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 * Some rmap walk that needs to find all ptes/hugepmds without false
273 * negatives (like migrate and split_huge_page) running concurrent
274 * with operations that copy or move pagetables (like mremap() and
275 * fork()) to be safe. They depend on the anon_vma "same_anon_vma"
276 * list to be in a certain order: the dst_vma must be placed after the
277 * src_vma in the list. This is always guaranteed by fork() but
278 * mremap() needs to call this function to enforce it in case the
279 * dst_vma isn't newly allocated and chained with the anon_vma_clone()
280 * function but just an extension of a pre-existing vma through
283 * NOTE: the same_anon_vma list can still be changed by other
284 * processes while mremap runs because mremap doesn't hold the
285 * anon_vma mutex to prevent modifications to the list while it
286 * runs. All we need to enforce is that the relative order of this
287 * process vmas isn't changing (we don't care about other vmas
288 * order). Each vma corresponds to an anon_vma_chain structure so
289 * there's no risk that other processes calling anon_vma_moveto_tail()
290 * and changing the same_anon_vma list under mremap() will screw with
291 * the relative order of this process vmas in the list, because we
292 * they can't alter the order of any vma that belongs to this
293 * process. And there can't be another anon_vma_moveto_tail() running
294 * concurrently with mremap() coming from this process because we hold
295 * the mmap_sem for the whole mremap(). fork() ordering dependency
296 * also shouldn't be affected because fork() only cares that the
297 * parent vmas are placed in the list before the child vmas and
298 * anon_vma_moveto_tail() won't reorder vmas from either the fork()
301 void anon_vma_moveto_tail(struct vm_area_struct
*dst
)
303 struct anon_vma_chain
*pavc
;
304 struct anon_vma
*root
= NULL
;
306 list_for_each_entry_reverse(pavc
, &dst
->anon_vma_chain
, same_vma
) {
307 struct anon_vma
*anon_vma
= pavc
->anon_vma
;
308 VM_BUG_ON(pavc
->vma
!= dst
);
309 root
= lock_anon_vma_root(root
, anon_vma
);
310 list_del(&pavc
->same_anon_vma
);
311 list_add_tail(&pavc
->same_anon_vma
, &anon_vma
->head
);
313 unlock_anon_vma_root(root
);
317 * Attach vma to its own anon_vma, as well as to the anon_vmas that
318 * the corresponding VMA in the parent process is attached to.
319 * Returns 0 on success, non-zero on failure.
321 int anon_vma_fork(struct vm_area_struct
*vma
, struct vm_area_struct
*pvma
)
323 struct anon_vma_chain
*avc
;
324 struct anon_vma
*anon_vma
;
326 /* Don't bother if the parent process has no anon_vma here. */
331 * First, attach the new VMA to the parent VMA's anon_vmas,
332 * so rmap can find non-COWed pages in child processes.
334 if (anon_vma_clone(vma
, pvma
))
337 /* Then add our own anon_vma. */
338 anon_vma
= anon_vma_alloc();
341 avc
= anon_vma_chain_alloc(GFP_KERNEL
);
343 goto out_error_free_anon_vma
;
346 * The root anon_vma's spinlock is the lock actually used when we
347 * lock any of the anon_vmas in this anon_vma tree.
349 anon_vma
->root
= pvma
->anon_vma
->root
;
351 * With refcounts, an anon_vma can stay around longer than the
352 * process it belongs to. The root anon_vma needs to be pinned until
353 * this anon_vma is freed, because the lock lives in the root.
355 get_anon_vma(anon_vma
->root
);
356 /* Mark this anon_vma as the one where our new (COWed) pages go. */
357 vma
->anon_vma
= anon_vma
;
358 anon_vma_lock(anon_vma
);
359 anon_vma_chain_link(vma
, avc
, anon_vma
);
360 anon_vma_unlock(anon_vma
);
364 out_error_free_anon_vma
:
365 put_anon_vma(anon_vma
);
367 unlink_anon_vmas(vma
);
371 void unlink_anon_vmas(struct vm_area_struct
*vma
)
373 struct anon_vma_chain
*avc
, *next
;
374 struct anon_vma
*root
= NULL
;
377 * Unlink each anon_vma chained to the VMA. This list is ordered
378 * from newest to oldest, ensuring the root anon_vma gets freed last.
380 list_for_each_entry_safe(avc
, next
, &vma
->anon_vma_chain
, same_vma
) {
381 struct anon_vma
*anon_vma
= avc
->anon_vma
;
383 root
= lock_anon_vma_root(root
, anon_vma
);
384 list_del(&avc
->same_anon_vma
);
387 * Leave empty anon_vmas on the list - we'll need
388 * to free them outside the lock.
390 if (list_empty(&anon_vma
->head
))
393 list_del(&avc
->same_vma
);
394 anon_vma_chain_free(avc
);
396 unlock_anon_vma_root(root
);
399 * Iterate the list once more, it now only contains empty and unlinked
400 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
401 * needing to acquire the anon_vma->root->mutex.
403 list_for_each_entry_safe(avc
, next
, &vma
->anon_vma_chain
, same_vma
) {
404 struct anon_vma
*anon_vma
= avc
->anon_vma
;
406 put_anon_vma(anon_vma
);
408 list_del(&avc
->same_vma
);
409 anon_vma_chain_free(avc
);
413 static void anon_vma_ctor(void *data
)
415 struct anon_vma
*anon_vma
= data
;
417 mutex_init(&anon_vma
->mutex
);
418 atomic_set(&anon_vma
->refcount
, 0);
419 INIT_LIST_HEAD(&anon_vma
->head
);
422 void __init
anon_vma_init(void)
424 anon_vma_cachep
= kmem_cache_create("anon_vma", sizeof(struct anon_vma
),
425 0, SLAB_DESTROY_BY_RCU
|SLAB_PANIC
, anon_vma_ctor
);
426 anon_vma_chain_cachep
= KMEM_CACHE(anon_vma_chain
, SLAB_PANIC
);
430 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
432 * Since there is no serialization what so ever against page_remove_rmap()
433 * the best this function can do is return a locked anon_vma that might
434 * have been relevant to this page.
436 * The page might have been remapped to a different anon_vma or the anon_vma
437 * returned may already be freed (and even reused).
439 * In case it was remapped to a different anon_vma, the new anon_vma will be a
440 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
441 * ensure that any anon_vma obtained from the page will still be valid for as
442 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
444 * All users of this function must be very careful when walking the anon_vma
445 * chain and verify that the page in question is indeed mapped in it
446 * [ something equivalent to page_mapped_in_vma() ].
