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_rwsem (while writing or truncating, not reading or faulting)
25 * mapping->invalidate_lock (in filemap_fault)
26 * page->flags PG_locked (lock_page) * (see hugetlbfs below)
27 * hugetlbfs_i_mmap_rwsem_key (in huge_pmd_share)
28 * mapping->i_mmap_rwsem
29 * hugetlb_fault_mutex (hugetlbfs specific page fault mutex)
31 * mm->page_table_lock or pte_lock
32 * swap_lock (in swap_duplicate, swap_info_get)
33 * mmlist_lock (in mmput, drain_mmlist and others)
34 * mapping->private_lock (in __set_page_dirty_buffers)
35 * lock_page_memcg move_lock (in __set_page_dirty_buffers)
36 * i_pages lock (widely used)
37 * lruvec->lru_lock (in lock_page_lruvec_irq)
38 * inode->i_lock (in set_page_dirty's __mark_inode_dirty)
39 * bdi.wb->list_lock (in set_page_dirty's __mark_inode_dirty)
40 * sb_lock (within inode_lock in fs/fs-writeback.c)
41 * i_pages lock (widely used, in set_page_dirty,
42 * in arch-dependent flush_dcache_mmap_lock,
43 * within bdi.wb->list_lock in __sync_single_inode)
45 * anon_vma->rwsem,mapping->i_mmap_rwsem (memory_failure, collect_procs_anon)
49 * * hugetlbfs PageHuge() pages take locks in this order:
50 * mapping->i_mmap_rwsem
51 * hugetlb_fault_mutex (hugetlbfs specific page fault mutex)
52 * page->flags PG_locked (lock_page)
56 #include <linux/sched/mm.h>
57 #include <linux/sched/task.h>
58 #include <linux/pagemap.h>
59 #include <linux/swap.h>
60 #include <linux/swapops.h>
61 #include <linux/slab.h>
62 #include <linux/init.h>
63 #include <linux/ksm.h>
64 #include <linux/rmap.h>
65 #include <linux/rcupdate.h>
66 #include <linux/export.h>
67 #include <linux/memcontrol.h>
68 #include <linux/mmu_notifier.h>
69 #include <linux/migrate.h>
70 #include <linux/hugetlb.h>
71 #include <linux/huge_mm.h>
72 #include <linux/backing-dev.h>
73 #include <linux/page_idle.h>
74 #include <linux/memremap.h>
75 #include <linux/userfaultfd_k.h>
77 #include <asm/tlbflush.h>
79 #include <trace/events/tlb.h>
83 static struct kmem_cache
*anon_vma_cachep
;
84 static struct kmem_cache
*anon_vma_chain_cachep
;
86 static inline struct anon_vma
*anon_vma_alloc(void)
88 struct anon_vma
*anon_vma
;
90 anon_vma
= kmem_cache_alloc(anon_vma_cachep
, GFP_KERNEL
);
92 atomic_set(&anon_vma
->refcount
, 1);
93 anon_vma
->num_children
= 0;
94 anon_vma
->num_active_vmas
= 0;
95 anon_vma
->parent
= anon_vma
;
97 * Initialise the anon_vma root to point to itself. If called
98 * from fork, the root will be reset to the parents anon_vma.
100 anon_vma
->root
= anon_vma
;
106 static inline void anon_vma_free(struct anon_vma
*anon_vma
)
108 VM_BUG_ON(atomic_read(&anon_vma
->refcount
));
111 * Synchronize against page_lock_anon_vma_read() such that
112 * we can safely hold the lock without the anon_vma getting
115 * Relies on the full mb implied by the atomic_dec_and_test() from
116 * put_anon_vma() against the acquire barrier implied by
117 * down_read_trylock() from page_lock_anon_vma_read(). This orders:
119 * page_lock_anon_vma_read() VS put_anon_vma()
120 * down_read_trylock() atomic_dec_and_test()
122 * atomic_read() rwsem_is_locked()
124 * LOCK should suffice since the actual taking of the lock must
125 * happen _before_ what follows.
128 if (rwsem_is_locked(&anon_vma
->root
->rwsem
)) {
129 anon_vma_lock_write(anon_vma
);
130 anon_vma_unlock_write(anon_vma
);
133 kmem_cache_free(anon_vma_cachep
, anon_vma
);
136 static inline struct anon_vma_chain
*anon_vma_chain_alloc(gfp_t gfp
)
138 return kmem_cache_alloc(anon_vma_chain_cachep
, gfp
);
141 static void anon_vma_chain_free(struct anon_vma_chain
*anon_vma_chain
)
143 kmem_cache_free(anon_vma_chain_cachep
, anon_vma_chain
);
146 static void anon_vma_chain_link(struct vm_area_struct
*vma
,
147 struct anon_vma_chain
*avc
,
148 struct anon_vma
*anon_vma
)
151 avc
->anon_vma
= anon_vma
;
152 list_add(&avc
->same_vma
, &vma
->anon_vma_chain
);
153 anon_vma_interval_tree_insert(avc
, &anon_vma
->rb_root
);
157 * __anon_vma_prepare - attach an anon_vma to a memory region
158 * @vma: the memory region in question
160 * This makes sure the memory mapping described by 'vma' has
161 * an 'anon_vma' attached to it, so that we can associate the
162 * anonymous pages mapped into it with that anon_vma.
164 * The common case will be that we already have one, which
165 * is handled inline by anon_vma_prepare(). But if
166 * not we either need to find an adjacent mapping that we
167 * can re-use the anon_vma from (very common when the only
168 * reason for splitting a vma has been mprotect()), or we
169 * allocate a new one.
171 * Anon-vma allocations are very subtle, because we may have
172 * optimistically looked up an anon_vma in page_lock_anon_vma_read()
173 * and that may actually touch the rwsem even in the newly
174 * allocated vma (it depends on RCU to make sure that the
175 * anon_vma isn't actually destroyed).
177 * As a result, we need to do proper anon_vma locking even
178 * for the new allocation. At the same time, we do not want
179 * to do any locking for the common case of already having
182 * This must be called with the mmap_lock held for reading.
184 int __anon_vma_prepare(struct vm_area_struct
*vma
)
186 struct mm_struct
*mm
= vma
->vm_mm
;
187 struct anon_vma
*anon_vma
, *allocated
;
188 struct anon_vma_chain
*avc
;
192 avc
= anon_vma_chain_alloc(GFP_KERNEL
);
196 anon_vma
= find_mergeable_anon_vma(vma
);
199 anon_vma
= anon_vma_alloc();
200 if (unlikely(!anon_vma
))
201 goto out_enomem_free_avc
;
202 anon_vma
->num_children
++; /* self-parent link for new root */
203 allocated
= anon_vma
;
206 anon_vma_lock_write(anon_vma
);
207 /* page_table_lock to protect against threads */
208 spin_lock(&mm
->page_table_lock
);
209 if (likely(!vma
->anon_vma
)) {
210 vma
->anon_vma
= anon_vma
;
211 anon_vma_chain_link(vma
, avc
, anon_vma
);
212 anon_vma
->num_active_vmas
++;
216 spin_unlock(&mm
->page_table_lock
);
217 anon_vma_unlock_write(anon_vma
);
219 if (unlikely(allocated
))
220 put_anon_vma(allocated
);
222 anon_vma_chain_free(avc
);
227 anon_vma_chain_free(avc
);
233 * This is a useful helper function for locking the anon_vma root as
234 * we traverse the vma->anon_vma_chain, looping over anon_vma's that
237 * Such anon_vma's should have the same root, so you'd expect to see
238 * just a single mutex_lock for the whole traversal.
240 static inline struct anon_vma
*lock_anon_vma_root(struct anon_vma
*root
, struct anon_vma
*anon_vma
)
242 struct anon_vma
*new_root
= anon_vma
->root
;
243 if (new_root
!= root
) {
244 if (WARN_ON_ONCE(root
))
245 up_write(&root
->rwsem
);
247 down_write(&root
->rwsem
);
252 static inline void unlock_anon_vma_root(struct anon_vma
*root
)
255 up_write(&root
->rwsem
);
259 * Attach the anon_vmas from src to dst.
260 * Returns 0 on success, -ENOMEM on failure.
262 * anon_vma_clone() is called by __vma_adjust(), __split_vma(), copy_vma() and
263 * anon_vma_fork(). The first three want an exact copy of src, while the last
264 * one, anon_vma_fork(), may try to reuse an existing anon_vma to prevent
265 * endless growth of anon_vma. Since dst->anon_vma is set to NULL before call,
266 * we can identify this case by checking (!dst->anon_vma && src->anon_vma).
268 * If (!dst->anon_vma && src->anon_vma) is true, this function tries to find
269 * and reuse existing anon_vma which has no vmas and only one child anon_vma.
270 * This prevents degradation of anon_vma hierarchy to endless linear chain in
271 * case of constantly forking task. On the other hand, an anon_vma with more
272 * than one child isn't reused even if there was no alive vma, thus rmap
273 * walker has a good chance of avoiding scanning the whole hierarchy when it
274 * searches where page is mapped.
276 int anon_vma_clone(struct vm_area_struct
*dst
, struct vm_area_struct
*src
)
278 struct anon_vma_chain
*avc
, *pavc
;
279 struct anon_vma
*root
= NULL
;
281 list_for_each_entry_reverse(pavc
, &src
->anon_vma_chain
, same_vma
) {
282 struct anon_vma
*anon_vma
;
284 avc
= anon_vma_chain_alloc(GFP_NOWAIT
| __GFP_NOWARN
);
285 if (unlikely(!avc
)) {
286 unlock_anon_vma_root(root
);
288 avc
= anon_vma_chain_alloc(GFP_KERNEL
);
292 anon_vma
= pavc
->anon_vma
;
293 root
= lock_anon_vma_root(root
, anon_vma
);
294 anon_vma_chain_link(dst
, avc
, anon_vma
);
297 * Reuse existing anon_vma if it has no vma and only one
300 * Root anon_vma is never reused:
301 * it has self-parent reference and at least one child.
