2 * mm/rmap.c - physical to virtual reverse mappings
4 * Copyright 2001, Rik van Riel <riel@conectiva.com.br>
5 * Released under the General Public License (GPL).
7 * Simple, low overhead reverse mapping scheme.
8 * Please try to keep this thing as modular as possible.
10 * Provides methods for unmapping each kind of mapped page:
11 * the anon methods track anonymous pages, and
12 * the file methods track pages belonging to an inode.
14 * Original design by Rik van Riel <riel@conectiva.com.br> 2001
15 * File methods by Dave McCracken <dmccr@us.ibm.com> 2003, 2004
16 * Anonymous methods by Andrea Arcangeli <andrea@suse.de> 2004
17 * Contributions by Hugh Dickins 2003, 2004
21 * Lock ordering in mm:
23 * inode->i_rwsem (while writing or truncating, not reading or faulting)
25 * mapping->invalidate_lock (in filemap_fault)
26 * page->flags PG_locked (lock_page) * (see hugetlbfs below)
27 * hugetlbfs_i_mmap_rwsem_key (in huge_pmd_share)
28 * mapping->i_mmap_rwsem
29 * hugetlb_fault_mutex (hugetlbfs specific page fault mutex)
31 * mm->page_table_lock or pte_lock
32 * swap_lock (in swap_duplicate, swap_info_get)
33 * mmlist_lock (in mmput, drain_mmlist and others)
34 * mapping->private_lock (in 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() pages take locks in this order:
50 * mapping->i_mmap_rwsem
51 * hugetlb_fault_mutex (hugetlbfs specific page fault mutex)
52 * page->flags PG_locked (lock_page)
56 #include <linux/sched/mm.h>
57 #include <linux/sched/task.h>
58 #include <linux/pagemap.h>
59 #include <linux/swap.h>
60 #include <linux/swapops.h>
61 #include <linux/slab.h>
62 #include <linux/init.h>
63 #include <linux/ksm.h>
64 #include <linux/rmap.h>
65 #include <linux/rcupdate.h>
66 #include <linux/export.h>
67 #include <linux/memcontrol.h>
68 #include <linux/mmu_notifier.h>
69 #include <linux/migrate.h>
70 #include <linux/hugetlb.h>
71 #include <linux/huge_mm.h>
72 #include <linux/backing-dev.h>
73 #include <linux/page_idle.h>
74 #include <linux/memremap.h>
75 #include <linux/userfaultfd_k.h>
76 #include <linux/mm_inline.h>
78 #include <asm/tlbflush.h>
80 #define CREATE_TRACE_POINTS
81 #include <trace/events/tlb.h>
82 #include <trace/events/migrate.h>
86 static struct kmem_cache *anon_vma_cachep;
87 static struct kmem_cache *anon_vma_chain_cachep;
89 static inline struct anon_vma *anon_vma_alloc(void)
91 struct anon_vma *anon_vma;
93 anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
95 atomic_set(&anon_vma->refcount, 1);
96 anon_vma->degree = 1; /* Reference for first vma */
97 anon_vma->parent = anon_vma;
99 * Initialise the anon_vma root to point to itself. If called
100 * from fork, the root will be reset to the parents anon_vma.
102 anon_vma->root = anon_vma;
108 static inline void anon_vma_free(struct anon_vma *anon_vma)
110 VM_BUG_ON(atomic_read(&anon_vma->refcount));
113 * Synchronize against folio_lock_anon_vma_read() such that
114 * we can safely hold the lock without the anon_vma getting
117 * Relies on the full mb implied by the atomic_dec_and_test() from
118 * put_anon_vma() against the acquire barrier implied by
119 * down_read_trylock() from folio_lock_anon_vma_read(). This orders:
121 * folio_lock_anon_vma_read() VS put_anon_vma()
122 * down_read_trylock() atomic_dec_and_test()
124 * atomic_read() rwsem_is_locked()
126 * LOCK should suffice since the actual taking of the lock must
127 * happen _before_ what follows.
130 if (rwsem_is_locked(&anon_vma->root->rwsem)) {
131 anon_vma_lock_write(anon_vma);
132 anon_vma_unlock_write(anon_vma);
135 kmem_cache_free(anon_vma_cachep, anon_vma);
138 static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp)
140 return kmem_cache_alloc(anon_vma_chain_cachep, gfp);
143 static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
145 kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
148 static void anon_vma_chain_link(struct vm_area_struct *vma,
149 struct anon_vma_chain *avc,
150 struct anon_vma *anon_vma)
153 avc->anon_vma = anon_vma;
154 list_add(&avc->same_vma, &vma->anon_vma_chain);
155 anon_vma_interval_tree_insert(avc, &anon_vma->rb_root);
159 * __anon_vma_prepare - attach an anon_vma to a memory region
160 * @vma: the memory region in question
162 * This makes sure the memory mapping described by 'vma' has
163 * an 'anon_vma' attached to it, so that we can associate the
164 * anonymous pages mapped into it with that anon_vma.
166 * The common case will be that we already have one, which
167 * is handled inline by anon_vma_prepare(). But if
168 * not we either need to find an adjacent mapping that we
169 * can re-use the anon_vma from (very common when the only
170 * reason for splitting a vma has been mprotect()), or we
171 * allocate a new one.
173 * Anon-vma allocations are very subtle, because we may have
174 * optimistically looked up an anon_vma in folio_lock_anon_vma_read()
175 * and that may actually touch the rwsem even in the newly
176 * allocated vma (it depends on RCU to make sure that the
177 * anon_vma isn't actually destroyed).
179 * As a result, we need to do proper anon_vma locking even
180 * for the new allocation. At the same time, we do not want
181 * to do any locking for the common case of already having
184 * This must be called with the mmap_lock held for reading.
186 int __anon_vma_prepare(struct vm_area_struct *vma)
188 struct mm_struct *mm = vma->vm_mm;
189 struct anon_vma *anon_vma, *allocated;
190 struct anon_vma_chain *avc;
194 avc = anon_vma_chain_alloc(GFP_KERNEL);
198 anon_vma = find_mergeable_anon_vma(vma);
201 anon_vma = anon_vma_alloc();
202 if (unlikely(!anon_vma))
203 goto out_enomem_free_avc;
204 allocated = anon_vma;
207 anon_vma_lock_write(anon_vma);
208 /* page_table_lock to protect against threads */
209 spin_lock(&mm->page_table_lock);
210 if (likely(!vma->anon_vma)) {
211 vma->anon_vma = anon_vma;
212 anon_vma_chain_link(vma, avc, anon_vma);
213 /* vma reference or self-parent link for new root */
218 spin_unlock(&mm->page_table_lock);
219 anon_vma_unlock_write(anon_vma);
221 if (unlikely(allocated))
222 put_anon_vma(allocated);
224 anon_vma_chain_free(avc);
229 anon_vma_chain_free(avc);
235 * This is a useful helper function for locking the anon_vma root as
236 * we traverse the vma->anon_vma_chain, looping over anon_vma's that
239 * Such anon_vma's should have the same root, so you'd expect to see
240 * just a single mutex_lock for the whole traversal.
242 static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
244 struct anon_vma *new_root = anon_vma->root;
245 if (new_root != root) {
246 if (WARN_ON_ONCE(root))
247 up_write(&root->rwsem);
249 down_write(&root->rwsem);
254 static inline void unlock_anon_vma_root(struct anon_vma *root)
257 up_write(&root->rwsem);
261 * Attach the anon_vmas from src to dst.
262 * Returns 0 on success, -ENOMEM on failure.
264 * anon_vma_clone() is called by __vma_adjust(), __split_vma(), copy_vma() and
265 * anon_vma_fork(). The first three want an exact copy of src, while the last
266 * one, anon_vma_fork(), may try to reuse an existing anon_vma to prevent
267 * endless growth of anon_vma. Since dst->anon_vma is set to NULL before call,
268 * we can identify this case by checking (!dst->anon_vma && src->anon_vma).
270 * If (!dst->anon_vma && src->anon_vma) is true, this function tries to find
271 * and reuse existing anon_vma which has no vmas and only one child anon_vma.
272 * This prevents degradation of anon_vma hierarchy to endless linear chain in
273 * case of constantly forking task. On the other hand, an anon_vma with more
274 * than one child isn't reused even if there was no alive vma, thus rmap
275 * walker has a good chance of avoiding scanning the whole hierarchy when it
276 * searches where page is mapped.
278 int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
280 struct anon_vma_chain *avc, *pavc;
281 struct anon_vma *root = NULL;
283 list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
284 struct anon_vma *anon_vma;
286 avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
287 if (unlikely(!avc)) {
288 unlock_anon_vma_root(root);
290 avc = anon_vma_chain_alloc(GFP_KERNEL);
294 anon_vma = pavc->anon_vma;
295 root = lock_anon_vma_root(root, anon_vma);
296 anon_vma_chain_link(dst, avc, anon_vma);
299 * Reuse existing anon_vma if its degree lower than two,
300 * that means it has no vma and only one anon_vma child.
302 * Do not choose parent anon_vma, otherwise first child
303 * will always reuse it. Root anon_vma is never reused:
304 * it has self-parent reference and at least one child.
306 if (!dst->anon_vma && src->anon_vma &&
307 anon_vma != src->anon_vma && anon_vma->degree < 2)
308 dst->anon_vma = anon_vma;
311 dst->anon_vma->degree++;
312 unlock_anon_vma_root(root);
317 * dst->anon_vma is dropped here otherwise its degree can be incorrectly
318 * decremented in unlink_anon_vmas().
