2 * mm/rmap.c - physical to virtual reverse mappings
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.
17 * Contributions by Hugh Dickins 2003, 2004
21 * Lock ordering in mm:
23 * inode->i_rwsem (while writing or truncating, not reading or faulting)
25 * mapping->invalidate_lock (in filemap_fault)
26 * page->flags PG_locked (lock_page)
27 * hugetlbfs_i_mmap_rwsem_key (in huge_pmd_share, see hugetlbfs below)
29 * mapping->i_mmap_rwsem
31 * mm->page_table_lock or pte_lock
32 * swap_lock (in swap_duplicate, swap_info_get)
33 * mmlist_lock (in mmput, drain_mmlist and others)
34 * mapping->private_lock (in block_dirty_folio)
35 * folio_lock_memcg move_lock (in block_dirty_folio)
36 * i_pages lock (widely used)
37 * lruvec->lru_lock (in folio_lruvec_lock_irq)
38 * inode->i_lock (in set_page_dirty's __mark_inode_dirty)
39 * bdi.wb->list_lock (in set_page_dirty's __mark_inode_dirty)
40 * sb_lock (within inode_lock in fs/fs-writeback.c)
41 * i_pages lock (widely used, in set_page_dirty,
42 * in arch-dependent flush_dcache_mmap_lock,
43 * within bdi.wb->list_lock in __sync_single_inode)
45 * anon_vma->rwsem,mapping->i_mmap_rwsem (memory_failure, collect_procs_anon)
49 * hugetlbfs PageHuge() take locks in this order:
50 * hugetlb_fault_mutex (hugetlbfs specific page fault mutex)
51 * vma_lock (hugetlb specific lock for pmd_sharing)
52 * mapping->i_mmap_rwsem (also used for hugetlb pmd sharing)
53 * page->flags PG_locked (lock_page)
57 #include <linux/sched/mm.h>
58 #include <linux/sched/task.h>
59 #include <linux/pagemap.h>
60 #include <linux/swap.h>
61 #include <linux/swapops.h>
62 #include <linux/slab.h>
63 #include <linux/init.h>
64 #include <linux/ksm.h>
65 #include <linux/rmap.h>
66 #include <linux/rcupdate.h>
67 #include <linux/export.h>
68 #include <linux/memcontrol.h>
69 #include <linux/mmu_notifier.h>
70 #include <linux/migrate.h>
71 #include <linux/hugetlb.h>
72 #include <linux/huge_mm.h>
73 #include <linux/backing-dev.h>
74 #include <linux/page_idle.h>
75 #include <linux/memremap.h>
76 #include <linux/userfaultfd_k.h>
77 #include <linux/mm_inline.h>
79 #include <asm/tlbflush.h>
81 #define CREATE_TRACE_POINTS
82 #include <trace/events/tlb.h>
83 #include <trace/events/migrate.h>
87 static struct kmem_cache *anon_vma_cachep;
88 static struct kmem_cache *anon_vma_chain_cachep;
90 static inline struct anon_vma *anon_vma_alloc(void)
92 struct anon_vma *anon_vma;
94 anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
96 atomic_set(&anon_vma->refcount, 1);
97 anon_vma->num_children = 0;
98 anon_vma->num_active_vmas = 0;
99 anon_vma->parent = anon_vma;
101 * Initialise the anon_vma root to point to itself. If called
102 * from fork, the root will be reset to the parents anon_vma.
104 anon_vma->root = anon_vma;
110 static inline void anon_vma_free(struct anon_vma *anon_vma)
112 VM_BUG_ON(atomic_read(&anon_vma->refcount));
115 * Synchronize against folio_lock_anon_vma_read() such that
116 * we can safely hold the lock without the anon_vma getting
119 * Relies on the full mb implied by the atomic_dec_and_test() from
120 * put_anon_vma() against the acquire barrier implied by
121 * down_read_trylock() from folio_lock_anon_vma_read(). This orders:
123 * folio_lock_anon_vma_read() VS put_anon_vma()
124 * down_read_trylock() atomic_dec_and_test()
126 * atomic_read() rwsem_is_locked()
128 * LOCK should suffice since the actual taking of the lock must
129 * happen _before_ what follows.
132 if (rwsem_is_locked(&anon_vma->root->rwsem)) {
133 anon_vma_lock_write(anon_vma);
134 anon_vma_unlock_write(anon_vma);
137 kmem_cache_free(anon_vma_cachep, anon_vma);
140 static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp)
142 return kmem_cache_alloc(anon_vma_chain_cachep, gfp);
145 static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
147 kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
150 static void anon_vma_chain_link(struct vm_area_struct *vma,
151 struct anon_vma_chain *avc,
152 struct anon_vma *anon_vma)
155 avc->anon_vma = anon_vma;
156 list_add(&avc->same_vma, &vma->anon_vma_chain);
157 anon_vma_interval_tree_insert(avc, &anon_vma->rb_root);
161 * __anon_vma_prepare - attach an anon_vma to a memory region
162 * @vma: the memory region in question
164 * This makes sure the memory mapping described by 'vma' has
165 * an 'anon_vma' attached to it, so that we can associate the
166 * anonymous pages mapped into it with that anon_vma.
168 * The common case will be that we already have one, which
169 * is handled inline by anon_vma_prepare(). But if
170 * not we either need to find an adjacent mapping that we
171 * can re-use the anon_vma from (very common when the only
172 * reason for splitting a vma has been mprotect()), or we
173 * allocate a new one.
175 * Anon-vma allocations are very subtle, because we may have
176 * optimistically looked up an anon_vma in folio_lock_anon_vma_read()
177 * and that may actually touch the rwsem even in the newly
178 * allocated vma (it depends on RCU to make sure that the
179 * anon_vma isn't actually destroyed).
181 * As a result, we need to do proper anon_vma locking even
182 * for the new allocation. At the same time, we do not want
183 * to do any locking for the common case of already having
186 * This must be called with the mmap_lock held for reading.
188 int __anon_vma_prepare(struct vm_area_struct *vma)
190 struct mm_struct *mm = vma->vm_mm;
191 struct anon_vma *anon_vma, *allocated;
192 struct anon_vma_chain *avc;
196 avc = anon_vma_chain_alloc(GFP_KERNEL);
200 anon_vma = find_mergeable_anon_vma(vma);
203 anon_vma = anon_vma_alloc();
204 if (unlikely(!anon_vma))
205 goto out_enomem_free_avc;
206 anon_vma->num_children++; /* self-parent link for new root */
207 allocated = anon_vma;
210 anon_vma_lock_write(anon_vma);
211 /* page_table_lock to protect against threads */
212 spin_lock(&mm->page_table_lock);
213 if (likely(!vma->anon_vma)) {
214 vma->anon_vma = anon_vma;
215 anon_vma_chain_link(vma, avc, anon_vma);
216 anon_vma->num_active_vmas++;
220 spin_unlock(&mm->page_table_lock);
221 anon_vma_unlock_write(anon_vma);
223 if (unlikely(allocated))
224 put_anon_vma(allocated);
226 anon_vma_chain_free(avc);
231 anon_vma_chain_free(avc);
237 * This is a useful helper function for locking the anon_vma root as
238 * we traverse the vma->anon_vma_chain, looping over anon_vma's that
241 * Such anon_vma's should have the same root, so you'd expect to see
242 * just a single mutex_lock for the whole traversal.
244 static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
246 struct anon_vma *new_root = anon_vma->root;
247 if (new_root != root) {
248 if (WARN_ON_ONCE(root))
249 up_write(&root->rwsem);
251 down_write(&root->rwsem);
256 static inline void unlock_anon_vma_root(struct anon_vma *root)
259 up_write(&root->rwsem);
263 * Attach the anon_vmas from src to dst.
264 * Returns 0 on success, -ENOMEM on failure.
266 * anon_vma_clone() is called by vma_expand(), vma_merge(), __split_vma(),
267 * copy_vma() and anon_vma_fork(). The first four want an exact copy of src,
268 * while the last one, anon_vma_fork(), may try to reuse an existing anon_vma to
269 * prevent endless growth of anon_vma. Since dst->anon_vma is set to NULL before
270 * call, we can identify this case by checking (!dst->anon_vma &&
273 * If (!dst->anon_vma && src->anon_vma) is true, this function tries to find
274 * and reuse existing anon_vma which has no vmas and only one child anon_vma.
275 * This prevents degradation of anon_vma hierarchy to endless linear chain in
276 * case of constantly forking task. On the other hand, an anon_vma with more
277 * than one child isn't reused even if there was no alive vma, thus rmap
278 * walker has a good chance of avoiding scanning the whole hierarchy when it
279 * searches where page is mapped.
281 int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
283 struct anon_vma_chain *avc, *pavc;
284 struct anon_vma *root = NULL;
286 list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
287 struct anon_vma *anon_vma;
289 avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
290 if (unlikely(!avc)) {
291 unlock_anon_vma_root(root);
293 avc = anon_vma_chain_alloc(GFP_KERNEL);
297 anon_vma = pavc->anon_vma;
298 root = lock_anon_vma_root(root, anon_vma);
299 anon_vma_chain_link(dst, avc, anon_vma);
302 * Reuse existing anon_vma if it has no vma and only one
305 * Root anon_vma is never reused:
306 * it has self-parent reference and at least one child.
308 if (!dst->anon_vma && src->anon_vma &&
309 anon_vma->num_children < 2 &&
310 anon_vma->num_active_vmas == 0)
311 dst->anon_vma = anon_vma;
314 dst->anon_vma->num_active_vmas++;
315 unlock_anon_vma_root(root);
320 * dst->anon_vma is dropped here otherwise its num_active_vmas can
321 * be incorrectly decremented in unlink_anon_vmas().
322 * We can safely do this because callers of anon_vma_clone() don't care
323 * about dst->anon_vma if anon_vma_clone() failed.
325 dst->anon_vma = NULL;
326 unlink_anon_vmas(dst);
331 * Attach vma to its own anon_vma, as well as to the anon_vmas that
332 * the corresponding VMA in the parent process is attached to.
333 * Returns 0 on success, non-zero on failure.
335 int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
337 struct anon_vma_chain *avc;
338 struct anon_vma *anon_vma;
341 /* Don't bother if the parent process has no anon_vma here. */
345 /* Drop inherited anon_vma, we'll reuse existing or allocate new. */
346 vma->anon_vma = NULL;
349 * First, attach the new VMA to the parent VMA's anon_vmas,
350 * so rmap can find non-COWed pages in child processes.
