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->rwsem,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>
59 #include <linux/backing-dev.h>
61 #include <asm/tlbflush.h>
65 #define MUTEX_RETRY_COUNT (65536)
66 #define MUTEX_RETRY_RESCHED (1024)
70 static struct kmem_cache
*anon_vma_cachep
;
71 static struct kmem_cache
*anon_vma_chain_cachep
;
73 static inline struct anon_vma
*anon_vma_alloc(void)
75 struct anon_vma
*anon_vma
;
77 anon_vma
= kmem_cache_alloc(anon_vma_cachep
, GFP_KERNEL
);
79 atomic_set(&anon_vma
->refcount
, 1);
81 * Initialise the anon_vma root to point to itself. If called
82 * from fork, the root will be reset to the parents anon_vma.
84 anon_vma
->root
= anon_vma
;
90 static inline void anon_vma_free(struct anon_vma
*anon_vma
)
92 VM_BUG_ON(atomic_read(&anon_vma
->refcount
));
95 * Synchronize against page_lock_anon_vma_read() such that
96 * we can safely hold the lock without the anon_vma getting
99 * Relies on the full mb implied by the atomic_dec_and_test() from
100 * put_anon_vma() against the acquire barrier implied by
101 * down_read_trylock() from page_lock_anon_vma_read(). This orders:
103 * page_lock_anon_vma_read() VS put_anon_vma()
104 * down_read_trylock() atomic_dec_and_test()
106 * atomic_read() rwsem_is_locked()
108 * LOCK should suffice since the actual taking of the lock must
109 * happen _before_ what follows.
112 if (anon_vma
->root
&& rwsem_is_locked(&anon_vma
->root
->rwsem
)) {
113 anon_vma_lock_write(anon_vma
);
114 anon_vma_unlock_write(anon_vma
);
117 kmem_cache_free(anon_vma_cachep
, anon_vma
);
120 static inline struct anon_vma_chain
*anon_vma_chain_alloc(gfp_t gfp
)
122 return kmem_cache_alloc(anon_vma_chain_cachep
, gfp
);
125 static void anon_vma_chain_free(struct anon_vma_chain
*anon_vma_chain
)
127 kmem_cache_free(anon_vma_chain_cachep
, anon_vma_chain
);
130 static void anon_vma_chain_link(struct vm_area_struct
*vma
,
131 struct anon_vma_chain
*avc
,
132 struct anon_vma
*anon_vma
)
135 avc
->anon_vma
= anon_vma
;
136 list_add(&avc
->same_vma
, &vma
->anon_vma_chain
);
137 anon_vma_interval_tree_insert(avc
, &anon_vma
->rb_root
);
141 * anon_vma_prepare - attach an anon_vma to a memory region
142 * @vma: the memory region in question
144 * This makes sure the memory mapping described by 'vma' has
145 * an 'anon_vma' attached to it, so that we can associate the
146 * anonymous pages mapped into it with that anon_vma.
148 * The common case will be that we already have one, but if
149 * not we either need to find an adjacent mapping that we
150 * can re-use the anon_vma from (very common when the only
151 * reason for splitting a vma has been mprotect()), or we
152 * allocate a new one.
154 * Anon-vma allocations are very subtle, because we may have
155 * optimistically looked up an anon_vma in page_lock_anon_vma_read()
156 * and that may actually touch the spinlock even in the newly
157 * allocated vma (it depends on RCU to make sure that the
158 * anon_vma isn't actually destroyed).
160 * As a result, we need to do proper anon_vma locking even
161 * for the new allocation. At the same time, we do not want
162 * to do any locking for the common case of already having
165 * This must be called with the mmap_sem held for reading.
167 int anon_vma_prepare(struct vm_area_struct
*vma
)
169 struct anon_vma
*anon_vma
= vma
->anon_vma
;
170 struct anon_vma_chain
*avc
;
173 if (unlikely(!anon_vma
)) {
174 struct mm_struct
*mm
= vma
->vm_mm
;
175 struct anon_vma
*allocated
;
177 avc
= anon_vma_chain_alloc(GFP_KERNEL
);
181 anon_vma
= find_mergeable_anon_vma(vma
);
184 anon_vma
= anon_vma_alloc();
185 if (unlikely(!anon_vma
))
186 goto out_enomem_free_avc
;
187 allocated
= anon_vma
;
190 anon_vma_lock_write(anon_vma
);
191 /* page_table_lock to protect against threads */
192 spin_lock(&mm
->page_table_lock
);
193 if (likely(!vma
->anon_vma
)) {
194 vma
->anon_vma
= anon_vma
;
195 anon_vma_chain_link(vma
, avc
, anon_vma
);
199 spin_unlock(&mm
->page_table_lock
);
200 anon_vma_unlock_write(anon_vma
);
202 if (unlikely(allocated
))
203 put_anon_vma(allocated
);
205 anon_vma_chain_free(avc
);
210 anon_vma_chain_free(avc
);
216 * This is a useful helper function for locking the anon_vma root as
217 * we traverse the vma->anon_vma_chain, looping over anon_vma's that
220 * Such anon_vma's should have the same root, so you'd expect to see
221 * just a single mutex_lock for the whole traversal.
223 static inline struct anon_vma
*lock_anon_vma_root(struct anon_vma
*root
, struct anon_vma
*anon_vma
)
225 struct anon_vma
*new_root
= anon_vma
->root
;
226 if (new_root
!= root
) {
227 if (WARN_ON_ONCE(root
))
228 up_write(&root
->rwsem
);
230 down_write(&root
->rwsem
);
235 static inline void unlock_anon_vma_root(struct anon_vma
*root
)
238 up_write(&root
->rwsem
);
242 * Attach the anon_vmas from src to dst.
243 * Returns 0 on success, -ENOMEM on failure.
245 int anon_vma_clone(struct vm_area_struct
*dst
, struct vm_area_struct
*src
)
247 struct anon_vma_chain
*avc
, *pavc
;
248 struct anon_vma
*root
= NULL
;
250 list_for_each_entry_reverse(pavc
, &src
->anon_vma_chain
, same_vma
) {
251 struct anon_vma
*anon_vma
;
253 avc
= anon_vma_chain_alloc(GFP_NOWAIT
| __GFP_NOWARN
);
254 if (unlikely(!avc
)) {
255 unlock_anon_vma_root(root
);
257 avc
= anon_vma_chain_alloc(GFP_KERNEL
);
261 anon_vma
= pavc
->anon_vma
;
262 root
= lock_anon_vma_root(root
, anon_vma
);
263 anon_vma_chain_link(dst
, avc
, anon_vma
);
265 unlock_anon_vma_root(root
);
269 unlink_anon_vmas(dst
);
274 * Attach vma to its own anon_vma, as well as to the anon_vmas that
275 * the corresponding VMA in the parent process is attached to.
276 * Returns 0 on success, non-zero on failure.
278 int anon_vma_fork(struct vm_area_struct
*vma
, struct vm_area_struct
*pvma
)
280 struct anon_vma_chain
*avc
;
281 struct anon_vma
*anon_vma
;
283 /* Don't bother if the parent process has no anon_vma here. */
288 * First, attach the new VMA to the parent VMA's anon_vmas,
289 * so rmap can find non-COWed pages in child processes.
