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>
21 static __always_inline
22 struct vm_area_struct *find_dst_vma(struct mm_struct *dst_mm,
23 unsigned long dst_start,
27 * Make sure that the dst range is both valid and fully within a
28 * single existing vma.
30 struct vm_area_struct *dst_vma;
32 dst_vma = find_vma(dst_mm, dst_start);
36 if (dst_start < dst_vma->vm_start ||
37 dst_start + len > dst_vma->vm_end)
41 * Check the vma is registered in uffd, this is required to
42 * enforce the VM_MAYWRITE check done at uffd registration
45 if (!dst_vma->vm_userfaultfd_ctx.ctx)
51 static int mcopy_atomic_pte(struct mm_struct *dst_mm,
53 struct vm_area_struct *dst_vma,
54 unsigned long dst_addr,
55 unsigned long src_addr,
59 pte_t _dst_pte, *dst_pte;
64 pgoff_t offset, max_off;
69 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, dst_vma, dst_addr);
73 page_kaddr = kmap_atomic(page);
74 ret = copy_from_user(page_kaddr,
75 (const void __user *) src_addr,
77 kunmap_atomic(page_kaddr);
79 /* fallback to copy_from_user outside mmap_lock */
83 /* don't free the page */
87 flush_dcache_page(page);
94 * The memory barrier inside __SetPageUptodate makes sure that
95 * preceding stores to the page contents become visible before
96 * the set_pte_at() write.
98 __SetPageUptodate(page);
101 if (mem_cgroup_charge(page, dst_mm, GFP_KERNEL))
104 _dst_pte = pte_mkdirty(mk_pte(page, dst_vma->vm_page_prot));
105 if (dst_vma->vm_flags & VM_WRITE) {
107 _dst_pte = pte_mkuffd_wp(_dst_pte);
109 _dst_pte = pte_mkwrite(_dst_pte);
112 dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
113 if (dst_vma->vm_file) {
114 /* the shmem MAP_PRIVATE case requires checking the i_size */
115 inode = dst_vma->vm_file->f_inode;
116 offset = linear_page_index(dst_vma, dst_addr);
117 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
119 if (unlikely(offset >= max_off))
120 goto out_release_uncharge_unlock;
123 if (!pte_none(*dst_pte))
124 goto out_release_uncharge_unlock;
126 inc_mm_counter(dst_mm, MM_ANONPAGES);
127 page_add_new_anon_rmap(page, dst_vma, dst_addr, false);
128 lru_cache_add_inactive_or_unevictable(page, dst_vma);
130 set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
132 /* No need to invalidate - it was non-present before */
133 update_mmu_cache(dst_vma, dst_addr, dst_pte);
135 pte_unmap_unlock(dst_pte, ptl);
139 out_release_uncharge_unlock:
140 pte_unmap_unlock(dst_pte, ptl);
146 static int mfill_zeropage_pte(struct mm_struct *dst_mm,
148 struct vm_area_struct *dst_vma,
149 unsigned long dst_addr)
151 pte_t _dst_pte, *dst_pte;
154 pgoff_t offset, max_off;
157 _dst_pte = pte_mkspecial(pfn_pte(my_zero_pfn(dst_addr),
158 dst_vma->vm_page_prot));
159 dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
160 if (dst_vma->vm_file) {
161 /* the shmem MAP_PRIVATE case requires checking the i_size */
162 inode = dst_vma->vm_file->f_inode;
163 offset = linear_page_index(dst_vma, dst_addr);
164 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
166 if (unlikely(offset >= max_off))
170 if (!pte_none(*dst_pte))
172 set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
173 /* No need to invalidate - it was non-present before */
174 update_mmu_cache(dst_vma, dst_addr, dst_pte);
177 pte_unmap_unlock(dst_pte, ptl);
181 static pmd_t *mm_alloc_pmd(struct mm_struct *mm, unsigned long address)
187 pgd = pgd_offset(mm, address);
188 p4d = p4d_alloc(mm, pgd, address);
191 pud = pud_alloc(mm, p4d, address);
195 * Note that we didn't run this because the pmd was
196 * missing, the *pmd may be already established and in
197 * turn it may also be a trans_huge_pmd.