319 * We can safely do this because callers of anon_vma_clone() don't care
320 * about dst->anon_vma if anon_vma_clone() failed.
322 dst->anon_vma = NULL;
323 unlink_anon_vmas(dst);
328 * Attach vma to its own anon_vma, as well as to the anon_vmas that
329 * the corresponding VMA in the parent process is attached to.
330 * Returns 0 on success, non-zero on failure.
332 int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
334 struct anon_vma_chain *avc;
335 struct anon_vma *anon_vma;
338 /* Don't bother if the parent process has no anon_vma here. */
342 /* Drop inherited anon_vma, we'll reuse existing or allocate new. */
343 vma->anon_vma = NULL;
346 * First, attach the new VMA to the parent VMA's anon_vmas,
347 * so rmap can find non-COWed pages in child processes.
349 error = anon_vma_clone(vma, pvma);
353 /* An existing anon_vma has been reused, all done then. */
357 /* Then add our own anon_vma. */
358 anon_vma = anon_vma_alloc();
361 avc = anon_vma_chain_alloc(GFP_KERNEL);
363 goto out_error_free_anon_vma;
366 * The root anon_vma's rwsem is the lock actually used when we
367 * lock any of the anon_vmas in this anon_vma tree.
369 anon_vma->root = pvma->anon_vma->root;
370 anon_vma->parent = pvma->anon_vma;
372 * With refcounts, an anon_vma can stay around longer than the
373 * process it belongs to. The root anon_vma needs to be pinned until
374 * this anon_vma is freed, because the lock lives in the root.
376 get_anon_vma(anon_vma->root);
377 /* Mark this anon_vma as the one where our new (COWed) pages go. */
378 vma->anon_vma = anon_vma;
379 anon_vma_lock_write(anon_vma);
380 anon_vma_chain_link(vma, avc, anon_vma);
381 anon_vma->parent->degree++;
382 anon_vma_unlock_write(anon_vma);
386 out_error_free_anon_vma:
387 put_anon_vma(anon_vma);
389 unlink_anon_vmas(vma);
393 void unlink_anon_vmas(struct vm_area_struct *vma)
395 struct anon_vma_chain *avc, *next;
396 struct anon_vma *root = NULL;
399 * Unlink each anon_vma chained to the VMA. This list is ordered
400 * from newest to oldest, ensuring the root anon_vma gets freed last.
402 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
403 struct anon_vma *anon_vma = avc->anon_vma;
405 root = lock_anon_vma_root(root, anon_vma);
406 anon_vma_interval_tree_remove(avc, &anon_vma->rb_root);
409 * Leave empty anon_vmas on the list - we'll need
410 * to free them outside the lock.
412 if (RB_EMPTY_ROOT(&anon_vma->rb_root.rb_root)) {
413 anon_vma->parent->degree--;
417 list_del(&avc->same_vma);
418 anon_vma_chain_free(avc);
421 vma->anon_vma->degree--;
424 * vma would still be needed after unlink, and anon_vma will be prepared
427 vma->anon_vma = NULL;
429 unlock_anon_vma_root(root);
432 * Iterate the list once more, it now only contains empty and unlinked
433 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
434 * needing to write-acquire the anon_vma->root->rwsem.
436 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
437 struct anon_vma *anon_vma = avc->anon_vma;
439 VM_WARN_ON(anon_vma->degree);
440 put_anon_vma(anon_vma);
442 list_del(&avc->same_vma);
443 anon_vma_chain_free(avc);
447 static void anon_vma_ctor(void *data)
449 struct anon_vma *anon_vma = data;
451 init_rwsem(&anon_vma->rwsem);
452 atomic_set(&anon_vma->refcount, 0);
453 anon_vma->rb_root = RB_ROOT_CACHED;
456 void __init anon_vma_init(void)
458 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
459 0, SLAB_TYPESAFE_BY_RCU|SLAB_PANIC|SLAB_ACCOUNT,
461 anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain,
462 SLAB_PANIC|SLAB_ACCOUNT);
466 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
468 * Since there is no serialization what so ever against page_remove_rmap()
469 * the best this function can do is return a refcount increased anon_vma
470 * that might have been relevant to this page.
472 * The page might have been remapped to a different anon_vma or the anon_vma
473 * returned may already be freed (and even reused).
475 * In case it was remapped to a different anon_vma, the new anon_vma will be a
476 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
477 * ensure that any anon_vma obtained from the page will still be valid for as
478 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
480 * All users of this function must be very careful when walking the anon_vma
481 * chain and verify that the page in question is indeed mapped in it
482 * [ something equivalent to page_mapped_in_vma() ].
484 * Since anon_vma's slab is SLAB_TYPESAFE_BY_RCU and we know from
485 * page_remove_rmap() that the anon_vma pointer from page->mapping is valid
486 * if there is a mapcount, we can dereference the anon_vma after observing
489 struct anon_vma *page_get_anon_vma(struct page *page)
491 struct anon_vma *anon_vma = NULL;
492 unsigned long anon_mapping;
495 anon_mapping = (unsigned long)READ_ONCE(page->mapping);
496 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
498 if (!page_mapped(page))
501 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
502 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
508 * If this page is still mapped, then its anon_vma cannot have been
509 * freed. But if it has been unmapped, we have no security against the
510 * anon_vma structure being freed and reused (for another anon_vma:
511 * SLAB_TYPESAFE_BY_RCU guarantees that - so the atomic_inc_not_zero()
512 * above cannot corrupt).
514 if (!page_mapped(page)) {
516 put_anon_vma(anon_vma);
526 * Similar to page_get_anon_vma() except it locks the anon_vma.
528 * Its a little more complex as it tries to keep the fast path to a single
529 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
530 * reference like with page_get_anon_vma() and then block on the mutex.
532 struct anon_vma *folio_lock_anon_vma_read(struct folio *folio)
534 struct anon_vma *anon_vma = NULL;
535 struct anon_vma *root_anon_vma;
536 unsigned long anon_mapping;
539 anon_mapping = (unsigned long)READ_ONCE(folio->mapping);
540 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
542 if (!folio_mapped(folio))
545 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
546 root_anon_vma = READ_ONCE(anon_vma->root);
547 if (down_read_trylock(&root_anon_vma->rwsem)) {
549 * If the folio is still mapped, then this anon_vma is still
550 * its anon_vma, and holding the mutex ensures that it will
551 * not go away, see anon_vma_free().
553 if (!folio_mapped(folio)) {
554 up_read(&root_anon_vma->rwsem);
560 /* trylock failed, we got to sleep */
561 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
566 if (!folio_mapped(folio)) {
568 put_anon_vma(anon_vma);
572 /* we pinned the anon_vma, its safe to sleep */
574 anon_vma_lock_read(anon_vma);
576 if (atomic_dec_and_test(&anon_vma->refcount)) {
578 * Oops, we held the last refcount, release the lock
579 * and bail -- can't simply use put_anon_vma() because
580 * we'll deadlock on the anon_vma_lock_write() recursion.
582 anon_vma_unlock_read(anon_vma);
583 __put_anon_vma(anon_vma);
594 void page_unlock_anon_vma_read(struct anon_vma *anon_vma)
596 anon_vma_unlock_read(anon_vma);
599 #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
601 * Flush TLB entries for recently unmapped pages from remote CPUs. It is
602 * important if a PTE was dirty when it was unmapped that it's flushed
603 * before any IO is initiated on the page to prevent lost writes. Similarly,
604 * it must be flushed before freeing to prevent data leakage.
606 void try_to_unmap_flush(void)
608 struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
610 if (!tlb_ubc->flush_required)
613 arch_tlbbatch_flush(&tlb_ubc->arch);
614 tlb_ubc->flush_required = false;
615 tlb_ubc->writable = false;
618 /* Flush iff there are potentially writable TLB entries that can race with IO */
619 void try_to_unmap_flush_dirty(void)
621 struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
623 if (tlb_ubc->writable)
624 try_to_unmap_flush();
628 * Bits 0-14 of mm->tlb_flush_batched record pending generations.
629 * Bits 16-30 of mm->tlb_flush_batched bit record flushed generations.
631 #define TLB_FLUSH_BATCH_FLUSHED_SHIFT 16
632 #define TLB_FLUSH_BATCH_PENDING_MASK \
633 ((1 << (TLB_FLUSH_BATCH_FLUSHED_SHIFT - 1)) - 1)
634 #define TLB_FLUSH_BATCH_PENDING_LARGE \
635 (TLB_FLUSH_BATCH_PENDING_MASK / 2)
637 static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
639 struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
642 arch_tlbbatch_add_mm(&tlb_ubc->arch, mm);
643 tlb_ubc->flush_required = true;
646 * Ensure compiler does not re-order the setting of tlb_flush_batched
647 * before the PTE is cleared.
650 batch = atomic_read(&mm->tlb_flush_batched);
652 if ((batch & TLB_FLUSH_BATCH_PENDING_MASK) > TLB_FLUSH_BATCH_PENDING_LARGE) {
654 * Prevent `pending' from catching up with `flushed' because of
655 * overflow. Reset `pending' and `flushed' to be 1 and 0 if
656 * `pending' becomes large.
