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 struct mem_cgroup *memcg;
60 pte_t _dst_pte, *dst_pte;
65 pgoff_t offset, max_off;
70 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, dst_vma, dst_addr);
74 page_kaddr = kmap_atomic(page);
75 ret = copy_from_user(page_kaddr,
76 (const void __user *) src_addr,
78 kunmap_atomic(page_kaddr);
80 /* fallback to copy_from_user outside mmap_sem */
84 /* don't free the page */
93 * The memory barrier inside __SetPageUptodate makes sure that
94 * preceding stores to the page contents become visible before
95 * the set_pte_at() write.
97 __SetPageUptodate(page);
100 if (mem_cgroup_try_charge(page, dst_mm, GFP_KERNEL, &memcg, false))
103 _dst_pte = pte_mkdirty(mk_pte(page, dst_vma->vm_page_prot));
104 if ((dst_vma->vm_flags & VM_WRITE) && !wp_copy)
105 _dst_pte = pte_mkwrite(_dst_pte);
107 dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
108 if (dst_vma->vm_file) {
109 /* the shmem MAP_PRIVATE case requires checking the i_size */
110 inode = dst_vma->vm_file->f_inode;
111 offset = linear_page_index(dst_vma, dst_addr);
112 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
114 if (unlikely(offset >= max_off))
115 goto out_release_uncharge_unlock;
118 if (!pte_none(*dst_pte))
119 goto out_release_uncharge_unlock;
121 inc_mm_counter(dst_mm, MM_ANONPAGES);
122 page_add_new_anon_rmap(page, dst_vma, dst_addr, false);
123 mem_cgroup_commit_charge(page, memcg, false, false);
124 lru_cache_add_active_or_unevictable(page, dst_vma);
126 set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
128 /* No need to invalidate - it was non-present before */
129 update_mmu_cache(dst_vma, dst_addr, dst_pte);
131 pte_unmap_unlock(dst_pte, ptl);
135 out_release_uncharge_unlock:
136 pte_unmap_unlock(dst_pte, ptl);
137 mem_cgroup_cancel_charge(page, memcg, false);
143 static int mfill_zeropage_pte(struct mm_struct *dst_mm,
145 struct vm_area_struct *dst_vma,
146 unsigned long dst_addr)
148 pte_t _dst_pte, *dst_pte;
151 pgoff_t offset, max_off;
154 _dst_pte = pte_mkspecial(pfn_pte(my_zero_pfn(dst_addr),
155 dst_vma->vm_page_prot));
156 dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
157 if (dst_vma->vm_file) {
158 /* the shmem MAP_PRIVATE case requires checking the i_size */
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);
163 if (unlikely(offset >= max_off))
167 if (!pte_none(*dst_pte))
169 set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
170 /* No need to invalidate - it was non-present before */
171 update_mmu_cache(dst_vma, dst_addr, dst_pte);
174 pte_unmap_unlock(dst_pte, ptl);
178 static pmd_t *mm_alloc_pmd(struct mm_struct *mm, unsigned long address)
184 pgd = pgd_offset(mm, address);
185 p4d = p4d_alloc(mm, pgd, address);
188 pud = pud_alloc(mm, p4d, address);
192 * Note that we didn't run this because the pmd was
193 * missing, the *pmd may be already established and in
194 * turn it may also be a trans_huge_pmd.
196 return pmd_alloc(mm, pud, address);
199 #ifdef CONFIG_HUGETLB_PAGE
201 * __mcopy_atomic processing for HUGETLB vmas. Note that this routine is
202 * called with mmap_sem held, it will release mmap_sem before returning.
204 static __always_inline ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
205 struct vm_area_struct *dst_vma,
206 unsigned long dst_start,
207 unsigned long src_start,
211 int vm_alloc_shared = dst_vma->vm_flags & VM_SHARED;
212 int vm_shared = dst_vma->vm_flags & VM_SHARED;
215 unsigned long src_addr, dst_addr;
218 unsigned long vma_hpagesize;
221 struct address_space *mapping;
224 * There is no default zero huge page for all huge page sizes as
225 * supported by hugetlb. A PMD_SIZE huge pages may exist as used
226 * by THP. Since we can not reliably insert a zero page, this
227 * feature is not supported.
230 up_read(&dst_mm->mmap_sem);
234 src_addr = src_start;
235 dst_addr = dst_start;
238 vma_hpagesize = vma_kernel_pagesize(dst_vma);
241 * Validate alignment based on huge page size
244 if (dst_start & (vma_hpagesize - 1) || len & (vma_hpagesize - 1))
249 * On routine entry dst_vma is set. If we had to drop mmap_sem and
250 * retry, dst_vma will be set to NULL and we must lookup again.
254 dst_vma = find_dst_vma(dst_mm, dst_start, len);
255 if (!dst_vma || !is_vm_hugetlb_page(dst_vma))
259 if (vma_hpagesize != vma_kernel_pagesize(dst_vma))
262 vm_shared = dst_vma->vm_flags & VM_SHARED;
266 * If not shared, ensure the dst_vma has a anon_vma.
270 if (unlikely(anon_vma_prepare(dst_vma)))
274 while (src_addr < src_start + len) {
277 BUG_ON(dst_addr >= dst_start + len);
280 * Serialize via i_mmap_rwsem and hugetlb_fault_mutex.
