1 // SPDX-License-Identifier: GPL-2.0-only
5 * Copyright (C) 2015 Red Hat, Inc.
9 #include <linux/sched/signal.h>
10 #include <linux/pagemap.h>
11 #include <linux/rmap.h>
12 #include <linux/swap.h>
13 #include <linux/swapops.h>
14 #include <linux/userfaultfd_k.h>
15 #include <linux/mmu_notifier.h>
16 #include <linux/hugetlb.h>
17 #include <linux/shmem_fs.h>
18 #include <asm/tlbflush.h>
22 static __always_inline
23 bool validate_dst_vma(struct vm_area_struct *dst_vma, unsigned long dst_end)
25 /* Make sure that the dst range is fully within dst_vma. */
26 if (dst_end > dst_vma->vm_end)
30 * Check the vma is registered in uffd, this is required to
31 * enforce the VM_MAYWRITE check done at uffd registration
34 if (!dst_vma->vm_userfaultfd_ctx.ctx)
40 static __always_inline
41 struct vm_area_struct *find_vma_and_prepare_anon(struct mm_struct *mm,
44 struct vm_area_struct *vma;
46 mmap_assert_locked(mm);
47 vma = vma_lookup(mm, addr);
49 vma = ERR_PTR(-ENOENT);
50 else if (!(vma->vm_flags & VM_SHARED) &&
51 unlikely(anon_vma_prepare(vma)))
52 vma = ERR_PTR(-ENOMEM);
57 #ifdef CONFIG_PER_VMA_LOCK
59 * uffd_lock_vma() - Lookup and lock vma corresponding to @address.
60 * @mm: mm to search vma in.
61 * @address: address that the vma should contain.
63 * Should be called without holding mmap_lock.
65 * Return: A locked vma containing @address, -ENOENT if no vma is found, or
66 * -ENOMEM if anon_vma couldn't be allocated.
68 static struct vm_area_struct *uffd_lock_vma(struct mm_struct *mm,
69 unsigned long address)
71 struct vm_area_struct *vma;
73 vma = lock_vma_under_rcu(mm, address);
76 * We know we're going to need to use anon_vma, so check
79 if (!(vma->vm_flags & VM_SHARED) && unlikely(!vma->anon_vma))
86 vma = find_vma_and_prepare_anon(mm, address);
89 * We cannot use vma_start_read() as it may fail due to
90 * false locked (see comment in vma_start_read()). We
91 * can avoid that by directly locking vm_lock under
92 * mmap_lock, which guarantees that nobody can lock the
93 * vma for write (vma_start_write()) under us.
95 down_read(&vma->vm_lock->lock);
102 static struct vm_area_struct *uffd_mfill_lock(struct mm_struct *dst_mm,
103 unsigned long dst_start,
106 struct vm_area_struct *dst_vma;
108 dst_vma = uffd_lock_vma(dst_mm, dst_start);
109 if (IS_ERR(dst_vma) || validate_dst_vma(dst_vma, dst_start + len))
112 vma_end_read(dst_vma);
113 return ERR_PTR(-ENOENT);
116 static void uffd_mfill_unlock(struct vm_area_struct *vma)
123 static struct vm_area_struct *uffd_mfill_lock(struct mm_struct *dst_mm,
124 unsigned long dst_start,
127 struct vm_area_struct *dst_vma;
129 mmap_read_lock(dst_mm);
130 dst_vma = find_vma_and_prepare_anon(dst_mm, dst_start);
134 if (validate_dst_vma(dst_vma, dst_start + len))
137 dst_vma = ERR_PTR(-ENOENT);
139 mmap_read_unlock(dst_mm);
143 static void uffd_mfill_unlock(struct vm_area_struct *vma)
145 mmap_read_unlock(vma->vm_mm);
149 /* Check if dst_addr is outside of file's size. Must be called with ptl held. */
150 static bool mfill_file_over_size(struct vm_area_struct *dst_vma,
151 unsigned long dst_addr)
154 pgoff_t offset, max_off;
156 if (!dst_vma->vm_file)
159 inode = dst_vma->vm_file->f_inode;
160 offset = linear_page_index(dst_vma, dst_addr);
161 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
162 return offset >= max_off;
166 * Install PTEs, to map dst_addr (within dst_vma) to page.
168 * This function handles both MCOPY_ATOMIC_NORMAL and _CONTINUE for both shmem
169 * and anon, and for both shared and private VMAs.
171 int mfill_atomic_install_pte(pmd_t *dst_pmd,
172 struct vm_area_struct *dst_vma,
173 unsigned long dst_addr, struct page *page,
174 bool newly_allocated, uffd_flags_t flags)
177 struct mm_struct *dst_mm = dst_vma->vm_mm;
178 pte_t _dst_pte, *dst_pte;
179 bool writable = dst_vma->vm_flags & VM_WRITE;
180 bool vm_shared = dst_vma->vm_flags & VM_SHARED;
182 struct folio *folio = page_folio(page);
183 bool page_in_cache = folio_mapping(folio);
185 _dst_pte = mk_pte(page, dst_vma->vm_page_prot);
186 _dst_pte = pte_mkdirty(_dst_pte);
187 if (page_in_cache && !vm_shared)
190 _dst_pte = pte_mkwrite(_dst_pte, dst_vma);
191 if (flags & MFILL_ATOMIC_WP)
192 _dst_pte = pte_mkuffd_wp(_dst_pte);
195 dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
199 if (mfill_file_over_size(dst_vma, dst_addr)) {
206 * We allow to overwrite a pte marker: consider when both MISSING|WP
207 * registered, we firstly wr-protect a none pte which has no page cache
208 * page backing it, then access the page.
210 if (!pte_none_mostly(ptep_get(dst_pte)))
214 /* Usually, cache pages are already added to LRU */
216 folio_add_lru(folio);
217 folio_add_file_rmap_pte(folio, page, dst_vma);
219 folio_add_new_anon_rmap(folio, dst_vma, dst_addr);
220 folio_add_lru_vma(folio, dst_vma);
224 * Must happen after rmap, as mm_counter() checks mapping (via
225 * PageAnon()), which is set by __page_set_anon_rmap().
