2 * mm/rmap.c - physical to virtual reverse mappings
4 * Copyright 2001, Rik van Riel <riel@conectiva.com.br>
5 * Released under the General Public License (GPL).
7 * Simple, low overhead reverse mapping scheme.
8 * Please try to keep this thing as modular as possible.
10 * Provides methods for unmapping each kind of mapped page:
11 * the anon methods track anonymous pages, and
12 * the file methods track pages belonging to an inode.
14 * Original design by Rik van Riel <riel@conectiva.com.br> 2001
15 * File methods by Dave McCracken <dmccr@us.ibm.com> 2003, 2004
16 * Anonymous methods by Andrea Arcangeli <andrea@suse.de> 2004
17 * Contributions by Hugh Dickins 2003, 2004
21 * Lock ordering in mm:
23 * inode->i_mutex (while writing or truncating, not reading or faulting)
25 * page->flags PG_locked (lock_page)
26 * hugetlbfs_i_mmap_rwsem_key (in huge_pmd_share)
27 * mapping->i_mmap_rwsem
29 * mm->page_table_lock or pte_lock
30 * pgdat->lru_lock (in mark_page_accessed, isolate_lru_page)
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 * mem_cgroup_{begin,end}_page_stat (memcg->move_lock)
35 * i_pages lock (widely used)
36 * inode->i_lock (in set_page_dirty's __mark_inode_dirty)
37 * bdi.wb->list_lock (in set_page_dirty's __mark_inode_dirty)
38 * sb_lock (within inode_lock in fs/fs-writeback.c)
39 * i_pages lock (widely used, in set_page_dirty,
40 * in arch-dependent flush_dcache_mmap_lock,
41 * within bdi.wb->list_lock in __sync_single_inode)
43 * anon_vma->rwsem,mapping->i_mutex (memory_failure, collect_procs_anon)
49 #include <linux/sched/mm.h>
50 #include <linux/sched/task.h>
51 #include <linux/pagemap.h>
52 #include <linux/swap.h>
53 #include <linux/swapops.h>
54 #include <linux/slab.h>
55 #include <linux/init.h>
56 #include <linux/ksm.h>
57 #include <linux/rmap.h>
58 #include <linux/rcupdate.h>
59 #include <linux/export.h>
60 #include <linux/memcontrol.h>
61 #include <linux/mmu_notifier.h>
62 #include <linux/migrate.h>
63 #include <linux/hugetlb.h>
64 #include <linux/huge_mm.h>
65 #include <linux/backing-dev.h>
66 #include <linux/page_idle.h>
67 #include <linux/memremap.h>
68 #include <linux/userfaultfd_k.h>
70 #include <asm/tlbflush.h>
72 #include <trace/events/tlb.h>
76 static struct kmem_cache
*anon_vma_cachep
;
77 static struct kmem_cache
*anon_vma_chain_cachep
;
79 static inline struct anon_vma
*anon_vma_alloc(void)
81 struct anon_vma
*anon_vma
;
83 anon_vma
= kmem_cache_alloc(anon_vma_cachep
, GFP_KERNEL
);
85 atomic_set(&anon_vma
->refcount
, 1);
86 anon_vma
->degree
= 1; /* Reference for first vma */
87 anon_vma
->parent
= anon_vma
;
89 * Initialise the anon_vma root to point to itself. If called
90 * from fork, the root will be reset to the parents anon_vma.
92 anon_vma
->root
= anon_vma
;
98 static inline void anon_vma_free(struct anon_vma
*anon_vma
)
100 VM_BUG_ON(atomic_read(&anon_vma
->refcount
));
103 * Synchronize against page_lock_anon_vma_read() such that
104 * we can safely hold the lock without the anon_vma getting
107 * Relies on the full mb implied by the atomic_dec_and_test() from
108 * put_anon_vma() against the acquire barrier implied by
109 * down_read_trylock() from page_lock_anon_vma_read(). This orders:
111 * page_lock_anon_vma_read() VS put_anon_vma()
112 * down_read_trylock() atomic_dec_and_test()
114 * atomic_read() rwsem_is_locked()
116 * LOCK should suffice since the actual taking of the lock must
117 * happen _before_ what follows.
120 if (rwsem_is_locked(&anon_vma
->root
->rwsem
)) {
121 anon_vma_lock_write(anon_vma
);
122 anon_vma_unlock_write(anon_vma
);
125 kmem_cache_free(anon_vma_cachep
, anon_vma
);
128 static inline struct anon_vma_chain
*anon_vma_chain_alloc(gfp_t gfp
)
130 return kmem_cache_alloc(anon_vma_chain_cachep
, gfp
);
133 static void anon_vma_chain_free(struct anon_vma_chain
*anon_vma_chain
)
135 kmem_cache_free(anon_vma_chain_cachep
, anon_vma_chain
);
138 static void anon_vma_chain_link(struct vm_area_struct
*vma
,
139 struct anon_vma_chain
*avc
,
140 struct anon_vma
*anon_vma
)
143 avc
->anon_vma
= anon_vma
;
144 list_add(&avc
->same_vma
, &vma
->anon_vma_chain
);
145 anon_vma_interval_tree_insert(avc
, &anon_vma
->rb_root
);
149 * __anon_vma_prepare - attach an anon_vma to a memory region
150 * @vma: the memory region in question
152 * This makes sure the memory mapping described by 'vma' has
153 * an 'anon_vma' attached to it, so that we can associate the
154 * anonymous pages mapped into it with that anon_vma.
156 * The common case will be that we already have one, which
157 * is handled inline by anon_vma_prepare(). But if
158 * not we either need to find an adjacent mapping that we
159 * can re-use the anon_vma from (very common when the only
160 * reason for splitting a vma has been mprotect()), or we
161 * allocate a new one.
163 * Anon-vma allocations are very subtle, because we may have
164 * optimistically looked up an anon_vma in page_lock_anon_vma_read()
165 * and that may actually touch the spinlock even in the newly
166 * allocated vma (it depends on RCU to make sure that the
167 * anon_vma isn't actually destroyed).
169 * As a result, we need to do proper anon_vma locking even
170 * for the new allocation. At the same time, we do not want
171 * to do any locking for the common case of already having
174 * This must be called with the mmap_sem held for reading.
176 int __anon_vma_prepare(struct vm_area_struct
*vma
)
178 struct mm_struct
*mm
= vma
->vm_mm
;
179 struct anon_vma
*anon_vma
, *allocated
;
180 struct anon_vma_chain
*avc
;
184 avc
= anon_vma_chain_alloc(GFP_KERNEL
);
188 anon_vma
= find_mergeable_anon_vma(vma
);
191 anon_vma
= anon_vma_alloc();
192 if (unlikely(!anon_vma
))
193 goto out_enomem_free_avc
;
194 allocated
= anon_vma
;
197 anon_vma_lock_write(anon_vma
);
198 /* page_table_lock to protect against threads */
199 spin_lock(&mm
->page_table_lock
);
200 if (likely(!vma
->anon_vma
)) {
201 vma
->anon_vma
= anon_vma
;
202 anon_vma_chain_link(vma
, avc
, anon_vma
);
203 /* vma reference or self-parent link for new root */
208 spin_unlock(&mm
->page_table_lock
);
209 anon_vma_unlock_write(anon_vma
);
211 if (unlikely(allocated
))
212 put_anon_vma(allocated
);
214 anon_vma_chain_free(avc
);
219 anon_vma_chain_free(avc
);
225 * This is a useful helper function for locking the anon_vma root as
226 * we traverse the vma->anon_vma_chain, looping over anon_vma's that
229 * Such anon_vma's should have the same root, so you'd expect to see
230 * just a single mutex_lock for the whole traversal.
