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
);
132 * It's critical to add new vmas to the tail of the anon_vma,
133 * see comment in huge_memory.c:__split_huge_page().
135 list_add_tail(&avc
->same_anon_vma
, &anon_vma
->head
);
139 * anon_vma_prepare - attach an anon_vma to a memory region
140 * @vma: the memory region in question
142 * This makes sure the memory mapping described by 'vma' has
143 * an 'anon_vma' attached to it, so that we can associate the
144 * anonymous pages mapped into it with that anon_vma.
146 * The common case will be that we already have one, but if
147 * not we either need to find an adjacent mapping that we
148 * can re-use the anon_vma from (very common when the only
149 * reason for splitting a vma has been mprotect()), or we
150 * allocate a new one.
152 * Anon-vma allocations are very subtle, because we may have
153 * optimistically looked up an anon_vma in page_lock_anon_vma()
154 * and that may actually touch the spinlock even in the newly
155 * allocated vma (it depends on RCU to make sure that the
156 * anon_vma isn't actually destroyed).
158 * As a result, we need to do proper anon_vma locking even
159 * for the new allocation. At the same time, we do not want
160 * to do any locking for the common case of already having
163 * This must be called with the mmap_sem held for reading.
165 int anon_vma_prepare(struct vm_area_struct
*vma
)
167 struct anon_vma
*anon_vma
= vma
->anon_vma
;
168 struct anon_vma_chain
*avc
;
171 if (unlikely(!anon_vma
)) {
172 struct mm_struct
*mm
= vma
->vm_mm
;
173 struct anon_vma
*allocated
;
175 avc
= anon_vma_chain_alloc(GFP_KERNEL
);
179 anon_vma
= find_mergeable_anon_vma(vma
);
182 anon_vma
= anon_vma_alloc();
183 if (unlikely(!anon_vma
))
184 goto out_enomem_free_avc
;
185 allocated
= anon_vma
;
188 anon_vma_lock(anon_vma
);
189 /* page_table_lock to protect against threads */
190 spin_lock(&mm
->page_table_lock
);
191 if (likely(!vma
->anon_vma
)) {
192 vma
->anon_vma
= anon_vma
;
193 anon_vma_chain_link(vma
, avc
, anon_vma
);
197 spin_unlock(&mm
->page_table_lock
);
198 anon_vma_unlock(anon_vma
);
200 if (unlikely(allocated
))
201 put_anon_vma(allocated
);
203 anon_vma_chain_free(avc
);
208 anon_vma_chain_free(avc
);
214 * This is a useful helper function for locking the anon_vma root as
215 * we traverse the vma->anon_vma_chain, looping over anon_vma's that
218 * Such anon_vma's should have the same root, so you'd expect to see
219 * just a single mutex_lock for the whole traversal.
221 static inline struct anon_vma
*lock_anon_vma_root(struct anon_vma
*root
, struct anon_vma
*anon_vma
)
223 struct anon_vma
*new_root
= anon_vma
->root
;
224 if (new_root
!= root
) {
225 if (WARN_ON_ONCE(root
))
226 mutex_unlock(&root
->mutex
);
228 mutex_lock(&root
->mutex
);
233 static inline void unlock_anon_vma_root(struct anon_vma
*root
)
236 mutex_unlock(&root
->mutex
);
240 * Attach the anon_vmas from src to dst.
241 * Returns 0 on success, -ENOMEM on failure.
243 int anon_vma_clone(struct vm_area_struct
*dst
, struct vm_area_struct
*src
)
245 struct anon_vma_chain
*avc
, *pavc
;
246 struct anon_vma
*root
= NULL
;
248 list_for_each_entry_reverse(pavc
, &src
->anon_vma_chain
, same_vma
) {
249 struct anon_vma
*anon_vma
;
251 avc
= anon_vma_chain_alloc(GFP_NOWAIT
| __GFP_NOWARN
);
252 if (unlikely(!avc
)) {
253 unlock_anon_vma_root(root
);
255 avc
= anon_vma_chain_alloc(GFP_KERNEL
);
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 unlink_anon_vmas(dst
);
272 * Attach vma to its own anon_vma, as well as to the anon_vmas that
273 * the corresponding VMA in the parent process is attached to.
274 * Returns 0 on success, non-zero on failure.
276 int anon_vma_fork(struct vm_area_struct
*vma
, struct vm_area_struct
*pvma
)
278 struct anon_vma_chain
*avc
;
279 struct anon_vma
*anon_vma
;
281 /* Don't bother if the parent process has no anon_vma here. */
286 * First, attach the new VMA to the parent VMA's anon_vmas,
287 * so rmap can find non-COWed pages in child processes.
289 if (anon_vma_clone(vma
, pvma
))
292 /* Then add our own anon_vma. */
293 anon_vma
= anon_vma_alloc();
296 avc
= anon_vma_chain_alloc(GFP_KERNEL
);
298 goto out_error_free_anon_vma
;
301 * The root anon_vma's spinlock is the lock actually used when we
302 * lock any of the anon_vmas in this anon_vma tree.
304 anon_vma
->root
= pvma
->anon_vma
->root
;
306 * With refcounts, an anon_vma can stay around longer than the
307 * process it belongs to. The root anon_vma needs to be pinned until
308 * this anon_vma is freed, because the lock lives in the root.
310 get_anon_vma(anon_vma
->root
);
311 /* Mark this anon_vma as the one where our new (COWed) pages go. */
312 vma
->anon_vma
= anon_vma
;
313 anon_vma_lock(anon_vma
);
314 anon_vma_chain_link(vma
, avc
, anon_vma
);
315 anon_vma_unlock(anon_vma
);
319 out_error_free_anon_vma
:
320 put_anon_vma(anon_vma
);
322 unlink_anon_vmas(vma
);
326 void unlink_anon_vmas(struct vm_area_struct
*vma
)
328 struct anon_vma_chain
*avc
, *next
;
329 struct anon_vma
*root
= NULL
;
332 * Unlink each anon_vma chained to the VMA. This list is ordered
333 * from newest to oldest, ensuring the root anon_vma gets freed last.
335 list_for_each_entry_safe(avc
, next
, &vma
->anon_vma_chain
, same_vma
) {
336 struct anon_vma
*anon_vma
= avc
->anon_vma
;
338 root
= lock_anon_vma_root(root
, anon_vma
);
339 list_del(&avc
->same_anon_vma
);
342 * Leave empty anon_vmas on the list - we'll need
343 * to free them outside the lock.