291 if (anon_vma_clone(vma
, pvma
))
294 /* Then add our own anon_vma. */
295 anon_vma
= anon_vma_alloc();
298 avc
= anon_vma_chain_alloc(GFP_KERNEL
);
300 goto out_error_free_anon_vma
;
303 * The root anon_vma's spinlock is the lock actually used when we
304 * lock any of the anon_vmas in this anon_vma tree.
306 anon_vma
->root
= pvma
->anon_vma
->root
;
308 * With refcounts, an anon_vma can stay around longer than the
309 * process it belongs to. The root anon_vma needs to be pinned until
310 * this anon_vma is freed, because the lock lives in the root.
312 get_anon_vma(anon_vma
->root
);
313 /* Mark this anon_vma as the one where our new (COWed) pages go. */
314 vma
->anon_vma
= anon_vma
;
315 anon_vma_lock_write(anon_vma
);
316 anon_vma_chain_link(vma
, avc
, anon_vma
);
317 anon_vma_unlock_write(anon_vma
);
321 out_error_free_anon_vma
:
322 put_anon_vma(anon_vma
);
324 unlink_anon_vmas(vma
);
328 void unlink_anon_vmas(struct vm_area_struct
*vma
)
330 struct anon_vma_chain
*avc
, *next
;
331 struct anon_vma
*root
= NULL
;
334 * Unlink each anon_vma chained to the VMA. This list is ordered
335 * from newest to oldest, ensuring the root anon_vma gets freed last.
337 list_for_each_entry_safe(avc
, next
, &vma
->anon_vma_chain
, same_vma
) {
338 struct anon_vma
*anon_vma
= avc
->anon_vma
;
340 root
= lock_anon_vma_root(root
, anon_vma
);
341 anon_vma_interval_tree_remove(avc
, &anon_vma
->rb_root
);
344 * Leave empty anon_vmas on the list - we'll need
345 * to free them outside the lock.
347 if (RB_EMPTY_ROOT(&anon_vma
->rb_root
))
350 list_del(&avc
->same_vma
);
351 anon_vma_chain_free(avc
);
353 unlock_anon_vma_root(root
);
356 * Iterate the list once more, it now only contains empty and unlinked
357 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
358 * needing to write-acquire the anon_vma->root->rwsem.
360 list_for_each_entry_safe(avc
, next
, &vma
->anon_vma_chain
, same_vma
) {
361 struct anon_vma
*anon_vma
= avc
->anon_vma
;
363 put_anon_vma(anon_vma
);
365 list_del(&avc
->same_vma
);
366 anon_vma_chain_free(avc
);
370 static void anon_vma_ctor(void *data
)
372 struct anon_vma
*anon_vma
= data
;
374 init_rwsem(&anon_vma
->rwsem
);
375 atomic_set(&anon_vma
->refcount
, 0);
376 anon_vma
->rb_root
= RB_ROOT
;
379 void __init
anon_vma_init(void)
381 anon_vma_cachep
= kmem_cache_create("anon_vma", sizeof(struct anon_vma
),
382 0, SLAB_DESTROY_BY_RCU
|SLAB_PANIC
, anon_vma_ctor
);
383 anon_vma_chain_cachep
= KMEM_CACHE(anon_vma_chain
, SLAB_PANIC
);
387 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
389 * Since there is no serialization what so ever against page_remove_rmap()
390 * the best this function can do is return a locked anon_vma that might
391 * have been relevant to this page.
393 * The page might have been remapped to a different anon_vma or the anon_vma
394 * returned may already be freed (and even reused).
396 * In case it was remapped to a different anon_vma, the new anon_vma will be a
397 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
398 * ensure that any anon_vma obtained from the page will still be valid for as
399 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
401 * All users of this function must be very careful when walking the anon_vma
402 * chain and verify that the page in question is indeed mapped in it
403 * [ something equivalent to page_mapped_in_vma() ].
405 * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap()
406 * that the anon_vma pointer from page->mapping is valid if there is a
407 * mapcount, we can dereference the anon_vma after observing those.
409 struct anon_vma
*page_get_anon_vma(struct page
*page
)
411 struct anon_vma
*anon_vma
= NULL
;
412 unsigned long anon_mapping
;
415 anon_mapping
= (unsigned long) ACCESS_ONCE(page
->mapping
);
416 if ((anon_mapping
& PAGE_MAPPING_FLAGS
) != PAGE_MAPPING_ANON
)
418 if (!page_mapped(page
))
421 anon_vma
= (struct anon_vma
*) (anon_mapping
- PAGE_MAPPING_ANON
);
422 if (!atomic_inc_not_zero(&anon_vma
->refcount
)) {
428 * If this page is still mapped, then its anon_vma cannot have been
429 * freed. But if it has been unmapped, we have no security against the
430 * anon_vma structure being freed and reused (for another anon_vma:
431 * SLAB_DESTROY_BY_RCU guarantees that - so the atomic_inc_not_zero()
432 * above cannot corrupt).
434 if (!page_mapped(page
)) {
436 put_anon_vma(anon_vma
);
446 * Similar to page_get_anon_vma() except it locks the anon_vma.
448 * Its a little more complex as it tries to keep the fast path to a single
449 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
450 * reference like with page_get_anon_vma() and then block on the mutex.
452 struct anon_vma
*page_lock_anon_vma_read(struct page
*page
)
454 struct anon_vma
*anon_vma
= NULL
;
455 struct anon_vma
*root_anon_vma
;
456 unsigned long anon_mapping
;
459 anon_mapping
= (unsigned long) ACCESS_ONCE(page
->mapping
);
460 if ((anon_mapping
& PAGE_MAPPING_FLAGS
) != PAGE_MAPPING_ANON
)
462 if (!page_mapped(page
))
465 anon_vma
= (struct anon_vma
*) (anon_mapping
- PAGE_MAPPING_ANON
);
466 root_anon_vma
= ACCESS_ONCE(anon_vma
->root
);
467 if (down_read_trylock(&root_anon_vma
->rwsem
)) {
469 * If the page is still mapped, then this anon_vma is still
470 * its anon_vma, and holding the mutex ensures that it will
471 * not go away, see anon_vma_free().
473 if (!page_mapped(page
)) {
474 up_read(&root_anon_vma
->rwsem
);
480 /* trylock failed, we got to sleep */
481 if (!atomic_inc_not_zero(&anon_vma
->refcount
)) {
486 if (!page_mapped(page
)) {
488 put_anon_vma(anon_vma
);
492 /* we pinned the anon_vma, its safe to sleep */
494 anon_vma_lock_read(anon_vma
);
496 if (atomic_dec_and_test(&anon_vma
->refcount
)) {
498 * Oops, we held the last refcount, release the lock
499 * and bail -- can't simply use put_anon_vma() because
500 * we'll deadlock on the anon_vma_lock_write() recursion.
