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) * (see huegtlbfs below)
26 * hugetlbfs_i_mmap_rwsem_key (in huge_pmd_share)
27 * mapping->i_mmap_rwsem
28 * hugetlb_fault_mutex (hugetlbfs specific page fault mutex)
30 * mm->page_table_lock or pte_lock
31 * swap_lock (in swap_duplicate, swap_info_get)
32 * mmlist_lock (in mmput, drain_mmlist and others)
33 * mapping->private_lock (in __set_page_dirty_buffers)
34 * lock_page_memcg move_lock (in __set_page_dirty_buffers)
35 * i_pages lock (widely used)
36 * lruvec->lru_lock (in lock_page_lruvec_irq)
37 * inode->i_lock (in set_page_dirty's __mark_inode_dirty)
38 * bdi.wb->list_lock (in set_page_dirty's __mark_inode_dirty)
39 * sb_lock (within inode_lock in fs/fs-writeback.c)
40 * i_pages lock (widely used, in set_page_dirty,
41 * in arch-dependent flush_dcache_mmap_lock,
42 * within bdi.wb->list_lock in __sync_single_inode)
44 * anon_vma->rwsem,mapping->i_mutex (memory_failure, collect_procs_anon)
48 * * hugetlbfs PageHuge() pages take locks in this order:
49 * mapping->i_mmap_rwsem
50 * hugetlb_fault_mutex (hugetlbfs specific page fault mutex)
51 * page->flags PG_locked (lock_page)
55 #include <linux/sched/mm.h>
56 #include <linux/sched/task.h>
57 #include <linux/pagemap.h>
58 #include <linux/swap.h>
59 #include <linux/swapops.h>
60 #include <linux/slab.h>
61 #include <linux/init.h>
62 #include <linux/ksm.h>
63 #include <linux/rmap.h>
64 #include <linux/rcupdate.h>
65 #include <linux/export.h>
66 #include <linux/memcontrol.h>
67 #include <linux/mmu_notifier.h>
68 #include <linux/migrate.h>
69 #include <linux/hugetlb.h>
70 #include <linux/huge_mm.h>
71 #include <linux/backing-dev.h>
72 #include <linux/page_idle.h>
73 #include <linux/memremap.h>
74 #include <linux/userfaultfd_k.h>
76 #include <asm/tlbflush.h>
78 #include <trace/events/tlb.h>
82 static struct kmem_cache
*anon_vma_cachep
;
83 static struct kmem_cache
*anon_vma_chain_cachep
;
85 static inline struct anon_vma
*anon_vma_alloc(void)
87 struct anon_vma
*anon_vma
;
89 anon_vma
= kmem_cache_alloc(anon_vma_cachep
, GFP_KERNEL
);
91 atomic_set(&anon_vma
->refcount
, 1);
92 anon_vma
->degree
= 1; /* Reference for first vma */
93 anon_vma
->parent
= anon_vma
;
95 * Initialise the anon_vma root to point to itself. If called
96 * from fork, the root will be reset to the parents anon_vma.
98 anon_vma
->root
= anon_vma
;
104 static inline void anon_vma_free(struct anon_vma
*anon_vma
)
106 VM_BUG_ON(atomic_read(&anon_vma
->refcount
));
109 * Synchronize against page_lock_anon_vma_read() such that
110 * we can safely hold the lock without the anon_vma getting
113 * Relies on the full mb implied by the atomic_dec_and_test() from
114 * put_anon_vma() against the acquire barrier implied by
115 * down_read_trylock() from page_lock_anon_vma_read(). This orders:
117 * page_lock_anon_vma_read() VS put_anon_vma()
118 * down_read_trylock() atomic_dec_and_test()
120 * atomic_read() rwsem_is_locked()
122 * LOCK should suffice since the actual taking of the lock must
123 * happen _before_ what follows.
126 if (rwsem_is_locked(&anon_vma
->root
->rwsem
)) {
127 anon_vma_lock_write(anon_vma
);
128 anon_vma_unlock_write(anon_vma
);
131 kmem_cache_free(anon_vma_cachep
, anon_vma
);
134 static inline struct anon_vma_chain
*anon_vma_chain_alloc(gfp_t gfp
)
136 return kmem_cache_alloc(anon_vma_chain_cachep
, gfp
);
139 static void anon_vma_chain_free(struct anon_vma_chain
*anon_vma_chain
)
141 kmem_cache_free(anon_vma_chain_cachep
, anon_vma_chain
);
144 static void anon_vma_chain_link(struct vm_area_struct
*vma
,
145 struct anon_vma_chain
*avc
,
146 struct anon_vma
*anon_vma
)
149 avc
->anon_vma
= anon_vma
;
150 list_add(&avc
->same_vma
, &vma
->anon_vma_chain
);
151 anon_vma_interval_tree_insert(avc
, &anon_vma
->rb_root
);
155 * __anon_vma_prepare - attach an anon_vma to a memory region
156 * @vma: the memory region in question
158 * This makes sure the memory mapping described by 'vma' has
159 * an 'anon_vma' attached to it, so that we can associate the
160 * anonymous pages mapped into it with that anon_vma.
162 * The common case will be that we already have one, which
163 * is handled inline by anon_vma_prepare(). But if
164 * not we either need to find an adjacent mapping that we
165 * can re-use the anon_vma from (very common when the only
166 * reason for splitting a vma has been mprotect()), or we
167 * allocate a new one.
169 * Anon-vma allocations are very subtle, because we may have
170 * optimistically looked up an anon_vma in page_lock_anon_vma_read()
171 * and that may actually touch the rwsem even in the newly
172 * allocated vma (it depends on RCU to make sure that the
173 * anon_vma isn't actually destroyed).
175 * As a result, we need to do proper anon_vma locking even
176 * for the new allocation. At the same time, we do not want
177 * to do any locking for the common case of already having
180 * This must be called with the mmap_lock held for reading.
182 int __anon_vma_prepare(struct vm_area_struct
*vma
)
184 struct mm_struct
*mm
= vma
->vm_mm
;
185 struct anon_vma
*anon_vma
, *allocated
;
186 struct anon_vma_chain
*avc
;
190 avc
= anon_vma_chain_alloc(GFP_KERNEL
);
194 anon_vma
= find_mergeable_anon_vma(vma
);
197 anon_vma
= anon_vma_alloc();
198 if (unlikely(!anon_vma
))
199 goto out_enomem_free_avc
;
200 allocated
= anon_vma
;
203 anon_vma_lock_write(anon_vma
);
204 /* page_table_lock to protect against threads */
205 spin_lock(&mm
->page_table_lock
);
206 if (likely(!vma
->anon_vma
)) {
207 vma
->anon_vma
= anon_vma
;
208 anon_vma_chain_link(vma
, avc
, anon_vma
);
209 /* vma reference or self-parent link for new root */
214 spin_unlock(&mm
->page_table_lock
);
215 anon_vma_unlock_write(anon_vma
);
217 if (unlikely(allocated
))
218 put_anon_vma(allocated
);
220 anon_vma_chain_free(avc
);
225 anon_vma_chain_free(avc
);
231 * This is a useful helper function for locking the anon_vma root as
232 * we traverse the vma->anon_vma_chain, looping over anon_vma's that
235 * Such anon_vma's should have the same root, so you'd expect to see
236 * just a single mutex_lock for the whole traversal.
238 static inline struct anon_vma
*lock_anon_vma_root(struct anon_vma
*root
, struct anon_vma
*anon_vma
)
240 struct anon_vma
*new_root
= anon_vma
->root
;
241 if (new_root
!= root
) {
242 if (WARN_ON_ONCE(root
))
243 up_write(&root
->rwsem
);
245 down_write(&root
->rwsem
);
250 static inline void unlock_anon_vma_root(struct anon_vma
*root
)
253 up_write(&root
->rwsem
);
257 * Attach the anon_vmas from src to dst.
258 * Returns 0 on success, -ENOMEM on failure.
