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)
25 * page->flags PG_locked (lock_page)
26 * mapping->i_mmap_mutex
28 * mm->page_table_lock or pte_lock
29 * zone->lru_lock (in mark_page_accessed, isolate_lru_page)
30 * swap_lock (in swap_duplicate, swap_info_get)
31 * mmlist_lock (in mmput, drain_mmlist and others)
32 * mapping->private_lock (in __set_page_dirty_buffers)
33 * inode->i_lock (in set_page_dirty's __mark_inode_dirty)
34 * bdi.wb->list_lock (in set_page_dirty's __mark_inode_dirty)
35 * sb_lock (within inode_lock in fs/fs-writeback.c)
36 * mapping->tree_lock (widely used, in set_page_dirty,
37 * in arch-dependent flush_dcache_mmap_lock,
38 * within bdi.wb->list_lock in __sync_single_inode)
40 * anon_vma->mutex,mapping->i_mutex (memory_failure, collect_procs_anon)
46 #include <linux/pagemap.h>
47 #include <linux/swap.h>
48 #include <linux/swapops.h>
49 #include <linux/slab.h>
50 #include <linux/init.h>
51 #include <linux/ksm.h>
52 #include <linux/rmap.h>
53 #include <linux/rcupdate.h>
54 #include <linux/export.h>
55 #include <linux/memcontrol.h>
56 #include <linux/mmu_notifier.h>
57 #include <linux/migrate.h>
58 #include <linux/hugetlb.h>
60 #include <asm/tlbflush.h>
64 static struct kmem_cache
*anon_vma_cachep
;
65 static struct kmem_cache
*anon_vma_chain_cachep
;
67 static inline struct anon_vma
*anon_vma_alloc(void)
69 struct anon_vma
*anon_vma
;
71 anon_vma
= kmem_cache_alloc(anon_vma_cachep
, GFP_KERNEL
);
73 atomic_set(&anon_vma
->refcount
, 1);
75 * Initialise the anon_vma root to point to itself. If called
76 * from fork, the root will be reset to the parents anon_vma.
78 anon_vma
->root
= anon_vma
;
84 static inline void anon_vma_free(struct anon_vma
*anon_vma
)
86 VM_BUG_ON(atomic_read(&anon_vma
->refcount
));
89 * Synchronize against page_lock_anon_vma() such that
90 * we can safely hold the lock without the anon_vma getting
93 * Relies on the full mb implied by the atomic_dec_and_test() from
94 * put_anon_vma() against the acquire barrier implied by
95 * mutex_trylock() from page_lock_anon_vma(). This orders:
97 * page_lock_anon_vma() VS put_anon_vma()
98 * mutex_trylock() atomic_dec_and_test()
100 * atomic_read() mutex_is_locked()
102 * LOCK should suffice since the actual taking of the lock must
103 * happen _before_ what follows.
105 if (mutex_is_locked(&anon_vma
->root
->mutex
)) {
106 anon_vma_lock(anon_vma
);
107 anon_vma_unlock(anon_vma
);
110 kmem_cache_free(anon_vma_cachep
, anon_vma
);
113 static inline struct anon_vma_chain
*anon_vma_chain_alloc(gfp_t gfp
)
115 return kmem_cache_alloc(anon_vma_chain_cachep
, gfp
);
118 static void anon_vma_chain_free(struct anon_vma_chain
*anon_vma_chain
)
120 kmem_cache_free(anon_vma_chain_cachep
, anon_vma_chain
);
123 static void anon_vma_chain_link(struct vm_area_struct
*vma
,
124 struct anon_vma_chain
*avc
,
125 struct anon_vma
*anon_vma
)
128 avc
->anon_vma
= anon_vma
;
129 list_add(&avc
->same_vma
, &vma
->anon_vma_chain
);
130 anon_vma_interval_tree_insert(avc
, &anon_vma
->rb_root
);
134 * anon_vma_prepare - attach an anon_vma to a memory region
135 * @vma: the memory region in question
137 * This makes sure the memory mapping described by 'vma' has
138 * an 'anon_vma' attached to it, so that we can associate the
139 * anonymous pages mapped into it with that anon_vma.
141 * The common case will be that we already have one, but if
142 * not we either need to find an adjacent mapping that we
143 * can re-use the anon_vma from (very common when the only
144 * reason for splitting a vma has been mprotect()), or we
145 * allocate a new one.
147 * Anon-vma allocations are very subtle, because we may have
148 * optimistically looked up an anon_vma in page_lock_anon_vma()
149 * and that may actually touch the spinlock even in the newly
150 * allocated vma (it depends on RCU to make sure that the
151 * anon_vma isn't actually destroyed).
153 * As a result, we need to do proper anon_vma locking even
154 * for the new allocation. At the same time, we do not want
155 * to do any locking for the common case of already having
158 * This must be called with the mmap_sem held for reading.
160 int anon_vma_prepare(struct vm_area_struct
*vma
)
162 struct anon_vma
*anon_vma
= vma
->anon_vma
;
163 struct anon_vma_chain
*avc
;
166 if (unlikely(!anon_vma
)) {
167 struct mm_struct
*mm
= vma
->vm_mm
;
168 struct anon_vma
*allocated
;
170 avc
= anon_vma_chain_alloc(GFP_KERNEL
);
174 anon_vma
= find_mergeable_anon_vma(vma
);
177 anon_vma
= anon_vma_alloc();
178 if (unlikely(!anon_vma
))
179 goto out_enomem_free_avc
;
180 allocated
= anon_vma
;
183 anon_vma_lock(anon_vma
);
184 /* page_table_lock to protect against threads */
185 spin_lock(&mm
->page_table_lock
);
186 if (likely(!vma
->anon_vma
)) {
187 vma
->anon_vma
= anon_vma
;
188 anon_vma_chain_link(vma
, avc
, anon_vma
);
192 spin_unlock(&mm
->page_table_lock
);
193 anon_vma_unlock(anon_vma
);
195 if (unlikely(allocated
))
196 put_anon_vma(allocated
);
198 anon_vma_chain_free(avc
);
203 anon_vma_chain_free(avc
);
209 * This is a useful helper function for locking the anon_vma root as
210 * we traverse the vma->anon_vma_chain, looping over anon_vma's that
213 * Such anon_vma's should have the same root, so you'd expect to see
214 * just a single mutex_lock for the whole traversal.
216 static inline struct anon_vma
*lock_anon_vma_root(struct anon_vma
*root
, struct anon_vma
*anon_vma
)
218 struct anon_vma
*new_root
= anon_vma
->root
;
219 if (new_root
!= root
) {
220 if (WARN_ON_ONCE(root
))
221 mutex_unlock(&root
->mutex
);
223 mutex_lock(&root
->mutex
);
228 static inline void unlock_anon_vma_root(struct anon_vma
*root
)
231 mutex_unlock(&root
->mutex
);
235 * Attach the anon_vmas from src to dst.
236 * Returns 0 on success, -ENOMEM on failure.