502 anon_vma_unlock_read(anon_vma
);
503 __put_anon_vma(anon_vma
);
514 void page_unlock_anon_vma_read(struct anon_vma
*anon_vma
)
516 anon_vma_unlock_read(anon_vma
);
520 * At what user virtual address is page expected in @vma?
522 static inline unsigned long
523 __vma_address(struct page
*page
, struct vm_area_struct
*vma
)
525 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
527 if (unlikely(is_vm_hugetlb_page(vma
)))
528 pgoff
= page
->index
<< huge_page_order(page_hstate(page
));
530 return vma
->vm_start
+ ((pgoff
- vma
->vm_pgoff
) << PAGE_SHIFT
);
534 vma_address(struct page
*page
, struct vm_area_struct
*vma
)
536 unsigned long address
= __vma_address(page
, vma
);
538 /* page should be within @vma mapping range */
539 VM_BUG_ON(address
< vma
->vm_start
|| address
>= vma
->vm_end
);
545 * At what user virtual address is page expected in vma?
546 * Caller should check the page is actually part of the vma.
548 unsigned long page_address_in_vma(struct page
*page
, struct vm_area_struct
*vma
)
550 unsigned long address
;
551 if (PageAnon(page
)) {
552 struct anon_vma
*page__anon_vma
= page_anon_vma(page
);
554 * Note: swapoff's unuse_vma() is more efficient with this
555 * check, and needs it to match anon_vma when KSM is active.
557 if (!vma
->anon_vma
|| !page__anon_vma
||
558 vma
->anon_vma
->root
!= page__anon_vma
->root
)
560 } else if (page
->mapping
&& !(vma
->vm_flags
& VM_NONLINEAR
)) {
562 vma
->vm_file
->f_mapping
!= page
->mapping
)
566 address
= __vma_address(page
, vma
);
567 if (unlikely(address
< vma
->vm_start
|| address
>= vma
->vm_end
))
572 pmd_t
*mm_find_pmd(struct mm_struct
*mm
, unsigned long address
)
578 pgd
= pgd_offset(mm
, address
);
579 if (!pgd_present(*pgd
))
582 pud
= pud_offset(pgd
, address
);
583 if (!pud_present(*pud
))
586 pmd
= pmd_offset(pud
, address
);
587 if (!pmd_present(*pmd
))
594 * Check that @page is mapped at @address into @mm.
596 * If @sync is false, page_check_address may perform a racy check to avoid
597 * the page table lock when the pte is not present (helpful when reclaiming
598 * highly shared pages).
600 * On success returns with pte mapped and locked.
602 pte_t
*__page_check_address(struct page
*page
, struct mm_struct
*mm
,
603 unsigned long address
, spinlock_t
**ptlp
, int sync
)
609 if (unlikely(PageHuge(page
))) {
610 /* when pud is not present, pte will be NULL */
611 pte
= huge_pte_offset(mm
, address
);
615 ptl
= &mm
->page_table_lock
;
619 pmd
= mm_find_pmd(mm
, address
);
623 if (pmd_trans_huge(*pmd
))
626 pte
= pte_offset_map(pmd
, address
);
627 /* Make a quick check before getting the lock */
628 if (!sync
&& !pte_present(*pte
)) {
633 ptl
= pte_lockptr(mm
, pmd
);
636 if (pte_present(*pte
) && page_to_pfn(page
) == pte_pfn(*pte
)) {
640 pte_unmap_unlock(pte
, ptl
);
645 * page_mapped_in_vma - check whether a page is really mapped in a VMA
646 * @page: the page to test
647 * @vma: the VMA to test
649 * Returns 1 if the page is mapped into the page tables of the VMA, 0
650 * if the page is not mapped into the page tables of this VMA. Only
651 * valid for normal file or anonymous VMAs.
653 int page_mapped_in_vma(struct page
*page
, struct vm_area_struct
*vma
)
655 unsigned long address
;
659 address
= __vma_address(page
, vma
);
660 if (unlikely(address
< vma
->vm_start
|| address
>= vma
->vm_end
))
662 pte
= page_check_address(page
, vma
->vm_mm
, address
, &ptl
, 1);
663 if (!pte
) /* the page is not in this mm */
665 pte_unmap_unlock(pte
, ptl
);
671 * Subfunctions of page_referenced: page_referenced_one called
672 * repeatedly from either page_referenced_anon or page_referenced_file.
674 int page_referenced_one(struct page
*page
, struct vm_area_struct
*vma
,
675 unsigned long address
, unsigned int *mapcount
,
676 unsigned long *vm_flags
)
678 struct mm_struct
*mm
= vma
->vm_mm
;
681 if (unlikely(PageTransHuge(page
))) {
684 spin_lock(&mm
->page_table_lock
);
686 * rmap might return false positives; we must filter
687 * these out using page_check_address_pmd().
689 pmd
= page_check_address_pmd(page
, mm
, address
,
690 PAGE_CHECK_ADDRESS_PMD_FLAG
);
692 spin_unlock(&mm
->page_table_lock
);
696 if (vma
->vm_flags
& VM_LOCKED
) {
697 spin_unlock(&mm
->page_table_lock
);
698 *mapcount
= 0; /* break early from loop */
699 *vm_flags
|= VM_LOCKED
;
703 /* go ahead even if the pmd is pmd_trans_splitting() */
704 if (pmdp_clear_flush_young_notify(vma
, address
, pmd
))
706 spin_unlock(&mm
->page_table_lock
);
712 * rmap might return false positives; we must filter
713 * these out using page_check_address().
715 pte
= page_check_address(page
, mm
, address
, &ptl
, 0);
719 if (vma
->vm_flags
& VM_LOCKED
) {
720 pte_unmap_unlock(pte
, ptl
);
721 *mapcount
= 0; /* break early from loop */
722 *vm_flags
|= VM_LOCKED
;
726 if (ptep_clear_flush_young_notify(vma
, address
, pte
)) {
728 * Don't treat a reference through a sequentially read
729 * mapping as such. If the page has been used in
730 * another mapping, we will catch it; if this other
731 * mapping is already gone, the unmap path will have
732 * set PG_referenced or activated the page.
734 if (likely(!VM_SequentialReadHint(vma
)))
737 pte_unmap_unlock(pte
, ptl
);
743 *vm_flags
|= vma
->vm_flags
;
748 static int page_referenced_anon(struct page
*page
,
749 struct mem_cgroup
*memcg
,
750 unsigned long *vm_flags
)
752 unsigned int mapcount
;
753 struct anon_vma
*anon_vma
;
755 struct anon_vma_chain
*avc
;
758 anon_vma
= page_lock_anon_vma_read(page
);
762 mapcount
= page_mapcount(page
);
763 pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
764 anon_vma_interval_tree_foreach(avc
, &anon_vma
->rb_root
, pgoff
, pgoff
) {
765 struct vm_area_struct
*vma
= avc
->vma
;
766 unsigned long address
= vma_address(page
, vma
);
768 * If we are reclaiming on behalf of a cgroup, skip
769 * counting on behalf of references from different
772 if (memcg
&& !mm_match_cgroup(vma
->vm_mm
, memcg
))
774 referenced
+= page_referenced_one(page
, vma
, address
,
775 &mapcount
, vm_flags
);
780 page_unlock_anon_vma_read(anon_vma
);
785 * page_referenced_file - referenced check for object-based rmap
786 * @page: the page we're checking references on.
