock_anon_vma(struct anon_vma
*anon_vma
)
507 anon_vma_unlock(anon_vma
);
511 * At what user virtual address is page expected in @vma?
512 * Returns virtual address or -EFAULT if page's index/offset is not
513 * within the range mapped the @vma.
516 vma_address(struct page
*page
, struct vm_area_struct
*vma
)
518 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
519 unsigned long address
;
521 if (unlikely(is_vm_hugetlb_page(vma
)))
522 pgoff
= page
->index
<< huge_page_order(page_hstate(page
));
523 address
= vma
->vm_start
+ ((pgoff
- vma
->vm_pgoff
) << PAGE_SHIFT
);
524 if (unlikely(address
< vma
->vm_start
|| address
>= vma
->vm_end
)) {
525 /* page should be within @vma mapping range */
532 * At what user virtual address is page expected in vma?
533 * Caller should check the page is actually part of the vma.
535 unsigned long page_address_in_vma(struct page
*page
, struct vm_area_struct
*vma
)
537 if (PageAnon(page
)) {
538 struct anon_vma
*page__anon_vma
= page_anon_vma(page
);
540 * Note: swapoff's unuse_vma() is more efficient with this
541 * check, and needs it to match anon_vma when KSM is active.
543 if (!vma
->anon_vma
|| !page__anon_vma
||
544 vma
->anon_vma
->root
!= page__anon_vma
->root
)
546 } else if (page
->mapping
&& !(vma
->vm_flags
& VM_NONLINEAR
)) {
548 vma
->vm_file
->f_mapping
!= page
->mapping
)
552 return vma_address(page
, vma
);
556 * Check that @page is mapped at @address into @mm.
558 * If @sync is false, page_check_address may perform a racy check to avoid
559 * the page table lock when the pte is not present (helpful when reclaiming
560 * highly shared pages).
562 * On success returns with pte mapped and locked.
564 pte_t
*__page_check_address(struct page
*page
, struct mm_struct
*mm
,
565 unsigned long address
, spinlock_t
**ptlp
, int sync
)
573 if (unlikely(PageHuge(page
))) {
574 pte
= huge_pte_offset(mm
, address
);
575 ptl
= &mm
->page_table_lock
;
579 pgd
= pgd_offset(mm
, address
);
580 if (!pgd_present(*pgd
))
583 pud
= pud_offset(pgd
, address
);
584 if (!pud_present(*pud
))
587 pmd
= pmd_offset(pud
, address
);
588 if (!pmd_present(*pmd
))
590 if (pmd_trans_huge(*pmd
))
593 pte
= pte_offset_map(pmd
, address
);
594 /* Make a quick check before getting the lock */
595 if (!sync
&& !pte_present(*pte
)) {
600 ptl
= pte_lockptr(mm
, pmd
);
603 if (pte_present(*pte
) && page_to_pfn(page
) == pte_pfn(*pte
)) {
607 pte_unmap_unlock(pte
, ptl
);
612 * page_mapped_in_vma - check whether a page is really mapped in a VMA
613 * @page: the page to test
614 * @vma: the VMA to test
616 * Returns 1 if the page is mapped into the page tables of the VMA, 0
617 * if the page is not mapped into the page tables of this VMA. Only
618 * valid for normal file or anonymous VMAs.
620 int page_mapped_in_vma(struct page
*page
, struct vm_area_struct
*vma
)
622 unsigned long address
;
626 address
= vma_address(page
, vma
);
627 if (address
== -EFAULT
) /* out of vma range */
629 pte
= page_check_address(page
, vma
->vm_mm
, address
, &ptl
, 1);
630 if (!pte
) /* the page is not in this mm */
632 pte_unmap_unlock(pte
, ptl
);
638 * Subfunctions of page_referenced: page_referenced_one called
639 * repeatedly from either page_referenced_anon or page_referenced_file.
641 int page_referenced_one(struct page
*page
, struct vm_area_struct
*vma
,
642 unsigned long address
, unsigned int *mapcount
,
643 unsigned long *vm_flags
)
645 struct mm_struct
*mm
= vma
->vm_mm
;
648 if (unlikely(PageTransHuge(page
))) {
651 spin_lock(&mm
->page_table_lock
);
653 * rmap might return false positives; we must filter
654 * these out using page_check_address_pmd().
656 pmd
= page_check_address_pmd(page
, mm
, address
,
657 PAGE_CHECK_ADDRESS_PMD_FLAG
);
659 spin_unlock(&mm
->page_table_lock
);
663 if (vma
->vm_flags
& VM_LOCKED
) {
664 spin_unlock(&mm
->page_table_lock
);
665 *mapcount
= 0; /* break early from loop */
666 *vm_flags
|= VM_LOCKED
;
670 /* go ahead even if the pmd is pmd_trans_splitting() */
671 if (pmdp_clear_flush_young_notify(vma
, address
, pmd
))
673 spin_unlock(&mm
->page_table_lock
);
679 * rmap might return false positives; we must filter
680 * these out using page_check_address().