448 * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap()
449 * that the anon_vma pointer from page->mapping is valid if there is a
450 * mapcount, we can dereference the anon_vma after observing those.
452 struct anon_vma
*page_get_anon_vma(struct page
*page
)
454 struct anon_vma
*anon_vma
= NULL
;
455 unsigned long anon_mapping
;
458 anon_mapping
= (unsigned long) ACCESS_ONCE(page
->mapping
);
459 if ((anon_mapping
& PAGE_MAPPING_FLAGS
) != PAGE_MAPPING_ANON
)
461 if (!page_mapped(page
))
464 anon_vma
= (struct anon_vma
*) (anon_mapping
- PAGE_MAPPING_ANON
);
465 if (!atomic_inc_not_zero(&anon_vma
->refcount
)) {
471 * If this page is still mapped, then its anon_vma cannot have been
472 * freed. But if it has been unmapped, we have no security against the
473 * anon_vma structure being freed and reused (for another anon_vma:
474 * SLAB_DESTROY_BY_RCU guarantees that - so the atomic_inc_not_zero()
475 * above cannot corrupt).
477 if (!page_mapped(page
)) {
478 put_anon_vma(anon_vma
);
488 * Similar to page_get_anon_vma() except it locks the anon_vma.
490 * Its a little more complex as it tries to keep the fast path to a single
491 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
492 * reference like with page_get_anon_vma() and then block on the mutex.
494 struct anon_vma
*page_lock_anon_vma(struct page
*page
)
496 struct anon_vma
*anon_vma
= NULL
;
497 struct anon_vma
*root_anon_vma
;
498 unsigned long anon_mapping
;
501 anon_mapping
= (unsigned long) ACCESS_ONCE(page
->mapping
);
502 if ((anon_mapping
& PAGE_MAPPING_FLAGS
) != PAGE_MAPPING_ANON
)
504 if (!page_mapped(page
))
507 anon_vma
= (struct anon_vma
*) (anon_mapping
- PAGE_MAPPING_ANON
);
508 root_anon_vma
= ACCESS_ONCE(anon_vma
->root
);
509 if (mutex_trylock(&root_anon_vma
->mutex
)) {
511 * If the page is still mapped, then this anon_vma is still
512 * its anon_vma, and holding the mutex ensures that it will
513 * not go away, see anon_vma_free().
515 if (!page_mapped(page
)) {
516 mutex_unlock(&root_anon_vma
->mutex
);
522 /* trylock failed, we got to sleep */
523 if (!atomic_inc_not_zero(&anon_vma
->refcount
)) {
528 if (!page_mapped(page
)) {
529 put_anon_vma(anon_vma
);
534 /* we pinned the anon_vma, its safe to sleep */
536 anon_vma_lock(anon_vma
);
538 if (atomic_dec_and_test(&anon_vma
->refcount
)) {
540 * Oops, we held the last refcount, release the lock
541 * and bail -- can't simply use put_anon_vma() because
542 * we'll deadlock on the anon_vma_lock() recursion.
544 anon_vma_unlock(anon_vma
);
545 __put_anon_vma(anon_vma
);
556 void page_unlock_anon_vma(struct anon_vma
*anon_vma
)
558 anon_vma_unlock(anon_vma
);
562 * At what user virtual address is page expected in @vma?
563 * Returns virtual address or -EFAULT if page's index/offset is not
564 * within the range mapped the @vma.
567 vma_address(struct page
*page
, struct vm_area_struct
*vma
)
569 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
570 unsigned long address
;
572 if (unlikely(is_vm_hugetlb_page(vma
)))
573 pgoff
= page
->index
<< huge_page_order(page_hstate(page
));
574 address
= vma
->vm_start
+ ((pgoff
- vma
->vm_pgoff
) << PAGE_SHIFT
);
575 if (unlikely(address
< vma
->vm_start
|| address
>= vma
->vm_end
)) {
576 /* page should be within @vma mapping range */
583 * At what user virtual address is page expected in vma?
584 * Caller should check the page is actually part of the vma.
586 unsigned long page_address_in_vma(struct page
*page
, struct vm_area_struct
*vma
)
588 if (PageAnon(page
)) {
589 struct anon_vma
*page__anon_vma
= page_anon_vma(page
);
591 * Note: swapoff's unuse_vma() is more efficient with this
592 * check, and needs it to match anon_vma when KSM is active.
594 if (!vma
->anon_vma
|| !page__anon_vma
||
595 vma
->anon_vma
->root
!= page__anon_vma
->root
)
597 } else if (page
->mapping
&& !(vma
->vm_flags
& VM_NONLINEAR
)) {
599 vma
->vm_file
->f_mapping
!= page
->mapping
)
603 return vma_address(page
, vma
);
607 * Check that @page is mapped at @address into @mm.
609 * If @sync is false, page_check_address may perform a racy check to avoid
610 * the page table lock when the pte is not present (helpful when reclaiming
611 * highly shared pages).
613 * On success returns with pte mapped and locked.
615 pte_t
*__page_check_address(struct page
*page
, struct mm_struct
*mm
,
616 unsigned long address
, spinlock_t
**ptlp
, int sync
)
624 if (unlikely(PageHuge(page
))) {
625 pte
= huge_pte_offset(mm
, address
);
626 ptl
= &mm
->page_table_lock
;
630 pgd
= pgd_offset(mm
, address
);
631 if (!pgd_present(*pgd
))
634 pud
= pud_offset(pgd
, address
);
635 if (!pud_present(*pud
))
638 pmd
= pmd_offset(pud
, address
);
639 if (!pmd_present(*pmd
))
641 if (pmd_trans_huge(*pmd
))
644 pte
= pte_offset_map(pmd
, address
);
645 /* Make a quick check before getting the lock */
646 if (!sync
&& !pte_present(*pte
)) {
651 ptl
= pte_lockptr(mm
, pmd
);
654 if (pte_present(*pte
) && page_to_pfn(page
) == pte_pfn(*pte
)) {
658 pte_unmap_unlock(pte
, ptl
);
663 * page_mapped_in_vma - check whether a page is really mapped in a VMA
664 * @page: the page to test
665 * @vma: the VMA to test
667 * Returns 1 if the page is mapped into the page tables of the VMA, 0
668 * if the page is not mapped into the page tables of this VMA. Only
669 * valid for normal file or anonymous VMAs.