232 static inline struct anon_vma
*lock_anon_vma_root(struct anon_vma
*root
, struct anon_vma
*anon_vma
)
234 struct anon_vma
*new_root
= anon_vma
->root
;
235 if (new_root
!= root
) {
236 if (WARN_ON_ONCE(root
))
237 up_write(&root
->rwsem
);
239 down_write(&root
->rwsem
);
244 static inline void unlock_anon_vma_root(struct anon_vma
*root
)
247 up_write(&root
->rwsem
);
251 * Attach the anon_vmas from src to dst.
252 * Returns 0 on success, -ENOMEM on failure.
254 * If dst->anon_vma is NULL this function tries to find and reuse existing
255 * anon_vma which has no vmas and only one child anon_vma. This prevents
256 * degradation of anon_vma hierarchy to endless linear chain in case of
257 * constantly forking task. On the other hand, an anon_vma with more than one
258 * child isn't reused even if there was no alive vma, thus rmap walker has a
259 * good chance of avoiding scanning the whole hierarchy when it searches where
262 int anon_vma_clone(struct vm_area_struct
*dst
, struct vm_area_struct
*src
)
264 struct anon_vma_chain
*avc
, *pavc
;
265 struct anon_vma
*root
= NULL
;
267 list_for_each_entry_reverse(pavc
, &src
->anon_vma_chain
, same_vma
) {
268 struct anon_vma
*anon_vma
;
270 avc
= anon_vma_chain_alloc(GFP_NOWAIT
| __GFP_NOWARN
);
271 if (unlikely(!avc
)) {
272 unlock_anon_vma_root(root
);
274 avc
= anon_vma_chain_alloc(GFP_KERNEL
);
278 anon_vma
= pavc
->anon_vma
;
279 root
= lock_anon_vma_root(root
, anon_vma
);
280 anon_vma_chain_link(dst
, avc
, anon_vma
);
283 * Reuse existing anon_vma if its degree lower than two,
284 * that means it has no vma and only one anon_vma child.
286 * Do not chose parent anon_vma, otherwise first child
287 * will always reuse it. Root anon_vma is never reused:
288 * it has self-parent reference and at least one child.
290 if (!dst
->anon_vma
&& anon_vma
!= src
->anon_vma
&&
291 anon_vma
->degree
< 2)
292 dst
->anon_vma
= anon_vma
;
295 dst
->anon_vma
->degree
++;
296 unlock_anon_vma_root(root
);
301 * dst->anon_vma is dropped here otherwise its degree can be incorrectly
302 * decremented in unlink_anon_vmas().
303 * We can safely do this because callers of anon_vma_clone() don't care
304 * about dst->anon_vma if anon_vma_clone() failed.
306 dst
->anon_vma
= NULL
;
307 unlink_anon_vmas(dst
);
312 * Attach vma to its own anon_vma, as well as to the anon_vmas that
313 * the corresponding VMA in the parent process is attached to.
314 * Returns 0 on success, non-zero on failure.
316 int anon_vma_fork(struct vm_area_struct
*vma
, struct vm_area_struct
*pvma
)
318 struct anon_vma_chain
*avc
;
319 struct anon_vma
*anon_vma
;
322 /* Don't bother if the parent process has no anon_vma here. */
326 /* Drop inherited anon_vma, we'll reuse existing or allocate new. */
327 vma
->anon_vma
= NULL
;
330 * First, attach the new VMA to the parent VMA's anon_vmas,
331 * so rmap can find non-COWed pages in child processes.
333 error
= anon_vma_clone(vma
, pvma
);
337 /* An existing anon_vma has been reused, all done then. */
341 /* Then add our own anon_vma. */
342 anon_vma
= anon_vma_alloc();
345 avc
= anon_vma_chain_alloc(GFP_KERNEL
);
347 goto out_error_free_anon_vma
;
350 * The root anon_vma's spinlock is the lock actually used when we
351 * lock any of the anon_vmas in this anon_vma tree.
353 anon_vma
->root
= pvma
->anon_vma
->root
;
354 anon_vma
->parent
= pvma
->anon_vma
;
356 * With refcounts, an anon_vma can stay around longer than the
357 * process it belongs to. The root anon_vma needs to be pinned until
358 * this anon_vma is freed, because the lock lives in the root.
360 get_anon_vma(anon_vma
->root
);
361 /* Mark this anon_vma as the one where our new (COWed) pages go. */
362 vma
->anon_vma
= anon_vma
;
363 anon_vma_lock_write(anon_vma
);
364 anon_vma_chain_link(vma
, avc
, anon_vma
);
365 anon_vma
->parent
->degree
++;
366 anon_vma_unlock_write(anon_vma
);
370 out_error_free_anon_vma
:
371 put_anon_vma(anon_vma
);
373 unlink_anon_vmas(vma
);
377 void unlink_anon_vmas(struct vm_area_struct
*vma
)
379 struct anon_vma_chain
*avc
, *next
;
380 struct anon_vma
*root
= NULL
;
383 * Unlink each anon_vma chained to the VMA. This list is ordered
384 * from newest to oldest, ensuring the root anon_vma gets freed last.
386 list_for_each_entry_safe(avc
, next
, &vma
->anon_vma_chain
, same_vma
) {
387 struct anon_vma
*anon_vma
= avc
->anon_vma
;
389 root
= lock_anon_vma_root(root
, anon_vma
);
390 anon_vma_interval_tree_remove(avc
, &anon_vma
->rb_root
);
393 * Leave empty anon_vmas on the list - we'll need
394 * to free them outside the lock.
396 if (RB_EMPTY_ROOT(&anon_vma
->rb_root
.rb_root
)) {
397 anon_vma
->parent
->degree
--;
401 list_del(&avc
->same_vma
);
402 anon_vma_chain_free(avc
);
405 vma
->anon_vma
->degree
--;
406 unlock_anon_vma_root(root
);
409 * Iterate the list once more, it now only contains empty and unlinked
410 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
411 * needing to write-acquire the anon_vma->root->rwsem.
413 list_for_each_entry_safe(avc
, next
, &vma
->anon_vma_chain
, same_vma
) {
414 struct anon_vma
*anon_vma
= avc
->anon_vma
;
416 VM_WARN_ON(anon_vma
->degree
);
417 put_anon_vma(anon_vma
);
419 list_del(&avc
->same_vma
);
420 anon_vma_chain_free(avc
);
424 static void anon_vma_ctor(void *data
)
426 struct anon_vma
*anon_vma
= data
;
428 init_rwsem(&anon_vma
->rwsem
);
429 atomic_set(&anon_vma
->refcount
, 0);
430 anon_vma
->rb_root
= RB_ROOT_CACHED
;
433 void __init
anon_vma_init(void)
435 anon_vma_cachep
= kmem_cache_create("anon_vma", sizeof(struct anon_vma
),
436 0, SLAB_TYPESAFE_BY_RCU
|SLAB_PANIC
|SLAB_ACCOUNT
,
438 anon_vma_chain_cachep
= KMEM_CACHE(anon_vma_chain
,
439 SLAB_PANIC
|SLAB_ACCOUNT
);
443 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
445 * Since there is no serialization what so ever against page_remove_rmap()
446 * the best this function can do is return a locked anon_vma that might
447 * have been relevant to this page.
449 * The page might have been remapped to a different anon_vma or the anon_vma
450 * returned may already be freed (and even reused).
452 * In case it was remapped to a different anon_vma, the new anon_vma will be a
453 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
454 * ensure that any anon_vma obtained from the page will still be valid for as
455 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
457 * All users of this function must be very careful when walking the anon_vma
458 * chain and verify that the page in question is indeed mapped in it
459 * [ something equivalent to page_mapped_in_vma() ].
461 * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap()
462 * that the anon_vma pointer from page->mapping is valid if there is a
463 * mapcount, we can dereference the anon_vma after observing those.