238 int anon_vma_clone(struct vm_area_struct
*dst
, struct vm_area_struct
*src
)
240 struct anon_vma_chain
*avc
, *pavc
;
241 struct anon_vma
*root
= NULL
;
243 list_for_each_entry_reverse(pavc
, &src
->anon_vma_chain
, same_vma
) {
244 struct anon_vma
*anon_vma
;
246 avc
= anon_vma_chain_alloc(GFP_NOWAIT
| __GFP_NOWARN
);
247 if (unlikely(!avc
)) {
248 unlock_anon_vma_root(root
);
250 avc
= anon_vma_chain_alloc(GFP_KERNEL
);
254 anon_vma
= pavc
->anon_vma
;
255 root
= lock_anon_vma_root(root
, anon_vma
);
256 anon_vma_chain_link(dst
, avc
, anon_vma
);
258 unlock_anon_vma_root(root
);
262 unlink_anon_vmas(dst
);
267 * Attach vma to its own anon_vma, as well as to the anon_vmas that
268 * the corresponding VMA in the parent process is attached to.
269 * Returns 0 on success, non-zero on failure.
271 int anon_vma_fork(struct vm_area_struct
*vma
, struct vm_area_struct
*pvma
)
273 struct anon_vma_chain
*avc
;
274 struct anon_vma
*anon_vma
;
276 /* Don't bother if the parent process has no anon_vma here. */
281 * First, attach the new VMA to the parent VMA's anon_vmas,
282 * so rmap can find non-COWed pages in child processes.
284 if (anon_vma_clone(vma
, pvma
))
287 /* Then add our own anon_vma. */
288 anon_vma
= anon_vma_alloc();
291 avc
= anon_vma_chain_alloc(GFP_KERNEL
);
293 goto out_error_free_anon_vma
;
296 * The root anon_vma's spinlock is the lock actually used when we
297 * lock any of the anon_vmas in this anon_vma tree.
299 anon_vma
->root
= pvma
->anon_vma
->root
;
301 * With refcounts, an anon_vma can stay around longer than the
302 * process it belongs to. The root anon_vma needs to be pinned until
303 * this anon_vma is freed, because the lock lives in the root.
305 get_anon_vma(anon_vma
->root
);
306 /* Mark this anon_vma as the one where our new (COWed) pages go. */
307 vma
->anon_vma
= anon_vma
;
308 anon_vma_lock(anon_vma
);
309 anon_vma_chain_link(vma
, avc
, anon_vma
);
310 anon_vma_unlock(anon_vma
);
314 out_error_free_anon_vma
:
315 put_anon_vma(anon_vma
);
317 unlink_anon_vmas(vma
);
321 void unlink_anon_vmas(struct vm_area_struct
*vma
)
323 struct anon_vma_chain
*avc
, *next
;
324 struct anon_vma
*root
= NULL
;
327 * Unlink each anon_vma chained to the VMA. This list is ordered
328 * from newest to oldest, ensuring the root anon_vma gets freed last.
330 list_for_each_entry_safe(avc
, next
, &vma
->anon_vma_chain
, same_vma
) {
331 struct anon_vma
*anon_vma
= avc
->anon_vma
;
333 root
= lock_anon_vma_root(root
, anon_vma
);
334 anon_vma_interval_tree_remove(avc
, &anon_vma
->rb_root
);
337 * Leave empty anon_vmas on the list - we'll need
338 * to free them outside the lock.
340 if (RB_EMPTY_ROOT(&anon_vma
->rb_root
))
343 list_del(&avc
->same_vma
);
344 anon_vma_chain_free(avc
);
346 unlock_anon_vma_root(root
);
349 * Iterate the list once more, it now only contains empty and unlinked
350 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
351 * needing to acquire the anon_vma->root->mutex.
353 list_for_each_entry_safe(avc
, next
, &vma
->anon_vma_chain
, same_vma
) {
354 struct anon_vma
*anon_vma
= avc
->anon_vma
;
356 put_anon_vma(anon_vma
);
358 list_del(&avc
->same_vma
);
359 anon_vma_chain_free(avc
);
363 static void anon_vma_ctor(void *data
)
365 struct anon_vma
*anon_vma
= data
;
367 mutex_init(&anon_vma
->mutex
);
368 atomic_set(&anon_vma
->refcount
, 0);
369 anon_vma
->rb_root
= RB_ROOT
;
372 void __init
anon_vma_init(void)
374 anon_vma_cachep
= kmem_cache_create("anon_vma", sizeof(struct anon_vma
),
375 0, SLAB_DESTROY_BY_RCU
|SLAB_PANIC
, anon_vma_ctor
);
376 anon_vma_chain_cachep
= KMEM_CACHE(anon_vma_chain
, SLAB_PANIC
);
380 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
382 * Since there is no serialization what so ever against page_remove_rmap()
383 * the best this function can do is return a locked anon_vma that might
384 * have been relevant to this page.
386 * The page might have been remapped to a different anon_vma or the anon_vma
387 * returned may already be freed (and even reused).
389 * In case it was remapped to a different anon_vma, the new anon_vma will be a
390 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
391 * ensure that any anon_vma obtained from the page will still be valid for as
392 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
394 * All users of this function must be very careful when walking the anon_vma
395 * chain and verify that the page in question is indeed mapped in it
396 * [ something equivalent to page_mapped_in_vma() ].
398 * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap()
399 * that the anon_vma pointer from page->mapping is valid if there is a
400 * mapcount, we can dereference the anon_vma after observing those.
402 struct anon_vma
*page_get_anon_vma(struct page
*page
)
404 struct anon_vma
*anon_vma
= NULL
;
405 unsigned long anon_mapping
;
408 anon_mapping
= (unsigned long) ACCESS_ONCE(page
->mapping
);
409 if ((anon_mapping
& PAGE_MAPPING_FLAGS
) != PAGE_MAPPING_ANON
)
411 if (!page_mapped(page
))
414 anon_vma
= (struct anon_vma
*) (anon_mapping
- PAGE_MAPPING_ANON
);
415 if (!atomic_inc_not_zero(&anon_vma
->refcount
)) {
421 * If this page is still mapped, then its anon_vma cannot have been
422 * freed. But if it has been unmapped, we have no security against the
423 * anon_vma structure being freed and reused (for another anon_vma:
424 * SLAB_DESTROY_BY_RCU guarantees that - so the atomic_inc_not_zero()
425 * above cannot corrupt).
427 if (!page_mapped(page
)) {
428 put_anon_vma(anon_vma
);
438 * Similar to page_get_anon_vma() except it locks the anon_vma.
440 * Its a little more complex as it tries to keep the fast path to a single
441 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
442 * reference like with page_get_anon_vma() and then block on the mutex.