260 * anon_vma_clone() is called by __vma_split(), __split_vma(), copy_vma() and
261 * anon_vma_fork(). The first three want an exact copy of src, while the last
262 * one, anon_vma_fork(), may try to reuse an existing anon_vma to prevent
263 * endless growth of anon_vma. Since dst->anon_vma is set to NULL before call,
264 * we can identify this case by checking (!dst->anon_vma && src->anon_vma).
266 * If (!dst->anon_vma && src->anon_vma) is true, this function tries to find
267 * and reuse existing anon_vma which has no vmas and only one child anon_vma.
268 * This prevents degradation of anon_vma hierarchy to endless linear chain in
269 * case of constantly forking task. On the other hand, an anon_vma with more
270 * than one child isn't reused even if there was no alive vma, thus rmap
271 * walker has a good chance of avoiding scanning the whole hierarchy when it
272 * searches where page is mapped.
274 int anon_vma_clone(struct vm_area_struct
*dst
, struct vm_area_struct
*src
)
276 struct anon_vma_chain
*avc
, *pavc
;
277 struct anon_vma
*root
= NULL
;
279 list_for_each_entry_reverse(pavc
, &src
->anon_vma_chain
, same_vma
) {
280 struct anon_vma
*anon_vma
;
282 avc
= anon_vma_chain_alloc(GFP_NOWAIT
| __GFP_NOWARN
);
283 if (unlikely(!avc
)) {
284 unlock_anon_vma_root(root
);
286 avc
= anon_vma_chain_alloc(GFP_KERNEL
);
290 anon_vma
= pavc
->anon_vma
;
291 root
= lock_anon_vma_root(root
, anon_vma
);
292 anon_vma_chain_link(dst
, avc
, anon_vma
);
295 * Reuse existing anon_vma if its degree lower than two,
296 * that means it has no vma and only one anon_vma child.
298 * Do not chose parent anon_vma, otherwise first child
299 * will always reuse it. Root anon_vma is never reused:
300 * it has self-parent reference and at least one child.
302 if (!dst
->anon_vma
&& src
->anon_vma
&&
303 anon_vma
!= src
->anon_vma
&& anon_vma
->degree
< 2)
304 dst
->anon_vma
= anon_vma
;
307 dst
->anon_vma
->degree
++;
308 unlock_anon_vma_root(root
);
313 * dst->anon_vma is dropped here otherwise its degree can be incorrectly
314 * decremented in unlink_anon_vmas().
315 * We can safely do this because callers of anon_vma_clone() don't care
316 * about dst->anon_vma if anon_vma_clone() failed.
318 dst
->anon_vma
= NULL
;
319 unlink_anon_vmas(dst
);
324 * Attach vma to its own anon_vma, as well as to the anon_vmas that
325 * the corresponding VMA in the parent process is attached to.
326 * Returns 0 on success, non-zero on failure.
328 int anon_vma_fork(struct vm_area_struct
*vma
, struct vm_area_struct
*pvma
)
330 struct anon_vma_chain
*avc
;
331 struct anon_vma
*anon_vma
;
334 /* Don't bother if the parent process has no anon_vma here. */
338 /* Drop inherited anon_vma, we'll reuse existing or allocate new. */
339 vma
->anon_vma
= NULL
;
342 * First, attach the new VMA to the parent VMA's anon_vmas,
343 * so rmap can find non-COWed pages in child processes.
345 error
= anon_vma_clone(vma
, pvma
);
349 /* An existing anon_vma has been reused, all done then. */
353 /* Then add our own anon_vma. */
354 anon_vma
= anon_vma_alloc();
357 avc
= anon_vma_chain_alloc(GFP_KERNEL
);
359 goto out_error_free_anon_vma
;
362 * The root anon_vma's rwsem is the lock actually used when we
363 * lock any of the anon_vmas in this anon_vma tree.
365 anon_vma
->root
= pvma
->anon_vma
->root
;
366 anon_vma
->parent
= pvma
->anon_vma
;
368 * With refcounts, an anon_vma can stay around longer than the
369 * process it belongs to. The root anon_vma needs to be pinned until
370 * this anon_vma is freed, because the lock lives in the root.
372 get_anon_vma(anon_vma
->root
);
373 /* Mark this anon_vma as the one where our new (COWed) pages go. */
374 vma
->anon_vma
= anon_vma
;
375 anon_vma_lock_write(anon_vma
);
376 anon_vma_chain_link(vma
, avc
, anon_vma
);
377 anon_vma
->parent
->degree
++;
378 anon_vma_unlock_write(anon_vma
);
382 out_error_free_anon_vma
:
383 put_anon_vma(anon_vma
);
385 unlink_anon_vmas(vma
);
389 void unlink_anon_vmas(struct vm_area_struct
*vma
)
391 struct anon_vma_chain
*avc
, *next
;
392 struct anon_vma
*root
= NULL
;
395 * Unlink each anon_vma chained to the VMA. This list is ordered
396 * from newest to oldest, ensuring the root anon_vma gets freed last.
398 list_for_each_entry_safe(avc
, next
, &vma
->anon_vma_chain
, same_vma
) {
399 struct anon_vma
*anon_vma
= avc
->anon_vma
;
401 root
= lock_anon_vma_root(root
, anon_vma
);
402 anon_vma_interval_tree_remove(avc
, &anon_vma
->rb_root
);
405 * Leave empty anon_vmas on the list - we'll need
406 * to free them outside the lock.
408 if (RB_EMPTY_ROOT(&anon_vma
->rb_root
.rb_root
)) {
409 anon_vma
->parent
->degree
--;
413 list_del(&avc
->same_vma
);
414 anon_vma_chain_free(avc
);
417 vma
->anon_vma
->degree
--;
420 * vma would still be needed after unlink, and anon_vma will be prepared
423 vma
->anon_vma
= NULL
;
425 unlock_anon_vma_root(root
);
428 * Iterate the list once more, it now only contains empty and unlinked
429 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
430 * needing to write-acquire the anon_vma->root->rwsem.
432 list_for_each_entry_safe(avc
, next
, &vma
->anon_vma_chain
, same_vma
) {
433 struct anon_vma
*anon_vma
= avc
->anon_vma
;
435 VM_WARN_ON(anon_vma
->degree
);
436 put_anon_vma(anon_vma
);
438 list_del(&avc
->same_vma
);
439 anon_vma_chain_free(avc
);
443 static void anon_vma_ctor(void *data
)
445 struct anon_vma
*anon_vma
= data
;
447 init_rwsem(&anon_vma
->rwsem
);
448 atomic_set(&anon_vma
->refcount
, 0);
449 anon_vma
->rb_root
= RB_ROOT_CACHED
;
452 void __init
anon_vma_init(void)
454 anon_vma_cachep
= kmem_cache_create("anon_vma", sizeof(struct anon_vma
),
455 0, SLAB_TYPESAFE_BY_RCU
|SLAB_PANIC
|SLAB_ACCOUNT
,
457 anon_vma_chain_cachep
= KMEM_CACHE(anon_vma_chain
,
458 SLAB_PANIC
|SLAB_ACCOUNT
);
462 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
464 * Since there is no serialization what so ever against page_remove_rmap()
465 * the best this function can do is return a refcount increased anon_vma
466 * that might have been relevant to this page.
468 * The page might have been remapped to a different anon_vma or the anon_vma
469 * returned may already be freed (and even reused).
471 * In case it was remapped to a different anon_vma, the new anon_vma will be a
472 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
473 * ensure that any anon_vma obtained from the page will still be valid for as
474 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
476 * All users of this function must be very careful when walking the anon_vma
477 * chain and verify that the page in question is indeed mapped in it
478 * [ something equivalent to page_mapped_in_vma() ].