465 struct anon_vma
*page_get_anon_vma(struct page
*page
)
467 struct anon_vma
*anon_vma
= NULL
;
468 unsigned long anon_mapping
;
471 anon_mapping
= (unsigned long)READ_ONCE(page
->mapping
);
472 if ((anon_mapping
& PAGE_MAPPING_FLAGS
) != PAGE_MAPPING_ANON
)
474 if (!page_mapped(page
))
477 anon_vma
= (struct anon_vma
*) (anon_mapping
- PAGE_MAPPING_ANON
);
478 if (!atomic_inc_not_zero(&anon_vma
->refcount
)) {
484 * If this page is still mapped, then its anon_vma cannot have been
485 * freed. But if it has been unmapped, we have no security against the
486 * anon_vma structure being freed and reused (for another anon_vma:
487 * SLAB_TYPESAFE_BY_RCU guarantees that - so the atomic_inc_not_zero()
488 * above cannot corrupt).
490 if (!page_mapped(page
)) {
492 put_anon_vma(anon_vma
);
502 * Similar to page_get_anon_vma() except it locks the anon_vma.
504 * Its a little more complex as it tries to keep the fast path to a single
505 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
506 * reference like with page_get_anon_vma() and then block on the mutex.
508 struct anon_vma
*page_lock_anon_vma_read(struct page
*page
)
510 struct anon_vma
*anon_vma
= NULL
;
511 struct anon_vma
*root_anon_vma
;
512 unsigned long anon_mapping
;
515 anon_mapping
= (unsigned long)READ_ONCE(page
->mapping
);
516 if ((anon_mapping
& PAGE_MAPPING_FLAGS
) != PAGE_MAPPING_ANON
)
518 if (!page_mapped(page
))
521 anon_vma
= (struct anon_vma
*) (anon_mapping
- PAGE_MAPPING_ANON
);
522 root_anon_vma
= READ_ONCE(anon_vma
->root
);
523 if (down_read_trylock(&root_anon_vma
->rwsem
)) {
525 * If the page is still mapped, then this anon_vma is still
526 * its anon_vma, and holding the mutex ensures that it will
527 * not go away, see anon_vma_free().
529 if (!page_mapped(page
)) {
530 up_read(&root_anon_vma
->rwsem
);
536 /* trylock failed, we got to sleep */
537 if (!atomic_inc_not_zero(&anon_vma
->refcount
)) {
542 if (!page_mapped(page
)) {
544 put_anon_vma(anon_vma
);
548 /* we pinned the anon_vma, its safe to sleep */
550 anon_vma_lock_read(anon_vma
);
552 if (atomic_dec_and_test(&anon_vma
->refcount
)) {
554 * Oops, we held the last refcount, release the lock
555 * and bail -- can't simply use put_anon_vma() because
556 * we'll deadlock on the anon_vma_lock_write() recursion.
558 anon_vma_unlock_read(anon_vma
);
559 __put_anon_vma(anon_vma
);
570 void page_unlock_anon_vma_read(struct anon_vma
*anon_vma
)
572 anon_vma_unlock_read(anon_vma
);
575 #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
577 * Flush TLB entries for recently unmapped pages from remote CPUs. It is
578 * important if a PTE was dirty when it was unmapped that it's flushed
579 * before any IO is initiated on the page to prevent lost writes. Similarly,
580 * it must be flushed before freeing to prevent data leakage.
582 void try_to_unmap_flush(void)
584 struct tlbflush_unmap_batch
*tlb_ubc
= ¤t
->tlb_ubc
;
586 if (!tlb_ubc
->flush_required
)
589 arch_tlbbatch_flush(&tlb_ubc
->arch
);
590 tlb_ubc
->flush_required
= false;
591 tlb_ubc
->writable
= false;
594 /* Flush iff there are potentially writable TLB entries that can race with IO */
595 void try_to_unmap_flush_dirty(void)
597 struct tlbflush_unmap_batch
*tlb_ubc
= ¤t
->tlb_ubc
;
599 if (tlb_ubc
->writable
)
600 try_to_unmap_flush();
603 static void set_tlb_ubc_flush_pending(struct mm_struct
*mm
, bool writable
)
605 struct tlbflush_unmap_batch
*tlb_ubc
= ¤t
->tlb_ubc
;
607 arch_tlbbatch_add_mm(&tlb_ubc
->arch
, mm
);
608 tlb_ubc
->flush_required
= true;
611 * Ensure compiler does not re-order the setting of tlb_flush_batched
612 * before the PTE is cleared.
615 mm
->tlb_flush_batched
= true;
618 * If the PTE was dirty then it's best to assume it's writable. The
619 * caller must use try_to_unmap_flush_dirty() or try_to_unmap_flush()
620 * before the page is queued for IO.
623 tlb_ubc
->writable
= true;
627 * Returns true if the TLB flush should be deferred to the end of a batch of
628 * unmap operations to reduce IPIs.
630 static bool should_defer_flush(struct mm_struct
*mm
, enum ttu_flags flags
)
632 bool should_defer
= false;
634 if (!(flags
& TTU_BATCH_FLUSH
))
637 /* If remote CPUs need to be flushed then defer batch the flush */
638 if (cpumask_any_but(mm_cpumask(mm
), get_cpu()) < nr_cpu_ids
)
646 * Reclaim unmaps pages under the PTL but do not flush the TLB prior to
647 * releasing the PTL if TLB flushes are batched. It's possible for a parallel
648 * operation such as mprotect or munmap to race between reclaim unmapping
649 * the page and flushing the page. If this race occurs, it potentially allows
650 * access to data via a stale TLB entry. Tracking all mm's that have TLB
651 * batching in flight would be expensive during reclaim so instead track
652 * whether TLB batching occurred in the past and if so then do a flush here
653 * if required. This will cost one additional flush per reclaim cycle paid
654 * by the first operation at risk such as mprotect and mumap.
656 * This must be called under the PTL so that an access to tlb_flush_batched
657 * that is potentially a "reclaim vs mprotect/munmap/etc" race will synchronise
660 void flush_tlb_batched_pending(struct mm_struct
*mm
)
662 if (mm
->tlb_flush_batched
) {
666 * Do not allow the compiler to re-order the clearing of
667 * tlb_flush_batched before the tlb is flushed.
670 mm
->tlb_flush_batched
= false;
674 static void set_tlb_ubc_flush_pending(struct mm_struct
*mm
, bool writable
)
678 static bool should_defer_flush(struct mm_struct
*mm
, enum ttu_flags flags
)
682 #endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
685 * At what user virtual address is page expected in vma?
686 * Caller should check the page is actually part of the vma.
688 unsigned long page_address_in_vma(struct page
*page
, struct vm_area_struct
*vma
)
690 if (PageAnon(page
)) {
691 struct anon_vma
*page__anon_vma
= page_anon_vma(page
);
693 * Note: swapoff's unuse_vma() is more efficient with this
694 * check, and needs it to match anon_vma when KSM is active.
696 if (!vma
->anon_vma
|| !page__anon_vma
||
697 vma
->anon_vma
->root
!= page__anon_vma
->root
)
699 } else if (!vma
->vm_file
) {
701 } else if (vma
->vm_file
->f_mapping
!= compound_head(page
)->mapping
) {
705 return vma_address(page
, vma
);
708 pmd_t
*mm_find_pmd(struct mm_struct
*mm
, unsigned long address
)
716 pgd
= pgd_offset(mm
, address
);
717 if (!pgd_present(*pgd
))
720 p4d
= p4d_offset(pgd
, address
);
721 if (!p4d_present(*p4d
))
724 pud
= pud_offset(p4d
, address
);
725 if (!pud_present(*pud
))
728 pmd
= pmd_offset(pud
, address
);
730 * Some THP functions use the sequence pmdp_huge_clear_flush(), set_pmd_at()
731 * without holding anon_vma lock for write. So when looking for a
732 * genuine pmde (in which to find pte), test present and !THP together.
