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 */
92 * The memory barrier inside __SetPageUptodate makes sure that
93 * preceding stores to the page contents become visible before
94 * the set_pte_at() write.
96 __SetPageUptodate(page
);
99 if (mem_cgroup_charge(page
, dst_mm
, GFP_KERNEL
))
102 _dst_pte
= pte_mkdirty(mk_pte(page
, dst_vma
->vm_page_prot
));
103 if (dst_vma
->vm_flags
& VM_WRITE
) {
105 _dst_pte
= pte_mkuffd_wp(_dst_pte
);
107 _dst_pte
= pte_mkwrite(_dst_pte
);
110 dst_pte
= pte_offset_map_lock(dst_mm
, dst_pmd
, dst_addr
, &ptl
);
111 if (dst_vma
->vm_file
) {
112 /* the shmem MAP_PRIVATE case requires checking the i_size */
113 inode
= dst_vma
->vm_file
->f_inode
;
114 offset
= linear_page_index(dst_vma
, dst_addr
);
115 max_off
= DIV_ROUND_UP(i_size_read(inode
), PAGE_SIZE
);
117 if (unlikely(offset
>= max_off
))
118 goto out_release_uncharge_unlock
;
121 if (!pte_none(*dst_pte
))
122 goto out_release_uncharge_unlock
;
124 inc_mm_counter(dst_mm
, MM_ANONPAGES
);
125 page_add_new_anon_rmap(page
, dst_vma
, dst_addr
, false);
126 lru_cache_add_inactive_or_unevictable(page
, dst_vma
);
128 set_pte_at(dst_mm
, dst_addr
, dst_pte
, _dst_pte
);
130 /* No need to invalidate - it was non-present before */
131 update_mmu_cache(dst_vma
, dst_addr
, dst_pte
);
133 pte_unmap_unlock(dst_pte
, ptl
);
137 out_release_uncharge_unlock
:
138 pte_unmap_unlock(dst_pte
, ptl
);
144 static int mfill_zeropage_pte(struct mm_struct
*dst_mm
,
146 struct vm_area_struct
*dst_vma
,
147 unsigned long dst_addr
)
149 pte_t _dst_pte
, *dst_pte
;
152 pgoff_t offset
, max_off
;
155 _dst_pte
= pte_mkspecial(pfn_pte(my_zero_pfn(dst_addr
),
156 dst_vma
->vm_page_prot
));
157 dst_pte
= pte_offset_map_lock(dst_mm
, dst_pmd
, dst_addr
, &ptl
);
158 if (dst_vma
->vm_file
) {
159 /* the shmem MAP_PRIVATE case requires checking the i_size */
160 inode
= dst_vma
->vm_file
->f_inode
;
161 offset
= linear_page_index(dst_vma
, dst_addr
);
162 max_off
= DIV_ROUND_UP(i_size_read(inode
), PAGE_SIZE
);
164 if (unlikely(offset
>= max_off
))
168 if (!pte_none(*dst_pte
))
170 set_pte_at(dst_mm
, dst_addr
, dst_pte
, _dst_pte
);
171 /* No need to invalidate - it was non-present before */
172 update_mmu_cache(dst_vma
, dst_addr
, dst_pte
);
175 pte_unmap_unlock(dst_pte
, ptl
);
179 static pmd_t
*mm_alloc_pmd(struct mm_struct
*mm
, unsigned long address
)
185 pgd
= pgd_offset(mm
, address
);
186 p4d
= p4d_alloc(mm
, pgd
, address
);
189 pud
= pud_alloc(mm
, p4d
, address
);
193 * Note that we didn't run this because the pmd was
194 * missing, the *pmd may be already established and in
195 * turn it may also be a trans_huge_pmd.
197 return pmd_alloc(mm
, pud
, address
);
200 #ifdef CONFIG_HUGETLB_PAGE
202 * __mcopy_atomic processing for HUGETLB vmas. Note that this routine is
203 * called with mmap_lock held, it will release mmap_lock before returning.
205 static __always_inline ssize_t
__mcopy_atomic_hugetlb(struct mm_struct
*dst_mm
,
206 struct vm_area_struct
*dst_vma
,
207 unsigned long dst_start
,
208 unsigned long src_start
,
210 enum mcopy_atomic_mode mode
)
212 int vm_alloc_shared
= dst_vma
->vm_flags
& VM_SHARED
;
213 int vm_shared
= dst_vma
->vm_flags
& VM_SHARED
;
216 unsigned long src_addr
, dst_addr
;
219 unsigned long vma_hpagesize
;
222 struct address_space
*mapping
;
225 * There is no default zero huge page for all huge page sizes as
226 * supported by hugetlb. A PMD_SIZE huge pages may exist as used
227 * by THP. Since we can not reliably insert a zero page, this
228 * feature is not supported.
