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 void unlink_anon_vmas(struct vm_area_struct
*vma
)
329 struct anon_vma_chain
*avc
, *next
;
330 struct anon_vma
*root
= NULL
;
333 * Unlink each anon_vma chained to the VMA. This list is ordered
334 * from newest to oldest, ensuring the root anon_vma gets freed last.
336 list_for_each_entry_safe(avc
, next
, &vma
->anon_vma_chain
, same_vma
) {
337 struct anon_vma
*anon_vma
= avc
->anon_vma
;
339 root
= lock_anon_vma_root(root
, anon_vma
);
340 list_del(&avc
->same_anon_vma
);
343 * Leave empty anon_vmas on the list - we'll need
344 * to free them outside the lock.
346 if (list_empty(&anon_vma
->head
))
349 list_del(&avc
->same_vma
);
350 anon_vma_chain_free(avc
);
352 unlock_anon_vma_root(root
);
355 * Iterate the list once more, it now only contains empty and unlinked
356 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
357 * needing to acquire the anon_vma->root->mutex.
359 list_for_each_entry_safe(avc
, next
, &vma
->anon_vma_chain
, same_vma
) {
360 struct anon_vma
*anon_vma
= avc
->anon_vma
;
362 put_anon_vma(anon_vma
);
364 list_del(&avc
->same_vma
);
365 anon_vma_chain_free(avc
);
369 static void anon_vma_ctor(void *data
)
371 struct anon_vma
*anon_vma
= data
;
373 mutex_init(&anon_vma
->mutex
);
374 atomic_set(&anon_vma
->refcount
, 0);
375 INIT_LIST_HEAD(&anon_vma
->head
);
378 void __init
anon_vma_init(void)
380 anon_vma_cachep
= kmem_cache_create("anon_vma", sizeof(struct anon_vma
),
381 0, SLAB_DESTROY_BY_RCU
|SLAB_PANIC
, anon_vma_ctor
);
382 anon_vma_chain_cachep
= KMEM_CACHE(anon_vma_chain
, SLAB_PANIC
);
386 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
388 * Since there is no serialization what so ever against page_remove_rmap()
389 * the best this function can do is return a locked anon_vma that might
390 * have been relevant to this page.
392 * The page might have been remapped to a different anon_vma or the anon_vma
393 * returned may already be freed (and even reused).
395 * In case it was remapped to a different anon_vma, the new anon_vma will be a
396 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
397 * ensure that any anon_vma obtained from the page will still be valid for as
398 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
400 * All users of this function must be very careful when walking the anon_vma
401 * chain and verify that the page in question is indeed mapped in it
402 * [ something equivalent to page_mapped_in_vma() ].
404 * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap()
405 * that the anon_vma pointer from page->mapping is valid if there is a
406 * mapcount, we can dereference the anon_vma after observing those.
408 struct anon_vma
*page_get_anon_vma(struct page
*page
)
410 struct anon_vma
*anon_vma
= NULL
;
411 unsigned long anon_mapping
;
414 anon_mapping
= (unsigned long) ACCESS_ONCE(page
->mapping
);
415 if ((anon_mapping
& PAGE_MAPPING_FLAGS
) != PAGE_MAPPING_ANON
)
417 if (!page_mapped(page
))
420 anon_vma
= (struct anon_vma
*) (anon_mapping
- PAGE_MAPPING_ANON
);
421 if (!atomic_inc_not_zero(&anon_vma
->refcount
)) {
427 * If this page is still mapped, then its anon_vma cannot have been
428 * freed. But if it has been unmapped, we have no security against the
429 * anon_vma structure being freed and reused (for another anon_vma:
430 * SLAB_DESTROY_BY_RCU guarantees that - so the atomic_inc_not_zero()
431 * above cannot corrupt).
433 if (!page_mapped(page
)) {
434 put_anon_vma(anon_vma
);
444 * Similar to page_get_anon_vma() except it locks the anon_vma.
446 * Its a little more complex as it tries to keep the fast path to a single
447 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
448 * reference like with page_get_anon_vma() and then block on the mutex.
450 struct anon_vma
*page_lock_anon_vma(struct page
*page
)
452 struct anon_vma
*anon_vma
= NULL
;
453 struct anon_vma
*root_anon_vma
;
454 unsigned long anon_mapping
;
457 anon_mapping
= (unsigned long) ACCESS_ONCE(page
->mapping
);
458 if ((anon_mapping
& PAGE_MAPPING_FLAGS
) != PAGE_MAPPING_ANON
)
460 if (!page_mapped(page
))
463 anon_vma
= (struct anon_vma
*) (anon_mapping
- PAGE_MAPPING_ANON
);
464 root_anon_vma
= ACCESS_ONCE(anon_vma
->root
);
465 if (mutex_trylock(&root_anon_vma
->mutex
)) {
467 * If the page is still mapped, then this anon_vma is still
468 * its anon_vma, and holding the mutex ensures that it will
469 * not go away, see anon_vma_free().
471 if (!page_mapped(page
)) {
472 mutex_unlock(&root_anon_vma
->mutex
);
478 /* trylock failed, we got to sleep */
479 if (!atomic_inc_not_zero(&anon_vma
->refcount
)) {
484 if (!page_mapped(page
)) {
485 put_anon_vma(anon_vma
);
490 /* we pinned the anon_vma, its safe to sleep */
492 anon_vma_lock(anon_vma
);
494 if (atomic_dec_and_test(&anon_vma
->refcount
)) {
496 * Oops, we held the last refcount, release the lock
497 * and bail -- can't simply use put_anon_vma() because
498 * we'll deadlock on the anon_vma_lock() recursion.
500 anon_vma_unlock(anon_vma
);
501 __put_anon_vma(anon_vma
);
512 void page_unlock_anon_vma(struct anon_vma
*anon_vma
)
514 anon_vma_unlock(anon_vma
);
518 * At what user virtual address is page expected in @vma?