658 nbatch = atomic_cmpxchg(&mm->tlb_flush_batched, batch, 1);
659 if (nbatch != batch) {
664 atomic_inc(&mm->tlb_flush_batched);
668 * If the PTE was dirty then it's best to assume it's writable. The
669 * caller must use try_to_unmap_flush_dirty() or try_to_unmap_flush()
670 * before the page is queued for IO.
673 tlb_ubc->writable = true;
677 * Returns true if the TLB flush should be deferred to the end of a batch of
678 * unmap operations to reduce IPIs.
680 static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
682 bool should_defer = false;
684 if (!(flags & TTU_BATCH_FLUSH))
687 /* If remote CPUs need to be flushed then defer batch the flush */
688 if (cpumask_any_but(mm_cpumask(mm), get_cpu()) < nr_cpu_ids)
696 * Reclaim unmaps pages under the PTL but do not flush the TLB prior to
697 * releasing the PTL if TLB flushes are batched. It's possible for a parallel
698 * operation such as mprotect or munmap to race between reclaim unmapping
699 * the page and flushing the page. If this race occurs, it potentially allows
700 * access to data via a stale TLB entry. Tracking all mm's that have TLB
701 * batching in flight would be expensive during reclaim so instead track
702 * whether TLB batching occurred in the past and if so then do a flush here
703 * if required. This will cost one additional flush per reclaim cycle paid
704 * by the first operation at risk such as mprotect and mumap.
706 * This must be called under the PTL so that an access to tlb_flush_batched
707 * that is potentially a "reclaim vs mprotect/munmap/etc" race will synchronise
710 void flush_tlb_batched_pending(struct mm_struct *mm)
712 int batch = atomic_read(&mm->tlb_flush_batched);
713 int pending = batch & TLB_FLUSH_BATCH_PENDING_MASK;
714 int flushed = batch >> TLB_FLUSH_BATCH_FLUSHED_SHIFT;
716 if (pending != flushed) {
719 * If the new TLB flushing is pending during flushing, leave
720 * mm->tlb_flush_batched as is, to avoid losing flushing.
722 atomic_cmpxchg(&mm->tlb_flush_batched, batch,
723 pending | (pending << TLB_FLUSH_BATCH_FLUSHED_SHIFT));
727 static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
731 static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
735 #endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
738 * At what user virtual address is page expected in vma?
739 * Caller should check the page is actually part of the vma.
741 unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
743 struct folio *folio = page_folio(page);
744 if (folio_test_anon(folio)) {
745 struct anon_vma *page__anon_vma = folio_anon_vma(folio);
747 * Note: swapoff's unuse_vma() is more efficient with this
748 * check, and needs it to match anon_vma when KSM is active.
750 if (!vma->anon_vma || !page__anon_vma ||
751 vma->anon_vma->root != page__anon_vma->root)
753 } else if (!vma->vm_file) {
755 } else if (vma->vm_file->f_mapping != folio->mapping) {
759 return vma_address(page, vma);
762 pmd_t *mm_find_pmd(struct mm_struct *mm, unsigned long address)
770 pgd = pgd_offset(mm, address);
771 if (!pgd_present(*pgd))
774 p4d = p4d_offset(pgd, address);
775 if (!p4d_present(*p4d))
778 pud = pud_offset(p4d, address);
779 if (!pud_present(*pud))
782 pmd = pmd_offset(pud, address);
784 * Some THP functions use the sequence pmdp_huge_clear_flush(), set_pmd_at()
785 * without holding anon_vma lock for write. So when looking for a
786 * genuine pmde (in which to find pte), test present and !THP together.
790 if (!pmd_present(pmde) || pmd_trans_huge(pmde))
796 struct folio_referenced_arg {
799 unsigned long vm_flags;
800 struct mem_cgroup *memcg;
803 * arg: folio_referenced_arg will be passed
805 static bool folio_referenced_one(struct folio *folio,
806 struct vm_area_struct *vma, unsigned long address, void *arg)
808 struct folio_referenced_arg *pra = arg;
809 DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0);
812 while (page_vma_mapped_walk(&pvmw)) {
813 address = pvmw.address;
815 if ((vma->vm_flags & VM_LOCKED) &&
816 (!folio_test_large(folio) || !pvmw.pte)) {
817 /* Restore the mlock which got missed */
818 mlock_vma_folio(folio, vma, !pvmw.pte);
819 page_vma_mapped_walk_done(&pvmw);
820 pra->vm_flags |= VM_LOCKED;
821 return false; /* To break the loop */
825 if (ptep_clear_flush_young_notify(vma, address,
828 * Don't treat a reference through
829 * a sequentially read mapping as such.
830 * If the folio has been used in another mapping,
831 * we will catch it; if this other mapping is
832 * already gone, the unmap path will have set
833 * the referenced flag or activated the folio.
835 if (likely(!(vma->vm_flags & VM_SEQ_READ)))
838 } else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
839 if (pmdp_clear_flush_young_notify(vma, address,
843 /* unexpected pmd-mapped folio? */
851 folio_clear_idle(folio);
852 if (folio_test_clear_young(folio))
857 pra->vm_flags |= vma->vm_flags & ~VM_LOCKED;
861 return false; /* To break the loop */
866 static bool invalid_folio_referenced_vma(struct vm_area_struct *vma, void *arg)
868 struct folio_referenced_arg *pra = arg;
869 struct mem_cgroup *memcg = pra->memcg;
871 if (!mm_match_cgroup(vma->vm_mm, memcg))
878 * folio_referenced() - Test if the folio was referenced.
879 * @folio: The folio to test.
880 * @is_locked: Caller holds lock on the folio.
881 * @memcg: target memory cgroup
882 * @vm_flags: A combination of all the vma->vm_flags which referenced the folio.
884 * Quick test_and_clear_referenced for all mappings of a folio,
886 * Return: The number of mappings which referenced the folio.
888 int folio_referenced(struct folio *folio, int is_locked,
889 struct mem_cgroup *memcg, unsigned long *vm_flags)
892 struct folio_referenced_arg pra = {
893 .mapcount = folio_mapcount(folio),
896 struct rmap_walk_control rwc = {
897 .rmap_one = folio_referenced_one,
899 .anon_lock = folio_lock_anon_vma_read,
906 if (!folio_raw_mapping(folio))
909 if (!is_locked && (!folio_test_anon(folio) || folio_test_ksm(folio))) {
910 we_locked = folio_trylock(folio);
916 * If we are reclaiming on behalf of a cgroup, skip
917 * counting on behalf of references from different
921 rwc.invalid_vma = invalid_folio_referenced_vma;
924 rmap_walk(folio, &rwc);
925 *vm_flags = pra.vm_flags;
930 return pra.referenced;
933 static int page_vma_mkclean_one(struct page_vma_mapped_walk *pvmw)
936 struct vm_area_struct *vma = pvmw->vma;
937 struct mmu_notifier_range range;
938 unsigned long address = pvmw->address;
941 * We have to assume the worse case ie pmd for invalidation. Note that
942 * the folio can not be freed from this function.
944 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
945 0, vma, vma->vm_mm, address,
946 vma_address_end(pvmw));
947 mmu_notifier_invalidate_range_start(&range);
949 while (page_vma_mapped_walk(pvmw)) {
952 address = pvmw->address;
955 pte_t *pte = pvmw->pte;
957 if (!pte_dirty(*pte) && !pte_write(*pte))
960 flush_cache_page(vma, address, pte_pfn(*pte));
961 entry = ptep_clear_flush(vma, address, pte);
962 entry = pte_wrprotect(entry);
963 entry = pte_mkclean(entry);
964 set_pte_at(vma->vm_mm, address, pte, entry);
967 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
968 pmd_t *pmd = pvmw->pmd;
971 if (!pmd_dirty(*pmd) && !pmd_write(*pmd))
974 flush_cache_range(vma, address,
975 address + HPAGE_PMD_SIZE);
976 entry = pmdp_invalidate(vma, address, pmd);
977 entry = pmd_wrprotect(entry);
978 entry = pmd_mkclean(entry);
979 set_pmd_at(vma->vm_mm, address, pmd, entry);
982 /* unexpected pmd-mapped folio? */
988 * No need to call mmu_notifier_invalidate_range() as we are
989 * downgrading page table protection not changing it to point
992 * See Documentation/vm/mmu_notifier.rst
998 mmu_notifier_invalidate_range_end(&range);
1003 static bool page_mkclean_one(struct folio *folio, struct vm_area_struct *vma,
1004 unsigned long address, void *arg)
1006 DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, PVMW_SYNC);
1009 *cleaned += page_vma_mkclean_one(&pvmw);
1014 static bool invalid_mkclean_vma(struct vm_area_struct *vma, void *arg)
1016 if (vma->vm_flags & VM_SHARED)
1022 int folio_mkclean(struct folio *folio)
1025 struct address_space *mapping;
1026 struct rmap_walk_control rwc = {
1027 .arg = (void *)&cleaned,
1028 .rmap_one = page_mkclean_one,
1029 .invalid_vma = invalid_mkclean_vma,
1032 BUG_ON(!folio_test_locked(folio));
1034 if (!folio_mapped(folio))
1037 mapping = folio_mapping(folio);
1041 rmap_walk(folio, &rwc);
1045 EXPORT_SYMBOL_GPL(folio_mkclean);
1048 * pfn_mkclean_range - Cleans the PTEs (including PMDs) mapped with range of
1049 * [@pfn, @pfn + @nr_pages) at the specific offset (@pgoff)
1050 * within the @vma of shared mappings. And since clean PTEs
1051 * should also be readonly, write protects them too.