352 error = anon_vma_clone(vma, pvma);
356 /* An existing anon_vma has been reused, all done then. */
360 /* Then add our own anon_vma. */
361 anon_vma = anon_vma_alloc();
364 anon_vma->num_active_vmas++;
365 avc = anon_vma_chain_alloc(GFP_KERNEL);
367 goto out_error_free_anon_vma;
370 * The root anon_vma's rwsem is the lock actually used when we
371 * lock any of the anon_vmas in this anon_vma tree.
373 anon_vma->root = pvma->anon_vma->root;
374 anon_vma->parent = pvma->anon_vma;
376 * With refcounts, an anon_vma can stay around longer than the
377 * process it belongs to. The root anon_vma needs to be pinned until
378 * this anon_vma is freed, because the lock lives in the root.
380 get_anon_vma(anon_vma->root);
381 /* Mark this anon_vma as the one where our new (COWed) pages go. */
382 vma->anon_vma = anon_vma;
383 anon_vma_lock_write(anon_vma);
384 anon_vma_chain_link(vma, avc, anon_vma);
385 anon_vma->parent->num_children++;
386 anon_vma_unlock_write(anon_vma);
390 out_error_free_anon_vma:
391 put_anon_vma(anon_vma);
393 unlink_anon_vmas(vma);
397 void unlink_anon_vmas(struct vm_area_struct *vma)
399 struct anon_vma_chain *avc, *next;
400 struct anon_vma *root = NULL;
403 * Unlink each anon_vma chained to the VMA. This list is ordered
404 * from newest to oldest, ensuring the root anon_vma gets freed last.
406 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
407 struct anon_vma *anon_vma = avc->anon_vma;
409 root = lock_anon_vma_root(root, anon_vma);
410 anon_vma_interval_tree_remove(avc, &anon_vma->rb_root);
413 * Leave empty anon_vmas on the list - we'll need
414 * to free them outside the lock.
416 if (RB_EMPTY_ROOT(&anon_vma->rb_root.rb_root)) {
417 anon_vma->parent->num_children--;
421 list_del(&avc->same_vma);
422 anon_vma_chain_free(avc);
425 vma->anon_vma->num_active_vmas--;
428 * vma would still be needed after unlink, and anon_vma will be prepared
431 vma->anon_vma = NULL;
433 unlock_anon_vma_root(root);
436 * Iterate the list once more, it now only contains empty and unlinked
437 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
438 * needing to write-acquire the anon_vma->root->rwsem.
440 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
441 struct anon_vma *anon_vma = avc->anon_vma;
443 VM_WARN_ON(anon_vma->num_children);
444 VM_WARN_ON(anon_vma->num_active_vmas);
445 put_anon_vma(anon_vma);
447 list_del(&avc->same_vma);
448 anon_vma_chain_free(avc);
452 static void anon_vma_ctor(void *data)
454 struct anon_vma *anon_vma = data;
456 init_rwsem(&anon_vma->rwsem);
457 atomic_set(&anon_vma->refcount, 0);
458 anon_vma->rb_root = RB_ROOT_CACHED;
461 void __init anon_vma_init(void)
463 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
464 0, SLAB_TYPESAFE_BY_RCU|SLAB_PANIC|SLAB_ACCOUNT,
466 anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain,
467 SLAB_PANIC|SLAB_ACCOUNT);
471 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
473 * Since there is no serialization what so ever against folio_remove_rmap_*()
474 * the best this function can do is return a refcount increased anon_vma
475 * that might have been relevant to this page.
477 * The page might have been remapped to a different anon_vma or the anon_vma
478 * returned may already be freed (and even reused).
480 * In case it was remapped to a different anon_vma, the new anon_vma will be a
481 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
482 * ensure that any anon_vma obtained from the page will still be valid for as
483 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
485 * All users of this function must be very careful when walking the anon_vma
486 * chain and verify that the page in question is indeed mapped in it
487 * [ something equivalent to page_mapped_in_vma() ].
489 * Since anon_vma's slab is SLAB_TYPESAFE_BY_RCU and we know from
490 * folio_remove_rmap_*() that the anon_vma pointer from page->mapping is valid
491 * if there is a mapcount, we can dereference the anon_vma after observing
494 * NOTE: the caller should normally hold folio lock when calling this. If
495 * not, the caller needs to double check the anon_vma didn't change after
496 * taking the anon_vma lock for either read or write (UFFDIO_MOVE can modify it
497 * concurrently without folio lock protection). See folio_lock_anon_vma_read()
498 * which has already covered that, and comment above remap_pages().
500 struct anon_vma *folio_get_anon_vma(struct folio *folio)
502 struct anon_vma *anon_vma = NULL;
503 unsigned long anon_mapping;
506 anon_mapping = (unsigned long)READ_ONCE(folio->mapping);
507 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
509 if (!folio_mapped(folio))
512 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
513 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
519 * If this folio is still mapped, then its anon_vma cannot have been
520 * freed. But if it has been unmapped, we have no security against the
521 * anon_vma structure being freed and reused (for another anon_vma:
522 * SLAB_TYPESAFE_BY_RCU guarantees that - so the atomic_inc_not_zero()
523 * above cannot corrupt).
525 if (!folio_mapped(folio)) {
527 put_anon_vma(anon_vma);
537 * Similar to folio_get_anon_vma() except it locks the anon_vma.
539 * Its a little more complex as it tries to keep the fast path to a single
540 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
541 * reference like with folio_get_anon_vma() and then block on the mutex
542 * on !rwc->try_lock case.
544 struct anon_vma *folio_lock_anon_vma_read(struct folio *folio,
545 struct rmap_walk_control *rwc)
547 struct anon_vma *anon_vma = NULL;
548 struct anon_vma *root_anon_vma;
549 unsigned long anon_mapping;
553 anon_mapping = (unsigned long)READ_ONCE(folio->mapping);
554 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
556 if (!folio_mapped(folio))
559 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
560 root_anon_vma = READ_ONCE(anon_vma->root);
561 if (down_read_trylock(&root_anon_vma->rwsem)) {
563 * folio_move_anon_rmap() might have changed the anon_vma as we
564 * might not hold the folio lock here.
566 if (unlikely((unsigned long)READ_ONCE(folio->mapping) !=
568 up_read(&root_anon_vma->rwsem);
574 * If the folio is still mapped, then this anon_vma is still
575 * its anon_vma, and holding the mutex ensures that it will
576 * not go away, see anon_vma_free().
578 if (!folio_mapped(folio)) {
579 up_read(&root_anon_vma->rwsem);
585 if (rwc && rwc->try_lock) {
587 rwc->contended = true;
591 /* trylock failed, we got to sleep */
592 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
597 if (!folio_mapped(folio)) {
599 put_anon_vma(anon_vma);
603 /* we pinned the anon_vma, its safe to sleep */
605 anon_vma_lock_read(anon_vma);
608 * folio_move_anon_rmap() might have changed the anon_vma as we might
609 * not hold the folio lock here.
611 if (unlikely((unsigned long)READ_ONCE(folio->mapping) !=
613 anon_vma_unlock_read(anon_vma);
614 put_anon_vma(anon_vma);
619 if (atomic_dec_and_test(&anon_vma->refcount)) {
621 * Oops, we held the last refcount, release the lock
622 * and bail -- can't simply use put_anon_vma() because
623 * we'll deadlock on the anon_vma_lock_write() recursion.
625 anon_vma_unlock_read(anon_vma);
626 __put_anon_vma(anon_vma);
637 #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
639 * Flush TLB entries for recently unmapped pages from remote CPUs. It is
640 * important if a PTE was dirty when it was unmapped that it's flushed
641 * before any IO is initiated on the page to prevent lost writes. Similarly,
642 * it must be flushed before freeing to prevent data leakage.
644 void try_to_unmap_flush(void)
646 struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
648 if (!tlb_ubc->flush_required)
651 arch_tlbbatch_flush(&tlb_ubc->arch);
652 tlb_ubc->flush_required = false;
653 tlb_ubc->writable = false;
656 /* Flush iff there are potentially writable TLB entries that can race with IO */
657 void try_to_unmap_flush_dirty(void)
659 struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
661 if (tlb_ubc->writable)
662 try_to_unmap_flush();
666 * Bits 0-14 of mm->tlb_flush_batched record pending generations.
667 * Bits 16-30 of mm->tlb_flush_batched bit record flushed generations.
669 #define TLB_FLUSH_BATCH_FLUSHED_SHIFT 16
670 #define TLB_FLUSH_BATCH_PENDING_MASK \
671 ((1 << (TLB_FLUSH_BATCH_FLUSHED_SHIFT - 1)) - 1)
672 #define TLB_FLUSH_BATCH_PENDING_LARGE \
673 (TLB_FLUSH_BATCH_PENDING_MASK / 2)
675 static void set_tlb_ubc_flush_pending(struct mm_struct *mm, pte_t pteval,
678 struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
680 bool writable = pte_dirty(pteval);
682 if (!pte_accessible(mm, pteval))
685 arch_tlbbatch_add_pending(&tlb_ubc->arch, mm, uaddr);
686 tlb_ubc->flush_required = true;
689 * Ensure compiler does not re-order the setting of tlb_flush_batched
690 * before the PTE is cleared.
693 batch = atomic_read(&mm->tlb_flush_batched);
695 if ((batch & TLB_FLUSH_BATCH_PENDING_MASK) > TLB_FLUSH_BATCH_PENDING_LARGE) {
697 * Prevent `pending' from catching up with `flushed' because of
698 * overflow. Reset `pending' and `flushed' to be 1 and 0 if
699 * `pending' becomes large.
701 if (!atomic_try_cmpxchg(&mm->tlb_flush_batched, &batch, 1))
704 atomic_inc(&mm->tlb_flush_batched);
708 * If the PTE was dirty then it's best to assume it's writable. The
709 * caller must use try_to_unmap_flush_dirty() or try_to_unmap_flush()
710 * before the page is queued for IO.
713 tlb_ubc->writable = true;
717 * Returns true if the TLB flush should be deferred to the end of a batch of
718 * unmap operations to reduce IPIs.
720 static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
722 if (!(flags & TTU_BATCH_FLUSH))
725 return arch_tlbbatch_should_defer(mm);
729 * Reclaim unmaps pages under the PTL but do not flush the TLB prior to
730 * releasing the PTL if TLB flushes are batched. It's possible for a parallel
731 * operation such as mprotect or munmap to race between reclaim unmapping
732 * the page and flushing the page. If this race occurs, it potentially allows
733 * access to data via a stale TLB entry. Tracking all mm's that have TLB
734 * batching in flight would be expensive during reclaim so instead track
735 * whether TLB batching occurred in the past and if so then do a flush here
736 * if required. This will cost one additional flush per reclaim cycle paid
737 * by the first operation at risk such as mprotect and mumap.