227 inc_mm_counter(dst_mm, mm_counter(folio));
229 set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
231 /* No need to invalidate - it was non-present before */
232 update_mmu_cache(dst_vma, dst_addr, dst_pte);
235 pte_unmap_unlock(dst_pte, ptl);
240 static int mfill_atomic_pte_copy(pmd_t *dst_pmd,
241 struct vm_area_struct *dst_vma,
242 unsigned long dst_addr,
243 unsigned long src_addr,
245 struct folio **foliop)
253 folio = vma_alloc_folio(GFP_HIGHUSER_MOVABLE, 0, dst_vma,
258 kaddr = kmap_local_folio(folio, 0);
260 * The read mmap_lock is held here. Despite the
261 * mmap_lock being read recursive a deadlock is still
262 * possible if a writer has taken a lock. For example:
264 * process A thread 1 takes read lock on own mmap_lock
265 * process A thread 2 calls mmap, blocks taking write lock
266 * process B thread 1 takes page fault, read lock on own mmap lock
267 * process B thread 2 calls mmap, blocks taking write lock
268 * process A thread 1 blocks taking read lock on process B
269 * process B thread 1 blocks taking read lock on process A
271 * Disable page faults to prevent potential deadlock
272 * and retry the copy outside the mmap_lock.
275 ret = copy_from_user(kaddr, (const void __user *) src_addr,
280 /* fallback to copy_from_user outside mmap_lock */
284 /* don't free the page */
288 flush_dcache_folio(folio);
295 * The memory barrier inside __folio_mark_uptodate makes sure that
296 * preceding stores to the page contents become visible before
297 * the set_pte_at() write.
299 __folio_mark_uptodate(folio);
302 if (mem_cgroup_charge(folio, dst_vma->vm_mm, GFP_KERNEL))
305 ret = mfill_atomic_install_pte(dst_pmd, dst_vma, dst_addr,
306 &folio->page, true, flags);
316 static int mfill_atomic_pte_zeroed_folio(pmd_t *dst_pmd,
317 struct vm_area_struct *dst_vma,
318 unsigned long dst_addr)
323 folio = vma_alloc_zeroed_movable_folio(dst_vma, dst_addr);
327 if (mem_cgroup_charge(folio, dst_vma->vm_mm, GFP_KERNEL))
331 * The memory barrier inside __folio_mark_uptodate makes sure that
332 * zeroing out the folio become visible before mapping the page
333 * using set_pte_at(). See do_anonymous_page().
335 __folio_mark_uptodate(folio);
337 ret = mfill_atomic_install_pte(dst_pmd, dst_vma, dst_addr,
338 &folio->page, true, 0);
348 static int mfill_atomic_pte_zeropage(pmd_t *dst_pmd,
349 struct vm_area_struct *dst_vma,
350 unsigned long dst_addr)
352 pte_t _dst_pte, *dst_pte;
356 if (mm_forbids_zeropage(dst_vma->vm_mm))
357 return mfill_atomic_pte_zeroed_folio(dst_pmd, dst_vma, dst_addr);
359 _dst_pte = pte_mkspecial(pfn_pte(my_zero_pfn(dst_addr),
360 dst_vma->vm_page_prot));
362 dst_pte = pte_offset_map_lock(dst_vma->vm_mm, dst_pmd, dst_addr, &ptl);
365 if (mfill_file_over_size(dst_vma, dst_addr)) {
370 if (!pte_none(ptep_get(dst_pte)))
372 set_pte_at(dst_vma->vm_mm, dst_addr, dst_pte, _dst_pte);
373 /* No need to invalidate - it was non-present before */
374 update_mmu_cache(dst_vma, dst_addr, dst_pte);
377 pte_unmap_unlock(dst_pte, ptl);
382 /* Handles UFFDIO_CONTINUE for all shmem VMAs (shared or private). */
383 static int mfill_atomic_pte_continue(pmd_t *dst_pmd,
384 struct vm_area_struct *dst_vma,
385 unsigned long dst_addr,
388 struct inode *inode = file_inode(dst_vma->vm_file);
389 pgoff_t pgoff = linear_page_index(dst_vma, dst_addr);
394 ret = shmem_get_folio(inode, pgoff, &folio, SGP_NOALLOC);
395 /* Our caller expects us to return -EFAULT if we failed to find folio */
405 page = folio_file_page(folio, pgoff);
406 if (PageHWPoison(page)) {
411 ret = mfill_atomic_install_pte(dst_pmd, dst_vma, dst_addr,
426 /* Handles UFFDIO_POISON for all non-hugetlb VMAs. */
427 static int mfill_atomic_pte_poison(pmd_t *dst_pmd,
428 struct vm_area_struct *dst_vma,
429 unsigned long dst_addr,
433 struct mm_struct *dst_mm = dst_vma->vm_mm;
434 pte_t _dst_pte, *dst_pte;
437 _dst_pte = make_pte_marker(PTE_MARKER_POISONED);
439 dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
443 if (mfill_file_over_size(dst_vma, dst_addr)) {
449 /* Refuse to overwrite any PTE, even a PTE marker (e.g. UFFD WP). */
450 if (!pte_none(ptep_get(dst_pte)))
453 set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
455 /* No need to invalidate - it was non-present before */
456 update_mmu_cache(dst_vma, dst_addr, dst_pte);
459 pte_unmap_unlock(dst_pte, ptl);
464 static pmd_t *mm_alloc_pmd(struct mm_struct *mm, unsigned long address)
470 pgd = pgd_offset(mm, address);
471 p4d = p4d_alloc(mm, pgd, address);
474 pud = pud_alloc(mm, p4d, address);
478 * Note that we didn't run this because the pmd was
479 * missing, the *pmd may be already established and in
480 * turn it may also be a trans_huge_pmd.
482 return pmd_alloc(mm, pud, address);
485 #ifdef CONFIG_HUGETLB_PAGE
487 * mfill_atomic processing for HUGETLB vmas. Note that this routine is
488 * called with either vma-lock or mmap_lock held, it will release the lock
491 static __always_inline ssize_t mfill_atomic_hugetlb(
492 struct userfaultfd_ctx *ctx,
493 struct vm_area_struct *dst_vma,
494 unsigned long dst_start,
495 unsigned long src_start,
499 struct mm_struct *dst_mm = dst_vma->vm_mm;
502 unsigned long src_addr, dst_addr;
505 unsigned long vma_hpagesize;
508 struct address_space *mapping;
511 * There is no default zero huge page for all huge page sizes as
512 * supported by hugetlb. A PMD_SIZE huge pages may exist as used
513 * by THP. Since we can not reliably insert a zero page, this
514 * feature is not supported.