345 if (list_empty(&anon_vma
->head
))
348 list_del(&avc
->same_vma
);
349 anon_vma_chain_free(avc
);
351 unlock_anon_vma_root(root
);
354 * Iterate the list once more, it now only contains empty and unlinked
355 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
356 * needing to acquire the anon_vma->root->mutex.
358 list_for_each_entry_safe(avc
, next
, &vma
->anon_vma_chain
, same_vma
) {
359 struct anon_vma
*anon_vma
= avc
->anon_vma
;
361 put_anon_vma(anon_vma
);
363 list_del(&avc
->same_vma
);
364 anon_vma_chain_free(avc
);
368 static void anon_vma_ctor(void *data
)
370 struct anon_vma
*anon_vma
= data
;
372 mutex_init(&anon_vma
->mutex
);
373 atomic_set(&anon_vma
->refcount
, 0);
374 INIT_LIST_HEAD(&anon_vma
->head
);
377 void __init
anon_vma_init(void)
379 anon_vma_cachep
= kmem_cache_create("anon_vma", sizeof(struct anon_vma
),
380 0, SLAB_DESTROY_BY_RCU
|SLAB_PANIC
, anon_vma_ctor
);
381 anon_vma_chain_cachep
= KMEM_CACHE(anon_vma_chain
, SLAB_PANIC
);
385 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
387 * Since there is no serialization what so ever against page_remove_rmap()
388 * the best this function can do is return a locked anon_vma that might
389 * have been relevant to this page.
391 * The page might have been remapped to a different anon_vma or the anon_vma
392 * returned may already be freed (and even reused).
394 * In case it was remapped to a different anon_vma, the new anon_vma will be a
395 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
396 * ensure that any anon_vma obtained from the page will still be valid for as
397 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
399 * All users of this function must be very careful when walking the anon_vma
400 * chain and verify that the page in question is indeed mapped in it
401 * [ something equivalent to page_mapped_in_vma() ].
403 * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap()
404 * that the anon_vma pointer from page->mapping is valid if there is a
405 * mapcount, we can dereference the anon_vma after observing those.
407 struct anon_vma
*page_get_anon_vma(struct page
*page
)
409 struct anon_vma
*anon_vma
= NULL
;
410 unsigned long anon_mapping
;
413 anon_mapping
= (unsigned long) ACCESS_ONCE(page
->mapping
);
414 if ((anon_mapping
& PAGE_MAPPING_FLAGS
) != PAGE_MAPPING_ANON
)
416 if (!page_mapped(page
))
419 anon_vma
= (struct anon_vma
*) (anon_mapping
- PAGE_MAPPING_ANON
);
420 if (!atomic_inc_not_zero(&anon_vma
->refcount
)) {
426 * If this page is still mapped, then its anon_vma cannot have been
427 * freed. But if it has been unmapped, we have no security against the
428 * anon_vma structure being freed and reused (for another anon_vma:
429 * SLAB_DESTROY_BY_RCU guarantees that - so the atomic_inc_not_zero()
430 * above cannot corrupt).
432 if (!page_mapped(page
)) {
433 put_anon_vma(anon_vma
);
443 * Similar to page_get_anon_vma() except it locks the anon_vma.
445 * Its a little more complex as it tries to keep the fast path to a single
446 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
447 * reference like with page_get_anon_vma() and then block on the mutex.
449 struct anon_vma
*page_lock_anon_vma(struct page
*page
)
451 struct anon_vma
*anon_vma
= NULL
;
452 struct anon_vma
*root_anon_vma
;
453 unsigned long anon_mapping
;
456 anon_mapping
= (unsigned long) ACCESS_ONCE(page
->mapping
);
457 if ((anon_mapping
& PAGE_MAPPING_FLAGS
) != PAGE_MAPPING_ANON
)
459 if (!page_mapped(page
))
462 anon_vma
= (struct anon_vma
*) (anon_mapping
- PAGE_MAPPING_ANON
);
463 root_anon_vma
= ACCESS_ONCE(anon_vma
->root
);
464 if (mutex_trylock(&root_anon_vma
->mutex
)) {
466 * If the page is still mapped, then this anon_vma is still
467 * its anon_vma, and holding the mutex ensures that it will
468 * not go away, see anon_vma_free().
470 if (!page_mapped(page
)) {
471 mutex_unlock(&root_anon_vma
->mutex
);
477 /* trylock failed, we got to sleep */
478 if (!atomic_inc_not_zero(&anon_vma
->refcount
)) {
483 if (!page_mapped(page
)) {
484 put_anon_vma(anon_vma
);
489 /* we pinned the anon_vma, its safe to sleep */
491 anon_vma_lock(anon_vma
);
493 if (atomic_dec_and_test(&anon_vma
->refcount
)) {
495 * Oops, we held the last refcount, release the lock
496 * and bail -- can't simply use put_anon_vma() because
497 * we'll deadlock on the anon_vma_lock() recursion.
499 anon_vma_unlock(anon_vma
);
500 __put_anon_vma(anon_vma
);
511 void page_unlock_anon_vma(struct anon_vma
*anon_vma
)
513 anon_vma_unlock(anon_vma
);
517 * At what user virtual address is page expected in @vma?
518 * Returns virtual address or -EFAULT if page's index/offset is not
519 * within the range mapped the @vma.
522 vma_address(struct page
*page
, struct vm_area_struct
*vma
)
524 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
525 unsigned long address
;
527 if (unlikely(is_vm_hugetlb_page(vma
)))
528 pgoff
= page
->index
<< huge_page_order(page_hstate(page
));
529 address
= vma
->vm_start
+ ((pgoff
- vma
->vm_pgoff
) << PAGE_SHIFT
);
530 if (unlikely(address
< vma
->vm_start
|| address
>= vma
->vm_end
)) {
531 /* page should be within @vma mapping range */
538 * At what user virtual address is page expected in vma?
539 * Caller should check the page is actually part of the vma.
541 unsigned long page_address_in_vma(struct page
*page
, struct vm_area_struct
*vma
)
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 return vma_address(page
, vma
);
562 * Check that @page is mapped at @address into @mm.
564 * If @sync is false, page_check_address may perform a racy check to avoid
565 * the page table lock when the pte is not present (helpful when reclaiming
566 * highly shared pages).
568 * On success returns with pte mapped and locked.