199 return pmd_alloc(mm, pud, address);
202 #ifdef CONFIG_HUGETLB_PAGE
204 * __mcopy_atomic processing for HUGETLB vmas. Note that this routine is
205 * called with mmap_lock held, it will release mmap_lock before returning.
207 static __always_inline ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
208 struct vm_area_struct *dst_vma,
209 unsigned long dst_start,
210 unsigned long src_start,
215 int vm_alloc_shared = dst_vma->vm_flags & VM_SHARED;
216 int vm_shared = dst_vma->vm_flags & VM_SHARED;
219 unsigned long src_addr, dst_addr;
222 unsigned long vma_hpagesize;
225 struct address_space *mapping;
228 * There is no default zero huge page for all huge page sizes as
229 * supported by hugetlb. A PMD_SIZE huge pages may exist as used
230 * by THP. Since we can not reliably insert a zero page, this
231 * feature is not supported.
234 mmap_read_unlock(dst_mm);
238 src_addr = src_start;
239 dst_addr = dst_start;
242 vma_hpagesize = vma_kernel_pagesize(dst_vma);
245 * Validate alignment based on huge page size
248 if (dst_start & (vma_hpagesize - 1) || len & (vma_hpagesize - 1))
253 * On routine entry dst_vma is set. If we had to drop mmap_lock and
254 * retry, dst_vma will be set to NULL and we must lookup again.
258 dst_vma = find_dst_vma(dst_mm, dst_start, len);
259 if (!dst_vma || !is_vm_hugetlb_page(dst_vma))
263 if (vma_hpagesize != vma_kernel_pagesize(dst_vma))
266 vm_shared = dst_vma->vm_flags & VM_SHARED;
270 * If not shared, ensure the dst_vma has a anon_vma.
274 if (unlikely(anon_vma_prepare(dst_vma)))
278 while (src_addr < src_start + len) {
281 BUG_ON(dst_addr >= dst_start + len);
284 * Serialize via i_mmap_rwsem and hugetlb_fault_mutex.
285 * i_mmap_rwsem ensures the dst_pte remains valid even
286 * in the case of shared pmds. fault mutex prevents
287 * races with other faulting threads.
289 mapping = dst_vma->vm_file->f_mapping;
290 i_mmap_lock_read(mapping);
291 idx = linear_page_index(dst_vma, dst_addr);
292 hash = hugetlb_fault_mutex_hash(mapping, idx);
293 mutex_lock(&hugetlb_fault_mutex_table[hash]);
296 dst_pte = huge_pte_alloc(dst_mm, dst_addr, vma_hpagesize);
298 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
299 i_mmap_unlock_read(mapping);
304 dst_pteval = huge_ptep_get(dst_pte);
305 if (!huge_pte_none(dst_pteval)) {
306 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
307 i_mmap_unlock_read(mapping);
311 err = hugetlb_mcopy_atomic_pte(dst_mm, dst_pte, dst_vma,
312 dst_addr, src_addr, &page);
314 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
315 i_mmap_unlock_read(mapping);
316 vm_alloc_shared = vm_shared;
320 if (unlikely(err == -ENOENT)) {
321 mmap_read_unlock(dst_mm);
324 err = copy_huge_page_from_user(page,
325 (const void __user *)src_addr,
326 vma_hpagesize / PAGE_SIZE,
332 mmap_read_lock(dst_mm);
334 * If memory mappings are changing because of non-cooperative
335 * operation (e.g. mremap) running in parallel, bail out and
336 * request the user to retry later
338 if (mmap_changing && READ_ONCE(*mmap_changing)) {
349 dst_addr += vma_hpagesize;
350 src_addr += vma_hpagesize;
351 copied += vma_hpagesize;
353 if (fatal_signal_pending(current))
361 mmap_read_unlock(dst_mm);
365 * We encountered an error and are about to free a newly
366 * allocated huge page.
368 * Reservation handling is very subtle, and is different for
369 * private and shared mappings. See the routine
370 * restore_reserve_on_error for details. Unfortunately, we
371 * can not call restore_reserve_on_error now as it would
372 * require holding mmap_lock.
374 * If a reservation for the page existed in the reservation
375 * map of a private mapping, the map was modified to indicate
376 * the reservation was consumed when the page was allocated.
377 * We clear the PagePrivate flag now so that the global
378 * reserve count will not be incremented in free_huge_page.