787 * @memcg: target memory control group
788 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
790 * For an object-based mapped page, find all the places it is mapped and
791 * check/clear the referenced flag. This is done by following the page->mapping
792 * pointer, then walking the chain of vmas it holds. It returns the number
793 * of references it found.
795 * This function is only called from page_referenced for object-based pages.
797 static int page_referenced_file(struct page
*page
,
798 struct mem_cgroup
*memcg
,
799 unsigned long *vm_flags
)
801 unsigned int mapcount
;
802 struct address_space
*mapping
= page
->mapping
;
803 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
804 struct vm_area_struct
*vma
;
808 * The caller's checks on page->mapping and !PageAnon have made
809 * sure that this is a file page: the check for page->mapping
810 * excludes the case just before it gets set on an anon page.
812 BUG_ON(PageAnon(page
));
815 * The page lock not only makes sure that page->mapping cannot
816 * suddenly be NULLified by truncation, it makes sure that the
817 * structure at mapping cannot be freed and reused yet,
818 * so we can safely take mapping->i_mmap_mutex.
820 BUG_ON(!PageLocked(page
));
823 if (!mutex_trylock(&mapping
->i_mmap_mutex
))
824 return 1; //put in active list
827 * i_mmap_mutex does not stabilize mapcount at all, but mapcount
828 * is more likely to be accurate if we note it after spinning.
830 mapcount
= page_mapcount(page
);
832 vma_interval_tree_foreach(vma
, &mapping
->i_mmap
, pgoff
, pgoff
) {
833 unsigned long address
= vma_address(page
, vma
);
835 * If we are reclaiming on behalf of a cgroup, skip
836 * counting on behalf of references from different
839 if (memcg
&& !mm_match_cgroup(vma
->vm_mm
, memcg
))
841 referenced
+= page_referenced_one(page
, vma
, address
,
842 &mapcount
, vm_flags
);
847 mutex_unlock(&mapping
->i_mmap_mutex
);
852 * page_referenced - test if the page was referenced
853 * @page: the page to test
854 * @is_locked: caller holds lock on the page
855 * @memcg: target memory cgroup
856 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
858 * Quick test_and_clear_referenced for all mappings to a page,
859 * returns the number of ptes which referenced the page.
861 int page_referenced(struct page
*page
,
863 struct mem_cgroup
*memcg
,
864 unsigned long *vm_flags
)
870 if (page_mapped(page
) && page_rmapping(page
)) {
871 if (!is_locked
&& (!PageAnon(page
) || PageKsm(page
))) {
872 we_locked
= trylock_page(page
);
878 if (unlikely(PageKsm(page
)))
879 referenced
+= page_referenced_ksm(page
, memcg
,
881 else if (PageAnon(page
))
882 referenced
+= page_referenced_anon(page
, memcg
,
884 else if (page
->mapping
)
885 referenced
+= page_referenced_file(page
, memcg
,
890 if (page_test_and_clear_young(page_to_pfn(page
)))
897 static int page_mkclean_one(struct page
*page
, struct vm_area_struct
*vma
,
898 unsigned long address
)
900 struct mm_struct
*mm
= vma
->vm_mm
;
905 pte
= page_check_address(page
, mm
, address
, &ptl
, 1);
909 if (pte_dirty(*pte
) || pte_write(*pte
)) {
912 flush_cache_page(vma
, address
, pte_pfn(*pte
));
913 entry
= ptep_clear_flush(vma
, address
, pte
);
914 entry
= pte_wrprotect(entry
);
915 entry
= pte_mkclean(entry
);
916 set_pte_at(mm
, address
, pte
, entry
);
920 pte_unmap_unlock(pte
, ptl
);
923 mmu_notifier_invalidate_page(mm
, address
);
928 static int page_mkclean_file(struct address_space
*mapping
, struct page
*page
)
930 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
931 struct vm_area_struct
*vma
;
934 BUG_ON(PageAnon(page
));
936 mutex_lock(&mapping
->i_mmap_mutex
);
937 vma_interval_tree_foreach(vma
, &mapping
->i_mmap
, pgoff
, pgoff
) {
938 if (vma
->vm_flags
& VM_SHARED
) {
939 unsigned long address
= vma_address(page
, vma
);
940 ret
+= page_mkclean_one(page
, vma
, address
);
943 mutex_unlock(&mapping
->i_mmap_mutex
);
947 int page_mkclean(struct page
*page
)
951 BUG_ON(!PageLocked(page
));
953 if (page_mapped(page
)) {
954 struct address_space
*mapping
= page_mapping(page
);
956 ret
= page_mkclean_file(mapping
, page
);
961 EXPORT_SYMBOL_GPL(page_mkclean
);
964 * page_move_anon_rmap - move a page to our anon_vma
965 * @page: the page to move to our anon_vma
966 * @vma: the vma the page belongs to
967 * @address: the user virtual address mapped
969 * When a page belongs exclusively to one process after a COW event,
970 * that page can be moved into the anon_vma that belongs to just that
971 * process, so the rmap code will not search the parent or sibling
974 void page_move_anon_rmap(struct page
*page
,
975 struct vm_area_struct
*vma
, unsigned long address
)
977 struct anon_vma
*anon_vma
= vma
->anon_vma
;
979 VM_BUG_ON(!PageLocked(page
));
980 VM_BUG_ON(!anon_vma
);
981 VM_BUG_ON(page
->index
!= linear_page_index(vma
, address
));
983 anon_vma
= (void *) anon_vma
+ PAGE_MAPPING_ANON
;
984 page
->mapping
= (struct address_space
*) anon_vma
;
988 * __page_set_anon_rmap - set up new anonymous rmap
989 * @page: Page to add to rmap
990 * @vma: VM area to add page to.
991 * @address: User virtual address of the mapping
992 * @exclusive: the page is exclusively owned by the current process
994 static void __page_set_anon_rmap(struct page
*page
,
995 struct vm_area_struct
*vma
, unsigned long address
, int exclusive
)
997 struct anon_vma
*anon_vma
= vma
->anon_vma
;
1005 * If the page isn't exclusively mapped into this vma,
1006 * we must use the _oldest_ possible anon_vma for the
1010 anon_vma
= anon_vma
->root
;
1012 anon_vma
= (void *) anon_vma
+ PAGE_MAPPING_ANON
;
1013 page
->mapping
= (struct address_space
*) anon_vma
;
1014 page
->index
= linear_page_index(vma
, address
);
1018 * __page_check_anon_rmap - sanity check anonymous rmap addition
1019 * @page: the page to add the mapping to
1020 * @vma: the vm area in which the mapping is added
1021 * @address: the user virtual address mapped
1023 static void __page_check_anon_rmap(struct page
*page
,
1024 struct vm_area_struct
*vma
, unsigned long address
)
1026 #ifdef CONFIG_DEBUG_VM
1028 * The page's anon-rmap details (mapping and index) are guaranteed to
1029 * be set up correctly at this point.