519 * Returns virtual address or -EFAULT if page's index/offset is not
520 * within the range mapped the @vma.
523 vma_address(struct page
*page
, struct vm_area_struct
*vma
)
525 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
526 unsigned long address
;
528 if (unlikely(is_vm_hugetlb_page(vma
)))
529 pgoff
= page
->index
<< huge_page_order(page_hstate(page
));
530 address
= vma
->vm_start
+ ((pgoff
- vma
->vm_pgoff
) << PAGE_SHIFT
);
531 if (unlikely(address
< vma
->vm_start
|| address
>= vma
->vm_end
)) {
532 /* page should be within @vma mapping range */
539 * At what user virtual address is page expected in vma?
540 * Caller should check the page is actually part of the vma.
542 unsigned long page_address_in_vma(struct page
*page
, struct vm_area_struct
*vma
)
544 if (PageAnon(page
)) {
545 struct anon_vma
*page__anon_vma
= page_anon_vma(page
);
547 * Note: swapoff's unuse_vma() is more efficient with this
548 * check, and needs it to match anon_vma when KSM is active.
550 if (!vma
->anon_vma
|| !page__anon_vma
||
551 vma
->anon_vma
->root
!= page__anon_vma
->root
)
553 } else if (page
->mapping
&& !(vma
->vm_flags
& VM_NONLINEAR
)) {
555 vma
->vm_file
->f_mapping
!= page
->mapping
)
559 return vma_address(page
, vma
);
563 * Check that @page is mapped at @address into @mm.
565 * If @sync is false, page_check_address may perform a racy check to avoid
566 * the page table lock when the pte is not present (helpful when reclaiming
567 * highly shared pages).
569 * On success returns with pte mapped and locked.
571 pte_t
*__page_check_address(struct page
*page
, struct mm_struct
*mm
,
572 unsigned long address
, spinlock_t
**ptlp
, int sync
)
580 if (unlikely(PageHuge(page
))) {
581 pte
= huge_pte_offset(mm
, address
);
582 ptl
= &mm
->page_table_lock
;
586 pgd
= pgd_offset(mm
, address
);
587 if (!pgd_present(*pgd
))
590 pud
= pud_offset(pgd
, address
);
591 if (!pud_present(*pud
))
594 pmd
= pmd_offset(pud
, address
);
595 if (!pmd_present(*pmd
))
597 if (pmd_trans_huge(*pmd
))
600 pte
= pte_offset_map(pmd
, address
);
601 /* Make a quick check before getting the lock */
602 if (!sync
&& !pte_present(*pte
)) {
607 ptl
= pte_lockptr(mm
, pmd
);
610 if (pte_present(*pte
) && page_to_pfn(page
) == pte_pfn(*pte
)) {
614 pte_unmap_unlock(pte
, ptl
);
619 * page_mapped_in_vma - check whether a page is really mapped in a VMA
620 * @page: the page to test
621 * @vma: the VMA to test
623 * Returns 1 if the page is mapped into the page tables of the VMA, 0
624 * if the page is not mapped into the page tables of this VMA. Only
625 * valid for normal file or anonymous VMAs.
627 int page_mapped_in_vma(struct page
*page
, struct vm_area_struct
*vma
)
629 unsigned long address
;
633 address
= vma_address(page
, vma
);
634 if (address
== -EFAULT
) /* out of vma range */
636 pte
= page_check_address(page
, vma
->vm_mm
, address
, &ptl
, 1);
637 if (!pte
) /* the page is not in this mm */
639 pte_unmap_unlock(pte
, ptl
);
645 * Subfunctions of page_referenced: page_referenced_one called
646 * repeatedly from either page_referenced_anon or page_referenced_file.
648 int page_referenced_one(struct page
*page
, struct vm_area_struct
*vma
,
649 unsigned long address
, unsigned int *mapcount
,
650 unsigned long *vm_flags
)
652 struct mm_struct
*mm
= vma
->vm_mm
;
655 if (unlikely(PageTransHuge(page
))) {
658 spin_lock(&mm
->page_table_lock
);
660 * rmap might return false positives; we must filter
661 * these out using page_check_address_pmd().
663 pmd
= page_check_address_pmd(page
, mm
, address
,
664 PAGE_CHECK_ADDRESS_PMD_FLAG
);
666 spin_unlock(&mm
->page_table_lock
);
670 if (vma
->vm_flags
& VM_LOCKED
) {
671 spin_unlock(&mm
->page_table_lock
);
672 *mapcount
= 0; /* break early from loop */
673 *vm_flags
|= VM_LOCKED
;
677 /* go ahead even if the pmd is pmd_trans_splitting() */
678 if (pmdp_clear_flush_young_notify(vma
, address
, pmd
))
680 spin_unlock(&mm
->page_table_lock
);
686 * rmap might return false positives; we must filter
687 * these out using page_check_address().
689 pte
= page_check_address(page
, mm
, address
, &ptl
, 0);
693 if (vma
->vm_flags
& VM_LOCKED
) {
694 pte_unmap_unlock(pte
, ptl
);
695 *mapcount
= 0; /* break early from loop */
696 *vm_flags
|= VM_LOCKED
;
700 if (ptep_clear_flush_young_notify(vma
, address
, pte
)) {
702 * Don't treat a reference through a sequentially read
703 * mapping as such. If the page has been used in
704 * another mapping, we will catch it; if this other
705 * mapping is already gone, the unmap path will have
706 * set PG_referenced or activated the page.