480 * Since anon_vma's slab is SLAB_TYPESAFE_BY_RCU and we know from
481 * page_remove_rmap() that the anon_vma pointer from page->mapping is valid
482 * if there is a mapcount, we can dereference the anon_vma after observing
485 struct anon_vma
*page_get_anon_vma(struct page
*page
)
487 struct anon_vma
*anon_vma
= NULL
;
488 unsigned long anon_mapping
;
491 anon_mapping
= (unsigned long)READ_ONCE(page
->mapping
);
492 if ((anon_mapping
& PAGE_MAPPING_FLAGS
) != PAGE_MAPPING_ANON
)
494 if (!page_mapped(page
))
497 anon_vma
= (struct anon_vma
*) (anon_mapping
- PAGE_MAPPING_ANON
);
498 if (!atomic_inc_not_zero(&anon_vma
->refcount
)) {
504 * If this page is still mapped, then its anon_vma cannot have been
505 * freed. But if it has been unmapped, we have no security against the
506 * anon_vma structure being freed and reused (for another anon_vma:
507 * SLAB_TYPESAFE_BY_RCU guarantees that - so the atomic_inc_not_zero()
508 * above cannot corrupt).
510 if (!page_mapped(page
)) {
512 put_anon_vma(anon_vma
);
522 * Similar to page_get_anon_vma() except it locks the anon_vma.
524 * Its a little more complex as it tries to keep the fast path to a single
525 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
526 * reference like with page_get_anon_vma() and then block on the mutex.
528 struct anon_vma
*page_lock_anon_vma_read(struct page
*page
)
530 struct anon_vma
*anon_vma
= NULL
;
531 struct anon_vma
*root_anon_vma
;
532 unsigned long anon_mapping
;
535 anon_mapping
= (unsigned long)READ_ONCE(page
->mapping
);
536 if ((anon_mapping
& PAGE_MAPPING_FLAGS
) != PAGE_MAPPING_ANON
)
538 if (!page_mapped(page
))
541 anon_vma
= (struct anon_vma
*) (anon_mapping
- PAGE_MAPPING_ANON
);
542 root_anon_vma
= READ_ONCE(anon_vma
->root
);
543 if (down_read_trylock(&root_anon_vma
->rwsem
)) {
545 * If the page is still mapped, then this anon_vma is still
546 * its anon_vma, and holding the mutex ensures that it will
547 * not go away, see anon_vma_free().
549 if (!page_mapped(page
)) {
550 up_read(&root_anon_vma
->rwsem
);
556 /* trylock failed, we got to sleep */
557 if (!atomic_inc_not_zero(&anon_vma
->refcount
)) {
562 if (!page_mapped(page
)) {
564 put_anon_vma(anon_vma
);
568 /* we pinned the anon_vma, its safe to sleep */
570 anon_vma_lock_read(anon_vma
);
572 if (atomic_dec_and_test(&anon_vma
->refcount
)) {
574 * Oops, we held the last refcount, release the lock
575 * and bail -- can't simply use put_anon_vma() because
576 * we'll deadlock on the anon_vma_lock_write() recursion.
578 anon_vma_unlock_read(anon_vma
);
579 __put_anon_vma(anon_vma
);
590 void page_unlock_anon_vma_read(struct anon_vma
*anon_vma
)
592 anon_vma_unlock_read(anon_vma
);
595 #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
597 * Flush TLB entries for recently unmapped pages from remote CPUs. It is
598 * important if a PTE was dirty when it was unmapped that it's flushed
599 * before any IO is initiated on the page to prevent lost writes. Similarly,
600 * it must be flushed before freeing to prevent data leakage.
602 void try_to_unmap_flush(void)
604 struct tlbflush_unmap_batch
*tlb_ubc
= ¤t
->tlb_ubc
;
606 if (!tlb_ubc
->flush_required
)
609 arch_tlbbatch_flush(&tlb_ubc
->arch
);
610 tlb_ubc
->flush_required
= false;
611 tlb_ubc
->writable
= false;
614 /* Flush iff there are potentially writable TLB entries that can race with IO */
615 void try_to_unmap_flush_dirty(void)
617 struct tlbflush_unmap_batch
*tlb_ubc
= ¤t
->tlb_ubc
;
619 if (tlb_ubc
->writable
)
620 try_to_unmap_flush();
623 static void set_tlb_ubc_flush_pending(struct mm_struct
*mm
, bool writable
)
625 struct tlbflush_unmap_batch
*tlb_ubc
= ¤t
->tlb_ubc
;
627 arch_tlbbatch_add_mm(&tlb_ubc
->arch
, mm
);
628 tlb_ubc
->flush_required
= true;
631 * Ensure compiler does not re-order the setting of tlb_flush_batched
632 * before the PTE is cleared.
635 mm
->tlb_flush_batched
= true;
638 * If the PTE was dirty then it's best to assume it's writable. The
639 * caller must use try_to_unmap_flush_dirty() or try_to_unmap_flush()
640 * before the page is queued for IO.
643 tlb_ubc
->writable
= true;
647 * Returns true if the TLB flush should be deferred to the end of a batch of
648 * unmap operations to reduce IPIs.
650 static bool should_defer_flush(struct mm_struct
*mm
, enum ttu_flags flags
)
652 bool should_defer
= false;
654 if (!(flags
& TTU_BATCH_FLUSH
))
657 /* If remote CPUs need to be flushed then defer batch the flush */
658 if (cpumask_any_but(mm_cpumask(mm
), get_cpu()) < nr_cpu_ids
)
666 * Reclaim unmaps pages under the PTL but do not flush the TLB prior to
667 * releasing the PTL if TLB flushes are batched. It's possible for a parallel
668 * operation such as mprotect or munmap to race between reclaim unmapping
669 * the page and flushing the page. If this race occurs, it potentially allows
670 * access to data via a stale TLB entry. Tracking all mm's that have TLB
671 * batching in flight would be expensive during reclaim so instead track
672 * whether TLB batching occurred in the past and if so then do a flush here
673 * if required. This will cost one additional flush per reclaim cycle paid
674 * by the first operation at risk such as mprotect and mumap.
676 * This must be called under the PTL so that an access to tlb_flush_batched
677 * that is potentially a "reclaim vs mprotect/munmap/etc" race will synchronise
680 void flush_tlb_batched_pending(struct mm_struct
*mm
)
682 if (data_race(mm
->tlb_flush_batched
)) {
686 * Do not allow the compiler to re-order the clearing of
687 * tlb_flush_batched before the tlb is flushed.
690 mm
->tlb_flush_batched
= false;
694 static void set_tlb_ubc_flush_pending(struct mm_struct
*mm
, bool writable
)
698 static bool should_defer_flush(struct mm_struct
*mm
, enum ttu_flags flags
)
702 #endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
705 * At what user virtual address is page expected in vma?
706 * Caller should check the page is actually part of the vma.
708 unsigned long page_address_in_vma(struct page
*page
, struct vm_area_struct
*vma
)
710 unsigned long address
;
711 if (PageAnon(page
)) {
712 struct anon_vma
*page__anon_vma
= page_anon_vma(page
);
714 * Note: swapoff's unuse_vma() is more efficient with this
715 * check, and needs it to match anon_vma when KSM is active.
717 if (!vma
->anon_vma
|| !page__anon_vma
||
718 vma
->anon_vma
->root
!= page__anon_vma
->root
)
720 } else if (page
->mapping
) {
721 if (!vma
->vm_file
|| vma
->vm_file
->f_mapping
!= page
->mapping
)
725 address
= __vma_address(page
, vma
);
726 if (unlikely(address
< vma
->vm_start
|| address
>= vma
->vm_end
))
731 pmd_t
*mm_find_pmd(struct mm_struct
*mm
, unsigned long address
)
739 pgd
= pgd_offset(mm
, address
);
740 if (!pgd_present(*pgd
))
743 p4d
= p4d_offset(pgd
, address
);
744 if (!p4d_present(*p4d
))
747 pud
= pud_offset(p4d
, address
);
748 if (!pud_present(*pud
))
751 pmd
= pmd_offset(pud
, address
);
753 * Some THP functions use the sequence pmdp_huge_clear_flush(), set_pmd_at()
754 * without holding anon_vma lock for write. So when looking for a
755 * genuine pmde (in which to find pte), test present and !THP together.