671 int page_mapped_in_vma(struct page
*page
, struct vm_area_struct
*vma
)
673 unsigned long address
;
677 address
= vma_address(page
, vma
);
678 if (address
== -EFAULT
) /* out of vma range */
680 pte
= page_check_address(page
, vma
->vm_mm
, address
, &ptl
, 1);
681 if (!pte
) /* the page is not in this mm */
683 pte_unmap_unlock(pte
, ptl
);
689 * Subfunctions of page_referenced: page_referenced_one called
690 * repeatedly from either page_referenced_anon or page_referenced_file.
692 int page_referenced_one(struct page
*page
, struct vm_area_struct
*vma
,
693 unsigned long address
, unsigned int *mapcount
,
694 unsigned long *vm_flags
)
696 struct mm_struct
*mm
= vma
->vm_mm
;
699 if (unlikely(PageTransHuge(page
))) {
702 spin_lock(&mm
->page_table_lock
);
704 * rmap might return false positives; we must filter
705 * these out using page_check_address_pmd().
707 pmd
= page_check_address_pmd(page
, mm
, address
,
708 PAGE_CHECK_ADDRESS_PMD_FLAG
);
710 spin_unlock(&mm
->page_table_lock
);
714 if (vma
->vm_flags
& VM_LOCKED
) {
715 spin_unlock(&mm
->page_table_lock
);
716 *mapcount
= 0; /* break early from loop */
717 *vm_flags
|= VM_LOCKED
;
721 /* go ahead even if the pmd is pmd_trans_splitting() */
722 if (pmdp_clear_flush_young_notify(vma
, address
, pmd
))
724 spin_unlock(&mm
->page_table_lock
);
730 * rmap might return false positives; we must filter
731 * these out using page_check_address().
733 pte
= page_check_address(page
, mm
, address
, &ptl
, 0);
737 if (vma
->vm_flags
& VM_LOCKED
) {
738 pte_unmap_unlock(pte
, ptl
);
739 *mapcount
= 0; /* break early from loop */
740 *vm_flags
|= VM_LOCKED
;
744 if (ptep_clear_flush_young_notify(vma
, address
, pte
)) {
746 * Don't treat a reference through a sequentially read
747 * mapping as such. If the page has been used in
748 * another mapping, we will catch it; if this other
749 * mapping is already gone, the unmap path will have
750 * set PG_referenced or activated the page.
752 if (likely(!VM_SequentialReadHint(vma
)))
755 pte_unmap_unlock(pte
, ptl
);
761 *vm_flags
|= vma
->vm_flags
;
766 static int page_referenced_anon(struct page
*page
,
767 struct mem_cgroup
*memcg
,
768 unsigned long *vm_flags
)
770 unsigned int mapcount
;
771 struct anon_vma
*anon_vma
;
772 struct anon_vma_chain
*avc
;
775 anon_vma
= page_lock_anon_vma(page
);
779 mapcount
= page_mapcount(page
);
780 list_for_each_entry(avc
, &anon_vma
->head
, same_anon_vma
) {
781 struct vm_area_struct
*vma
= avc
->vma
;
782 unsigned long address
= vma_address(page
, vma
);
783 if (address
== -EFAULT
)
786 * If we are reclaiming on behalf of a cgroup, skip
787 * counting on behalf of references from different
790 if (memcg
&& !mm_match_cgroup(vma
->vm_mm
, memcg
))
792 referenced
+= page_referenced_one(page
, vma
, address
,
793 &mapcount
, vm_flags
);
798 page_unlock_anon_vma(anon_vma
);
803 * page_referenced_file - referenced check for object-based rmap
804 * @page: the page we're checking references on.
805 * @memcg: target memory control group
806 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
808 * For an object-based mapped page, find all the places it is mapped and
809 * check/clear the referenced flag. This is done by following the page->mapping
810 * pointer, then walking the chain of vmas it holds. It returns the number
811 * of references it found.
813 * This function is only called from page_referenced for object-based pages.
815 static int page_referenced_file(struct page
*page
,
816 struct mem_cgroup
*memcg
,
817 unsigned long *vm_flags
)
819 unsigned int mapcount
;
820 struct address_space
*mapping
= page
->mapping
;
821 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
822 struct vm_area_struct
*vma
;
826 * The caller's checks on page->mapping and !PageAnon have made
827 * sure that this is a file page: the check for page->mapping
828 * excludes the case just before it gets set on an anon page.
830 BUG_ON(PageAnon(page
));
833 * The page lock not only makes sure that page->mapping cannot
834 * suddenly be NULLified by truncation, it makes sure that the
835 * structure at mapping cannot be freed and reused yet,
836 * so we can safely take mapping->i_mmap_mutex.
838 BUG_ON(!PageLocked(page
));
840 mutex_lock(&mapping
->i_mmap_mutex
);
843 * i_mmap_mutex does not stabilize mapcount at all, but mapcount
844 * is more likely to be accurate if we note it after spinning.
846 mapcount
= page_mapcount(page
);
848 vma_interval_tree_foreach(vma
, &mapping
->i_mmap
, pgoff
, pgoff
) {
849 unsigned long address
= vma_address(page
, vma
);
850 if (address
== -EFAULT
)
853 * If we are reclaiming on behalf of a cgroup, skip
854 * counting on behalf of references from different
857 if (memcg
&& !mm_match_cgroup(vma
->vm_mm
, memcg
))
859 referenced
+= page_referenced_one(page
, vma
, address
,
860 &mapcount
, vm_flags
);
865 mutex_unlock(&mapping
->i_mmap_mutex
);
870 * page_referenced - test if the page was referenced
871 * @page: the page to test
872 * @is_locked: caller holds lock on the page
873 * @memcg: target memory cgroup
874 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
876 * Quick test_and_clear_referenced for all mappings to a page,
877 * returns the number of ptes which referenced the page.