444 struct anon_vma
*page_lock_anon_vma(struct page
*page
)
446 struct anon_vma
*anon_vma
= NULL
;
447 struct anon_vma
*root_anon_vma
;
448 unsigned long anon_mapping
;
451 anon_mapping
= (unsigned long) ACCESS_ONCE(page
->mapping
);
452 if ((anon_mapping
& PAGE_MAPPING_FLAGS
) != PAGE_MAPPING_ANON
)
454 if (!page_mapped(page
))
457 anon_vma
= (struct anon_vma
*) (anon_mapping
- PAGE_MAPPING_ANON
);
458 root_anon_vma
= ACCESS_ONCE(anon_vma
->root
);
459 if (mutex_trylock(&root_anon_vma
->mutex
)) {
461 * If the page is still mapped, then this anon_vma is still
462 * its anon_vma, and holding the mutex ensures that it will
463 * not go away, see anon_vma_free().
465 if (!page_mapped(page
)) {
466 mutex_unlock(&root_anon_vma
->mutex
);
472 /* trylock failed, we got to sleep */
473 if (!atomic_inc_not_zero(&anon_vma
->refcount
)) {
478 if (!page_mapped(page
)) {
479 put_anon_vma(anon_vma
);
484 /* we pinned the anon_vma, its safe to sleep */
486 anon_vma_lock(anon_vma
);
488 if (atomic_dec_and_test(&anon_vma
->refcount
)) {
490 * Oops, we held the last refcount, release the lock
491 * and bail -- can't simply use put_anon_vma() because
492 * we'll deadlock on the anon_vma_lock() recursion.
494 anon_vma_unlock(anon_vma
);
495 __put_anon_vma(anon_vma
);
506 void page_unlock_anon_vma(struct anon_vma
*anon_vma
)
508 anon_vma_unlock(anon_vma
);
512 * At what user virtual address is page expected in @vma?
514 static inline unsigned long
515 __vma_address(struct page
*page
, struct vm_area_struct
*vma
)
517 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
519 if (unlikely(is_vm_hugetlb_page(vma
)))
520 pgoff
= page
->index
<< huge_page_order(page_hstate(page
));
522 return vma
->vm_start
+ ((pgoff
- vma
->vm_pgoff
) << PAGE_SHIFT
);
526 vma_address(struct page
*page
, struct vm_area_struct
*vma
)
528 unsigned long address
= __vma_address(page
, vma
);
530 /* page should be within @vma mapping range */
531 VM_BUG_ON(address
< vma
->vm_start
|| address
>= vma
->vm_end
);
537 * At what user virtual address is page expected in vma?
538 * Caller should check the page is actually part of the vma.
540 unsigned long page_address_in_vma(struct page
*page
, struct vm_area_struct
*vma
)
542 unsigned long address
;
543 if (PageAnon(page
)) {
544 struct anon_vma
*page__anon_vma
= page_anon_vma(page
);
546 * Note: swapoff's unuse_vma() is more efficient with this
547 * check, and needs it to match anon_vma when KSM is active.
549 if (!vma
->anon_vma
|| !page__anon_vma
||
550 vma
->anon_vma
->root
!= page__anon_vma
->root
)
552 } else if (page
->mapping
&& !(vma
->vm_flags
& VM_NONLINEAR
)) {
554 vma
->vm_file
->f_mapping
!= page
->mapping
)
558 address
= __vma_address(page
, vma
);
559 if (unlikely(address
< vma
->vm_start
|| address
>= vma
->vm_end
))
565 * Check that @page is mapped at @address into @mm.
567 * If @sync is false, page_check_address may perform a racy check to avoid
568 * the page table lock when the pte is not present (helpful when reclaiming
569 * highly shared pages).
571 * On success returns with pte mapped and locked.
573 pte_t
*__page_check_address(struct page
*page
, struct mm_struct
*mm
,
574 unsigned long address
, spinlock_t
**ptlp
, int sync
)
582 if (unlikely(PageHuge(page
))) {
583 pte
= huge_pte_offset(mm
, address
);
584 ptl
= &mm
->page_table_lock
;
588 pgd
= pgd_offset(mm
, address
);
589 if (!pgd_present(*pgd
))
592 pud
= pud_offset(pgd
, address
);
593 if (!pud_present(*pud
))
596 pmd
= pmd_offset(pud
, address
);
597 if (!pmd_present(*pmd
))
599 if (pmd_trans_huge(*pmd
))
602 pte
= pte_offset_map(pmd
, address
);
603 /* Make a quick check before getting the lock */
604 if (!sync
&& !pte_present(*pte
)) {
609 ptl
= pte_lockptr(mm
, pmd
);
612 if (pte_present(*pte
) && page_to_pfn(page
) == pte_pfn(*pte
)) {
616 pte_unmap_unlock(pte
, ptl
);
621 * page_mapped_in_vma - check whether a page is really mapped in a VMA
622 * @page: the page to test
623 * @vma: the VMA to test
625 * Returns 1 if the page is mapped into the page tables of the VMA, 0
626 * if the page is not mapped into the page tables of this VMA. Only
627 * valid for normal file or anonymous VMAs.
629 int page_mapped_in_vma(struct page
*page
, struct vm_area_struct
*vma
)
631 unsigned long address
;
635 address
= __vma_address(page
, vma
);
636 if (unlikely(address
< vma
->vm_start
|| address
>= vma
->vm_end
))
638 pte
= page_check_address(page
, vma
->vm_mm
, address
, &ptl
, 1);
639 if (!pte
) /* the page is not in this mm */
641 pte_unmap_unlock(pte
, ptl
);
647 * Subfunctions of page_referenced: page_referenced_one called
648 * repeatedly from either page_referenced_anon or page_referenced_file.
650 int page_referenced_one(struct page
*page
, struct vm_area_struct
*vma
,
651 unsigned long address
, unsigned int *mapcount
,
652 unsigned long *vm_flags
)
654 struct mm_struct
*mm
= vma
->vm_mm
;
657 if (unlikely(PageTransHuge(page
))) {
660 spin_lock(&mm
->page_table_lock
);
662 * rmap might return false positives; we must filter
663 * these out using page_check_address_pmd().
665 pmd
= page_check_address_pmd(page
, mm
, address
,
666 PAGE_CHECK_ADDRESS_PMD_FLAG
);
668 spin_unlock(&mm
->page_table_lock
);
672 if (vma
->vm_flags
& VM_LOCKED
) {
673 spin_unlock(&mm
->page_table_lock
);
674 *mapcount
= 0; /* break early from loop */
675 *vm_flags
|= VM_LOCKED
;
679 /* go ahead even if the pmd is pmd_trans_splitting() */
680 if (pmdp_clear_flush_young_notify(vma
, address
, pmd
))
682 spin_unlock(&mm
->page_table_lock
);
688 * rmap might return false positives; we must filter
689 * these out using page_check_address().