230 if (mode
== MCOPY_ATOMIC_ZEROPAGE
) {
231 mmap_read_unlock(dst_mm
);
235 src_addr
= src_start
;
236 dst_addr
= dst_start
;
239 vma_hpagesize
= vma_kernel_pagesize(dst_vma
);
242 * Validate alignment based on huge page size
245 if (dst_start
& (vma_hpagesize
- 1) || len
& (vma_hpagesize
- 1))
250 * On routine entry dst_vma is set. If we had to drop mmap_lock and
251 * retry, dst_vma will be set to NULL and we must lookup again.
255 dst_vma
= find_dst_vma(dst_mm
, dst_start
, len
);
256 if (!dst_vma
|| !is_vm_hugetlb_page(dst_vma
))
260 if (vma_hpagesize
!= vma_kernel_pagesize(dst_vma
))
263 vm_shared
= dst_vma
->vm_flags
& VM_SHARED
;
267 * If not shared, ensure the dst_vma has a anon_vma.
271 if (unlikely(anon_vma_prepare(dst_vma
)))
275 while (src_addr
< src_start
+ len
) {
276 BUG_ON(dst_addr
>= dst_start
+ len
);
279 * Serialize via i_mmap_rwsem and hugetlb_fault_mutex.
280 * i_mmap_rwsem ensures the dst_pte remains valid even
281 * in the case of shared pmds. fault mutex prevents
282 * races with other faulting threads.
284 mapping
= dst_vma
->vm_file
->f_mapping
;
285 i_mmap_lock_read(mapping
);
286 idx
= linear_page_index(dst_vma
, dst_addr
);
287 hash
= hugetlb_fault_mutex_hash(mapping
, idx
);
288 mutex_lock(&hugetlb_fault_mutex_table
[hash
]);
291 dst_pte
= huge_pte_alloc(dst_mm
, dst_vma
, dst_addr
, vma_hpagesize
);
293 mutex_unlock(&hugetlb_fault_mutex_table
[hash
]);
294 i_mmap_unlock_read(mapping
);
298 if (mode
!= MCOPY_ATOMIC_CONTINUE
&&
299 !huge_pte_none(huge_ptep_get(dst_pte
))) {
301 mutex_unlock(&hugetlb_fault_mutex_table
[hash
]);
302 i_mmap_unlock_read(mapping
);
306 err
= hugetlb_mcopy_atomic_pte(dst_mm
, dst_pte
, dst_vma
,
307 dst_addr
, src_addr
, mode
, &page
);
309 mutex_unlock(&hugetlb_fault_mutex_table
[hash
]);
310 i_mmap_unlock_read(mapping
);
311 vm_alloc_shared
= vm_shared
;
315 if (unlikely(err
== -ENOENT
)) {
316 mmap_read_unlock(dst_mm
);
319 err
= copy_huge_page_from_user(page
,
320 (const void __user
*)src_addr
,
321 vma_hpagesize
/ PAGE_SIZE
,
327 mmap_read_lock(dst_mm
);
335 dst_addr
+= vma_hpagesize
;
336 src_addr
+= vma_hpagesize
;
337 copied
+= vma_hpagesize
;
339 if (fatal_signal_pending(current
))
347 mmap_read_unlock(dst_mm
);
351 * We encountered an error and are about to free a newly
352 * allocated huge page.
354 * Reservation handling is very subtle, and is different for
355 * private and shared mappings. See the routine
356 * restore_reserve_on_error for details. Unfortunately, we
357 * can not call restore_reserve_on_error now as it would
358 * require holding mmap_lock.
360 * If a reservation for the page existed in the reservation
361 * map of a private mapping, the map was modified to indicate
362 * the reservation was consumed when the page was allocated.
363 * We clear the HPageRestoreReserve flag now so that the global
364 * reserve count will not be incremented in free_huge_page.
365 * The reservation map will still indicate the reservation
366 * was consumed and possibly prevent later page allocation.
367 * This is better than leaking a global reservation. If no
368 * reservation existed, it is still safe to clear
369 * HPageRestoreReserve as no adjustments to reservation counts
370 * were made during allocation.
372 * The reservation map for shared mappings indicates which
373 * pages have reservations. When a huge page is allocated
374 * for an address with a reservation, no change is made to
375 * the reserve map. In this case HPageRestoreReserve will be
376 * set to indicate that the global reservation count should be
377 * incremented when the page is freed. This is the desired
378 * behavior. However, when a huge page is allocated for an
379 * address without a reservation a reservation entry is added
380 * to the reservation map, and HPageRestoreReserve will not be
381 * set. When the page is freed, the global reserve count will
382 * NOT be incremented and it will appear as though we have
383 * leaked reserved page. In this case, set HPageRestoreReserve
384 * so that the global reserve count will be incremented to
385 * match the reservation map entry which was created.