516 if (uffd_flags_mode_is(flags, MFILL_ATOMIC_ZEROPAGE)) {
517 up_read(&ctx->map_changing_lock);
518 uffd_mfill_unlock(dst_vma);
522 src_addr = src_start;
523 dst_addr = dst_start;
526 vma_hpagesize = vma_kernel_pagesize(dst_vma);
529 * Validate alignment based on huge page size
532 if (dst_start & (vma_hpagesize - 1) || len & (vma_hpagesize - 1))
537 * On routine entry dst_vma is set. If we had to drop mmap_lock and
538 * retry, dst_vma will be set to NULL and we must lookup again.
541 dst_vma = uffd_mfill_lock(dst_mm, dst_start, len);
542 if (IS_ERR(dst_vma)) {
543 err = PTR_ERR(dst_vma);
548 if (!is_vm_hugetlb_page(dst_vma))
552 if (vma_hpagesize != vma_kernel_pagesize(dst_vma))
556 * If memory mappings are changing because of non-cooperative
557 * operation (e.g. mremap) running in parallel, bail out and
558 * request the user to retry later
560 down_read(&ctx->map_changing_lock);
562 if (atomic_read(&ctx->mmap_changing))
566 while (src_addr < src_start + len) {
567 BUG_ON(dst_addr >= dst_start + len);
570 * Serialize via vma_lock and hugetlb_fault_mutex.
571 * vma_lock ensures the dst_pte remains valid even
572 * in the case of shared pmds. fault mutex prevents
573 * races with other faulting threads.
575 idx = linear_page_index(dst_vma, dst_addr);
576 mapping = dst_vma->vm_file->f_mapping;
577 hash = hugetlb_fault_mutex_hash(mapping, idx);
578 mutex_lock(&hugetlb_fault_mutex_table[hash]);
579 hugetlb_vma_lock_read(dst_vma);
582 dst_pte = huge_pte_alloc(dst_mm, dst_vma, dst_addr, vma_hpagesize);
584 hugetlb_vma_unlock_read(dst_vma);
585 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
589 if (!uffd_flags_mode_is(flags, MFILL_ATOMIC_CONTINUE) &&
590 !huge_pte_none_mostly(huge_ptep_get(dst_pte))) {
592 hugetlb_vma_unlock_read(dst_vma);
593 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
597 err = hugetlb_mfill_atomic_pte(dst_pte, dst_vma, dst_addr,
598 src_addr, flags, &folio);
600 hugetlb_vma_unlock_read(dst_vma);
601 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
605 if (unlikely(err == -ENOENT)) {
606 up_read(&ctx->map_changing_lock);
607 uffd_mfill_unlock(dst_vma);
610 err = copy_folio_from_user(folio,
611 (const void __user *)src_addr, true);
623 dst_addr += vma_hpagesize;
624 src_addr += vma_hpagesize;
625 copied += vma_hpagesize;
627 if (fatal_signal_pending(current))
635 up_read(&ctx->map_changing_lock);
637 uffd_mfill_unlock(dst_vma);
643 BUG_ON(!copied && !err);
644 return copied ? copied : err;
646 #else /* !CONFIG_HUGETLB_PAGE */
647 /* fail at build time if gcc attempts to use this */
648 extern ssize_t mfill_atomic_hugetlb(struct userfaultfd_ctx *ctx,
649 struct vm_area_struct *dst_vma,
650 unsigned long dst_start,
651 unsigned long src_start,
654 #endif /* CONFIG_HUGETLB_PAGE */
656 static __always_inline ssize_t mfill_atomic_pte(pmd_t *dst_pmd,
657 struct vm_area_struct *dst_vma,
658 unsigned long dst_addr,
659 unsigned long src_addr,
661 struct folio **foliop)
665 if (uffd_flags_mode_is(flags, MFILL_ATOMIC_CONTINUE)) {
666 return mfill_atomic_pte_continue(dst_pmd, dst_vma,
668 } else if (uffd_flags_mode_is(flags, MFILL_ATOMIC_POISON)) {
669 return mfill_atomic_pte_poison(dst_pmd, dst_vma,
674 * The normal page fault path for a shmem will invoke the
675 * fault, fill the hole in the file and COW it right away. The
676 * result generates plain anonymous memory. So when we are
677 * asked to fill an hole in a MAP_PRIVATE shmem mapping, we'll
678 * generate anonymous memory directly without actually filling
679 * the hole. For the MAP_PRIVATE case the robustness check
680 * only happens in the pagetable (to verify it's still none)
681 * and not in the radix tree.
683 if (!(dst_vma->vm_flags & VM_SHARED)) {
684 if (uffd_flags_mode_is(flags, MFILL_ATOMIC_COPY))
685 err = mfill_atomic_pte_copy(dst_pmd, dst_vma,
689 err = mfill_atomic_pte_zeropage(dst_pmd,
692 err = shmem_mfill_atomic_pte(dst_pmd, dst_vma,
700 static __always_inline ssize_t mfill_atomic(struct userfaultfd_ctx *ctx,
701 unsigned long dst_start,
702 unsigned long src_start,
706 struct mm_struct *dst_mm = ctx->mm;
707 struct vm_area_struct *dst_vma;
710 unsigned long src_addr, dst_addr;
715 * Sanitize the command parameters:
717 BUG_ON(dst_start & ~PAGE_MASK);
718 BUG_ON(len & ~PAGE_MASK);
720 /* Does the address range wrap, or is the span zero-sized? */
721 BUG_ON(src_start + len <= src_start);
722 BUG_ON(dst_start + len <= dst_start);
724 src_addr = src_start;
725 dst_addr = dst_start;
730 * Make sure the vma is not shared, that the dst range is
731 * both valid and fully within a single existing vma.