1031 * We have exclusion against page_add_anon_rmap because the caller
1032 * always holds the page locked, except if called from page_dup_rmap,
1033 * in which case the page is already known to be setup.
1035 * We have exclusion against page_add_new_anon_rmap because those pages
1036 * are initially only visible via the pagetables, and the pte is locked
1037 * over the call to page_add_new_anon_rmap.
1039 BUG_ON(page_anon_vma(page
)->root
!= vma
->anon_vma
->root
);
1040 BUG_ON(page
->index
!= linear_page_index(vma
, address
));
1045 * page_add_anon_rmap - add pte mapping to an anonymous page
1046 * @page: the page to add the mapping to
1047 * @vma: the vm area in which the mapping is added
1048 * @address: the user virtual address mapped
1050 * The caller needs to hold the pte lock, and the page must be locked in
1051 * the anon_vma case: to serialize mapping,index checking after setting,
1052 * and to ensure that PageAnon is not being upgraded racily to PageKsm
1053 * (but PageKsm is never downgraded to PageAnon).
1055 void page_add_anon_rmap(struct page
*page
,
1056 struct vm_area_struct
*vma
, unsigned long address
)
1058 do_page_add_anon_rmap(page
, vma
, address
, 0);
1062 * Special version of the above for do_swap_page, which often runs
1063 * into pages that are exclusively owned by the current process.
1064 * Everybody else should continue to use page_add_anon_rmap above.
1066 void do_page_add_anon_rmap(struct page
*page
,
1067 struct vm_area_struct
*vma
, unsigned long address
, int exclusive
)
1069 int first
= atomic_inc_and_test(&page
->_mapcount
);
1071 if (!PageTransHuge(page
))
1072 __inc_zone_page_state(page
, NR_ANON_PAGES
);
1074 __inc_zone_page_state(page
,
1075 NR_ANON_TRANSPARENT_HUGEPAGES
);
1077 if (unlikely(PageKsm(page
)))
1080 VM_BUG_ON(!PageLocked(page
));
1081 /* address might be in next vma when migration races vma_adjust */
1083 __page_set_anon_rmap(page
, vma
, address
, exclusive
);
1085 __page_check_anon_rmap(page
, vma
, address
);
1089 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1090 * @page: the page to add the mapping to
1091 * @vma: the vm area in which the mapping is added
1092 * @address: the user virtual address mapped
1094 * Same as page_add_anon_rmap but must only be called on *new* pages.
1095 * This means the inc-and-test can be bypassed.
1096 * Page does not have to be locked.
1098 void page_add_new_anon_rmap(struct page
*page
,
1099 struct vm_area_struct
*vma
, unsigned long address
)
1101 VM_BUG_ON(address
< vma
->vm_start
|| address
>= vma
->vm_end
);
1102 SetPageSwapBacked(page
);
1103 atomic_set(&page
->_mapcount
, 0); /* increment count (starts at -1) */
1104 if (!PageTransHuge(page
))
1105 __inc_zone_page_state(page
, NR_ANON_PAGES
);
1107 __inc_zone_page_state(page
, NR_ANON_TRANSPARENT_HUGEPAGES
);
1108 __page_set_anon_rmap(page
, vma
, address
, 1);
1109 if (!mlocked_vma_newpage(vma
, page
))
1110 lru_cache_add_lru(page
, LRU_ACTIVE_ANON
);
1112 add_page_to_unevictable_list(page
);
1116 * page_add_file_rmap - add pte mapping to a file page
1117 * @page: the page to add the mapping to
1119 * The caller needs to hold the pte lock.
1121 void page_add_file_rmap(struct page
*page
)
1124 unsigned long flags
;
1126 mem_cgroup_begin_update_page_stat(page
, &locked
, &flags
);
1127 if (atomic_inc_and_test(&page
->_mapcount
)) {
1128 __inc_zone_page_state(page
, NR_FILE_MAPPED
);
1129 mem_cgroup_inc_page_stat(page
, MEMCG_NR_FILE_MAPPED
);
1131 mem_cgroup_end_update_page_stat(page
, &locked
, &flags
);
1135 * page_remove_rmap - take down pte mapping from a page
1136 * @page: page to remove mapping from
1138 * The caller needs to hold the pte lock.
1140 void page_remove_rmap(struct page
*page
)
1142 bool anon
= PageAnon(page
);
1144 unsigned long flags
;
1147 * The anon case has no mem_cgroup page_stat to update; but may
1148 * uncharge_page() below, where the lock ordering can deadlock if
1149 * we hold the lock against page_stat move: so avoid it on anon.
1152 mem_cgroup_begin_update_page_stat(page
, &locked
, &flags
);
1154 /* page still mapped by someone else? */
1155 if (!atomic_add_negative(-1, &page
->_mapcount
))
1159 * Hugepages are not counted in NR_ANON_PAGES nor NR_FILE_MAPPED
1160 * and not charged by memcg for now.
1162 if (unlikely(PageHuge(page
)))
1165 mem_cgroup_uncharge_page(page
);
1166 if (!PageTransHuge(page
))
1167 __dec_zone_page_state(page
, NR_ANON_PAGES
);
1169 __dec_zone_page_state(page
,
1170 NR_ANON_TRANSPARENT_HUGEPAGES
);
1172 __dec_zone_page_state(page
, NR_FILE_MAPPED
);
1173 mem_cgroup_dec_page_stat(page
, MEMCG_NR_FILE_MAPPED
);
1174 mem_cgroup_end_update_page_stat(page
, &locked
, &flags
);
1176 if (unlikely(PageMlocked(page
)))
1177 clear_page_mlock(page
);
1179 * It would be tidy to reset the PageAnon mapping here,
1180 * but that might overwrite a racing page_add_anon_rmap
1181 * which increments mapcount after us but sets mapping
1182 * before us: so leave the reset to free_hot_cold_page,
1183 * and remember that it's only reliable while mapped.
1184 * Leaving it set also helps swapoff to reinstate ptes
1185 * faster for those pages still in swapcache.
1190 mem_cgroup_end_update_page_stat(page
, &locked
, &flags
);
1194 * Subfunctions of try_to_unmap: try_to_unmap_one called
1195 * repeatedly from try_to_unmap_ksm, try_to_unmap_anon or try_to_unmap_file.
1197 int try_to_unmap_one(struct page
*page
, struct vm_area_struct
*vma
,
1198 unsigned long address
, enum ttu_flags flags
)
1200 struct mm_struct
*mm
= vma
->vm_mm
;
1204 int ret
= SWAP_AGAIN
;
1206 pte
= page_check_address(page
, mm
, address
, &ptl
, 0);
1211 * If the page is mlock()d, we cannot swap it out.
1212 * If it's recently referenced (perhaps page_referenced
1213 * skipped over this mm) then we should reactivate it.