387 * Note that vm_alloc_shared is based on the flags of the vma
388 * for which the page was originally allocated. dst_vma could
389 * be different or NULL on error.
392 SetHPageRestoreReserve(page
);
394 ClearHPageRestoreReserve(page
);
399 BUG_ON(!copied
&& !err
);
400 return copied
? copied
: err
;
402 #else /* !CONFIG_HUGETLB_PAGE */
403 /* fail at build time if gcc attempts to use this */
404 extern ssize_t
__mcopy_atomic_hugetlb(struct mm_struct
*dst_mm
,
405 struct vm_area_struct
*dst_vma
,
406 unsigned long dst_start
,
407 unsigned long src_start
,
409 enum mcopy_atomic_mode mode
);
410 #endif /* CONFIG_HUGETLB_PAGE */
412 static __always_inline ssize_t
mfill_atomic_pte(struct mm_struct
*dst_mm
,
414 struct vm_area_struct
*dst_vma
,
415 unsigned long dst_addr
,
416 unsigned long src_addr
,
424 * The normal page fault path for a shmem will invoke the
425 * fault, fill the hole in the file and COW it right away. The
426 * result generates plain anonymous memory. So when we are
427 * asked to fill an hole in a MAP_PRIVATE shmem mapping, we'll
428 * generate anonymous memory directly without actually filling
429 * the hole. For the MAP_PRIVATE case the robustness check
430 * only happens in the pagetable (to verify it's still none)
431 * and not in the radix tree.
433 if (!(dst_vma
->vm_flags
& VM_SHARED
)) {
435 err
= mcopy_atomic_pte(dst_mm
, dst_pmd
, dst_vma
,
436 dst_addr
, src_addr
, page
,
439 err
= mfill_zeropage_pte(dst_mm
, dst_pmd
,
442 VM_WARN_ON_ONCE(wp_copy
);
444 err
= shmem_mcopy_atomic_pte(dst_mm
, dst_pmd
,
448 err
= shmem_mfill_zeropage_pte(dst_mm
, dst_pmd
,
455 static __always_inline ssize_t
__mcopy_atomic(struct mm_struct
*dst_mm
,
456 unsigned long dst_start
,
457 unsigned long src_start
,
459 enum mcopy_atomic_mode mcopy_mode
,
463 struct vm_area_struct
*dst_vma
;
466 unsigned long src_addr
, dst_addr
;
470 bool zeropage
= (mcopy_mode
== MCOPY_ATOMIC_ZEROPAGE
);
473 * Sanitize the command parameters:
475 BUG_ON(dst_start
& ~PAGE_MASK
);
476 BUG_ON(len
& ~PAGE_MASK
);
478 /* Does the address range wrap, or is the span zero-sized? */
479 BUG_ON(src_start
+ len
<= src_start
);
480 BUG_ON(dst_start
+ len
<= dst_start
);
482 src_addr
= src_start
;
483 dst_addr
= dst_start
;
487 mmap_read_lock(dst_mm
);
490 * If memory mappings are changing because of non-cooperative
491 * operation (e.g. mremap) running in parallel, bail out and
492 * request the user to retry later
495 if (mmap_changing
&& READ_ONCE(*mmap_changing
))
499 * Make sure the vma is not shared, that the dst range is
500 * both valid and fully within a single existing vma.
503 dst_vma
= find_dst_vma(dst_mm
, dst_start
, len
);
509 * shmem_zero_setup is invoked in mmap for MAP_ANONYMOUS|MAP_SHARED but
510 * it will overwrite vm_ops, so vma_is_anonymous must return false.
512 if (WARN_ON_ONCE(vma_is_anonymous(dst_vma
) &&
513 dst_vma
->vm_flags
& VM_SHARED
))
517 * validate 'mode' now that we know the dst_vma: don't allow
518 * a wrprotect copy if the userfaultfd didn't register as WP.