691 pte
= page_check_address(page
, mm
, address
, &ptl
, 0);
695 if (vma
->vm_flags
& VM_LOCKED
) {
696 pte_unmap_unlock(pte
, ptl
);
697 *mapcount
= 0; /* break early from loop */
698 *vm_flags
|= VM_LOCKED
;
702 if (ptep_clear_flush_young_notify(vma
, address
, pte
)) {
704 * Don't treat a reference through a sequentially read
705 * mapping as such. If the page has been used in
706 * another mapping, we will catch it; if this other
707 * mapping is already gone, the unmap path will have
708 * set PG_referenced or activated the page.
710 if (likely(!VM_SequentialReadHint(vma
)))
713 pte_unmap_unlock(pte
, ptl
);
719 *vm_flags
|= vma
->vm_flags
;
724 static int page_referenced_anon(struct page
*page
,
725 struct mem_cgroup
*memcg
,
726 unsigned long *vm_flags
)
728 unsigned int mapcount
;
729 struct anon_vma
*anon_vma
;
731 struct anon_vma_chain
*avc
;
734 anon_vma
= page_lock_anon_vma(page
);
738 mapcount
= page_mapcount(page
);
739 pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
740 anon_vma_interval_tree_foreach(avc
, &anon_vma
->rb_root
, pgoff
, pgoff
) {
741 struct vm_area_struct
*vma
= avc
->vma
;
742 unsigned long address
= vma_address(page
, vma
);
744 * If we are reclaiming on behalf of a cgroup, skip
745 * counting on behalf of references from different
748 if (memcg
&& !mm_match_cgroup(vma
->vm_mm
, memcg
))
750 referenced
+= page_referenced_one(page
, vma
, address
,
751 &mapcount
, vm_flags
);
756 page_unlock_anon_vma(anon_vma
);
761 * page_referenced_file - referenced check for object-based rmap
762 * @page: the page we're checking references on.
763 * @memcg: target memory control group
764 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
766 * For an object-based mapped page, find all the places it is mapped and
767 * check/clear the referenced flag. This is done by following the page->mapping
768 * pointer, then walking the chain of vmas it holds. It returns the number
769 * of references it found.
771 * This function is only called from page_referenced for object-based pages.
773 static int page_referenced_file(struct page
*page
,
774 struct mem_cgroup
*memcg
,
775 unsigned long *vm_flags
)
777 unsigned int mapcount
;
778 struct address_space
*mapping
= page
->mapping
;
779 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
780 struct vm_area_struct
*vma
;
784 * The caller's checks on page->mapping and !PageAnon have made
785 * sure that this is a file page: the check for page->mapping
786 * excludes the case just before it gets set on an anon page.
788 BUG_ON(PageAnon(page
));
791 * The page lock not only makes sure that page->mapping cannot
792 * suddenly be NULLified by truncation, it makes sure that the
793 * structure at mapping cannot be freed and reused yet,
794 * so we can safely take mapping->i_mmap_mutex.
796 BUG_ON(!PageLocked(page
));
798 mutex_lock(&mapping
->i_mmap_mutex
);
801 * i_mmap_mutex does not stabilize mapcount at all, but mapcount
802 * is more likely to be accurate if we note it after spinning.
804 mapcount
= page_mapcount(page
);
806 vma_interval_tree_foreach(vma
, &mapping
->i_mmap
, pgoff
, pgoff
) {
807 unsigned long address
= vma_address(page
, vma
);
809 * If we are reclaiming on behalf of a cgroup, skip
810 * counting on behalf of references from different
813 if (memcg
&& !mm_match_cgroup(vma
->vm_mm
, memcg
))
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 * @memcg: target memory cgroup
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
*memcg
,
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
, memcg
,
855 else if (PageAnon(page
))
856 referenced
+= page_referenced_anon(page
, memcg
,
858 else if (page
->mapping
)
859 referenced
+= page_referenced_file(page
, memcg
,
864 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
;
905 BUG_ON(PageAnon(page
));
907 mutex_lock(&mapping
->i_mmap_mutex
);
908 vma_interval_tree_foreach(vma
, &mapping
->i_mmap
, pgoff
, pgoff
) {
909 if (vma
->vm_flags
& VM_SHARED
) {
910 unsigned long address
= vma_address(page
, vma
);
911 ret
+= page_mkclean_one(page
, vma
, address
);
914 mutex_unlock(&mapping
->i_mmap_mutex
);
918 int page_mkclean(struct page
*page
)
922 BUG_ON(!PageLocked(page
));
924 if (page_mapped(page
)) {
925 struct address_space
*mapping
= page_mapping(page
);
927 ret
= page_mkclean_file(mapping
, page
);
928 if (page_test_and_clear_dirty(page_to_pfn(page
), 1))
935 EXPORT_SYMBOL_GPL(page_mkclean
);
938 * page_move_anon_rmap - move a page to our anon_vma
939 * @page: the page to move to our anon_vma
940 * @vma: the vma the page belongs to
941 * @address: the user virtual address mapped
943 * When a page belongs exclusively to one process after a COW event,
944 * that page can be moved into the anon_vma that belongs to just that
945 * process, so the rmap code will not search the parent or sibling
948 void page_move_anon_rmap(struct page
*page
,
949 struct vm_area_struct
*vma
, unsigned long address
)
951 struct anon_vma
*anon_vma
= vma
->anon_vma
;
953 VM_BUG_ON(!PageLocked(page
));
954 VM_BUG_ON(!anon_vma
);
955 VM_BUG_ON(page
->index
!= linear_page_index(vma
, address
));
957 anon_vma
= (void *) anon_vma
+ PAGE_MAPPING_ANON
;
958 page
->mapping
= (struct address_space
*) anon_vma
;
962 * __page_set_anon_rmap - set up new anonymous rmap
963 * @page: Page to add to rmap
964 * @vma: VM area to add page to.
965 * @address: User virtual address of the mapping
966 * @exclusive: the page is exclusively owned by the current process
968 static void __page_set_anon_rmap(struct page
*page
,
969 struct vm_area_struct
*vma
, unsigned long address
, int exclusive
)
971 struct anon_vma
*anon_vma
= vma
->anon_vma
;
979 * If the page isn't exclusively mapped into this vma,
980 * we must use the _oldest_ possible anon_vma for the
984 anon_vma
= anon_vma
->root
;
986 anon_vma
= (void *) anon_vma
+ PAGE_MAPPING_ANON
;
987 page
->mapping
= (struct address_space
*) anon_vma
;
988 page
->index
= linear_page_index(vma
, address
);
992 * __page_check_anon_rmap - sanity check anonymous rmap addition
993 * @page: the page to add the mapping to
994 * @vma: the vm area in which the mapping is added
995 * @address: the user virtual address mapped
997 static void __page_check_anon_rmap(struct page
*page
,
998 struct vm_area_struct
*vma
, unsigned long address
)
1000 #ifdef CONFIG_DEBUG_VM
1002 * The page's anon-rmap details (mapping and index) are guaranteed to
1003 * be set up correctly at this point.
1005 * We have exclusion against page_add_anon_rmap because the caller
1006 * always holds the page locked, except if called from page_dup_rmap,
1007 * in which case the page is already known to be setup.