708 if (likely(!VM_SequentialReadHint(vma
)))
711 pte_unmap_unlock(pte
, ptl
);
714 /* Pretend the page is referenced if the task has the
715 swap token and is in the middle of a page fault. */
716 if (mm
!= current
->mm
&& has_swap_token(mm
) &&
717 rwsem_is_locked(&mm
->mmap_sem
))
723 *vm_flags
|= vma
->vm_flags
;
728 static int page_referenced_anon(struct page
*page
,
729 struct mem_cgroup
*mem_cont
,
730 unsigned long *vm_flags
)
732 unsigned int mapcount
;
733 struct anon_vma
*anon_vma
;
734 struct anon_vma_chain
*avc
;
737 anon_vma
= page_lock_anon_vma(page
);
741 mapcount
= page_mapcount(page
);
742 list_for_each_entry(avc
, &anon_vma
->head
, same_anon_vma
) {
743 struct vm_area_struct
*vma
= avc
->vma
;
744 unsigned long address
= vma_address(page
, vma
);
745 if (address
== -EFAULT
)
748 * If we are reclaiming on behalf of a cgroup, skip
749 * counting on behalf of references from different
752 if (mem_cont
&& !mm_match_cgroup(vma
->vm_mm
, mem_cont
))
754 referenced
+= page_referenced_one(page
, vma
, address
,
755 &mapcount
, vm_flags
);
760 page_unlock_anon_vma(anon_vma
);
765 * page_referenced_file - referenced check for object-based rmap
766 * @page: the page we're checking references on.
767 * @mem_cont: target memory controller
768 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
770 * For an object-based mapped page, find all the places it is mapped and
771 * check/clear the referenced flag. This is done by following the page->mapping
772 * pointer, then walking the chain of vmas it holds. It returns the number
773 * of references it found.
775 * This function is only called from page_referenced for object-based pages.
777 static int page_referenced_file(struct page
*page
,
778 struct mem_cgroup
*mem_cont
,
779 unsigned long *vm_flags
)
781 unsigned int mapcount
;
782 struct address_space
*mapping
= page
->mapping
;
783 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
784 struct vm_area_struct
*vma
;
785 struct prio_tree_iter iter
;
789 * The caller's checks on page->mapping and !PageAnon have made
790 * sure that this is a file page: the check for page->mapping
791 * excludes the case just before it gets set on an anon page.
793 BUG_ON(PageAnon(page
));
796 * The page lock not only makes sure that page->mapping cannot
797 * suddenly be NULLified by truncation, it makes sure that the
798 * structure at mapping cannot be freed and reused yet,
799 * so we can safely take mapping->i_mmap_mutex.
801 BUG_ON(!PageLocked(page
));
803 mutex_lock(&mapping
->i_mmap_mutex
);
806 * i_mmap_mutex does not stabilize mapcount at all, but mapcount
807 * is more likely to be accurate if we note it after spinning.
809 mapcount
= page_mapcount(page
);
811 vma_prio_tree_foreach(vma
, &iter
, &mapping
->i_mmap
, pgoff
, pgoff
) {
812 unsigned long address
= vma_address(page
, vma
);
813 if (address
== -EFAULT
)
816 * If we are reclaiming on behalf of a cgroup, skip
817 * counting on behalf of references from different
820 if (mem_cont
&& !mm_match_cgroup(vma
->vm_mm
, mem_cont
))
822 referenced
+= page_referenced_one(page
, vma
, address
,
823 &mapcount
, vm_flags
);
828 mutex_unlock(&mapping
->i_mmap_mutex
);
833 * page_referenced - test if the page was referenced
834 * @page: the page to test
835 * @is_locked: caller holds lock on the page
836 * @mem_cont: target memory controller
837 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
839 * Quick test_and_clear_referenced for all mappings to a page,
840 * returns the number of ptes which referenced the page.
842 int page_referenced(struct page
*page
,
844 struct mem_cgroup
*mem_cont
,
845 unsigned long *vm_flags
)
851 if (page_mapped(page
) && page_rmapping(page
)) {
852 if (!is_locked
&& (!PageAnon(page
) || PageKsm(page
))) {
853 we_locked
= trylock_page(page
);
859 if (unlikely(PageKsm(page
)))
860 referenced
+= page_referenced_ksm(page
, mem_cont
,
862 else if (PageAnon(page
))
863 referenced
+= page_referenced_anon(page
, mem_cont
,
865 else if (page
->mapping
)
866 referenced
+= page_referenced_file(page
, mem_cont
,
872 if (page_test_and_clear_young(page_to_pfn(page
)))
878 static int page_mkclean_one(struct page
*page
, struct vm_area_struct
*vma
,
879 unsigned long address
)
881 struct mm_struct
*mm
= vma
->vm_mm
;
886 pte
= page_check_address(page
, mm
, address
, &ptl
, 1);
890 if (pte_dirty(*pte
) || pte_write(*pte
)) {
893 flush_cache_page(vma
, address
, pte_pfn(*pte
));
894 entry
= ptep_clear_flush_notify(vma
, address
, pte
);
895 entry
= pte_wrprotect(entry
);
896 entry
= pte_mkclean(entry
);
897 set_pte_at(mm
, address
, pte
, entry
);
901 pte_unmap_unlock(pte
, ptl
);
906 static int page_mkclean_file(struct address_space
*mapping
, struct page
*page
)
908 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
909 struct vm_area_struct
*vma
;
910 struct prio_tree_iter iter
;
913 BUG_ON(PageAnon(page
));
915 mutex_lock(&mapping
->i_mmap_mutex
);
916 vma_prio_tree_foreach(vma
, &iter
, &mapping
->i_mmap
, pgoff
, pgoff
) {
917 if (vma
->vm_flags
& VM_SHARED
) {
918 unsigned long address
= vma_address(page
, vma
);
919 if (address
== -EFAULT
)
921 ret
+= page_mkclean_one(page
, vma
, address
);
924 mutex_unlock(&mapping
->i_mmap_mutex
);
928 int page_mkclean(struct page
*page
)
932 BUG_ON(!PageLocked(page
));
934 if (page_mapped(page
)) {
935 struct address_space
*mapping
= page_mapping(page
);
937 ret
= page_mkclean_file(mapping
, page
);
938 if (page_test_and_clear_dirty(page_to_pfn(page
), 1))
945 EXPORT_SYMBOL_GPL(page_mkclean
);
948 * page_move_anon_rmap - move a page to our anon_vma
949 * @page: the page to move to our anon_vma
950 * @vma: the vma the page belongs to
951 * @address: the user virtual address mapped
953 * When a page belongs exclusively to one process after a COW event,
954 * that page can be moved into the anon_vma that belongs to just that
955 * process, so the rmap code will not search the parent or sibling
958 void page_move_anon_rmap(struct page
*page
,
959 struct vm_area_struct
*vma
, unsigned long address
)
961 struct anon_vma
*anon_vma
= vma
->anon_vma
;
963 VM_BUG_ON(!PageLocked(page
));
964 VM_BUG_ON(!anon_vma
);
965 VM_BUG_ON(page
->index
!= linear_page_index(vma
, address
));
967 anon_vma
= (void *) anon_vma
+ PAGE_MAPPING_ANON
;
968 page
->mapping
= (struct address_space
*) anon_vma
;
972 * __page_set_anon_rmap - set up new anonymous rmap
973 * @page: Page to add to rmap
974 * @vma: VM area to add page to.