759 if (!pmd_present(pmde
) || pmd_trans_huge(pmde
))
765 struct page_referenced_arg
{
768 unsigned long vm_flags
;
769 struct mem_cgroup
*memcg
;
772 * arg: page_referenced_arg will be passed
774 static bool page_referenced_one(struct page
*page
, struct vm_area_struct
*vma
,
775 unsigned long address
, void *arg
)
777 struct page_referenced_arg
*pra
= arg
;
778 struct page_vma_mapped_walk pvmw
= {
785 while (page_vma_mapped_walk(&pvmw
)) {
786 address
= pvmw
.address
;
788 if (vma
->vm_flags
& VM_LOCKED
) {
789 page_vma_mapped_walk_done(&pvmw
);
790 pra
->vm_flags
|= VM_LOCKED
;
791 return false; /* To break the loop */
795 if (ptep_clear_flush_young_notify(vma
, address
,
798 * Don't treat a reference through
799 * a sequentially read mapping as such.
800 * If the page has been used in another mapping,
801 * we will catch it; if this other mapping is
802 * already gone, the unmap path will have set
803 * PG_referenced or activated the page.
805 if (likely(!(vma
->vm_flags
& VM_SEQ_READ
)))
808 } else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE
)) {
809 if (pmdp_clear_flush_young_notify(vma
, address
,
813 /* unexpected pmd-mapped page? */
821 clear_page_idle(page
);
822 if (test_and_clear_page_young(page
))
827 pra
->vm_flags
|= vma
->vm_flags
;
831 return false; /* To break the loop */
836 static bool invalid_page_referenced_vma(struct vm_area_struct
*vma
, void *arg
)
838 struct page_referenced_arg
*pra
= arg
;
839 struct mem_cgroup
*memcg
= pra
->memcg
;
841 if (!mm_match_cgroup(vma
->vm_mm
, memcg
))
848 * page_referenced - test if the page was referenced
849 * @page: the page to test
850 * @is_locked: caller holds lock on the page
851 * @memcg: target memory cgroup
852 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
854 * Quick test_and_clear_referenced for all mappings to a page,
855 * returns the number of ptes which referenced the page.
857 int page_referenced(struct page
*page
,
859 struct mem_cgroup
*memcg
,
860 unsigned long *vm_flags
)
863 struct page_referenced_arg pra
= {
864 .mapcount
= total_mapcount(page
),
867 struct rmap_walk_control rwc
= {
868 .rmap_one
= page_referenced_one
,
870 .anon_lock
= page_lock_anon_vma_read
,
877 if (!page_rmapping(page
))
880 if (!is_locked
&& (!PageAnon(page
) || PageKsm(page
))) {
881 we_locked
= trylock_page(page
);
887 * If we are reclaiming on behalf of a cgroup, skip
888 * counting on behalf of references from different
892 rwc
.invalid_vma
= invalid_page_referenced_vma
;
895 rmap_walk(page
, &rwc
);
896 *vm_flags
= pra
.vm_flags
;
901 return pra
.referenced
;
904 static bool page_mkclean_one(struct page
*page
, struct vm_area_struct
*vma
,
905 unsigned long address
, void *arg
)
907 struct page_vma_mapped_walk pvmw
= {
913 struct mmu_notifier_range range
;
917 * We have to assume the worse case ie pmd for invalidation. Note that
918 * the page can not be free from this function.
920 mmu_notifier_range_init(&range
, MMU_NOTIFY_PROTECTION_PAGE
,
921 0, vma
, vma
->vm_mm
, address
,
922 min(vma
->vm_end
, address
+ page_size(page
)));
923 mmu_notifier_invalidate_range_start(&range
);
925 while (page_vma_mapped_walk(&pvmw
)) {
928 address
= pvmw
.address
;
931 pte_t
*pte
= pvmw
.pte
;
933 if (!pte_dirty(*pte
) && !pte_write(*pte
))
936 flush_cache_page(vma
, address
, pte_pfn(*pte
));
937 entry
= ptep_clear_flush(vma
, address
, pte
);
938 entry
= pte_wrprotect(entry
);
939 entry
= pte_mkclean(entry
);
940 set_pte_at(vma
->vm_mm
, address
, pte
, entry
);
943 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
944 pmd_t
*pmd
= pvmw
.pmd
;
947 if (!pmd_dirty(*pmd
) && !pmd_write(*pmd
))
950 flush_cache_page(vma
, address
, page_to_pfn(page
));
951 entry
= pmdp_invalidate(vma
, address
, pmd
);
952 entry
= pmd_wrprotect(entry
);
953 entry
= pmd_mkclean(entry
);
954 set_pmd_at(vma
->vm_mm
, address
, pmd
, entry
);
957 /* unexpected pmd-mapped page? */
963 * No need to call mmu_notifier_invalidate_range() as we are
964 * downgrading page table protection not changing it to point
967 * See Documentation/vm/mmu_notifier.rst
973 mmu_notifier_invalidate_range_end(&range
);
978 static bool invalid_mkclean_vma(struct vm_area_struct
*vma
, void *arg
)
980 if (vma
->vm_flags
& VM_SHARED
)
986 int page_mkclean(struct page
*page
)
989 struct address_space
*mapping
;
990 struct rmap_walk_control rwc
= {
991 .arg
= (void *)&cleaned
,
992 .rmap_one
= page_mkclean_one
,
993 .invalid_vma
= invalid_mkclean_vma
,
996 BUG_ON(!PageLocked(page
));
998 if (!page_mapped(page
))
1001 mapping
= page_mapping(page
);
1005 rmap_walk(page
, &rwc
);
1009 EXPORT_SYMBOL_GPL(page_mkclean
);
1012 * page_move_anon_rmap - move a page to our anon_vma
1013 * @page: the page to move to our anon_vma
1014 * @vma: the vma the page belongs to
1016 * When a page belongs exclusively to one process after a COW event,
1017 * that page can be moved into the anon_vma that belongs to just that
1018 * process, so the rmap code will not search the parent or sibling
1021 void page_move_anon_rmap(struct page
*page
, struct vm_area_struct
*vma
)
1023 struct anon_vma
*anon_vma
= vma
->anon_vma
;
1025 page
= compound_head(page
);
1027 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
1028 VM_BUG_ON_VMA(!anon_vma
, vma
);
1030 anon_vma
= (void *) anon_vma
+ PAGE_MAPPING_ANON
;
1032 * Ensure that anon_vma and the PAGE_MAPPING_ANON bit are written
1033 * simultaneously, so a concurrent reader (eg page_referenced()'s
1034 * PageAnon()) will not see one without the other.
1036 WRITE_ONCE(page
->mapping
, (struct address_space
*) anon_vma
);
1040 * __page_set_anon_rmap - set up new anonymous rmap
1041 * @page: Page or Hugepage to add to rmap
1042 * @vma: VM area to add page to.
1043 * @address: User virtual address of the mapping
1044 * @exclusive: the page is exclusively owned by the current process
1046 static void __page_set_anon_rmap(struct page
*page
,
1047 struct vm_area_struct
*vma
, unsigned long address
, int exclusive
)
1049 struct anon_vma
*anon_vma
= vma
->anon_vma
;
1057 * If the page isn't exclusively mapped into this vma,
1058 * we must use the _oldest_ possible anon_vma for the
1062 anon_vma
= anon_vma
->root
;
1065 * page_idle does a lockless/optimistic rmap scan on page->mapping.
1066 * Make sure the compiler doesn't split the stores of anon_vma and
1067 * the PAGE_MAPPING_ANON type identifier, otherwise the rmap code
1068 * could mistake the mapping for a struct address_space and crash.