1009 * We have exclusion against page_add_new_anon_rmap because those pages
1010 * are initially only visible via the pagetables, and the pte is locked
1011 * over the call to page_add_new_anon_rmap.
1013 BUG_ON(page_anon_vma(page
)->root
!= vma
->anon_vma
->root
);
1014 BUG_ON(page
->index
!= linear_page_index(vma
, address
));
1019 * page_add_anon_rmap - add pte mapping to an anonymous page
1020 * @page: the page to add the mapping to
1021 * @vma: the vm area in which the mapping is added
1022 * @address: the user virtual address mapped
1024 * The caller needs to hold the pte lock, and the page must be locked in
1025 * the anon_vma case: to serialize mapping,index checking after setting,
1026 * and to ensure that PageAnon is not being upgraded racily to PageKsm
1027 * (but PageKsm is never downgraded to PageAnon).
1029 void page_add_anon_rmap(struct page
*page
,
1030 struct vm_area_struct
*vma
, unsigned long address
)
1032 do_page_add_anon_rmap(page
, vma
, address
, 0);
1036 * Special version of the above for do_swap_page, which often runs
1037 * into pages that are exclusively owned by the current process.
1038 * Everybody else should continue to use page_add_anon_rmap above.
1040 void do_page_add_anon_rmap(struct page
*page
,
1041 struct vm_area_struct
*vma
, unsigned long address
, int exclusive
)
1043 int first
= atomic_inc_and_test(&page
->_mapcount
);
1045 if (!PageTransHuge(page
))
1046 __inc_zone_page_state(page
, NR_ANON_PAGES
);
1048 __inc_zone_page_state(page
,
1049 NR_ANON_TRANSPARENT_HUGEPAGES
);
1051 if (unlikely(PageKsm(page
)))
1054 VM_BUG_ON(!PageLocked(page
));
1055 /* address might be in next vma when migration races vma_adjust */
1057 __page_set_anon_rmap(page
, vma
, address
, exclusive
);
1059 __page_check_anon_rmap(page
, vma
, address
);
1063 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1064 * @page: the page to add the mapping to
1065 * @vma: the vm area in which the mapping is added
1066 * @address: the user virtual address mapped
1068 * Same as page_add_anon_rmap but must only be called on *new* pages.
1069 * This means the inc-and-test can be bypassed.
1070 * Page does not have to be locked.
1072 void page_add_new_anon_rmap(struct page
*page
,
1073 struct vm_area_struct
*vma
, unsigned long address
)
1075 VM_BUG_ON(address
< vma
->vm_start
|| address
>= vma
->vm_end
);
1076 SetPageSwapBacked(page
);
1077 atomic_set(&page
->_mapcount
, 0); /* increment count (starts at -1) */
1078 if (!PageTransHuge(page
))
1079 __inc_zone_page_state(page
, NR_ANON_PAGES
);
1081 __inc_zone_page_state(page
, NR_ANON_TRANSPARENT_HUGEPAGES
);
1082 __page_set_anon_rmap(page
, vma
, address
, 1);
1083 if (!mlocked_vma_newpage(vma
, page
))
1084 lru_cache_add_lru(page
, LRU_ACTIVE_ANON
);
1086 add_page_to_unevictable_list(page
);
1090 * page_add_file_rmap - add pte mapping to a file page
1091 * @page: the page to add the mapping to
1093 * The caller needs to hold the pte lock.
1095 void page_add_file_rmap(struct page
*page
)
1098 unsigned long flags
;
1100 mem_cgroup_begin_update_page_stat(page
, &locked
, &flags
);
1101 if (atomic_inc_and_test(&page
->_mapcount
)) {
1102 __inc_zone_page_state(page
, NR_FILE_MAPPED
);
1103 mem_cgroup_inc_page_stat(page
, MEMCG_NR_FILE_MAPPED
);
1105 mem_cgroup_end_update_page_stat(page
, &locked
, &flags
);
1109 * page_remove_rmap - take down pte mapping from a page
1110 * @page: page to remove mapping from
1112 * The caller needs to hold the pte lock.
1114 void page_remove_rmap(struct page
*page
)
1116 bool anon
= PageAnon(page
);
1118 unsigned long flags
;
1121 * The anon case has no mem_cgroup page_stat to update; but may
1122 * uncharge_page() below, where the lock ordering can deadlock if
1123 * we hold the lock against page_stat move: so avoid it on anon.
1126 mem_cgroup_begin_update_page_stat(page
, &locked
, &flags
);
1128 /* page still mapped by someone else? */
1129 if (!atomic_add_negative(-1, &page
->_mapcount
))
1133 * Now that the last pte has gone, s390 must transfer dirty
1134 * flag from storage key to struct page. We can usually skip
1135 * this if the page is anon, so about to be freed; but perhaps
1136 * not if it's in swapcache - there might be another pte slot
1137 * containing the swap entry, but page not yet written to swap.
1139 if ((!anon
|| PageSwapCache(page
)) &&
1140 page_test_and_clear_dirty(page_to_pfn(page
), 1))
1141 set_page_dirty(page
);
1143 * Hugepages are not counted in NR_ANON_PAGES nor NR_FILE_MAPPED
1144 * and not charged by memcg for now.
1146 if (unlikely(PageHuge(page
)))
1149 mem_cgroup_uncharge_page(page
);
1150 if (!PageTransHuge(page
))
1151 __dec_zone_page_state(page
, NR_ANON_PAGES
);
1153 __dec_zone_page_state(page
,
1154 NR_ANON_TRANSPARENT_HUGEPAGES
);
1156 __dec_zone_page_state(page
, NR_FILE_MAPPED
);
1157 mem_cgroup_dec_page_stat(page
, MEMCG_NR_FILE_MAPPED
);
1158 mem_cgroup_end_update_page_stat(page
, &locked
, &flags
);
1160 if (unlikely(PageMlocked(page
)))
1161 clear_page_mlock(page
);
1163 * It would be tidy to reset the PageAnon mapping here,
1164 * but that might overwrite a racing page_add_anon_rmap
1165 * which increments mapcount after us but sets mapping
1166 * before us: so leave the reset to free_hot_cold_page,
1167 * and remember that it's only reliable while mapped.
1168 * Leaving it set also helps swapoff to reinstate ptes
1169 * faster for those pages still in swapcache.
1174 mem_cgroup_end_update_page_stat(page
, &locked
, &flags
);
1178 * Subfunctions of try_to_unmap: try_to_unmap_one called
1179 * repeatedly from try_to_unmap_ksm, try_to_unmap_anon or try_to_unmap_file.
1181 int try_to_unmap_one(struct page
*page
, struct vm_area_struct
*vma
,
1182 unsigned long address
, enum ttu_flags flags
)
1184 struct mm_struct
*mm
= vma
->vm_mm
;
1188 int ret
= SWAP_AGAIN
;
1190 pte
= page_check_address(page
, mm
, address
, &ptl
, 0);
1195 * If the page is mlock()d, we cannot swap it out.