879 int page_referenced(struct page
*page
,
881 struct mem_cgroup
*memcg
,
882 unsigned long *vm_flags
)
888 if (page_mapped(page
) && page_rmapping(page
)) {
889 if (!is_locked
&& (!PageAnon(page
) || PageKsm(page
))) {
890 we_locked
= trylock_page(page
);
896 if (unlikely(PageKsm(page
)))
897 referenced
+= page_referenced_ksm(page
, memcg
,
899 else if (PageAnon(page
))
900 referenced
+= page_referenced_anon(page
, memcg
,
902 else if (page
->mapping
)
903 referenced
+= page_referenced_file(page
, memcg
,
908 if (page_test_and_clear_young(page_to_pfn(page
)))
915 static int page_mkclean_one(struct page
*page
, struct vm_area_struct
*vma
,
916 unsigned long address
)
918 struct mm_struct
*mm
= vma
->vm_mm
;
923 pte
= page_check_address(page
, mm
, address
, &ptl
, 1);
927 if (pte_dirty(*pte
) || pte_write(*pte
)) {
930 flush_cache_page(vma
, address
, pte_pfn(*pte
));
931 entry
= ptep_clear_flush_notify(vma
, address
, pte
);
932 entry
= pte_wrprotect(entry
);
933 entry
= pte_mkclean(entry
);
934 set_pte_at(mm
, address
, pte
, entry
);
938 pte_unmap_unlock(pte
, ptl
);
943 static int page_mkclean_file(struct address_space
*mapping
, struct page
*page
)
945 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
946 struct vm_area_struct
*vma
;
949 BUG_ON(PageAnon(page
));
951 mutex_lock(&mapping
->i_mmap_mutex
);
952 vma_interval_tree_foreach(vma
, &mapping
->i_mmap
, pgoff
, pgoff
) {
953 if (vma
->vm_flags
& VM_SHARED
) {
954 unsigned long address
= vma_address(page
, vma
);
955 if (address
== -EFAULT
)
957 ret
+= page_mkclean_one(page
, vma
, address
);
960 mutex_unlock(&mapping
->i_mmap_mutex
);
964 int page_mkclean(struct page
*page
)
968 BUG_ON(!PageLocked(page
));
970 if (page_mapped(page
)) {
971 struct address_space
*mapping
= page_mapping(page
);
973 ret
= page_mkclean_file(mapping
, page
);
974 if (page_test_and_clear_dirty(page_to_pfn(page
), 1))
981 EXPORT_SYMBOL_GPL(page_mkclean
);
984 * page_move_anon_rmap - move a page to our anon_vma
985 * @page: the page to move to our anon_vma
986 * @vma: the vma the page belongs to
987 * @address: the user virtual address mapped
989 * When a page belongs exclusively to one process after a COW event,
990 * that page can be moved into the anon_vma that belongs to just that
991 * process, so the rmap code will not search the parent or sibling
994 void page_move_anon_rmap(struct page
*page
,
995 struct vm_area_struct
*vma
, unsigned long address
)
997 struct anon_vma
*anon_vma
= vma
->anon_vma
;
999 VM_BUG_ON(!PageLocked(page
));
1000 VM_BUG_ON(!anon_vma
);
1001 VM_BUG_ON(page
->index
!= linear_page_index(vma
, address
));
1003 anon_vma
= (void *) anon_vma
+ PAGE_MAPPING_ANON
;
1004 page
->mapping
= (struct address_space
*) anon_vma
;
1008 * __page_set_anon_rmap - set up new anonymous rmap
1009 * @page: Page to add to rmap
1010 * @vma: VM area to add page to.
1011 * @address: User virtual address of the mapping
1012 * @exclusive: the page is exclusively owned by the current process
1014 static void __page_set_anon_rmap(struct page
*page
,
1015 struct vm_area_struct
*vma
, unsigned long address
, int exclusive
)
1017 struct anon_vma
*anon_vma
= vma
->anon_vma
;
1025 * If the page isn't exclusively mapped into this vma,
1026 * we must use the _oldest_ possible anon_vma for the
1030 anon_vma
= anon_vma
->root
;
1032 anon_vma
= (void *) anon_vma
+ PAGE_MAPPING_ANON
;
1033 page
->mapping
= (struct address_space
*) anon_vma
;
1034 page
->index
= linear_page_index(vma
, address
);
1038 * __page_check_anon_rmap - sanity check anonymous rmap addition
1039 * @page: the page to add the mapping to
1040 * @vma: the vm area in which the mapping is added
1041 * @address: the user virtual address mapped
1043 static void __page_check_anon_rmap(struct page
*page
,
1044 struct vm_area_struct
*vma
, unsigned long address
)
1046 #ifdef CONFIG_DEBUG_VM
1048 * The page's anon-rmap details (mapping and index) are guaranteed to
1049 * be set up correctly at this point.
1051 * We have exclusion against page_add_anon_rmap because the caller
1052 * always holds the page locked, except if called from page_dup_rmap,
1053 * in which case the page is already known to be setup.
1055 * We have exclusion against page_add_new_anon_rmap because those pages
1056 * are initially only visible via the pagetables, and the pte is locked
1057 * over the call to page_add_new_anon_rmap.
1059 BUG_ON(page_anon_vma(page
)->root
!= vma
->anon_vma
->root
);
1060 BUG_ON(page
->index
!= linear_page_index(vma
, address
));
1065 * page_add_anon_rmap - add pte mapping to an anonymous page
1066 * @page: the page to add the mapping to
1067 * @vma: the vm area in which the mapping is added
1068 * @address: the user virtual address mapped
1070 * The caller needs to hold the pte lock, and the page must be locked in
1071 * the anon_vma case: to serialize mapping,index checking after setting,
1072 * and to ensure that PageAnon is not being upgraded racily to PageKsm
1073 * (but PageKsm is never downgraded to PageAnon).
1075 void page_add_anon_rmap(struct page
*page
,
1076 struct vm_area_struct
*vma
, unsigned long address
)
1078 do_page_add_anon_rmap(page
, vma
, address
, 0);
1082 * Special version of the above for do_swap_page, which often runs
1083 * into pages that are exclusively owned by the current process.
1084 * Everybody else should continue to use page_add_anon_rmap above.
1086 void do_page_add_anon_rmap(struct page
*page
,
1087 struct vm_area_struct
*vma
, unsigned long address
, int exclusive
)
1089 int first
= atomic_inc_and_test(&page
->_mapcount
);
1091 if (!PageTransHuge(page
))
1092 __inc_zone_page_state(page
, NR_ANON_PAGES
);
1094 __inc_zone_page_state(page
,
1095 NR_ANON_TRANSPARENT_HUGEPAGES
);
1097 if (unlikely(PageKsm(page
)))
1100 VM_BUG_ON(!PageLocked(page
));
1101 /* address might be in next vma when migration races vma_adjust */
1103 __page_set_anon_rmap(page
, vma
, address
, exclusive
);
1105 __page_check_anon_rmap(page
, vma
, address
);
1109 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1110 * @page: the page to add the mapping to
1111 * @vma: the vm area in which the mapping is added
1112 * @address: the user virtual address mapped
1114 * Same as page_add_anon_rmap but must only be called on *new* pages.