682 pte
= page_check_address(page
, mm
, address
, &ptl
, 0);
686 if (vma
->vm_flags
& VM_LOCKED
) {
687 pte_unmap_unlock(pte
, ptl
);
688 *mapcount
= 0; /* break early from loop */
689 *vm_flags
|= VM_LOCKED
;
693 if (ptep_clear_flush_young_notify(vma
, address
, pte
)) {
695 * Don't treat a reference through a sequentially read
696 * mapping as such. If the page has been used in
697 * another mapping, we will catch it; if this other
698 * mapping is already gone, the unmap path will have
699 * set PG_referenced or activated the page.
701 if (likely(!VM_SequentialReadHint(vma
)))
704 pte_unmap_unlock(pte
, ptl
);
707 /* Pretend the page is referenced if the task has the
708 swap token and is in the middle of a page fault. */
709 if (mm
!= current
->mm
&& has_swap_token(mm
) &&
710 rwsem_is_locked(&mm
->mmap_sem
))
716 *vm_flags
|= vma
->vm_flags
;
721 static int page_referenced_anon(struct page
*page
,
722 struct mem_cgroup
*mem_cont
,
723 unsigned long *vm_flags
)
725 unsigned int mapcount
;
726 struct anon_vma
*anon_vma
;
727 struct anon_vma_chain
*avc
;
730 anon_vma
= page_lock_anon_vma(page
);
734 mapcount
= page_mapcount(page
);
735 list_for_each_entry(avc
, &anon_vma
->head
, same_anon_vma
) {
736 struct vm_area_struct
*vma
= avc
->vma
;
737 unsigned long address
= vma_address(page
, vma
);
738 if (address
== -EFAULT
)
741 * If we are reclaiming on behalf of a cgroup, skip
742 * counting on behalf of references from different
745 if (mem_cont
&& !mm_match_cgroup(vma
->vm_mm
, mem_cont
))
747 referenced
+= page_referenced_one(page
, vma
, address
,
748 &mapcount
, vm_flags
);
753 page_unlock_anon_vma(anon_vma
);
758 * page_referenced_file - referenced check for object-based rmap
759 * @page: the page we're checking references on.
760 * @mem_cont: target memory controller
761 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
763 * For an object-based mapped page, find all the places it is mapped and
764 * check/clear the referenced flag. This is done by following the page->mapping
765 * pointer, then walking the chain of vmas it holds. It returns the number
766 * of references it found.
768 * This function is only called from page_referenced for object-based pages.
770 static int page_referenced_file(struct page
*page
,
771 struct mem_cgroup
*mem_cont
,
772 unsigned long *vm_flags
)
774 unsigned int mapcount
;
775 struct address_space
*mapping
= page
->mapping
;
776 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
777 struct vm_area_struct
*vma
;
778 struct prio_tree_iter iter
;
782 * The caller's checks on page->mapping and !PageAnon have made
783 * sure that this is a file page: the check for page->mapping
784 * excludes the case just before it gets set on an anon page.
786 BUG_ON(PageAnon(page
));
789 * The page lock not only makes sure that page->mapping cannot
790 * suddenly be NULLified by truncation, it makes sure that the
791 * structure at mapping cannot be freed and reused yet,
792 * so we can safely take mapping->i_mmap_mutex.
794 BUG_ON(!PageLocked(page
));
796 mutex_lock(&mapping
->i_mmap_mutex
);
799 * i_mmap_mutex does not stabilize mapcount at all, but mapcount
800 * is more likely to be accurate if we note it after spinning.
802 mapcount
= page_mapcount(page
);
804 vma_prio_tree_foreach(vma
, &iter
, &mapping
->i_mmap
, pgoff
, pgoff
) {
805 unsigned long address
= vma_address(page
, vma
);
806 if (address
== -EFAULT
)
809 * If we are reclaiming on behalf of a cgroup, skip
810 * counting on behalf of references from different
813 if (mem_cont
&& !mm_match_cgroup(vma
->vm_mm
, mem_cont
))
815 referenced
+= page_referenced_one(page
, vma
, address
,
816 &mapcount
, vm_flags
);
821 mutex_unlock(&mapping
->i_mmap_mutex
);
826 * page_referenced - test if the page was referenced
827 * @page: the page to test
828 * @is_locked: caller holds lock on the page
829 * @mem_cont: target memory controller
830 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
832 * Quick test_and_clear_referenced for all mappings to a page,
833 * returns the number of ptes which referenced the page.