1196 * If it's recently referenced (perhaps page_referenced
1197 * skipped over this mm) then we should reactivate it.
1199 if (!(flags
& TTU_IGNORE_MLOCK
)) {
1200 if (vma
->vm_flags
& VM_LOCKED
)
1203 if (TTU_ACTION(flags
) == TTU_MUNLOCK
)
1206 if (!(flags
& TTU_IGNORE_ACCESS
)) {
1207 if (ptep_clear_flush_young_notify(vma
, address
, pte
)) {
1213 /* Nuke the page table entry. */
1214 flush_cache_page(vma
, address
, page_to_pfn(page
));
1215 pteval
= ptep_clear_flush_notify(vma
, address
, pte
);
1217 /* Move the dirty bit to the physical page now the pte is gone. */
1218 if (pte_dirty(pteval
))
1219 set_page_dirty(page
);
1221 /* Update high watermark before we lower rss */
1222 update_hiwater_rss(mm
);
1224 if (PageHWPoison(page
) && !(flags
& TTU_IGNORE_HWPOISON
)) {
1226 dec_mm_counter(mm
, MM_ANONPAGES
);
1228 dec_mm_counter(mm
, MM_FILEPAGES
);
1229 set_pte_at(mm
, address
, pte
,
1230 swp_entry_to_pte(make_hwpoison_entry(page
)));
1231 } else if (PageAnon(page
)) {
1232 swp_entry_t entry
= { .val
= page_private(page
) };
1234 if (PageSwapCache(page
)) {
1236 * Store the swap location in the pte.
1237 * See handle_pte_fault() ...
1239 if (swap_duplicate(entry
) < 0) {
1240 set_pte_at(mm
, address
, pte
, pteval
);
1244 if (list_empty(&mm
->mmlist
)) {
1245 spin_lock(&mmlist_lock
);
1246 if (list_empty(&mm
->mmlist
))
1247 list_add(&mm
->mmlist
, &init_mm
.mmlist
);
1248 spin_unlock(&mmlist_lock
);
1250 dec_mm_counter(mm
, MM_ANONPAGES
);
1251 inc_mm_counter(mm
, MM_SWAPENTS
);
1252 } else if (IS_ENABLED(CONFIG_MIGRATION
)) {
1254 * Store the pfn of the page in a special migration
1255 * pte. do_swap_page() will wait until the migration
1256 * pte is removed and then restart fault handling.
1258 BUG_ON(TTU_ACTION(flags
) != TTU_MIGRATION
);
1259 entry
= make_migration_entry(page
, pte_write(pteval
));
1261 set_pte_at(mm
, address
, pte
, swp_entry_to_pte(entry
));
1262 BUG_ON(pte_file(*pte
));
1263 } else if (IS_ENABLED(CONFIG_MIGRATION
) &&
1264 (TTU_ACTION(flags
) == TTU_MIGRATION
)) {
1265 /* Establish migration entry for a file page */
1267 entry
= make_migration_entry(page
, pte_write(pteval
));
1268 set_pte_at(mm
, address
, pte
, swp_entry_to_pte(entry
));
1270 dec_mm_counter(mm
, MM_FILEPAGES
);
1272 page_remove_rmap(page
);
1273 page_cache_release(page
);
1276 pte_unmap_unlock(pte
, ptl
);
1281 pte_unmap_unlock(pte
, ptl
);
1285 * We need mmap_sem locking, Otherwise VM_LOCKED check makes
1286 * unstable result and race. Plus, We can't wait here because
1287 * we now hold anon_vma->mutex or mapping->i_mmap_mutex.
1288 * if trylock failed, the page remain in evictable lru and later
1289 * vmscan could retry to move the page to unevictable lru if the
1290 * page is actually mlocked.
1292 if (down_read_trylock(&vma
->vm_mm
->mmap_sem
)) {
1293 if (vma
->vm_flags
& VM_LOCKED
) {
1294 mlock_vma_page(page
);
1297 up_read(&vma
->vm_mm
->mmap_sem
);
1303 * objrmap doesn't work for nonlinear VMAs because the assumption that
1304 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
1305 * Consequently, given a particular page and its ->index, we cannot locate the
1306 * ptes which are mapping that page without an exhaustive linear search.
1308 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
1309 * maps the file to which the target page belongs. The ->vm_private_data field
1310 * holds the current cursor into that scan. Successive searches will circulate
1311 * around the vma's virtual address space.
1313 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
1314 * more scanning pressure is placed against them as well. Eventually pages
1315 * will become fully unmapped and are eligible for eviction.
1317 * For very sparsely populated VMAs this is a little inefficient - chances are
1318 * there there won't be many ptes located within the scan cluster. In this case
1319 * maybe we could scan further - to the end of the pte page, perhaps.
1321 * Mlocked pages: check VM_LOCKED under mmap_sem held for read, if we can
1322 * acquire it without blocking. If vma locked, mlock the pages in the cluster,
1323 * rather than unmapping them. If we encounter the "check_page" that vmscan is
1324 * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN.
1326 #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
1327 #define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
1329 static int try_to_unmap_cluster(unsigned long cursor
, unsigned int *mapcount
,
1330 struct vm_area_struct
*vma
, struct page
*check_page
)
1332 struct mm_struct
*mm
= vma
->vm_mm
;
1340 unsigned long address
;
1342 int ret
= SWAP_AGAIN
;
1345 address
= (vma
->vm_start
+ cursor
) & CLUSTER_MASK
;
1346 end
= address
+ CLUSTER_SIZE
;
1347 if (address
< vma
->vm_start
)
1348 address
= vma
->vm_start
;
1349 if (end
> vma
->vm_end
)
1352 pgd
= pgd_offset(mm
, address
);
1353 if (!pgd_present(*pgd
))
1356 pud
= pud_offset(pgd
, address
);
1357 if (!pud_present(*pud
))
1360 pmd
= pmd_offset(pud
, address
);
1361 if (!pmd_present(*pmd
))
1365 * If we can acquire the mmap_sem for read, and vma is VM_LOCKED,
1366 * keep the sem while scanning the cluster for mlocking pages.