379 * The reservation map will still indicate the reservation
380 * was consumed and possibly prevent later page allocation.
381 * This is better than leaking a global reservation. If no
382 * reservation existed, it is still safe to clear PagePrivate
383 * as no adjustments to reservation counts were made during
386 * The reservation map for shared mappings indicates which
387 * pages have reservations. When a huge page is allocated
388 * for an address with a reservation, no change is made to
389 * the reserve map. In this case PagePrivate will be set
390 * to indicate that the global reservation count should be
391 * incremented when the page is freed. This is the desired
392 * behavior. However, when a huge page is allocated for an
393 * address without a reservation a reservation entry is added
394 * to the reservation map, and PagePrivate will not be set.
395 * When the page is freed, the global reserve count will NOT
396 * be incremented and it will appear as though we have leaked
397 * reserved page. In this case, set PagePrivate so that the
398 * global reserve count will be incremented to match the
399 * reservation map entry which was created.
401 * Note that vm_alloc_shared is based on the flags of the vma
402 * for which the page was originally allocated. dst_vma could
403 * be different or NULL on error.
406 SetPagePrivate(page);
408 ClearPagePrivate(page);
413 BUG_ON(!copied && !err);
414 return copied ? copied : err;
416 #else /* !CONFIG_HUGETLB_PAGE */
417 /* fail at build time if gcc attempts to use this */
418 extern ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
419 struct vm_area_struct *dst_vma,
420 unsigned long dst_start,
421 unsigned long src_start,
425 #endif /* CONFIG_HUGETLB_PAGE */
427 static __always_inline ssize_t mfill_atomic_pte(struct mm_struct *dst_mm,
429 struct vm_area_struct *dst_vma,
430 unsigned long dst_addr,
431 unsigned long src_addr,
439 * The normal page fault path for a shmem will invoke the
440 * fault, fill the hole in the file and COW it right away. The
441 * result generates plain anonymous memory. So when we are
442 * asked to fill an hole in a MAP_PRIVATE shmem mapping, we'll
443 * generate anonymous memory directly without actually filling
444 * the hole. For the MAP_PRIVATE case the robustness check
445 * only happens in the pagetable (to verify it's still none)
446 * and not in the radix tree.
448 if (!(dst_vma->vm_flags & VM_SHARED)) {
450 err = mcopy_atomic_pte(dst_mm, dst_pmd, dst_vma,
451 dst_addr, src_addr, page,
454 err = mfill_zeropage_pte(dst_mm, dst_pmd,
457 VM_WARN_ON_ONCE(wp_copy);
459 err = shmem_mcopy_atomic_pte(dst_mm, dst_pmd,
463 err = shmem_mfill_zeropage_pte(dst_mm, dst_pmd,
470 static __always_inline ssize_t __mcopy_atomic(struct mm_struct *dst_mm,
471 unsigned long dst_start,
472 unsigned long src_start,
478 struct vm_area_struct *dst_vma;
481 unsigned long src_addr, dst_addr;
487 * Sanitize the command parameters:
489 BUG_ON(dst_start & ~PAGE_MASK);
490 BUG_ON(len & ~PAGE_MASK);
492 /* Does the address range wrap, or is the span zero-sized? */
493 BUG_ON(src_start + len <= src_start);
494 BUG_ON(dst_start + len <= dst_start);
496 src_addr = src_start;
497 dst_addr = dst_start;
501 mmap_read_lock(dst_mm);
504 * If memory mappings are changing because of non-cooperative
505 * operation (e.g. mremap) running in parallel, bail out and
506 * request the user to retry later
509 if (mmap_changing && READ_ONCE(*mmap_changing))
513 * Make sure the vma is not shared, that the dst range is
514 * both valid and fully within a single existing vma.
517 dst_vma = find_dst_vma(dst_mm, dst_start, len);
523 * shmem_zero_setup is invoked in mmap for MAP_ANONYMOUS|MAP_SHARED but
524 * it will overwrite vm_ops, so vma_is_anonymous must return false.
526 if (WARN_ON_ONCE(vma_is_anonymous(dst_vma) &&
527 dst_vma->vm_flags & VM_SHARED))
531 * validate 'mode' now that we know the dst_vma: don't allow
532 * a wrprotect copy if the userfaultfd didn't register as WP.