835 int page_referenced(struct page
*page
,
837 struct mem_cgroup
*mem_cont
,
838 unsigned long *vm_flags
)
844 if (page_mapped(page
) && page_rmapping(page
)) {
845 if (!is_locked
&& (!PageAnon(page
) || PageKsm(page
))) {
846 we_locked
= trylock_page(page
);
852 if (unlikely(PageKsm(page
)))
853 referenced
+= page_referenced_ksm(page
, mem_cont
,
855 else if (PageAnon(page
))
856 referenced
+= page_referenced_anon(page
, mem_cont
,
858 else if (page
->mapping
)
859 referenced
+= page_referenced_file(page
, mem_cont
,
865 if (page_test_and_clear_young(page_to_pfn(page
)))
871 static int page_mkclean_one(struct page
*page
, struct vm_area_struct
*vma
,
872 unsigned long address
)
874 struct mm_struct
*mm
= vma
->vm_mm
;
879 pte
= page_check_address(page
, mm
, address
, &ptl
, 1);
883 if (pte_dirty(*pte
) || pte_write(*pte
)) {
886 flush_cache_page(vma
, address
, pte_pfn(*pte
));
887 entry
= ptep_clear_flush_notify(vma
, address
, pte
);
888 entry
= pte_wrprotect(entry
);
889 entry
= pte_mkclean(entry
);
890 set_pte_at(mm
, address
, pte
, entry
);
894 pte_unmap_unlock(pte
, ptl
);
899 static int page_mkclean_file(struct address_space
*mapping
, struct page
*page
)
901 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
902 struct vm_area_struct
*vma
;
903 struct prio_tree_iter iter
;
906 BUG_ON(PageAnon(page
));
908 mutex_lock(&mapping
->i_mmap_mutex
);
909 vma_prio_tree_foreach(vma
, &iter
, &mapping
->i_mmap
, pgoff
, pgoff
) {
910 if (vma
->vm_flags
& VM_SHARED
) {
911 unsigned long address
= vma_address(page
, vma
);
912 if (address
== -EFAULT
)
914 ret
+= page_mkclean_one(page
, vma
, address
);
917 mutex_unlock(&mapping
->i_mmap_mutex
);
921 int page_mkclean(struct page
*page
)
925 BUG_ON(!PageLocked(page
));
927 if (page_mapped(page
)) {
928 struct address_space
*mapping
= page_mapping(page
);
930 ret
= page_mkclean_file(mapping
, page
);
931 if (page_test_and_clear_dirty(page_to_pfn(page
), 1))
938 EXPORT_SYMBOL_GPL(page_mkclean
);
941 * page_move_anon_rmap - move a page to our anon_vma
942 * @page: the page to move to our anon_vma
943 * @vma: the vma the page belongs to
944 * @address: the user virtual address mapped
946 * When a page belongs exclusively to one process after a COW event,
947 * that page can be moved into the anon_vma that belongs to just that
948 * process, so the rmap code will not search the parent or sibling
951 void page_move_anon_rmap(struct page
*page
,
952 struct vm_area_struct
*vma
, unsigned long address
)
954 struct anon_vma
*anon_vma
= vma
->anon_vma
;
956 VM_BUG_ON(!PageLocked(page
));
957 VM_BUG_ON(!anon_vma
);
958 VM_BUG_ON(page
->index
!= linear_page_index(vma
, address
));
960 anon_vma
= (void *) anon_vma
+ PAGE_MAPPING_ANON
;
961 page
->mapping
= (struct address_space
*) anon_vma
;
965 * __page_set_anon_rmap - set up new anonymous rmap
966 * @page: Page to add to rmap
967 * @vma: VM area to add page to.
968 * @address: User virtual address of the mapping
969 * @exclusive: the page is exclusively owned by the current process
971 static void __page_set_anon_rmap(struct page
*page
,
972 struct vm_area_struct
*vma
, unsigned long address
, int exclusive
)
974 struct anon_vma
*anon_vma
= vma
->anon_vma
;
982 * If the page isn't exclusively mapped into this vma,
983 * we must use the _oldest_ possible anon_vma for the
987 anon_vma
= anon_vma
->root
;
989 anon_vma
= (void *) anon_vma
+ PAGE_MAPPING_ANON
;
990 page
->mapping
= (struct address_space
*) anon_vma
;
991 page
->index
= linear_page_index(vma
, address
);
995 * __page_check_anon_rmap - sanity check anonymous rmap addition
996 * @page: the page to add the mapping to
997 * @vma: the vm area in which the mapping is added
998 * @address: the user virtual address mapped
1000 static void __page_check_anon_rmap(struct page
*page
,
1001 struct vm_area_struct
*vma
, unsigned long address
)
1003 #ifdef CONFIG_DEBUG_VM
1005 * The page's anon-rmap details (mapping and index) are guaranteed to
1006 * be set up correctly at this point.
1008 * We have exclusion against page_add_anon_rmap because the caller
1009 * always holds the page locked, except if called from page_dup_rmap,
1010 * in which case the page is already known to be setup.
1012 * We have exclusion against page_add_new_anon_rmap because those pages
1013 * are initially only visible via the pagetables, and the pte is locked
1014 * over the call to page_add_new_anon_rmap.
1016 BUG_ON(page_anon_vma(page
)->root
!= vma
->anon_vma
->root
);
1017 BUG_ON(page
->index
!= linear_page_index(vma
, address
));
1022 * page_add_anon_rmap - add pte mapping to an anonymous page
1023 * @page: the page to add the mapping to
1024 * @vma: the vm area in which the mapping is added
1025 * @address: the user virtual address mapped
1027 * The caller needs to hold the pte lock, and the page must be locked in
1028 * the anon_vma case: to serialize mapping,index checking after setting,
1029 * and to ensure that PageAnon is not being upgraded racily to PageKsm
1030 * (but PageKsm is never downgraded to PageAnon).
1032 void page_add_anon_rmap(struct page
*page
,
1033 struct vm_area_struct
*vma
, unsigned long address
)
1035 do_page_add_anon_rmap(page
, vma
, address
, 0);
1039 * Special version of the above for do_swap_page, which often runs
1040 * into pages that are exclusively owned by the current process.