739 * This must be called under the PTL so that an access to tlb_flush_batched
740 * that is potentially a "reclaim vs mprotect/munmap/etc" race will synchronise
743 void flush_tlb_batched_pending(struct mm_struct *mm)
745 int batch = atomic_read(&mm->tlb_flush_batched);
746 int pending = batch & TLB_FLUSH_BATCH_PENDING_MASK;
747 int flushed = batch >> TLB_FLUSH_BATCH_FLUSHED_SHIFT;
749 if (pending != flushed) {
750 arch_flush_tlb_batched_pending(mm);
752 * If the new TLB flushing is pending during flushing, leave
753 * mm->tlb_flush_batched as is, to avoid losing flushing.
755 atomic_cmpxchg(&mm->tlb_flush_batched, batch,
756 pending | (pending << TLB_FLUSH_BATCH_FLUSHED_SHIFT));
760 static void set_tlb_ubc_flush_pending(struct mm_struct *mm, pte_t pteval,
765 static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
769 #endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
772 * At what user virtual address is page expected in vma?
773 * Caller should check the page is actually part of the vma.
775 unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
777 struct folio *folio = page_folio(page);
778 if (folio_test_anon(folio)) {
779 struct anon_vma *page__anon_vma = folio_anon_vma(folio);
781 * Note: swapoff's unuse_vma() is more efficient with this
782 * check, and needs it to match anon_vma when KSM is active.
784 if (!vma->anon_vma || !page__anon_vma ||
785 vma->anon_vma->root != page__anon_vma->root)
787 } else if (!vma->vm_file) {
789 } else if (vma->vm_file->f_mapping != folio->mapping) {
793 return vma_address(page, vma);
797 * Returns the actual pmd_t* where we expect 'address' to be mapped from, or
798 * NULL if it doesn't exist. No guarantees / checks on what the pmd_t*
801 pmd_t *mm_find_pmd(struct mm_struct *mm, unsigned long address)
808 pgd = pgd_offset(mm, address);
809 if (!pgd_present(*pgd))
812 p4d = p4d_offset(pgd, address);
813 if (!p4d_present(*p4d))
816 pud = pud_offset(p4d, address);
817 if (!pud_present(*pud))
820 pmd = pmd_offset(pud, address);
825 struct folio_referenced_arg {
828 unsigned long vm_flags;
829 struct mem_cgroup *memcg;
833 * arg: folio_referenced_arg will be passed
835 static bool folio_referenced_one(struct folio *folio,
836 struct vm_area_struct *vma, unsigned long address, void *arg)
838 struct folio_referenced_arg *pra = arg;
839 DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0);
841 unsigned long start = address, ptes = 0;
843 while (page_vma_mapped_walk(&pvmw)) {
844 address = pvmw.address;
846 if (vma->vm_flags & VM_LOCKED) {
847 if (!folio_test_large(folio) || !pvmw.pte) {
848 /* Restore the mlock which got missed */
849 mlock_vma_folio(folio, vma);
850 page_vma_mapped_walk_done(&pvmw);
851 pra->vm_flags |= VM_LOCKED;
852 return false; /* To break the loop */
855 * For large folio fully mapped to VMA, will
856 * be handled after the pvmw loop.
858 * For large folio cross VMA boundaries, it's
859 * expected to be picked by page reclaim. But
860 * should skip reference of pages which are in
861 * the range of VM_LOCKED vma. As page reclaim
862 * should just count the reference of pages out
863 * the range of VM_LOCKED vma.
871 if (lru_gen_enabled() &&
872 pte_young(ptep_get(pvmw.pte))) {
873 lru_gen_look_around(&pvmw);
877 if (ptep_clear_flush_young_notify(vma, address,
880 } else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
881 if (pmdp_clear_flush_young_notify(vma, address,
885 /* unexpected pmd-mapped folio? */
892 if ((vma->vm_flags & VM_LOCKED) &&
893 folio_test_large(folio) &&
894 folio_within_vma(folio, vma)) {
895 unsigned long s_align, e_align;
897 s_align = ALIGN_DOWN(start, PMD_SIZE);
898 e_align = ALIGN_DOWN(start + folio_size(folio) - 1, PMD_SIZE);
900 /* folio doesn't cross page table boundary and fully mapped */
901 if ((s_align == e_align) && (ptes == folio_nr_pages(folio))) {
902 /* Restore the mlock which got missed */
903 mlock_vma_folio(folio, vma);
904 pra->vm_flags |= VM_LOCKED;
905 return false; /* To break the loop */
910 folio_clear_idle(folio);
911 if (folio_test_clear_young(folio))
916 pra->vm_flags |= vma->vm_flags & ~VM_LOCKED;
920 return false; /* To break the loop */
925 static bool invalid_folio_referenced_vma(struct vm_area_struct *vma, void *arg)
927 struct folio_referenced_arg *pra = arg;
928 struct mem_cgroup *memcg = pra->memcg;
931 * Ignore references from this mapping if it has no recency. If the
932 * folio has been used in another mapping, we will catch it; if this
933 * other mapping is already gone, the unmap path will have set the
934 * referenced flag or activated the folio in zap_pte_range().
936 if (!vma_has_recency(vma))
940 * If we are reclaiming on behalf of a cgroup, skip counting on behalf
941 * of references from different cgroups.
943 if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
950 * folio_referenced() - Test if the folio was referenced.
951 * @folio: The folio to test.
952 * @is_locked: Caller holds lock on the folio.
953 * @memcg: target memory cgroup
954 * @vm_flags: A combination of all the vma->vm_flags which referenced the folio.
956 * Quick test_and_clear_referenced for all mappings of a folio,
958 * Return: The number of mappings which referenced the folio. Return -1 if
959 * the function bailed out due to rmap lock contention.
961 int folio_referenced(struct folio *folio, int is_locked,
962 struct mem_cgroup *memcg, unsigned long *vm_flags)
965 struct folio_referenced_arg pra = {
966 .mapcount = folio_mapcount(folio),
969 struct rmap_walk_control rwc = {
970 .rmap_one = folio_referenced_one,
972 .anon_lock = folio_lock_anon_vma_read,
974 .invalid_vma = invalid_folio_referenced_vma,
981 if (!folio_raw_mapping(folio))
984 if (!is_locked && (!folio_test_anon(folio) || folio_test_ksm(folio))) {
985 we_locked = folio_trylock(folio);
990 rmap_walk(folio, &rwc);
991 *vm_flags = pra.vm_flags;
996 return rwc.contended ? -1 : pra.referenced;
999 static int page_vma_mkclean_one(struct page_vma_mapped_walk *pvmw)
1002 struct vm_area_struct *vma = pvmw->vma;
1003 struct mmu_notifier_range range;
1004 unsigned long address = pvmw->address;
1007 * We have to assume the worse case ie pmd for invalidation. Note that
1008 * the folio can not be freed from this function.
1010 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE, 0,
1011 vma->vm_mm, address, vma_address_end(pvmw));
1012 mmu_notifier_invalidate_range_start(&range);
1014 while (page_vma_mapped_walk(pvmw)) {
1017 address = pvmw->address;
1019 pte_t *pte = pvmw->pte;
1020 pte_t entry = ptep_get(pte);
1022 if (!pte_dirty(entry) && !pte_write(entry))
1025 flush_cache_page(vma, address, pte_pfn(entry));
1026 entry = ptep_clear_flush(vma, address, pte);
1027 entry = pte_wrprotect(entry);
1028 entry = pte_mkclean(entry);
1029 set_pte_at(vma->vm_mm, address, pte, entry);
1032 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1033 pmd_t *pmd = pvmw->pmd;
1036 if (!pmd_dirty(*pmd) && !pmd_write(*pmd))
1039 flush_cache_range(vma, address,
1040 address + HPAGE_PMD_SIZE);
1041 entry = pmdp_invalidate(vma, address, pmd);
1042 entry = pmd_wrprotect(entry);
1043 entry = pmd_mkclean(entry);
1044 set_pmd_at(vma->vm_mm, address, pmd, entry);
1047 /* unexpected pmd-mapped folio? */
1056 mmu_notifier_invalidate_range_end(&range);
1061 static bool page_mkclean_one(struct folio *folio, struct vm_area_struct *vma,
1062 unsigned long address, void *arg)
1064 DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, PVMW_SYNC);
1067 *cleaned += page_vma_mkclean_one(&pvmw);
1072 static bool invalid_mkclean_vma(struct vm_area_struct *vma, void *arg)
1074 if (vma->vm_flags & VM_SHARED)
1080 int folio_mkclean(struct folio *folio)
1083 struct address_space *mapping;
1084 struct rmap_walk_control rwc = {
1085 .arg = (void *)&cleaned,
1086 .rmap_one = page_mkclean_one,
1087 .invalid_vma = invalid_mkclean_vma,
1090 BUG_ON(!folio_test_locked(folio));
1092 if (!folio_mapped(folio))
1095 mapping = folio_mapping(folio);
1099 rmap_walk(folio, &rwc);
1103 EXPORT_SYMBOL_GPL(folio_mkclean);
1106 * pfn_mkclean_range - Cleans the PTEs (including PMDs) mapped with range of
1107 * [@pfn, @pfn + @nr_pages) at the specific offset (@pgoff)
1108 * within the @vma of shared mappings. And since clean PTEs
1109 * should also be readonly, write protects them too.
1111 * @nr_pages: number of physically contiguous pages srarting with @pfn.
1112 * @pgoff: page offset that the @pfn mapped with.
1113 * @vma: vma that @pfn mapped within.
1115 * Returns the number of cleaned PTEs (including PMDs).
1117 int pfn_mkclean_range(unsigned long pfn, unsigned long nr_pages, pgoff_t pgoff,
1118 struct vm_area_struct *vma)
1120 struct page_vma_mapped_walk pvmw = {
1122 .nr_pages = nr_pages,
1128 if (invalid_mkclean_vma(vma, NULL))
1131 pvmw.address = vma_pgoff_address(pgoff, nr_pages, vma);
1132 VM_BUG_ON_VMA(pvmw.address == -EFAULT, vma);
1134 return page_vma_mkclean_one(&pvmw);
1137 int folio_total_mapcount(struct folio *folio)
1139 int mapcount = folio_entire_mapcount(folio);
1143 /* In the common case, avoid the loop when no pages mapped by PTE */
1144 if (folio_nr_pages_mapped(folio) == 0)
1147 * Add all the PTE mappings of those pages mapped by PTE.
1148 * Limit the loop to folio_nr_pages_mapped()?