975 * @address: User virtual address of the mapping
976 * @exclusive: the page is exclusively owned by the current process
978 static void __page_set_anon_rmap(struct page
*page
,
979 struct vm_area_struct
*vma
, unsigned long address
, int exclusive
)
981 struct anon_vma
*anon_vma
= vma
->anon_vma
;
989 * If the page isn't exclusively mapped into this vma,
990 * we must use the _oldest_ possible anon_vma for the
994 anon_vma
= anon_vma
->root
;
996 anon_vma
= (void *) anon_vma
+ PAGE_MAPPING_ANON
;
997 page
->mapping
= (struct address_space
*) anon_vma
;
998 page
->index
= linear_page_index(vma
, address
);
1002 * __page_check_anon_rmap - sanity check anonymous rmap addition
1003 * @page: the page to add the mapping to
1004 * @vma: the vm area in which the mapping is added
1005 * @address: the user virtual address mapped
1007 static void __page_check_anon_rmap(struct page
*page
,
1008 struct vm_area_struct
*vma
, unsigned long address
)
1010 #ifdef CONFIG_DEBUG_VM
1012 * The page's anon-rmap details (mapping and index) are guaranteed to
1013 * be set up correctly at this point.
1015 * We have exclusion against page_add_anon_rmap because the caller
1016 * always holds the page locked, except if called from page_dup_rmap,
1017 * in which case the page is already known to be setup.
1019 * We have exclusion against page_add_new_anon_rmap because those pages
1020 * are initially only visible via the pagetables, and the pte is locked
1021 * over the call to page_add_new_anon_rmap.
1023 BUG_ON(page_anon_vma(page
)->root
!= vma
->anon_vma
->root
);
1024 BUG_ON(page
->index
!= linear_page_index(vma
, address
));
1029 * page_add_anon_rmap - add pte mapping to an anonymous page
1030 * @page: the page to add the mapping to
1031 * @vma: the vm area in which the mapping is added
1032 * @address: the user virtual address mapped
1034 * The caller needs to hold the pte lock, and the page must be locked in
1035 * the anon_vma case: to serialize mapping,index checking after setting,
1036 * and to ensure that PageAnon is not being upgraded racily to PageKsm
1037 * (but PageKsm is never downgraded to PageAnon).
1039 void page_add_anon_rmap(struct page
*page
,
1040 struct vm_area_struct
*vma
, unsigned long address
)
1042 do_page_add_anon_rmap(page
, vma
, address
, 0);
1046 * Special version of the above for do_swap_page, which often runs
1047 * into pages that are exclusively owned by the current process.
1048 * Everybody else should continue to use page_add_anon_rmap above.
1050 void do_page_add_anon_rmap(struct page
*page
,
1051 struct vm_area_struct
*vma
, unsigned long address
, int exclusive
)
1053 int first
= atomic_inc_and_test(&page
->_mapcount
);
1055 if (!PageTransHuge(page
))
1056 __inc_zone_page_state(page
, NR_ANON_PAGES
);
1058 __inc_zone_page_state(page
,
1059 NR_ANON_TRANSPARENT_HUGEPAGES
);
1061 if (unlikely(PageKsm(page
)))
1064 VM_BUG_ON(!PageLocked(page
));
1065 /* address might be in next vma when migration races vma_adjust */
1067 __page_set_anon_rmap(page
, vma
, address
, exclusive
);
1069 __page_check_anon_rmap(page
, vma
, address
);
1073 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1074 * @page: the page to add the mapping to
1075 * @vma: the vm area in which the mapping is added
1076 * @address: the user virtual address mapped
1078 * Same as page_add_anon_rmap but must only be called on *new* pages.
1079 * This means the inc-and-test can be bypassed.
1080 * Page does not have to be locked.
1082 void page_add_new_anon_rmap(struct page
*page
,
1083 struct vm_area_struct
*vma
, unsigned long address
)
1085 VM_BUG_ON(address
< vma
->vm_start
|| address
>= vma
->vm_end
);
1086 SetPageSwapBacked(page
);
1087 atomic_set(&page
->_mapcount
, 0); /* increment count (starts at -1) */
1088 if (!PageTransHuge(page
))
1089 __inc_zone_page_state(page
, NR_ANON_PAGES
);
1091 __inc_zone_page_state(page
, NR_ANON_TRANSPARENT_HUGEPAGES
);
1092 __page_set_anon_rmap(page
, vma
, address
, 1);
1093 if (page_evictable(page
, vma
))
1094 lru_cache_add_lru(page
, LRU_ACTIVE_ANON
);
1096 add_page_to_unevictable_list(page
);
1100 * page_add_file_rmap - add pte mapping to a file page
1101 * @page: the page to add the mapping to
1103 * The caller needs to hold the pte lock.