1215 if (!(flags
& TTU_IGNORE_MLOCK
)) {
1216 if (vma
->vm_flags
& VM_LOCKED
)
1219 if (TTU_ACTION(flags
) == TTU_MUNLOCK
)
1222 if (!(flags
& TTU_IGNORE_ACCESS
)) {
1223 if (ptep_clear_flush_young_notify(vma
, address
, pte
)) {
1229 /* Nuke the page table entry. */
1230 flush_cache_page(vma
, address
, page_to_pfn(page
));
1231 pteval
= ptep_clear_flush(vma
, address
, pte
);
1233 /* Move the dirty bit to the physical page now the pte is gone. */
1234 if (pte_dirty(pteval
))
1235 set_page_dirty(page
);
1237 /* Update high watermark before we lower rss */
1238 update_hiwater_rss(mm
);
1240 if (PageHWPoison(page
) && !(flags
& TTU_IGNORE_HWPOISON
)) {
1241 if (!PageHuge(page
)) {
1243 dec_mm_counter(mm
, MM_ANONPAGES
);
1245 dec_mm_counter(mm
, MM_FILEPAGES
);
1247 set_pte_at(mm
, address
, pte
,
1248 swp_entry_to_pte(make_hwpoison_entry(page
)));
1249 } else if (PageAnon(page
)) {
1250 swp_entry_t entry
= { .val
= page_private(page
) };
1252 if (PageSwapCache(page
)) {
1254 * Store the swap location in the pte.
1255 * See handle_pte_fault() ...
1257 if (swap_duplicate(entry
) < 0) {
1258 set_pte_at(mm
, address
, pte
, pteval
);
1262 if (list_empty(&mm
->mmlist
)) {
1263 spin_lock(&mmlist_lock
);
1264 if (list_empty(&mm
->mmlist
))
1265 list_add(&mm
->mmlist
, &init_mm
.mmlist
);
1266 spin_unlock(&mmlist_lock
);
1268 dec_mm_counter(mm
, MM_ANONPAGES
);
1269 inc_mm_counter(mm
, MM_SWAPENTS
);
1270 } else if (IS_ENABLED(CONFIG_MIGRATION
)) {
1272 * Store the pfn of the page in a special migration
1273 * pte. do_swap_page() will wait until the migration
1274 * pte is removed and then restart fault handling.
1276 BUG_ON(TTU_ACTION(flags
) != TTU_MIGRATION
);
1277 entry
= make_migration_entry(page
, pte_write(pteval
));
1279 set_pte_at(mm
, address
, pte
, swp_entry_to_pte(entry
));
1280 BUG_ON(pte_file(*pte
));
1281 } else if (IS_ENABLED(CONFIG_MIGRATION
) &&
1282 (TTU_ACTION(flags
) == TTU_MIGRATION
)) {
1283 /* Establish migration entry for a file page */
1285 entry
= make_migration_entry(page
, pte_write(pteval
));
1286 set_pte_at(mm
, address
, pte
, swp_entry_to_pte(entry
));
1288 dec_mm_counter(mm
, MM_FILEPAGES
);
1290 page_remove_rmap(page
);
1291 page_cache_release(page
);
1294 pte_unmap_unlock(pte
, ptl
);
1295 if (ret
!= SWAP_FAIL
)
1296 mmu_notifier_invalidate_page(mm
, address
);
1301 pte_unmap_unlock(pte
, ptl
);
1305 * We need mmap_sem locking, Otherwise VM_LOCKED check makes
1306 * unstable result and race. Plus, We can't wait here because
1307 * we now hold anon_vma->rwsem or mapping->i_mmap_mutex.
1308 * if trylock failed, the page remain in evictable lru and later
1309 * vmscan could retry to move the page to unevictable lru if the
1310 * page is actually mlocked.
1312 if (down_read_trylock(&vma
->vm_mm
->mmap_sem
)) {
1313 if (vma
->vm_flags
& VM_LOCKED
) {
1314 mlock_vma_page(page
);
1317 up_read(&vma
->vm_mm
->mmap_sem
);
1323 * objrmap doesn't work for nonlinear VMAs because the assumption that
1324 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
1325 * Consequently, given a particular page and its ->index, we cannot locate the
1326 * ptes which are mapping that page without an exhaustive linear search.
1328 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
1329 * maps the file to which the target page belongs. The ->vm_private_data field
1330 * holds the current cursor into that scan. Successive searches will circulate
1331 * around the vma's virtual address space.
1333 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
1334 * more scanning pressure is placed against them as well. Eventually pages
1335 * will become fully unmapped and are eligible for eviction.
1337 * For very sparsely populated VMAs this is a little inefficient - chances are
1338 * there there won't be many ptes located within the scan cluster. In this case
1339 * maybe we could scan further - to the end of the pte page, perhaps.
1341 * Mlocked pages: check VM_LOCKED under mmap_sem held for read, if we can
1342 * acquire it without blocking. If vma locked, mlock the pages in the cluster,
1343 * rather than unmapping them. If we encounter the "check_page" that vmscan is
1344 * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN.
1346 #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
1347 #define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
1349 static int try_to_unmap_cluster(unsigned long cursor
, unsigned int *mapcount
,
1350 struct vm_area_struct
*vma
, struct page
*check_page
)
1352 struct mm_struct
*mm
= vma
->vm_mm
;
1358 unsigned long address
;
1359 unsigned long mmun_start
; /* For mmu_notifiers */
1360 unsigned long mmun_end
; /* For mmu_notifiers */
1362 int ret
= SWAP_AGAIN
;
1365 address
= (vma
->vm_start
+ cursor
) & CLUSTER_MASK
;
1366 end
= address
+ CLUSTER_SIZE
;
1367 if (address
< vma
->vm_start
)
1368 address
= vma
->vm_start
;
1369 if (end
> vma
->vm_end
)
1372 pmd
= mm_find_pmd(mm
, address
);
1376 mmun_start
= address
;
1378 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
1381 * If we can acquire the mmap_sem for read, and vma is VM_LOCKED,
1382 * keep the sem while scanning the cluster for mlocking pages.
1384 if (down_read_trylock(&vma
->vm_mm
->mmap_sem
)) {
1385 locked_vma
= (vma
->vm_flags
& VM_LOCKED
);
1387 up_read(&vma
->vm_mm
->mmap_sem
); /* don't need it */
1390 pte
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
1392 /* Update high watermark before we lower rss */
1393 update_hiwater_rss(mm
);
1395 for (; address
< end
; pte
++, address
+= PAGE_SIZE
) {
1396 if (!pte_present(*pte
))
1398 page
= vm_normal_page(vma
, address
, *pte
);
1399 BUG_ON(!page
|| PageAnon(page
));
1402 if (page
== check_page
) {
1403 /* we know we have check_page locked */
1404 mlock_vma_page(page
);
1406 } else if (trylock_page(page
)) {
1408 * If we can lock the page, perform mlock.
1409 * Otherwise leave the page alone, it will be
1410 * eventually encountered again later.