570 pte_t
*__page_check_address(struct page
*page
, struct mm_struct
*mm
,
571 unsigned long address
, spinlock_t
**ptlp
, int sync
)
579 if (unlikely(PageHuge(page
))) {
580 pte
= huge_pte_offset(mm
, address
);
581 ptl
= &mm
->page_table_lock
;
585 pgd
= pgd_offset(mm
, address
);
586 if (!pgd_present(*pgd
))
589 pud
= pud_offset(pgd
, address
);
590 if (!pud_present(*pud
))
593 pmd
= pmd_offset(pud
, address
);
594 if (!pmd_present(*pmd
))
596 if (pmd_trans_huge(*pmd
))
599 pte
= pte_offset_map(pmd
, address
);
600 /* Make a quick check before getting the lock */
601 if (!sync
&& !pte_present(*pte
)) {
606 ptl
= pte_lockptr(mm
, pmd
);
609 if (pte_present(*pte
) && page_to_pfn(page
) == pte_pfn(*pte
)) {
613 pte_unmap_unlock(pte
, ptl
);
618 * page_mapped_in_vma - check whether a page is really mapped in a VMA
619 * @page: the page to test
620 * @vma: the VMA to test
622 * Returns 1 if the page is mapped into the page tables of the VMA, 0
623 * if the page is not mapped into the page tables of this VMA. Only
624 * valid for normal file or anonymous VMAs.
626 int page_mapped_in_vma(struct page
*page
, struct vm_area_struct
*vma
)
628 unsigned long address
;
632 address
= vma_address(page
, vma
);
633 if (address
== -EFAULT
) /* out of vma range */
635 pte
= page_check_address(page
, vma
->vm_mm
, address
, &ptl
, 1);
636 if (!pte
) /* the page is not in this mm */
638 pte_unmap_unlock(pte
, ptl
);
644 * Subfunctions of page_referenced: page_referenced_one called
645 * repeatedly from either page_referenced_anon or page_referenced_file.
647 int page_referenced_one(struct page
*page
, struct vm_area_struct
*vma
,
648 unsigned long address
, unsigned int *mapcount
,
649 unsigned long *vm_flags
)
651 struct mm_struct
*mm
= vma
->vm_mm
;
654 if (unlikely(PageTransHuge(page
))) {
657 spin_lock(&mm
->page_table_lock
);
659 * rmap might return false positives; we must filter
660 * these out using page_check_address_pmd().
662 pmd
= page_check_address_pmd(page
, mm
, address
,
663 PAGE_CHECK_ADDRESS_PMD_FLAG
);
665 spin_unlock(&mm
->page_table_lock
);
669 if (vma
->vm_flags
& VM_LOCKED
) {
670 spin_unlock(&mm
->page_table_lock
);
671 *mapcount
= 0; /* break early from loop */
672 *vm_flags
|= VM_LOCKED
;
676 /* go ahead even if the pmd is pmd_trans_splitting() */
677 if (pmdp_clear_flush_young_notify(vma
, address
, pmd
))
679 spin_unlock(&mm
->page_table_lock
);
685 * rmap might return false positives; we must filter
686 * these out using page_check_address().
688 pte
= page_check_address(page
, mm
, address
, &ptl
, 0);
692 if (vma
->vm_flags
& VM_LOCKED
) {
693 pte_unmap_unlock(pte
, ptl
);
694 *mapcount
= 0; /* break early from loop */
695 *vm_flags
|= VM_LOCKED
;
699 if (ptep_clear_flush_young_notify(vma
, address
, pte
)) {
701 * Don't treat a reference through a sequentially read
702 * mapping as such. If the page has been used in
703 * another mapping, we will catch it; if this other
704 * mapping is already gone, the unmap path will have
705 * set PG_referenced or activated the page.
707 if (likely(!VM_SequentialReadHint(vma
)))
710 pte_unmap_unlock(pte
, ptl
);
716 *vm_flags
|= vma
->vm_flags
;
721 static int page_referenced_anon(struct page
*page
,
722 struct mem_cgroup
*memcg
,
723 unsigned long *vm_flags
)
725 unsigned int mapcount
;
726 struct anon_vma
*anon_vma
;
727 struct anon_vma_chain
*avc
;
730 anon_vma
= page_lock_anon_vma(page
);
734 mapcount
= page_mapcount(page
);
735 list_for_each_entry(avc
, &anon_vma
->head
, same_anon_vma
) {
736 struct vm_area_struct
*vma
= avc
->vma
;
737 unsigned long address
= vma_address(page
, vma
);
738 if (address
== -EFAULT
)
741 * If we are reclaiming on behalf of a cgroup, skip
742 * counting on behalf of references from different
745 if (memcg
&& !mm_match_cgroup(vma
->vm_mm
, memcg
))
747 referenced
+= page_referenced_one(page
, vma
, address
,
748 &mapcount
, vm_flags
);
753 page_unlock_anon_vma(anon_vma
);
758 * page_referenced_file - referenced check for object-based rmap
759 * @page: the page we're checking references on.
760 * @memcg: target memory control group
761 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
763 * For an object-based mapped page, find all the places it is mapped and
764 * check/clear the referenced flag. This is done by following the page->mapping
765 * pointer, then walking the chain of vmas it holds. It returns the number
766 * of references it found.
768 * This function is only called from page_referenced for object-based pages.
770 static int page_referenced_file(struct page
*page
,
771 struct mem_cgroup
*memcg
,
772 unsigned long *vm_flags
)
774 unsigned int mapcount
;
775 struct address_space
*mapping
= page
->mapping
;
776 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
777 struct vm_area_struct
*vma
;
781 * The caller's checks on page->mapping and !PageAnon have made
782 * sure that this is a file page: the check for page->mapping
783 * excludes the case just before it gets set on an anon page.
785 BUG_ON(PageAnon(page
));
788 * The page lock not only makes sure that page->mapping cannot
789 * suddenly be NULLified by truncation, it makes sure that the
790 * structure at mapping cannot be freed and reused yet,
791 * so we can safely take mapping->i_mmap_mutex.
793 BUG_ON(!PageLocked(page
));
795 mutex_lock(&mapping
->i_mmap_mutex
);
798 * i_mmap_mutex does not stabilize mapcount at all, but mapcount
799 * is more likely to be accurate if we note it after spinning.