1070 anon_vma
= (void *) anon_vma
+ PAGE_MAPPING_ANON
;
1071 WRITE_ONCE(page
->mapping
, (struct address_space
*) anon_vma
);
1072 page
->index
= linear_page_index(vma
, address
);
1076 * __page_check_anon_rmap - sanity check anonymous rmap addition
1077 * @page: the page to add the mapping to
1078 * @vma: the vm area in which the mapping is added
1079 * @address: the user virtual address mapped
1081 static void __page_check_anon_rmap(struct page
*page
,
1082 struct vm_area_struct
*vma
, unsigned long address
)
1085 * The page's anon-rmap details (mapping and index) are guaranteed to
1086 * be set up correctly at this point.
1088 * We have exclusion against page_add_anon_rmap because the caller
1089 * always holds the page locked.
1091 * We have exclusion against page_add_new_anon_rmap because those pages
1092 * are initially only visible via the pagetables, and the pte is locked
1093 * over the call to page_add_new_anon_rmap.
1095 VM_BUG_ON_PAGE(page_anon_vma(page
)->root
!= vma
->anon_vma
->root
, page
);
1096 VM_BUG_ON_PAGE(page_to_pgoff(page
) != linear_page_index(vma
, address
),
1101 * page_add_anon_rmap - add pte mapping to an anonymous page
1102 * @page: the page to add the mapping to
1103 * @vma: the vm area in which the mapping is added
1104 * @address: the user virtual address mapped
1105 * @compound: charge the page as compound or small page
1107 * The caller needs to hold the pte lock, and the page must be locked in
1108 * the anon_vma case: to serialize mapping,index checking after setting,
1109 * and to ensure that PageAnon is not being upgraded racily to PageKsm
1110 * (but PageKsm is never downgraded to PageAnon).
1112 void page_add_anon_rmap(struct page
*page
,
1113 struct vm_area_struct
*vma
, unsigned long address
, bool compound
)
1115 do_page_add_anon_rmap(page
, vma
, address
, compound
? RMAP_COMPOUND
: 0);
1119 * Special version of the above for do_swap_page, which often runs
1120 * into pages that are exclusively owned by the current process.
1121 * Everybody else should continue to use page_add_anon_rmap above.
1123 void do_page_add_anon_rmap(struct page
*page
,
1124 struct vm_area_struct
*vma
, unsigned long address
, int flags
)
1126 bool compound
= flags
& RMAP_COMPOUND
;
1129 if (unlikely(PageKsm(page
)))
1130 lock_page_memcg(page
);
1132 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
1136 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
1137 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
1138 mapcount
= compound_mapcount_ptr(page
);
1139 first
= atomic_inc_and_test(mapcount
);
1141 first
= atomic_inc_and_test(&page
->_mapcount
);
1145 int nr
= compound
? thp_nr_pages(page
) : 1;
1147 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1148 * these counters are not modified in interrupt context, and
1149 * pte lock(a spinlock) is held, which implies preemption
1153 __mod_lruvec_page_state(page
, NR_ANON_THPS
, nr
);
1154 __mod_lruvec_page_state(page
, NR_ANON_MAPPED
, nr
);
1157 if (unlikely(PageKsm(page
))) {
1158 unlock_page_memcg(page
);
1162 /* address might be in next vma when migration races vma_adjust */
1164 __page_set_anon_rmap(page
, vma
, address
,
1165 flags
& RMAP_EXCLUSIVE
);
1167 __page_check_anon_rmap(page
, vma
, address
);
1171 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1172 * @page: the page to add the mapping to
1173 * @vma: the vm area in which the mapping is added
1174 * @address: the user virtual address mapped
1175 * @compound: charge the page as compound or small page
1177 * Same as page_add_anon_rmap but must only be called on *new* pages.
1178 * This means the inc-and-test can be bypassed.
1179 * Page does not have to be locked.
1181 void page_add_new_anon_rmap(struct page
*page
,
1182 struct vm_area_struct
*vma
, unsigned long address
, bool compound
)
1184 int nr
= compound
? thp_nr_pages(page
) : 1;
1186 VM_BUG_ON_VMA(address
< vma
->vm_start
|| address
>= vma
->vm_end
, vma
);
1187 __SetPageSwapBacked(page
);
1189 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
1190 /* increment count (starts at -1) */
1191 atomic_set(compound_mapcount_ptr(page
), 0);
1192 if (hpage_pincount_available(page
))
1193 atomic_set(compound_pincount_ptr(page
), 0);
1195 __mod_lruvec_page_state(page
, NR_ANON_THPS
, nr
);
1197 /* Anon THP always mapped first with PMD */
1198 VM_BUG_ON_PAGE(PageTransCompound(page
), page
);
1199 /* increment count (starts at -1) */
1200 atomic_set(&page
->_mapcount
, 0);
1202 __mod_lruvec_page_state(page
, NR_ANON_MAPPED
, nr
);
1203 __page_set_anon_rmap(page
, vma
, address
, 1);
1207 * page_add_file_rmap - add pte mapping to a file page
1208 * @page: the page to add the mapping to
1209 * @compound: charge the page as compound or small page
1211 * The caller needs to hold the pte lock.
1213 void page_add_file_rmap(struct page
*page
, bool compound
)
1217 VM_BUG_ON_PAGE(compound
&& !PageTransHuge(page
), page
);
1218 lock_page_memcg(page
);
1219 if (compound
&& PageTransHuge(page
)) {
1220 int nr_pages
= thp_nr_pages(page
);
1222 for (i
= 0, nr
= 0; i
< nr_pages
; i
++) {
1223 if (atomic_inc_and_test(&page
[i
]._mapcount
))
1226 if (!atomic_inc_and_test(compound_mapcount_ptr(page
)))
1228 if (PageSwapBacked(page
))
1229 __mod_lruvec_page_state(page
, NR_SHMEM_PMDMAPPED
,
1232 __mod_lruvec_page_state(page
, NR_FILE_PMDMAPPED
,
1235 if (PageTransCompound(page
) && page_mapping(page
)) {
1236 VM_WARN_ON_ONCE(!PageLocked(page
));
1238 SetPageDoubleMap(compound_head(page
));
1239 if (PageMlocked(page
))
1240 clear_page_mlock(compound_head(page
));
1242 if (!atomic_inc_and_test(&page
->_mapcount
))
1245 __mod_lruvec_page_state(page
, NR_FILE_MAPPED
, nr
);
1247 unlock_page_memcg(page
);
1250 static void page_remove_file_rmap(struct page
*page
, bool compound
)
1254 VM_BUG_ON_PAGE(compound
&& !PageHead(page
), page
);
1256 /* Hugepages are not counted in NR_FILE_MAPPED for now. */
1257 if (unlikely(PageHuge(page
))) {
1258 /* hugetlb pages are always mapped with pmds */
1259 atomic_dec(compound_mapcount_ptr(page
));
1263 /* page still mapped by someone else? */
1264 if (compound
&& PageTransHuge(page
)) {
1265 int nr_pages
= thp_nr_pages(page
);
1267 for (i
= 0, nr
= 0; i
< nr_pages
; i
++) {
1268 if (atomic_add_negative(-1, &page
[i
]._mapcount
))
1271 if (!atomic_add_negative(-1, compound_mapcount_ptr(page
)))
1273 if (PageSwapBacked(page
))
1274 __mod_lruvec_page_state(page
, NR_SHMEM_PMDMAPPED
,
1277 __mod_lruvec_page_state(page
, NR_FILE_PMDMAPPED
,
1280 if (!atomic_add_negative(-1, &page
->_mapcount
))
1285 * We use the irq-unsafe __{inc|mod}_lruvec_page_state because
1286 * these counters are not modified in interrupt context, and
1287 * pte lock(a spinlock) is held, which implies preemption disabled.