1412 mlock_vma_page(page
);
1415 continue; /* don't unmap */
1418 if (ptep_clear_flush_young_notify(vma
, address
, pte
))
1421 /* Nuke the page table entry. */
1422 flush_cache_page(vma
, address
, pte_pfn(*pte
));
1423 pteval
= ptep_clear_flush(vma
, address
, pte
);
1425 /* If nonlinear, store the file page offset in the pte. */
1426 if (page
->index
!= linear_page_index(vma
, address
))
1427 set_pte_at(mm
, address
, pte
, pgoff_to_pte(page
->index
));
1429 /* Move the dirty bit to the physical page now the pte is gone. */
1430 if (pte_dirty(pteval
))
1431 set_page_dirty(page
);
1433 page_remove_rmap(page
);
1434 page_cache_release(page
);
1435 dec_mm_counter(mm
, MM_FILEPAGES
);
1438 pte_unmap_unlock(pte
- 1, ptl
);
1439 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
1441 up_read(&vma
->vm_mm
->mmap_sem
);
1445 bool is_vma_temporary_stack(struct vm_area_struct
*vma
)
1447 int maybe_stack
= vma
->vm_flags
& (VM_GROWSDOWN
| VM_GROWSUP
);
1452 if ((vma
->vm_flags
& VM_STACK_INCOMPLETE_SETUP
) ==
1453 VM_STACK_INCOMPLETE_SETUP
)
1460 * try_to_unmap_anon - unmap or unlock anonymous page using the object-based
1462 * @page: the page to unmap/unlock
1463 * @flags: action and flags
1465 * Find all the mappings of a page using the mapping pointer and the vma chains
1466 * contained in the anon_vma struct it points to.
1468 * This function is only called from try_to_unmap/try_to_munlock for
1470 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1471 * where the page was found will be held for write. So, we won't recheck
1472 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1475 static int try_to_unmap_anon(struct page
*page
, enum ttu_flags flags
)
1477 struct anon_vma
*anon_vma
;
1479 struct anon_vma_chain
*avc
;
1480 int ret
= SWAP_AGAIN
;
1482 anon_vma
= page_lock_anon_vma_read(page
);
1486 pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
1487 anon_vma_interval_tree_foreach(avc
, &anon_vma
->rb_root
, pgoff
, pgoff
) {
1488 struct vm_area_struct
*vma
= avc
->vma
;
1489 unsigned long address
;
1492 * During exec, a temporary VMA is setup and later moved.
1493 * The VMA is moved under the anon_vma lock but not the
1494 * page tables leading to a race where migration cannot
1495 * find the migration ptes. Rather than increasing the
1496 * locking requirements of exec(), migration skips
1497 * temporary VMAs until after exec() completes.
1499 if (IS_ENABLED(CONFIG_MIGRATION
) && (flags
& TTU_MIGRATION
) &&
1500 is_vma_temporary_stack(vma
))
1503 address
= vma_address(page
, vma
);
1504 ret
= try_to_unmap_one(page
, vma
, address
, flags
);
1505 if (ret
!= SWAP_AGAIN
|| !page_mapped(page
))
1509 page_unlock_anon_vma_read(anon_vma
);
1514 * try_to_unmap_file - unmap/unlock file page using the object-based rmap method
1515 * @page: the page to unmap/unlock
1516 * @flags: action and flags
1518 * Find all the mappings of a page using the mapping pointer and the vma chains
1519 * contained in the address_space struct it points to.
1521 * This function is only called from try_to_unmap/try_to_munlock for
1522 * object-based pages.
1523 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1524 * where the page was found will be held for write. So, we won't recheck
1525 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1528 static int try_to_unmap_file(struct page
*page
, enum ttu_flags flags
)
1530 struct address_space
*mapping
= page
->mapping
;
1531 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
1532 struct vm_area_struct
*vma
;
1533 int ret
= SWAP_AGAIN
;
1534 unsigned long cursor
;
1535 unsigned long max_nl_cursor
= 0;
1536 unsigned long max_nl_size
= 0;
1537 unsigned int mapcount
;
1542 pgoff
= page
->index
<< compound_order(page
);
1544 while (!mutex_trylock(&mapping
->i_mmap_mutex
)) {
1546 if (!(retry
% MUTEX_RETRY_RESCHED
))
1548 if (retry
> MUTEX_RETRY_COUNT
) {
1549 printk(KERN_ERR
">> failed to lock i_mmap_mutex in try_to_unmap_file <<\n");
1555 vma_interval_tree_foreach(vma
, &mapping
->i_mmap
, pgoff
, pgoff
) {
1556 unsigned long address
= vma_address(page
, vma
);
1557 ret
= try_to_unmap_one(page
, vma
, address
, flags
);
1558 if (ret
!= SWAP_AGAIN
|| !page_mapped(page
))
1562 if (list_empty(&mapping
->i_mmap_nonlinear
))
1566 * We don't bother to try to find the munlocked page in nonlinears.
1567 * It's costly. Instead, later, page reclaim logic may call
1568 * try_to_unmap(TTU_MUNLOCK) and recover PG_mlocked lazily.
1570 if (TTU_ACTION(flags
) == TTU_MUNLOCK
)
1573 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
,
1575 cursor
= (unsigned long) vma
->vm_private_data
;
1576 if (cursor
> max_nl_cursor
)
1577 max_nl_cursor
= cursor
;
1578 cursor
= vma
->vm_end
- vma
->vm_start
;
1579 if (cursor
> max_nl_size
)
1580 max_nl_size
= cursor
;
1583 if (max_nl_size
== 0) { /* all nonlinears locked or reserved ? */
1589 * We don't try to search for this page in the nonlinear vmas,
1590 * and page_referenced wouldn't have found it anyway. Instead
1591 * just walk the nonlinear vmas trying to age and unmap some.
1592 * The mapcount of the page we came in with is irrelevant,
1593 * but even so use it as a guide to how hard we should try?
1595 mapcount
= page_mapcount(page
);
1600 max_nl_size
= (max_nl_size
+ CLUSTER_SIZE
- 1) & CLUSTER_MASK
;
1601 if (max_nl_cursor
== 0)
1602 max_nl_cursor
= CLUSTER_SIZE
;
1605 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
,
1607 cursor
= (unsigned long) vma
->vm_private_data
;
1608 while ( cursor
< max_nl_cursor
&&
1609 cursor
< vma
->vm_end
- vma
->vm_start
) {
1610 if (try_to_unmap_cluster(cursor
, &mapcount
,
1611 vma
, page
) == SWAP_MLOCK
)
1613 cursor
+= CLUSTER_SIZE
;
1614 vma
->vm_private_data
= (void *) cursor
;
1615 if ((int)mapcount
<= 0)
1618 vma
->vm_private_data
= (void *) max_nl_cursor
;
1621 max_nl_cursor
+= CLUSTER_SIZE
;
1622 } while (max_nl_cursor
<= max_nl_size
);
1625 * Don't loop forever (perhaps all the remaining pages are
1626 * in locked vmas). Reset cursor on all unreserved nonlinear
1627 * vmas, now forgetting on which ones it had fallen behind.