1053 * @nr_pages: number of physically contiguous pages srarting with @pfn.
1054 * @pgoff: page offset that the @pfn mapped with.
1055 * @vma: vma that @pfn mapped within.
1057 * Returns the number of cleaned PTEs (including PMDs).
1059 int pfn_mkclean_range(unsigned long pfn, unsigned long nr_pages, pgoff_t pgoff,
1060 struct vm_area_struct *vma)
1062 struct page_vma_mapped_walk pvmw = {
1064 .nr_pages = nr_pages,
1070 if (invalid_mkclean_vma(vma, NULL))
1073 pvmw.address = vma_pgoff_address(pgoff, nr_pages, vma);
1074 VM_BUG_ON_VMA(pvmw.address == -EFAULT, vma);
1076 return page_vma_mkclean_one(&pvmw);
1080 * page_move_anon_rmap - move a page to our anon_vma
1081 * @page: the page to move to our anon_vma
1082 * @vma: the vma the page belongs to
1084 * When a page belongs exclusively to one process after a COW event,
1085 * that page can be moved into the anon_vma that belongs to just that
1086 * process, so the rmap code will not search the parent or sibling
1089 void page_move_anon_rmap(struct page *page, struct vm_area_struct *vma)
1091 struct anon_vma *anon_vma = vma->anon_vma;
1092 struct page *subpage = page;
1094 page = compound_head(page);
1096 VM_BUG_ON_PAGE(!PageLocked(page), page);
1097 VM_BUG_ON_VMA(!anon_vma, vma);
1099 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1101 * Ensure that anon_vma and the PAGE_MAPPING_ANON bit are written
1102 * simultaneously, so a concurrent reader (eg folio_referenced()'s
1103 * folio_test_anon()) will not see one without the other.
1105 WRITE_ONCE(page->mapping, (struct address_space *) anon_vma);
1106 SetPageAnonExclusive(subpage);
1110 * __page_set_anon_rmap - set up new anonymous rmap
1111 * @page: Page or Hugepage to add to rmap
1112 * @vma: VM area to add page to.
1113 * @address: User virtual address of the mapping
1114 * @exclusive: the page is exclusively owned by the current process
1116 static void __page_set_anon_rmap(struct page *page,
1117 struct vm_area_struct *vma, unsigned long address, int exclusive)
1119 struct anon_vma *anon_vma = vma->anon_vma;
1127 * If the page isn't exclusively mapped into this vma,
1128 * we must use the _oldest_ possible anon_vma for the
1132 anon_vma = anon_vma->root;
1135 * page_idle does a lockless/optimistic rmap scan on page->mapping.
1136 * Make sure the compiler doesn't split the stores of anon_vma and
1137 * the PAGE_MAPPING_ANON type identifier, otherwise the rmap code
1138 * could mistake the mapping for a struct address_space and crash.
1140 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1141 WRITE_ONCE(page->mapping, (struct address_space *) anon_vma);
1142 page->index = linear_page_index(vma, address);
1145 SetPageAnonExclusive(page);
1149 * __page_check_anon_rmap - sanity check anonymous rmap addition
1150 * @page: the page to add the mapping to
1151 * @vma: the vm area in which the mapping is added
1152 * @address: the user virtual address mapped
1154 static void __page_check_anon_rmap(struct page *page,
1155 struct vm_area_struct *vma, unsigned long address)
1157 struct folio *folio = page_folio(page);
1159 * The page's anon-rmap details (mapping and index) are guaranteed to
1160 * be set up correctly at this point.
1162 * We have exclusion against page_add_anon_rmap because the caller
1163 * always holds the page locked.
1165 * We have exclusion against page_add_new_anon_rmap because those pages
1166 * are initially only visible via the pagetables, and the pte is locked
1167 * over the call to page_add_new_anon_rmap.
1169 VM_BUG_ON_FOLIO(folio_anon_vma(folio)->root != vma->anon_vma->root,
1171 VM_BUG_ON_PAGE(page_to_pgoff(page) != linear_page_index(vma, address),
1176 * page_add_anon_rmap - add pte mapping to an anonymous page
1177 * @page: the page to add the mapping to
1178 * @vma: the vm area in which the mapping is added
1179 * @address: the user virtual address mapped
1180 * @flags: the rmap flags
1182 * The caller needs to hold the pte lock, and the page must be locked in
1183 * the anon_vma case: to serialize mapping,index checking after setting,
1184 * and to ensure that PageAnon is not being upgraded racily to PageKsm
1185 * (but PageKsm is never downgraded to PageAnon).
1187 void page_add_anon_rmap(struct page *page,
1188 struct vm_area_struct *vma, unsigned long address, rmap_t flags)
1190 bool compound = flags & RMAP_COMPOUND;
1193 if (unlikely(PageKsm(page)))
1194 lock_page_memcg(page);
1196 VM_BUG_ON_PAGE(!PageLocked(page), page);
1200 VM_BUG_ON_PAGE(!PageLocked(page), page);
1201 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
1202 mapcount = compound_mapcount_ptr(page);
1203 first = atomic_inc_and_test(mapcount);
1205 first = atomic_inc_and_test(&page->_mapcount);
1207 VM_BUG_ON_PAGE(!first && (flags & RMAP_EXCLUSIVE), page);
1208 VM_BUG_ON_PAGE(!first && PageAnonExclusive(page), page);
1211 int nr = compound ? thp_nr_pages(page) : 1;
1213 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1214 * these counters are not modified in interrupt context, and
1215 * pte lock(a spinlock) is held, which implies preemption
1219 __mod_lruvec_page_state(page, NR_ANON_THPS, nr);
1220 __mod_lruvec_page_state(page, NR_ANON_MAPPED, nr);
1223 if (unlikely(PageKsm(page)))
1224 unlock_page_memcg(page);
1226 /* address might be in next vma when migration races vma_adjust */
1228 __page_set_anon_rmap(page, vma, address,
1229 !!(flags & RMAP_EXCLUSIVE));
1231 __page_check_anon_rmap(page, vma, address);
1233 mlock_vma_page(page, vma, compound);
1237 * page_add_new_anon_rmap - add mapping to a new anonymous page
1238 * @page: the page to add the mapping to
1239 * @vma: the vm area in which the mapping is added
1240 * @address: the user virtual address mapped
1242 * If it's a compound page, it is accounted as a compound page. As the page
1243 * is new, it's assume to get mapped exclusively by a single process.
1245 * Same as page_add_anon_rmap but must only be called on *new* pages.
1246 * This means the inc-and-test can be bypassed.
1247 * Page does not have to be locked.
1249 void page_add_new_anon_rmap(struct page *page,
1250 struct vm_area_struct *vma, unsigned long address)
1252 const bool compound = PageCompound(page);
1253 int nr = compound ? thp_nr_pages(page) : 1;
1255 VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
1256 __SetPageSwapBacked(page);
1258 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
1259 /* increment count (starts at -1) */
1260 atomic_set(compound_mapcount_ptr(page), 0);
1261 atomic_set(compound_pincount_ptr(page), 0);
1263 __mod_lruvec_page_state(page, NR_ANON_THPS, nr);
1265 /* increment count (starts at -1) */
1266 atomic_set(&page->_mapcount, 0);
1268 __mod_lruvec_page_state(page, NR_ANON_MAPPED, nr);
1269 __page_set_anon_rmap(page, vma, address, 1);
1273 * page_add_file_rmap - add pte mapping to a file page
1274 * @page: the page to add the mapping to
1275 * @vma: the vm area in which the mapping is added
1276 * @compound: charge the page as compound or small page
1278 * The caller needs to hold the pte lock.
1280 void page_add_file_rmap(struct page *page,
1281 struct vm_area_struct *vma, bool compound)
1285 VM_BUG_ON_PAGE(compound && !PageTransHuge(page), page);
1286 lock_page_memcg(page);
1287 if (compound && PageTransHuge(page)) {
1288 int nr_pages = thp_nr_pages(page);
1290 for (i = 0; i < nr_pages; i++) {
1291 if (atomic_inc_and_test(&page[i]._mapcount))
1294 if (!atomic_inc_and_test(compound_mapcount_ptr(page)))
1298 * It is racy to ClearPageDoubleMap in page_remove_file_rmap();
1299 * but page lock is held by all page_add_file_rmap() compound
1300 * callers, and SetPageDoubleMap below warns if !PageLocked:
1301 * so here is a place that DoubleMap can be safely cleared.