733 dst_vma = uffd_mfill_lock(dst_mm, dst_start, len);
734 if (IS_ERR(dst_vma)) {
735 err = PTR_ERR(dst_vma);
740 * If memory mappings are changing because of non-cooperative
741 * operation (e.g. mremap) running in parallel, bail out and
742 * request the user to retry later
744 down_read(&ctx->map_changing_lock);
746 if (atomic_read(&ctx->mmap_changing))
751 * shmem_zero_setup is invoked in mmap for MAP_ANONYMOUS|MAP_SHARED but
752 * it will overwrite vm_ops, so vma_is_anonymous must return false.
754 if (WARN_ON_ONCE(vma_is_anonymous(dst_vma) &&
755 dst_vma->vm_flags & VM_SHARED))
759 * validate 'mode' now that we know the dst_vma: don't allow
760 * a wrprotect copy if the userfaultfd didn't register as WP.
762 if ((flags & MFILL_ATOMIC_WP) && !(dst_vma->vm_flags & VM_UFFD_WP))
766 * If this is a HUGETLB vma, pass off to appropriate routine
768 if (is_vm_hugetlb_page(dst_vma))
769 return mfill_atomic_hugetlb(ctx, dst_vma, dst_start,
770 src_start, len, flags);
772 if (!vma_is_anonymous(dst_vma) && !vma_is_shmem(dst_vma))
774 if (!vma_is_shmem(dst_vma) &&
775 uffd_flags_mode_is(flags, MFILL_ATOMIC_CONTINUE))
778 while (src_addr < src_start + len) {
781 BUG_ON(dst_addr >= dst_start + len);
783 dst_pmd = mm_alloc_pmd(dst_mm, dst_addr);
784 if (unlikely(!dst_pmd)) {
789 dst_pmdval = pmdp_get_lockless(dst_pmd);
791 * If the dst_pmd is mapped as THP don't
792 * override it and just be strict.
794 if (unlikely(pmd_trans_huge(dst_pmdval))) {
798 if (unlikely(pmd_none(dst_pmdval)) &&
799 unlikely(__pte_alloc(dst_mm, dst_pmd))) {
803 /* If an huge pmd materialized from under us fail */
804 if (unlikely(pmd_trans_huge(*dst_pmd))) {
809 BUG_ON(pmd_none(*dst_pmd));
810 BUG_ON(pmd_trans_huge(*dst_pmd));
812 err = mfill_atomic_pte(dst_pmd, dst_vma, dst_addr,
813 src_addr, flags, &folio);
816 if (unlikely(err == -ENOENT)) {
819 up_read(&ctx->map_changing_lock);
820 uffd_mfill_unlock(dst_vma);
823 kaddr = kmap_local_folio(folio, 0);
824 err = copy_from_user(kaddr,
825 (const void __user *) src_addr,
832 flush_dcache_folio(folio);
838 dst_addr += PAGE_SIZE;
839 src_addr += PAGE_SIZE;
842 if (fatal_signal_pending(current))
850 up_read(&ctx->map_changing_lock);
851 uffd_mfill_unlock(dst_vma);
857 BUG_ON(!copied && !err);
858 return copied ? copied : err;
861 ssize_t mfill_atomic_copy(struct userfaultfd_ctx *ctx, unsigned long dst_start,
862 unsigned long src_start, unsigned long len,
865 return mfill_atomic(ctx, dst_start, src_start, len,
866 uffd_flags_set_mode(flags, MFILL_ATOMIC_COPY));
869 ssize_t mfill_atomic_zeropage(struct userfaultfd_ctx *ctx,
873 return mfill_atomic(ctx, start, 0, len,
874 uffd_flags_set_mode(0, MFILL_ATOMIC_ZEROPAGE));
877 ssize_t mfill_atomic_continue(struct userfaultfd_ctx *ctx, unsigned long start,
878 unsigned long len, uffd_flags_t flags)
882 * A caller might reasonably assume that UFFDIO_CONTINUE contains an
883 * smp_wmb() to ensure that any writes to the about-to-be-mapped page by
884 * the thread doing the UFFDIO_CONTINUE are guaranteed to be visible to
885 * subsequent loads from the page through the newly mapped address range.
889 return mfill_atomic(ctx, start, 0, len,
890 uffd_flags_set_mode(flags, MFILL_ATOMIC_CONTINUE));
893 ssize_t mfill_atomic_poison(struct userfaultfd_ctx *ctx, unsigned long start,
894 unsigned long len, uffd_flags_t flags)
896 return mfill_atomic(ctx, start, 0, len,
897 uffd_flags_set_mode(flags, MFILL_ATOMIC_POISON));
900 long uffd_wp_range(struct vm_area_struct *dst_vma,
901 unsigned long start, unsigned long len, bool enable_wp)
903 unsigned int mm_cp_flags;
904 struct mmu_gather tlb;
907 VM_WARN_ONCE(start < dst_vma->vm_start || start + len > dst_vma->vm_end,
908 "The address range exceeds VMA boundary.\n");
910 mm_cp_flags = MM_CP_UFFD_WP;
912 mm_cp_flags = MM_CP_UFFD_WP_RESOLVE;
915 * vma->vm_page_prot already reflects that uffd-wp is enabled for this
916 * VMA (see userfaultfd_set_vm_flags()) and that all PTEs are supposed
917 * to be write-protected as default whenever protection changes.
918 * Try upgrading write permissions manually.