303 if (!dst
->anon_vma
&& src
->anon_vma
&&
304 anon_vma
->num_children
< 2 &&
305 anon_vma
->num_active_vmas
== 0)
306 dst
->anon_vma
= anon_vma
;
309 dst
->anon_vma
->num_active_vmas
++;
310 unlock_anon_vma_root(root
);
315 * dst->anon_vma is dropped here otherwise its degree can be incorrectly
316 * decremented in unlink_anon_vmas().
317 * We can safely do this because callers of anon_vma_clone() don't care
318 * about dst->anon_vma if anon_vma_clone() failed.
320 dst
->anon_vma
= NULL
;
321 unlink_anon_vmas(dst
);
326 * Attach vma to its own anon_vma, as well as to the anon_vmas that
327 * the corresponding VMA in the parent process is attached to.
328 * Returns 0 on success, non-zero on failure.
330 int anon_vma_fork(struct vm_area_struct
*vma
, struct vm_area_struct
*pvma
)
332 struct anon_vma_chain
*avc
;
333 struct anon_vma
*anon_vma
;
336 /* Don't bother if the parent process has no anon_vma here. */
340 /* Drop inherited anon_vma, we'll reuse existing or allocate new. */
341 vma
->anon_vma
= NULL
;
344 * First, attach the new VMA to the parent VMA's anon_vmas,
345 * so rmap can find non-COWed pages in child processes.
347 error
= anon_vma_clone(vma
, pvma
);
351 /* An existing anon_vma has been reused, all done then. */
355 /* Then add our own anon_vma. */
356 anon_vma
= anon_vma_alloc();
359 anon_vma
->num_active_vmas
++;
360 avc
= anon_vma_chain_alloc(GFP_KERNEL
);
362 goto out_error_free_anon_vma
;
365 * The root anon_vma's rwsem is the lock actually used when we
366 * lock any of the anon_vmas in this anon_vma tree.
368 anon_vma
->root
= pvma
->anon_vma
->root
;
369 anon_vma
->parent
= pvma
->anon_vma
;
371 * With refcounts, an anon_vma can stay around longer than the
372 * process it belongs to. The root anon_vma needs to be pinned until
373 * this anon_vma is freed, because the lock lives in the root.
375 get_anon_vma(anon_vma
->root
);
376 /* Mark this anon_vma as the one where our new (COWed) pages go. */
377 vma
->anon_vma
= anon_vma
;
378 anon_vma_lock_write(anon_vma
);
379 anon_vma_chain_link(vma
, avc
, anon_vma
);
380 anon_vma
->parent
->num_children
++;
381 anon_vma_unlock_write(anon_vma
);
385 out_error_free_anon_vma
:
386 put_anon_vma(anon_vma
);
388 unlink_anon_vmas(vma
);
392 void unlink_anon_vmas(struct vm_area_struct
*vma
)
394 struct anon_vma_chain
*avc
, *next
;
395 struct anon_vma
*root
= NULL
;
398 * Unlink each anon_vma chained to the VMA. This list is ordered
399 * from newest to oldest, ensuring the root anon_vma gets freed last.
401 list_for_each_entry_safe(avc
, next
, &vma
->anon_vma_chain
, same_vma
) {
402 struct anon_vma
*anon_vma
= avc
->anon_vma
;
404 root
= lock_anon_vma_root(root
, anon_vma
);
405 anon_vma_interval_tree_remove(avc
, &anon_vma
->rb_root
);
408 * Leave empty anon_vmas on the list - we'll need
409 * to free them outside the lock.
411 if (RB_EMPTY_ROOT(&anon_vma
->rb_root
.rb_root
)) {
412 anon_vma
->parent
->num_children
--;
416 list_del(&avc
->same_vma
);
417 anon_vma_chain_free(avc
);
420 vma
->anon_vma
->num_active_vmas
--;
423 * vma would still be needed after unlink, and anon_vma will be prepared
426 vma
->anon_vma
= NULL
;
428 unlock_anon_vma_root(root
);
431 * Iterate the list once more, it now only contains empty and unlinked
432 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
433 * needing to write-acquire the anon_vma->root->rwsem.
435 list_for_each_entry_safe(avc
, next
, &vma
->anon_vma_chain
, same_vma
) {
436 struct anon_vma
*anon_vma
= avc
->anon_vma
;
438 VM_WARN_ON(anon_vma
->num_children
);
439 VM_WARN_ON(anon_vma
->num_active_vmas
);
440 put_anon_vma(anon_vma
);
442 list_del(&avc
->same_vma
);
443 anon_vma_chain_free(avc
);
447 static void anon_vma_ctor(void *data
)
449 struct anon_vma
*anon_vma
= data
;
451 init_rwsem(&anon_vma
->rwsem
);
452 atomic_set(&anon_vma
->refcount
, 0);
453 anon_vma
->rb_root
= RB_ROOT_CACHED
;
456 void __init
anon_vma_init(void)
458 anon_vma_cachep
= kmem_cache_create("anon_vma", sizeof(struct anon_vma
),
459 0, SLAB_TYPESAFE_BY_RCU
|SLAB_PANIC
|SLAB_ACCOUNT
,
461 anon_vma_chain_cachep
= KMEM_CACHE(anon_vma_chain
,
462 SLAB_PANIC
|SLAB_ACCOUNT
);
466 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
468 * Since there is no serialization what so ever against page_remove_rmap()
469 * the best this function can do is return a refcount increased anon_vma
470 * that might have been relevant to this page.
472 * The page might have been remapped to a different anon_vma or the anon_vma
473 * returned may already be freed (and even reused).
475 * In case it was remapped to a different anon_vma, the new anon_vma will be a
476 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
477 * ensure that any anon_vma obtained from the page will still be valid for as
478 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
480 * All users of this function must be very careful when walking the anon_vma
481 * chain and verify that the page in question is indeed mapped in it
482 * [ something equivalent to page_mapped_in_vma() ].
484 * Since anon_vma's slab is SLAB_TYPESAFE_BY_RCU and we know from
485 * page_remove_rmap() that the anon_vma pointer from page->mapping is valid
486 * if there is a mapcount, we can dereference the anon_vma after observing
489 struct anon_vma
*page_get_anon_vma(struct page
*page
)
491 struct anon_vma
*anon_vma
= NULL
;
492 unsigned long anon_mapping
;
495 anon_mapping
= (unsigned long)READ_ONCE(page
->mapping
);
496 if ((anon_mapping
& PAGE_MAPPING_FLAGS
) != PAGE_MAPPING_ANON
)
498 if (!page_mapped(page
))
501 anon_vma
= (struct anon_vma
*) (anon_mapping
- PAGE_MAPPING_ANON
);
502 if (!atomic_inc_not_zero(&anon_vma
->refcount
)) {
508 * If this page is still mapped, then its anon_vma cannot have been
509 * freed. But if it has been unmapped, we have no security against the
510 * anon_vma structure being freed and reused (for another anon_vma:
511 * SLAB_TYPESAFE_BY_RCU guarantees that - so the atomic_inc_not_zero()
512 * above cannot corrupt).
514 if (!page_mapped(page
)) {
516 put_anon_vma(anon_vma
);
526 * Similar to page_get_anon_vma() except it locks the anon_vma.
528 * Its a little more complex as it tries to keep the fast path to a single
529 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
530 * reference like with page_get_anon_vma() and then block on the mutex.
532 struct anon_vma
*page_lock_anon_vma_read(struct page
*page
)
534 struct anon_vma
*anon_vma
= NULL
;
535 struct anon_vma
*root_anon_vma
;
536 unsigned long anon_mapping
;
539 anon_mapping
= (unsigned long)READ_ONCE(page
->mapping
);
540 if ((anon_mapping
& PAGE_MAPPING_FLAGS
) != PAGE_MAPPING_ANON
)
542 if (!page_mapped(page
))
545 anon_vma
= (struct anon_vma
*) (anon_mapping
- PAGE_MAPPING_ANON
);
546 root_anon_vma
= READ_ONCE(anon_vma
->root
);
547 if (down_read_trylock(&root_anon_vma
->rwsem
)) {
549 * If the page is still mapped, then this anon_vma is still
550 * its anon_vma, and holding the mutex ensures that it will
551 * not go away, see anon_vma_free().
553 if (!page_mapped(page
)) {
554 up_read(&root_anon_vma
->rwsem
);
560 /* trylock failed, we got to sleep */
561 if (!atomic_inc_not_zero(&anon_vma
->refcount
)) {
566 if (!page_mapped(page
)) {
568 put_anon_vma(anon_vma
);
572 /* we pinned the anon_vma, its safe to sleep */
574 anon_vma_lock_read(anon_vma
);
576 if (atomic_dec_and_test(&anon_vma
->refcount
)) {
578 * Oops, we held the last refcount, release the lock
579 * and bail -- can't simply use put_anon_vma() because
580 * we'll deadlock on the anon_vma_lock_write() recursion.
582 anon_vma_unlock_read(anon_vma
);
583 __put_anon_vma(anon_vma
);
594 void page_unlock_anon_vma_read(struct anon_vma
*anon_vma
)
596 anon_vma_unlock_read(anon_vma
);
599 #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
601 * Flush TLB entries for recently unmapped pages from remote CPUs. It is
602 * important if a PTE was dirty when it was unmapped that it's flushed
603 * before any IO is initiated on the page to prevent lost writes. Similarly,
604 * it must be flushed before freeing to prevent data leakage.