1368 if (down_read_trylock(&vma
->vm_mm
->mmap_sem
)) {
1369 locked_vma
= (vma
->vm_flags
& VM_LOCKED
);
1371 up_read(&vma
->vm_mm
->mmap_sem
); /* don't need it */
1374 pte
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
1376 /* Update high watermark before we lower rss */
1377 update_hiwater_rss(mm
);
1379 for (; address
< end
; pte
++, address
+= PAGE_SIZE
) {
1380 if (!pte_present(*pte
))
1382 page
= vm_normal_page(vma
, address
, *pte
);
1383 BUG_ON(!page
|| PageAnon(page
));
1386 mlock_vma_page(page
); /* no-op if already mlocked */
1387 if (page
== check_page
)
1389 continue; /* don't unmap */
1392 if (ptep_clear_flush_young_notify(vma
, address
, pte
))
1395 /* Nuke the page table entry. */
1396 flush_cache_page(vma
, address
, pte_pfn(*pte
));
1397 pteval
= ptep_clear_flush_notify(vma
, address
, pte
);
1399 /* If nonlinear, store the file page offset in the pte. */
1400 if (page
->index
!= linear_page_index(vma
, address
))
1401 set_pte_at(mm
, address
, pte
, pgoff_to_pte(page
->index
));
1403 /* Move the dirty bit to the physical page now the pte is gone. */
1404 if (pte_dirty(pteval
))
1405 set_page_dirty(page
);
1407 page_remove_rmap(page
);
1408 page_cache_release(page
);
1409 dec_mm_counter(mm
, MM_FILEPAGES
);
1412 pte_unmap_unlock(pte
- 1, ptl
);
1414 up_read(&vma
->vm_mm
->mmap_sem
);
1418 bool is_vma_temporary_stack(struct vm_area_struct
*vma
)
1420 int maybe_stack
= vma
->vm_flags
& (VM_GROWSDOWN
| VM_GROWSUP
);
1425 if ((vma
->vm_flags
& VM_STACK_INCOMPLETE_SETUP
) ==
1426 VM_STACK_INCOMPLETE_SETUP
)
1433 * try_to_unmap_anon - unmap or unlock anonymous page using the object-based
1435 * @page: the page to unmap/unlock
1436 * @flags: action and flags
1438 * Find all the mappings of a page using the mapping pointer and the vma chains
1439 * contained in the anon_vma struct it points to.
1441 * This function is only called from try_to_unmap/try_to_munlock for
1443 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1444 * where the page was found will be held for write. So, we won't recheck
1445 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1448 static int try_to_unmap_anon(struct page
*page
, enum ttu_flags flags
)
1450 struct anon_vma
*anon_vma
;
1452 struct anon_vma_chain
*avc
;
1453 int ret
= SWAP_AGAIN
;
1455 anon_vma
= page_lock_anon_vma(page
);
1459 pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
1460 anon_vma_interval_tree_foreach(avc
, &anon_vma
->rb_root
, pgoff
, pgoff
) {
1461 struct vm_area_struct
*vma
= avc
->vma
;
1462 unsigned long address
;
1465 * During exec, a temporary VMA is setup and later moved.
1466 * The VMA is moved under the anon_vma lock but not the
1467 * page tables leading to a race where migration cannot
1468 * find the migration ptes. Rather than increasing the
1469 * locking requirements of exec(), migration skips
1470 * temporary VMAs until after exec() completes.
1472 if (IS_ENABLED(CONFIG_MIGRATION
) && (flags
& TTU_MIGRATION
) &&
1473 is_vma_temporary_stack(vma
))
1476 address
= vma_address(page
, vma
);
1477 ret
= try_to_unmap_one(page
, vma
, address
, flags
);
1478 if (ret
!= SWAP_AGAIN
|| !page_mapped(page
))
1482 page_unlock_anon_vma(anon_vma
);
1487 * try_to_unmap_file - unmap/unlock file page using the object-based rmap method
1488 * @page: the page to unmap/unlock
1489 * @flags: action and flags
1491 * Find all the mappings of a page using the mapping pointer and the vma chains
1492 * contained in the address_space struct it points to.
1494 * This function is only called from try_to_unmap/try_to_munlock for
1495 * object-based pages.
1496 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1497 * where the page was found will be held for write. So, we won't recheck
1498 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1501 static int try_to_unmap_file(struct page
*page
, enum ttu_flags flags
)
1503 struct address_space
*mapping
= page
->mapping
;
1504 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
1505 struct vm_area_struct
*vma
;
1506 int ret
= SWAP_AGAIN
;
1507 unsigned long cursor
;
1508 unsigned long max_nl_cursor
= 0;
1509 unsigned long max_nl_size
= 0;
1510 unsigned int mapcount
;
1512 mutex_lock(&mapping
->i_mmap_mutex
);
1513 vma_interval_tree_foreach(vma
, &mapping
->i_mmap
, pgoff
, pgoff
) {
1514 unsigned long address
= vma_address(page
, vma
);
1515 ret
= try_to_unmap_one(page
, vma
, address
, flags
);
1516 if (ret
!= SWAP_AGAIN
|| !page_mapped(page
))
1520 if (list_empty(&mapping
->i_mmap_nonlinear
))
1524 * We don't bother to try to find the munlocked page in nonlinears.
1525 * It's costly. Instead, later, page reclaim logic may call
1526 * try_to_unmap(TTU_MUNLOCK) and recover PG_mlocked lazily.
1528 if (TTU_ACTION(flags
) == TTU_MUNLOCK
)
1531 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
,
1533 cursor
= (unsigned long) vma
->vm_private_data
;
1534 if (cursor
> max_nl_cursor
)
1535 max_nl_cursor
= cursor
;
1536 cursor
= vma
->vm_end
- vma
->vm_start
;
1537 if (cursor
> max_nl_size
)
1538 max_nl_size
= cursor
;
1541 if (max_nl_size
== 0) { /* all nonlinears locked or reserved ? */
1547 * We don't try to search for this page in the nonlinear vmas,
1548 * and page_referenced wouldn't have found it anyway. Instead
1549 * just walk the nonlinear vmas trying to age and unmap some.
1550 * The mapcount of the page we came in with is irrelevant,
1551 * but even so use it as a guide to how hard we should try?
1553 mapcount
= page_mapcount(page
);
1558 max_nl_size
= (max_nl_size
+ CLUSTER_SIZE
- 1) & CLUSTER_MASK
;
1559 if (max_nl_cursor
== 0)
1560 max_nl_cursor
= CLUSTER_SIZE
;
1563 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
,
1565 cursor
= (unsigned long) vma
->vm_private_data
;
1566 while ( cursor
< max_nl_cursor
&&
1567 cursor
< vma
->vm_end
- vma
->vm_start
) {
1568 if (try_to_unmap_cluster(cursor
, &mapcount
,
1569 vma
, page
) == SWAP_MLOCK
)
1571 cursor
+= CLUSTER_SIZE
;
1572 vma
->vm_private_data
= (void *) cursor
;
1573 if ((int)mapcount
<= 0)
1576 vma
->vm_private_data
= (void *) max_nl_cursor
;
1579 max_nl_cursor
+= CLUSTER_SIZE
;
1580 } while (max_nl_cursor
<= max_nl_size
);
1583 * Don't loop forever (perhaps all the remaining pages are
1584 * in locked vmas). Reset cursor on all unreserved nonlinear
1585 * vmas, now forgetting on which ones it had fallen behind.
1587 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
, shared
.nonlinear
)
1588 vma
->vm_private_data
= NULL
;
1590 mutex_unlock(&mapping
->i_mmap_mutex
);
1595 * try_to_unmap - try to remove all page table mappings to a page
1596 * @page: the page to get unmapped
1597 * @flags: action and flags
1599 * Tries to remove all the page table entries which are mapping this
1600 * page, used in the pageout path. Caller must hold the page lock.