1115 * This means the inc-and-test can be bypassed.
1116 * Page does not have to be locked.
1118 void page_add_new_anon_rmap(struct page
*page
,
1119 struct vm_area_struct
*vma
, unsigned long address
)
1121 VM_BUG_ON(address
< vma
->vm_start
|| address
>= vma
->vm_end
);
1122 SetPageSwapBacked(page
);
1123 atomic_set(&page
->_mapcount
, 0); /* increment count (starts at -1) */
1124 if (!PageTransHuge(page
))
1125 __inc_zone_page_state(page
, NR_ANON_PAGES
);
1127 __inc_zone_page_state(page
, NR_ANON_TRANSPARENT_HUGEPAGES
);
1128 __page_set_anon_rmap(page
, vma
, address
, 1);
1129 if (page_evictable(page
, vma
))
1130 lru_cache_add_lru(page
, LRU_ACTIVE_ANON
);
1132 add_page_to_unevictable_list(page
);
1136 * page_add_file_rmap - add pte mapping to a file page
1137 * @page: the page to add the mapping to
1139 * The caller needs to hold the pte lock.
1141 void page_add_file_rmap(struct page
*page
)
1144 unsigned long flags
;
1146 mem_cgroup_begin_update_page_stat(page
, &locked
, &flags
);
1147 if (atomic_inc_and_test(&page
->_mapcount
)) {
1148 __inc_zone_page_state(page
, NR_FILE_MAPPED
);
1149 mem_cgroup_inc_page_stat(page
, MEMCG_NR_FILE_MAPPED
);
1151 mem_cgroup_end_update_page_stat(page
, &locked
, &flags
);
1155 * page_remove_rmap - take down pte mapping from a page
1156 * @page: page to remove mapping from
1158 * The caller needs to hold the pte lock.
1160 void page_remove_rmap(struct page
*page
)
1162 bool anon
= PageAnon(page
);
1164 unsigned long flags
;
1167 * The anon case has no mem_cgroup page_stat to update; but may
1168 * uncharge_page() below, where the lock ordering can deadlock if
1169 * we hold the lock against page_stat move: so avoid it on anon.
1172 mem_cgroup_begin_update_page_stat(page
, &locked
, &flags
);
1174 /* page still mapped by someone else? */
1175 if (!atomic_add_negative(-1, &page
->_mapcount
))
1179 * Now that the last pte has gone, s390 must transfer dirty
1180 * flag from storage key to struct page. We can usually skip
1181 * this if the page is anon, so about to be freed; but perhaps
1182 * not if it's in swapcache - there might be another pte slot
1183 * containing the swap entry, but page not yet written to swap.
1185 if ((!anon
|| PageSwapCache(page
)) &&
1186 page_test_and_clear_dirty(page_to_pfn(page
), 1))
1187 set_page_dirty(page
);
1189 * Hugepages are not counted in NR_ANON_PAGES nor NR_FILE_MAPPED
1190 * and not charged by memcg for now.
1192 if (unlikely(PageHuge(page
)))
1195 mem_cgroup_uncharge_page(page
);
1196 if (!PageTransHuge(page
))
1197 __dec_zone_page_state(page
, NR_ANON_PAGES
);
1199 __dec_zone_page_state(page
,
1200 NR_ANON_TRANSPARENT_HUGEPAGES
);
1202 __dec_zone_page_state(page
, NR_FILE_MAPPED
);
1203 mem_cgroup_dec_page_stat(page
, MEMCG_NR_FILE_MAPPED
);
1206 * It would be tidy to reset the PageAnon mapping here,
1207 * but that might overwrite a racing page_add_anon_rmap
1208 * which increments mapcount after us but sets mapping
1209 * before us: so leave the reset to free_hot_cold_page,
1210 * and remember that it's only reliable while mapped.
1211 * Leaving it set also helps swapoff to reinstate ptes
1212 * faster for those pages still in swapcache.
1216 mem_cgroup_end_update_page_stat(page
, &locked
, &flags
);
1220 * Subfunctions of try_to_unmap: try_to_unmap_one called
1221 * repeatedly from try_to_unmap_ksm, try_to_unmap_anon or try_to_unmap_file.
1223 int try_to_unmap_one(struct page
*page
, struct vm_area_struct
*vma
,
1224 unsigned long address
, enum ttu_flags flags
)
1226 struct mm_struct
*mm
= vma
->vm_mm
;
1230 int ret
= SWAP_AGAIN
;
1232 pte
= page_check_address(page
, mm
, address
, &ptl
, 0);
1237 * If the page is mlock()d, we cannot swap it out.
1238 * If it's recently referenced (perhaps page_referenced
1239 * skipped over this mm) then we should reactivate it.
1241 if (!(flags
& TTU_IGNORE_MLOCK
)) {
1242 if (vma
->vm_flags
& VM_LOCKED
)
1245 if (TTU_ACTION(flags
) == TTU_MUNLOCK
)
1248 if (!(flags
& TTU_IGNORE_ACCESS
)) {
1249 if (ptep_clear_flush_young_notify(vma
, address
, pte
)) {
1255 /* Nuke the page table entry. */
1256 flush_cache_page(vma
, address
, page_to_pfn(page
));
1257 pteval
= ptep_clear_flush_notify(vma
, address
, pte
);
1259 /* Move the dirty bit to the physical page now the pte is gone. */
1260 if (pte_dirty(pteval
))
1261 set_page_dirty(page
);
1263 /* Update high watermark before we lower rss */
1264 update_hiwater_rss(mm
);
1266 if (PageHWPoison(page
) && !(flags
& TTU_IGNORE_HWPOISON
)) {
1268 dec_mm_counter(mm
, MM_ANONPAGES
);
1270 dec_mm_counter(mm
, MM_FILEPAGES
);
1271 set_pte_at(mm
, address
, pte
,
1272 swp_entry_to_pte(make_hwpoison_entry(page
)));
1273 } else if (PageAnon(page
)) {
1274 swp_entry_t entry
= { .val
= page_private(page
) };
1276 if (PageSwapCache(page
)) {
1278 * Store the swap location in the pte.
1279 * See handle_pte_fault() ...