606 void try_to_unmap_flush(void)
608 struct tlbflush_unmap_batch
*tlb_ubc
= ¤t
->tlb_ubc
;
610 if (!tlb_ubc
->flush_required
)
613 arch_tlbbatch_flush(&tlb_ubc
->arch
);
614 tlb_ubc
->flush_required
= false;
615 tlb_ubc
->writable
= false;
618 /* Flush iff there are potentially writable TLB entries that can race with IO */
619 void try_to_unmap_flush_dirty(void)
621 struct tlbflush_unmap_batch
*tlb_ubc
= ¤t
->tlb_ubc
;
623 if (tlb_ubc
->writable
)
624 try_to_unmap_flush();
627 static void set_tlb_ubc_flush_pending(struct mm_struct
*mm
, bool writable
)
629 struct tlbflush_unmap_batch
*tlb_ubc
= ¤t
->tlb_ubc
;
631 arch_tlbbatch_add_mm(&tlb_ubc
->arch
, mm
);
632 tlb_ubc
->flush_required
= true;
635 * Ensure compiler does not re-order the setting of tlb_flush_batched
636 * before the PTE is cleared.
639 mm
->tlb_flush_batched
= true;
642 * If the PTE was dirty then it's best to assume it's writable. The
643 * caller must use try_to_unmap_flush_dirty() or try_to_unmap_flush()
644 * before the page is queued for IO.
647 tlb_ubc
->writable
= true;
651 * Returns true if the TLB flush should be deferred to the end of a batch of
652 * unmap operations to reduce IPIs.
654 static bool should_defer_flush(struct mm_struct
*mm
, enum ttu_flags flags
)
656 bool should_defer
= false;
658 if (!(flags
& TTU_BATCH_FLUSH
))
661 /* If remote CPUs need to be flushed then defer batch the flush */
662 if (cpumask_any_but(mm_cpumask(mm
), get_cpu()) < nr_cpu_ids
)
670 * Reclaim unmaps pages under the PTL but do not flush the TLB prior to
671 * releasing the PTL if TLB flushes are batched. It's possible for a parallel
672 * operation such as mprotect or munmap to race between reclaim unmapping
673 * the page and flushing the page. If this race occurs, it potentially allows
674 * access to data via a stale TLB entry. Tracking all mm's that have TLB
675 * batching in flight would be expensive during reclaim so instead track
676 * whether TLB batching occurred in the past and if so then do a flush here
677 * if required. This will cost one additional flush per reclaim cycle paid
678 * by the first operation at risk such as mprotect and mumap.
680 * This must be called under the PTL so that an access to tlb_flush_batched
681 * that is potentially a "reclaim vs mprotect/munmap/etc" race will synchronise
684 void flush_tlb_batched_pending(struct mm_struct
*mm
)
686 if (data_race(mm
->tlb_flush_batched
)) {
690 * Do not allow the compiler to re-order the clearing of
691 * tlb_flush_batched before the tlb is flushed.
694 mm
->tlb_flush_batched
= false;
698 static void set_tlb_ubc_flush_pending(struct mm_struct
*mm
, bool writable
)
702 static bool should_defer_flush(struct mm_struct
*mm
, enum ttu_flags flags
)
706 #endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
709 * At what user virtual address is page expected in vma?
710 * Caller should check the page is actually part of the vma.
712 unsigned long page_address_in_vma(struct page
*page
, struct vm_area_struct
*vma
)
714 if (PageAnon(page
)) {
715 struct anon_vma
*page__anon_vma
= page_anon_vma(page
);
717 * Note: swapoff's unuse_vma() is more efficient with this
718 * check, and needs it to match anon_vma when KSM is active.
720 if (!vma
->anon_vma
|| !page__anon_vma
||
721 vma
->anon_vma
->root
!= page__anon_vma
->root
)
723 } else if (!vma
->vm_file
) {
725 } else if (vma
->vm_file
->f_mapping
!= compound_head(page
)->mapping
) {
729 return vma_address(page
, vma
);
732 pmd_t
*mm_find_pmd(struct mm_struct
*mm
, unsigned long address
)
740 pgd
= pgd_offset(mm
, address
);
741 if (!pgd_present(*pgd
))
744 p4d
= p4d_offset(pgd
, address
);
745 if (!p4d_present(*p4d
))
748 pud
= pud_offset(p4d
, address
);
749 if (!pud_present(*pud
))
752 pmd
= pmd_offset(pud
, address
);
754 * Some THP functions use the sequence pmdp_huge_clear_flush(), set_pmd_at()
755 * without holding anon_vma lock for write. So when looking for a
756 * genuine pmde (in which to find pte), test present and !THP together.
760 if (!pmd_present(pmde
) || pmd_trans_huge(pmde
))
766 struct page_referenced_arg
{
769 unsigned long vm_flags
;
770 struct mem_cgroup
*memcg
;
773 * arg: page_referenced_arg will be passed
775 static bool page_referenced_one(struct page
*page
, struct vm_area_struct
*vma
,
776 unsigned long address
, void *arg
)
778 struct page_referenced_arg
*pra
= arg
;
779 struct page_vma_mapped_walk pvmw
= {
786 while (page_vma_mapped_walk(&pvmw
)) {
787 address
= pvmw
.address
;
789 if (vma
->vm_flags
& VM_LOCKED
) {
790 page_vma_mapped_walk_done(&pvmw
);
791 pra
->vm_flags
|= VM_LOCKED
;
792 return false; /* To break the loop */
796 if (ptep_clear_flush_young_notify(vma
, address
,
799 * Don't treat a reference through
800 * a sequentially read mapping as such.
801 * If the page has been used in another mapping,
802 * we will catch it; if this other mapping is
803 * already gone, the unmap path will have set
804 * PG_referenced or activated the page.
806 if (likely(!(vma
->vm_flags
& VM_SEQ_READ
)))
809 } else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE
)) {
810 if (pmdp_clear_flush_young_notify(vma
, address
,
814 /* unexpected pmd-mapped page? */
822 clear_page_idle(page
);
823 if (test_and_clear_page_young(page
))
828 pra
->vm_flags
|= vma
->vm_flags
;
832 return false; /* To break the loop */
837 static bool invalid_page_referenced_vma(struct vm_area_struct
*vma
, void *arg
)
839 struct page_referenced_arg
*pra
= arg
;
840 struct mem_cgroup
*memcg
= pra
->memcg
;
842 if (!mm_match_cgroup(vma
->vm_mm
, memcg
))
849 * page_referenced - test if the page was referenced
850 * @page: the page to test
851 * @is_locked: caller holds lock on the page
852 * @memcg: target memory cgroup
853 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
855 * Quick test_and_clear_referenced for all mappings to a page,
856 * returns the number of ptes which referenced the page.
858 int page_referenced(struct page
*page
,
860 struct mem_cgroup
*memcg
,
861 unsigned long *vm_flags
)
864 struct page_referenced_arg pra
= {
865 .mapcount
= total_mapcount(page
),
868 struct rmap_walk_control rwc
= {
869 .rmap_one
= page_referenced_one
,
871 .anon_lock
= page_lock_anon_vma_read
,
878 if (!page_rmapping(page
))
881 if (!is_locked
&& (!PageAnon(page
) || PageKsm(page
))) {
882 we_locked
= trylock_page(page
);
888 * If we are reclaiming on behalf of a cgroup, skip
889 * counting on behalf of references from different
893 rwc
.invalid_vma
= invalid_page_referenced_vma
;
896 rmap_walk(page
, &rwc
);
897 *vm_flags
= pra
.vm_flags
;
902 return pra
.referenced
;
905 static bool page_mkclean_one(struct page
*page
, struct vm_area_struct
*vma
,
906 unsigned long address
, void *arg
)
908 struct page_vma_mapped_walk pvmw
= {
914 struct mmu_notifier_range range
;
918 * We have to assume the worse case ie pmd for invalidation. Note that
919 * the page can not be free from this function.
921 mmu_notifier_range_init(&range
, MMU_NOTIFY_PROTECTION_PAGE
,
922 0, vma
, vma
->vm_mm
, address
,
923 vma_address_end(page
, vma
));
924 mmu_notifier_invalidate_range_start(&range
);
926 while (page_vma_mapped_walk(&pvmw
)) {
929 address
= pvmw
.address
;
932 pte_t
*pte
= pvmw
.pte
;
934 if (!pte_dirty(*pte
) && !pte_write(*pte
))
937 flush_cache_page(vma
, address
, pte_pfn(*pte
));
938 entry
= ptep_clear_flush(vma
, address
, pte
);
939 entry
= pte_wrprotect(entry
);
940 entry
= pte_mkclean(entry
);
941 set_pte_at(vma
->vm_mm
, address
, pte
, entry
);
944 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
945 pmd_t
*pmd
= pvmw
.pmd
;
948 if (!pmd_dirty(*pmd
) && !pmd_write(*pmd
))
951 flush_cache_page(vma
, address
, page_to_pfn(page
));
952 entry
= pmdp_invalidate(vma
, address
, pmd
);
953 entry
= pmd_wrprotect(entry
);
954 entry
= pmd_mkclean(entry
);
955 set_pmd_at(vma
->vm_mm
, address
, pmd
, entry
);
958 /* unexpected pmd-mapped page? */
964 * No need to call mmu_notifier_invalidate_range() as we are
965 * downgrading page table protection not changing it to point
968 * See Documentation/vm/mmu_notifier.rst
974 mmu_notifier_invalidate_range_end(&range
);
979 static bool invalid_mkclean_vma(struct vm_area_struct
*vma
, void *arg
)
981 if (vma
->vm_flags
& VM_SHARED
)
987 int page_mkclean(struct page
*page
)
990 struct address_space
*mapping
;
991 struct rmap_walk_control rwc
= {
992 .arg
= (void *)&cleaned
,
993 .rmap_one
= page_mkclean_one
,
994 .invalid_vma
= invalid_mkclean_vma
,
997 BUG_ON(!PageLocked(page
));
999 if (!page_mapped(page
))
1002 mapping
= page_mapping(page
);
1006 rmap_walk(page
, &rwc
);
1010 EXPORT_SYMBOL_GPL(page_mkclean
);
1013 * page_move_anon_rmap - move a page to our anon_vma
1014 * @page: the page to move to our anon_vma
1015 * @vma: the vma the page belongs to
1017 * When a page belongs exclusively to one process after a COW event,
1018 * that page can be moved into the anon_vma that belongs to just that
1019 * process, so the rmap code will not search the parent or sibling
1022 void page_move_anon_rmap(struct page
*page
, struct vm_area_struct
*vma
)
1024 struct anon_vma
*anon_vma
= vma
->anon_vma
;
1026 page
= compound_head(page
);
1028 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
1029 VM_BUG_ON_VMA(!anon_vma
, vma
);
1031 anon_vma
= (void *) anon_vma
+ PAGE_MAPPING_ANON
;
1033 * Ensure that anon_vma and the PAGE_MAPPING_ANON bit are written
1034 * simultaneously, so a concurrent reader (eg page_referenced()'s
1035 * PageAnon()) will not see one without the other.