1289 __mod_lruvec_page_state(page
, NR_FILE_MAPPED
, -nr
);
1291 if (unlikely(PageMlocked(page
)))
1292 clear_page_mlock(page
);
1295 static void page_remove_anon_compound_rmap(struct page
*page
)
1299 if (!atomic_add_negative(-1, compound_mapcount_ptr(page
)))
1302 /* Hugepages are not counted in NR_ANON_PAGES for now. */
1303 if (unlikely(PageHuge(page
)))
1306 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE
))
1309 __mod_lruvec_page_state(page
, NR_ANON_THPS
, -thp_nr_pages(page
));
1311 if (TestClearPageDoubleMap(page
)) {
1313 * Subpages can be mapped with PTEs too. Check how many of
1314 * them are still mapped.
1316 for (i
= 0, nr
= 0; i
< thp_nr_pages(page
); i
++) {
1317 if (atomic_add_negative(-1, &page
[i
]._mapcount
))
1322 * Queue the page for deferred split if at least one small
1323 * page of the compound page is unmapped, but at least one
1324 * small page is still mapped.
1326 if (nr
&& nr
< thp_nr_pages(page
))
1327 deferred_split_huge_page(page
);
1329 nr
= thp_nr_pages(page
);
1332 if (unlikely(PageMlocked(page
)))
1333 clear_page_mlock(page
);
1336 __mod_lruvec_page_state(page
, NR_ANON_MAPPED
, -nr
);
1340 * page_remove_rmap - take down pte mapping from a page
1341 * @page: page to remove mapping from
1342 * @compound: uncharge the page as compound or small page
1344 * The caller needs to hold the pte lock.
1346 void page_remove_rmap(struct page
*page
, bool compound
)
1348 lock_page_memcg(page
);
1350 if (!PageAnon(page
)) {
1351 page_remove_file_rmap(page
, compound
);
1356 page_remove_anon_compound_rmap(page
);
1360 /* page still mapped by someone else? */
1361 if (!atomic_add_negative(-1, &page
->_mapcount
))
1365 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1366 * these counters are not modified in interrupt context, and
1367 * pte lock(a spinlock) is held, which implies preemption disabled.
1369 __dec_lruvec_page_state(page
, NR_ANON_MAPPED
);
1371 if (unlikely(PageMlocked(page
)))
1372 clear_page_mlock(page
);
1374 if (PageTransCompound(page
))
1375 deferred_split_huge_page(compound_head(page
));
1378 * It would be tidy to reset the PageAnon mapping here,
1379 * but that might overwrite a racing page_add_anon_rmap
1380 * which increments mapcount after us but sets mapping
1381 * before us: so leave the reset to free_unref_page,
1382 * and remember that it's only reliable while mapped.
1383 * Leaving it set also helps swapoff to reinstate ptes
1384 * faster for those pages still in swapcache.
1387 unlock_page_memcg(page
);
1391 * @arg: enum ttu_flags will be passed to this argument
1393 static bool try_to_unmap_one(struct page
*page
, struct vm_area_struct
*vma
,
1394 unsigned long address
, void *arg
)
1396 struct mm_struct
*mm
= vma
->vm_mm
;
1397 struct page_vma_mapped_walk pvmw
= {
1403 struct page
*subpage
;
1405 struct mmu_notifier_range range
;
1406 enum ttu_flags flags
= (enum ttu_flags
)(long)arg
;
1408 /* munlock has nothing to gain from examining un-locked vmas */
1409 if ((flags
& TTU_MUNLOCK
) && !(vma
->vm_flags
& VM_LOCKED
))
1412 if (IS_ENABLED(CONFIG_MIGRATION
) && (flags
& TTU_MIGRATION
) &&
1413 is_zone_device_page(page
) && !is_device_private_page(page
))
1416 if (flags
& TTU_SPLIT_HUGE_PMD
) {
1417 split_huge_pmd_address(vma
, address
,
1418 flags
& TTU_SPLIT_FREEZE
, page
);
1422 * For THP, we have to assume the worse case ie pmd for invalidation.
1423 * For hugetlb, it could be much worse if we need to do pud
1424 * invalidation in the case of pmd sharing.
1426 * Note that the page can not be free in this function as call of
1427 * try_to_unmap() must hold a reference on the page.
1429 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
, vma
->vm_mm
,
1431 min(vma
->vm_end
, address
+ page_size(page
)));
1432 if (PageHuge(page
)) {
1434 * If sharing is possible, start and end will be adjusted
1437 adjust_range_if_pmd_sharing_possible(vma
, &range
.start
,
1440 mmu_notifier_invalidate_range_start(&range
);
1442 while (page_vma_mapped_walk(&pvmw
)) {
1443 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1444 /* PMD-mapped THP migration entry */
1445 if (!pvmw
.pte
&& (flags
& TTU_MIGRATION
)) {
1446 VM_BUG_ON_PAGE(PageHuge(page
) || !PageTransCompound(page
), page
);
1448 set_pmd_migration_entry(&pvmw
, page
);
1454 * If the page is mlock()d, we cannot swap it out.
1455 * If it's recently referenced (perhaps page_referenced
1456 * skipped over this mm) then we should reactivate it.
1458 if (!(flags
& TTU_IGNORE_MLOCK
)) {
1459 if (vma
->vm_flags
& VM_LOCKED
) {
1460 /* PTE-mapped THP are never mlocked */
1461 if (!PageTransCompound(page
)) {
1463 * Holding pte lock, we do *not* need
1466 mlock_vma_page(page
);
1469 page_vma_mapped_walk_done(&pvmw
);
1472 if (flags
& TTU_MUNLOCK
)
1476 /* Unexpected PMD-mapped THP? */
1477 VM_BUG_ON_PAGE(!pvmw
.pte
, page
);
1479 subpage
= page
- page_to_pfn(page
) + pte_pfn(*pvmw
.pte
);
1480 address
= pvmw
.address
;
1482 if (PageHuge(page
) && !PageAnon(page
)) {
1484 * To call huge_pmd_unshare, i_mmap_rwsem must be
1485 * held in write mode. Caller needs to explicitly
1486 * do this outside rmap routines.
1488 VM_BUG_ON(!(flags
& TTU_RMAP_LOCKED
));
1489 if (huge_pmd_unshare(mm
, vma
, &address
, pvmw
.pte
)) {
1491 * huge_pmd_unshare unmapped an entire PMD
1492 * page. There is no way of knowing exactly
1493 * which PMDs may be cached for this mm, so
1494 * we must flush them all. start/end were
1495 * already adjusted above to cover this range.
1497 flush_cache_range(vma
, range
.start
, range
.end
);
1498 flush_tlb_range(vma
, range
.start
, range
.end
);
1499 mmu_notifier_invalidate_range(mm
, range
.start
,
1503 * The ref count of the PMD page was dropped
1504 * which is part of the way map counting
1505 * is done for shared PMDs. Return 'true'
1506 * here. When there is no other sharing,
1507 * huge_pmd_unshare returns false and we will
1508 * unmap the actual page and drop map count
1511 page_vma_mapped_walk_done(&pvmw
);
1516 if (IS_ENABLED(CONFIG_MIGRATION
) &&
1517 (flags
& TTU_MIGRATION
) &&
1518 is_zone_device_page(page
)) {
1522 pteval
= ptep_get_and_clear(mm
, pvmw
.address
, pvmw
.pte
);
1525 * Store the pfn of the page in a special migration
1526 * pte. do_swap_page() will wait until the migration
1527 * pte is removed and then restart fault handling.
1529 entry
= make_migration_entry(page
, 0);
1530 swp_pte
= swp_entry_to_pte(entry
);
1533 * pteval maps a zone device page and is therefore
1536 if (pte_swp_soft_dirty(pteval
))
1537 swp_pte
= pte_swp_mksoft_dirty(swp_pte
);
1538 if (pte_swp_uffd_wp(pteval
))
1539 swp_pte
= pte_swp_mkuffd_wp(swp_pte
);
1540 set_pte_at(mm
, pvmw
.address
, pvmw
.pte
, swp_pte
);
1542 * No need to invalidate here it will synchronize on
1543 * against the special swap migration pte.