534 wp_copy = mode & UFFDIO_COPY_MODE_WP;
535 if (wp_copy && !(dst_vma->vm_flags & VM_UFFD_WP))
539 * If this is a HUGETLB vma, pass off to appropriate routine
541 if (is_vm_hugetlb_page(dst_vma))
542 return __mcopy_atomic_hugetlb(dst_mm, dst_vma, dst_start,
543 src_start, len, mmap_changing,
546 if (!vma_is_anonymous(dst_vma) && !vma_is_shmem(dst_vma))
550 * Ensure the dst_vma has a anon_vma or this page
551 * would get a NULL anon_vma when moved in the
555 if (!(dst_vma->vm_flags & VM_SHARED) &&
556 unlikely(anon_vma_prepare(dst_vma)))
559 while (src_addr < src_start + len) {
562 BUG_ON(dst_addr >= dst_start + len);
564 dst_pmd = mm_alloc_pmd(dst_mm, dst_addr);
565 if (unlikely(!dst_pmd)) {
570 dst_pmdval = pmd_read_atomic(dst_pmd);
572 * If the dst_pmd is mapped as THP don't
573 * override it and just be strict.
575 if (unlikely(pmd_trans_huge(dst_pmdval))) {
579 if (unlikely(pmd_none(dst_pmdval)) &&
580 unlikely(__pte_alloc(dst_mm, dst_pmd))) {
584 /* If an huge pmd materialized from under us fail */
585 if (unlikely(pmd_trans_huge(*dst_pmd))) {
590 BUG_ON(pmd_none(*dst_pmd));
591 BUG_ON(pmd_trans_huge(*dst_pmd));
593 err = mfill_atomic_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
594 src_addr, &page, zeropage, wp_copy);
597 if (unlikely(err == -ENOENT)) {
600 mmap_read_unlock(dst_mm);
603 page_kaddr = kmap(page);
604 err = copy_from_user(page_kaddr,
605 (const void __user *) src_addr,
612 flush_dcache_page(page);
618 dst_addr += PAGE_SIZE;
619 src_addr += PAGE_SIZE;
622 if (fatal_signal_pending(current))
630 mmap_read_unlock(dst_mm);
636 BUG_ON(!copied && !err);
637 return copied ? copied : err;
640 ssize_t mcopy_atomic(struct mm_struct *dst_mm, unsigned long dst_start,
641 unsigned long src_start, unsigned long len,
642 bool *mmap_changing, __u64 mode)
644 return __mcopy_atomic(dst_mm, dst_start, src_start, len, false,
645 mmap_changing, mode);
648 ssize_t mfill_zeropage(struct mm_struct *dst_mm, unsigned long start,
649 unsigned long len, bool *mmap_changing)
651 return __mcopy_atomic(dst_mm, start, 0, len, true, mmap_changing, 0);
654 int mwriteprotect_range(struct mm_struct *dst_mm, unsigned long start,
655 unsigned long len, bool enable_wp, bool *mmap_changing)
657 struct vm_area_struct *dst_vma;
662 * Sanitize the command parameters:
664 BUG_ON(start & ~PAGE_MASK);
665 BUG_ON(len & ~PAGE_MASK);
667 /* Does the address range wrap, or is the span zero-sized? */
668 BUG_ON(start + len <= start);
670 mmap_read_lock(dst_mm);
673 * If memory mappings are changing because of non-cooperative
674 * operation (e.g. mremap) running in parallel, bail out and
675 * request the user to retry later
678 if (mmap_changing && READ_ONCE(*mmap_changing))
682 dst_vma = find_dst_vma(dst_mm, start, len);
684 * Make sure the vma is not shared, that the dst range is
685 * both valid and fully within a single existing vma.
687 if (!dst_vma || (dst_vma->vm_flags & VM_SHARED))
689 if (!userfaultfd_wp(dst_vma))
691 if (!vma_is_anonymous(dst_vma))
695 newprot = vm_get_page_prot(dst_vma->vm_flags & ~(VM_WRITE));
697 newprot = vm_get_page_prot(dst_vma->vm_flags);
699 change_protection(dst_vma, start, start + len, newprot,
700 enable_wp ? MM_CP_UFFD_WP : MM_CP_UFFD_WP_RESOLVE);
704 mmap_read_unlock(dst_mm);