1303 VM_WARN_ON_ONCE(!PageLocked(page));
1304 if (nr == nr_pages && PageDoubleMap(page))
1305 ClearPageDoubleMap(page);
1307 if (PageSwapBacked(page))
1308 __mod_lruvec_page_state(page, NR_SHMEM_PMDMAPPED,
1311 __mod_lruvec_page_state(page, NR_FILE_PMDMAPPED,
1314 if (PageTransCompound(page) && page_mapping(page)) {
1315 VM_WARN_ON_ONCE(!PageLocked(page));
1316 SetPageDoubleMap(compound_head(page));
1318 if (atomic_inc_and_test(&page->_mapcount))
1323 __mod_lruvec_page_state(page, NR_FILE_MAPPED, nr);
1324 unlock_page_memcg(page);
1326 mlock_vma_page(page, vma, compound);
1329 static void page_remove_file_rmap(struct page *page, bool compound)
1333 VM_BUG_ON_PAGE(compound && !PageHead(page), page);
1335 /* Hugepages are not counted in NR_FILE_MAPPED for now. */
1336 if (unlikely(PageHuge(page))) {
1337 /* hugetlb pages are always mapped with pmds */
1338 atomic_dec(compound_mapcount_ptr(page));
1342 /* page still mapped by someone else? */
1343 if (compound && PageTransHuge(page)) {
1344 int nr_pages = thp_nr_pages(page);
1346 for (i = 0; i < nr_pages; i++) {
1347 if (atomic_add_negative(-1, &page[i]._mapcount))
1350 if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
1352 if (PageSwapBacked(page))
1353 __mod_lruvec_page_state(page, NR_SHMEM_PMDMAPPED,
1356 __mod_lruvec_page_state(page, NR_FILE_PMDMAPPED,
1359 if (atomic_add_negative(-1, &page->_mapcount))
1364 __mod_lruvec_page_state(page, NR_FILE_MAPPED, -nr);
1367 static void page_remove_anon_compound_rmap(struct page *page)
1371 if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
1374 /* Hugepages are not counted in NR_ANON_PAGES for now. */
1375 if (unlikely(PageHuge(page)))
1378 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE))
1381 __mod_lruvec_page_state(page, NR_ANON_THPS, -thp_nr_pages(page));
1383 if (TestClearPageDoubleMap(page)) {
1385 * Subpages can be mapped with PTEs too. Check how many of
1386 * them are still mapped.
1388 for (i = 0, nr = 0; i < thp_nr_pages(page); i++) {
1389 if (atomic_add_negative(-1, &page[i]._mapcount))
1394 * Queue the page for deferred split if at least one small
1395 * page of the compound page is unmapped, but at least one
1396 * small page is still mapped.
1398 if (nr && nr < thp_nr_pages(page))
1399 deferred_split_huge_page(page);
1401 nr = thp_nr_pages(page);
1405 __mod_lruvec_page_state(page, NR_ANON_MAPPED, -nr);
1409 * page_remove_rmap - take down pte mapping from a page
1410 * @page: page to remove mapping from
1411 * @vma: the vm area from which the mapping is removed
1412 * @compound: uncharge the page as compound or small page
1414 * The caller needs to hold the pte lock.
1416 void page_remove_rmap(struct page *page,
1417 struct vm_area_struct *vma, bool compound)
1419 lock_page_memcg(page);
1421 if (!PageAnon(page)) {
1422 page_remove_file_rmap(page, compound);
1427 page_remove_anon_compound_rmap(page);
1431 /* page still mapped by someone else? */
1432 if (!atomic_add_negative(-1, &page->_mapcount))
1436 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1437 * these counters are not modified in interrupt context, and
1438 * pte lock(a spinlock) is held, which implies preemption disabled.
1440 __dec_lruvec_page_state(page, NR_ANON_MAPPED);
1442 if (PageTransCompound(page))
1443 deferred_split_huge_page(compound_head(page));
1446 * It would be tidy to reset the PageAnon mapping here,
1447 * but that might overwrite a racing page_add_anon_rmap
1448 * which increments mapcount after us but sets mapping
1449 * before us: so leave the reset to free_unref_page,
1450 * and remember that it's only reliable while mapped.
1451 * Leaving it set also helps swapoff to reinstate ptes
1452 * faster for those pages still in swapcache.
1455 unlock_page_memcg(page);
1457 munlock_vma_page(page, vma, compound);
1461 * @arg: enum ttu_flags will be passed to this argument
1463 static bool try_to_unmap_one(struct folio *folio, struct vm_area_struct *vma,
1464 unsigned long address, void *arg)
1466 struct mm_struct *mm = vma->vm_mm;
1467 DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0);
1469 struct page *subpage;
1470 bool anon_exclusive, ret = true;
1471 struct mmu_notifier_range range;
1472 enum ttu_flags flags = (enum ttu_flags)(long)arg;
1475 * When racing against e.g. zap_pte_range() on another cpu,
1476 * in between its ptep_get_and_clear_full() and page_remove_rmap(),
1477 * try_to_unmap() may return before page_mapped() has become false,
1478 * if page table locking is skipped: use TTU_SYNC to wait for that.
1480 if (flags & TTU_SYNC)
1481 pvmw.flags = PVMW_SYNC;
1483 if (flags & TTU_SPLIT_HUGE_PMD)
1484 split_huge_pmd_address(vma, address, false, folio);
1487 * For THP, we have to assume the worse case ie pmd for invalidation.
1488 * For hugetlb, it could be much worse if we need to do pud
1489 * invalidation in the case of pmd sharing.
1491 * Note that the folio can not be freed in this function as call of
1492 * try_to_unmap() must hold a reference on the folio.
1494 range.end = vma_address_end(&pvmw);
1495 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1496 address, range.end);
1497 if (folio_test_hugetlb(folio)) {
1499 * If sharing is possible, start and end will be adjusted
1502 adjust_range_if_pmd_sharing_possible(vma, &range.start,
1505 mmu_notifier_invalidate_range_start(&range);
1507 while (page_vma_mapped_walk(&pvmw)) {
1508 /* Unexpected PMD-mapped THP? */
1509 VM_BUG_ON_FOLIO(!pvmw.pte, folio);
1512 * If the folio is in an mlock()d vma, we must not swap it out.
1514 if (!(flags & TTU_IGNORE_MLOCK) &&
1515 (vma->vm_flags & VM_LOCKED)) {
1516 /* Restore the mlock which got missed */
1517 mlock_vma_folio(folio, vma, false);
1518 page_vma_mapped_walk_done(&pvmw);
1523 subpage = folio_page(folio,
1524 pte_pfn(*pvmw.pte) - folio_pfn(folio));
1525 address = pvmw.address;
1526 anon_exclusive = folio_test_anon(folio) &&
1527 PageAnonExclusive(subpage);
1529 if (folio_test_hugetlb(folio)) {
1531 * huge_pmd_unshare may unmap an entire PMD page.
1532 * There is no way of knowing exactly which PMDs may
1533 * be cached for this mm, so we must flush them all.
1534 * start/end were already adjusted above to cover this
1537 flush_cache_range(vma, range.start, range.end);
1539 if (!folio_test_anon(folio)) {
1541 * To call huge_pmd_unshare, i_mmap_rwsem must be
1542 * held in write mode. Caller needs to explicitly
1543 * do this outside rmap routines.
1545 VM_BUG_ON(!(flags & TTU_RMAP_LOCKED));
1547 if (huge_pmd_unshare(mm, vma, &address, pvmw.pte)) {
1548 flush_tlb_range(vma, range.start, range.end);
1549 mmu_notifier_invalidate_range(mm, range.start,
1553 * The ref count of the PMD page was dropped
1554 * which is part of the way map counting
1555 * is done for shared PMDs. Return 'true'
1556 * here. When there is no other sharing,
1557 * huge_pmd_unshare returns false and we will
1558 * unmap the actual page and drop map count
1561 page_vma_mapped_walk_done(&pvmw);
1566 flush_cache_page(vma, address, pte_pfn(*pvmw.pte));
1570 * Nuke the page table entry. When having to clear
1571 * PageAnonExclusive(), we always have to flush.
1573 if (should_defer_flush(mm, flags) && !anon_exclusive) {
1575 * We clear the PTE but do not flush so potentially
1576 * a remote CPU could still be writing to the folio.
1577 * If the entry was previously clean then the
1578 * architecture must guarantee that a clear->dirty
1579 * transition on a cached TLB entry is written through
1580 * and traps if the PTE is unmapped.
1582 pteval = ptep_get_and_clear(mm, address, pvmw.pte);
1584 set_tlb_ubc_flush_pending(mm, pte_dirty(pteval));
1586 pteval = ptep_clear_flush(vma, address, pvmw.pte);
1590 * Now the pte is cleared. If this pte was uffd-wp armed,
1591 * we may want to replace a none pte with a marker pte if
1592 * it's file-backed, so we don't lose the tracking info.
1594 pte_install_uffd_wp_if_needed(vma, address, pvmw.pte, pteval);
1596 /* Set the dirty flag on the folio now the pte is gone. */
1597 if (pte_dirty(pteval))
1598 folio_mark_dirty(folio);
1600 /* Update high watermark before we lower rss */
1601 update_hiwater_rss(mm);
1603 if (PageHWPoison(subpage) && !(flags & TTU_IGNORE_HWPOISON)) {
1604 pteval = swp_entry_to_pte(make_hwpoison_entry(subpage));
1605 if (folio_test_hugetlb(folio)) {
1606 hugetlb_count_sub(folio_nr_pages(folio), mm);
1607 set_huge_swap_pte_at(mm, address,
1609 vma_mmu_pagesize(vma));
1611 dec_mm_counter(mm, mm_counter(&folio->page));
1612 set_pte_at(mm, address, pvmw.pte, pteval);
1615 } else if (pte_unused(pteval) && !userfaultfd_armed(vma)) {
1617 * The guest indicated that the page content is of no
1618 * interest anymore. Simply discard the pte, vmscan
1619 * will take care of the rest.
1620 * A future reference will then fault in a new zero
1621 * page. When userfaultfd is active, we must not drop
1622 * this page though, as its main user (postcopy
1623 * migration) will not expect userfaults on already
1626 dec_mm_counter(mm, mm_counter(&folio->page));
1627 /* We have to invalidate as we cleared the pte */
1628 mmu_notifier_invalidate_range(mm, address,
1629 address + PAGE_SIZE);
1630 } else if (folio_test_anon(folio)) {
1631 swp_entry_t entry = { .val = page_private(subpage) };
1634 * Store the swap location in the pte.