1041 * Everybody else should continue to use page_add_anon_rmap above.
1043 void do_page_add_anon_rmap(struct page
*page
,
1044 struct vm_area_struct
*vma
, unsigned long address
, int exclusive
)
1046 int first
= atomic_inc_and_test(&page
->_mapcount
);
1048 if (!PageTransHuge(page
))
1049 __inc_zone_page_state(page
, NR_ANON_PAGES
);
1051 __inc_zone_page_state(page
,
1052 NR_ANON_TRANSPARENT_HUGEPAGES
);
1054 if (unlikely(PageKsm(page
)))
1057 VM_BUG_ON(!PageLocked(page
));
1058 /* address might be in next vma when migration races vma_adjust */
1060 __page_set_anon_rmap(page
, vma
, address
, exclusive
);
1062 __page_check_anon_rmap(page
, vma
, address
);
1066 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1067 * @page: the page to add the mapping to
1068 * @vma: the vm area in which the mapping is added
1069 * @address: the user virtual address mapped
1071 * Same as page_add_anon_rmap but must only be called on *new* pages.
1072 * This means the inc-and-test can be bypassed.
1073 * Page does not have to be locked.
1075 void page_add_new_anon_rmap(struct page
*page
,
1076 struct vm_area_struct
*vma
, unsigned long address
)
1078 VM_BUG_ON(address
< vma
->vm_start
|| address
>= vma
->vm_end
);
1079 SetPageSwapBacked(page
);
1080 atomic_set(&page
->_mapcount
, 0); /* increment count (starts at -1) */
1081 if (!PageTransHuge(page
))
1082 __inc_zone_page_state(page
, NR_ANON_PAGES
);
1084 __inc_zone_page_state(page
, NR_ANON_TRANSPARENT_HUGEPAGES
);
1085 __page_set_anon_rmap(page
, vma
, address
, 1);
1086 if (page_evictable(page
, vma
))
1087 lru_cache_add_lru(page
, LRU_ACTIVE_ANON
);
1089 add_page_to_unevictable_list(page
);
1093 * page_add_file_rmap - add pte mapping to a file page
1094 * @page: the page to add the mapping to
1096 * The caller needs to hold the pte lock.
1098 void page_add_file_rmap(struct page
*page
)
1100 if (atomic_inc_and_test(&page
->_mapcount
)) {
1101 __inc_zone_page_state(page
, NR_FILE_MAPPED
);
1102 mem_cgroup_inc_page_stat(page
, MEMCG_NR_FILE_MAPPED
);
1107 * page_remove_rmap - take down pte mapping from a page
1108 * @page: page to remove mapping from
1110 * The caller needs to hold the pte lock.
1112 void page_remove_rmap(struct page
*page
)
1114 /* page still mapped by someone else? */
1115 if (!atomic_add_negative(-1, &page
->_mapcount
))
1119 * Now that the last pte has gone, s390 must transfer dirty
1120 * flag from storage key to struct page. We can usually skip
1121 * this if the page is anon, so about to be freed; but perhaps
1122 * not if it's in swapcache - there might be another pte slot
1123 * containing the swap entry, but page not yet written to swap.
1125 if ((!PageAnon(page
) || PageSwapCache(page
)) &&
1126 page_test_and_clear_dirty(page_to_pfn(page
), 1))
1127 set_page_dirty(page
);
1129 * Hugepages are not counted in NR_ANON_PAGES nor NR_FILE_MAPPED
1130 * and not charged by memcg for now.
1132 if (unlikely(PageHuge(page
)))
1134 if (PageAnon(page
)) {
1135 mem_cgroup_uncharge_page(page
);
1136 if (!PageTransHuge(page
))
1137 __dec_zone_page_state(page
, NR_ANON_PAGES
);
1139 __dec_zone_page_state(page
,
1140 NR_ANON_TRANSPARENT_HUGEPAGES
);
1142 __dec_zone_page_state(page
, NR_FILE_MAPPED
);
1143 mem_cgroup_dec_page_stat(page
, MEMCG_NR_FILE_MAPPED
);
1146 * It would be tidy to reset the PageAnon mapping here,
1147 * but that might overwrite a racing page_add_anon_rmap
1148 * which increments mapcount after us but sets mapping
1149 * before us: so leave the reset to free_hot_cold_page,
1150 * and remember that it's only reliable while mapped.
1151 * Leaving it set also helps swapoff to reinstate ptes
1152 * faster for those pages still in swapcache.
1157 * Subfunctions of try_to_unmap: try_to_unmap_one called
1158 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
1160 int try_to_unmap_one(struct page
*page
, struct vm_area_struct
*vma
,
1161 unsigned long address
, enum ttu_flags flags
)
1163 struct mm_struct
*mm
= vma
->vm_mm
;
1167 int ret
= SWAP_AGAIN
;
1169 pte
= page_check_address(page
, mm
, address
, &ptl
, 0);
1174 * If the page is mlock()d, we cannot swap it out.
1175 * If it's recently referenced (perhaps page_referenced
1176 * skipped over this mm) then we should reactivate it.