1545 * The assignment to subpage above was computed from a
1546 * swap PTE which results in an invalid pointer.
1547 * Since only PAGE_SIZE pages can currently be
1548 * migrated, just set it to page. This will need to be
1549 * changed when hugepage migrations to device private
1550 * memory are supported.
1556 /* Nuke the page table entry. */
1557 flush_cache_page(vma
, address
, pte_pfn(*pvmw
.pte
));
1558 if (should_defer_flush(mm
, flags
)) {
1560 * We clear the PTE but do not flush so potentially
1561 * a remote CPU could still be writing to the page.
1562 * If the entry was previously clean then the
1563 * architecture must guarantee that a clear->dirty
1564 * transition on a cached TLB entry is written through
1565 * and traps if the PTE is unmapped.
1567 pteval
= ptep_get_and_clear(mm
, address
, pvmw
.pte
);
1569 set_tlb_ubc_flush_pending(mm
, pte_dirty(pteval
));
1571 pteval
= ptep_clear_flush(vma
, address
, pvmw
.pte
);
1574 /* Move the dirty bit to the page. Now the pte is gone. */
1575 if (pte_dirty(pteval
))
1576 set_page_dirty(page
);
1578 /* Update high watermark before we lower rss */
1579 update_hiwater_rss(mm
);
1581 if (PageHWPoison(page
) && !(flags
& TTU_IGNORE_HWPOISON
)) {
1582 pteval
= swp_entry_to_pte(make_hwpoison_entry(subpage
));
1583 if (PageHuge(page
)) {
1584 hugetlb_count_sub(compound_nr(page
), mm
);
1585 set_huge_swap_pte_at(mm
, address
,
1587 vma_mmu_pagesize(vma
));
1589 dec_mm_counter(mm
, mm_counter(page
));
1590 set_pte_at(mm
, address
, pvmw
.pte
, pteval
);
1593 } else if (pte_unused(pteval
) && !userfaultfd_armed(vma
)) {
1595 * The guest indicated that the page content is of no
1596 * interest anymore. Simply discard the pte, vmscan
1597 * will take care of the rest.
1598 * A future reference will then fault in a new zero
1599 * page. When userfaultfd is active, we must not drop
1600 * this page though, as its main user (postcopy
1601 * migration) will not expect userfaults on already
1604 dec_mm_counter(mm
, mm_counter(page
));
1605 /* We have to invalidate as we cleared the pte */
1606 mmu_notifier_invalidate_range(mm
, address
,
1607 address
+ PAGE_SIZE
);
1608 } else if (IS_ENABLED(CONFIG_MIGRATION
) &&
1609 (flags
& (TTU_MIGRATION
|TTU_SPLIT_FREEZE
))) {
1613 if (arch_unmap_one(mm
, vma
, address
, pteval
) < 0) {
1614 set_pte_at(mm
, address
, pvmw
.pte
, pteval
);
1616 page_vma_mapped_walk_done(&pvmw
);
1621 * Store the pfn of the page in a special migration
1622 * pte. do_swap_page() will wait until the migration
1623 * pte is removed and then restart fault handling.
1625 entry
= make_migration_entry(subpage
,
1627 swp_pte
= swp_entry_to_pte(entry
);
1628 if (pte_soft_dirty(pteval
))
1629 swp_pte
= pte_swp_mksoft_dirty(swp_pte
);
1630 if (pte_uffd_wp(pteval
))
1631 swp_pte
= pte_swp_mkuffd_wp(swp_pte
);
1632 set_pte_at(mm
, address
, pvmw
.pte
, swp_pte
);
1634 * No need to invalidate here it will synchronize on
1635 * against the special swap migration pte.
1637 } else if (PageAnon(page
)) {
1638 swp_entry_t entry
= { .val
= page_private(subpage
) };
1641 * Store the swap location in the pte.
1642 * See handle_pte_fault() ...
1644 if (unlikely(PageSwapBacked(page
) != PageSwapCache(page
))) {
1647 /* We have to invalidate as we cleared the pte */
1648 mmu_notifier_invalidate_range(mm
, address
,
1649 address
+ PAGE_SIZE
);
1650 page_vma_mapped_walk_done(&pvmw
);
1654 /* MADV_FREE page check */
1655 if (!PageSwapBacked(page
)) {
1656 if (!PageDirty(page
)) {
1657 /* Invalidate as we cleared the pte */
1658 mmu_notifier_invalidate_range(mm
,
1659 address
, address
+ PAGE_SIZE
);
1660 dec_mm_counter(mm
, MM_ANONPAGES
);
1665 * If the page was redirtied, it cannot be
1666 * discarded. Remap the page to page table.
1668 set_pte_at(mm
, address
, pvmw
.pte
, pteval
);
1669 SetPageSwapBacked(page
);
1671 page_vma_mapped_walk_done(&pvmw
);
1675 if (swap_duplicate(entry
) < 0) {
1676 set_pte_at(mm
, address
, pvmw
.pte
, pteval
);
1678 page_vma_mapped_walk_done(&pvmw
);
1681 if (arch_unmap_one(mm
, vma
, address
, pteval
) < 0) {
1682 set_pte_at(mm
, address
, pvmw
.pte
, pteval
);
1684 page_vma_mapped_walk_done(&pvmw
);
1687 if (list_empty(&mm
->mmlist
)) {
1688 spin_lock(&mmlist_lock
);
1689 if (list_empty(&mm
->mmlist
))
1690 list_add(&mm
->mmlist
, &init_mm
.mmlist
);
1691 spin_unlock(&mmlist_lock
);
1693 dec_mm_counter(mm
, MM_ANONPAGES
);
1694 inc_mm_counter(mm
, MM_SWAPENTS
);
1695 swp_pte
= swp_entry_to_pte(entry
);
1696 if (pte_soft_dirty(pteval
))
1697 swp_pte
= pte_swp_mksoft_dirty(swp_pte
);
1698 if (pte_uffd_wp(pteval
))
1699 swp_pte
= pte_swp_mkuffd_wp(swp_pte
);
1700 set_pte_at(mm
, address
, pvmw
.pte
, swp_pte
);
1701 /* Invalidate as we cleared the pte */
1702 mmu_notifier_invalidate_range(mm
, address
,
1703 address
+ PAGE_SIZE
);
1706 * This is a locked file-backed page, thus it cannot
1707 * be removed from the page cache and replaced by a new
1708 * page before mmu_notifier_invalidate_range_end, so no
1709 * concurrent thread might update its page table to
1710 * point at new page while a device still is using this
1713 * See Documentation/vm/mmu_notifier.rst
1715 dec_mm_counter(mm
, mm_counter_file(page
));
1719 * No need to call mmu_notifier_invalidate_range() it has be
1720 * done above for all cases requiring it to happen under page
1721 * table lock before mmu_notifier_invalidate_range_end()
1723 * See Documentation/vm/mmu_notifier.rst
1725 page_remove_rmap(subpage
, PageHuge(page
));
1729 mmu_notifier_invalidate_range_end(&range
);
1734 static bool invalid_migration_vma(struct vm_area_struct
*vma
, void *arg
)
1736 return vma_is_temporary_stack(vma
);
1739 static int page_not_mapped(struct page
*page
)
1741 return !page_mapped(page
);
1745 * try_to_unmap - try to remove all page table mappings to a page
1746 * @page: the page to get unmapped
1747 * @flags: action and flags
1749 * Tries to remove all the page table entries which are mapping this
1750 * page, used in the pageout path. Caller must hold the page lock.
1752 * If unmap is successful, return true. Otherwise, false.
1754 bool try_to_unmap(struct page
*page
, enum ttu_flags flags
)
1756 struct rmap_walk_control rwc
= {
1757 .rmap_one
= try_to_unmap_one
,
1758 .arg
= (void *)flags
,
1759 .done
= page_not_mapped
,
1760 .anon_lock
= page_lock_anon_vma_read
,
1764 * During exec, a temporary VMA is setup and later moved.