1629 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
, shared
.nonlinear
)
1630 vma
->vm_private_data
= NULL
;
1632 mutex_unlock(&mapping
->i_mmap_mutex
);
1637 * try_to_unmap - try to remove all page table mappings to a page
1638 * @page: the page to get unmapped
1639 * @flags: action and flags
1641 * Tries to remove all the page table entries which are mapping this
1642 * page, used in the pageout path. Caller must hold the page lock.
1643 * Return values are:
1645 * SWAP_SUCCESS - we succeeded in removing all mappings
1646 * SWAP_AGAIN - we missed a mapping, try again later
1647 * SWAP_FAIL - the page is unswappable
1648 * SWAP_MLOCK - page is mlocked.
1650 int try_to_unmap(struct page
*page
, enum ttu_flags flags
)
1654 BUG_ON(!PageLocked(page
));
1655 VM_BUG_ON(!PageHuge(page
) && PageTransHuge(page
));
1657 if (unlikely(PageKsm(page
)))
1658 ret
= try_to_unmap_ksm(page
, flags
);
1659 else if (PageAnon(page
))
1660 ret
= try_to_unmap_anon(page
, flags
);
1662 ret
= try_to_unmap_file(page
, flags
);
1663 if (ret
!= SWAP_MLOCK
&& !page_mapped(page
))
1669 * try_to_munlock - try to munlock a page
1670 * @page: the page to be munlocked
1672 * Called from munlock code. Checks all of the VMAs mapping the page
1673 * to make sure nobody else has this page mlocked. The page will be
1674 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1676 * Return values are:
1678 * SWAP_AGAIN - no vma is holding page mlocked, or,
1679 * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem
1680 * SWAP_FAIL - page cannot be located at present
1681 * SWAP_MLOCK - page is now mlocked.
1683 int try_to_munlock(struct page
*page
)
1685 VM_BUG_ON(!PageLocked(page
) || PageLRU(page
));
1687 if (unlikely(PageKsm(page
)))
1688 return try_to_unmap_ksm(page
, TTU_MUNLOCK
);
1689 else if (PageAnon(page
))
1690 return try_to_unmap_anon(page
, TTU_MUNLOCK
);
1692 return try_to_unmap_file(page
, TTU_MUNLOCK
);
1695 void __put_anon_vma(struct anon_vma
*anon_vma
)
1697 struct anon_vma
*root
= anon_vma
->root
;
1699 anon_vma_free(anon_vma
);
1700 if (root
&& root
!= anon_vma
&& atomic_dec_and_test(&root
->refcount
))
1701 anon_vma_free(root
);
1704 #ifdef CONFIG_MIGRATION
1706 * rmap_walk() and its helpers rmap_walk_anon() and rmap_walk_file():
1707 * Called by migrate.c to remove migration ptes, but might be used more later.
1709 static int rmap_walk_anon(struct page
*page
, int (*rmap_one
)(struct page
*,
1710 struct vm_area_struct
*, unsigned long, void *), void *arg
)
1712 struct anon_vma
*anon_vma
;
1713 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
1714 struct anon_vma_chain
*avc
;
1715 int ret
= SWAP_AGAIN
;
1718 * Note: remove_migration_ptes() cannot use page_lock_anon_vma_read()
1719 * because that depends on page_mapped(); but not all its usages
1720 * are holding mmap_sem. Users without mmap_sem are required to
1721 * take a reference count to prevent the anon_vma disappearing
1723 anon_vma
= page_anon_vma(page
);
1726 anon_vma_lock_read(anon_vma
);
1727 anon_vma_interval_tree_foreach(avc
, &anon_vma
->rb_root
, pgoff
, pgoff
) {
1728 struct vm_area_struct
*vma
= avc
->vma
;
1729 unsigned long address
= vma_address(page
, vma
);
1730 ret
= rmap_one(page
, vma
, address
, arg
);
1731 if (ret
!= SWAP_AGAIN
)
1734 anon_vma_unlock_read(anon_vma
);
1738 static int rmap_walk_file(struct page
*page
, int (*rmap_one
)(struct page
*,
1739 struct vm_area_struct
*, unsigned long, void *), void *arg
)
1741 struct address_space
*mapping
= page
->mapping
;
1742 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
1743 struct vm_area_struct
*vma
;
1744 int ret
= SWAP_AGAIN
;
1748 mutex_lock(&mapping
->i_mmap_mutex
);
1749 vma_interval_tree_foreach(vma
, &mapping
->i_mmap
, pgoff
, pgoff
) {
1750 unsigned long address
= vma_address(page
, vma
);
1751 ret
= rmap_one(page
, vma
, address
, arg
);
1752 if (ret
!= SWAP_AGAIN
)
1756 * No nonlinear handling: being always shared, nonlinear vmas
1757 * never contain migration ptes. Decide what to do about this
1758 * limitation to linear when we need rmap_walk() on nonlinear.
1760 mutex_unlock(&mapping
->i_mmap_mutex
);
1764 int rmap_walk(struct page
*page
, int (*rmap_one
)(struct page
*,
1765 struct vm_area_struct
*, unsigned long, void *), void *arg
)
1767 VM_BUG_ON(!PageLocked(page
));
1769 if (unlikely(PageKsm(page
)))
1770 return rmap_walk_ksm(page
, rmap_one
, arg
);
1771 else if (PageAnon(page
))
1772 return rmap_walk_anon(page
, rmap_one
, arg
);
1774 return rmap_walk_file(page
, rmap_one
, arg
);
1776 #endif /* CONFIG_MIGRATION */
1778 #ifdef CONFIG_HUGETLB_PAGE
1780 * The following three functions are for anonymous (private mapped) hugepages.
1781 * Unlike common anonymous pages, anonymous hugepages have no accounting code
1782 * and no lru code, because we handle hugepages differently from common pages.
1784 static void __hugepage_set_anon_rmap(struct page
*page
,
1785 struct vm_area_struct
*vma
, unsigned long address
, int exclusive
)
1787 struct anon_vma
*anon_vma
= vma
->anon_vma
;
1794 anon_vma
= anon_vma
->root
;
1796 anon_vma
= (void *) anon_vma
+ PAGE_MAPPING_ANON
;
1797 page
->mapping
= (struct address_space
*) anon_vma
;
1798 page
->index
= linear_page_index(vma
, address
);
1801 void hugepage_add_anon_rmap(struct page
*page
,
1802 struct vm_area_struct
*vma
, unsigned long address
)
1804 struct anon_vma
*anon_vma
= vma
->anon_vma
;
1807 BUG_ON(!PageLocked(page
));
1809 /* address might be in next vma when migration races vma_adjust */
1810 first
= atomic_inc_and_test(&page
->_mapcount
);
1812 __hugepage_set_anon_rmap(page
, vma
, address
, 0);
1815 void hugepage_add_new_anon_rmap(struct page
*page
,
1816 struct vm_area_struct
*vma
, unsigned long address
)
1818 BUG_ON(address
< vma
->vm_start
|| address
>= vma
->vm_end
);
1819 atomic_set(&page
->_mapcount
, 0);
1820 __hugepage_set_anon_rmap(page
, vma
, address
, 1);
1822 #endif /* CONFIG_HUGETLB_PAGE */