801 mapcount
= page_mapcount(page
);
803 vma_interval_tree_foreach(vma
, &mapping
->i_mmap
, pgoff
, pgoff
) {
804 unsigned long address
= vma_address(page
, vma
);
805 if (address
== -EFAULT
)
808 * If we are reclaiming on behalf of a cgroup, skip
809 * counting on behalf of references from different
812 if (memcg
&& !mm_match_cgroup(vma
->vm_mm
, memcg
))
814 referenced
+= page_referenced_one(page
, vma
, address
,
815 &mapcount
, vm_flags
);
820 mutex_unlock(&mapping
->i_mmap_mutex
);
825 * page_referenced - test if the page was referenced
826 * @page: the page to test
827 * @is_locked: caller holds lock on the page
828 * @memcg: target memory cgroup
829 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
831 * Quick test_and_clear_referenced for all mappings to a page,
832 * returns the number of ptes which referenced the page.
834 int page_referenced(struct page
*page
,
836 struct mem_cgroup
*memcg
,
837 unsigned long *vm_flags
)
843 if (page_mapped(page
) && page_rmapping(page
)) {
844 if (!is_locked
&& (!PageAnon(page
) || PageKsm(page
))) {
845 we_locked
= trylock_page(page
);
851 if (unlikely(PageKsm(page
)))
852 referenced
+= page_referenced_ksm(page
, memcg
,
854 else if (PageAnon(page
))
855 referenced
+= page_referenced_anon(page
, memcg
,
857 else if (page
->mapping
)
858 referenced
+= page_referenced_file(page
, memcg
,
863 if (page_test_and_clear_young(page_to_pfn(page
)))
870 static int page_mkclean_one(struct page
*page
, struct vm_area_struct
*vma
,
871 unsigned long address
)
873 struct mm_struct
*mm
= vma
->vm_mm
;
878 pte
= page_check_address(page
, mm
, address
, &ptl
, 1);
882 if (pte_dirty(*pte
) || pte_write(*pte
)) {
885 flush_cache_page(vma
, address
, pte_pfn(*pte
));
886 entry
= ptep_clear_flush_notify(vma
, address
, pte
);
887 entry
= pte_wrprotect(entry
);
888 entry
= pte_mkclean(entry
);
889 set_pte_at(mm
, address
, pte
, entry
);
893 pte_unmap_unlock(pte
, ptl
);
898 static int page_mkclean_file(struct address_space
*mapping
, struct page
*page
)
900 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
901 struct vm_area_struct
*vma
;
904 BUG_ON(PageAnon(page
));
906 mutex_lock(&mapping
->i_mmap_mutex
);
907 vma_interval_tree_foreach(vma
, &mapping
->i_mmap
, pgoff
, pgoff
) {
908 if (vma
->vm_flags
& VM_SHARED
) {
909 unsigned long address
= vma_address(page
, vma
);
910 if (address
== -EFAULT
)
912 ret
+= page_mkclean_one(page
, vma
, address
);
915 mutex_unlock(&mapping
->i_mmap_mutex
);
919 int page_mkclean(struct page
*page
)
923 BUG_ON(!PageLocked(page
));
925 if (page_mapped(page
)) {
926 struct address_space
*mapping
= page_mapping(page
);
928 ret
= page_mkclean_file(mapping
, page
);
929 if (page_test_and_clear_dirty(page_to_pfn(page
), 1))
936 EXPORT_SYMBOL_GPL(page_mkclean
);
939 * page_move_anon_rmap - move a page to our anon_vma
940 * @page: the page to move to our anon_vma
941 * @vma: the vma the page belongs to
942 * @address: the user virtual address mapped
944 * When a page belongs exclusively to one process after a COW event,
945 * that page can be moved into the anon_vma that belongs to just that
946 * process, so the rmap code will not search the parent or sibling
949 void page_move_anon_rmap(struct page
*page
,
950 struct vm_area_struct
*vma
, unsigned long address
)
952 struct anon_vma
*anon_vma
= vma
->anon_vma
;
954 VM_BUG_ON(!PageLocked(page
));
955 VM_BUG_ON(!anon_vma
);
956 VM_BUG_ON(page
->index
!= linear_page_index(vma
, address
));
958 anon_vma
= (void *) anon_vma
+ PAGE_MAPPING_ANON
;
959 page
->mapping
= (struct address_space
*) anon_vma
;
963 * __page_set_anon_rmap - set up new anonymous rmap
964 * @page: Page to add to rmap
965 * @vma: VM area to add page to.
966 * @address: User virtual address of the mapping
967 * @exclusive: the page is exclusively owned by the current process
969 static void __page_set_anon_rmap(struct page
*page
,
970 struct vm_area_struct
*vma
, unsigned long address
, int exclusive
)
972 struct anon_vma
*anon_vma
= vma
->anon_vma
;
980 * If the page isn't exclusively mapped into this vma,
981 * we must use the _oldest_ possible anon_vma for the
985 anon_vma
= anon_vma
->root
;
987 anon_vma
= (void *) anon_vma
+ PAGE_MAPPING_ANON
;
988 page
->mapping
= (struct address_space
*) anon_vma
;
989 page
->index
= linear_page_index(vma
, address
);
993 * __page_check_anon_rmap - sanity check anonymous rmap addition
994 * @page: the page to add the mapping to
995 * @vma: the vm area in which the mapping is added
996 * @address: the user virtual address mapped
998 static void __page_check_anon_rmap(struct page
*page
,
999 struct vm_area_struct
*vma
, unsigned long address
)
1001 #ifdef CONFIG_DEBUG_VM
1003 * The page's anon-rmap details (mapping and index) are guaranteed to
1004 * be set up correctly at this point.
1006 * We have exclusion against page_add_anon_rmap because the caller
1007 * always holds the page locked, except if called from page_dup_rmap,
1008 * in which case the page is already known to be setup.
1010 * We have exclusion against page_add_new_anon_rmap because those pages
1011 * are initially only visible via the pagetables, and the pte is locked
1012 * over the call to page_add_new_anon_rmap.
1014 BUG_ON(page_anon_vma(page
)->root
!= vma
->anon_vma
->root
);
1015 BUG_ON(page
->index
!= linear_page_index(vma
, address
));
1020 * page_add_anon_rmap - add pte mapping to an anonymous page
1021 * @page: the page to add the mapping to
1022 * @vma: the vm area in which the mapping is added
1023 * @address: the user virtual address mapped
1025 * The caller needs to hold the pte lock, and the page must be locked in
1026 * the anon_vma case: to serialize mapping,index checking after setting,
1027 * and to ensure that PageAnon is not being upgraded racily to PageKsm
1028 * (but PageKsm is never downgraded to PageAnon).