736 if (!pmd_present(pmde
) || pmd_trans_huge(pmde
))
742 struct page_referenced_arg
{
745 unsigned long vm_flags
;
746 struct mem_cgroup
*memcg
;
749 * arg: page_referenced_arg will be passed
751 static bool page_referenced_one(struct page
*page
, struct vm_area_struct
*vma
,
752 unsigned long address
, void *arg
)
754 struct page_referenced_arg
*pra
= arg
;
755 struct page_vma_mapped_walk pvmw
= {
762 while (page_vma_mapped_walk(&pvmw
)) {
763 address
= pvmw
.address
;
765 if (vma
->vm_flags
& VM_LOCKED
) {
766 page_vma_mapped_walk_done(&pvmw
);
767 pra
->vm_flags
|= VM_LOCKED
;
768 return false; /* To break the loop */
772 if (ptep_clear_flush_young_notify(vma
, address
,
775 * Don't treat a reference through
776 * a sequentially read mapping as such.
777 * If the page has been used in another mapping,
778 * we will catch it; if this other mapping is
779 * already gone, the unmap path will have set
780 * PG_referenced or activated the page.
782 if (likely(!(vma
->vm_flags
& VM_SEQ_READ
)))
785 } else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE
)) {
786 if (pmdp_clear_flush_young_notify(vma
, address
,
790 /* unexpected pmd-mapped page? */
798 clear_page_idle(page
);
799 if (test_and_clear_page_young(page
))
804 pra
->vm_flags
|= vma
->vm_flags
;
808 return false; /* To break the loop */
813 static bool invalid_page_referenced_vma(struct vm_area_struct
*vma
, void *arg
)
815 struct page_referenced_arg
*pra
= arg
;
816 struct mem_cgroup
*memcg
= pra
->memcg
;
818 if (!mm_match_cgroup(vma
->vm_mm
, memcg
))
825 * page_referenced - test if the page was referenced
826 * @page: the page to test
827 * @is_locked: caller holds lock on the page
828 * @memcg: target memory cgroup
829 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
831 * Quick test_and_clear_referenced for all mappings to a page,
832 * returns the number of ptes which referenced the page.
834 int page_referenced(struct page
*page
,
836 struct mem_cgroup
*memcg
,
837 unsigned long *vm_flags
)
840 struct page_referenced_arg pra
= {
841 .mapcount
= total_mapcount(page
),
844 struct rmap_walk_control rwc
= {
845 .rmap_one
= page_referenced_one
,
847 .anon_lock
= page_lock_anon_vma_read
,
854 if (!page_rmapping(page
))
857 if (!is_locked
&& (!PageAnon(page
) || PageKsm(page
))) {
858 we_locked
= trylock_page(page
);
864 * If we are reclaiming on behalf of a cgroup, skip
865 * counting on behalf of references from different
869 rwc
.invalid_vma
= invalid_page_referenced_vma
;
872 rmap_walk(page
, &rwc
);
873 *vm_flags
= pra
.vm_flags
;
878 return pra
.referenced
;
881 static bool page_mkclean_one(struct page
*page
, struct vm_area_struct
*vma
,
882 unsigned long address
, void *arg
)
884 struct page_vma_mapped_walk pvmw
= {
890 struct mmu_notifier_range range
;
894 * We have to assume the worse case ie pmd for invalidation. Note that
895 * the page can not be free from this function.
897 mmu_notifier_range_init(&range
, MMU_NOTIFY_PROTECTION_PAGE
,
898 0, vma
, vma
->vm_mm
, address
,
899 vma_address_end(page
, vma
));
900 mmu_notifier_invalidate_range_start(&range
);
902 while (page_vma_mapped_walk(&pvmw
)) {
905 address
= pvmw
.address
;
908 pte_t
*pte
= pvmw
.pte
;
910 if (!pte_dirty(*pte
) && !pte_write(*pte
))
913 flush_cache_page(vma
, address
, pte_pfn(*pte
));
914 entry
= ptep_clear_flush(vma
, address
, pte
);
915 entry
= pte_wrprotect(entry
);
916 entry
= pte_mkclean(entry
);
917 set_pte_at(vma
->vm_mm
, address
, pte
, entry
);
920 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
921 pmd_t
*pmd
= pvmw
.pmd
;
924 if (!pmd_dirty(*pmd
) && !pmd_write(*pmd
))
927 flush_cache_page(vma
, address
, page_to_pfn(page
));
928 entry
= pmdp_invalidate(vma
, address
, pmd
);
929 entry
= pmd_wrprotect(entry
);
930 entry
= pmd_mkclean(entry
);
931 set_pmd_at(vma
->vm_mm
, address
, pmd
, entry
);
934 /* unexpected pmd-mapped page? */
940 * No need to call mmu_notifier_invalidate_range() as we are
941 * downgrading page table protection not changing it to point
944 * See Documentation/vm/mmu_notifier.rst
950 mmu_notifier_invalidate_range_end(&range
);
955 static bool invalid_mkclean_vma(struct vm_area_struct
*vma
, void *arg
)
957 if (vma
->vm_flags
& VM_SHARED
)
963 int page_mkclean(struct page
*page
)
966 struct address_space
*mapping
;
967 struct rmap_walk_control rwc
= {
968 .arg
= (void *)&cleaned
,
969 .rmap_one
= page_mkclean_one
,
970 .invalid_vma
= invalid_mkclean_vma
,
973 BUG_ON(!PageLocked(page
));
975 if (!page_mapped(page
))
978 mapping
= page_mapping(page
);
982 rmap_walk(page
, &rwc
);
986 EXPORT_SYMBOL_GPL(page_mkclean
);
989 * page_move_anon_rmap - move a page to our anon_vma
990 * @page: the page to move to our anon_vma
991 * @vma: the vma the page belongs to
993 * When a page belongs exclusively to one process after a COW event,
994 * that page can be moved into the anon_vma that belongs to just that
995 * process, so the rmap code will not search the parent or sibling
998 void page_move_anon_rmap(struct page
*page
, struct vm_area_struct
*vma
)
1000 struct anon_vma
*anon_vma
= vma
->anon_vma
;
1002 page
= compound_head(page
);
1004 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
1005 VM_BUG_ON_VMA(!anon_vma
, vma
);
1007 anon_vma
= (void *) anon_vma
+ PAGE_MAPPING_ANON
;
1009 * Ensure that anon_vma and the PAGE_MAPPING_ANON bit are written
1010 * simultaneously, so a concurrent reader (eg page_referenced()'s
1011 * PageAnon()) will not see one without the other.
1013 WRITE_ONCE(page
->mapping
, (struct address_space
*) anon_vma
);
1017 * __page_set_anon_rmap - set up new anonymous rmap
1018 * @page: Page or Hugepage to add to rmap
1019 * @vma: VM area to add page to.
1020 * @address: User virtual address of the mapping
1021 * @exclusive: the page is exclusively owned by the current process
1023 static void __page_set_anon_rmap(struct page
*page
,
1024 struct vm_area_struct
*vma
, unsigned long address
, int exclusive
)
1026 struct anon_vma
*anon_vma
= vma
->anon_vma
;
1034 * If the page isn't exclusively mapped into this vma,
1035 * we must use the _oldest_ possible anon_vma for the
1039 anon_vma
= anon_vma
->root
;
1041 anon_vma
= (void *) anon_vma
+ PAGE_MAPPING_ANON
;
1042 page
->mapping
= (struct address_space
*) anon_vma
;
1043 page
->index
= linear_page_index(vma
, address
);
1047 * __page_check_anon_rmap - sanity check anonymous rmap addition
1048 * @page: the page to add the mapping to
1049 * @vma: the vm area in which the mapping is added
1050 * @address: the user virtual address mapped
1052 static void __page_check_anon_rmap(struct page
*page
,
1053 struct vm_area_struct
*vma
, unsigned long address
)
1055 #ifdef CONFIG_DEBUG_VM
1057 * The page's anon-rmap details (mapping and index) are guaranteed to
1058 * be set up correctly at this point.