1105 void page_add_file_rmap(struct page
*page
)
1107 if (atomic_inc_and_test(&page
->_mapcount
)) {
1108 __inc_zone_page_state(page
, NR_FILE_MAPPED
);
1109 mem_cgroup_inc_page_stat(page
, MEMCG_NR_FILE_MAPPED
);
1114 * page_remove_rmap - take down pte mapping from a page
1115 * @page: page to remove mapping from
1117 * The caller needs to hold the pte lock.
1119 void page_remove_rmap(struct page
*page
)
1121 /* page still mapped by someone else? */
1122 if (!atomic_add_negative(-1, &page
->_mapcount
))
1126 * Now that the last pte has gone, s390 must transfer dirty
1127 * flag from storage key to struct page. We can usually skip
1128 * this if the page is anon, so about to be freed; but perhaps
1129 * not if it's in swapcache - there might be another pte slot
1130 * containing the swap entry, but page not yet written to swap.
1132 if ((!PageAnon(page
) || PageSwapCache(page
)) &&
1133 page_test_and_clear_dirty(page_to_pfn(page
), 1))
1134 set_page_dirty(page
);
1136 * Hugepages are not counted in NR_ANON_PAGES nor NR_FILE_MAPPED
1137 * and not charged by memcg for now.
1139 if (unlikely(PageHuge(page
)))
1141 if (PageAnon(page
)) {
1142 mem_cgroup_uncharge_page(page
);
1143 if (!PageTransHuge(page
))
1144 __dec_zone_page_state(page
, NR_ANON_PAGES
);
1146 __dec_zone_page_state(page
,
1147 NR_ANON_TRANSPARENT_HUGEPAGES
);
1149 __dec_zone_page_state(page
, NR_FILE_MAPPED
);
1150 mem_cgroup_dec_page_stat(page
, MEMCG_NR_FILE_MAPPED
);
1153 * It would be tidy to reset the PageAnon mapping here,
1154 * but that might overwrite a racing page_add_anon_rmap
1155 * which increments mapcount after us but sets mapping
1156 * before us: so leave the reset to free_hot_cold_page,
1157 * and remember that it's only reliable while mapped.
1158 * Leaving it set also helps swapoff to reinstate ptes
1159 * faster for those pages still in swapcache.
1164 * Subfunctions of try_to_unmap: try_to_unmap_one called
1165 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
1167 int try_to_unmap_one(struct page
*page
, struct vm_area_struct
*vma
,
1168 unsigned long address
, enum ttu_flags flags
)
1170 struct mm_struct
*mm
= vma
->vm_mm
;
1174 int ret
= SWAP_AGAIN
;
1176 pte
= page_check_address(page
, mm
, address
, &ptl
, 0);
1181 * If the page is mlock()d, we cannot swap it out.
1182 * If it's recently referenced (perhaps page_referenced
1183 * skipped over this mm) then we should reactivate it.
1185 if (!(flags
& TTU_IGNORE_MLOCK
)) {
1186 if (vma
->vm_flags
& VM_LOCKED
)
1189 if (TTU_ACTION(flags
) == TTU_MUNLOCK
)
1192 if (!(flags
& TTU_IGNORE_ACCESS
)) {
1193 if (ptep_clear_flush_young_notify(vma
, address
, pte
)) {
1199 /* Nuke the page table entry. */
1200 flush_cache_page(vma
, address
, page_to_pfn(page
));
1201 pteval
= ptep_clear_flush_notify(vma
, address
, pte
);
1203 /* Move the dirty bit to the physical page now the pte is gone. */
1204 if (pte_dirty(pteval
))
1205 set_page_dirty(page
);
1207 /* Update high watermark before we lower rss */
1208 update_hiwater_rss(mm
);
1210 if (PageHWPoison(page
) && !(flags
& TTU_IGNORE_HWPOISON
)) {
1212 dec_mm_counter(mm
, MM_ANONPAGES
);
1214 dec_mm_counter(mm
, MM_FILEPAGES
);
1215 set_pte_at(mm
, address
, pte
,
1216 swp_entry_to_pte(make_hwpoison_entry(page
)));
1217 } else if (PageAnon(page
)) {
1218 swp_entry_t entry
= { .val
= page_private(page
) };
1220 if (PageSwapCache(page
)) {
1222 * Store the swap location in the pte.
1223 * See handle_pte_fault() ...
1225 if (swap_duplicate(entry
) < 0) {
1226 set_pte_at(mm
, address
, pte
, pteval
);
1230 if (list_empty(&mm
->mmlist
)) {
1231 spin_lock(&mmlist_lock
);
1232 if (list_empty(&mm
->mmlist
))
1233 list_add(&mm
->mmlist
, &init_mm
.mmlist
);
1234 spin_unlock(&mmlist_lock
);
1236 dec_mm_counter(mm
, MM_ANONPAGES
);
1237 inc_mm_counter(mm
, MM_SWAPENTS
);
1238 } else if (PAGE_MIGRATION
) {
1240 * Store the pfn of the page in a special migration
1241 * pte. do_swap_page() will wait until the migration
1242 * pte is removed and then restart fault handling.
1244 BUG_ON(TTU_ACTION(flags
) != TTU_MIGRATION
);
1245 entry
= make_migration_entry(page
, pte_write(pteval
));
1247 set_pte_at(mm
, address
, pte
, swp_entry_to_pte(entry
));
1248 BUG_ON(pte_file(*pte
));
1249 } else if (PAGE_MIGRATION
&& (TTU_ACTION(flags
) == TTU_MIGRATION
)) {
1250 /* Establish migration entry for a file page */
1252 entry
= make_migration_entry(page
, pte_write(pteval
));
1253 set_pte_at(mm
, address
, pte
, swp_entry_to_pte(entry
));
1255 dec_mm_counter(mm
, MM_FILEPAGES
);
1257 page_remove_rmap(page
);
1258 page_cache_release(page
);
1261 pte_unmap_unlock(pte
, ptl
);
1266 pte_unmap_unlock(pte
, ptl
);
1270 * We need mmap_sem locking, Otherwise VM_LOCKED check makes
1271 * unstable result and race. Plus, We can't wait here because
1272 * we now hold anon_vma->mutex or mapping->i_mmap_mutex.