1635 * See handle_pte_fault() ...
1637 if (unlikely(folio_test_swapbacked(folio) !=
1638 folio_test_swapcache(folio))) {
1641 /* We have to invalidate as we cleared the pte */
1642 mmu_notifier_invalidate_range(mm, address,
1643 address + PAGE_SIZE);
1644 page_vma_mapped_walk_done(&pvmw);
1648 /* MADV_FREE page check */
1649 if (!folio_test_swapbacked(folio)) {
1650 int ref_count, map_count;
1653 * Synchronize with gup_pte_range():
1654 * - clear PTE; barrier; read refcount
1655 * - inc refcount; barrier; read PTE
1659 ref_count = folio_ref_count(folio);
1660 map_count = folio_mapcount(folio);
1663 * Order reads for page refcount and dirty flag
1664 * (see comments in __remove_mapping()).
1669 * The only page refs must be one from isolation
1670 * plus the rmap(s) (dropped by discard:).
1672 if (ref_count == 1 + map_count &&
1673 !folio_test_dirty(folio)) {
1674 /* Invalidate as we cleared the pte */
1675 mmu_notifier_invalidate_range(mm,
1676 address, address + PAGE_SIZE);
1677 dec_mm_counter(mm, MM_ANONPAGES);
1682 * If the folio was redirtied, it cannot be
1683 * discarded. Remap the page to page table.
1685 set_pte_at(mm, address, pvmw.pte, pteval);
1686 folio_set_swapbacked(folio);
1688 page_vma_mapped_walk_done(&pvmw);
1692 if (swap_duplicate(entry) < 0) {
1693 set_pte_at(mm, address, pvmw.pte, pteval);
1695 page_vma_mapped_walk_done(&pvmw);
1698 if (arch_unmap_one(mm, vma, address, pteval) < 0) {
1700 set_pte_at(mm, address, pvmw.pte, pteval);
1702 page_vma_mapped_walk_done(&pvmw);
1705 if (anon_exclusive &&
1706 page_try_share_anon_rmap(subpage)) {
1708 set_pte_at(mm, address, pvmw.pte, pteval);
1710 page_vma_mapped_walk_done(&pvmw);
1714 * Note: We *don't* remember if the page was mapped
1715 * exclusively in the swap pte if the architecture
1716 * doesn't support __HAVE_ARCH_PTE_SWP_EXCLUSIVE. In
1717 * that case, swapin code has to re-determine that
1718 * manually and might detect the page as possibly
1719 * shared, for example, if there are other references on
1720 * the page or if the page is under writeback. We made
1721 * sure that there are no GUP pins on the page that
1722 * would rely on it, so for GUP pins this is fine.
1724 if (list_empty(&mm->mmlist)) {
1725 spin_lock(&mmlist_lock);
1726 if (list_empty(&mm->mmlist))
1727 list_add(&mm->mmlist, &init_mm.mmlist);
1728 spin_unlock(&mmlist_lock);
1730 dec_mm_counter(mm, MM_ANONPAGES);
1731 inc_mm_counter(mm, MM_SWAPENTS);
1732 swp_pte = swp_entry_to_pte(entry);
1734 swp_pte = pte_swp_mkexclusive(swp_pte);
1735 if (pte_soft_dirty(pteval))
1736 swp_pte = pte_swp_mksoft_dirty(swp_pte);
1737 if (pte_uffd_wp(pteval))
1738 swp_pte = pte_swp_mkuffd_wp(swp_pte);
1739 set_pte_at(mm, address, pvmw.pte, swp_pte);
1740 /* Invalidate as we cleared the pte */
1741 mmu_notifier_invalidate_range(mm, address,
1742 address + PAGE_SIZE);
1745 * This is a locked file-backed folio,
1746 * so it cannot be removed from the page
1747 * cache and replaced by a new folio before
1748 * mmu_notifier_invalidate_range_end, so no
1749 * concurrent thread might update its page table
1750 * to point at a new folio while a device is
1751 * still using this folio.
1753 * See Documentation/vm/mmu_notifier.rst
1755 dec_mm_counter(mm, mm_counter_file(&folio->page));
1759 * No need to call mmu_notifier_invalidate_range() it has be
1760 * done above for all cases requiring it to happen under page
1761 * table lock before mmu_notifier_invalidate_range_end()
1763 * See Documentation/vm/mmu_notifier.rst
1765 page_remove_rmap(subpage, vma, folio_test_hugetlb(folio));
1766 if (vma->vm_flags & VM_LOCKED)
1767 mlock_page_drain_local();
1771 mmu_notifier_invalidate_range_end(&range);
1776 static bool invalid_migration_vma(struct vm_area_struct *vma, void *arg)
1778 return vma_is_temporary_stack(vma);
1781 static int page_not_mapped(struct folio *folio)
1783 return !folio_mapped(folio);
1787 * try_to_unmap - Try to remove all page table mappings to a folio.
1788 * @folio: The folio to unmap.
1789 * @flags: action and flags
1791 * Tries to remove all the page table entries which are mapping this
1792 * folio. It is the caller's responsibility to check if the folio is
1793 * still mapped if needed (use TTU_SYNC to prevent accounting races).
1795 * Context: Caller must hold the folio lock.
1797 void try_to_unmap(struct folio *folio, enum ttu_flags flags)
1799 struct rmap_walk_control rwc = {
1800 .rmap_one = try_to_unmap_one,
1801 .arg = (void *)flags,
1802 .done = page_not_mapped,
1803 .anon_lock = folio_lock_anon_vma_read,
1806 if (flags & TTU_RMAP_LOCKED)
1807 rmap_walk_locked(folio, &rwc);
1809 rmap_walk(folio, &rwc);
1813 * @arg: enum ttu_flags will be passed to this argument.
1815 * If TTU_SPLIT_HUGE_PMD is specified any PMD mappings will be split into PTEs
1816 * containing migration entries.
1818 static bool try_to_migrate_one(struct folio *folio, struct vm_area_struct *vma,
1819 unsigned long address, void *arg)
1821 struct mm_struct *mm = vma->vm_mm;
1822 DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0);
1824 struct page *subpage;
1825 bool anon_exclusive, ret = true;
1826 struct mmu_notifier_range range;
1827 enum ttu_flags flags = (enum ttu_flags)(long)arg;
1830 * When racing against e.g. zap_pte_range() on another cpu,
1831 * in between its ptep_get_and_clear_full() and page_remove_rmap(),
1832 * try_to_migrate() may return before page_mapped() has become false,
1833 * if page table locking is skipped: use TTU_SYNC to wait for that.
1835 if (flags & TTU_SYNC)
1836 pvmw.flags = PVMW_SYNC;
1839 * unmap_page() in mm/huge_memory.c is the only user of migration with
1840 * TTU_SPLIT_HUGE_PMD and it wants to freeze.
1842 if (flags & TTU_SPLIT_HUGE_PMD)
1843 split_huge_pmd_address(vma, address, true, folio);
1846 * For THP, we have to assume the worse case ie pmd for invalidation.
1847 * For hugetlb, it could be much worse if we need to do pud
1848 * invalidation in the case of pmd sharing.
1850 * Note that the page can not be free in this function as call of
1851 * try_to_unmap() must hold a reference on the page.
1853 range.end = vma_address_end(&pvmw);
1854 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1855 address, range.end);
1856 if (folio_test_hugetlb(folio)) {
1858 * If sharing is possible, start and end will be adjusted
1861 adjust_range_if_pmd_sharing_possible(vma, &range.start,
1864 mmu_notifier_invalidate_range_start(&range);
1866 while (page_vma_mapped_walk(&pvmw)) {
1867 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1868 /* PMD-mapped THP migration entry */
1870 subpage = folio_page(folio,
1871 pmd_pfn(*pvmw.pmd) - folio_pfn(folio));
1872 VM_BUG_ON_FOLIO(folio_test_hugetlb(folio) ||
1873 !folio_test_pmd_mappable(folio), folio);
1875 if (set_pmd_migration_entry(&pvmw, subpage)) {
1877 page_vma_mapped_walk_done(&pvmw);
1884 /* Unexpected PMD-mapped THP? */
1885 VM_BUG_ON_FOLIO(!pvmw.pte, folio);
1887 subpage = folio_page(folio,
1888 pte_pfn(*pvmw.pte) - folio_pfn(folio));
1889 address = pvmw.address;
1890 anon_exclusive = folio_test_anon(folio) &&
1891 PageAnonExclusive(subpage);
1893 if (folio_test_hugetlb(folio)) {
1895 * huge_pmd_unshare may unmap an entire PMD page.
1896 * There is no way of knowing exactly which PMDs may
1897 * be cached for this mm, so we must flush them all.
1898 * start/end were already adjusted above to cover this
1901 flush_cache_range(vma, range.start, range.end);
1903 if (!folio_test_anon(folio)) {
1905 * To call huge_pmd_unshare, i_mmap_rwsem must be
1906 * held in write mode. Caller needs to explicitly
1907 * do this outside rmap routines.