1060 * We have exclusion against page_add_anon_rmap because the caller
1061 * always holds the page locked, except if called from page_dup_rmap,
1062 * in which case the page is already known to be setup.
1064 * We have exclusion against page_add_new_anon_rmap because those pages
1065 * are initially only visible via the pagetables, and the pte is locked
1066 * over the call to page_add_new_anon_rmap.
1068 BUG_ON(page_anon_vma(page
)->root
!= vma
->anon_vma
->root
);
1069 BUG_ON(page_to_pgoff(page
) != linear_page_index(vma
, address
));
1074 * page_add_anon_rmap - add pte mapping to an anonymous page
1075 * @page: the page to add the mapping to
1076 * @vma: the vm area in which the mapping is added
1077 * @address: the user virtual address mapped
1078 * @compound: charge the page as compound or small page
1080 * The caller needs to hold the pte lock, and the page must be locked in
1081 * the anon_vma case: to serialize mapping,index checking after setting,
1082 * and to ensure that PageAnon is not being upgraded racily to PageKsm
1083 * (but PageKsm is never downgraded to PageAnon).
1085 void page_add_anon_rmap(struct page
*page
,
1086 struct vm_area_struct
*vma
, unsigned long address
, bool compound
)
1088 do_page_add_anon_rmap(page
, vma
, address
, compound
? RMAP_COMPOUND
: 0);
1092 * Special version of the above for do_swap_page, which often runs
1093 * into pages that are exclusively owned by the current process.
1094 * Everybody else should continue to use page_add_anon_rmap above.
1096 void do_page_add_anon_rmap(struct page
*page
,
1097 struct vm_area_struct
*vma
, unsigned long address
, int flags
)
1099 bool compound
= flags
& RMAP_COMPOUND
;
1104 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
1105 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
1106 mapcount
= compound_mapcount_ptr(page
);
1107 first
= atomic_inc_and_test(mapcount
);
1109 first
= atomic_inc_and_test(&page
->_mapcount
);
1113 int nr
= compound
? hpage_nr_pages(page
) : 1;
1115 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1116 * these counters are not modified in interrupt context, and
1117 * pte lock(a spinlock) is held, which implies preemption
1121 __inc_node_page_state(page
, NR_ANON_THPS
);
1122 __mod_node_page_state(page_pgdat(page
), NR_ANON_MAPPED
, nr
);
1124 if (unlikely(PageKsm(page
)))
1127 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
1129 /* address might be in next vma when migration races vma_adjust */
1131 __page_set_anon_rmap(page
, vma
, address
,
1132 flags
& RMAP_EXCLUSIVE
);
1134 __page_check_anon_rmap(page
, vma
, address
);
1138 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1139 * @page: the page to add the mapping to
1140 * @vma: the vm area in which the mapping is added
1141 * @address: the user virtual address mapped
1142 * @compound: charge the page as compound or small page
1144 * Same as page_add_anon_rmap but must only be called on *new* pages.
1145 * This means the inc-and-test can be bypassed.
1146 * Page does not have to be locked.
1148 void page_add_new_anon_rmap(struct page
*page
,
1149 struct vm_area_struct
*vma
, unsigned long address
, bool compound
)
1151 int nr
= compound
? hpage_nr_pages(page
) : 1;
1153 VM_BUG_ON_VMA(address
< vma
->vm_start
|| address
>= vma
->vm_end
, vma
);
1154 __SetPageSwapBacked(page
);
1156 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
1157 /* increment count (starts at -1) */
1158 atomic_set(compound_mapcount_ptr(page
), 0);
1159 __inc_node_page_state(page
, NR_ANON_THPS
);
1161 /* Anon THP always mapped first with PMD */
1162 VM_BUG_ON_PAGE(PageTransCompound(page
), page
);
1163 /* increment count (starts at -1) */
1164 atomic_set(&page
->_mapcount
, 0);
1166 __mod_node_page_state(page_pgdat(page
), NR_ANON_MAPPED
, nr
);
1167 __page_set_anon_rmap(page
, vma
, address
, 1);
1171 * page_add_file_rmap - add pte mapping to a file page
1172 * @page: the page to add the mapping to
1173 * @compound: charge the page as compound or small page
1175 * The caller needs to hold the pte lock.
1177 void page_add_file_rmap(struct page
*page
, bool compound
)
1181 VM_BUG_ON_PAGE(compound
&& !PageTransHuge(page
), page
);
1182 lock_page_memcg(page
);
1183 if (compound
&& PageTransHuge(page
)) {
1184 for (i
= 0, nr
= 0; i
< HPAGE_PMD_NR
; i
++) {
1185 if (atomic_inc_and_test(&page
[i
]._mapcount
))
1188 if (!atomic_inc_and_test(compound_mapcount_ptr(page
)))
1190 if (PageSwapBacked(page
))
1191 __inc_node_page_state(page
, NR_SHMEM_PMDMAPPED
);
1193 __inc_node_page_state(page
, NR_FILE_PMDMAPPED
);
1195 if (PageTransCompound(page
) && page_mapping(page
)) {
1196 VM_WARN_ON_ONCE(!PageLocked(page
));
1198 SetPageDoubleMap(compound_head(page
));
1199 if (PageMlocked(page
))
1200 clear_page_mlock(compound_head(page
));
1202 if (!atomic_inc_and_test(&page
->_mapcount
))
1205 __mod_lruvec_page_state(page
, NR_FILE_MAPPED
, nr
);
1207 unlock_page_memcg(page
);
1210 static void page_remove_file_rmap(struct page
*page
, bool compound
)
1214 VM_BUG_ON_PAGE(compound
&& !PageHead(page
), page
);
1215 lock_page_memcg(page
);
1217 /* Hugepages are not counted in NR_FILE_MAPPED for now. */
1218 if (unlikely(PageHuge(page
))) {
1219 /* hugetlb pages are always mapped with pmds */
1220 atomic_dec(compound_mapcount_ptr(page
));
1224 /* page still mapped by someone else? */
1225 if (compound
&& PageTransHuge(page
)) {
1226 for (i
= 0, nr
= 0; i
< HPAGE_PMD_NR
; i
++) {
1227 if (atomic_add_negative(-1, &page
[i
]._mapcount
))
1230 if (!atomic_add_negative(-1, compound_mapcount_ptr(page
)))
1232 if (PageSwapBacked(page
))
1233 __dec_node_page_state(page
, NR_SHMEM_PMDMAPPED
);
1235 __dec_node_page_state(page
, NR_FILE_PMDMAPPED
);
1237 if (!atomic_add_negative(-1, &page
->_mapcount
))
1242 * We use the irq-unsafe __{inc|mod}_lruvec_page_state because
1243 * these counters are not modified in interrupt context, and
1244 * pte lock(a spinlock) is held, which implies preemption disabled.
1246 __mod_lruvec_page_state(page
, NR_FILE_MAPPED
, -nr
);
1248 if (unlikely(PageMlocked(page
)))
1249 clear_page_mlock(page
);
1251 unlock_page_memcg(page
);
1254 static void page_remove_anon_compound_rmap(struct page
*page
)
1258 if (!atomic_add_negative(-1, compound_mapcount_ptr(page
)))
1261 /* Hugepages are not counted in NR_ANON_PAGES for now. */
1262 if (unlikely(PageHuge(page
)))
1265 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE
))
1268 __dec_node_page_state(page
, NR_ANON_THPS
);
1270 if (TestClearPageDoubleMap(page
)) {
1272 * Subpages can be mapped with PTEs too. Check how many of
1273 * themi are still mapped.