1273 * if trylock failed, the page remain in evictable lru and later
1274 * vmscan could retry to move the page to unevictable lru if the
1275 * page is actually mlocked.
1277 if (down_read_trylock(&vma
->vm_mm
->mmap_sem
)) {
1278 if (vma
->vm_flags
& VM_LOCKED
) {
1279 mlock_vma_page(page
);
1282 up_read(&vma
->vm_mm
->mmap_sem
);
1288 * objrmap doesn't work for nonlinear VMAs because the assumption that
1289 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
1290 * Consequently, given a particular page and its ->index, we cannot locate the
1291 * ptes which are mapping that page without an exhaustive linear search.
1293 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
1294 * maps the file to which the target page belongs. The ->vm_private_data field
1295 * holds the current cursor into that scan. Successive searches will circulate
1296 * around the vma's virtual address space.
1298 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
1299 * more scanning pressure is placed against them as well. Eventually pages
1300 * will become fully unmapped and are eligible for eviction.
1302 * For very sparsely populated VMAs this is a little inefficient - chances are
1303 * there there won't be many ptes located within the scan cluster. In this case
1304 * maybe we could scan further - to the end of the pte page, perhaps.
1306 * Mlocked pages: check VM_LOCKED under mmap_sem held for read, if we can
1307 * acquire it without blocking. If vma locked, mlock the pages in the cluster,
1308 * rather than unmapping them. If we encounter the "check_page" that vmscan is
1309 * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN.
1311 #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
1312 #define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
1314 static int try_to_unmap_cluster(unsigned long cursor
, unsigned int *mapcount
,
1315 struct vm_area_struct
*vma
, struct page
*check_page
)
1317 struct mm_struct
*mm
= vma
->vm_mm
;
1325 unsigned long address
;
1327 int ret
= SWAP_AGAIN
;
1330 address
= (vma
->vm_start
+ cursor
) & CLUSTER_MASK
;
1331 end
= address
+ CLUSTER_SIZE
;
1332 if (address
< vma
->vm_start
)
1333 address
= vma
->vm_start
;
1334 if (end
> vma
->vm_end
)
1337 pgd
= pgd_offset(mm
, address
);
1338 if (!pgd_present(*pgd
))
1341 pud
= pud_offset(pgd
, address
);
1342 if (!pud_present(*pud
))
1345 pmd
= pmd_offset(pud
, address
);
1346 if (!pmd_present(*pmd
))
1350 * If we can acquire the mmap_sem for read, and vma is VM_LOCKED,
1351 * keep the sem while scanning the cluster for mlocking pages.
1353 if (down_read_trylock(&vma
->vm_mm
->mmap_sem
)) {
1354 locked_vma
= (vma
->vm_flags
& VM_LOCKED
);
1356 up_read(&vma
->vm_mm
->mmap_sem
); /* don't need it */
1359 pte
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
1361 /* Update high watermark before we lower rss */
1362 update_hiwater_rss(mm
);
1364 for (; address
< end
; pte
++, address
+= PAGE_SIZE
) {
1365 if (!pte_present(*pte
))
1367 page
= vm_normal_page(vma
, address
, *pte
);
1368 BUG_ON(!page
|| PageAnon(page
));
1371 mlock_vma_page(page
); /* no-op if already mlocked */
1372 if (page
== check_page
)
1374 continue; /* don't unmap */
1377 if (ptep_clear_flush_young_notify(vma
, address
, pte
))
1380 /* Nuke the page table entry. */
1381 flush_cache_page(vma
, address
, pte_pfn(*pte
));
1382 pteval
= ptep_clear_flush_notify(vma
, address
, pte
);
1384 /* If nonlinear, store the file page offset in the pte. */
1385 if (page
->index
!= linear_page_index(vma
, address
))
1386 set_pte_at(mm
, address
, pte
, pgoff_to_pte(page
->index
));
1388 /* Move the dirty bit to the physical page now the pte is gone. */
1389 if (pte_dirty(pteval
))
1390 set_page_dirty(page
);
1392 page_remove_rmap(page
);
1393 page_cache_release(page
);
1394 dec_mm_counter(mm
, MM_FILEPAGES
);
1397 pte_unmap_unlock(pte
- 1, ptl
);
1399 up_read(&vma
->vm_mm
->mmap_sem
);
1403 bool is_vma_temporary_stack(struct vm_area_struct
*vma
)
1405 int maybe_stack
= vma
->vm_flags
& (VM_GROWSDOWN
| VM_GROWSUP
);
1410 if ((vma
->vm_flags
& VM_STACK_INCOMPLETE_SETUP
) ==
1411 VM_STACK_INCOMPLETE_SETUP
)
1418 * try_to_unmap_anon - unmap or unlock anonymous page using the object-based
1420 * @page: the page to unmap/unlock
1421 * @flags: action and flags
1423 * Find all the mappings of a page using the mapping pointer and the vma chains
1424 * contained in the anon_vma struct it points to.
1426 * This function is only called from try_to_unmap/try_to_munlock for
1428 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1429 * where the page was found will be held for write. So, we won't recheck
1430 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1433 static int try_to_unmap_anon(struct page
*page
, enum ttu_flags flags
)
1435 struct anon_vma
*anon_vma
;
1436 struct anon_vma_chain
*avc
;
1437 int ret
= SWAP_AGAIN
;
1439 anon_vma
= page_lock_anon_vma(page
);
1443 list_for_each_entry(avc
, &anon_vma
->head
, same_anon_vma
) {
1444 struct vm_area_struct
*vma
= avc
->vma
;
1445 unsigned long address
;
1448 * During exec, a temporary VMA is setup and later moved.