920 if (!enable_wp && vma_wants_manual_pte_write_upgrade(dst_vma))
921 mm_cp_flags |= MM_CP_TRY_CHANGE_WRITABLE;
922 tlb_gather_mmu(&tlb, dst_vma->vm_mm);
923 ret = change_protection(&tlb, dst_vma, start, start + len, mm_cp_flags);
924 tlb_finish_mmu(&tlb);
929 int mwriteprotect_range(struct userfaultfd_ctx *ctx, unsigned long start,
930 unsigned long len, bool enable_wp)
932 struct mm_struct *dst_mm = ctx->mm;
933 unsigned long end = start + len;
934 unsigned long _start, _end;
935 struct vm_area_struct *dst_vma;
936 unsigned long page_mask;
938 VMA_ITERATOR(vmi, dst_mm, start);
941 * Sanitize the command parameters:
943 BUG_ON(start & ~PAGE_MASK);
944 BUG_ON(len & ~PAGE_MASK);
946 /* Does the address range wrap, or is the span zero-sized? */
947 BUG_ON(start + len <= start);
949 mmap_read_lock(dst_mm);
952 * If memory mappings are changing because of non-cooperative
953 * operation (e.g. mremap) running in parallel, bail out and
954 * request the user to retry later
956 down_read(&ctx->map_changing_lock);
958 if (atomic_read(&ctx->mmap_changing))
962 for_each_vma_range(vmi, dst_vma, end) {
964 if (!userfaultfd_wp(dst_vma)) {
969 if (is_vm_hugetlb_page(dst_vma)) {
971 page_mask = vma_kernel_pagesize(dst_vma) - 1;
972 if ((start & page_mask) || (len & page_mask))
976 _start = max(dst_vma->vm_start, start);
977 _end = min(dst_vma->vm_end, end);
979 err = uffd_wp_range(dst_vma, _start, _end - _start, enable_wp);
981 /* Return 0 on success, <0 on failures */
987 up_read(&ctx->map_changing_lock);
988 mmap_read_unlock(dst_mm);
993 void double_pt_lock(spinlock_t *ptl1,
1001 /* exchange ptl1 and ptl2 */
1006 /* lock in virtual address order to avoid lock inversion */
1009 spin_lock_nested(ptl2, SINGLE_DEPTH_NESTING);
1014 void double_pt_unlock(spinlock_t *ptl1,
1027 static int move_present_pte(struct mm_struct *mm,
1028 struct vm_area_struct *dst_vma,
1029 struct vm_area_struct *src_vma,
1030 unsigned long dst_addr, unsigned long src_addr,
1031 pte_t *dst_pte, pte_t *src_pte,
1032 pte_t orig_dst_pte, pte_t orig_src_pte,
1033 spinlock_t *dst_ptl, spinlock_t *src_ptl,
1034 struct folio *src_folio)
1038 double_pt_lock(dst_ptl, src_ptl);
1040 if (!pte_same(ptep_get(src_pte), orig_src_pte) ||
1041 !pte_same(ptep_get(dst_pte), orig_dst_pte)) {
1045 if (folio_test_large(src_folio) ||
1046 folio_maybe_dma_pinned(src_folio) ||
1047 !PageAnonExclusive(&src_folio->page)) {
1052 orig_src_pte = ptep_clear_flush(src_vma, src_addr, src_pte);
1053 /* Folio got pinned from under us. Put it back and fail the move. */
1054 if (folio_maybe_dma_pinned(src_folio)) {
1055 set_pte_at(mm, src_addr, src_pte, orig_src_pte);
1060 folio_move_anon_rmap(src_folio, dst_vma);
1061 src_folio->index = linear_page_index(dst_vma, dst_addr);
1063 orig_dst_pte = mk_pte(&src_folio->page, dst_vma->vm_page_prot);
1064 /* Follow mremap() behavior and treat the entry dirty after the move */
1065 orig_dst_pte = pte_mkwrite(pte_mkdirty(orig_dst_pte), dst_vma);
1067 set_pte_at(mm, dst_addr, dst_pte, orig_dst_pte);
1069 double_pt_unlock(dst_ptl, src_ptl);
1073 static int move_swap_pte(struct mm_struct *mm,
1074 unsigned long dst_addr, unsigned long src_addr,
1075 pte_t *dst_pte, pte_t *src_pte,
1076 pte_t orig_dst_pte, pte_t orig_src_pte,
1077 spinlock_t *dst_ptl, spinlock_t *src_ptl)
1079 if (!pte_swp_exclusive(orig_src_pte))
1082 double_pt_lock(dst_ptl, src_ptl);
1084 if (!pte_same(ptep_get(src_pte), orig_src_pte) ||
1085 !pte_same(ptep_get(dst_pte), orig_dst_pte)) {
1086 double_pt_unlock(dst_ptl, src_ptl);
1090 orig_src_pte = ptep_get_and_clear(mm, src_addr, src_pte);
1091 set_pte_at(mm, dst_addr, dst_pte, orig_src_pte);
1092 double_pt_unlock(dst_ptl, src_ptl);
1097 static int move_zeropage_pte(struct mm_struct *mm,
1098 struct vm_area_struct *dst_vma,
1099 struct vm_area_struct *src_vma,
1100 unsigned long dst_addr, unsigned long src_addr,
1101 pte_t *dst_pte, pte_t *src_pte,
1102 pte_t orig_dst_pte, pte_t orig_src_pte,
1103 spinlock_t *dst_ptl, spinlock_t *src_ptl)
1107 double_pt_lock(dst_ptl, src_ptl);
1108 if (!pte_same(ptep_get(src_pte), orig_src_pte) ||
1109 !pte_same(ptep_get(dst_pte), orig_dst_pte)) {
1110 double_pt_unlock(dst_ptl, src_ptl);
1114 zero_pte = pte_mkspecial(pfn_pte(my_zero_pfn(dst_addr),
1115 dst_vma->vm_page_prot));
1116 ptep_clear_flush(src_vma, src_addr, src_pte);
1117 set_pte_at(mm, dst_addr, dst_pte, zero_pte);
1118 double_pt_unlock(dst_ptl, src_ptl);
1125 * The mmap_lock for reading is held by the caller. Just move the page
1126 * from src_pmd to dst_pmd if possible, and return true if succeeded
1127 * in moving the page.
1129 static int move_pages_pte(struct mm_struct *mm, pmd_t *dst_pmd, pmd_t *src_pmd,
1130 struct vm_area_struct *dst_vma,
1131 struct vm_area_struct *src_vma,
1132 unsigned long dst_addr, unsigned long src_addr,
1136 pte_t orig_src_pte, orig_dst_pte;
1137 pte_t src_folio_pte;
1138 spinlock_t *src_ptl, *dst_ptl;
1139 pte_t *src_pte = NULL;
1140 pte_t *dst_pte = NULL;
1142 struct folio *src_folio = NULL;
1143 struct anon_vma *src_anon_vma = NULL;
1144 struct mmu_notifier_range range;
1147 flush_cache_range(src_vma, src_addr, src_addr + PAGE_SIZE);
1148 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm,
1149 src_addr, src_addr + PAGE_SIZE);
1150 mmu_notifier_invalidate_range_start(&range);
1152 dst_pte = pte_offset_map_nolock(mm, dst_pmd, dst_addr, &dst_ptl);
1154 /* Retry if a huge pmd materialized from under us */
1155 if (unlikely(!dst_pte)) {
1160 src_pte = pte_offset_map_nolock(mm, src_pmd, src_addr, &src_ptl);
1163 * We held the mmap_lock for reading so MADV_DONTNEED
1164 * can zap transparent huge pages under us, or the
1165 * transparent huge page fault can establish new
1166 * transparent huge pages under us.