1037 WRITE_ONCE(page
->mapping
, (struct address_space
*) anon_vma
);
1041 * __page_set_anon_rmap - set up new anonymous rmap
1042 * @page: Page or Hugepage to add to rmap
1043 * @vma: VM area to add page to.
1044 * @address: User virtual address of the mapping
1045 * @exclusive: the page is exclusively owned by the current process
1047 static void __page_set_anon_rmap(struct page
*page
,
1048 struct vm_area_struct
*vma
, unsigned long address
, int exclusive
)
1050 struct anon_vma
*anon_vma
= vma
->anon_vma
;
1058 * If the page isn't exclusively mapped into this vma,
1059 * we must use the _oldest_ possible anon_vma for the
1063 anon_vma
= anon_vma
->root
;
1066 * page_idle does a lockless/optimistic rmap scan on page->mapping.
1067 * Make sure the compiler doesn't split the stores of anon_vma and
1068 * the PAGE_MAPPING_ANON type identifier, otherwise the rmap code
1069 * could mistake the mapping for a struct address_space and crash.
1071 anon_vma
= (void *) anon_vma
+ PAGE_MAPPING_ANON
;
1072 WRITE_ONCE(page
->mapping
, (struct address_space
*) anon_vma
);
1073 page
->index
= linear_page_index(vma
, address
);
1077 * __page_check_anon_rmap - sanity check anonymous rmap addition
1078 * @page: the page to add the mapping to
1079 * @vma: the vm area in which the mapping is added
1080 * @address: the user virtual address mapped
1082 static void __page_check_anon_rmap(struct page
*page
,
1083 struct vm_area_struct
*vma
, unsigned long address
)
1086 * The page's anon-rmap details (mapping and index) are guaranteed to
1087 * be set up correctly at this point.
1089 * We have exclusion against page_add_anon_rmap because the caller
1090 * always holds the page locked.
1092 * We have exclusion against page_add_new_anon_rmap because those pages
1093 * are initially only visible via the pagetables, and the pte is locked
1094 * over the call to page_add_new_anon_rmap.
1096 VM_BUG_ON_PAGE(page_anon_vma(page
)->root
!= vma
->anon_vma
->root
, page
);
1097 VM_BUG_ON_PAGE(page_to_pgoff(page
) != linear_page_index(vma
, address
),
1102 * page_add_anon_rmap - add pte mapping to an anonymous page
1103 * @page: the page to add the mapping to
1104 * @vma: the vm area in which the mapping is added
1105 * @address: the user virtual address mapped
1106 * @compound: charge the page as compound or small page
1108 * The caller needs to hold the pte lock, and the page must be locked in
1109 * the anon_vma case: to serialize mapping,index checking after setting,
1110 * and to ensure that PageAnon is not being upgraded racily to PageKsm
1111 * (but PageKsm is never downgraded to PageAnon).
1113 void page_add_anon_rmap(struct page
*page
,
1114 struct vm_area_struct
*vma
, unsigned long address
, bool compound
)
1116 do_page_add_anon_rmap(page
, vma
, address
, compound
? RMAP_COMPOUND
: 0);
1120 * Special version of the above for do_swap_page, which often runs
1121 * into pages that are exclusively owned by the current process.
1122 * Everybody else should continue to use page_add_anon_rmap above.
1124 void do_page_add_anon_rmap(struct page
*page
,
1125 struct vm_area_struct
*vma
, unsigned long address
, int flags
)
1127 bool compound
= flags
& RMAP_COMPOUND
;
1130 if (unlikely(PageKsm(page
)))
1131 lock_page_memcg(page
);
1133 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
1137 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
1138 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
1139 mapcount
= compound_mapcount_ptr(page
);
1140 first
= atomic_inc_and_test(mapcount
);
1142 first
= atomic_inc_and_test(&page
->_mapcount
);
1146 int nr
= compound
? thp_nr_pages(page
) : 1;
1148 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1149 * these counters are not modified in interrupt context, and
1150 * pte lock(a spinlock) is held, which implies preemption
1154 __mod_lruvec_page_state(page
, NR_ANON_THPS
, nr
);
1155 __mod_lruvec_page_state(page
, NR_ANON_MAPPED
, nr
);
1158 if (unlikely(PageKsm(page
))) {
1159 unlock_page_memcg(page
);
1163 /* address might be in next vma when migration races vma_adjust */
1165 __page_set_anon_rmap(page
, vma
, address
,
1166 flags
& RMAP_EXCLUSIVE
);
1168 __page_check_anon_rmap(page
, vma
, address
);
1172 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1173 * @page: the page to add the mapping to
1174 * @vma: the vm area in which the mapping is added
1175 * @address: the user virtual address mapped
1176 * @compound: charge the page as compound or small page
1178 * Same as page_add_anon_rmap but must only be called on *new* pages.
1179 * This means the inc-and-test can be bypassed.
1180 * Page does not have to be locked.
1182 void page_add_new_anon_rmap(struct page
*page
,
1183 struct vm_area_struct
*vma
, unsigned long address
, bool compound
)
1185 int nr
= compound
? thp_nr_pages(page
) : 1;
1187 VM_BUG_ON_VMA(address
< vma
->vm_start
|| address
>= vma
->vm_end
, vma
);
1188 __SetPageSwapBacked(page
);
1190 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
1191 /* increment count (starts at -1) */
1192 atomic_set(compound_mapcount_ptr(page
), 0);
1193 if (hpage_pincount_available(page
))
1194 atomic_set(compound_pincount_ptr(page
), 0);
1196 __mod_lruvec_page_state(page
, NR_ANON_THPS
, nr
);
1198 /* Anon THP always mapped first with PMD */
1199 VM_BUG_ON_PAGE(PageTransCompound(page
), page
);
1200 /* increment count (starts at -1) */
1201 atomic_set(&page
->_mapcount
, 0);
1203 __mod_lruvec_page_state(page
, NR_ANON_MAPPED
, nr
);
1204 __page_set_anon_rmap(page
, vma
, address
, 1);
1208 * page_add_file_rmap - add pte mapping to a file page
1209 * @page: the page to add the mapping to
1210 * @compound: charge the page as compound or small page
1212 * The caller needs to hold the pte lock.
1214 void page_add_file_rmap(struct page
*page
, bool compound
)
1218 VM_BUG_ON_PAGE(compound
&& !PageTransHuge(page
), page
);
1219 lock_page_memcg(page
);
1220 if (compound
&& PageTransHuge(page
)) {
1221 int nr_pages
= thp_nr_pages(page
);
1223 for (i
= 0, nr
= 0; i
< nr_pages
; i
++) {
1224 if (atomic_inc_and_test(&page
[i
]._mapcount
))
1227 if (!atomic_inc_and_test(compound_mapcount_ptr(page
)))
1229 if (PageSwapBacked(page
))
1230 __mod_lruvec_page_state(page
, NR_SHMEM_PMDMAPPED
,
1233 __mod_lruvec_page_state(page
, NR_FILE_PMDMAPPED
,
1236 if (PageTransCompound(page
) && page_mapping(page
)) {
1237 struct page
*head
= compound_head(page
);
1239 VM_WARN_ON_ONCE(!PageLocked(page
));
1241 SetPageDoubleMap(head
);
1242 if (PageMlocked(page
))
1243 clear_page_mlock(head
);
1245 if (!atomic_inc_and_test(&page
->_mapcount
))
1248 __mod_lruvec_page_state(page
, NR_FILE_MAPPED
, nr
);
1250 unlock_page_memcg(page
);
1253 static void page_remove_file_rmap(struct page
*page
, bool compound
)
1257 VM_BUG_ON_PAGE(compound
&& !PageHead(page
), page
);
1259 /* Hugepages are not counted in NR_FILE_MAPPED for now. */
1260 if (unlikely(PageHuge(page
))) {
1261 /* hugetlb pages are always mapped with pmds */
1262 atomic_dec(compound_mapcount_ptr(page
));
1266 /* page still mapped by someone else? */
1267 if (compound
&& PageTransHuge(page
)) {
1268 int nr_pages
= thp_nr_pages(page
);
1270 for (i
= 0, nr
= 0; i
< nr_pages
; i
++) {
1271 if (atomic_add_negative(-1, &page
[i
]._mapcount
))
1274 if (!atomic_add_negative(-1, compound_mapcount_ptr(page
)))
1276 if (PageSwapBacked(page
))
1277 __mod_lruvec_page_state(page
, NR_SHMEM_PMDMAPPED
,
1280 __mod_lruvec_page_state(page
, NR_FILE_PMDMAPPED
,
1283 if (!atomic_add_negative(-1, &page
->_mapcount
))
1288 * We use the irq-unsafe __{inc|mod}_lruvec_page_state because
1289 * these counters are not modified in interrupt context, and
1290 * pte lock(a spinlock) is held, which implies preemption disabled.