1149 * Perhaps: given all the raciness, that may be a good or a bad idea.
1151 nr_pages = folio_nr_pages(folio);
1152 for (i = 0; i < nr_pages; i++)
1153 mapcount += atomic_read(&folio_page(folio, i)->_mapcount);
1155 /* But each of those _mapcounts was based on -1 */
1156 mapcount += nr_pages;
1160 static __always_inline unsigned int __folio_add_rmap(struct folio *folio,
1161 struct page *page, int nr_pages, enum rmap_level level,
1164 atomic_t *mapped = &folio->_nr_pages_mapped;
1167 __folio_rmap_sanity_checks(folio, page, nr_pages, level);
1170 case RMAP_LEVEL_PTE:
1172 first = atomic_inc_and_test(&page->_mapcount);
1173 if (first && folio_test_large(folio)) {
1174 first = atomic_inc_return_relaxed(mapped);
1175 first = (first < ENTIRELY_MAPPED);
1180 } while (page++, --nr_pages > 0);
1182 case RMAP_LEVEL_PMD:
1183 first = atomic_inc_and_test(&folio->_entire_mapcount);
1185 nr = atomic_add_return_relaxed(ENTIRELY_MAPPED, mapped);
1186 if (likely(nr < ENTIRELY_MAPPED + ENTIRELY_MAPPED)) {
1187 *nr_pmdmapped = folio_nr_pages(folio);
1188 nr = *nr_pmdmapped - (nr & FOLIO_PAGES_MAPPED);
1189 /* Raced ahead of a remove and another add? */
1190 if (unlikely(nr < 0))
1193 /* Raced ahead of a remove of ENTIRELY_MAPPED */
1203 * folio_move_anon_rmap - move a folio to our anon_vma
1204 * @folio: The folio to move to our anon_vma
1205 * @vma: The vma the folio belongs to
1207 * When a folio belongs exclusively to one process after a COW event,
1208 * that folio can be moved into the anon_vma that belongs to just that
1209 * process, so the rmap code will not search the parent or sibling processes.
1211 void folio_move_anon_rmap(struct folio *folio, struct vm_area_struct *vma)
1213 void *anon_vma = vma->anon_vma;
1215 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
1216 VM_BUG_ON_VMA(!anon_vma, vma);
1218 anon_vma += PAGE_MAPPING_ANON;
1220 * Ensure that anon_vma and the PAGE_MAPPING_ANON bit are written
1221 * simultaneously, so a concurrent reader (eg folio_referenced()'s
1222 * folio_test_anon()) will not see one without the other.
1224 WRITE_ONCE(folio->mapping, anon_vma);
1228 * __folio_set_anon - set up a new anonymous rmap for a folio
1229 * @folio: The folio to set up the new anonymous rmap for.
1230 * @vma: VM area to add the folio to.
1231 * @address: User virtual address of the mapping
1232 * @exclusive: Whether the folio is exclusive to the process.
1234 static void __folio_set_anon(struct folio *folio, struct vm_area_struct *vma,
1235 unsigned long address, bool exclusive)
1237 struct anon_vma *anon_vma = vma->anon_vma;
1242 * If the folio isn't exclusive to this vma, we must use the _oldest_
1243 * possible anon_vma for the folio mapping!
1246 anon_vma = anon_vma->root;
1249 * page_idle does a lockless/optimistic rmap scan on folio->mapping.
1250 * Make sure the compiler doesn't split the stores of anon_vma and
1251 * the PAGE_MAPPING_ANON type identifier, otherwise the rmap code
1252 * could mistake the mapping for a struct address_space and crash.
1254 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1255 WRITE_ONCE(folio->mapping, (struct address_space *) anon_vma);
1256 folio->index = linear_page_index(vma, address);
1260 * __page_check_anon_rmap - sanity check anonymous rmap addition
1261 * @folio: The folio containing @page.
1262 * @page: the page to check the mapping of
1263 * @vma: the vm area in which the mapping is added
1264 * @address: the user virtual address mapped
1266 static void __page_check_anon_rmap(struct folio *folio, struct page *page,
1267 struct vm_area_struct *vma, unsigned long address)
1270 * The page's anon-rmap details (mapping and index) are guaranteed to
1271 * be set up correctly at this point.
1273 * We have exclusion against folio_add_anon_rmap_*() because the caller
1274 * always holds the page locked.
1276 * We have exclusion against folio_add_new_anon_rmap because those pages
1277 * are initially only visible via the pagetables, and the pte is locked
1278 * over the call to folio_add_new_anon_rmap.
1280 VM_BUG_ON_FOLIO(folio_anon_vma(folio)->root != vma->anon_vma->root,
1282 VM_BUG_ON_PAGE(page_to_pgoff(page) != linear_page_index(vma, address),
1286 static __always_inline void __folio_add_anon_rmap(struct folio *folio,
1287 struct page *page, int nr_pages, struct vm_area_struct *vma,
1288 unsigned long address, rmap_t flags, enum rmap_level level)
1290 int i, nr, nr_pmdmapped = 0;
1292 nr = __folio_add_rmap(folio, page, nr_pages, level, &nr_pmdmapped);
1294 __lruvec_stat_mod_folio(folio, NR_ANON_THPS, nr_pmdmapped);
1296 __lruvec_stat_mod_folio(folio, NR_ANON_MAPPED, nr);
1298 if (unlikely(!folio_test_anon(folio))) {
1299 VM_WARN_ON_FOLIO(!folio_test_locked(folio), folio);
1301 * For a PTE-mapped large folio, we only know that the single
1302 * PTE is exclusive. Further, __folio_set_anon() might not get
1303 * folio->index right when not given the address of the head
1306 VM_WARN_ON_FOLIO(folio_test_large(folio) &&
1307 level != RMAP_LEVEL_PMD, folio);
1308 __folio_set_anon(folio, vma, address,
1309 !!(flags & RMAP_EXCLUSIVE));
1310 } else if (likely(!folio_test_ksm(folio))) {
1311 __page_check_anon_rmap(folio, page, vma, address);
1314 if (flags & RMAP_EXCLUSIVE) {
1316 case RMAP_LEVEL_PTE:
1317 for (i = 0; i < nr_pages; i++)
1318 SetPageAnonExclusive(page + i);
1320 case RMAP_LEVEL_PMD:
1321 SetPageAnonExclusive(page);
1325 for (i = 0; i < nr_pages; i++) {
1326 struct page *cur_page = page + i;
1328 /* While PTE-mapping a THP we have a PMD and a PTE mapping. */
1329 VM_WARN_ON_FOLIO((atomic_read(&cur_page->_mapcount) > 0 ||
1330 (folio_test_large(folio) &&
1331 folio_entire_mapcount(folio) > 1)) &&
1332 PageAnonExclusive(cur_page), folio);
1336 * For large folio, only mlock it if it's fully mapped to VMA. It's
1337 * not easy to check whether the large folio is fully mapped to VMA
1338 * here. Only mlock normal 4K folio and leave page reclaim to handle
1341 if (!folio_test_large(folio))
1342 mlock_vma_folio(folio, vma);
1346 * folio_add_anon_rmap_ptes - add PTE mappings to a page range of an anon folio
1347 * @folio: The folio to add the mappings to
1348 * @page: The first page to add
1349 * @nr_pages: The number of pages which will be mapped
1350 * @vma: The vm area in which the mappings are added
1351 * @address: The user virtual address of the first page to map
1352 * @flags: The rmap flags
1354 * The page range of folio is defined by [first_page, first_page + nr_pages)
1356 * The caller needs to hold the page table lock, and the page must be locked in
1357 * the anon_vma case: to serialize mapping,index checking after setting,
1358 * and to ensure that an anon folio is not being upgraded racily to a KSM folio
1359 * (but KSM folios are never downgraded).
1361 void folio_add_anon_rmap_ptes(struct folio *folio, struct page *page,
1362 int nr_pages, struct vm_area_struct *vma, unsigned long address,
1365 __folio_add_anon_rmap(folio, page, nr_pages, vma, address, flags,
1370 * folio_add_anon_rmap_pmd - add a PMD mapping to a page range of an anon folio
1371 * @folio: The folio to add the mapping to
1372 * @page: The first page to add
1373 * @vma: The vm area in which the mapping is added
1374 * @address: The user virtual address of the first page to map
1375 * @flags: The rmap flags
1377 * The page range of folio is defined by [first_page, first_page + HPAGE_PMD_NR)
1379 * The caller needs to hold the page table lock, and the page must be locked in
1380 * the anon_vma case: to serialize mapping,index checking after setting.
1382 void folio_add_anon_rmap_pmd(struct folio *folio, struct page *page,
1383 struct vm_area_struct *vma, unsigned long address, rmap_t flags)
1385 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1386 __folio_add_anon_rmap(folio, page, HPAGE_PMD_NR, vma, address, flags,
1394 * folio_add_new_anon_rmap - Add mapping to a new anonymous folio.
1395 * @folio: The folio to add the mapping to.
1396 * @vma: the vm area in which the mapping is added
1397 * @address: the user virtual address mapped
1399 * Like folio_add_anon_rmap_*() but must only be called on *new* folios.
1400 * This means the inc-and-test can be bypassed.
1401 * The folio does not have to be locked.
1403 * If the folio is pmd-mappable, it is accounted as a THP. As the folio
1404 * is new, it's assumed to be mapped exclusively by a single process.
1406 void folio_add_new_anon_rmap(struct folio *folio, struct vm_area_struct *vma,
1407 unsigned long address)
1409 int nr = folio_nr_pages(folio);
1411 VM_WARN_ON_FOLIO(folio_test_hugetlb(folio), folio);
1412 VM_BUG_ON_VMA(address < vma->vm_start ||
1413 address + (nr << PAGE_SHIFT) > vma->vm_end, vma);
1414 __folio_set_swapbacked(folio);
1415 __folio_set_anon(folio, vma, address, true);
1417 if (likely(!folio_test_large(folio))) {
1418 /* increment count (starts at -1) */
1419 atomic_set(&folio->_mapcount, 0);
1420 SetPageAnonExclusive(&folio->page);
1421 } else if (!folio_test_pmd_mappable(folio)) {
1424 for (i = 0; i < nr; i++) {
1425 struct page *page = folio_page(folio, i);
1427 /* increment count (starts at -1) */
1428 atomic_set(&page->_mapcount, 0);
1429 SetPageAnonExclusive(page);
1432 atomic_set(&folio->_nr_pages_mapped, nr);
1434 /* increment count (starts at -1) */
1435 atomic_set(&folio->_entire_mapcount, 0);
1436 atomic_set(&folio->_nr_pages_mapped, ENTIRELY_MAPPED);
1437 SetPageAnonExclusive(&folio->page);
1438 __lruvec_stat_mod_folio(folio, NR_ANON_THPS, nr);
1441 __lruvec_stat_mod_folio(folio, NR_ANON_MAPPED, nr);
1444 static __always_inline void __folio_add_file_rmap(struct folio *folio,
1445 struct page *page, int nr_pages, struct vm_area_struct *vma,
1446 enum rmap_level level)
1448 int nr, nr_pmdmapped = 0;
1450 VM_WARN_ON_FOLIO(folio_test_anon(folio), folio);
1452 nr = __folio_add_rmap(folio, page, nr_pages, level, &nr_pmdmapped);
1454 __lruvec_stat_mod_folio(folio, folio_test_swapbacked(folio) ?