1909 VM_BUG_ON(!(flags & TTU_RMAP_LOCKED));
1911 if (huge_pmd_unshare(mm, vma, &address, pvmw.pte)) {
1912 flush_tlb_range(vma, range.start, range.end);
1913 mmu_notifier_invalidate_range(mm, range.start,
1917 * The ref count of the PMD page was dropped
1918 * which is part of the way map counting
1919 * is done for shared PMDs. Return 'true'
1920 * here. When there is no other sharing,
1921 * huge_pmd_unshare returns false and we will
1922 * unmap the actual page and drop map count
1925 page_vma_mapped_walk_done(&pvmw);
1930 flush_cache_page(vma, address, pte_pfn(*pvmw.pte));
1933 /* Nuke the page table entry. */
1934 pteval = ptep_clear_flush(vma, address, pvmw.pte);
1936 /* Set the dirty flag on the folio now the pte is gone. */
1937 if (pte_dirty(pteval))
1938 folio_mark_dirty(folio);
1940 /* Update high watermark before we lower rss */
1941 update_hiwater_rss(mm);
1943 if (folio_is_zone_device(folio)) {
1944 unsigned long pfn = folio_pfn(folio);
1949 BUG_ON(page_try_share_anon_rmap(subpage));
1952 * Store the pfn of the page in a special migration
1953 * pte. do_swap_page() will wait until the migration
1954 * pte is removed and then restart fault handling.
1956 entry = pte_to_swp_entry(pteval);
1957 if (is_writable_device_private_entry(entry))
1958 entry = make_writable_migration_entry(pfn);
1959 else if (anon_exclusive)
1960 entry = make_readable_exclusive_migration_entry(pfn);
1962 entry = make_readable_migration_entry(pfn);
1963 swp_pte = swp_entry_to_pte(entry);
1966 * pteval maps a zone device page and is therefore
1969 if (pte_swp_soft_dirty(pteval))
1970 swp_pte = pte_swp_mksoft_dirty(swp_pte);
1971 if (pte_swp_uffd_wp(pteval))
1972 swp_pte = pte_swp_mkuffd_wp(swp_pte);
1973 set_pte_at(mm, pvmw.address, pvmw.pte, swp_pte);
1974 trace_set_migration_pte(pvmw.address, pte_val(swp_pte),
1975 compound_order(&folio->page));
1977 * No need to invalidate here it will synchronize on
1978 * against the special swap migration pte.
1980 * The assignment to subpage above was computed from a
1981 * swap PTE which results in an invalid pointer.
1982 * Since only PAGE_SIZE pages can currently be
1983 * migrated, just set it to page. This will need to be
1984 * changed when hugepage migrations to device private
1985 * memory are supported.
1987 subpage = &folio->page;
1988 } else if (PageHWPoison(subpage)) {
1989 pteval = swp_entry_to_pte(make_hwpoison_entry(subpage));
1990 if (folio_test_hugetlb(folio)) {
1991 hugetlb_count_sub(folio_nr_pages(folio), mm);
1992 set_huge_swap_pte_at(mm, address,
1994 vma_mmu_pagesize(vma));
1996 dec_mm_counter(mm, mm_counter(&folio->page));
1997 set_pte_at(mm, address, pvmw.pte, pteval);
2000 } else if (pte_unused(pteval) && !userfaultfd_armed(vma)) {
2002 * The guest indicated that the page content is of no
2003 * interest anymore. Simply discard the pte, vmscan
2004 * will take care of the rest.
2005 * A future reference will then fault in a new zero
2006 * page. When userfaultfd is active, we must not drop
2007 * this page though, as its main user (postcopy
2008 * migration) will not expect userfaults on already
2011 dec_mm_counter(mm, mm_counter(&folio->page));
2012 /* We have to invalidate as we cleared the pte */
2013 mmu_notifier_invalidate_range(mm, address,
2014 address + PAGE_SIZE);
2019 if (arch_unmap_one(mm, vma, address, pteval) < 0) {
2020 set_pte_at(mm, address, pvmw.pte, pteval);
2022 page_vma_mapped_walk_done(&pvmw);
2025 VM_BUG_ON_PAGE(pte_write(pteval) && folio_test_anon(folio) &&
2026 !anon_exclusive, subpage);
2027 if (anon_exclusive &&
2028 page_try_share_anon_rmap(subpage)) {
2029 set_pte_at(mm, address, pvmw.pte, pteval);
2031 page_vma_mapped_walk_done(&pvmw);
2036 * Store the pfn of the page in a special migration
2037 * pte. do_swap_page() will wait until the migration
2038 * pte is removed and then restart fault handling.
2040 if (pte_write(pteval))
2041 entry = make_writable_migration_entry(
2042 page_to_pfn(subpage));
2043 else if (anon_exclusive)
2044 entry = make_readable_exclusive_migration_entry(
2045 page_to_pfn(subpage));
2047 entry = make_readable_migration_entry(
2048 page_to_pfn(subpage));
2050 swp_pte = swp_entry_to_pte(entry);
2051 if (pte_soft_dirty(pteval))
2052 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2053 if (pte_uffd_wp(pteval))
2054 swp_pte = pte_swp_mkuffd_wp(swp_pte);
2055 set_pte_at(mm, address, pvmw.pte, swp_pte);
2056 trace_set_migration_pte(address, pte_val(swp_pte),
2057 compound_order(&folio->page));
2059 * No need to invalidate here it will synchronize on
2060 * against the special swap migration pte.
2065 * No need to call mmu_notifier_invalidate_range() it has be
2066 * done above for all cases requiring it to happen under page
2067 * table lock before mmu_notifier_invalidate_range_end()
2069 * See Documentation/vm/mmu_notifier.rst
2071 page_remove_rmap(subpage, vma, folio_test_hugetlb(folio));
2072 if (vma->vm_flags & VM_LOCKED)
2073 mlock_page_drain_local();
2077 mmu_notifier_invalidate_range_end(&range);
2083 * try_to_migrate - try to replace all page table mappings with swap entries
2084 * @folio: the folio to replace page table entries for
2085 * @flags: action and flags
2087 * Tries to remove all the page table entries which are mapping this folio and
2088 * replace them with special swap entries. Caller must hold the folio lock.
2090 void try_to_migrate(struct folio *folio, enum ttu_flags flags)
2092 struct rmap_walk_control rwc = {
2093 .rmap_one = try_to_migrate_one,
2094 .arg = (void *)flags,
2095 .done = page_not_mapped,
2096 .anon_lock = folio_lock_anon_vma_read,
2100 * Migration always ignores mlock and only supports TTU_RMAP_LOCKED and
2101 * TTU_SPLIT_HUGE_PMD and TTU_SYNC flags.
2103 if (WARN_ON_ONCE(flags & ~(TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD |
2107 if (folio_is_zone_device(folio) && !folio_is_device_private(folio))
2111 * During exec, a temporary VMA is setup and later moved.
2112 * The VMA is moved under the anon_vma lock but not the
2113 * page tables leading to a race where migration cannot
2114 * find the migration ptes. Rather than increasing the
2115 * locking requirements of exec(), migration skips
2116 * temporary VMAs until after exec() completes.
2118 if (!folio_test_ksm(folio) && folio_test_anon(folio))
2119 rwc.invalid_vma = invalid_migration_vma;
2121 if (flags & TTU_RMAP_LOCKED)
2122 rmap_walk_locked(folio, &rwc);
2124 rmap_walk(folio, &rwc);
2127 #ifdef CONFIG_DEVICE_PRIVATE
2128 struct make_exclusive_args {
2129 struct mm_struct *mm;
2130 unsigned long address;
2135 static bool page_make_device_exclusive_one(struct folio *folio,
2136 struct vm_area_struct *vma, unsigned long address, void *priv)
2138 struct mm_struct *mm = vma->vm_mm;
2139 DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0);
2140 struct make_exclusive_args *args = priv;
2142 struct page *subpage;
2144 struct mmu_notifier_range range;
2148 mmu_notifier_range_init_owner(&range, MMU_NOTIFY_EXCLUSIVE, 0, vma,
2149 vma->vm_mm, address, min(vma->vm_end,
2150 address + folio_size(folio)),
2152 mmu_notifier_invalidate_range_start(&range);
2154 while (page_vma_mapped_walk(&pvmw)) {
2155 /* Unexpected PMD-mapped THP? */
2156 VM_BUG_ON_FOLIO(!pvmw.pte, folio);
2158 if (!pte_present(*pvmw.pte)) {
2160 page_vma_mapped_walk_done(&pvmw);
2164 subpage = folio_page(folio,
2165 pte_pfn(*pvmw.pte) - folio_pfn(folio));
2166 address = pvmw.address;
2168 /* Nuke the page table entry. */
2169 flush_cache_page(vma, address, pte_pfn(*pvmw.pte));
2170 pteval = ptep_clear_flush(vma, address, pvmw.pte);
2172 /* Set the dirty flag on the folio now the pte is gone. */
2173 if (pte_dirty(pteval))
2174 folio_mark_dirty(folio);
2177 * Check that our target page is still mapped at the expected
2180 if (args->mm == mm && args->address == address &&
2185 * Store the pfn of the page in a special migration
2186 * pte. do_swap_page() will wait until the migration
2187 * pte is removed and then restart fault handling.
2189 if (pte_write(pteval))
2190 entry = make_writable_device_exclusive_entry(
2191 page_to_pfn(subpage));
2193 entry = make_readable_device_exclusive_entry(
2194 page_to_pfn(subpage));
2195 swp_pte = swp_entry_to_pte(entry);
2196 if (pte_soft_dirty(pteval))
2197 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2198 if (pte_uffd_wp(pteval))
2199 swp_pte = pte_swp_mkuffd_wp(swp_pte);
2201 set_pte_at(mm, address, pvmw.pte, swp_pte);
2204 * There is a reference on the page for the swap entry which has
2205 * been removed, so shouldn't take another.