1168 if (unlikely(!src_pte)) {
1173 /* Sanity checks before the operation */
1174 if (WARN_ON_ONCE(pmd_none(*dst_pmd)) || WARN_ON_ONCE(pmd_none(*src_pmd)) ||
1175 WARN_ON_ONCE(pmd_trans_huge(*dst_pmd)) || WARN_ON_ONCE(pmd_trans_huge(*src_pmd))) {
1181 orig_dst_pte = ptep_get(dst_pte);
1182 spin_unlock(dst_ptl);
1183 if (!pte_none(orig_dst_pte)) {
1189 orig_src_pte = ptep_get(src_pte);
1190 spin_unlock(src_ptl);
1191 if (pte_none(orig_src_pte)) {
1192 if (!(mode & UFFDIO_MOVE_MODE_ALLOW_SRC_HOLES))
1194 else /* nothing to do to move a hole */
1199 /* If PTE changed after we locked the folio them start over */
1200 if (src_folio && unlikely(!pte_same(src_folio_pte, orig_src_pte))) {
1205 if (pte_present(orig_src_pte)) {
1206 if (is_zero_pfn(pte_pfn(orig_src_pte))) {
1207 err = move_zeropage_pte(mm, dst_vma, src_vma,
1208 dst_addr, src_addr, dst_pte, src_pte,
1209 orig_dst_pte, orig_src_pte,
1215 * Pin and lock both source folio and anon_vma. Since we are in
1216 * RCU read section, we can't block, so on contention have to
1217 * unmap the ptes, obtain the lock and retry.
1220 struct folio *folio;
1223 * Pin the page while holding the lock to be sure the
1224 * page isn't freed under us
1227 if (!pte_same(orig_src_pte, ptep_get(src_pte))) {
1228 spin_unlock(src_ptl);
1233 folio = vm_normal_folio(src_vma, src_addr, orig_src_pte);
1234 if (!folio || !PageAnonExclusive(&folio->page)) {
1235 spin_unlock(src_ptl);
1242 src_folio_pte = orig_src_pte;
1243 spin_unlock(src_ptl);
1245 if (!folio_trylock(src_folio)) {
1246 pte_unmap(&orig_src_pte);
1247 pte_unmap(&orig_dst_pte);
1248 src_pte = dst_pte = NULL;
1249 /* now we can block and wait */
1250 folio_lock(src_folio);
1254 if (WARN_ON_ONCE(!folio_test_anon(src_folio))) {
1260 /* at this point we have src_folio locked */
1261 if (folio_test_large(src_folio)) {
1262 /* split_folio() can block */
1263 pte_unmap(&orig_src_pte);
1264 pte_unmap(&orig_dst_pte);
1265 src_pte = dst_pte = NULL;
1266 err = split_folio(src_folio);
1269 /* have to reacquire the folio after it got split */
1270 folio_unlock(src_folio);
1271 folio_put(src_folio);
1276 if (!src_anon_vma) {
1278 * folio_referenced walks the anon_vma chain
1279 * without the folio lock. Serialize against it with
1280 * the anon_vma lock, the folio lock is not enough.
1282 src_anon_vma = folio_get_anon_vma(src_folio);
1283 if (!src_anon_vma) {
1284 /* page was unmapped from under us */
1288 if (!anon_vma_trylock_write(src_anon_vma)) {
1289 pte_unmap(&orig_src_pte);
1290 pte_unmap(&orig_dst_pte);
1291 src_pte = dst_pte = NULL;
1292 /* now we can block and wait */
1293 anon_vma_lock_write(src_anon_vma);
1298 err = move_present_pte(mm, dst_vma, src_vma,
1299 dst_addr, src_addr, dst_pte, src_pte,
1300 orig_dst_pte, orig_src_pte,
1301 dst_ptl, src_ptl, src_folio);
1303 entry = pte_to_swp_entry(orig_src_pte);
1304 if (non_swap_entry(entry)) {
1305 if (is_migration_entry(entry)) {
1306 pte_unmap(&orig_src_pte);
1307 pte_unmap(&orig_dst_pte);
1308 src_pte = dst_pte = NULL;
1309 migration_entry_wait(mm, src_pmd, src_addr);
1316 err = move_swap_pte(mm, dst_addr, src_addr,
1318 orig_dst_pte, orig_src_pte,
1324 anon_vma_unlock_write(src_anon_vma);
1325 put_anon_vma(src_anon_vma);
1328 folio_unlock(src_folio);
1329 folio_put(src_folio);
1335 mmu_notifier_invalidate_range_end(&range);
1340 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1341 static inline bool move_splits_huge_pmd(unsigned long dst_addr,
1342 unsigned long src_addr,
1343 unsigned long src_end)
1345 return (src_addr & ~HPAGE_PMD_MASK) || (dst_addr & ~HPAGE_PMD_MASK) ||
1346 src_end - src_addr < HPAGE_PMD_SIZE;
1349 static inline bool move_splits_huge_pmd(unsigned long dst_addr,
1350 unsigned long src_addr,
1351 unsigned long src_end)
1353 /* This is unreachable anyway, just to avoid warnings when HPAGE_PMD_SIZE==0 */
1358 static inline bool vma_move_compatible(struct vm_area_struct *vma)
1360 return !(vma->vm_flags & (VM_PFNMAP | VM_IO | VM_HUGETLB |
1361 VM_MIXEDMAP | VM_SHADOW_STACK));
1364 static int validate_move_areas(struct userfaultfd_ctx *ctx,
1365 struct vm_area_struct *src_vma,
1366 struct vm_area_struct *dst_vma)
1368 /* Only allow moving if both have the same access and protection */
1369 if ((src_vma->vm_flags & VM_ACCESS_FLAGS) != (dst_vma->vm_flags & VM_ACCESS_FLAGS) ||
1370 pgprot_val(src_vma->vm_page_prot) != pgprot_val(dst_vma->vm_page_prot))
1373 /* Only allow moving if both are mlocked or both aren't */
1374 if ((src_vma->vm_flags & VM_LOCKED) != (dst_vma->vm_flags & VM_LOCKED))
1378 * For now, we keep it simple and only move between writable VMAs.