1455 NR_SHMEM_PMDMAPPED : NR_FILE_PMDMAPPED, nr_pmdmapped);
1457 __lruvec_stat_mod_folio(folio, NR_FILE_MAPPED, nr);
1459 /* See comments in folio_add_anon_rmap_*() */
1460 if (!folio_test_large(folio))
1461 mlock_vma_folio(folio, vma);
1465 * folio_add_file_rmap_ptes - add PTE mappings to a page range of a folio
1466 * @folio: The folio to add the mappings to
1467 * @page: The first page to add
1468 * @nr_pages: The number of pages that will be mapped using PTEs
1469 * @vma: The vm area in which the mappings are added
1471 * The page range of the folio is defined by [page, page + nr_pages)
1473 * The caller needs to hold the page table lock.
1475 void folio_add_file_rmap_ptes(struct folio *folio, struct page *page,
1476 int nr_pages, struct vm_area_struct *vma)
1478 __folio_add_file_rmap(folio, page, nr_pages, vma, RMAP_LEVEL_PTE);
1482 * folio_add_file_rmap_pmd - add a PMD mapping to a page range of a folio
1483 * @folio: The folio to add the mapping to
1484 * @page: The first page to add
1485 * @vma: The vm area in which the mapping is added
1487 * The page range of the folio is defined by [page, page + HPAGE_PMD_NR)
1489 * The caller needs to hold the page table lock.
1491 void folio_add_file_rmap_pmd(struct folio *folio, struct page *page,
1492 struct vm_area_struct *vma)
1494 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1495 __folio_add_file_rmap(folio, page, HPAGE_PMD_NR, vma, RMAP_LEVEL_PMD);
1501 static __always_inline void __folio_remove_rmap(struct folio *folio,
1502 struct page *page, int nr_pages, struct vm_area_struct *vma,
1503 enum rmap_level level)
1505 atomic_t *mapped = &folio->_nr_pages_mapped;
1506 int last, nr = 0, nr_pmdmapped = 0;
1507 enum node_stat_item idx;
1509 __folio_rmap_sanity_checks(folio, page, nr_pages, level);
1512 case RMAP_LEVEL_PTE:
1514 last = atomic_add_negative(-1, &page->_mapcount);
1515 if (last && folio_test_large(folio)) {
1516 last = atomic_dec_return_relaxed(mapped);
1517 last = (last < ENTIRELY_MAPPED);
1522 } while (page++, --nr_pages > 0);
1524 case RMAP_LEVEL_PMD:
1525 last = atomic_add_negative(-1, &folio->_entire_mapcount);
1527 nr = atomic_sub_return_relaxed(ENTIRELY_MAPPED, mapped);
1528 if (likely(nr < ENTIRELY_MAPPED)) {
1529 nr_pmdmapped = folio_nr_pages(folio);
1530 nr = nr_pmdmapped - (nr & FOLIO_PAGES_MAPPED);
1531 /* Raced ahead of another remove and an add? */
1532 if (unlikely(nr < 0))
1535 /* An add of ENTIRELY_MAPPED raced ahead */
1543 if (folio_test_anon(folio))
1545 else if (folio_test_swapbacked(folio))
1546 idx = NR_SHMEM_PMDMAPPED;
1548 idx = NR_FILE_PMDMAPPED;
1549 __lruvec_stat_mod_folio(folio, idx, -nr_pmdmapped);
1552 idx = folio_test_anon(folio) ? NR_ANON_MAPPED : NR_FILE_MAPPED;
1553 __lruvec_stat_mod_folio(folio, idx, -nr);
1556 * Queue anon large folio for deferred split if at least one
1557 * page of the folio is unmapped and at least one page
1560 if (folio_test_large(folio) && folio_test_anon(folio))
1561 if (level == RMAP_LEVEL_PTE || nr < nr_pmdmapped)
1562 deferred_split_folio(folio);
1566 * It would be tidy to reset folio_test_anon mapping when fully
1567 * unmapped, but that might overwrite a racing folio_add_anon_rmap_*()
1568 * which increments mapcount after us but sets mapping before us:
1569 * so leave the reset to free_pages_prepare, and remember that
1570 * it's only reliable while mapped.
1573 munlock_vma_folio(folio, vma);
1577 * folio_remove_rmap_ptes - remove PTE mappings from a page range of a folio
1578 * @folio: The folio to remove the mappings from
1579 * @page: The first page to remove
1580 * @nr_pages: The number of pages that will be removed from the mapping
1581 * @vma: The vm area from which the mappings are removed
1583 * The page range of the folio is defined by [page, page + nr_pages)
1585 * The caller needs to hold the page table lock.
1587 void folio_remove_rmap_ptes(struct folio *folio, struct page *page,
1588 int nr_pages, struct vm_area_struct *vma)
1590 __folio_remove_rmap(folio, page, nr_pages, vma, RMAP_LEVEL_PTE);
1594 * folio_remove_rmap_pmd - remove a PMD mapping from a page range of a folio
1595 * @folio: The folio to remove the mapping from
1596 * @page: The first page to remove
1597 * @vma: The vm area from which the mapping is removed
1599 * The page range of the folio is defined by [page, page + HPAGE_PMD_NR)
1601 * The caller needs to hold the page table lock.
1603 void folio_remove_rmap_pmd(struct folio *folio, struct page *page,
1604 struct vm_area_struct *vma)
1606 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1607 __folio_remove_rmap(folio, page, HPAGE_PMD_NR, vma, RMAP_LEVEL_PMD);
1614 * @arg: enum ttu_flags will be passed to this argument
1616 static bool try_to_unmap_one(struct folio *folio, struct vm_area_struct *vma,
1617 unsigned long address, void *arg)
1619 struct mm_struct *mm = vma->vm_mm;
1620 DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0);
1622 struct page *subpage;
1623 bool anon_exclusive, ret = true;
1624 struct mmu_notifier_range range;
1625 enum ttu_flags flags = (enum ttu_flags)(long)arg;
1627 unsigned long hsz = 0;
1630 * When racing against e.g. zap_pte_range() on another cpu,
1631 * in between its ptep_get_and_clear_full() and folio_remove_rmap_*(),
1632 * try_to_unmap() may return before page_mapped() has become false,
1633 * if page table locking is skipped: use TTU_SYNC to wait for that.
1635 if (flags & TTU_SYNC)
1636 pvmw.flags = PVMW_SYNC;
1638 if (flags & TTU_SPLIT_HUGE_PMD)
1639 split_huge_pmd_address(vma, address, false, folio);
1642 * For THP, we have to assume the worse case ie pmd for invalidation.
1643 * For hugetlb, it could be much worse if we need to do pud
1644 * invalidation in the case of pmd sharing.
1646 * Note that the folio can not be freed in this function as call of
1647 * try_to_unmap() must hold a reference on the folio.
1649 range.end = vma_address_end(&pvmw);
1650 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
1651 address, range.end);
1652 if (folio_test_hugetlb(folio)) {
1654 * If sharing is possible, start and end will be adjusted
1657 adjust_range_if_pmd_sharing_possible(vma, &range.start,
1660 /* We need the huge page size for set_huge_pte_at() */
1661 hsz = huge_page_size(hstate_vma(vma));
1663 mmu_notifier_invalidate_range_start(&range);
1665 while (page_vma_mapped_walk(&pvmw)) {
1666 /* Unexpected PMD-mapped THP? */
1667 VM_BUG_ON_FOLIO(!pvmw.pte, folio);
1670 * If the folio is in an mlock()d vma, we must not swap it out.
1672 if (!(flags & TTU_IGNORE_MLOCK) &&
1673 (vma->vm_flags & VM_LOCKED)) {
1674 /* Restore the mlock which got missed */
1675 if (!folio_test_large(folio))
1676 mlock_vma_folio(folio, vma);
1677 page_vma_mapped_walk_done(&pvmw);
1682 pfn = pte_pfn(ptep_get(pvmw.pte));
1683 subpage = folio_page(folio, pfn - folio_pfn(folio));
1684 address = pvmw.address;
1685 anon_exclusive = folio_test_anon(folio) &&
1686 PageAnonExclusive(subpage);
1688 if (folio_test_hugetlb(folio)) {
1689 bool anon = folio_test_anon(folio);
1692 * The try_to_unmap() is only passed a hugetlb page
1693 * in the case where the hugetlb page is poisoned.
1695 VM_BUG_ON_PAGE(!PageHWPoison(subpage), subpage);
1697 * huge_pmd_unshare may unmap an entire PMD page.
1698 * There is no way of knowing exactly which PMDs may
1699 * be cached for this mm, so we must flush them all.
1700 * start/end were already adjusted above to cover this
1703 flush_cache_range(vma, range.start, range.end);
1706 * To call huge_pmd_unshare, i_mmap_rwsem must be
1707 * held in write mode. Caller needs to explicitly
1708 * do this outside rmap routines.
1710 * We also must hold hugetlb vma_lock in write mode.
1711 * Lock order dictates acquiring vma_lock BEFORE
1712 * i_mmap_rwsem. We can only try lock here and fail
1716 VM_BUG_ON(!(flags & TTU_RMAP_LOCKED));
1717 if (!hugetlb_vma_trylock_write(vma)) {
1718 page_vma_mapped_walk_done(&pvmw);
1722 if (huge_pmd_unshare(mm, vma, address, pvmw.pte)) {
1723 hugetlb_vma_unlock_write(vma);
1724 flush_tlb_range(vma,
1725 range.start, range.end);
1727 * The ref count of the PMD page was
1728 * dropped which is part of the way map
1729 * counting is done for shared PMDs.