1765 * The VMA is moved under the anon_vma lock but not the
1766 * page tables leading to a race where migration cannot
1767 * find the migration ptes. Rather than increasing the
1768 * locking requirements of exec(), migration skips
1769 * temporary VMAs until after exec() completes.
1771 if ((flags
& (TTU_MIGRATION
|TTU_SPLIT_FREEZE
))
1772 && !PageKsm(page
) && PageAnon(page
))
1773 rwc
.invalid_vma
= invalid_migration_vma
;
1775 if (flags
& TTU_RMAP_LOCKED
)
1776 rmap_walk_locked(page
, &rwc
);
1778 rmap_walk(page
, &rwc
);
1780 return !page_mapcount(page
) ? true : false;
1784 * try_to_munlock - try to munlock a page
1785 * @page: the page to be munlocked
1787 * Called from munlock code. Checks all of the VMAs mapping the page
1788 * to make sure nobody else has this page mlocked. The page will be
1789 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1792 void try_to_munlock(struct page
*page
)
1794 struct rmap_walk_control rwc
= {
1795 .rmap_one
= try_to_unmap_one
,
1796 .arg
= (void *)TTU_MUNLOCK
,
1797 .done
= page_not_mapped
,
1798 .anon_lock
= page_lock_anon_vma_read
,
1802 VM_BUG_ON_PAGE(!PageLocked(page
) || PageLRU(page
), page
);
1803 VM_BUG_ON_PAGE(PageCompound(page
) && PageDoubleMap(page
), page
);
1805 rmap_walk(page
, &rwc
);
1808 void __put_anon_vma(struct anon_vma
*anon_vma
)
1810 struct anon_vma
*root
= anon_vma
->root
;
1812 anon_vma_free(anon_vma
);
1813 if (root
!= anon_vma
&& atomic_dec_and_test(&root
->refcount
))
1814 anon_vma_free(root
);
1817 static struct anon_vma
*rmap_walk_anon_lock(struct page
*page
,
1818 struct rmap_walk_control
*rwc
)
1820 struct anon_vma
*anon_vma
;
1823 return rwc
->anon_lock(page
);
1826 * Note: remove_migration_ptes() cannot use page_lock_anon_vma_read()
1827 * because that depends on page_mapped(); but not all its usages
1828 * are holding mmap_lock. Users without mmap_lock are required to
1829 * take a reference count to prevent the anon_vma disappearing
1831 anon_vma
= page_anon_vma(page
);
1835 anon_vma_lock_read(anon_vma
);
1840 * rmap_walk_anon - do something to anonymous page using the object-based
1842 * @page: the page to be handled
1843 * @rwc: control variable according to each walk type
1845 * Find all the mappings of a page using the mapping pointer and the vma chains
1846 * contained in the anon_vma struct it points to.
1848 * When called from try_to_munlock(), the mmap_lock of the mm containing the vma
1849 * where the page was found will be held for write. So, we won't recheck
1850 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1853 static void rmap_walk_anon(struct page
*page
, struct rmap_walk_control
*rwc
,
1856 struct anon_vma
*anon_vma
;
1857 pgoff_t pgoff_start
, pgoff_end
;
1858 struct anon_vma_chain
*avc
;
1861 anon_vma
= page_anon_vma(page
);
1862 /* anon_vma disappear under us? */
1863 VM_BUG_ON_PAGE(!anon_vma
, page
);
1865 anon_vma
= rmap_walk_anon_lock(page
, rwc
);
1870 pgoff_start
= page_to_pgoff(page
);
1871 pgoff_end
= pgoff_start
+ thp_nr_pages(page
) - 1;
1872 anon_vma_interval_tree_foreach(avc
, &anon_vma
->rb_root
,
1873 pgoff_start
, pgoff_end
) {
1874 struct vm_area_struct
*vma
= avc
->vma
;
1875 unsigned long address
= vma_address(page
, vma
);
1879 if (rwc
->invalid_vma
&& rwc
->invalid_vma(vma
, rwc
->arg
))
1882 if (!rwc
->rmap_one(page
, vma
, address
, rwc
->arg
))
1884 if (rwc
->done
&& rwc
->done(page
))
1889 anon_vma_unlock_read(anon_vma
);
1893 * rmap_walk_file - do something to file page using the object-based rmap method
1894 * @page: the page to be handled
1895 * @rwc: control variable according to each walk type
1897 * Find all the mappings of a page using the mapping pointer and the vma chains
1898 * contained in the address_space struct it points to.
1900 * When called from try_to_munlock(), the mmap_lock of the mm containing the vma
1901 * where the page was found will be held for write. So, we won't recheck
1902 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1905 static void rmap_walk_file(struct page
*page
, struct rmap_walk_control
*rwc
,
1908 struct address_space
*mapping
= page_mapping(page
);
1909 pgoff_t pgoff_start
, pgoff_end
;
1910 struct vm_area_struct
*vma
;
1913 * The page lock not only makes sure that page->mapping cannot
1914 * suddenly be NULLified by truncation, it makes sure that the
1915 * structure at mapping cannot be freed and reused yet,
1916 * so we can safely take mapping->i_mmap_rwsem.
1918 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
1923 pgoff_start
= page_to_pgoff(page
);
1924 pgoff_end
= pgoff_start
+ thp_nr_pages(page
) - 1;
1926 i_mmap_lock_read(mapping
);
1927 vma_interval_tree_foreach(vma
, &mapping
->i_mmap
,
1928 pgoff_start
, pgoff_end
) {
1929 unsigned long address
= vma_address(page
, vma
);
1933 if (rwc
->invalid_vma
&& rwc
->invalid_vma(vma
, rwc
->arg
))
1936 if (!rwc
->rmap_one(page
, vma
, address
, rwc
->arg
))
1938 if (rwc
->done
&& rwc
->done(page
))
1944 i_mmap_unlock_read(mapping
);
1947 void rmap_walk(struct page
*page
, struct rmap_walk_control
*rwc
)
1949 if (unlikely(PageKsm(page
)))
1950 rmap_walk_ksm(page
, rwc
);
1951 else if (PageAnon(page
))
1952 rmap_walk_anon(page
, rwc
, false);
1954 rmap_walk_file(page
, rwc
, false);
1957 /* Like rmap_walk, but caller holds relevant rmap lock */
1958 void rmap_walk_locked(struct page
*page
, struct rmap_walk_control
*rwc
)
1960 /* no ksm support for now */
1961 VM_BUG_ON_PAGE(PageKsm(page
), page
);
1963 rmap_walk_anon(page
, rwc
, true);
1965 rmap_walk_file(page
, rwc
, true);
1968 #ifdef CONFIG_HUGETLB_PAGE
1970 * The following two functions are for anonymous (private mapped) hugepages.
1971 * Unlike common anonymous pages, anonymous hugepages have no accounting code
1972 * and no lru code, because we handle hugepages differently from common pages.
1974 void hugepage_add_anon_rmap(struct page
*page
,
1975 struct vm_area_struct
*vma
, unsigned long address
)
1977 struct anon_vma
*anon_vma
= vma
->anon_vma
;
1980 BUG_ON(!PageLocked(page
));
1982 /* address might be in next vma when migration races vma_adjust */
1983 first
= atomic_inc_and_test(compound_mapcount_ptr(page
));
1985 __page_set_anon_rmap(page
, vma
, address
, 0);
1988 void hugepage_add_new_anon_rmap(struct page
*page
,
1989 struct vm_area_struct
*vma
, unsigned long address
)
1991 BUG_ON(address
< vma
->vm_start
|| address
>= vma
->vm_end
);
1992 atomic_set(compound_mapcount_ptr(page
), 0);
1993 if (hpage_pincount_available(page
))
1994 atomic_set(compound_pincount_ptr(page
), 0);
1996 __page_set_anon_rmap(page
, vma
, address
, 1);
1998 #endif /* CONFIG_HUGETLB_PAGE */