1449 * The VMA is moved under the anon_vma lock but not the
1450 * page tables leading to a race where migration cannot
1451 * find the migration ptes. Rather than increasing the
1452 * locking requirements of exec(), migration skips
1453 * temporary VMAs until after exec() completes.
1455 if (PAGE_MIGRATION
&& (flags
& TTU_MIGRATION
) &&
1456 is_vma_temporary_stack(vma
))
1459 address
= vma_address(page
, vma
);
1460 if (address
== -EFAULT
)
1462 ret
= try_to_unmap_one(page
, vma
, address
, flags
);
1463 if (ret
!= SWAP_AGAIN
|| !page_mapped(page
))
1467 page_unlock_anon_vma(anon_vma
);
1472 * try_to_unmap_file - unmap/unlock file page using the object-based rmap method
1473 * @page: the page to unmap/unlock
1474 * @flags: action and flags
1476 * Find all the mappings of a page using the mapping pointer and the vma chains
1477 * contained in the address_space struct it points to.
1479 * This function is only called from try_to_unmap/try_to_munlock for
1480 * object-based pages.
1481 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1482 * where the page was found will be held for write. So, we won't recheck
1483 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1486 static int try_to_unmap_file(struct page
*page
, enum ttu_flags flags
)
1488 struct address_space
*mapping
= page
->mapping
;
1489 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
1490 struct vm_area_struct
*vma
;
1491 struct prio_tree_iter iter
;
1492 int ret
= SWAP_AGAIN
;
1493 unsigned long cursor
;
1494 unsigned long max_nl_cursor
= 0;
1495 unsigned long max_nl_size
= 0;
1496 unsigned int mapcount
;
1498 mutex_lock(&mapping
->i_mmap_mutex
);
1499 vma_prio_tree_foreach(vma
, &iter
, &mapping
->i_mmap
, pgoff
, pgoff
) {
1500 unsigned long address
= vma_address(page
, vma
);
1501 if (address
== -EFAULT
)
1503 ret
= try_to_unmap_one(page
, vma
, address
, flags
);
1504 if (ret
!= SWAP_AGAIN
|| !page_mapped(page
))
1508 if (list_empty(&mapping
->i_mmap_nonlinear
))
1512 * We don't bother to try to find the munlocked page in nonlinears.
1513 * It's costly. Instead, later, page reclaim logic may call
1514 * try_to_unmap(TTU_MUNLOCK) and recover PG_mlocked lazily.
1516 if (TTU_ACTION(flags
) == TTU_MUNLOCK
)
1519 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
,
1520 shared
.vm_set
.list
) {
1521 cursor
= (unsigned long) vma
->vm_private_data
;
1522 if (cursor
> max_nl_cursor
)
1523 max_nl_cursor
= cursor
;
1524 cursor
= vma
->vm_end
- vma
->vm_start
;
1525 if (cursor
> max_nl_size
)
1526 max_nl_size
= cursor
;
1529 if (max_nl_size
== 0) { /* all nonlinears locked or reserved ? */
1535 * We don't try to search for this page in the nonlinear vmas,
1536 * and page_referenced wouldn't have found it anyway. Instead
1537 * just walk the nonlinear vmas trying to age and unmap some.
1538 * The mapcount of the page we came in with is irrelevant,
1539 * but even so use it as a guide to how hard we should try?
1541 mapcount
= page_mapcount(page
);
1546 max_nl_size
= (max_nl_size
+ CLUSTER_SIZE
- 1) & CLUSTER_MASK
;
1547 if (max_nl_cursor
== 0)
1548 max_nl_cursor
= CLUSTER_SIZE
;
1551 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
,
1552 shared
.vm_set
.list
) {
1553 cursor
= (unsigned long) vma
->vm_private_data
;
1554 while ( cursor
< max_nl_cursor
&&
1555 cursor
< vma
->vm_end
- vma
->vm_start
) {
1556 if (try_to_unmap_cluster(cursor
, &mapcount
,
1557 vma
, page
) == SWAP_MLOCK
)
1559 cursor
+= CLUSTER_SIZE
;
1560 vma
->vm_private_data
= (void *) cursor
;
1561 if ((int)mapcount
<= 0)
1564 vma
->vm_private_data
= (void *) max_nl_cursor
;
1567 max_nl_cursor
+= CLUSTER_SIZE
;
1568 } while (max_nl_cursor
<= max_nl_size
);
1571 * Don't loop forever (perhaps all the remaining pages are
1572 * in locked vmas). Reset cursor on all unreserved nonlinear
1573 * vmas, now forgetting on which ones it had fallen behind.
1575 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
, shared
.vm_set
.list
)
1576 vma
->vm_private_data
= NULL
;
1578 mutex_unlock(&mapping
->i_mmap_mutex
);
1583 * try_to_unmap - try to remove all page table mappings to a page
1584 * @page: the page to get unmapped
1585 * @flags: action and flags
1587 * Tries to remove all the page table entries which are mapping this
1588 * page, used in the pageout path. Caller must hold the page lock.
1589 * Return values are:
1591 * SWAP_SUCCESS - we succeeded in removing all mappings
1592 * SWAP_AGAIN - we missed a mapping, try again later
1593 * SWAP_FAIL - the page is unswappable
1594 * SWAP_MLOCK - page is mlocked.
1596 int try_to_unmap(struct page
*page
, enum ttu_flags flags
)
1600 BUG_ON(!PageLocked(page
));
1601 VM_BUG_ON(!PageHuge(page
) && PageTransHuge(page
));
1603 if (unlikely(PageKsm(page
)))
1604 ret
= try_to_unmap_ksm(page
, flags
);
1605 else if (PageAnon(page
))
1606 ret
= try_to_unmap_anon(page
, flags
);
1608 ret
= try_to_unmap_file(page
, flags
);
1609 if (ret
!= SWAP_MLOCK
&& !page_mapped(page
))
1615 * try_to_munlock - try to munlock a page
1616 * @page: the page to be munlocked
1618 * Called from munlock code. Checks all of the VMAs mapping the page
1619 * to make sure nobody else has this page mlocked. The page will be
1620 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1622 * Return values are:
1624 * SWAP_AGAIN - no vma is holding page mlocked, or,
1625 * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem
1626 * SWAP_FAIL - page cannot be located at present
1627 * SWAP_MLOCK - page is now mlocked.