1379 * Access flags are equal, therefore cheching only the source is enough.
1381 if (!(src_vma->vm_flags & VM_WRITE))
1384 /* Check if vma flags indicate content which can be moved */
1385 if (!vma_move_compatible(src_vma) || !vma_move_compatible(dst_vma))
1388 /* Ensure dst_vma is registered in uffd we are operating on */
1389 if (!dst_vma->vm_userfaultfd_ctx.ctx ||
1390 dst_vma->vm_userfaultfd_ctx.ctx != ctx)
1393 /* Only allow moving across anonymous vmas */
1394 if (!vma_is_anonymous(src_vma) || !vma_is_anonymous(dst_vma))
1400 static __always_inline
1401 int find_vmas_mm_locked(struct mm_struct *mm,
1402 unsigned long dst_start,
1403 unsigned long src_start,
1404 struct vm_area_struct **dst_vmap,
1405 struct vm_area_struct **src_vmap)
1407 struct vm_area_struct *vma;
1409 mmap_assert_locked(mm);
1410 vma = find_vma_and_prepare_anon(mm, dst_start);
1412 return PTR_ERR(vma);
1415 /* Skip finding src_vma if src_start is in dst_vma */
1416 if (src_start >= vma->vm_start && src_start < vma->vm_end)
1419 vma = vma_lookup(mm, src_start);
1427 #ifdef CONFIG_PER_VMA_LOCK
1428 static int uffd_move_lock(struct mm_struct *mm,
1429 unsigned long dst_start,
1430 unsigned long src_start,
1431 struct vm_area_struct **dst_vmap,
1432 struct vm_area_struct **src_vmap)
1434 struct vm_area_struct *vma;
1437 vma = uffd_lock_vma(mm, dst_start);
1439 return PTR_ERR(vma);
1443 * Skip finding src_vma if src_start is in dst_vma. This also ensures
1444 * that we don't lock the same vma twice.
1446 if (src_start >= vma->vm_start && src_start < vma->vm_end) {
1452 * Using uffd_lock_vma() to get src_vma can lead to following deadlock:
1456 * vma_start_read(dst_vma)
1457 * mmap_write_lock(mm)
1458 * vma_start_write(src_vma)
1459 * vma_start_read(src_vma)
1460 * mmap_read_lock(mm)
1461 * vma_start_write(dst_vma)
1463 *src_vmap = lock_vma_under_rcu(mm, src_start);
1464 if (likely(*src_vmap))
1467 /* Undo any locking and retry in mmap_lock critical section */
1468 vma_end_read(*dst_vmap);
1471 err = find_vmas_mm_locked(mm, dst_start, src_start, dst_vmap, src_vmap);
1474 * See comment in uffd_lock_vma() as to why not using
1475 * vma_start_read() here.
1477 down_read(&(*dst_vmap)->vm_lock->lock);
1478 if (*dst_vmap != *src_vmap)
1479 down_read_nested(&(*src_vmap)->vm_lock->lock,
1480 SINGLE_DEPTH_NESTING);
1482 mmap_read_unlock(mm);
1486 static void uffd_move_unlock(struct vm_area_struct *dst_vma,
1487 struct vm_area_struct *src_vma)
1489 vma_end_read(src_vma);
1490 if (src_vma != dst_vma)
1491 vma_end_read(dst_vma);
1496 static int uffd_move_lock(struct mm_struct *mm,
1497 unsigned long dst_start,
1498 unsigned long src_start,
1499 struct vm_area_struct **dst_vmap,
1500 struct vm_area_struct **src_vmap)
1505 err = find_vmas_mm_locked(mm, dst_start, src_start, dst_vmap, src_vmap);
1507 mmap_read_unlock(mm);
1511 static void uffd_move_unlock(struct vm_area_struct *dst_vma,
1512 struct vm_area_struct *src_vma)
1514 mmap_assert_locked(src_vma->vm_mm);
1515 mmap_read_unlock(dst_vma->vm_mm);
1520 * move_pages - move arbitrary anonymous pages of an existing vma
1521 * @ctx: pointer to the userfaultfd context
1522 * @dst_start: start of the destination virtual memory range
1523 * @src_start: start of the source virtual memory range
1524 * @len: length of the virtual memory range
1525 * @mode: flags from uffdio_move.mode
1527 * It will either use the mmap_lock in read mode or per-vma locks
1529 * move_pages() remaps arbitrary anonymous pages atomically in zero
1530 * copy. It only works on non shared anonymous pages because those can
1531 * be relocated without generating non linear anon_vmas in the rmap
1534 * It provides a zero copy mechanism to handle userspace page faults.
1535 * The source vma pages should have mapcount == 1, which can be
1536 * enforced by using madvise(MADV_DONTFORK) on src vma.
1538 * The thread receiving the page during the userland page fault
1539 * will receive the faulting page in the source vma through the network,
1540 * storage or any other I/O device (MADV_DONTFORK in the source vma
1541 * avoids move_pages() to fail with -EBUSY if the process forks before
1542 * move_pages() is called), then it will call move_pages() to map the
1543 * page in the faulting address in the destination vma.
1545 * This userfaultfd command works purely via pagetables, so it's the
1546 * most efficient way to move physical non shared anonymous pages
1547 * across different virtual addresses. Unlike mremap()/mmap()/munmap()
1548 * it does not create any new vmas. The mapping in the destination
1549 * address is atomic.
1551 * It only works if the vma protection bits are identical from the
1552 * source and destination vma.
1554 * It can remap non shared anonymous pages within the same vma too.
1556 * If the source virtual memory range has any unmapped holes, or if
1557 * the destination virtual memory range is not a whole unmapped hole,
1558 * move_pages() will fail respectively with -ENOENT or -EEXIST. This
1559 * provides a very strict behavior to avoid any chance of memory
1560 * corruption going unnoticed if there are userland race conditions.