1292 __mod_lruvec_page_state(page
, NR_FILE_MAPPED
, -nr
);
1294 if (unlikely(PageMlocked(page
)))
1295 clear_page_mlock(page
);
1298 static void page_remove_anon_compound_rmap(struct page
*page
)
1302 if (!atomic_add_negative(-1, compound_mapcount_ptr(page
)))
1305 /* Hugepages are not counted in NR_ANON_PAGES for now. */
1306 if (unlikely(PageHuge(page
)))
1309 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE
))
1312 __mod_lruvec_page_state(page
, NR_ANON_THPS
, -thp_nr_pages(page
));
1314 if (TestClearPageDoubleMap(page
)) {
1316 * Subpages can be mapped with PTEs too. Check how many of
1317 * them are still mapped.
1319 for (i
= 0, nr
= 0; i
< thp_nr_pages(page
); i
++) {
1320 if (atomic_add_negative(-1, &page
[i
]._mapcount
))
1325 * Queue the page for deferred split if at least one small
1326 * page of the compound page is unmapped, but at least one
1327 * small page is still mapped.
1329 if (nr
&& nr
< thp_nr_pages(page
))
1330 deferred_split_huge_page(page
);
1332 nr
= thp_nr_pages(page
);
1335 if (unlikely(PageMlocked(page
)))
1336 clear_page_mlock(page
);
1339 __mod_lruvec_page_state(page
, NR_ANON_MAPPED
, -nr
);
1343 * page_remove_rmap - take down pte mapping from a page
1344 * @page: page to remove mapping from
1345 * @compound: uncharge the page as compound or small page
1347 * The caller needs to hold the pte lock.
1349 void page_remove_rmap(struct page
*page
, bool compound
)
1351 lock_page_memcg(page
);
1353 if (!PageAnon(page
)) {
1354 page_remove_file_rmap(page
, compound
);
1359 page_remove_anon_compound_rmap(page
);
1363 /* page still mapped by someone else? */
1364 if (!atomic_add_negative(-1, &page
->_mapcount
))
1368 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1369 * these counters are not modified in interrupt context, and
1370 * pte lock(a spinlock) is held, which implies preemption disabled.
1372 __dec_lruvec_page_state(page
, NR_ANON_MAPPED
);
1374 if (unlikely(PageMlocked(page
)))
1375 clear_page_mlock(page
);
1377 if (PageTransCompound(page
))
1378 deferred_split_huge_page(compound_head(page
));
1381 * It would be tidy to reset the PageAnon mapping here,
1382 * but that might overwrite a racing page_add_anon_rmap
1383 * which increments mapcount after us but sets mapping
1384 * before us: so leave the reset to free_unref_page,
1385 * and remember that it's only reliable while mapped.
1386 * Leaving it set also helps swapoff to reinstate ptes
1387 * faster for those pages still in swapcache.
1390 unlock_page_memcg(page
);
1394 * @arg: enum ttu_flags will be passed to this argument
1396 static bool try_to_unmap_one(struct page
*page
, struct vm_area_struct
*vma
,
1397 unsigned long address
, void *arg
)
1399 struct mm_struct
*mm
= vma
->vm_mm
;
1400 struct page_vma_mapped_walk pvmw
= {
1406 struct page
*subpage
;
1408 struct mmu_notifier_range range
;
1409 enum ttu_flags flags
= (enum ttu_flags
)(long)arg
;
1412 * When racing against e.g. zap_pte_range() on another cpu,
1413 * in between its ptep_get_and_clear_full() and page_remove_rmap(),
1414 * try_to_unmap() may return before page_mapped() has become false,
1415 * if page table locking is skipped: use TTU_SYNC to wait for that.
1417 if (flags
& TTU_SYNC
)
1418 pvmw
.flags
= PVMW_SYNC
;
1420 if (flags
& TTU_SPLIT_HUGE_PMD
)
1421 split_huge_pmd_address(vma
, address
, false, page
);
1424 * For THP, we have to assume the worse case ie pmd for invalidation.
1425 * For hugetlb, it could be much worse if we need to do pud
1426 * invalidation in the case of pmd sharing.
1428 * Note that the page can not be free in this function as call of
1429 * try_to_unmap() must hold a reference on the page.
1431 range
.end
= PageKsm(page
) ?
1432 address
+ PAGE_SIZE
: vma_address_end(page
, vma
);
1433 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
, vma
->vm_mm
,
1434 address
, range
.end
);
1435 if (PageHuge(page
)) {
1437 * If sharing is possible, start and end will be adjusted
1440 adjust_range_if_pmd_sharing_possible(vma
, &range
.start
,
1443 mmu_notifier_invalidate_range_start(&range
);
1445 while (page_vma_mapped_walk(&pvmw
)) {
1447 * If the page is mlock()d, we cannot swap it out.
1449 if (!(flags
& TTU_IGNORE_MLOCK
) &&
1450 (vma
->vm_flags
& VM_LOCKED
)) {
1452 * PTE-mapped THP are never marked as mlocked: so do
1453 * not set it on a DoubleMap THP, nor on an Anon THP
1454 * (which may still be PTE-mapped after DoubleMap was
1455 * cleared). But stop unmapping even in those cases.
1457 if (!PageTransCompound(page
) || (PageHead(page
) &&
1458 !PageDoubleMap(page
) && !PageAnon(page
)))
1459 mlock_vma_page(page
);
1460 page_vma_mapped_walk_done(&pvmw
);
1465 /* Unexpected PMD-mapped THP? */
1466 VM_BUG_ON_PAGE(!pvmw
.pte
, page
);
1468 subpage
= page
- page_to_pfn(page
) + pte_pfn(*pvmw
.pte
);
1469 address
= pvmw
.address
;
1471 if (PageHuge(page
) && !PageAnon(page
)) {
1473 * To call huge_pmd_unshare, i_mmap_rwsem must be
1474 * held in write mode. Caller needs to explicitly
1475 * do this outside rmap routines.
1477 VM_BUG_ON(!(flags
& TTU_RMAP_LOCKED
));
1478 if (huge_pmd_unshare(mm
, vma
, &address
, pvmw
.pte
)) {
1480 * huge_pmd_unshare unmapped an entire PMD
1481 * page. There is no way of knowing exactly
1482 * which PMDs may be cached for this mm, so
1483 * we must flush them all. start/end were
1484 * already adjusted above to cover this range.
1486 flush_cache_range(vma
, range
.start
, range
.end
);
1487 flush_tlb_range(vma
, range
.start
, range
.end
);
1488 mmu_notifier_invalidate_range(mm
, range
.start
,
1492 * The ref count of the PMD page was dropped
1493 * which is part of the way map counting
1494 * is done for shared PMDs. Return 'true'
1495 * here. When there is no other sharing,
1496 * huge_pmd_unshare returns false and we will
1497 * unmap the actual page and drop map count
1500 page_vma_mapped_walk_done(&pvmw
);
1505 /* Nuke the page table entry. */
1506 flush_cache_page(vma
, address
, pte_pfn(*pvmw
.pte
));
1507 if (should_defer_flush(mm
, flags
)) {
1509 * We clear the PTE but do not flush so potentially
1510 * a remote CPU could still be writing to the page.
1511 * If the entry was previously clean then the
1512 * architecture must guarantee that a clear->dirty
1513 * transition on a cached TLB entry is written through
1514 * and traps if the PTE is unmapped.
1516 pteval
= ptep_get_and_clear(mm
, address
, pvmw
.pte
);
1518 set_tlb_ubc_flush_pending(mm
, pte_dirty(pteval
));
1520 pteval
= ptep_clear_flush(vma
, address
, pvmw
.pte
);
1523 /* Move the dirty bit to the page. Now the pte is gone. */
1524 if (pte_dirty(pteval
))
1525 set_page_dirty(page
);
1527 /* Update high watermark before we lower rss */
1528 update_hiwater_rss(mm
);
1530 if (PageHWPoison(page
) && !(flags
& TTU_IGNORE_HWPOISON
)) {
1531 pteval
= swp_entry_to_pte(make_hwpoison_entry(subpage
));
1532 if (PageHuge(page
)) {
1533 hugetlb_count_sub(compound_nr(page
), mm
);
1534 set_huge_swap_pte_at(mm
, address
,
1536 vma_mmu_pagesize(vma
));
1538 dec_mm_counter(mm
, mm_counter(page
));
1539 set_pte_at(mm
, address
, pvmw
.pte
, pteval
);
1542 } else if (pte_unused(pteval
) && !userfaultfd_armed(vma
)) {
1544 * The guest indicated that the page content is of no
1545 * interest anymore. Simply discard the pte, vmscan
1546 * will take care of the rest.
1547 * A future reference will then fault in a new zero
1548 * page. When userfaultfd is active, we must not drop
1549 * this page though, as its main user (postcopy
1550 * migration) will not expect userfaults on already
1553 dec_mm_counter(mm
, mm_counter(page
));
1554 /* We have to invalidate as we cleared the pte */
1555 mmu_notifier_invalidate_range(mm
, address
,
1556 address
+ PAGE_SIZE
);
1557 } else if (PageAnon(page
)) {
1558 swp_entry_t entry
= { .val
= page_private(subpage
) };
1561 * Store the swap location in the pte.