1281 if (swap_duplicate(entry
) < 0) {
1282 set_pte_at(mm
, address
, pte
, pteval
);
1286 if (list_empty(&mm
->mmlist
)) {
1287 spin_lock(&mmlist_lock
);
1288 if (list_empty(&mm
->mmlist
))
1289 list_add(&mm
->mmlist
, &init_mm
.mmlist
);
1290 spin_unlock(&mmlist_lock
);
1292 dec_mm_counter(mm
, MM_ANONPAGES
);
1293 inc_mm_counter(mm
, MM_SWAPENTS
);
1294 } else if (IS_ENABLED(CONFIG_MIGRATION
)) {
1296 * Store the pfn of the page in a special migration
1297 * pte. do_swap_page() will wait until the migration
1298 * pte is removed and then restart fault handling.
1300 BUG_ON(TTU_ACTION(flags
) != TTU_MIGRATION
);
1301 entry
= make_migration_entry(page
, pte_write(pteval
));
1303 set_pte_at(mm
, address
, pte
, swp_entry_to_pte(entry
));
1304 BUG_ON(pte_file(*pte
));
1305 } else if (IS_ENABLED(CONFIG_MIGRATION
) &&
1306 (TTU_ACTION(flags
) == TTU_MIGRATION
)) {
1307 /* Establish migration entry for a file page */
1309 entry
= make_migration_entry(page
, pte_write(pteval
));
1310 set_pte_at(mm
, address
, pte
, swp_entry_to_pte(entry
));
1312 dec_mm_counter(mm
, MM_FILEPAGES
);
1314 page_remove_rmap(page
);
1315 page_cache_release(page
);
1318 pte_unmap_unlock(pte
, ptl
);
1323 pte_unmap_unlock(pte
, ptl
);
1327 * We need mmap_sem locking, Otherwise VM_LOCKED check makes
1328 * unstable result and race. Plus, We can't wait here because
1329 * we now hold anon_vma->mutex or mapping->i_mmap_mutex.
1330 * if trylock failed, the page remain in evictable lru and later
1331 * vmscan could retry to move the page to unevictable lru if the
1332 * page is actually mlocked.
1334 if (down_read_trylock(&vma
->vm_mm
->mmap_sem
)) {
1335 if (vma
->vm_flags
& VM_LOCKED
) {
1336 mlock_vma_page(page
);
1339 up_read(&vma
->vm_mm
->mmap_sem
);
1345 * objrmap doesn't work for nonlinear VMAs because the assumption that
1346 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
1347 * Consequently, given a particular page and its ->index, we cannot locate the
1348 * ptes which are mapping that page without an exhaustive linear search.
1350 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
1351 * maps the file to which the target page belongs. The ->vm_private_data field
1352 * holds the current cursor into that scan. Successive searches will circulate
1353 * around the vma's virtual address space.
1355 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
1356 * more scanning pressure is placed against them as well. Eventually pages
1357 * will become fully unmapped and are eligible for eviction.
1359 * For very sparsely populated VMAs this is a little inefficient - chances are
1360 * there there won't be many ptes located within the scan cluster. In this case
1361 * maybe we could scan further - to the end of the pte page, perhaps.
1363 * Mlocked pages: check VM_LOCKED under mmap_sem held for read, if we can
1364 * acquire it without blocking. If vma locked, mlock the pages in the cluster,
1365 * rather than unmapping them. If we encounter the "check_page" that vmscan is
1366 * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN.
1368 #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
1369 #define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
1371 static int try_to_unmap_cluster(unsigned long cursor
, unsigned int *mapcount
,
1372 struct vm_area_struct
*vma
, struct page
*check_page
)
1374 struct mm_struct
*mm
= vma
->vm_mm
;
1382 unsigned long address
;
1384 int ret
= SWAP_AGAIN
;
1387 address
= (vma
->vm_start
+ cursor
) & CLUSTER_MASK
;
1388 end
= address
+ CLUSTER_SIZE
;
1389 if (address
< vma
->vm_start
)
1390 address
= vma
->vm_start
;
1391 if (end
> vma
->vm_end
)
1394 pgd
= pgd_offset(mm
, address
);
1395 if (!pgd_present(*pgd
))
1398 pud
= pud_offset(pgd
, address
);
1399 if (!pud_present(*pud
))
1402 pmd
= pmd_offset(pud
, address
);
1403 if (!pmd_present(*pmd
))
1407 * If we can acquire the mmap_sem for read, and vma is VM_LOCKED,
1408 * keep the sem while scanning the cluster for mlocking pages.
1410 if (down_read_trylock(&vma
->vm_mm
->mmap_sem
)) {
1411 locked_vma
= (vma
->vm_flags
& VM_LOCKED
);
1413 up_read(&vma
->vm_mm
->mmap_sem
); /* don't need it */
1416 pte
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
1418 /* Update high watermark before we lower rss */
1419 update_hiwater_rss(mm
);
1421 for (; address
< end
; pte
++, address
+= PAGE_SIZE
) {
1422 if (!pte_present(*pte
))
1424 page
= vm_normal_page(vma
, address
, *pte
);
1425 BUG_ON(!page
|| PageAnon(page
));
1428 mlock_vma_page(page
); /* no-op if already mlocked */
1429 if (page
== check_page
)
1431 continue; /* don't unmap */
1434 if (ptep_clear_flush_young_notify(vma
, address
, pte
))
1437 /* Nuke the page table entry. */
1438 flush_cache_page(vma
, address
, pte_pfn(*pte
));
1439 pteval
= ptep_clear_flush_notify(vma
, address
, pte
);
1441 /* If nonlinear, store the file page offset in the pte. */
1442 if (page
->index
!= linear_page_index(vma
, address
))
1443 set_pte_at(mm
, address
, pte
, pgoff_to_pte(page
->index
));
1445 /* Move the dirty bit to the physical page now the pte is gone. */
1446 if (pte_dirty(pteval
))
1447 set_page_dirty(page
);
1449 page_remove_rmap(page
);
1450 page_cache_release(page
);
1451 dec_mm_counter(mm
, MM_FILEPAGES
);
1454 pte_unmap_unlock(pte
- 1, ptl
);
1456 up_read(&vma
->vm_mm
->mmap_sem
);
1460 bool is_vma_temporary_stack(struct vm_area_struct
*vma
)
1462 int maybe_stack
= vma
->vm_flags
& (VM_GROWSDOWN
| VM_GROWSUP
);
1467 if ((vma
->vm_flags
& VM_STACK_INCOMPLETE_SETUP
) ==
1468 VM_STACK_INCOMPLETE_SETUP
)
1475 * try_to_unmap_anon - unmap or unlock anonymous page using the object-based
1477 * @page: the page to unmap/unlock
1478 * @flags: action and flags
1480 * Find all the mappings of a page using the mapping pointer and the vma chains
1481 * contained in the anon_vma struct it points to.