1178 if (!(flags
& TTU_IGNORE_MLOCK
)) {
1179 if (vma
->vm_flags
& VM_LOCKED
)
1182 if (TTU_ACTION(flags
) == TTU_MUNLOCK
)
1185 if (!(flags
& TTU_IGNORE_ACCESS
)) {
1186 if (ptep_clear_flush_young_notify(vma
, address
, pte
)) {
1192 /* Nuke the page table entry. */
1193 flush_cache_page(vma
, address
, page_to_pfn(page
));
1194 pteval
= ptep_clear_flush_notify(vma
, address
, pte
);
1196 /* Move the dirty bit to the physical page now the pte is gone. */
1197 if (pte_dirty(pteval
))
1198 set_page_dirty(page
);
1200 /* Update high watermark before we lower rss */
1201 update_hiwater_rss(mm
);
1203 if (PageHWPoison(page
) && !(flags
& TTU_IGNORE_HWPOISON
)) {
1205 dec_mm_counter(mm
, MM_ANONPAGES
);
1207 dec_mm_counter(mm
, MM_FILEPAGES
);
1208 set_pte_at(mm
, address
, pte
,
1209 swp_entry_to_pte(make_hwpoison_entry(page
)));
1210 } else if (PageAnon(page
)) {
1211 swp_entry_t entry
= { .val
= page_private(page
) };
1213 if (PageSwapCache(page
)) {
1215 * Store the swap location in the pte.
1216 * See handle_pte_fault() ...
1218 if (swap_duplicate(entry
) < 0) {
1219 set_pte_at(mm
, address
, pte
, pteval
);
1223 if (list_empty(&mm
->mmlist
)) {
1224 spin_lock(&mmlist_lock
);
1225 if (list_empty(&mm
->mmlist
))
1226 list_add(&mm
->mmlist
, &init_mm
.mmlist
);
1227 spin_unlock(&mmlist_lock
);
1229 dec_mm_counter(mm
, MM_ANONPAGES
);
1230 inc_mm_counter(mm
, MM_SWAPENTS
);
1231 } else if (PAGE_MIGRATION
) {
1233 * Store the pfn of the page in a special migration
1234 * pte. do_swap_page() will wait until the migration
1235 * pte is removed and then restart fault handling.
1237 BUG_ON(TTU_ACTION(flags
) != TTU_MIGRATION
);
1238 entry
= make_migration_entry(page
, pte_write(pteval
));
1240 set_pte_at(mm
, address
, pte
, swp_entry_to_pte(entry
));
1241 BUG_ON(pte_file(*pte
));
1242 } else if (PAGE_MIGRATION
&& (TTU_ACTION(flags
) == TTU_MIGRATION
)) {
1243 /* Establish migration entry for a file page */
1245 entry
= make_migration_entry(page
, pte_write(pteval
));
1246 set_pte_at(mm
, address
, pte
, swp_entry_to_pte(entry
));
1248 dec_mm_counter(mm
, MM_FILEPAGES
);
1250 page_remove_rmap(page
);
1251 page_cache_release(page
);
1254 pte_unmap_unlock(pte
, ptl
);
1259 pte_unmap_unlock(pte
, ptl
);
1263 * We need mmap_sem locking, Otherwise VM_LOCKED check makes
1264 * unstable result and race. Plus, We can't wait here because
1265 * we now hold anon_vma->mutex or mapping->i_mmap_mutex.
1266 * if trylock failed, the page remain in evictable lru and later
1267 * vmscan could retry to move the page to unevictable lru if the
1268 * page is actually mlocked.
1270 if (down_read_trylock(&vma
->vm_mm
->mmap_sem
)) {
1271 if (vma
->vm_flags
& VM_LOCKED
) {
1272 mlock_vma_page(page
);
1275 up_read(&vma
->vm_mm
->mmap_sem
);
1281 * objrmap doesn't work for nonlinear VMAs because the assumption that
1282 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
1283 * Consequently, given a particular page and its ->index, we cannot locate the
1284 * ptes which are mapping that page without an exhaustive linear search.
1286 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
1287 * maps the file to which the target page belongs. The ->vm_private_data field
1288 * holds the current cursor into that scan. Successive searches will circulate
1289 * around the vma's virtual address space.
1291 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
1292 * more scanning pressure is placed against them as well. Eventually pages
1293 * will become fully unmapped and are eligible for eviction.
1295 * For very sparsely populated VMAs this is a little inefficient - chances are
1296 * there there won't be many ptes located within the scan cluster. In this case
1297 * maybe we could scan further - to the end of the pte page, perhaps.
1299 * Mlocked pages: check VM_LOCKED under mmap_sem held for read, if we can
1300 * acquire it without blocking. If vma locked, mlock the pages in the cluster,
1301 * rather than unmapping them. If we encounter the "check_page" that vmscan is
1302 * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN.