1030 void page_add_anon_rmap(struct page
*page
,
1031 struct vm_area_struct
*vma
, unsigned long address
)
1033 do_page_add_anon_rmap(page
, vma
, address
, 0);
1037 * Special version of the above for do_swap_page, which often runs
1038 * into pages that are exclusively owned by the current process.
1039 * Everybody else should continue to use page_add_anon_rmap above.
1041 void do_page_add_anon_rmap(struct page
*page
,
1042 struct vm_area_struct
*vma
, unsigned long address
, int exclusive
)
1044 int first
= atomic_inc_and_test(&page
->_mapcount
);
1046 if (!PageTransHuge(page
))
1047 __inc_zone_page_state(page
, NR_ANON_PAGES
);
1049 __inc_zone_page_state(page
,
1050 NR_ANON_TRANSPARENT_HUGEPAGES
);
1052 if (unlikely(PageKsm(page
)))
1055 VM_BUG_ON(!PageLocked(page
));
1056 /* address might be in next vma when migration races vma_adjust */
1058 __page_set_anon_rmap(page
, vma
, address
, exclusive
);
1060 __page_check_anon_rmap(page
, vma
, address
);
1064 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1065 * @page: the page to add the mapping to
1066 * @vma: the vm area in which the mapping is added
1067 * @address: the user virtual address mapped
1069 * Same as page_add_anon_rmap but must only be called on *new* pages.
1070 * This means the inc-and-test can be bypassed.
1071 * Page does not have to be locked.
1073 void page_add_new_anon_rmap(struct page
*page
,
1074 struct vm_area_struct
*vma
, unsigned long address
)
1076 VM_BUG_ON(address
< vma
->vm_start
|| address
>= vma
->vm_end
);
1077 SetPageSwapBacked(page
);
1078 atomic_set(&page
->_mapcount
, 0); /* increment count (starts at -1) */
1079 if (!PageTransHuge(page
))
1080 __inc_zone_page_state(page
, NR_ANON_PAGES
);
1082 __inc_zone_page_state(page
, NR_ANON_TRANSPARENT_HUGEPAGES
);
1083 __page_set_anon_rmap(page
, vma
, address
, 1);
1084 if (page_evictable(page
, vma
))
1085 lru_cache_add_lru(page
, LRU_ACTIVE_ANON
);
1087 add_page_to_unevictable_list(page
);
1091 * page_add_file_rmap - add pte mapping to a file page
1092 * @page: the page to add the mapping to
1094 * The caller needs to hold the pte lock.
1096 void page_add_file_rmap(struct page
*page
)
1099 unsigned long flags
;
1101 mem_cgroup_begin_update_page_stat(page
, &locked
, &flags
);
1102 if (atomic_inc_and_test(&page
->_mapcount
)) {
1103 __inc_zone_page_state(page
, NR_FILE_MAPPED
);
1104 mem_cgroup_inc_page_stat(page
, MEMCG_NR_FILE_MAPPED
);
1106 mem_cgroup_end_update_page_stat(page
, &locked
, &flags
);
1110 * page_remove_rmap - take down pte mapping from a page
1111 * @page: page to remove mapping from
1113 * The caller needs to hold the pte lock.
1115 void page_remove_rmap(struct page
*page
)
1117 bool anon
= PageAnon(page
);
1119 unsigned long flags
;
1122 * The anon case has no mem_cgroup page_stat to update; but may
1123 * uncharge_page() below, where the lock ordering can deadlock if
1124 * we hold the lock against page_stat move: so avoid it on anon.
1127 mem_cgroup_begin_update_page_stat(page
, &locked
, &flags
);
1129 /* page still mapped by someone else? */
1130 if (!atomic_add_negative(-1, &page
->_mapcount
))
1134 * Now that the last pte has gone, s390 must transfer dirty
1135 * flag from storage key to struct page. We can usually skip
1136 * this if the page is anon, so about to be freed; but perhaps
1137 * not if it's in swapcache - there might be another pte slot
1138 * containing the swap entry, but page not yet written to swap.
1140 if ((!anon
|| PageSwapCache(page
)) &&
1141 page_test_and_clear_dirty(page_to_pfn(page
), 1))
1142 set_page_dirty(page
);
1144 * Hugepages are not counted in NR_ANON_PAGES nor NR_FILE_MAPPED
1145 * and not charged by memcg for now.
1147 if (unlikely(PageHuge(page
)))
1150 mem_cgroup_uncharge_page(page
);
1151 if (!PageTransHuge(page
))
1152 __dec_zone_page_state(page
, NR_ANON_PAGES
);
1154 __dec_zone_page_state(page
,
1155 NR_ANON_TRANSPARENT_HUGEPAGES
);
1157 __dec_zone_page_state(page
, NR_FILE_MAPPED
);
1158 mem_cgroup_dec_page_stat(page
, MEMCG_NR_FILE_MAPPED
);
1161 * It would be tidy to reset the PageAnon mapping here,
1162 * but that might overwrite a racing page_add_anon_rmap
1163 * which increments mapcount after us but sets mapping
1164 * before us: so leave the reset to free_hot_cold_page,
1165 * and remember that it's only reliable while mapped.
1166 * Leaving it set also helps swapoff to reinstate ptes
1167 * faster for those pages still in swapcache.
1171 mem_cgroup_end_update_page_stat(page
, &locked
, &flags
);
1175 * Subfunctions of try_to_unmap: try_to_unmap_one called
1176 * repeatedly from try_to_unmap_ksm, try_to_unmap_anon or try_to_unmap_file.
1178 int try_to_unmap_one(struct page
*page
, struct vm_area_struct
*vma
,
1179 unsigned long address
, enum ttu_flags flags
)
1181 struct mm_struct
*mm
= vma
->vm_mm
;
1185 int ret
= SWAP_AGAIN
;
1187 pte
= page_check_address(page
, mm
, address
, &ptl
, 0);
1192 * If the page is mlock()d, we cannot swap it out.
1193 * If it's recently referenced (perhaps page_referenced
1194 * skipped over this mm) then we should reactivate it.