1601 * Return values are:
1603 * SWAP_SUCCESS - we succeeded in removing all mappings
1604 * SWAP_AGAIN - we missed a mapping, try again later
1605 * SWAP_FAIL - the page is unswappable
1606 * SWAP_MLOCK - page is mlocked.
1608 int try_to_unmap(struct page
*page
, enum ttu_flags flags
)
1612 BUG_ON(!PageLocked(page
));
1613 VM_BUG_ON(!PageHuge(page
) && PageTransHuge(page
));
1615 if (unlikely(PageKsm(page
)))
1616 ret
= try_to_unmap_ksm(page
, flags
);
1617 else if (PageAnon(page
))
1618 ret
= try_to_unmap_anon(page
, flags
);
1620 ret
= try_to_unmap_file(page
, flags
);
1621 if (ret
!= SWAP_MLOCK
&& !page_mapped(page
))
1627 * try_to_munlock - try to munlock a page
1628 * @page: the page to be munlocked
1630 * Called from munlock code. Checks all of the VMAs mapping the page
1631 * to make sure nobody else has this page mlocked. The page will be
1632 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1634 * Return values are:
1636 * SWAP_AGAIN - no vma is holding page mlocked, or,
1637 * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem
1638 * SWAP_FAIL - page cannot be located at present
1639 * SWAP_MLOCK - page is now mlocked.
1641 int try_to_munlock(struct page
*page
)
1643 VM_BUG_ON(!PageLocked(page
) || PageLRU(page
));
1645 if (unlikely(PageKsm(page
)))
1646 return try_to_unmap_ksm(page
, TTU_MUNLOCK
);
1647 else if (PageAnon(page
))
1648 return try_to_unmap_anon(page
, TTU_MUNLOCK
);
1650 return try_to_unmap_file(page
, TTU_MUNLOCK
);
1653 void __put_anon_vma(struct anon_vma
*anon_vma
)
1655 struct anon_vma
*root
= anon_vma
->root
;
1657 if (root
!= anon_vma
&& atomic_dec_and_test(&root
->refcount
))
1658 anon_vma_free(root
);
1660 anon_vma_free(anon_vma
);
1663 #ifdef CONFIG_MIGRATION
1665 * rmap_walk() and its helpers rmap_walk_anon() and rmap_walk_file():
1666 * Called by migrate.c to remove migration ptes, but might be used more later.
1668 static int rmap_walk_anon(struct page
*page
, int (*rmap_one
)(struct page
*,
1669 struct vm_area_struct
*, unsigned long, void *), void *arg
)
1671 struct anon_vma
*anon_vma
;
1672 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
1673 struct anon_vma_chain
*avc
;
1674 int ret
= SWAP_AGAIN
;
1677 * Note: remove_migration_ptes() cannot use page_lock_anon_vma()
1678 * because that depends on page_mapped(); but not all its usages
1679 * are holding mmap_sem. Users without mmap_sem are required to
1680 * take a reference count to prevent the anon_vma disappearing
1682 anon_vma
= page_anon_vma(page
);
1685 anon_vma_lock(anon_vma
);
1686 anon_vma_interval_tree_foreach(avc
, &anon_vma
->rb_root
, pgoff
, pgoff
) {
1687 struct vm_area_struct
*vma
= avc
->vma
;
1688 unsigned long address
= vma_address(page
, vma
);
1689 ret
= rmap_one(page
, vma
, address
, arg
);
1690 if (ret
!= SWAP_AGAIN
)
1693 anon_vma_unlock(anon_vma
);
1697 static int rmap_walk_file(struct page
*page
, int (*rmap_one
)(struct page
*,
1698 struct vm_area_struct
*, unsigned long, void *), void *arg
)
1700 struct address_space
*mapping
= page
->mapping
;
1701 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
1702 struct vm_area_struct
*vma
;
1703 int ret
= SWAP_AGAIN
;
1707 mutex_lock(&mapping
->i_mmap_mutex
);
1708 vma_interval_tree_foreach(vma
, &mapping
->i_mmap
, pgoff
, pgoff
) {
1709 unsigned long address
= vma_address(page
, vma
);
1710 ret
= rmap_one(page
, vma
, address
, arg
);
1711 if (ret
!= SWAP_AGAIN
)
1715 * No nonlinear handling: being always shared, nonlinear vmas
1716 * never contain migration ptes. Decide what to do about this
1717 * limitation to linear when we need rmap_walk() on nonlinear.
1719 mutex_unlock(&mapping
->i_mmap_mutex
);
1723 int rmap_walk(struct page
*page
, int (*rmap_one
)(struct page
*,
1724 struct vm_area_struct
*, unsigned long, void *), void *arg
)
1726 VM_BUG_ON(!PageLocked(page
));
1728 if (unlikely(PageKsm(page
)))
1729 return rmap_walk_ksm(page
, rmap_one
, arg
);
1730 else if (PageAnon(page
))
1731 return rmap_walk_anon(page
, rmap_one
, arg
);
1733 return rmap_walk_file(page
, rmap_one
, arg
);
1735 #endif /* CONFIG_MIGRATION */
1737 #ifdef CONFIG_HUGETLB_PAGE
1739 * The following three functions are for anonymous (private mapped) hugepages.
1740 * Unlike common anonymous pages, anonymous hugepages have no accounting code
1741 * and no lru code, because we handle hugepages differently from common pages.
1743 static void __hugepage_set_anon_rmap(struct page
*page
,
1744 struct vm_area_struct
*vma
, unsigned long address
, int exclusive
)
1746 struct anon_vma
*anon_vma
= vma
->anon_vma
;
1753 anon_vma
= anon_vma
->root
;
1755 anon_vma
= (void *) anon_vma
+ PAGE_MAPPING_ANON
;
1756 page
->mapping
= (struct address_space
*) anon_vma
;
1757 page
->index
= linear_page_index(vma
, address
);
1760 void hugepage_add_anon_rmap(struct page
*page
,
1761 struct vm_area_struct
*vma
, unsigned long address
)
1763 struct anon_vma
*anon_vma
= vma
->anon_vma
;
1766 BUG_ON(!PageLocked(page
));
1768 /* address might be in next vma when migration races vma_adjust */
1769 first
= atomic_inc_and_test(&page
->_mapcount
);
1771 __hugepage_set_anon_rmap(page
, vma
, address
, 0);
1774 void hugepage_add_new_anon_rmap(struct page
*page
,
1775 struct vm_area_struct
*vma
, unsigned long address
)
1777 BUG_ON(address
< vma
->vm_start
|| address
>= vma
->vm_end
);
1778 atomic_set(&page
->_mapcount
, 0);
1779 __hugepage_set_anon_rmap(page
, vma
, address
, 1);
1781 #endif /* CONFIG_HUGETLB_PAGE */