1561 * Only one thread should resolve the userland page fault at any given
1562 * time for any given faulting address. This means that if two threads
1563 * try to both call move_pages() on the same destination address at the
1564 * same time, the second thread will get an explicit error from this
1567 * The command retval will return "len" is successful. The command
1568 * however can be interrupted by fatal signals or errors. If
1569 * interrupted it will return the number of bytes successfully
1570 * remapped before the interruption if any, or the negative error if
1571 * none. It will never return zero. Either it will return an error or
1572 * an amount of bytes successfully moved. If the retval reports a
1573 * "short" remap, the move_pages() command should be repeated by
1574 * userland with src+retval, dst+reval, len-retval if it wants to know
1575 * about the error that interrupted it.
1577 * The UFFDIO_MOVE_MODE_ALLOW_SRC_HOLES flag can be specified to
1578 * prevent -ENOENT errors to materialize if there are holes in the
1579 * source virtual range that is being remapped. The holes will be
1580 * accounted as successfully remapped in the retval of the
1581 * command. This is mostly useful to remap hugepage naturally aligned
1582 * virtual regions without knowing if there are transparent hugepage
1583 * in the regions or not, but preventing the risk of having to split
1584 * the hugepmd during the remap.
1586 * If there's any rmap walk that is taking the anon_vma locks without
1587 * first obtaining the folio lock (the only current instance is
1588 * folio_referenced), they will have to verify if the folio->mapping
1589 * has changed after taking the anon_vma lock. If it changed they
1590 * should release the lock and retry obtaining a new anon_vma, because
1591 * it means the anon_vma was changed by move_pages() before the lock
1592 * could be obtained. This is the only additional complexity added to
1593 * the rmap code to provide this anonymous page remapping functionality.
1595 ssize_t move_pages(struct userfaultfd_ctx *ctx, unsigned long dst_start,
1596 unsigned long src_start, unsigned long len, __u64 mode)
1598 struct mm_struct *mm = ctx->mm;
1599 struct vm_area_struct *src_vma, *dst_vma;
1600 unsigned long src_addr, dst_addr;
1601 pmd_t *src_pmd, *dst_pmd;
1605 /* Sanitize the command parameters. */
1606 if (WARN_ON_ONCE(src_start & ~PAGE_MASK) ||
1607 WARN_ON_ONCE(dst_start & ~PAGE_MASK) ||
1608 WARN_ON_ONCE(len & ~PAGE_MASK))
1611 /* Does the address range wrap, or is the span zero-sized? */
1612 if (WARN_ON_ONCE(src_start + len <= src_start) ||
1613 WARN_ON_ONCE(dst_start + len <= dst_start))
1616 err = uffd_move_lock(mm, dst_start, src_start, &dst_vma, &src_vma);
1620 /* Re-check after taking map_changing_lock */
1622 down_read(&ctx->map_changing_lock);
1623 if (likely(atomic_read(&ctx->mmap_changing)))
1626 * Make sure the vma is not shared, that the src and dst remap
1627 * ranges are both valid and fully within a single existing
1631 if (src_vma->vm_flags & VM_SHARED)
1633 if (src_start + len > src_vma->vm_end)
1636 if (dst_vma->vm_flags & VM_SHARED)
1638 if (dst_start + len > dst_vma->vm_end)
1641 err = validate_move_areas(ctx, src_vma, dst_vma);
1645 for (src_addr = src_start, dst_addr = dst_start;
1646 src_addr < src_start + len;) {
1649 unsigned long step_size;
1652 * Below works because anonymous area would not have a
1653 * transparent huge PUD. If file-backed support is added,
1654 * that case would need to be handled here.
1656 src_pmd = mm_find_pmd(mm, src_addr);
1657 if (unlikely(!src_pmd)) {
1658 if (!(mode & UFFDIO_MOVE_MODE_ALLOW_SRC_HOLES)) {
1662 src_pmd = mm_alloc_pmd(mm, src_addr);
1663 if (unlikely(!src_pmd)) {
1668 dst_pmd = mm_alloc_pmd(mm, dst_addr);
1669 if (unlikely(!dst_pmd)) {
1674 dst_pmdval = pmdp_get_lockless(dst_pmd);
1676 * If the dst_pmd is mapped as THP don't override it and just
1677 * be strict. If dst_pmd changes into TPH after this check, the
1678 * move_pages_huge_pmd() will detect the change and retry
1679 * while move_pages_pte() will detect the change and fail.
1681 if (unlikely(pmd_trans_huge(dst_pmdval))) {
1686 ptl = pmd_trans_huge_lock(src_pmd, src_vma);
1688 if (pmd_devmap(*src_pmd)) {
1694 /* Check if we can move the pmd without splitting it. */
1695 if (move_splits_huge_pmd(dst_addr, src_addr, src_start + len) ||
1696 !pmd_none(dst_pmdval)) {
1697 struct folio *folio = pmd_folio(*src_pmd);
1699 if (!folio || (!is_huge_zero_folio(folio) &&
1700 !PageAnonExclusive(&folio->page))) {
1707 split_huge_pmd(src_vma, src_pmd, src_addr);
1708 /* The folio will be split by move_pages_pte() */
1712 err = move_pages_huge_pmd(mm, dst_pmd, src_pmd,
1713 dst_pmdval, dst_vma, src_vma,
1714 dst_addr, src_addr);
1715 step_size = HPAGE_PMD_SIZE;
1717 if (pmd_none(*src_pmd)) {
1718 if (!(mode & UFFDIO_MOVE_MODE_ALLOW_SRC_HOLES)) {
1722 if (unlikely(__pte_alloc(mm, src_pmd))) {
1728 if (unlikely(pte_alloc(mm, dst_pmd))) {
1733 err = move_pages_pte(mm, dst_pmd, src_pmd,
1735 dst_addr, src_addr, mode);
1736 step_size = PAGE_SIZE;
1741 if (fatal_signal_pending(current)) {
1742 /* Do not override an error */
1743 if (!err || err == -EAGAIN)
1754 /* Proceed to the next page */
1755 dst_addr += step_size;
1756 src_addr += step_size;
1761 up_read(&ctx->map_changing_lock);
1762 uffd_move_unlock(dst_vma, src_vma);
1764 VM_WARN_ON(moved < 0);
1765 VM_WARN_ON(err > 0);
1766 VM_WARN_ON(!moved && !err);
1767 return moved ? moved : err;