1483 * This function is only called from try_to_unmap/try_to_munlock for
1485 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1486 * where the page was found will be held for write. So, we won't recheck
1487 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1490 static int try_to_unmap_anon(struct page
*page
, enum ttu_flags flags
)
1492 struct anon_vma
*anon_vma
;
1493 struct anon_vma_chain
*avc
;
1494 int ret
= SWAP_AGAIN
;
1496 anon_vma
= page_lock_anon_vma(page
);
1500 list_for_each_entry(avc
, &anon_vma
->head
, same_anon_vma
) {
1501 struct vm_area_struct
*vma
= avc
->vma
;
1502 unsigned long address
;
1505 * During exec, a temporary VMA is setup and later moved.
1506 * The VMA is moved under the anon_vma lock but not the
1507 * page tables leading to a race where migration cannot
1508 * find the migration ptes. Rather than increasing the
1509 * locking requirements of exec(), migration skips
1510 * temporary VMAs until after exec() completes.
1512 if (IS_ENABLED(CONFIG_MIGRATION
) && (flags
& TTU_MIGRATION
) &&
1513 is_vma_temporary_stack(vma
))
1516 address
= vma_address(page
, vma
);
1517 if (address
== -EFAULT
)
1519 ret
= try_to_unmap_one(page
, vma
, address
, flags
);
1520 if (ret
!= SWAP_AGAIN
|| !page_mapped(page
))
1524 page_unlock_anon_vma(anon_vma
);
1529 * try_to_unmap_file - unmap/unlock file page using the object-based rmap method
1530 * @page: the page to unmap/unlock
1531 * @flags: action and flags
1533 * Find all the mappings of a page using the mapping pointer and the vma chains
1534 * contained in the address_space struct it points to.
1536 * This function is only called from try_to_unmap/try_to_munlock for
1537 * object-based pages.
1538 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1539 * where the page was found will be held for write. So, we won't recheck
1540 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1543 static int try_to_unmap_file(struct page
*page
, enum ttu_flags flags
)
1545 struct address_space
*mapping
= page
->mapping
;
1546 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
1547 struct vm_area_struct
*vma
;
1548 int ret
= SWAP_AGAIN
;
1549 unsigned long cursor
;
1550 unsigned long max_nl_cursor
= 0;
1551 unsigned long max_nl_size
= 0;
1552 unsigned int mapcount
;
1554 mutex_lock(&mapping
->i_mmap_mutex
);
1555 vma_interval_tree_foreach(vma
, &mapping
->i_mmap
, pgoff
, pgoff
) {
1556 unsigned long address
= vma_address(page
, vma
);
1557 if (address
== -EFAULT
)
1559 ret
= try_to_unmap_one(page
, vma
, address
, flags
);
1560 if (ret
!= SWAP_AGAIN
|| !page_mapped(page
))
1564 if (list_empty(&mapping
->i_mmap_nonlinear
))
1568 * We don't bother to try to find the munlocked page in nonlinears.
1569 * It's costly. Instead, later, page reclaim logic may call
1570 * try_to_unmap(TTU_MUNLOCK) and recover PG_mlocked lazily.
1572 if (TTU_ACTION(flags
) == TTU_MUNLOCK
)
1575 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
,
1577 cursor
= (unsigned long) vma
->vm_private_data
;
1578 if (cursor
> max_nl_cursor
)
1579 max_nl_cursor
= cursor
;
1580 cursor
= vma
->vm_end
- vma
->vm_start
;
1581 if (cursor
> max_nl_size
)
1582 max_nl_size
= cursor
;
1585 if (max_nl_size
== 0) { /* all nonlinears locked or reserved ? */
1591 * We don't try to search for this page in the nonlinear vmas,
1592 * and page_referenced wouldn't have found it anyway. Instead
1593 * just walk the nonlinear vmas trying to age and unmap some.
1594 * The mapcount of the page we came in with is irrelevant,
1595 * but even so use it as a guide to how hard we should try?
1597 mapcount
= page_mapcount(page
);
1602 max_nl_size
= (max_nl_size
+ CLUSTER_SIZE
- 1) & CLUSTER_MASK
;
1603 if (max_nl_cursor
== 0)
1604 max_nl_cursor
= CLUSTER_SIZE
;
1607 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
,
1609 cursor
= (unsigned long) vma
->vm_private_data
;
1610 while ( cursor
< max_nl_cursor
&&
1611 cursor
< vma
->vm_end
- vma
->vm_start
) {
1612 if (try_to_unmap_cluster(cursor
, &mapcount
,
1613 vma
, page
) == SWAP_MLOCK
)
1615 cursor
+= CLUSTER_SIZE
;
1616 vma
->vm_private_data
= (void *) cursor
;
1617 if ((int)mapcount
<= 0)
1620 vma
->vm_private_data
= (void *) max_nl_cursor
;
1623 max_nl_cursor
+= CLUSTER_SIZE
;
1624 } while (max_nl_cursor
<= max_nl_size
);
1627 * Don't loop forever (perhaps all the remaining pages are
1628 * in locked vmas). Reset cursor on all unreserved nonlinear
1629 * vmas, now forgetting on which ones it had fallen behind.
1631 list_for_each_entry(vma
, &mapping
->i_mmap_nonlinear
, shared
.nonlinear
)
1632 vma
->vm_private_data
= NULL
;
1634 mutex_unlock(&mapping
->i_mmap_mutex
);
1639 * try_to_unmap - try to remove all page table mappings to a page
1640 * @page: the page to get unmapped
1641 * @flags: action and flags
1643 * Tries to remove all the page table entries which are mapping this
1644 * page, used in the pageout path. Caller must hold the page lock.
1645 * Return values are:
1647 * SWAP_SUCCESS - we succeeded in removing all mappings
1648 * SWAP_AGAIN - we missed a mapping, try again later
1649 * SWAP_FAIL - the page is unswappable
1650 * SWAP_MLOCK - page is mlocked.