2207 page_remove_rmap(subpage, vma, false);
2210 mmu_notifier_invalidate_range_end(&range);
2216 * folio_make_device_exclusive - Mark the folio exclusively owned by a device.
2217 * @folio: The folio to replace page table entries for.
2218 * @mm: The mm_struct where the folio is expected to be mapped.
2219 * @address: Address where the folio is expected to be mapped.
2220 * @owner: passed to MMU_NOTIFY_EXCLUSIVE range notifier callbacks
2222 * Tries to remove all the page table entries which are mapping this
2223 * folio and replace them with special device exclusive swap entries to
2224 * grant a device exclusive access to the folio.
2226 * Context: Caller must hold the folio lock.
2227 * Return: false if the page is still mapped, or if it could not be unmapped
2228 * from the expected address. Otherwise returns true (success).
2230 static bool folio_make_device_exclusive(struct folio *folio,
2231 struct mm_struct *mm, unsigned long address, void *owner)
2233 struct make_exclusive_args args = {
2239 struct rmap_walk_control rwc = {
2240 .rmap_one = page_make_device_exclusive_one,
2241 .done = page_not_mapped,
2242 .anon_lock = folio_lock_anon_vma_read,
2247 * Restrict to anonymous folios for now to avoid potential writeback
2250 if (!folio_test_anon(folio))
2253 rmap_walk(folio, &rwc);
2255 return args.valid && !folio_mapcount(folio);
2259 * make_device_exclusive_range() - Mark a range for exclusive use by a device
2260 * @mm: mm_struct of associated target process
2261 * @start: start of the region to mark for exclusive device access
2262 * @end: end address of region
2263 * @pages: returns the pages which were successfully marked for exclusive access
2264 * @owner: passed to MMU_NOTIFY_EXCLUSIVE range notifier to allow filtering
2266 * Returns: number of pages found in the range by GUP. A page is marked for
2267 * exclusive access only if the page pointer is non-NULL.
2269 * This function finds ptes mapping page(s) to the given address range, locks
2270 * them and replaces mappings with special swap entries preventing userspace CPU
2271 * access. On fault these entries are replaced with the original mapping after
2272 * calling MMU notifiers.
2274 * A driver using this to program access from a device must use a mmu notifier
2275 * critical section to hold a device specific lock during programming. Once
2276 * programming is complete it should drop the page lock and reference after
2277 * which point CPU access to the page will revoke the exclusive access.
2279 int make_device_exclusive_range(struct mm_struct *mm, unsigned long start,
2280 unsigned long end, struct page **pages,
2283 long npages = (end - start) >> PAGE_SHIFT;
2286 npages = get_user_pages_remote(mm, start, npages,
2287 FOLL_GET | FOLL_WRITE | FOLL_SPLIT_PMD,
2292 for (i = 0; i < npages; i++, start += PAGE_SIZE) {
2293 struct folio *folio = page_folio(pages[i]);
2294 if (PageTail(pages[i]) || !folio_trylock(folio)) {
2300 if (!folio_make_device_exclusive(folio, mm, start, owner)) {
2301 folio_unlock(folio);
2309 EXPORT_SYMBOL_GPL(make_device_exclusive_range);
2312 void __put_anon_vma(struct anon_vma *anon_vma)
2314 struct anon_vma *root = anon_vma->root;
2316 anon_vma_free(anon_vma);
2317 if (root != anon_vma && atomic_dec_and_test(&root->refcount))
2318 anon_vma_free(root);
2321 static struct anon_vma *rmap_walk_anon_lock(struct folio *folio,
2322 const struct rmap_walk_control *rwc)
2324 struct anon_vma *anon_vma;
2327 return rwc->anon_lock(folio);
2330 * Note: remove_migration_ptes() cannot use folio_lock_anon_vma_read()
2331 * because that depends on page_mapped(); but not all its usages
2332 * are holding mmap_lock. Users without mmap_lock are required to
2333 * take a reference count to prevent the anon_vma disappearing
2335 anon_vma = folio_anon_vma(folio);
2339 anon_vma_lock_read(anon_vma);
2344 * rmap_walk_anon - do something to anonymous page using the object-based
2346 * @page: the page to be handled
2347 * @rwc: control variable according to each walk type
2349 * Find all the mappings of a page using the mapping pointer and the vma chains
2350 * contained in the anon_vma struct it points to.
2352 static void rmap_walk_anon(struct folio *folio,
2353 const struct rmap_walk_control *rwc, bool locked)
2355 struct anon_vma *anon_vma;
2356 pgoff_t pgoff_start, pgoff_end;
2357 struct anon_vma_chain *avc;
2360 anon_vma = folio_anon_vma(folio);
2361 /* anon_vma disappear under us? */
2362 VM_BUG_ON_FOLIO(!anon_vma, folio);
2364 anon_vma = rmap_walk_anon_lock(folio, rwc);
2369 pgoff_start = folio_pgoff(folio);
2370 pgoff_end = pgoff_start + folio_nr_pages(folio) - 1;
2371 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root,
2372 pgoff_start, pgoff_end) {
2373 struct vm_area_struct *vma = avc->vma;
2374 unsigned long address = vma_address(&folio->page, vma);
2376 VM_BUG_ON_VMA(address == -EFAULT, vma);
2379 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2382 if (!rwc->rmap_one(folio, vma, address, rwc->arg))
2384 if (rwc->done && rwc->done(folio))
2389 anon_vma_unlock_read(anon_vma);
2393 * rmap_walk_file - do something to file page using the object-based rmap method
2394 * @page: the page to be handled
2395 * @rwc: control variable according to each walk type
2397 * Find all the mappings of a page using the mapping pointer and the vma chains
2398 * contained in the address_space struct it points to.
2400 static void rmap_walk_file(struct folio *folio,
2401 const struct rmap_walk_control *rwc, bool locked)
2403 struct address_space *mapping = folio_mapping(folio);
2404 pgoff_t pgoff_start, pgoff_end;
2405 struct vm_area_struct *vma;
2408 * The page lock not only makes sure that page->mapping cannot
2409 * suddenly be NULLified by truncation, it makes sure that the
2410 * structure at mapping cannot be freed and reused yet,
2411 * so we can safely take mapping->i_mmap_rwsem.
2413 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
2418 pgoff_start = folio_pgoff(folio);
2419 pgoff_end = pgoff_start + folio_nr_pages(folio) - 1;
2421 i_mmap_lock_read(mapping);
2422 vma_interval_tree_foreach(vma, &mapping->i_mmap,
2423 pgoff_start, pgoff_end) {
2424 unsigned long address = vma_address(&folio->page, vma);
2426 VM_BUG_ON_VMA(address == -EFAULT, vma);
2429 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2432 if (!rwc->rmap_one(folio, vma, address, rwc->arg))
2434 if (rwc->done && rwc->done(folio))
2440 i_mmap_unlock_read(mapping);
2443 void rmap_walk(struct folio *folio, const struct rmap_walk_control *rwc)
2445 if (unlikely(folio_test_ksm(folio)))
2446 rmap_walk_ksm(folio, rwc);
2447 else if (folio_test_anon(folio))
2448 rmap_walk_anon(folio, rwc, false);
2450 rmap_walk_file(folio, rwc, false);
2453 /* Like rmap_walk, but caller holds relevant rmap lock */
2454 void rmap_walk_locked(struct folio *folio, const struct rmap_walk_control *rwc)
2456 /* no ksm support for now */
2457 VM_BUG_ON_FOLIO(folio_test_ksm(folio), folio);
2458 if (folio_test_anon(folio))
2459 rmap_walk_anon(folio, rwc, true);
2461 rmap_walk_file(folio, rwc, true);
2464 #ifdef CONFIG_HUGETLB_PAGE
2466 * The following two functions are for anonymous (private mapped) hugepages.
2467 * Unlike common anonymous pages, anonymous hugepages have no accounting code
2468 * and no lru code, because we handle hugepages differently from common pages.
2470 * RMAP_COMPOUND is ignored.
2472 void hugepage_add_anon_rmap(struct page *page, struct vm_area_struct *vma,
2473 unsigned long address, rmap_t flags)
2475 struct anon_vma *anon_vma = vma->anon_vma;
2478 BUG_ON(!PageLocked(page));
2480 /* address might be in next vma when migration races vma_adjust */
2481 first = atomic_inc_and_test(compound_mapcount_ptr(page));
2482 VM_BUG_ON_PAGE(!first && (flags & RMAP_EXCLUSIVE), page);
2483 VM_BUG_ON_PAGE(!first && PageAnonExclusive(page), page);
2485 __page_set_anon_rmap(page, vma, address,
2486 !!(flags & RMAP_EXCLUSIVE));
2489 void hugepage_add_new_anon_rmap(struct page *page,
2490 struct vm_area_struct *vma, unsigned long address)
2492 BUG_ON(address < vma->vm_start || address >= vma->vm_end);
2493 atomic_set(compound_mapcount_ptr(page), 0);
2494 atomic_set(compound_pincount_ptr(page), 0);
2496 __page_set_anon_rmap(page, vma, address, 1);
2498 #endif /* CONFIG_HUGETLB_PAGE */