1562 * See handle_pte_fault() ...
1564 if (unlikely(PageSwapBacked(page
) != PageSwapCache(page
))) {
1567 /* We have to invalidate as we cleared the pte */
1568 mmu_notifier_invalidate_range(mm
, address
,
1569 address
+ PAGE_SIZE
);
1570 page_vma_mapped_walk_done(&pvmw
);
1574 /* MADV_FREE page check */
1575 if (!PageSwapBacked(page
)) {
1576 int ref_count
, map_count
;
1579 * Synchronize with gup_pte_range():
1580 * - clear PTE; barrier; read refcount
1581 * - inc refcount; barrier; read PTE
1585 ref_count
= page_ref_count(page
);
1586 map_count
= page_mapcount(page
);
1589 * Order reads for page refcount and dirty flag
1590 * (see comments in __remove_mapping()).
1595 * The only page refs must be one from isolation
1596 * plus the rmap(s) (dropped by discard:).
1598 if (ref_count
== 1 + map_count
&&
1600 /* Invalidate as we cleared the pte */
1601 mmu_notifier_invalidate_range(mm
,
1602 address
, address
+ PAGE_SIZE
);
1603 dec_mm_counter(mm
, MM_ANONPAGES
);
1608 * If the page was redirtied, it cannot be
1609 * discarded. Remap the page to page table.
1611 set_pte_at(mm
, address
, pvmw
.pte
, pteval
);
1612 SetPageSwapBacked(page
);
1614 page_vma_mapped_walk_done(&pvmw
);
1618 if (swap_duplicate(entry
) < 0) {
1619 set_pte_at(mm
, address
, pvmw
.pte
, pteval
);
1621 page_vma_mapped_walk_done(&pvmw
);
1624 if (arch_unmap_one(mm
, vma
, address
, pteval
) < 0) {
1625 set_pte_at(mm
, address
, pvmw
.pte
, pteval
);
1627 page_vma_mapped_walk_done(&pvmw
);
1630 if (list_empty(&mm
->mmlist
)) {
1631 spin_lock(&mmlist_lock
);
1632 if (list_empty(&mm
->mmlist
))
1633 list_add(&mm
->mmlist
, &init_mm
.mmlist
);
1634 spin_unlock(&mmlist_lock
);
1636 dec_mm_counter(mm
, MM_ANONPAGES
);
1637 inc_mm_counter(mm
, MM_SWAPENTS
);
1638 swp_pte
= swp_entry_to_pte(entry
);
1639 if (pte_soft_dirty(pteval
))
1640 swp_pte
= pte_swp_mksoft_dirty(swp_pte
);
1641 if (pte_uffd_wp(pteval
))
1642 swp_pte
= pte_swp_mkuffd_wp(swp_pte
);
1643 set_pte_at(mm
, address
, pvmw
.pte
, swp_pte
);
1644 /* Invalidate as we cleared the pte */
1645 mmu_notifier_invalidate_range(mm
, address
,
1646 address
+ PAGE_SIZE
);
1649 * This is a locked file-backed page, thus it cannot
1650 * be removed from the page cache and replaced by a new
1651 * page before mmu_notifier_invalidate_range_end, so no
1652 * concurrent thread might update its page table to
1653 * point at new page while a device still is using this
1656 * See Documentation/vm/mmu_notifier.rst
1658 dec_mm_counter(mm
, mm_counter_file(page
));
1662 * No need to call mmu_notifier_invalidate_range() it has be
1663 * done above for all cases requiring it to happen under page
1664 * table lock before mmu_notifier_invalidate_range_end()
1666 * See Documentation/vm/mmu_notifier.rst
1668 page_remove_rmap(subpage
, PageHuge(page
));
1672 mmu_notifier_invalidate_range_end(&range
);
1677 static bool invalid_migration_vma(struct vm_area_struct
*vma
, void *arg
)
1679 return vma_is_temporary_stack(vma
);
1682 static int page_not_mapped(struct page
*page
)
1684 return !page_mapped(page
);
1688 * try_to_unmap - try to remove all page table mappings to a page
1689 * @page: the page to get unmapped
1690 * @flags: action and flags
1692 * Tries to remove all the page table entries which are mapping this
1693 * page, used in the pageout path. Caller must hold the page lock.
1695 * It is the caller's responsibility to check if the page is still
1696 * mapped when needed (use TTU_SYNC to prevent accounting races).
1698 void try_to_unmap(struct page
*page
, enum ttu_flags flags
)
1700 struct rmap_walk_control rwc
= {
1701 .rmap_one
= try_to_unmap_one
,
1702 .arg
= (void *)flags
,
1703 .done
= page_not_mapped
,
1704 .anon_lock
= page_lock_anon_vma_read
,
1707 if (flags
& TTU_RMAP_LOCKED
)
1708 rmap_walk_locked(page
, &rwc
);
1710 rmap_walk(page
, &rwc
);
1714 * @arg: enum ttu_flags will be passed to this argument.
1716 * If TTU_SPLIT_HUGE_PMD is specified any PMD mappings will be split into PTEs
1717 * containing migration entries.
1719 static bool try_to_migrate_one(struct page
*page
, struct vm_area_struct
*vma
,
1720 unsigned long address
, void *arg
)
1722 struct mm_struct
*mm
= vma
->vm_mm
;
1723 struct page_vma_mapped_walk pvmw
= {
1729 struct page
*subpage
;
1731 struct mmu_notifier_range range
;
1732 enum ttu_flags flags
= (enum ttu_flags
)(long)arg
;
1735 * When racing against e.g. zap_pte_range() on another cpu,
1736 * in between its ptep_get_and_clear_full() and page_remove_rmap(),
1737 * try_to_migrate() may return before page_mapped() has become false,
1738 * if page table locking is skipped: use TTU_SYNC to wait for that.
1740 if (flags
& TTU_SYNC
)
1741 pvmw
.flags
= PVMW_SYNC
;
1744 * unmap_page() in mm/huge_memory.c is the only user of migration with
1745 * TTU_SPLIT_HUGE_PMD and it wants to freeze.
1747 if (flags
& TTU_SPLIT_HUGE_PMD
)
1748 split_huge_pmd_address(vma
, address
, true, page
);
1751 * For THP, we have to assume the worse case ie pmd for invalidation.
1752 * For hugetlb, it could be much worse if we need to do pud
1753 * invalidation in the case of pmd sharing.
1755 * Note that the page can not be free in this function as call of
1756 * try_to_unmap() must hold a reference on the page.
1758 range
.end
= PageKsm(page
) ?
1759 address
+ PAGE_SIZE
: vma_address_end(page
, vma
);
1760 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
, vma
->vm_mm
,
1761 address
, range
.end
);
1762 if (PageHuge(page
)) {
1764 * If sharing is possible, start and end will be adjusted
1767 adjust_range_if_pmd_sharing_possible(vma
, &range
.start
,
1770 mmu_notifier_invalidate_range_start(&range
);
1772 while (page_vma_mapped_walk(&pvmw
)) {
1773 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1774 /* PMD-mapped THP migration entry */
1776 VM_BUG_ON_PAGE(PageHuge(page
) ||
1777 !PageTransCompound(page
), page
);
1779 set_pmd_migration_entry(&pvmw
, page
);
1784 /* Unexpected PMD-mapped THP? */
1785 VM_BUG_ON_PAGE(!pvmw
.pte
, page
);
1787 subpage
= page
- page_to_pfn(page
) + pte_pfn(*pvmw
.pte
);
1788 address
= pvmw
.address
;
1790 if (PageHuge(page
) && !PageAnon(page
)) {
1792 * To call huge_pmd_unshare, i_mmap_rwsem must be
1793 * held in write mode. Caller needs to explicitly
1794 * do this outside rmap routines.
1796 VM_BUG_ON(!(flags
& TTU_RMAP_LOCKED
));
1797 if (huge_pmd_unshare(mm
, vma
, &address
, pvmw
.pte
)) {
1799 * huge_pmd_unshare unmapped an entire PMD
1800 * page. There is no way of knowing exactly
1801 * which PMDs may be cached for this mm, so
1802 * we must flush them all. start/end were
1803 * already adjusted above to cover this range.
1805 flush_cache_range(vma
, range
.start
, range
.end
);
1806 flush_tlb_range(vma
, range
.start
, range
.end
);
1807 mmu_notifier_invalidate_range(mm
, range
.start
,
1811 * The ref count of the PMD page was dropped
1812 * which is part of the way map counting
1813 * is done for shared PMDs. Return 'true'
1814 * here. When there is no other sharing,
1815 * huge_pmd_unshare returns false and we will
1816 * unmap the actual page and drop map count
1819 page_vma_mapped_walk_done(&pvmw
);
1824 /* Nuke the page table entry. */
1825 flush_cache_page(vma
, address
, pte_pfn(*pvmw
.pte
));
1826 pteval
= ptep_clear_flush(vma
, address
, pvmw
.pte
);
1828 /* Move the dirty bit to the page. Now the pte is gone. */
1829 if (pte_dirty(pteval
))
1830 set_page_dirty(page
);
1832 /* Update high watermark before we lower rss */
1833 update_hiwater_rss(mm
);
1835 if (is_zone_device_page(page
)) {
1840 * Store the pfn of the page in a special migration
1841 * pte. do_swap_page() will wait until the migration
1842 * pte is removed and then restart fault handling.