1304 #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
1305 #define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
1307 static int try_to_unmap_cluster(unsigned long cursor
, unsigned int *mapcount
,
1308 struct vm_area_struct
*vma
, struct page
*check_page
)
1310 struct mm_struct
*mm
= vma
->vm_mm
;
1318 unsigned long address
;
1320 int ret
= SWAP_AGAIN
;
1323 address
= (vma
->vm_start
+ cursor
) & CLUSTER_MASK
;
1324 end
= address
+ CLUSTER_SIZE
;
1325 if (address
< vma
->vm_start
)
1326 address
= vma
->vm_start
;
1327 if (end
> vma
->vm_end
)
1330 pgd
= pgd_offset(mm
, address
);
1331 if (!pgd_present(*pgd
))
1334 pud
= pud_offset(pgd
, address
);
1335 if (!pud_present(*pud
))
1338 pmd
= pmd_offset(pud
, address
);
1339 if (!pmd_present(*pmd
))
1343 * If we can acquire the mmap_sem for read, and vma is VM_LOCKED,
1344 * keep the sem while scanning the cluster for mlocking pages.
1346 if (down_read_trylock(&vma
->vm_mm
->mmap_sem
)) {
1347 locked_vma
= (vma
->vm_flags
& VM_LOCKED
);
1349 up_read(&vma
->vm_mm
->mmap_sem
); /* don't need it */
1352 pte
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
1354 /* Update high watermark before we lower rss */
1355 update_hiwater_rss(mm
);
1357 for (; address
< end
; pte
++, address
+= PAGE_SIZE
) {
1358 if (!pte_present(*pte
))
1360 page
= vm_normal_page(vma
, address
, *pte
);
1361 BUG_ON(!page
|| PageAnon(page
));
1364 mlock_vma_page(page
); /* no-op if already mlocked */
1365 if (page
== check_page
)
1367 continue; /* don't unmap */
1370 if (ptep_clear_flush_young_notify(vma
, address
, pte
))
1373 /* Nuke the page table entry. */
1374 flush_cache_page(vma
, address
, pte_pfn(*pte
));
1375 pteval
= ptep_clear_flush_notify(vma
, address
, pte
);
1377 /* If nonlinear, store the file page offset in the pte. */
1378 if (page
->index
!= linear_page_index(vma
, address
))
1379 set_pte_at(mm
, address
, pte
, pgoff_to_pte(page
->index
));
1381 /* Move the dirty bit to the physical page now the pte is gone. */
1382 if (pte_dirty(pteval
))
1383 set_page_dirty(page
);
1385 page_remove_rmap(page
);
1386 page_cache_release(page
);
1387 dec_mm_counter(mm
, MM_FILEPAGES
);
1390 pte_unmap_unlock(pte
- 1, ptl
);
1392 up_read(&vma
->vm_mm
->mmap_sem
);
1396 bool is_vma_temporary_stack(struct vm_area_struct
*vma
)
1398 int maybe_stack
= vma
->vm_flags
& (VM_GROWSDOWN
| VM_GROWSUP
);
1403 if ((vma
->vm_flags
& VM_STACK_INCOMPLETE_SETUP
) ==
1404 VM_STACK_INCOMPLETE_SETUP
)
1411 * try_to_unmap_anon - unmap or unlock anonymous page using the object-based
1413 * @page: the page to unmap/unlock
1414 * @flags: action and flags
1416 * Find all the mappings of a page using the mapping pointer and the vma chains
1417 * contained in the anon_vma struct it points to.
1419 * This function is only called from try_to_unmap/try_to_munlock for
1421 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1422 * where the page was found will be held for write. So, we won't recheck
1423 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1426 static int try_to_unmap_anon(struct page
*page
, enum ttu_flags flags
)
1428 struct anon_vma
*anon_vma
;
1429 struct anon_vma_chain
*avc
;
1430 int ret
= SWAP_AGAIN
;
1432 anon_vma
= page_lock_anon_vma(page
);
1436 list_for_each_entry(avc
, &anon_vma
->head
, same_anon_vma
) {
1437 struct vm_area_struct
*vma
= avc
->vma
;
1438 unsigned long address
;
1441 * During exec, a temporary VMA is setup and later moved.
1442 * The VMA is moved under the anon_vma lock but not the
1443 * page tables leading to a race where migration cannot
1444 * find the migration ptes. Rather than increasing the
1445 * locking requirements of exec(), migration skips
1446 * temporary VMAs until after exec() completes.
1448 if (PAGE_MIGRATION
&& (flags
& TTU_MIGRATION
) &&
1449 is_vma_temporary_stack(vma
))
1452 address
= vma_address(page
, vma
);
1453 if (address
== -EFAULT
)
1455 ret
= try_to_unmap_one(page
, vma
, address
, flags
);
1456 if (ret
!= SWAP_AGAIN
|| !page_mapped(page
))
1460 page_unlock_anon_vma(anon_vma
);
1465 * try_to_unmap_file - unmap/unlock file page using the object-based rmap method
1466 * @page: the page to unmap/unlock
1467 * @flags: action and flags
1469 * Find all the mappings of a page using the mapping pointer and the vma chains
1470 * contained in the address_space struct it points to.
1472 * This function is only called from try_to_unmap/try_to_munlock for
1473 * object-based pages.