1730 * Return 'true' here. When there is
1731 * no other sharing, huge_pmd_unshare
1732 * returns false and we will unmap the
1733 * actual page and drop map count
1736 page_vma_mapped_walk_done(&pvmw);
1739 hugetlb_vma_unlock_write(vma);
1741 pteval = huge_ptep_clear_flush(vma, address, pvmw.pte);
1743 flush_cache_page(vma, address, pfn);
1744 /* Nuke the page table entry. */
1745 if (should_defer_flush(mm, flags)) {
1747 * We clear the PTE but do not flush so potentially
1748 * a remote CPU could still be writing to the folio.
1749 * If the entry was previously clean then the
1750 * architecture must guarantee that a clear->dirty
1751 * transition on a cached TLB entry is written through
1752 * and traps if the PTE is unmapped.
1754 pteval = ptep_get_and_clear(mm, address, pvmw.pte);
1756 set_tlb_ubc_flush_pending(mm, pteval, address);
1758 pteval = ptep_clear_flush(vma, address, pvmw.pte);
1763 * Now the pte is cleared. If this pte was uffd-wp armed,
1764 * we may want to replace a none pte with a marker pte if
1765 * it's file-backed, so we don't lose the tracking info.
1767 pte_install_uffd_wp_if_needed(vma, address, pvmw.pte, pteval);
1769 /* Set the dirty flag on the folio now the pte is gone. */
1770 if (pte_dirty(pteval))
1771 folio_mark_dirty(folio);
1773 /* Update high watermark before we lower rss */
1774 update_hiwater_rss(mm);
1776 if (PageHWPoison(subpage) && (flags & TTU_HWPOISON)) {
1777 pteval = swp_entry_to_pte(make_hwpoison_entry(subpage));
1778 if (folio_test_hugetlb(folio)) {
1779 hugetlb_count_sub(folio_nr_pages(folio), mm);
1780 set_huge_pte_at(mm, address, pvmw.pte, pteval,
1783 dec_mm_counter(mm, mm_counter(folio));
1784 set_pte_at(mm, address, pvmw.pte, pteval);
1787 } else if (pte_unused(pteval) && !userfaultfd_armed(vma)) {
1789 * The guest indicated that the page content is of no
1790 * interest anymore. Simply discard the pte, vmscan
1791 * will take care of the rest.
1792 * A future reference will then fault in a new zero
1793 * page. When userfaultfd is active, we must not drop
1794 * this page though, as its main user (postcopy
1795 * migration) will not expect userfaults on already
1798 dec_mm_counter(mm, mm_counter(folio));
1799 } else if (folio_test_anon(folio)) {
1800 swp_entry_t entry = page_swap_entry(subpage);
1803 * Store the swap location in the pte.
1804 * See handle_pte_fault() ...
1806 if (unlikely(folio_test_swapbacked(folio) !=
1807 folio_test_swapcache(folio))) {
1810 page_vma_mapped_walk_done(&pvmw);
1814 /* MADV_FREE page check */
1815 if (!folio_test_swapbacked(folio)) {
1816 int ref_count, map_count;
1819 * Synchronize with gup_pte_range():
1820 * - clear PTE; barrier; read refcount
1821 * - inc refcount; barrier; read PTE
1825 ref_count = folio_ref_count(folio);
1826 map_count = folio_mapcount(folio);
1829 * Order reads for page refcount and dirty flag
1830 * (see comments in __remove_mapping()).
1835 * The only page refs must be one from isolation
1836 * plus the rmap(s) (dropped by discard:).
1838 if (ref_count == 1 + map_count &&
1839 !folio_test_dirty(folio)) {
1840 dec_mm_counter(mm, MM_ANONPAGES);
1845 * If the folio was redirtied, it cannot be
1846 * discarded. Remap the page to page table.
1848 set_pte_at(mm, address, pvmw.pte, pteval);
1849 folio_set_swapbacked(folio);
1851 page_vma_mapped_walk_done(&pvmw);
1855 if (swap_duplicate(entry) < 0) {
1856 set_pte_at(mm, address, pvmw.pte, pteval);
1858 page_vma_mapped_walk_done(&pvmw);
1861 if (arch_unmap_one(mm, vma, address, pteval) < 0) {
1863 set_pte_at(mm, address, pvmw.pte, pteval);
1865 page_vma_mapped_walk_done(&pvmw);
1869 /* See folio_try_share_anon_rmap(): clear PTE first. */
1870 if (anon_exclusive &&
1871 folio_try_share_anon_rmap_pte(folio, subpage)) {
1873 set_pte_at(mm, address, pvmw.pte, pteval);
1875 page_vma_mapped_walk_done(&pvmw);
1878 if (list_empty(&mm->mmlist)) {
1879 spin_lock(&mmlist_lock);
1880 if (list_empty(&mm->mmlist))
1881 list_add(&mm->mmlist, &init_mm.mmlist);
1882 spin_unlock(&mmlist_lock);
1884 dec_mm_counter(mm, MM_ANONPAGES);
1885 inc_mm_counter(mm, MM_SWAPENTS);
1886 swp_pte = swp_entry_to_pte(entry);
1888 swp_pte = pte_swp_mkexclusive(swp_pte);
1889 if (pte_soft_dirty(pteval))
1890 swp_pte = pte_swp_mksoft_dirty(swp_pte);
1891 if (pte_uffd_wp(pteval))
1892 swp_pte = pte_swp_mkuffd_wp(swp_pte);
1893 set_pte_at(mm, address, pvmw.pte, swp_pte);
1896 * This is a locked file-backed folio,
1897 * so it cannot be removed from the page
1898 * cache and replaced by a new folio before
1899 * mmu_notifier_invalidate_range_end, so no
1900 * concurrent thread might update its page table
1901 * to point at a new folio while a device is
1902 * still using this folio.
1904 * See Documentation/mm/mmu_notifier.rst
1906 dec_mm_counter(mm, mm_counter_file(folio));
1909 if (unlikely(folio_test_hugetlb(folio)))
1910 hugetlb_remove_rmap(folio);
1912 folio_remove_rmap_pte(folio, subpage, vma);
1913 if (vma->vm_flags & VM_LOCKED)
1914 mlock_drain_local();
1918 mmu_notifier_invalidate_range_end(&range);
1923 static bool invalid_migration_vma(struct vm_area_struct *vma, void *arg)
1925 return vma_is_temporary_stack(vma);
1928 static int folio_not_mapped(struct folio *folio)
1930 return !folio_mapped(folio);
1934 * try_to_unmap - Try to remove all page table mappings to a folio.
1935 * @folio: The folio to unmap.
1936 * @flags: action and flags
1938 * Tries to remove all the page table entries which are mapping this
1939 * folio. It is the caller's responsibility to check if the folio is
1940 * still mapped if needed (use TTU_SYNC to prevent accounting races).
1942 * Context: Caller must hold the folio lock.
1944 void try_to_unmap(struct folio *folio, enum ttu_flags flags)
1946 struct rmap_walk_control rwc = {
1947 .rmap_one = try_to_unmap_one,
1948 .arg = (void *)flags,
1949 .done = folio_not_mapped,
1950 .anon_lock = folio_lock_anon_vma_read,
1953 if (flags & TTU_RMAP_LOCKED)
1954 rmap_walk_locked(folio, &rwc);
1956 rmap_walk(folio, &rwc);
1960 * @arg: enum ttu_flags will be passed to this argument.
1962 * If TTU_SPLIT_HUGE_PMD is specified any PMD mappings will be split into PTEs
1963 * containing migration entries.
1965 static bool try_to_migrate_one(struct folio *folio, struct vm_area_struct *vma,
1966 unsigned long address, void *arg)
1968 struct mm_struct *mm = vma->vm_mm;
1969 DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0);
1971 struct page *subpage;
1972 bool anon_exclusive, ret = true;
1973 struct mmu_notifier_range range;
1974 enum ttu_flags flags = (enum ttu_flags)(long)arg;
1976 unsigned long hsz = 0;
1979 * When racing against e.g. zap_pte_range() on another cpu,
1980 * in between its ptep_get_and_clear_full() and folio_remove_rmap_*(),
1981 * try_to_migrate() may return before page_mapped() has become false,
1982 * if page table locking is skipped: use TTU_SYNC to wait for that.
1984 if (flags & TTU_SYNC)
1985 pvmw.flags = PVMW_SYNC;
1988 * unmap_page() in mm/huge_memory.c is the only user of migration with
1989 * TTU_SPLIT_HUGE_PMD and it wants to freeze.
1991 if (flags & TTU_SPLIT_HUGE_PMD)
1992 split_huge_pmd_address(vma, address, true, folio);
1995 * For THP, we have to assume the worse case ie pmd for invalidation.
1996 * For hugetlb, it could be much worse if we need to do pud
1997 * invalidation in the case of pmd sharing.
1999 * Note that the page can not be free in this function as call of
2000 * try_to_unmap() must hold a reference on the page.
2002 range.end = vma_address_end(&pvmw);
2003 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
2004 address, range.end);
2005 if (folio_test_hugetlb(folio)) {
2007 * If sharing is possible, start and end will be adjusted
2010 adjust_range_if_pmd_sharing_possible(vma, &range.start,
2013 /* We need the huge page size for set_huge_pte_at() */
2014 hsz = huge_page_size(hstate_vma(vma));
2016 mmu_notifier_invalidate_range_start(&range);
2018 while (page_vma_mapped_walk(&pvmw)) {
2019 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2020 /* PMD-mapped THP migration entry */
2022 subpage = folio_page(folio,
2023 pmd_pfn(*pvmw.pmd) - folio_pfn(folio));
2024 VM_BUG_ON_FOLIO(folio_test_hugetlb(folio) ||
2025 !folio_test_pmd_mappable(folio), folio);
2027 if (set_pmd_migration_entry(&pvmw, subpage)) {
2029 page_vma_mapped_walk_done(&pvmw);
2036 /* Unexpected PMD-mapped THP? */
2037 VM_BUG_ON_FOLIO(!pvmw.pte, folio);
2039 pfn = pte_pfn(ptep_get(pvmw.pte));
2041 if (folio_is_zone_device(folio)) {
2043 * Our PTE is a non-present device exclusive entry and
2044 * calculating the subpage as for the common case would
2045 * result in an invalid pointer.
2047 * Since only PAGE_SIZE pages can currently be
2048 * migrated, just set it to page. This will need to be
2049 * changed when hugepage migrations to device private
2050 * memory are supported.
2052 VM_BUG_ON_FOLIO(folio_nr_pages(folio) > 1, folio);
2053 subpage = &folio->page;
2055 subpage = folio_page(folio, pfn - folio_pfn(folio));
2057 address = pvmw.address;
2058 anon_exclusive = folio_test_anon(folio) &&
2059 PageAnonExclusive(subpage);
2061 if (folio_test_hugetlb(folio)) {
2062 bool anon = folio_test_anon(folio);
2065 * huge_pmd_unshare may unmap an entire PMD page.