1844 entry
= make_readable_migration_entry(
1846 swp_pte
= swp_entry_to_pte(entry
);
1849 * pteval maps a zone device page and is therefore
1852 if (pte_swp_soft_dirty(pteval
))
1853 swp_pte
= pte_swp_mksoft_dirty(swp_pte
);
1854 if (pte_swp_uffd_wp(pteval
))
1855 swp_pte
= pte_swp_mkuffd_wp(swp_pte
);
1856 set_pte_at(mm
, pvmw
.address
, pvmw
.pte
, swp_pte
);
1858 * No need to invalidate here it will synchronize on
1859 * against the special swap migration pte.
1861 * The assignment to subpage above was computed from a
1862 * swap PTE which results in an invalid pointer.
1863 * Since only PAGE_SIZE pages can currently be
1864 * migrated, just set it to page. This will need to be
1865 * changed when hugepage migrations to device private
1866 * memory are supported.
1869 } else if (PageHWPoison(page
)) {
1870 pteval
= swp_entry_to_pte(make_hwpoison_entry(subpage
));
1871 if (PageHuge(page
)) {
1872 hugetlb_count_sub(compound_nr(page
), mm
);
1873 set_huge_swap_pte_at(mm
, address
,
1875 vma_mmu_pagesize(vma
));
1877 dec_mm_counter(mm
, mm_counter(page
));
1878 set_pte_at(mm
, address
, pvmw
.pte
, pteval
);
1881 } else if (pte_unused(pteval
) && !userfaultfd_armed(vma
)) {
1883 * The guest indicated that the page content is of no
1884 * interest anymore. Simply discard the pte, vmscan
1885 * will take care of the rest.
1886 * A future reference will then fault in a new zero
1887 * page. When userfaultfd is active, we must not drop
1888 * this page though, as its main user (postcopy
1889 * migration) will not expect userfaults on already
1892 dec_mm_counter(mm
, mm_counter(page
));
1893 /* We have to invalidate as we cleared the pte */
1894 mmu_notifier_invalidate_range(mm
, address
,
1895 address
+ PAGE_SIZE
);
1900 if (arch_unmap_one(mm
, vma
, address
, pteval
) < 0) {
1901 set_pte_at(mm
, address
, pvmw
.pte
, pteval
);
1903 page_vma_mapped_walk_done(&pvmw
);
1908 * Store the pfn of the page in a special migration
1909 * pte. do_swap_page() will wait until the migration
1910 * pte is removed and then restart fault handling.
1912 if (pte_write(pteval
))
1913 entry
= make_writable_migration_entry(
1914 page_to_pfn(subpage
));
1916 entry
= make_readable_migration_entry(
1917 page_to_pfn(subpage
));
1919 swp_pte
= swp_entry_to_pte(entry
);
1920 if (pte_soft_dirty(pteval
))
1921 swp_pte
= pte_swp_mksoft_dirty(swp_pte
);
1922 if (pte_uffd_wp(pteval
))
1923 swp_pte
= pte_swp_mkuffd_wp(swp_pte
);
1924 set_pte_at(mm
, address
, pvmw
.pte
, swp_pte
);
1926 * No need to invalidate here it will synchronize on
1927 * against the special swap migration pte.
1932 * No need to call mmu_notifier_invalidate_range() it has be
1933 * done above for all cases requiring it to happen under page
1934 * table lock before mmu_notifier_invalidate_range_end()
1936 * See Documentation/vm/mmu_notifier.rst
1938 page_remove_rmap(subpage
, PageHuge(page
));
1942 mmu_notifier_invalidate_range_end(&range
);
1948 * try_to_migrate - try to replace all page table mappings with swap entries
1949 * @page: the page to replace page table entries for
1950 * @flags: action and flags
1952 * Tries to remove all the page table entries which are mapping this page and
1953 * replace them with special swap entries. Caller must hold the page lock.
1955 void try_to_migrate(struct page
*page
, enum ttu_flags flags
)
1957 struct rmap_walk_control rwc
= {
1958 .rmap_one
= try_to_migrate_one
,
1959 .arg
= (void *)flags
,
1960 .done
= page_not_mapped
,
1961 .anon_lock
= page_lock_anon_vma_read
,
1965 * Migration always ignores mlock and only supports TTU_RMAP_LOCKED and
1966 * TTU_SPLIT_HUGE_PMD and TTU_SYNC flags.
1968 if (WARN_ON_ONCE(flags
& ~(TTU_RMAP_LOCKED
| TTU_SPLIT_HUGE_PMD
|
1972 if (is_zone_device_page(page
) && !is_device_private_page(page
))
1976 * During exec, a temporary VMA is setup and later moved.
1977 * The VMA is moved under the anon_vma lock but not the
1978 * page tables leading to a race where migration cannot
1979 * find the migration ptes. Rather than increasing the
1980 * locking requirements of exec(), migration skips
1981 * temporary VMAs until after exec() completes.
1983 if (!PageKsm(page
) && PageAnon(page
))
1984 rwc
.invalid_vma
= invalid_migration_vma
;
1986 if (flags
& TTU_RMAP_LOCKED
)
1987 rmap_walk_locked(page
, &rwc
);
1989 rmap_walk(page
, &rwc
);
1993 * Walks the vma's mapping a page and mlocks the page if any locked vma's are
1994 * found. Once one is found the page is locked and the scan can be terminated.
1996 static bool page_mlock_one(struct page
*page
, struct vm_area_struct
*vma
,
1997 unsigned long address
, void *unused
)
1999 struct page_vma_mapped_walk pvmw
= {
2005 /* An un-locked vma doesn't have any pages to lock, continue the scan */
2006 if (!(vma
->vm_flags
& VM_LOCKED
))
2009 while (page_vma_mapped_walk(&pvmw
)) {
2011 * Need to recheck under the ptl to serialise with
2012 * __munlock_pagevec_fill() after VM_LOCKED is cleared in
2013 * munlock_vma_pages_range().
2015 if (vma
->vm_flags
& VM_LOCKED
) {
2017 * PTE-mapped THP are never marked as mlocked; but
2018 * this function is never called on a DoubleMap THP,
2019 * nor on an Anon THP (which may still be PTE-mapped
2020 * after DoubleMap was cleared).
2022 mlock_vma_page(page
);
2024 * No need to scan further once the page is marked
2027 page_vma_mapped_walk_done(&pvmw
);
2036 * page_mlock - try to mlock a page
2037 * @page: the page to be mlocked
2039 * Called from munlock code. Checks all of the VMAs mapping the page and mlocks
2040 * the page if any are found. The page will be returned with PG_mlocked cleared
2041 * if it is not mapped by any locked vmas.
2043 void page_mlock(struct page
*page
)
2045 struct rmap_walk_control rwc
= {
2046 .rmap_one
= page_mlock_one
,
2047 .done
= page_not_mapped
,
2048 .anon_lock
= page_lock_anon_vma_read
,
2052 VM_BUG_ON_PAGE(!PageLocked(page
) || PageLRU(page
), page
);
2053 VM_BUG_ON_PAGE(PageCompound(page
) && PageDoubleMap(page
), page
);
2055 /* Anon THP are only marked as mlocked when singly mapped */
2056 if (PageTransCompound(page
) && PageAnon(page
))
2059 rmap_walk(page
, &rwc
);
2062 #ifdef CONFIG_DEVICE_PRIVATE
2063 struct make_exclusive_args
{
2064 struct mm_struct
*mm
;
2065 unsigned long address
;
2070 static bool page_make_device_exclusive_one(struct page
*page
,
2071 struct vm_area_struct
*vma
, unsigned long address
, void *priv
)
2073 struct mm_struct
*mm
= vma
->vm_mm
;
2074 struct page_vma_mapped_walk pvmw
= {
2079 struct make_exclusive_args
*args
= priv
;
2081 struct page
*subpage
;
2083 struct mmu_notifier_range range
;
2087 mmu_notifier_range_init_owner(&range
, MMU_NOTIFY_EXCLUSIVE
, 0, vma
,
2088 vma
->vm_mm
, address
, min(vma
->vm_end
,
2089 address
+ page_size(page
)), args
->owner
);
2090 mmu_notifier_invalidate_range_start(&range
);
2092 while (page_vma_mapped_walk(&pvmw
)) {
2093 /* Unexpected PMD-mapped THP? */
2094 VM_BUG_ON_PAGE(!pvmw
.pte
, page
);
2096 if (!pte_present(*pvmw
.pte
)) {
2098 page_vma_mapped_walk_done(&pvmw
);
2102 subpage
= page
- page_to_pfn(page
) + pte_pfn(*pvmw
.pte
);
2103 address
= pvmw
.address
;
2105 /* Nuke the page table entry. */
2106 flush_cache_page(vma
, address
, pte_pfn(*pvmw
.pte
));
2107 pteval
= ptep_clear_flush(vma
, address
, pvmw
.pte
);
2109 /* Move the dirty bit to the page. Now the pte is gone. */
2110 if (pte_dirty(pteval
))
2111 set_page_dirty(page
);
2114 * Check that our target page is still mapped at the expected
2117 if (args
->mm
== mm
&& args
->address
== address
&&
2122 * Store the pfn of the page in a special migration
2123 * pte. do_swap_page() will wait until the migration
2124 * pte is removed and then restart fault handling.
2126 if (pte_write(pteval
))
2127 entry
= make_writable_device_exclusive_entry(
2128 page_to_pfn(subpage
));
2130 entry
= make_readable_device_exclusive_entry(
2131 page_to_pfn(subpage
));
2132 swp_pte
= swp_entry_to_pte(entry
);
2133 if (pte_soft_dirty(pteval
))
2134 swp_pte
= pte_swp_mksoft_dirty(swp_pte
);
2135 if (pte_uffd_wp(pteval
))
2136 swp_pte
= pte_swp_mkuffd_wp(swp_pte
);
2138 set_pte_at(mm
, address
, pvmw
.pte
, swp_pte
);
2141 * There is a reference on the page for the swap entry which has
2142 * been removed, so shouldn't take another.