1275 for (i
= 0, nr
= 0; i
< HPAGE_PMD_NR
; i
++) {
1276 if (atomic_add_negative(-1, &page
[i
]._mapcount
))
1283 if (unlikely(PageMlocked(page
)))
1284 clear_page_mlock(page
);
1287 __mod_node_page_state(page_pgdat(page
), NR_ANON_MAPPED
, -nr
);
1288 deferred_split_huge_page(page
);
1293 * page_remove_rmap - take down pte mapping from a page
1294 * @page: page to remove mapping from
1295 * @compound: uncharge the page as compound or small page
1297 * The caller needs to hold the pte lock.
1299 void page_remove_rmap(struct page
*page
, bool compound
)
1301 if (!PageAnon(page
))
1302 return page_remove_file_rmap(page
, compound
);
1305 return page_remove_anon_compound_rmap(page
);
1307 /* page still mapped by someone else? */
1308 if (!atomic_add_negative(-1, &page
->_mapcount
))
1312 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1313 * these counters are not modified in interrupt context, and
1314 * pte lock(a spinlock) is held, which implies preemption disabled.
1316 __dec_node_page_state(page
, NR_ANON_MAPPED
);
1318 if (unlikely(PageMlocked(page
)))
1319 clear_page_mlock(page
);
1321 if (PageTransCompound(page
))
1322 deferred_split_huge_page(compound_head(page
));
1325 * It would be tidy to reset the PageAnon mapping here,
1326 * but that might overwrite a racing page_add_anon_rmap
1327 * which increments mapcount after us but sets mapping
1328 * before us: so leave the reset to free_unref_page,
1329 * and remember that it's only reliable while mapped.
1330 * Leaving it set also helps swapoff to reinstate ptes
1331 * faster for those pages still in swapcache.
1336 * @arg: enum ttu_flags will be passed to this argument
1338 static bool try_to_unmap_one(struct page
*page
, struct vm_area_struct
*vma
,
1339 unsigned long address
, void *arg
)
1341 struct mm_struct
*mm
= vma
->vm_mm
;
1342 struct page_vma_mapped_walk pvmw
= {
1348 struct page
*subpage
;
1350 struct mmu_notifier_range range
;
1351 enum ttu_flags flags
= (enum ttu_flags
)arg
;
1354 * When racing against e.g. zap_pte_range() on another cpu,
1355 * in between its ptep_get_and_clear_full() and page_remove_rmap(),
1356 * try_to_unmap() may return false when it is about to become true,
1357 * if page table locking is skipped: use TTU_SYNC to wait for that.
1359 if (flags
& TTU_SYNC
)
1360 pvmw
.flags
= PVMW_SYNC
;
1362 /* munlock has nothing to gain from examining un-locked vmas */
1363 if ((flags
& TTU_MUNLOCK
) && !(vma
->vm_flags
& VM_LOCKED
))
1366 if (IS_ENABLED(CONFIG_MIGRATION
) && (flags
& TTU_MIGRATION
) &&
1367 is_zone_device_page(page
) && !is_device_private_page(page
))
1370 if (flags
& TTU_SPLIT_HUGE_PMD
) {
1371 split_huge_pmd_address(vma
, address
,
1372 flags
& TTU_SPLIT_FREEZE
, page
);
1376 * For THP, we have to assume the worse case ie pmd for invalidation.
1377 * For hugetlb, it could be much worse if we need to do pud
1378 * invalidation in the case of pmd sharing.
1380 * Note that the page can not be free in this function as call of
1381 * try_to_unmap() must hold a reference on the page.
1383 range
.end
= PageKsm(page
) ?
1384 address
+ PAGE_SIZE
: vma_address_end(page
, vma
);
1385 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
, vma
->vm_mm
,
1386 address
, range
.end
);
1387 if (PageHuge(page
)) {
1389 * If sharing is possible, start and end will be adjusted
1392 adjust_range_if_pmd_sharing_possible(vma
, &range
.start
,
1395 mmu_notifier_invalidate_range_start(&range
);
1397 while (page_vma_mapped_walk(&pvmw
)) {
1398 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1399 /* PMD-mapped THP migration entry */
1400 if (!pvmw
.pte
&& (flags
& TTU_MIGRATION
)) {
1401 VM_BUG_ON_PAGE(PageHuge(page
) || !PageTransCompound(page
), page
);
1403 set_pmd_migration_entry(&pvmw
, page
);
1409 * If the page is mlock()d, we cannot swap it out.
1410 * If it's recently referenced (perhaps page_referenced
1411 * skipped over this mm) then we should reactivate it.
1413 if (!(flags
& TTU_IGNORE_MLOCK
)) {
1414 if (vma
->vm_flags
& VM_LOCKED
) {
1415 /* PTE-mapped THP are never mlocked */
1416 if (!PageTransCompound(page
)) {
1418 * Holding pte lock, we do *not* need
1421 mlock_vma_page(page
);
1424 page_vma_mapped_walk_done(&pvmw
);
1427 if (flags
& TTU_MUNLOCK
)
1431 /* Unexpected PMD-mapped THP? */
1432 VM_BUG_ON_PAGE(!pvmw
.pte
, page
);
1434 subpage
= page
- page_to_pfn(page
) + pte_pfn(*pvmw
.pte
);
1435 address
= pvmw
.address
;
1437 if (PageHuge(page
)) {
1438 if (huge_pmd_unshare(mm
, &address
, pvmw
.pte
)) {
1440 * huge_pmd_unshare unmapped an entire PMD
1441 * page. There is no way of knowing exactly
1442 * which PMDs may be cached for this mm, so
1443 * we must flush them all. start/end were
1444 * already adjusted above to cover this range.
1446 flush_cache_range(vma
, range
.start
, range
.end
);
1447 flush_tlb_range(vma
, range
.start
, range
.end
);
1448 mmu_notifier_invalidate_range(mm
, range
.start
,
1452 * The ref count of the PMD page was dropped
1453 * which is part of the way map counting
1454 * is done for shared PMDs. Return 'true'
1455 * here. When there is no other sharing,
1456 * huge_pmd_unshare returns false and we will
1457 * unmap the actual page and drop map count
1460 page_vma_mapped_walk_done(&pvmw
);
1465 if (IS_ENABLED(CONFIG_MIGRATION
) &&
1466 (flags
& TTU_MIGRATION
) &&
1467 is_zone_device_page(page
)) {
1471 pteval
= ptep_get_and_clear(mm
, pvmw
.address
, pvmw
.pte
);
1474 * Store the pfn of the page in a special migration
1475 * pte. do_swap_page() will wait until the migration
1476 * pte is removed and then restart fault handling.
1478 entry
= make_migration_entry(page
, 0);
1479 swp_pte
= swp_entry_to_pte(entry
);
1480 if (pte_soft_dirty(pteval
))
1481 swp_pte
= pte_swp_mksoft_dirty(swp_pte
);
1482 set_pte_at(mm
, pvmw
.address
, pvmw
.pte
, swp_pte
);
1484 * No need to invalidate here it will synchronize on
1485 * against the special swap migration pte.
1487 * The assignment to subpage above was computed from a
1488 * swap PTE which results in an invalid pointer.
1489 * Since only PAGE_SIZE pages can currently be
1490 * migrated, just set it to page. This will need to be
1491 * changed when hugepage migrations to device private
1492 * memory are supported.
1498 if (!(flags
& TTU_IGNORE_ACCESS
)) {
1499 if (ptep_clear_flush_young_notify(vma
, address
,
1502 page_vma_mapped_walk_done(&pvmw
);
1507 /* Nuke the page table entry. */
1508 flush_cache_page(vma
, address
, pte_pfn(*pvmw
.pte
));
1509 if (should_defer_flush(mm
, flags
)) {
1511 * We clear the PTE but do not flush so potentially
1512 * a remote CPU could still be writing to the page.
1513 * If the entry was previously clean then the
1514 * architecture must guarantee that a clear->dirty
1515 * transition on a cached TLB entry is written through
1516 * and traps if the PTE is unmapped.