1474 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1475 * where the page was found will be held for write. So, we won't recheck
1476 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1479 static int try_to_unmap_file(struct page
*page
, enum ttu_flags flags
)
1481 struct address_space
*mapping
= page
->mapping
;
1482 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
1483 struct vm_area_struct
*vma
;
1484 struct prio_tree_iter iter
;
1485 int ret
= SWAP_AGAIN
;
1486 unsigned long cursor
;
1487 unsigned long max_nl_cursor
= 0;
1488 unsigned long max_nl_size
= 0;
1489 unsigned int mapcount
;
1491 mutex_lock(&mapping
->i_mmap_mutex
);
1492 vma_prio_tree_foreach(vma
, &iter
, &mapping
->i_mmap
, pgoff
, pgoff
) {
1493 unsigned long address
= vma_address(page
, vma
);
1494 if (address
== -EFAULT
)
1496 ret
= try_to_unmap_one(page
, vma
, address
, flags
);
1497 if (ret
!= SWAP_AGAIN
|| !page_mapped(page
))
1501 if (list_empty(&mapping
->i_mmap_nonlinear
))
1505 * We don't bother to try to find the munlocked page in nonlinears.
1506 * It's costly. Instead, later, page reclaim logic may call
1507 * try_to_unmap(TTU_MUNLOCK) and recover PG_mlocked lazily.
1509 if (TTU_ACTION(flags
) == TTU_MUNLOCK
)
1512 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
,
1513 shared
.vm_set
.list
) {
1514 cursor
= (unsigned long) vma
->vm_private_data
;
1515 if (cursor
> max_nl_cursor
)
1516 max_nl_cursor
= cursor
;
1517 cursor
= vma
->vm_end
- vma
->vm_start
;
1518 if (cursor
> max_nl_size
)
1519 max_nl_size
= cursor
;
1522 if (max_nl_size
== 0) { /* all nonlinears locked or reserved ? */
1528 * We don't try to search for this page in the nonlinear vmas,
1529 * and page_referenced wouldn't have found it anyway. Instead
1530 * just walk the nonlinear vmas trying to age and unmap some.
1531 * The mapcount of the page we came in with is irrelevant,
1532 * but even so use it as a guide to how hard we should try?
1534 mapcount
= page_mapcount(page
);
1539 max_nl_size
= (max_nl_size
+ CLUSTER_SIZE
- 1) & CLUSTER_MASK
;
1540 if (max_nl_cursor
== 0)
1541 max_nl_cursor
= CLUSTER_SIZE
;
1544 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
,
1545 shared
.vm_set
.list
) {
1546 cursor
= (unsigned long) vma
->vm_private_data
;
1547 while ( cursor
< max_nl_cursor
&&
1548 cursor
< vma
->vm_end
- vma
->vm_start
) {
1549 if (try_to_unmap_cluster(cursor
, &mapcount
,
1550 vma
, page
) == SWAP_MLOCK
)
1552 cursor
+= CLUSTER_SIZE
;
1553 vma
->vm_private_data
= (void *) cursor
;
1554 if ((int)mapcount
<= 0)
1557 vma
->vm_private_data
= (void *) max_nl_cursor
;
1560 max_nl_cursor
+= CLUSTER_SIZE
;
1561 } while (max_nl_cursor
<= max_nl_size
);
1564 * Don't loop forever (perhaps all the remaining pages are
1565 * in locked vmas). Reset cursor on all unreserved nonlinear
1566 * vmas, now forgetting on which ones it had fallen behind.
1568 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
, shared
.vm_set
.list
)
1569 vma
->vm_private_data
= NULL
;
1571 mutex_unlock(&mapping
->i_mmap_mutex
);
1576 * try_to_unmap - try to remove all page table mappings to a page
1577 * @page: the page to get unmapped
1578 * @flags: action and flags
1580 * Tries to remove all the page table entries which are mapping this
1581 * page, used in the pageout path. Caller must hold the page lock.
1582 * Return values are:
1584 * SWAP_SUCCESS - we succeeded in removing all mappings
1585 * SWAP_AGAIN - we missed a mapping, try again later
1586 * SWAP_FAIL - the page is unswappable
1587 * SWAP_MLOCK - page is mlocked.
1589 int try_to_unmap(struct page
*page
, enum ttu_flags flags
)
1593 BUG_ON(!PageLocked(page
));
1594 VM_BUG_ON(!PageHuge(page
) && PageTransHuge(page
));
1596 if (unlikely(PageKsm(page
)))
1597 ret
= try_to_unmap_ksm(page
, flags
);
1598 else if (PageAnon(page
))
1599 ret
= try_to_unmap_anon(page
, flags
);
1601 ret
= try_to_unmap_file(page
, flags
);
1602 if (ret
!= SWAP_MLOCK
&& !page_mapped(page
))
1608 * try_to_munlock - try to munlock a page
1609 * @page: the page to be munlocked
1611 * Called from munlock code. Checks all of the VMAs mapping the page
1612 * to make sure nobody else has this page mlocked. The page will be
1613 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1615 * Return values are:
1617 * SWAP_AGAIN - no vma is holding page mlocked, or,
1618 * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem
1619 * SWAP_FAIL - page cannot be located at present
1620 * SWAP_MLOCK - page is now mlocked.