2066 * There is no way of knowing exactly which PMDs may
2067 * be cached for this mm, so we must flush them all.
2068 * start/end were already adjusted above to cover this
2071 flush_cache_range(vma, range.start, range.end);
2074 * To call huge_pmd_unshare, i_mmap_rwsem must be
2075 * held in write mode. Caller needs to explicitly
2076 * do this outside rmap routines.
2078 * We also must hold hugetlb vma_lock in write mode.
2079 * Lock order dictates acquiring vma_lock BEFORE
2080 * i_mmap_rwsem. We can only try lock here and
2081 * fail if unsuccessful.
2084 VM_BUG_ON(!(flags & TTU_RMAP_LOCKED));
2085 if (!hugetlb_vma_trylock_write(vma)) {
2086 page_vma_mapped_walk_done(&pvmw);
2090 if (huge_pmd_unshare(mm, vma, address, pvmw.pte)) {
2091 hugetlb_vma_unlock_write(vma);
2092 flush_tlb_range(vma,
2093 range.start, range.end);
2096 * The ref count of the PMD page was
2097 * dropped which is part of the way map
2098 * counting is done for shared PMDs.
2099 * Return 'true' here. When there is
2100 * no other sharing, huge_pmd_unshare
2101 * returns false and we will unmap the
2102 * actual page and drop map count
2105 page_vma_mapped_walk_done(&pvmw);
2108 hugetlb_vma_unlock_write(vma);
2110 /* Nuke the hugetlb page table entry */
2111 pteval = huge_ptep_clear_flush(vma, address, pvmw.pte);
2113 flush_cache_page(vma, address, pfn);
2114 /* Nuke the page table entry. */
2115 if (should_defer_flush(mm, flags)) {
2117 * We clear the PTE but do not flush so potentially
2118 * a remote CPU could still be writing to the folio.
2119 * If the entry was previously clean then the
2120 * architecture must guarantee that a clear->dirty
2121 * transition on a cached TLB entry is written through
2122 * and traps if the PTE is unmapped.
2124 pteval = ptep_get_and_clear(mm, address, pvmw.pte);
2126 set_tlb_ubc_flush_pending(mm, pteval, address);
2128 pteval = ptep_clear_flush(vma, address, pvmw.pte);
2132 /* Set the dirty flag on the folio now the pte is gone. */
2133 if (pte_dirty(pteval))
2134 folio_mark_dirty(folio);
2136 /* Update high watermark before we lower rss */
2137 update_hiwater_rss(mm);
2139 if (folio_is_device_private(folio)) {
2140 unsigned long pfn = folio_pfn(folio);
2145 WARN_ON_ONCE(folio_try_share_anon_rmap_pte(folio,
2149 * Store the pfn of the page in a special migration
2150 * pte. do_swap_page() will wait until the migration
2151 * pte is removed and then restart fault handling.
2153 entry = pte_to_swp_entry(pteval);
2154 if (is_writable_device_private_entry(entry))
2155 entry = make_writable_migration_entry(pfn);
2156 else if (anon_exclusive)
2157 entry = make_readable_exclusive_migration_entry(pfn);
2159 entry = make_readable_migration_entry(pfn);
2160 swp_pte = swp_entry_to_pte(entry);
2163 * pteval maps a zone device page and is therefore
2166 if (pte_swp_soft_dirty(pteval))
2167 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2168 if (pte_swp_uffd_wp(pteval))
2169 swp_pte = pte_swp_mkuffd_wp(swp_pte);
2170 set_pte_at(mm, pvmw.address, pvmw.pte, swp_pte);
2171 trace_set_migration_pte(pvmw.address, pte_val(swp_pte),
2172 folio_order(folio));
2174 * No need to invalidate here it will synchronize on
2175 * against the special swap migration pte.
2177 } else if (PageHWPoison(subpage)) {
2178 pteval = swp_entry_to_pte(make_hwpoison_entry(subpage));
2179 if (folio_test_hugetlb(folio)) {
2180 hugetlb_count_sub(folio_nr_pages(folio), mm);
2181 set_huge_pte_at(mm, address, pvmw.pte, pteval,
2184 dec_mm_counter(mm, mm_counter(folio));
2185 set_pte_at(mm, address, pvmw.pte, pteval);
2188 } else if (pte_unused(pteval) && !userfaultfd_armed(vma)) {
2190 * The guest indicated that the page content is of no
2191 * interest anymore. Simply discard the pte, vmscan
2192 * will take care of the rest.
2193 * A future reference will then fault in a new zero
2194 * page. When userfaultfd is active, we must not drop
2195 * this page though, as its main user (postcopy
2196 * migration) will not expect userfaults on already
2199 dec_mm_counter(mm, mm_counter(folio));
2204 if (arch_unmap_one(mm, vma, address, pteval) < 0) {
2205 if (folio_test_hugetlb(folio))
2206 set_huge_pte_at(mm, address, pvmw.pte,
2209 set_pte_at(mm, address, pvmw.pte, pteval);
2211 page_vma_mapped_walk_done(&pvmw);
2214 VM_BUG_ON_PAGE(pte_write(pteval) && folio_test_anon(folio) &&
2215 !anon_exclusive, subpage);
2217 /* See folio_try_share_anon_rmap_pte(): clear PTE first. */
2218 if (folio_test_hugetlb(folio)) {
2219 if (anon_exclusive &&
2220 hugetlb_try_share_anon_rmap(folio)) {
2221 set_huge_pte_at(mm, address, pvmw.pte,
2224 page_vma_mapped_walk_done(&pvmw);
2227 } else if (anon_exclusive &&
2228 folio_try_share_anon_rmap_pte(folio, subpage)) {
2229 set_pte_at(mm, address, pvmw.pte, pteval);
2231 page_vma_mapped_walk_done(&pvmw);
2236 * Store the pfn of the page in a special migration
2237 * pte. do_swap_page() will wait until the migration
2238 * pte is removed and then restart fault handling.
2240 if (pte_write(pteval))
2241 entry = make_writable_migration_entry(
2242 page_to_pfn(subpage));
2243 else if (anon_exclusive)
2244 entry = make_readable_exclusive_migration_entry(
2245 page_to_pfn(subpage));
2247 entry = make_readable_migration_entry(
2248 page_to_pfn(subpage));
2249 if (pte_young(pteval))
2250 entry = make_migration_entry_young(entry);
2251 if (pte_dirty(pteval))
2252 entry = make_migration_entry_dirty(entry);
2253 swp_pte = swp_entry_to_pte(entry);
2254 if (pte_soft_dirty(pteval))
2255 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2256 if (pte_uffd_wp(pteval))
2257 swp_pte = pte_swp_mkuffd_wp(swp_pte);
2258 if (folio_test_hugetlb(folio))
2259 set_huge_pte_at(mm, address, pvmw.pte, swp_pte,
2262 set_pte_at(mm, address, pvmw.pte, swp_pte);
2263 trace_set_migration_pte(address, pte_val(swp_pte),
2264 folio_order(folio));
2266 * No need to invalidate here it will synchronize on
2267 * against the special swap migration pte.
2271 if (unlikely(folio_test_hugetlb(folio)))
2272 hugetlb_remove_rmap(folio);
2274 folio_remove_rmap_pte(folio, subpage, vma);
2275 if (vma->vm_flags & VM_LOCKED)
2276 mlock_drain_local();
2280 mmu_notifier_invalidate_range_end(&range);
2286 * try_to_migrate - try to replace all page table mappings with swap entries
2287 * @folio: the folio to replace page table entries for
2288 * @flags: action and flags
2290 * Tries to remove all the page table entries which are mapping this folio and
2291 * replace them with special swap entries. Caller must hold the folio lock.
2293 void try_to_migrate(struct folio *folio, enum ttu_flags flags)
2295 struct rmap_walk_control rwc = {
2296 .rmap_one = try_to_migrate_one,
2297 .arg = (void *)flags,
2298 .done = folio_not_mapped,
2299 .anon_lock = folio_lock_anon_vma_read,
2303 * Migration always ignores mlock and only supports TTU_RMAP_LOCKED and
2304 * TTU_SPLIT_HUGE_PMD, TTU_SYNC, and TTU_BATCH_FLUSH flags.
2306 if (WARN_ON_ONCE(flags & ~(TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD |
2307 TTU_SYNC | TTU_BATCH_FLUSH)))
2310 if (folio_is_zone_device(folio) &&
2311 (!folio_is_device_private(folio) && !folio_is_device_coherent(folio)))
2315 * During exec, a temporary VMA is setup and later moved.
2316 * The VMA is moved under the anon_vma lock but not the
2317 * page tables leading to a race where migration cannot
2318 * find the migration ptes. Rather than increasing the
2319 * locking requirements of exec(), migration skips
2320 * temporary VMAs until after exec() completes.
2322 if (!folio_test_ksm(folio) && folio_test_anon(folio))
2323 rwc.invalid_vma = invalid_migration_vma;
2325 if (flags & TTU_RMAP_LOCKED)
2326 rmap_walk_locked(folio, &rwc);
2328 rmap_walk(folio, &rwc);
2331 #ifdef CONFIG_DEVICE_PRIVATE
2332 struct make_exclusive_args {
2333 struct mm_struct *mm;
2334 unsigned long address;
2339 static bool page_make_device_exclusive_one(struct folio *folio,
2340 struct vm_area_struct *vma, unsigned long address, void *priv)
2342 struct mm_struct *mm = vma->vm_mm;
2343 DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0);
2344 struct make_exclusive_args *args = priv;
2346 struct page *subpage;
2348 struct mmu_notifier_range range;
2353 mmu_notifier_range_init_owner(&range, MMU_NOTIFY_EXCLUSIVE, 0,
2354 vma->vm_mm, address, min(vma->vm_end,
2355 address + folio_size(folio)),
2357 mmu_notifier_invalidate_range_start(&range);
2359 while (page_vma_mapped_walk(&pvmw)) {
2360 /* Unexpected PMD-mapped THP? */
2361 VM_BUG_ON_FOLIO(!pvmw.pte, folio);
2363 ptent = ptep_get(pvmw.pte);
2364 if (!pte_present(ptent)) {
2366 page_vma_mapped_walk_done(&pvmw);
2370 subpage = folio_page(folio,
2371 pte_pfn(ptent) - folio_pfn(folio));
2372 address = pvmw.address;
2374 /* Nuke the page table entry. */
2375 flush_cache_page(vma, address, pte_pfn(ptent));
2376 pteval = ptep_clear_flush(vma, address, pvmw.pte);
2378 /* Set the dirty flag on the folio now the pte is gone. */
2379 if (pte_dirty(pteval))
2380 folio_mark_dirty(folio);
2383 * Check that our target page is still mapped at the expected
2386 if (args->mm == mm && args->address == address &&
2391 * Store the pfn of the page in a special migration
2392 * pte. do_swap_page() will wait until the migration
2393 * pte is removed and then restart fault handling.