1518 pteval
= ptep_get_and_clear(mm
, address
, pvmw
.pte
);
1520 set_tlb_ubc_flush_pending(mm
, pte_dirty(pteval
));
1522 pteval
= ptep_clear_flush(vma
, address
, pvmw
.pte
);
1525 /* Move the dirty bit to the page. Now the pte is gone. */
1526 if (pte_dirty(pteval
))
1527 set_page_dirty(page
);
1529 /* Update high watermark before we lower rss */
1530 update_hiwater_rss(mm
);
1532 if (PageHWPoison(page
) && !(flags
& TTU_IGNORE_HWPOISON
)) {
1533 pteval
= swp_entry_to_pte(make_hwpoison_entry(subpage
));
1534 if (PageHuge(page
)) {
1535 hugetlb_count_sub(compound_nr(page
), mm
);
1536 set_huge_swap_pte_at(mm
, address
,
1538 vma_mmu_pagesize(vma
));
1540 dec_mm_counter(mm
, mm_counter(page
));
1541 set_pte_at(mm
, address
, pvmw
.pte
, pteval
);
1544 } else if (pte_unused(pteval
) && !userfaultfd_armed(vma
)) {
1546 * The guest indicated that the page content is of no
1547 * interest anymore. Simply discard the pte, vmscan
1548 * will take care of the rest.
1549 * A future reference will then fault in a new zero
1550 * page. When userfaultfd is active, we must not drop
1551 * this page though, as its main user (postcopy
1552 * migration) will not expect userfaults on already
1555 dec_mm_counter(mm
, mm_counter(page
));
1556 /* We have to invalidate as we cleared the pte */
1557 mmu_notifier_invalidate_range(mm
, address
,
1558 address
+ PAGE_SIZE
);
1559 } else if (IS_ENABLED(CONFIG_MIGRATION
) &&
1560 (flags
& (TTU_MIGRATION
|TTU_SPLIT_FREEZE
))) {
1564 if (arch_unmap_one(mm
, vma
, address
, pteval
) < 0) {
1565 set_pte_at(mm
, address
, pvmw
.pte
, pteval
);
1567 page_vma_mapped_walk_done(&pvmw
);
1572 * Store the pfn of the page in a special migration
1573 * pte. do_swap_page() will wait until the migration
1574 * pte is removed and then restart fault handling.
1576 entry
= make_migration_entry(subpage
,
1578 swp_pte
= swp_entry_to_pte(entry
);
1579 if (pte_soft_dirty(pteval
))
1580 swp_pte
= pte_swp_mksoft_dirty(swp_pte
);
1581 set_pte_at(mm
, address
, pvmw
.pte
, swp_pte
);
1583 * No need to invalidate here it will synchronize on
1584 * against the special swap migration pte.
1586 } else if (PageAnon(page
)) {
1587 swp_entry_t entry
= { .val
= page_private(subpage
) };
1590 * Store the swap location in the pte.
1591 * See handle_pte_fault() ...
1593 if (unlikely(PageSwapBacked(page
) != PageSwapCache(page
))) {
1596 /* We have to invalidate as we cleared the pte */
1597 mmu_notifier_invalidate_range(mm
, address
,
1598 address
+ PAGE_SIZE
);
1599 page_vma_mapped_walk_done(&pvmw
);
1603 /* MADV_FREE page check */
1604 if (!PageSwapBacked(page
)) {
1605 int ref_count
, map_count
;
1608 * Synchronize with gup_pte_range():
1609 * - clear PTE; barrier; read refcount
1610 * - inc refcount; barrier; read PTE
1614 ref_count
= page_ref_count(page
);
1615 map_count
= page_mapcount(page
);
1618 * Order reads for page refcount and dirty flag
1619 * (see comments in __remove_mapping()).
1624 * The only page refs must be one from isolation
1625 * plus the rmap(s) (dropped by discard:).
1627 if (ref_count
== 1 + map_count
&&
1629 /* Invalidate as we cleared the pte */
1630 mmu_notifier_invalidate_range(mm
,
1631 address
, address
+ PAGE_SIZE
);
1632 dec_mm_counter(mm
, MM_ANONPAGES
);
1637 * If the page was redirtied, it cannot be
1638 * discarded. Remap the page to page table.
1640 set_pte_at(mm
, address
, pvmw
.pte
, pteval
);
1641 SetPageSwapBacked(page
);
1643 page_vma_mapped_walk_done(&pvmw
);
1647 if (swap_duplicate(entry
) < 0) {
1648 set_pte_at(mm
, address
, pvmw
.pte
, pteval
);
1650 page_vma_mapped_walk_done(&pvmw
);
1653 if (arch_unmap_one(mm
, vma
, address
, pteval
) < 0) {
1654 set_pte_at(mm
, address
, pvmw
.pte
, pteval
);
1656 page_vma_mapped_walk_done(&pvmw
);
1659 if (list_empty(&mm
->mmlist
)) {
1660 spin_lock(&mmlist_lock
);
1661 if (list_empty(&mm
->mmlist
))
1662 list_add(&mm
->mmlist
, &init_mm
.mmlist
);
1663 spin_unlock(&mmlist_lock
);
1665 dec_mm_counter(mm
, MM_ANONPAGES
);
1666 inc_mm_counter(mm
, MM_SWAPENTS
);
1667 swp_pte
= swp_entry_to_pte(entry
);
1668 if (pte_soft_dirty(pteval
))
1669 swp_pte
= pte_swp_mksoft_dirty(swp_pte
);
1670 set_pte_at(mm
, address
, pvmw
.pte
, swp_pte
);
1671 /* Invalidate as we cleared the pte */
1672 mmu_notifier_invalidate_range(mm
, address
,
1673 address
+ PAGE_SIZE
);
1676 * This is a locked file-backed page, thus it cannot
1677 * be removed from the page cache and replaced by a new
1678 * page before mmu_notifier_invalidate_range_end, so no
1679 * concurrent thread might update its page table to
1680 * point at new page while a device still is using this
1683 * See Documentation/vm/mmu_notifier.rst
1685 dec_mm_counter(mm
, mm_counter_file(page
));
1689 * No need to call mmu_notifier_invalidate_range() it has be
1690 * done above for all cases requiring it to happen under page
1691 * table lock before mmu_notifier_invalidate_range_end()
1693 * See Documentation/vm/mmu_notifier.rst
1695 page_remove_rmap(subpage
, PageHuge(page
));
1699 mmu_notifier_invalidate_range_end(&range
);
1704 bool is_vma_temporary_stack(struct vm_area_struct
*vma
)
1706 int maybe_stack
= vma
->vm_flags
& (VM_GROWSDOWN
| VM_GROWSUP
);
1711 if ((vma
->vm_flags
& VM_STACK_INCOMPLETE_SETUP
) ==
1712 VM_STACK_INCOMPLETE_SETUP
)
1718 static bool invalid_migration_vma(struct vm_area_struct
*vma
, void *arg
)
1720 return is_vma_temporary_stack(vma
);
1723 static int page_not_mapped(struct page
*page
)
1725 return !page_mapped(page
);
1729 * try_to_unmap - try to remove all page table mappings to a page
1730 * @page: the page to get unmapped
1731 * @flags: action and flags
1733 * Tries to remove all the page table entries which are mapping this
1734 * page, used in the pageout path. Caller must hold the page lock.
1736 * If unmap is successful, return true. Otherwise, false.
1738 bool try_to_unmap(struct page
*page
, enum ttu_flags flags
)
1740 struct rmap_walk_control rwc
= {
1741 .rmap_one
= try_to_unmap_one
,
1742 .arg
= (void *)flags
,
1743 .done
= page_not_mapped
,
1744 .anon_lock
= page_lock_anon_vma_read
,
1748 * During exec, a temporary VMA is setup and later moved.
1749 * The VMA is moved under the anon_vma lock but not the
1750 * page tables leading to a race where migration cannot
1751 * find the migration ptes. Rather than increasing the
1752 * locking requirements of exec(), migration skips
1753 * temporary VMAs until after exec() completes.