1196 if (!(flags
& TTU_IGNORE_MLOCK
)) {
1197 if (vma
->vm_flags
& VM_LOCKED
)
1200 if (TTU_ACTION(flags
) == TTU_MUNLOCK
)
1203 if (!(flags
& TTU_IGNORE_ACCESS
)) {
1204 if (ptep_clear_flush_young_notify(vma
, address
, pte
)) {
1210 /* Nuke the page table entry. */
1211 flush_cache_page(vma
, address
, page_to_pfn(page
));
1212 pteval
= ptep_clear_flush_notify(vma
, address
, pte
);
1214 /* Move the dirty bit to the physical page now the pte is gone. */
1215 if (pte_dirty(pteval
))
1216 set_page_dirty(page
);
1218 /* Update high watermark before we lower rss */
1219 update_hiwater_rss(mm
);
1221 if (PageHWPoison(page
) && !(flags
& TTU_IGNORE_HWPOISON
)) {
1223 dec_mm_counter(mm
, MM_ANONPAGES
);
1225 dec_mm_counter(mm
, MM_FILEPAGES
);
1226 set_pte_at(mm
, address
, pte
,
1227 swp_entry_to_pte(make_hwpoison_entry(page
)));
1228 } else if (PageAnon(page
)) {
1229 swp_entry_t entry
= { .val
= page_private(page
) };
1231 if (PageSwapCache(page
)) {
1233 * Store the swap location in the pte.
1234 * See handle_pte_fault() ...
1236 if (swap_duplicate(entry
) < 0) {
1237 set_pte_at(mm
, address
, pte
, pteval
);
1241 if (list_empty(&mm
->mmlist
)) {
1242 spin_lock(&mmlist_lock
);
1243 if (list_empty(&mm
->mmlist
))
1244 list_add(&mm
->mmlist
, &init_mm
.mmlist
);
1245 spin_unlock(&mmlist_lock
);
1247 dec_mm_counter(mm
, MM_ANONPAGES
);
1248 inc_mm_counter(mm
, MM_SWAPENTS
);
1249 } else if (IS_ENABLED(CONFIG_MIGRATION
)) {
1251 * Store the pfn of the page in a special migration
1252 * pte. do_swap_page() will wait until the migration
1253 * pte is removed and then restart fault handling.
1255 BUG_ON(TTU_ACTION(flags
) != TTU_MIGRATION
);
1256 entry
= make_migration_entry(page
, pte_write(pteval
));
1258 set_pte_at(mm
, address
, pte
, swp_entry_to_pte(entry
));
1259 BUG_ON(pte_file(*pte
));
1260 } else if (IS_ENABLED(CONFIG_MIGRATION
) &&
1261 (TTU_ACTION(flags
) == TTU_MIGRATION
)) {
1262 /* Establish migration entry for a file page */
1264 entry
= make_migration_entry(page
, pte_write(pteval
));
1265 set_pte_at(mm
, address
, pte
, swp_entry_to_pte(entry
));
1267 dec_mm_counter(mm
, MM_FILEPAGES
);
1269 page_remove_rmap(page
);
1270 page_cache_release(page
);
1273 pte_unmap_unlock(pte
, ptl
);
1278 pte_unmap_unlock(pte
, ptl
);
1282 * We need mmap_sem locking, Otherwise VM_LOCKED check makes
1283 * unstable result and race. Plus, We can't wait here because
1284 * we now hold anon_vma->mutex or mapping->i_mmap_mutex.
1285 * if trylock failed, the page remain in evictable lru and later
1286 * vmscan could retry to move the page to unevictable lru if the
1287 * page is actually mlocked.
1289 if (down_read_trylock(&vma
->vm_mm
->mmap_sem
)) {
1290 if (vma
->vm_flags
& VM_LOCKED
) {
1291 mlock_vma_page(page
);
1294 up_read(&vma
->vm_mm
->mmap_sem
);
1300 * objrmap doesn't work for nonlinear VMAs because the assumption that
1301 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
1302 * Consequently, given a particular page and its ->index, we cannot locate the
1303 * ptes which are mapping that page without an exhaustive linear search.
1305 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
1306 * maps the file to which the target page belongs. The ->vm_private_data field
1307 * holds the current cursor into that scan. Successive searches will circulate
1308 * around the vma's virtual address space.
1310 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
1311 * more scanning pressure is placed against them as well. Eventually pages
1312 * will become fully unmapped and are eligible for eviction.
1314 * For very sparsely populated VMAs this is a little inefficient - chances are
1315 * there there won't be many ptes located within the scan cluster. In this case
1316 * maybe we could scan further - to the end of the pte page, perhaps.
1318 * Mlocked pages: check VM_LOCKED under mmap_sem held for read, if we can
1319 * acquire it without blocking. If vma locked, mlock the pages in the cluster,
1320 * rather than unmapping them. If we encounter the "check_page" that vmscan is
1321 * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN.
1323 #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
1324 #define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
1326 static int try_to_unmap_cluster(unsigned long cursor
, unsigned int *mapcount
,
1327 struct vm_area_struct
*vma
, struct page
*check_page
)
1329 struct mm_struct
*mm
= vma
->vm_mm
;
1337 unsigned long address
;
1339 int ret
= SWAP_AGAIN
;
1342 address
= (vma
->vm_start
+ cursor
) & CLUSTER_MASK
;
1343 end
= address
+ CLUSTER_SIZE
;
1344 if (address
< vma
->vm_start
)
1345 address
= vma
->vm_start
;
1346 if (end
> vma
->vm_end
)
1349 pgd
= pgd_offset(mm
, address
);
1350 if (!pgd_present(*pgd
))
1353 pud
= pud_offset(pgd
, address
);
1354 if (!pud_present(*pud
))
1357 pmd
= pmd_offset(pud
, address
);
1358 if (!pmd_present(*pmd
))
1362 * If we can acquire the mmap_sem for read, and vma is VM_LOCKED,
1363 * keep the sem while scanning the cluster for mlocking pages.