520 wp_copy
= mode
& UFFDIO_COPY_MODE_WP
;
521 if (wp_copy
&& !(dst_vma
->vm_flags
& VM_UFFD_WP
))
525 * If this is a HUGETLB vma, pass off to appropriate routine
527 if (is_vm_hugetlb_page(dst_vma
))
528 return __mcopy_atomic_hugetlb(dst_mm
, dst_vma
, dst_start
,
529 src_start
, len
, mcopy_mode
);
531 if (!vma_is_anonymous(dst_vma
) && !vma_is_shmem(dst_vma
))
533 if (mcopy_mode
== MCOPY_ATOMIC_CONTINUE
)
537 * Ensure the dst_vma has a anon_vma or this page
538 * would get a NULL anon_vma when moved in the
542 if (!(dst_vma
->vm_flags
& VM_SHARED
) &&
543 unlikely(anon_vma_prepare(dst_vma
)))
546 while (src_addr
< src_start
+ len
) {
549 BUG_ON(dst_addr
>= dst_start
+ len
);
551 dst_pmd
= mm_alloc_pmd(dst_mm
, dst_addr
);
552 if (unlikely(!dst_pmd
)) {
557 dst_pmdval
= pmd_read_atomic(dst_pmd
);
559 * If the dst_pmd is mapped as THP don't
560 * override it and just be strict.
562 if (unlikely(pmd_trans_huge(dst_pmdval
))) {
566 if (unlikely(pmd_none(dst_pmdval
)) &&
567 unlikely(__pte_alloc(dst_mm
, dst_pmd
))) {
571 /* If an huge pmd materialized from under us fail */
572 if (unlikely(pmd_trans_huge(*dst_pmd
))) {
577 BUG_ON(pmd_none(*dst_pmd
));
578 BUG_ON(pmd_trans_huge(*dst_pmd
));
580 err
= mfill_atomic_pte(dst_mm
, dst_pmd
, dst_vma
, dst_addr
,
581 src_addr
, &page
, zeropage
, wp_copy
);
584 if (unlikely(err
== -ENOENT
)) {
587 mmap_read_unlock(dst_mm
);
590 page_kaddr
= kmap(page
);
591 err
= copy_from_user(page_kaddr
,
592 (const void __user
*) src_addr
,
604 dst_addr
+= PAGE_SIZE
;
605 src_addr
+= PAGE_SIZE
;
608 if (fatal_signal_pending(current
))
616 mmap_read_unlock(dst_mm
);
622 BUG_ON(!copied
&& !err
);
623 return copied
? copied
: err
;
626 ssize_t
mcopy_atomic(struct mm_struct
*dst_mm
, unsigned long dst_start
,
627 unsigned long src_start
, unsigned long len
,
628 bool *mmap_changing
, __u64 mode
)
630 return __mcopy_atomic(dst_mm
, dst_start
, src_start
, len
,
631 MCOPY_ATOMIC_NORMAL
, mmap_changing
, mode
);
634 ssize_t
mfill_zeropage(struct mm_struct
*dst_mm
, unsigned long start
,
635 unsigned long len
, bool *mmap_changing
)
637 return __mcopy_atomic(dst_mm
, start
, 0, len
, MCOPY_ATOMIC_ZEROPAGE
,
641 ssize_t
mcopy_continue(struct mm_struct
*dst_mm
, unsigned long start
,
642 unsigned long len
, bool *mmap_changing
)
644 return __mcopy_atomic(dst_mm
, start
, 0, len
, MCOPY_ATOMIC_CONTINUE
,
648 int mwriteprotect_range(struct mm_struct
*dst_mm
, unsigned long start
,
649 unsigned long len
, bool enable_wp
, bool *mmap_changing
)
651 struct vm_area_struct
*dst_vma
;
656 * Sanitize the command parameters:
658 BUG_ON(start
& ~PAGE_MASK
);
659 BUG_ON(len
& ~PAGE_MASK
);
661 /* Does the address range wrap, or is the span zero-sized? */
662 BUG_ON(start
+ len
<= start
);
664 mmap_read_lock(dst_mm
);
667 * If memory mappings are changing because of non-cooperative
668 * operation (e.g. mremap) running in parallel, bail out and
669 * request the user to retry later
672 if (mmap_changing
&& READ_ONCE(*mmap_changing
))
676 dst_vma
= find_dst_vma(dst_mm
, start
, len
);
678 * Make sure the vma is not shared, that the dst range is
679 * both valid and fully within a single existing vma.
681 if (!dst_vma
|| (dst_vma
->vm_flags
& VM_SHARED
))
683 if (!userfaultfd_wp(dst_vma
))
685 if (!vma_is_anonymous(dst_vma
))
689 newprot
= vm_get_page_prot(dst_vma
->vm_flags
& ~(VM_WRITE
));
691 newprot
= vm_get_page_prot(dst_vma
->vm_flags
);
693 change_protection(dst_vma
, start
, start
+ len
, newprot
,
694 enable_wp
? MM_CP_UFFD_WP
: MM_CP_UFFD_WP_RESOLVE
);
698 mmap_read_unlock(dst_mm
);