1755 if ((flags
& (TTU_MIGRATION
|TTU_SPLIT_FREEZE
))
1756 && !PageKsm(page
) && PageAnon(page
))
1757 rwc
.invalid_vma
= invalid_migration_vma
;
1759 if (flags
& TTU_RMAP_LOCKED
)
1760 rmap_walk_locked(page
, &rwc
);
1762 rmap_walk(page
, &rwc
);
1765 * When racing against e.g. zap_pte_range() on another cpu,
1766 * in between its ptep_get_and_clear_full() and page_remove_rmap(),
1767 * try_to_unmap() may return false when it is about to become true,
1768 * if page table locking is skipped: use TTU_SYNC to wait for that.
1770 return !page_mapcount(page
);
1774 * try_to_munlock - try to munlock a page
1775 * @page: the page to be munlocked
1777 * Called from munlock code. Checks all of the VMAs mapping the page
1778 * to make sure nobody else has this page mlocked. The page will be
1779 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1782 void try_to_munlock(struct page
*page
)
1784 struct rmap_walk_control rwc
= {
1785 .rmap_one
= try_to_unmap_one
,
1786 .arg
= (void *)TTU_MUNLOCK
,
1787 .done
= page_not_mapped
,
1788 .anon_lock
= page_lock_anon_vma_read
,
1792 VM_BUG_ON_PAGE(!PageLocked(page
) || PageLRU(page
), page
);
1793 VM_BUG_ON_PAGE(PageCompound(page
) && PageDoubleMap(page
), page
);
1795 rmap_walk(page
, &rwc
);
1798 void __put_anon_vma(struct anon_vma
*anon_vma
)
1800 struct anon_vma
*root
= anon_vma
->root
;
1802 anon_vma_free(anon_vma
);
1803 if (root
!= anon_vma
&& atomic_dec_and_test(&root
->refcount
))
1804 anon_vma_free(root
);
1807 static struct anon_vma
*rmap_walk_anon_lock(struct page
*page
,
1808 struct rmap_walk_control
*rwc
)
1810 struct anon_vma
*anon_vma
;
1813 return rwc
->anon_lock(page
);
1816 * Note: remove_migration_ptes() cannot use page_lock_anon_vma_read()
1817 * because that depends on page_mapped(); but not all its usages
1818 * are holding mmap_sem. Users without mmap_sem are required to
1819 * take a reference count to prevent the anon_vma disappearing
1821 anon_vma
= page_anon_vma(page
);
1825 anon_vma_lock_read(anon_vma
);
1830 * rmap_walk_anon - do something to anonymous page using the object-based
1832 * @page: the page to be handled
1833 * @rwc: control variable according to each walk type
1835 * Find all the mappings of a page using the mapping pointer and the vma chains
1836 * contained in the anon_vma struct it points to.
1838 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1839 * where the page was found will be held for write. So, we won't recheck
1840 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1843 static void rmap_walk_anon(struct page
*page
, struct rmap_walk_control
*rwc
,
1846 struct anon_vma
*anon_vma
;
1847 pgoff_t pgoff_start
, pgoff_end
;
1848 struct anon_vma_chain
*avc
;
1851 anon_vma
= page_anon_vma(page
);
1852 /* anon_vma disappear under us? */
1853 VM_BUG_ON_PAGE(!anon_vma
, page
);
1855 anon_vma
= rmap_walk_anon_lock(page
, rwc
);
1860 pgoff_start
= page_to_pgoff(page
);
1861 pgoff_end
= pgoff_start
+ hpage_nr_pages(page
) - 1;
1862 anon_vma_interval_tree_foreach(avc
, &anon_vma
->rb_root
,
1863 pgoff_start
, pgoff_end
) {
1864 struct vm_area_struct
*vma
= avc
->vma
;
1865 unsigned long address
= vma_address(page
, vma
);
1867 VM_BUG_ON_VMA(address
== -EFAULT
, vma
);
1870 if (rwc
->invalid_vma
&& rwc
->invalid_vma(vma
, rwc
->arg
))
1873 if (!rwc
->rmap_one(page
, vma
, address
, rwc
->arg
))
1875 if (rwc
->done
&& rwc
->done(page
))
1880 anon_vma_unlock_read(anon_vma
);
1884 * rmap_walk_file - do something to file page using the object-based rmap method
1885 * @page: the page to be handled
1886 * @rwc: control variable according to each walk type
1888 * Find all the mappings of a page using the mapping pointer and the vma chains
1889 * contained in the address_space struct it points to.
1891 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1892 * where the page was found will be held for write. So, we won't recheck
1893 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1896 static void rmap_walk_file(struct page
*page
, struct rmap_walk_control
*rwc
,
1899 struct address_space
*mapping
= page_mapping(page
);
1900 pgoff_t pgoff_start
, pgoff_end
;
1901 struct vm_area_struct
*vma
;
1904 * The page lock not only makes sure that page->mapping cannot
1905 * suddenly be NULLified by truncation, it makes sure that the
1906 * structure at mapping cannot be freed and reused yet,
1907 * so we can safely take mapping->i_mmap_rwsem.
1909 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
1914 pgoff_start
= page_to_pgoff(page
);
1915 pgoff_end
= pgoff_start
+ hpage_nr_pages(page
) - 1;
1917 i_mmap_lock_read(mapping
);
1918 vma_interval_tree_foreach(vma
, &mapping
->i_mmap
,
1919 pgoff_start
, pgoff_end
) {
1920 unsigned long address
= vma_address(page
, vma
);
1922 VM_BUG_ON_VMA(address
== -EFAULT
, vma
);
1925 if (rwc
->invalid_vma
&& rwc
->invalid_vma(vma
, rwc
->arg
))
1928 if (!rwc
->rmap_one(page
, vma
, address
, rwc
->arg
))
1930 if (rwc
->done
&& rwc
->done(page
))
1936 i_mmap_unlock_read(mapping
);
1939 void rmap_walk(struct page
*page
, struct rmap_walk_control
*rwc
)
1941 if (unlikely(PageKsm(page
)))
1942 rmap_walk_ksm(page
, rwc
);
1943 else if (PageAnon(page
))
1944 rmap_walk_anon(page
, rwc
, false);
1946 rmap_walk_file(page
, rwc
, false);
1949 /* Like rmap_walk, but caller holds relevant rmap lock */
1950 void rmap_walk_locked(struct page
*page
, struct rmap_walk_control
*rwc
)
1952 /* no ksm support for now */
1953 VM_BUG_ON_PAGE(PageKsm(page
), page
);
1955 rmap_walk_anon(page
, rwc
, true);
1957 rmap_walk_file(page
, rwc
, true);
1960 #ifdef CONFIG_HUGETLB_PAGE
1962 * The following two functions are for anonymous (private mapped) hugepages.
1963 * Unlike common anonymous pages, anonymous hugepages have no accounting code
1964 * and no lru code, because we handle hugepages differently from common pages.
1966 void hugepage_add_anon_rmap(struct page
*page
,
1967 struct vm_area_struct
*vma
, unsigned long address
)
1969 struct anon_vma
*anon_vma
= vma
->anon_vma
;
1972 BUG_ON(!PageLocked(page
));
1974 /* address might be in next vma when migration races vma_adjust */
1975 first
= atomic_inc_and_test(compound_mapcount_ptr(page
));
1977 __page_set_anon_rmap(page
, vma
, address
, 0);
1980 void hugepage_add_new_anon_rmap(struct page
*page
,
1981 struct vm_area_struct
*vma
, unsigned long address
)
1983 BUG_ON(address
< vma
->vm_start
|| address
>= vma
->vm_end
);
1984 atomic_set(compound_mapcount_ptr(page
), 0);
1985 __page_set_anon_rmap(page
, vma
, address
, 1);
1987 #endif /* CONFIG_HUGETLB_PAGE */