1365 if (down_read_trylock(&vma
->vm_mm
->mmap_sem
)) {
1366 locked_vma
= (vma
->vm_flags
& VM_LOCKED
);
1368 up_read(&vma
->vm_mm
->mmap_sem
); /* don't need it */
1371 pte
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
1373 /* Update high watermark before we lower rss */
1374 update_hiwater_rss(mm
);
1376 for (; address
< end
; pte
++, address
+= PAGE_SIZE
) {
1377 if (!pte_present(*pte
))
1379 page
= vm_normal_page(vma
, address
, *pte
);
1380 BUG_ON(!page
|| PageAnon(page
));
1383 mlock_vma_page(page
); /* no-op if already mlocked */
1384 if (page
== check_page
)
1386 continue; /* don't unmap */
1389 if (ptep_clear_flush_young_notify(vma
, address
, pte
))
1392 /* Nuke the page table entry. */
1393 flush_cache_page(vma
, address
, pte_pfn(*pte
));
1394 pteval
= ptep_clear_flush_notify(vma
, address
, pte
);
1396 /* If nonlinear, store the file page offset in the pte. */
1397 if (page
->index
!= linear_page_index(vma
, address
))
1398 set_pte_at(mm
, address
, pte
, pgoff_to_pte(page
->index
));
1400 /* Move the dirty bit to the physical page now the pte is gone. */
1401 if (pte_dirty(pteval
))
1402 set_page_dirty(page
);
1404 page_remove_rmap(page
);
1405 page_cache_release(page
);
1406 dec_mm_counter(mm
, MM_FILEPAGES
);
1409 pte_unmap_unlock(pte
- 1, ptl
);
1411 up_read(&vma
->vm_mm
->mmap_sem
);
1415 bool is_vma_temporary_stack(struct vm_area_struct
*vma
)
1417 int maybe_stack
= vma
->vm_flags
& (VM_GROWSDOWN
| VM_GROWSUP
);
1422 if ((vma
->vm_flags
& VM_STACK_INCOMPLETE_SETUP
) ==
1423 VM_STACK_INCOMPLETE_SETUP
)
1430 * try_to_unmap_anon - unmap or unlock anonymous page using the object-based
1432 * @page: the page to unmap/unlock
1433 * @flags: action and flags
1435 * Find all the mappings of a page using the mapping pointer and the vma chains
1436 * contained in the anon_vma struct it points to.
1438 * This function is only called from try_to_unmap/try_to_munlock for
1440 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1441 * where the page was found will be held for write. So, we won't recheck
1442 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1445 static int try_to_unmap_anon(struct page
*page
, enum ttu_flags flags
)
1447 struct anon_vma
*anon_vma
;
1448 struct anon_vma_chain
*avc
;
1449 int ret
= SWAP_AGAIN
;
1451 anon_vma
= page_lock_anon_vma(page
);
1455 list_for_each_entry(avc
, &anon_vma
->head
, same_anon_vma
) {
1456 struct vm_area_struct
*vma
= avc
->vma
;
1457 unsigned long address
;
1460 * During exec, a temporary VMA is setup and later moved.
1461 * The VMA is moved under the anon_vma lock but not the
1462 * page tables leading to a race where migration cannot
1463 * find the migration ptes. Rather than increasing the
1464 * locking requirements of exec(), migration skips
1465 * temporary VMAs until after exec() completes.
1467 if (IS_ENABLED(CONFIG_MIGRATION
) && (flags
& TTU_MIGRATION
) &&
1468 is_vma_temporary_stack(vma
))
1471 address
= vma_address(page
, vma
);
1472 if (address
== -EFAULT
)
1474 ret
= try_to_unmap_one(page
, vma
, address
, flags
);
1475 if (ret
!= SWAP_AGAIN
|| !page_mapped(page
))
1479 page_unlock_anon_vma(anon_vma
);
1484 * try_to_unmap_file - unmap/unlock file page using the object-based rmap method
1485 * @page: the page to unmap/unlock
1486 * @flags: action and flags
1488 * Find all the mappings of a page using the mapping pointer and the vma chains
1489 * contained in the address_space struct it points to.
1491 * This function is only called from try_to_unmap/try_to_munlock for
1492 * object-based pages.
1493 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1494 * where the page was found will be held for write. So, we won't recheck
1495 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1498 static int try_to_unmap_file(struct page
*page
, enum ttu_flags flags
)
1500 struct address_space
*mapping
= page
->mapping
;
1501 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
1502 struct vm_area_struct
*vma
;
1503 int ret
= SWAP_AGAIN
;
1504 unsigned long cursor
;
1505 unsigned long max_nl_cursor
= 0;
1506 unsigned long max_nl_size
= 0;
1507 unsigned int mapcount
;
1509 mutex_lock(&mapping
->i_mmap_mutex
);
1510 vma_interval_tree_foreach(vma
, &mapping
->i_mmap
, pgoff
, pgoff
) {
1511 unsigned long address
= vma_address(page
, vma
);
1512 if (address
== -EFAULT
)
1514 ret
= try_to_unmap_one(page
, vma
, address
, flags
);
1515 if (ret
!= SWAP_AGAIN
|| !page_mapped(page
))
1519 if (list_empty(&mapping
->i_mmap_nonlinear
))
1523 * We don't bother to try to find the munlocked page in nonlinears.
1524 * It's costly. Instead, later, page reclaim logic may call
1525 * try_to_unmap(TTU_MUNLOCK) and recover PG_mlocked lazily.
1527 if (TTU_ACTION(flags
) == TTU_MUNLOCK
)
1530 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
,
1532 cursor
= (unsigned long) vma
->vm_private_data
;
1533 if (cursor
> max_nl_cursor
)
1534 max_nl_cursor
= cursor
;
1535 cursor
= vma
->vm_end
- vma
->vm_start
;
1536 if (cursor
> max_nl_size
)
1537 max_nl_size
= cursor
;
1540 if (max_nl_size
== 0) { /* all nonlinears locked or reserved ? */
1546 * We don't try to search for this page in the nonlinear vmas,
1547 * and page_referenced wouldn't have found it anyway. Instead
1548 * just walk the nonlinear vmas trying to age and unmap some.
1549 * The mapcount of the page we came in with is irrelevant,
1550 * but even so use it as a guide to how hard we should try?
1552 mapcount
= page_mapcount(page
);
1557 max_nl_size
= (max_nl_size
+ CLUSTER_SIZE
- 1) & CLUSTER_MASK
;
1558 if (max_nl_cursor
== 0)
1559 max_nl_cursor
= CLUSTER_SIZE
;
1562 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
,
1564 cursor
= (unsigned long) vma
->vm_private_data
;
1565 while ( cursor
< max_nl_cursor
&&
1566 cursor
< vma
->vm_end
- vma
->vm_start
) {
1567 if (try_to_unmap_cluster(cursor
, &mapcount
,
1568 vma
, page
) == SWAP_MLOCK
)
1570 cursor
+= CLUSTER_SIZE
;
1571 vma
->vm_private_data
= (void *) cursor
;
1572 if ((int)mapcount
<= 0)
1575 vma
->vm_private_data
= (void *) max_nl_cursor
;
1578 max_nl_cursor
+= CLUSTER_SIZE
;
1579 } while (max_nl_cursor
<= max_nl_size
);
1582 * Don't loop forever (perhaps all the remaining pages are
1583 * in locked vmas). Reset cursor on all unreserved nonlinear
1584 * vmas, now forgetting on which ones it had fallen behind.