1652 int try_to_unmap(struct page
*page
, enum ttu_flags flags
)
1656 BUG_ON(!PageLocked(page
));
1657 VM_BUG_ON(!PageHuge(page
) && PageTransHuge(page
));
1659 if (unlikely(PageKsm(page
)))
1660 ret
= try_to_unmap_ksm(page
, flags
);
1661 else if (PageAnon(page
))
1662 ret
= try_to_unmap_anon(page
, flags
);
1664 ret
= try_to_unmap_file(page
, flags
);
1665 if (ret
!= SWAP_MLOCK
&& !page_mapped(page
))
1671 * try_to_munlock - try to munlock a page
1672 * @page: the page to be munlocked
1674 * Called from munlock code. Checks all of the VMAs mapping the page
1675 * to make sure nobody else has this page mlocked. The page will be
1676 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1678 * Return values are:
1680 * SWAP_AGAIN - no vma is holding page mlocked, or,
1681 * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem
1682 * SWAP_FAIL - page cannot be located at present
1683 * SWAP_MLOCK - page is now mlocked.
1685 int try_to_munlock(struct page
*page
)
1687 VM_BUG_ON(!PageLocked(page
) || PageLRU(page
));
1689 if (unlikely(PageKsm(page
)))
1690 return try_to_unmap_ksm(page
, TTU_MUNLOCK
);
1691 else if (PageAnon(page
))
1692 return try_to_unmap_anon(page
, TTU_MUNLOCK
);
1694 return try_to_unmap_file(page
, TTU_MUNLOCK
);
1697 void __put_anon_vma(struct anon_vma
*anon_vma
)
1699 struct anon_vma
*root
= anon_vma
->root
;
1701 if (root
!= anon_vma
&& atomic_dec_and_test(&root
->refcount
))
1702 anon_vma_free(root
);
1704 anon_vma_free(anon_vma
);
1707 #ifdef CONFIG_MIGRATION
1709 * rmap_walk() and its helpers rmap_walk_anon() and rmap_walk_file():
1710 * Called by migrate.c to remove migration ptes, but might be used more later.
1712 static int rmap_walk_anon(struct page
*page
, int (*rmap_one
)(struct page
*,
1713 struct vm_area_struct
*, unsigned long, void *), void *arg
)
1715 struct anon_vma
*anon_vma
;
1716 struct anon_vma_chain
*avc
;
1717 int ret
= SWAP_AGAIN
;
1720 * Note: remove_migration_ptes() cannot use page_lock_anon_vma()
1721 * because that depends on page_mapped(); but not all its usages
1722 * are holding mmap_sem. Users without mmap_sem are required to
1723 * take a reference count to prevent the anon_vma disappearing
1725 anon_vma
= page_anon_vma(page
);
1728 anon_vma_lock(anon_vma
);
1729 list_for_each_entry(avc
, &anon_vma
->head
, same_anon_vma
) {
1730 struct vm_area_struct
*vma
= avc
->vma
;
1731 unsigned long address
= vma_address(page
, vma
);
1732 if (address
== -EFAULT
)
1734 ret
= rmap_one(page
, vma
, address
, arg
);
1735 if (ret
!= SWAP_AGAIN
)
1738 anon_vma_unlock(anon_vma
);
1742 static int rmap_walk_file(struct page
*page
, int (*rmap_one
)(struct page
*,
1743 struct vm_area_struct
*, unsigned long, void *), void *arg
)
1745 struct address_space
*mapping
= page
->mapping
;
1746 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
1747 struct vm_area_struct
*vma
;
1748 int ret
= SWAP_AGAIN
;
1752 mutex_lock(&mapping
->i_mmap_mutex
);
1753 vma_interval_tree_foreach(vma
, &mapping
->i_mmap
, pgoff
, pgoff
) {
1754 unsigned long address
= vma_address(page
, vma
);
1755 if (address
== -EFAULT
)
1757 ret
= rmap_one(page
, vma
, address
, arg
);
1758 if (ret
!= SWAP_AGAIN
)
1762 * No nonlinear handling: being always shared, nonlinear vmas
1763 * never contain migration ptes. Decide what to do about this
1764 * limitation to linear when we need rmap_walk() on nonlinear.
1766 mutex_unlock(&mapping
->i_mmap_mutex
);
1770 int rmap_walk(struct page
*page
, int (*rmap_one
)(struct page
*,
1771 struct vm_area_struct
*, unsigned long, void *), void *arg
)
1773 VM_BUG_ON(!PageLocked(page
));
1775 if (unlikely(PageKsm(page
)))
1776 return rmap_walk_ksm(page
, rmap_one
, arg
);
1777 else if (PageAnon(page
))
1778 return rmap_walk_anon(page
, rmap_one
, arg
);
1780 return rmap_walk_file(page
, rmap_one
, arg
);
1782 #endif /* CONFIG_MIGRATION */
1784 #ifdef CONFIG_HUGETLB_PAGE
1786 * The following three functions are for anonymous (private mapped) hugepages.
1787 * Unlike common anonymous pages, anonymous hugepages have no accounting code
1788 * and no lru code, because we handle hugepages differently from common pages.
1790 static void __hugepage_set_anon_rmap(struct page
*page
,
1791 struct vm_area_struct
*vma
, unsigned long address
, int exclusive
)
1793 struct anon_vma
*anon_vma
= vma
->anon_vma
;
1800 anon_vma
= anon_vma
->root
;
1802 anon_vma
= (void *) anon_vma
+ PAGE_MAPPING_ANON
;
1803 page
->mapping
= (struct address_space
*) anon_vma
;
1804 page
->index
= linear_page_index(vma
, address
);
1807 void hugepage_add_anon_rmap(struct page
*page
,
1808 struct vm_area_struct
*vma
, unsigned long address
)
1810 struct anon_vma
*anon_vma
= vma
->anon_vma
;
1813 BUG_ON(!PageLocked(page
));
1815 /* address might be in next vma when migration races vma_adjust */
1816 first
= atomic_inc_and_test(&page
->_mapcount
);
1818 __hugepage_set_anon_rmap(page
, vma
, address
, 0);
1821 void hugepage_add_new_anon_rmap(struct page
*page
,
1822 struct vm_area_struct
*vma
, unsigned long address
)
1824 BUG_ON(address
< vma
->vm_start
|| address
>= vma
->vm_end
);
1825 atomic_set(&page
->_mapcount
, 0);
1826 __hugepage_set_anon_rmap(page
, vma
, address
, 1);
1828 #endif /* CONFIG_HUGETLB_PAGE */