2395 if (pte_write(pteval))
2396 entry = make_writable_device_exclusive_entry(
2397 page_to_pfn(subpage));
2399 entry = make_readable_device_exclusive_entry(
2400 page_to_pfn(subpage));
2401 swp_pte = swp_entry_to_pte(entry);
2402 if (pte_soft_dirty(pteval))
2403 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2404 if (pte_uffd_wp(pteval))
2405 swp_pte = pte_swp_mkuffd_wp(swp_pte);
2407 set_pte_at(mm, address, pvmw.pte, swp_pte);
2410 * There is a reference on the page for the swap entry which has
2411 * been removed, so shouldn't take another.
2413 folio_remove_rmap_pte(folio, subpage, vma);
2416 mmu_notifier_invalidate_range_end(&range);
2422 * folio_make_device_exclusive - Mark the folio exclusively owned by a device.
2423 * @folio: The folio to replace page table entries for.
2424 * @mm: The mm_struct where the folio is expected to be mapped.
2425 * @address: Address where the folio is expected to be mapped.
2426 * @owner: passed to MMU_NOTIFY_EXCLUSIVE range notifier callbacks
2428 * Tries to remove all the page table entries which are mapping this
2429 * folio and replace them with special device exclusive swap entries to
2430 * grant a device exclusive access to the folio.
2432 * Context: Caller must hold the folio lock.
2433 * Return: false if the page is still mapped, or if it could not be unmapped
2434 * from the expected address. Otherwise returns true (success).
2436 static bool folio_make_device_exclusive(struct folio *folio,
2437 struct mm_struct *mm, unsigned long address, void *owner)
2439 struct make_exclusive_args args = {
2445 struct rmap_walk_control rwc = {
2446 .rmap_one = page_make_device_exclusive_one,
2447 .done = folio_not_mapped,
2448 .anon_lock = folio_lock_anon_vma_read,
2453 * Restrict to anonymous folios for now to avoid potential writeback
2456 if (!folio_test_anon(folio))
2459 rmap_walk(folio, &rwc);
2461 return args.valid && !folio_mapcount(folio);
2465 * make_device_exclusive_range() - Mark a range for exclusive use by a device
2466 * @mm: mm_struct of associated target process
2467 * @start: start of the region to mark for exclusive device access
2468 * @end: end address of region
2469 * @pages: returns the pages which were successfully marked for exclusive access
2470 * @owner: passed to MMU_NOTIFY_EXCLUSIVE range notifier to allow filtering
2472 * Returns: number of pages found in the range by GUP. A page is marked for
2473 * exclusive access only if the page pointer is non-NULL.
2475 * This function finds ptes mapping page(s) to the given address range, locks
2476 * them and replaces mappings with special swap entries preventing userspace CPU
2477 * access. On fault these entries are replaced with the original mapping after
2478 * calling MMU notifiers.
2480 * A driver using this to program access from a device must use a mmu notifier
2481 * critical section to hold a device specific lock during programming. Once
2482 * programming is complete it should drop the page lock and reference after
2483 * which point CPU access to the page will revoke the exclusive access.
2485 int make_device_exclusive_range(struct mm_struct *mm, unsigned long start,
2486 unsigned long end, struct page **pages,
2489 long npages = (end - start) >> PAGE_SHIFT;
2492 npages = get_user_pages_remote(mm, start, npages,
2493 FOLL_GET | FOLL_WRITE | FOLL_SPLIT_PMD,
2498 for (i = 0; i < npages; i++, start += PAGE_SIZE) {
2499 struct folio *folio = page_folio(pages[i]);
2500 if (PageTail(pages[i]) || !folio_trylock(folio)) {
2506 if (!folio_make_device_exclusive(folio, mm, start, owner)) {
2507 folio_unlock(folio);
2515 EXPORT_SYMBOL_GPL(make_device_exclusive_range);
2518 void __put_anon_vma(struct anon_vma *anon_vma)
2520 struct anon_vma *root = anon_vma->root;
2522 anon_vma_free(anon_vma);
2523 if (root != anon_vma && atomic_dec_and_test(&root->refcount))
2524 anon_vma_free(root);
2527 static struct anon_vma *rmap_walk_anon_lock(struct folio *folio,
2528 struct rmap_walk_control *rwc)
2530 struct anon_vma *anon_vma;
2533 return rwc->anon_lock(folio, rwc);
2536 * Note: remove_migration_ptes() cannot use folio_lock_anon_vma_read()
2537 * because that depends on page_mapped(); but not all its usages
2538 * are holding mmap_lock. Users without mmap_lock are required to
2539 * take a reference count to prevent the anon_vma disappearing
2541 anon_vma = folio_anon_vma(folio);
2545 if (anon_vma_trylock_read(anon_vma))
2548 if (rwc->try_lock) {
2550 rwc->contended = true;
2554 anon_vma_lock_read(anon_vma);
2560 * rmap_walk_anon - do something to anonymous page using the object-based
2562 * @folio: the folio to be handled
2563 * @rwc: control variable according to each walk type
2564 * @locked: caller holds relevant rmap lock
2566 * Find all the mappings of a folio using the mapping pointer and the vma
2567 * chains contained in the anon_vma struct it points to.
2569 static void rmap_walk_anon(struct folio *folio,
2570 struct rmap_walk_control *rwc, bool locked)
2572 struct anon_vma *anon_vma;
2573 pgoff_t pgoff_start, pgoff_end;
2574 struct anon_vma_chain *avc;
2577 anon_vma = folio_anon_vma(folio);
2578 /* anon_vma disappear under us? */
2579 VM_BUG_ON_FOLIO(!anon_vma, folio);
2581 anon_vma = rmap_walk_anon_lock(folio, rwc);
2586 pgoff_start = folio_pgoff(folio);
2587 pgoff_end = pgoff_start + folio_nr_pages(folio) - 1;
2588 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root,
2589 pgoff_start, pgoff_end) {
2590 struct vm_area_struct *vma = avc->vma;
2591 unsigned long address = vma_address(&folio->page, vma);
2593 VM_BUG_ON_VMA(address == -EFAULT, vma);
2596 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2599 if (!rwc->rmap_one(folio, vma, address, rwc->arg))
2601 if (rwc->done && rwc->done(folio))
2606 anon_vma_unlock_read(anon_vma);
2610 * rmap_walk_file - do something to file page using the object-based rmap method
2611 * @folio: the folio to be handled
2612 * @rwc: control variable according to each walk type
2613 * @locked: caller holds relevant rmap lock
2615 * Find all the mappings of a folio using the mapping pointer and the vma chains
2616 * contained in the address_space struct it points to.
2618 static void rmap_walk_file(struct folio *folio,
2619 struct rmap_walk_control *rwc, bool locked)
2621 struct address_space *mapping = folio_mapping(folio);
2622 pgoff_t pgoff_start, pgoff_end;
2623 struct vm_area_struct *vma;
2626 * The page lock not only makes sure that page->mapping cannot
2627 * suddenly be NULLified by truncation, it makes sure that the
2628 * structure at mapping cannot be freed and reused yet,
2629 * so we can safely take mapping->i_mmap_rwsem.
2631 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
2636 pgoff_start = folio_pgoff(folio);
2637 pgoff_end = pgoff_start + folio_nr_pages(folio) - 1;
2639 if (i_mmap_trylock_read(mapping))
2642 if (rwc->try_lock) {
2643 rwc->contended = true;
2647 i_mmap_lock_read(mapping);
2650 vma_interval_tree_foreach(vma, &mapping->i_mmap,
2651 pgoff_start, pgoff_end) {
2652 unsigned long address = vma_address(&folio->page, vma);
2654 VM_BUG_ON_VMA(address == -EFAULT, vma);
2657 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2660 if (!rwc->rmap_one(folio, vma, address, rwc->arg))
2662 if (rwc->done && rwc->done(folio))
2668 i_mmap_unlock_read(mapping);
2671 void rmap_walk(struct folio *folio, struct rmap_walk_control *rwc)
2673 if (unlikely(folio_test_ksm(folio)))
2674 rmap_walk_ksm(folio, rwc);
2675 else if (folio_test_anon(folio))
2676 rmap_walk_anon(folio, rwc, false);
2678 rmap_walk_file(folio, rwc, false);
2681 /* Like rmap_walk, but caller holds relevant rmap lock */
2682 void rmap_walk_locked(struct folio *folio, struct rmap_walk_control *rwc)
2684 /* no ksm support for now */
2685 VM_BUG_ON_FOLIO(folio_test_ksm(folio), folio);
2686 if (folio_test_anon(folio))
2687 rmap_walk_anon(folio, rwc, true);
2689 rmap_walk_file(folio, rwc, true);
2692 #ifdef CONFIG_HUGETLB_PAGE
2694 * The following two functions are for anonymous (private mapped) hugepages.
2695 * Unlike common anonymous pages, anonymous hugepages have no accounting code
2696 * and no lru code, because we handle hugepages differently from common pages.
2698 void hugetlb_add_anon_rmap(struct folio *folio, struct vm_area_struct *vma,
2699 unsigned long address, rmap_t flags)
2701 VM_WARN_ON_FOLIO(!folio_test_hugetlb(folio), folio);
2702 VM_WARN_ON_FOLIO(!folio_test_anon(folio), folio);
2704 atomic_inc(&folio->_entire_mapcount);
2705 if (flags & RMAP_EXCLUSIVE)
2706 SetPageAnonExclusive(&folio->page);
2707 VM_WARN_ON_FOLIO(folio_entire_mapcount(folio) > 1 &&
2708 PageAnonExclusive(&folio->page), folio);
2711 void hugetlb_add_new_anon_rmap(struct folio *folio,
2712 struct vm_area_struct *vma, unsigned long address)
2714 VM_WARN_ON_FOLIO(!folio_test_hugetlb(folio), folio);
2716 BUG_ON(address < vma->vm_start || address >= vma->vm_end);
2717 /* increment count (starts at -1) */
2718 atomic_set(&folio->_entire_mapcount, 0);
2719 folio_clear_hugetlb_restore_reserve(folio);
2720 __folio_set_anon(folio, vma, address, true);
2721 SetPageAnonExclusive(&folio->page);
2723 #endif /* CONFIG_HUGETLB_PAGE */