1586 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
, shared
.nonlinear
)
1587 vma
->vm_private_data
= NULL
;
1589 mutex_unlock(&mapping
->i_mmap_mutex
);
1594 * try_to_unmap - try to remove all page table mappings to a page
1595 * @page: the page to get unmapped
1596 * @flags: action and flags
1598 * Tries to remove all the page table entries which are mapping this
1599 * page, used in the pageout path. Caller must hold the page lock.
1600 * Return values are:
1602 * SWAP_SUCCESS - we succeeded in removing all mappings
1603 * SWAP_AGAIN - we missed a mapping, try again later
1604 * SWAP_FAIL - the page is unswappable
1605 * SWAP_MLOCK - page is mlocked.
1607 int try_to_unmap(struct page
*page
, enum ttu_flags flags
)
1611 BUG_ON(!PageLocked(page
));
1612 VM_BUG_ON(!PageHuge(page
) && PageTransHuge(page
));
1614 if (unlikely(PageKsm(page
)))
1615 ret
= try_to_unmap_ksm(page
, flags
);
1616 else if (PageAnon(page
))
1617 ret
= try_to_unmap_anon(page
, flags
);
1619 ret
= try_to_unmap_file(page
, flags
);
1620 if (ret
!= SWAP_MLOCK
&& !page_mapped(page
))
1626 * try_to_munlock - try to munlock a page
1627 * @page: the page to be munlocked
1629 * Called from munlock code. Checks all of the VMAs mapping the page
1630 * to make sure nobody else has this page mlocked. The page will be
1631 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1633 * Return values are:
1635 * SWAP_AGAIN - no vma is holding page mlocked, or,
1636 * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem
1637 * SWAP_FAIL - page cannot be located at present
1638 * SWAP_MLOCK - page is now mlocked.
1640 int try_to_munlock(struct page
*page
)
1642 VM_BUG_ON(!PageLocked(page
) || PageLRU(page
));
1644 if (unlikely(PageKsm(page
)))
1645 return try_to_unmap_ksm(page
, TTU_MUNLOCK
);
1646 else if (PageAnon(page
))
1647 return try_to_unmap_anon(page
, TTU_MUNLOCK
);
1649 return try_to_unmap_file(page
, TTU_MUNLOCK
);
1652 void __put_anon_vma(struct anon_vma
*anon_vma
)
1654 struct anon_vma
*root
= anon_vma
->root
;
1656 if (root
!= anon_vma
&& atomic_dec_and_test(&root
->refcount
))
1657 anon_vma_free(root
);
1659 anon_vma_free(anon_vma
);
1662 #ifdef CONFIG_MIGRATION
1664 * rmap_walk() and its helpers rmap_walk_anon() and rmap_walk_file():
1665 * Called by migrate.c to remove migration ptes, but might be used more later.
1667 static int rmap_walk_anon(struct page
*page
, int (*rmap_one
)(struct page
*,
1668 struct vm_area_struct
*, unsigned long, void *), void *arg
)
1670 struct anon_vma
*anon_vma
;
1671 struct anon_vma_chain
*avc
;
1672 int ret
= SWAP_AGAIN
;
1675 * Note: remove_migration_ptes() cannot use page_lock_anon_vma()
1676 * because that depends on page_mapped(); but not all its usages
1677 * are holding mmap_sem. Users without mmap_sem are required to
1678 * take a reference count to prevent the anon_vma disappearing
1680 anon_vma
= page_anon_vma(page
);
1683 anon_vma_lock(anon_vma
);
1684 list_for_each_entry(avc
, &anon_vma
->head
, same_anon_vma
) {
1685 struct vm_area_struct
*vma
= avc
->vma
;
1686 unsigned long address
= vma_address(page
, vma
);
1687 if (address
== -EFAULT
)
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 if (address
== -EFAULT
)
1712 ret
= rmap_one(page
, vma
, address
, arg
);
1713 if (ret
!= SWAP_AGAIN
)
1717 * No nonlinear handling: being always shared, nonlinear vmas
1718 * never contain migration ptes. Decide what to do about this
1719 * limitation to linear when we need rmap_walk() on nonlinear.
1721 mutex_unlock(&mapping
->i_mmap_mutex
);
1725 int rmap_walk(struct page
*page
, int (*rmap_one
)(struct page
*,
1726 struct vm_area_struct
*, unsigned long, void *), void *arg
)
1728 VM_BUG_ON(!PageLocked(page
));
1730 if (unlikely(PageKsm(page
)))
1731 return rmap_walk_ksm(page
, rmap_one
, arg
);
1732 else if (PageAnon(page
))
1733 return rmap_walk_anon(page
, rmap_one
, arg
);
1735 return rmap_walk_file(page
, rmap_one
, arg
);
1737 #endif /* CONFIG_MIGRATION */
1739 #ifdef CONFIG_HUGETLB_PAGE
1741 * The following three functions are for anonymous (private mapped) hugepages.
1742 * Unlike common anonymous pages, anonymous hugepages have no accounting code
1743 * and no lru code, because we handle hugepages differently from common pages.
1745 static void __hugepage_set_anon_rmap(struct page
*page
,
1746 struct vm_area_struct
*vma
, unsigned long address
, int exclusive
)
1748 struct anon_vma
*anon_vma
= vma
->anon_vma
;
1755 anon_vma
= anon_vma
->root
;
1757 anon_vma
= (void *) anon_vma
+ PAGE_MAPPING_ANON
;
1758 page
->mapping
= (struct address_space
*) anon_vma
;
1759 page
->index
= linear_page_index(vma
, address
);
1762 void hugepage_add_anon_rmap(struct page
*page
,
1763 struct vm_area_struct
*vma
, unsigned long address
)
1765 struct anon_vma
*anon_vma
= vma
->anon_vma
;
1768 BUG_ON(!PageLocked(page
));
1770 /* address might be in next vma when migration races vma_adjust */
1771 first
= atomic_inc_and_test(&page
->_mapcount
);
1773 __hugepage_set_anon_rmap(page
, vma
, address
, 0);
1776 void hugepage_add_new_anon_rmap(struct page
*page
,
1777 struct vm_area_struct
*vma
, unsigned long address
)
1779 BUG_ON(address
< vma
->vm_start
|| address
>= vma
->vm_end
);
1780 atomic_set(&page
->_mapcount
, 0);
1781 __hugepage_set_anon_rmap(page
, vma
, address
, 1);
1783 #endif /* CONFIG_HUGETLB_PAGE */