1629 int try_to_munlock(struct page
*page
)
1631 VM_BUG_ON(!PageLocked(page
) || PageLRU(page
));
1633 if (unlikely(PageKsm(page
)))
1634 return try_to_unmap_ksm(page
, TTU_MUNLOCK
);
1635 else if (PageAnon(page
))
1636 return try_to_unmap_anon(page
, TTU_MUNLOCK
);
1638 return try_to_unmap_file(page
, TTU_MUNLOCK
);
1641 void __put_anon_vma(struct anon_vma
*anon_vma
)
1643 struct anon_vma
*root
= anon_vma
->root
;
1645 if (root
!= anon_vma
&& atomic_dec_and_test(&root
->refcount
))
1646 anon_vma_free(root
);
1648 anon_vma_free(anon_vma
);
1651 #ifdef CONFIG_MIGRATION
1653 * rmap_walk() and its helpers rmap_walk_anon() and rmap_walk_file():
1654 * Called by migrate.c to remove migration ptes, but might be used more later.
1656 static int rmap_walk_anon(struct page
*page
, int (*rmap_one
)(struct page
*,
1657 struct vm_area_struct
*, unsigned long, void *), void *arg
)
1659 struct anon_vma
*anon_vma
;
1660 struct anon_vma_chain
*avc
;
1661 int ret
= SWAP_AGAIN
;
1664 * Note: remove_migration_ptes() cannot use page_lock_anon_vma()
1665 * because that depends on page_mapped(); but not all its usages
1666 * are holding mmap_sem. Users without mmap_sem are required to
1667 * take a reference count to prevent the anon_vma disappearing
1669 anon_vma
= page_anon_vma(page
);
1672 anon_vma_lock(anon_vma
);
1673 list_for_each_entry(avc
, &anon_vma
->head
, same_anon_vma
) {
1674 struct vm_area_struct
*vma
= avc
->vma
;
1675 unsigned long address
= vma_address(page
, vma
);
1676 if (address
== -EFAULT
)
1678 ret
= rmap_one(page
, vma
, address
, arg
);
1679 if (ret
!= SWAP_AGAIN
)
1682 anon_vma_unlock(anon_vma
);
1686 static int rmap_walk_file(struct page
*page
, int (*rmap_one
)(struct page
*,
1687 struct vm_area_struct
*, unsigned long, void *), void *arg
)
1689 struct address_space
*mapping
= page
->mapping
;
1690 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
1691 struct vm_area_struct
*vma
;
1692 struct prio_tree_iter iter
;
1693 int ret
= SWAP_AGAIN
;
1697 mutex_lock(&mapping
->i_mmap_mutex
);
1698 vma_prio_tree_foreach(vma
, &iter
, &mapping
->i_mmap
, pgoff
, pgoff
) {
1699 unsigned long address
= vma_address(page
, vma
);
1700 if (address
== -EFAULT
)
1702 ret
= rmap_one(page
, vma
, address
, arg
);
1703 if (ret
!= SWAP_AGAIN
)
1707 * No nonlinear handling: being always shared, nonlinear vmas
1708 * never contain migration ptes. Decide what to do about this
1709 * limitation to linear when we need rmap_walk() on nonlinear.
1711 mutex_unlock(&mapping
->i_mmap_mutex
);
1715 int rmap_walk(struct page
*page
, int (*rmap_one
)(struct page
*,
1716 struct vm_area_struct
*, unsigned long, void *), void *arg
)
1718 VM_BUG_ON(!PageLocked(page
));
1720 if (unlikely(PageKsm(page
)))
1721 return rmap_walk_ksm(page
, rmap_one
, arg
);
1722 else if (PageAnon(page
))
1723 return rmap_walk_anon(page
, rmap_one
, arg
);
1725 return rmap_walk_file(page
, rmap_one
, arg
);
1727 #endif /* CONFIG_MIGRATION */
1729 #ifdef CONFIG_HUGETLB_PAGE
1731 * The following three functions are for anonymous (private mapped) hugepages.
1732 * Unlike common anonymous pages, anonymous hugepages have no accounting code
1733 * and no lru code, because we handle hugepages differently from common pages.
1735 static void __hugepage_set_anon_rmap(struct page
*page
,
1736 struct vm_area_struct
*vma
, unsigned long address
, int exclusive
)
1738 struct anon_vma
*anon_vma
= vma
->anon_vma
;
1745 anon_vma
= anon_vma
->root
;
1747 anon_vma
= (void *) anon_vma
+ PAGE_MAPPING_ANON
;
1748 page
->mapping
= (struct address_space
*) anon_vma
;
1749 page
->index
= linear_page_index(vma
, address
);
1752 void hugepage_add_anon_rmap(struct page
*page
,
1753 struct vm_area_struct
*vma
, unsigned long address
)
1755 struct anon_vma
*anon_vma
= vma
->anon_vma
;
1758 BUG_ON(!PageLocked(page
));
1760 /* address might be in next vma when migration races vma_adjust */
1761 first
= atomic_inc_and_test(&page
->_mapcount
);
1763 __hugepage_set_anon_rmap(page
, vma
, address
, 0);
1766 void hugepage_add_new_anon_rmap(struct page
*page
,
1767 struct vm_area_struct
*vma
, unsigned long address
)
1769 BUG_ON(address
< vma
->vm_start
|| address
>= vma
->vm_end
);
1770 atomic_set(&page
->_mapcount
, 0);
1771 __hugepage_set_anon_rmap(page
, vma
, address
, 1);
1773 #endif /* CONFIG_HUGETLB_PAGE */