2144 page_remove_rmap(subpage
, false);
2147 mmu_notifier_invalidate_range_end(&range
);
2153 * page_make_device_exclusive - mark the page exclusively owned by a device
2154 * @page: the page to replace page table entries for
2155 * @mm: the mm_struct where the page is expected to be mapped
2156 * @address: address where the page is expected to be mapped
2157 * @owner: passed to MMU_NOTIFY_EXCLUSIVE range notifier callbacks
2159 * Tries to remove all the page table entries which are mapping this page and
2160 * replace them with special device exclusive swap entries to grant a device
2161 * exclusive access to the page. Caller must hold the page lock.
2163 * Returns false if the page is still mapped, or if it could not be unmapped
2164 * from the expected address. Otherwise returns true (success).
2166 static bool page_make_device_exclusive(struct page
*page
, struct mm_struct
*mm
,
2167 unsigned long address
, void *owner
)
2169 struct make_exclusive_args args
= {
2175 struct rmap_walk_control rwc
= {
2176 .rmap_one
= page_make_device_exclusive_one
,
2177 .done
= page_not_mapped
,
2178 .anon_lock
= page_lock_anon_vma_read
,
2183 * Restrict to anonymous pages for now to avoid potential writeback
2184 * issues. Also tail pages shouldn't be passed to rmap_walk so skip
2187 if (!PageAnon(page
) || PageTail(page
))
2190 rmap_walk(page
, &rwc
);
2192 return args
.valid
&& !page_mapcount(page
);
2196 * make_device_exclusive_range() - Mark a range for exclusive use by a device
2197 * @mm: mm_struct of assoicated target process
2198 * @start: start of the region to mark for exclusive device access
2199 * @end: end address of region
2200 * @pages: returns the pages which were successfully marked for exclusive access
2201 * @owner: passed to MMU_NOTIFY_EXCLUSIVE range notifier to allow filtering
2203 * Returns: number of pages found in the range by GUP. A page is marked for
2204 * exclusive access only if the page pointer is non-NULL.
2206 * This function finds ptes mapping page(s) to the given address range, locks
2207 * them and replaces mappings with special swap entries preventing userspace CPU
2208 * access. On fault these entries are replaced with the original mapping after
2209 * calling MMU notifiers.
2211 * A driver using this to program access from a device must use a mmu notifier
2212 * critical section to hold a device specific lock during programming. Once
2213 * programming is complete it should drop the page lock and reference after
2214 * which point CPU access to the page will revoke the exclusive access.
2216 int make_device_exclusive_range(struct mm_struct
*mm
, unsigned long start
,
2217 unsigned long end
, struct page
**pages
,
2220 long npages
= (end
- start
) >> PAGE_SHIFT
;
2223 npages
= get_user_pages_remote(mm
, start
, npages
,
2224 FOLL_GET
| FOLL_WRITE
| FOLL_SPLIT_PMD
,
2229 for (i
= 0; i
< npages
; i
++, start
+= PAGE_SIZE
) {
2230 if (!trylock_page(pages
[i
])) {
2236 if (!page_make_device_exclusive(pages
[i
], mm
, start
, owner
)) {
2237 unlock_page(pages
[i
]);
2245 EXPORT_SYMBOL_GPL(make_device_exclusive_range
);
2248 void __put_anon_vma(struct anon_vma
*anon_vma
)
2250 struct anon_vma
*root
= anon_vma
->root
;
2252 anon_vma_free(anon_vma
);
2253 if (root
!= anon_vma
&& atomic_dec_and_test(&root
->refcount
))
2254 anon_vma_free(root
);
2257 static struct anon_vma
*rmap_walk_anon_lock(struct page
*page
,
2258 struct rmap_walk_control
*rwc
)
2260 struct anon_vma
*anon_vma
;
2263 return rwc
->anon_lock(page
);
2266 * Note: remove_migration_ptes() cannot use page_lock_anon_vma_read()
2267 * because that depends on page_mapped(); but not all its usages
2268 * are holding mmap_lock. Users without mmap_lock are required to
2269 * take a reference count to prevent the anon_vma disappearing
2271 anon_vma
= page_anon_vma(page
);
2275 anon_vma_lock_read(anon_vma
);
2280 * rmap_walk_anon - do something to anonymous page using the object-based
2282 * @page: the page to be handled
2283 * @rwc: control variable according to each walk type
2285 * Find all the mappings of a page using the mapping pointer and the vma chains
2286 * contained in the anon_vma struct it points to.
2288 * When called from page_mlock(), the mmap_lock of the mm containing the vma
2289 * where the page was found will be held for write. So, we won't recheck
2290 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
2293 static void rmap_walk_anon(struct page
*page
, struct rmap_walk_control
*rwc
,
2296 struct anon_vma
*anon_vma
;
2297 pgoff_t pgoff_start
, pgoff_end
;
2298 struct anon_vma_chain
*avc
;
2301 anon_vma
= page_anon_vma(page
);
2302 /* anon_vma disappear under us? */
2303 VM_BUG_ON_PAGE(!anon_vma
, page
);
2305 anon_vma
= rmap_walk_anon_lock(page
, rwc
);
2310 pgoff_start
= page_to_pgoff(page
);
2311 pgoff_end
= pgoff_start
+ thp_nr_pages(page
) - 1;
2312 anon_vma_interval_tree_foreach(avc
, &anon_vma
->rb_root
,
2313 pgoff_start
, pgoff_end
) {
2314 struct vm_area_struct
*vma
= avc
->vma
;
2315 unsigned long address
= vma_address(page
, vma
);
2317 VM_BUG_ON_VMA(address
== -EFAULT
, vma
);
2320 if (rwc
->invalid_vma
&& rwc
->invalid_vma(vma
, rwc
->arg
))
2323 if (!rwc
->rmap_one(page
, vma
, address
, rwc
->arg
))
2325 if (rwc
->done
&& rwc
->done(page
))
2330 anon_vma_unlock_read(anon_vma
);
2334 * rmap_walk_file - do something to file page using the object-based rmap method
2335 * @page: the page to be handled
2336 * @rwc: control variable according to each walk type
2338 * Find all the mappings of a page using the mapping pointer and the vma chains
2339 * contained in the address_space struct it points to.
2341 * When called from page_mlock(), the mmap_lock of the mm containing the vma
2342 * where the page was found will be held for write. So, we won't recheck
2343 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
2346 static void rmap_walk_file(struct page
*page
, struct rmap_walk_control
*rwc
,
2349 struct address_space
*mapping
= page_mapping(page
);
2350 pgoff_t pgoff_start
, pgoff_end
;
2351 struct vm_area_struct
*vma
;
2354 * The page lock not only makes sure that page->mapping cannot
2355 * suddenly be NULLified by truncation, it makes sure that the
2356 * structure at mapping cannot be freed and reused yet,
2357 * so we can safely take mapping->i_mmap_rwsem.
2359 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
2364 pgoff_start
= page_to_pgoff(page
);
2365 pgoff_end
= pgoff_start
+ thp_nr_pages(page
) - 1;
2367 i_mmap_lock_read(mapping
);
2368 vma_interval_tree_foreach(vma
, &mapping
->i_mmap
,
2369 pgoff_start
, pgoff_end
) {
2370 unsigned long address
= vma_address(page
, vma
);
2372 VM_BUG_ON_VMA(address
== -EFAULT
, vma
);
2375 if (rwc
->invalid_vma
&& rwc
->invalid_vma(vma
, rwc
->arg
))
2378 if (!rwc
->rmap_one(page
, vma
, address
, rwc
->arg
))
2380 if (rwc
->done
&& rwc
->done(page
))
2386 i_mmap_unlock_read(mapping
);
2389 void rmap_walk(struct page
*page
, struct rmap_walk_control
*rwc
)
2391 if (unlikely(PageKsm(page
)))
2392 rmap_walk_ksm(page
, rwc
);
2393 else if (PageAnon(page
))
2394 rmap_walk_anon(page
, rwc
, false);
2396 rmap_walk_file(page
, rwc
, false);
2399 /* Like rmap_walk, but caller holds relevant rmap lock */
2400 void rmap_walk_locked(struct page
*page
, struct rmap_walk_control
*rwc
)
2402 /* no ksm support for now */
2403 VM_BUG_ON_PAGE(PageKsm(page
), page
);
2405 rmap_walk_anon(page
, rwc
, true);
2407 rmap_walk_file(page
, rwc
, true);
2410 #ifdef CONFIG_HUGETLB_PAGE
2412 * The following two functions are for anonymous (private mapped) hugepages.
2413 * Unlike common anonymous pages, anonymous hugepages have no accounting code
2414 * and no lru code, because we handle hugepages differently from common pages.
2416 void hugepage_add_anon_rmap(struct page
*page
,
2417 struct vm_area_struct
*vma
, unsigned long address
)
2419 struct anon_vma
*anon_vma
= vma
->anon_vma
;
2422 BUG_ON(!PageLocked(page
));
2424 /* address might be in next vma when migration races vma_adjust */
2425 first
= atomic_inc_and_test(compound_mapcount_ptr(page
));
2427 __page_set_anon_rmap(page
, vma
, address
, 0);
2430 void hugepage_add_new_anon_rmap(struct page
*page
,
2431 struct vm_area_struct
*vma
, unsigned long address
)
2433 BUG_ON(address
< vma
->vm_start
|| address
>= vma
->vm_end
);
2434 atomic_set(compound_mapcount_ptr(page
), 0);
2435 if (hpage_pincount_available(page
))
2436 atomic_set(compound_pincount_ptr(page
), 0);
2438 __page_set_anon_rmap(page
, vma
, address
, 1);
2440 #endif /* CONFIG_HUGETLB_PAGE */