1622 int try_to_munlock(struct page
*page
)
1624 VM_BUG_ON(!PageLocked(page
) || PageLRU(page
));
1626 if (unlikely(PageKsm(page
)))
1627 return try_to_unmap_ksm(page
, TTU_MUNLOCK
);
1628 else if (PageAnon(page
))
1629 return try_to_unmap_anon(page
, TTU_MUNLOCK
);
1631 return try_to_unmap_file(page
, TTU_MUNLOCK
);
1634 void __put_anon_vma(struct anon_vma
*anon_vma
)
1636 struct anon_vma
*root
= anon_vma
->root
;
1638 if (root
!= anon_vma
&& atomic_dec_and_test(&root
->refcount
))
1639 anon_vma_free(root
);
1641 anon_vma_free(anon_vma
);
1644 #ifdef CONFIG_MIGRATION
1646 * rmap_walk() and its helpers rmap_walk_anon() and rmap_walk_file():
1647 * Called by migrate.c to remove migration ptes, but might be used more later.
1649 static int rmap_walk_anon(struct page
*page
, int (*rmap_one
)(struct page
*,
1650 struct vm_area_struct
*, unsigned long, void *), void *arg
)
1652 struct anon_vma
*anon_vma
;
1653 struct anon_vma_chain
*avc
;
1654 int ret
= SWAP_AGAIN
;
1657 * Note: remove_migration_ptes() cannot use page_lock_anon_vma()
1658 * because that depends on page_mapped(); but not all its usages
1659 * are holding mmap_sem. Users without mmap_sem are required to
1660 * take a reference count to prevent the anon_vma disappearing
1662 anon_vma
= page_anon_vma(page
);
1665 anon_vma_lock(anon_vma
);
1666 list_for_each_entry(avc
, &anon_vma
->head
, same_anon_vma
) {
1667 struct vm_area_struct
*vma
= avc
->vma
;
1668 unsigned long address
= vma_address(page
, vma
);
1669 if (address
== -EFAULT
)
1671 ret
= rmap_one(page
, vma
, address
, arg
);
1672 if (ret
!= SWAP_AGAIN
)
1675 anon_vma_unlock(anon_vma
);
1679 static int rmap_walk_file(struct page
*page
, int (*rmap_one
)(struct page
*,
1680 struct vm_area_struct
*, unsigned long, void *), void *arg
)
1682 struct address_space
*mapping
= page
->mapping
;
1683 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
1684 struct vm_area_struct
*vma
;
1685 struct prio_tree_iter iter
;
1686 int ret
= SWAP_AGAIN
;
1690 mutex_lock(&mapping
->i_mmap_mutex
);
1691 vma_prio_tree_foreach(vma
, &iter
, &mapping
->i_mmap
, pgoff
, pgoff
) {
1692 unsigned long address
= vma_address(page
, vma
);
1693 if (address
== -EFAULT
)
1695 ret
= rmap_one(page
, vma
, address
, arg
);
1696 if (ret
!= SWAP_AGAIN
)
1700 * No nonlinear handling: being always shared, nonlinear vmas
1701 * never contain migration ptes. Decide what to do about this
1702 * limitation to linear when we need rmap_walk() on nonlinear.
1704 mutex_unlock(&mapping
->i_mmap_mutex
);
1708 int rmap_walk(struct page
*page
, int (*rmap_one
)(struct page
*,
1709 struct vm_area_struct
*, unsigned long, void *), void *arg
)
1711 VM_BUG_ON(!PageLocked(page
));
1713 if (unlikely(PageKsm(page
)))
1714 return rmap_walk_ksm(page
, rmap_one
, arg
);
1715 else if (PageAnon(page
))
1716 return rmap_walk_anon(page
, rmap_one
, arg
);
1718 return rmap_walk_file(page
, rmap_one
, arg
);
1720 #endif /* CONFIG_MIGRATION */
1722 #ifdef CONFIG_HUGETLB_PAGE
1724 * The following three functions are for anonymous (private mapped) hugepages.
1725 * Unlike common anonymous pages, anonymous hugepages have no accounting code
1726 * and no lru code, because we handle hugepages differently from common pages.
1728 static void __hugepage_set_anon_rmap(struct page
*page
,
1729 struct vm_area_struct
*vma
, unsigned long address
, int exclusive
)
1731 struct anon_vma
*anon_vma
= vma
->anon_vma
;
1738 anon_vma
= anon_vma
->root
;
1740 anon_vma
= (void *) anon_vma
+ PAGE_MAPPING_ANON
;
1741 page
->mapping
= (struct address_space
*) anon_vma
;
1742 page
->index
= linear_page_index(vma
, address
);
1745 void hugepage_add_anon_rmap(struct page
*page
,
1746 struct vm_area_struct
*vma
, unsigned long address
)
1748 struct anon_vma
*anon_vma
= vma
->anon_vma
;
1751 BUG_ON(!PageLocked(page
));
1753 /* address might be in next vma when migration races vma_adjust */
1754 first
= atomic_inc_and_test(&page
->_mapcount
);
1756 __hugepage_set_anon_rmap(page
, vma
, address
, 0);
1759 void hugepage_add_new_anon_rmap(struct page
*page
,
1760 struct vm_area_struct
*vma
, unsigned long address
)
1762 BUG_ON(address
< vma
->vm_start
|| address
>= vma
->vm_end
);
1763 atomic_set(&page
->_mapcount
, 0);
1764 __hugepage_set_anon_rmap(page
, vma
, address
, 1);
1766 #endif /* CONFIG_HUGETLB_PAGE */