281 * i_mmap_rwsem ensures the dst_pte remains valid even
282 * in the case of shared pmds. fault mutex prevents
283 * races with other faulting threads.
285 mapping = dst_vma->vm_file->f_mapping;
286 i_mmap_lock_read(mapping);
287 idx = linear_page_index(dst_vma, dst_addr);
288 hash = hugetlb_fault_mutex_hash(mapping, idx);
289 mutex_lock(&hugetlb_fault_mutex_table[hash]);
292 dst_pte = huge_pte_alloc(dst_mm, dst_addr, vma_hpagesize);
294 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
295 i_mmap_unlock_read(mapping);
300 dst_pteval = huge_ptep_get(dst_pte);
301 if (!huge_pte_none(dst_pteval)) {
302 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
303 i_mmap_unlock_read(mapping);
307 err = hugetlb_mcopy_atomic_pte(dst_mm, dst_pte, dst_vma,
308 dst_addr, src_addr, &page);
310 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
311 i_mmap_unlock_read(mapping);
312 vm_alloc_shared = vm_shared;
316 if (unlikely(err == -ENOENT)) {
317 up_read(&dst_mm->mmap_sem);
320 err = copy_huge_page_from_user(page,
321 (const void __user *)src_addr,
322 vma_hpagesize / PAGE_SIZE,
328 down_read(&dst_mm->mmap_sem);
336 dst_addr += vma_hpagesize;
337 src_addr += vma_hpagesize;
338 copied += vma_hpagesize;
340 if (fatal_signal_pending(current))
348 up_read(&dst_mm->mmap_sem);
352 * We encountered an error and are about to free a newly
353 * allocated huge page.
355 * Reservation handling is very subtle, and is different for
356 * private and shared mappings. See the routine
357 * restore_reserve_on_error for details. Unfortunately, we
358 * can not call restore_reserve_on_error now as it would
359 * require holding mmap_sem.
361 * If a reservation for the page existed in the reservation
362 * map of a private mapping, the map was modified to indicate
363 * the reservation was consumed when the page was allocated.
364 * We clear the PagePrivate flag now so that the global
365 * reserve count will not be incremented in free_huge_page.
366 * The reservation map will still indicate the reservation
367 * was consumed and possibly prevent later page allocation.
368 * This is better than leaking a global reservation. If no
369 * reservation existed, it is still safe to clear PagePrivate
370 * as no adjustments to reservation counts were made during
373 * The reservation map for shared mappings indicates which
374 * pages have reservations. When a huge page is allocated
375 * for an address with a reservation, no change is made to
376 * the reserve map. In this case PagePrivate will be set
377 * to indicate that the global reservation count should be
378 * incremented when the page is freed. This is the desired
379 * behavior. However, when a huge page is allocated for an
380 * address without a reservation a reservation entry is added
381 * to the reservation map, and PagePrivate will not be set.
382 * When the page is freed, the global reserve count will NOT
383 * be incremented and it will appear as though we have leaked
384 * reserved page. In this case, set PagePrivate so that the
385 * global reserve count will be incremented to match the
386 * reservation map entry which was created.
388 * Note that vm_alloc_shared is based on the flags of the vma
389 * for which the page was originally allocated. dst_vma could
390 * be different or NULL on error.
393 SetPagePrivate(page);
395 ClearPagePrivate(page);
400 BUG_ON(!copied && !err);
401 return copied ? copied : err;
403 #else /* !CONFIG_HUGETLB_PAGE */
404 /* fail at build time if gcc attempts to use this */
405 extern ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
406 struct vm_area_struct *dst_vma,
407 unsigned long dst_start,
408 unsigned long src_start,
411 #endif /* CONFIG_HUGETLB_PAGE */
413 static __always_inline ssize_t mfill_atomic_pte(struct mm_struct *dst_mm,
415 struct vm_area_struct *dst_vma,
416 unsigned long dst_addr,
417 unsigned long src_addr,
425 * The normal page fault path for a shmem will invoke the
426 * fault, fill the hole in the file and COW it right away. The
427 * result generates plain anonymous memory. So when we are
428 * asked to fill an hole in a MAP_PRIVATE shmem mapping, we'll
429 * generate anonymous memory directly without actually filling
430 * the hole. For the MAP_PRIVATE case the robustness check
431 * only happens in the pagetable (to verify it's still none)
432 * and not in the radix tree.
434 if (!(dst_vma->vm_flags & VM_SHARED)) {
436 err = mcopy_atomic_pte(dst_mm, dst_pmd, dst_vma,
437 dst_addr, src_addr, page,
440 err = mfill_zeropage_pte(dst_mm, dst_pmd,
443 VM_WARN_ON_ONCE(wp_copy);
445 err = shmem_mcopy_atomic_pte(dst_mm, dst_pmd,
449 err = shmem_mfill_zeropage_pte(dst_mm, dst_pmd,
456 static __always_inline ssize_t __mcopy_atomic(struct mm_struct *dst_mm,
457 unsigned long dst_start,
458 unsigned long src_start,
464 struct vm_area_struct *dst_vma;
467 unsigned long src_addr, dst_addr;
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 down_read(&dst_mm->mmap_sem);
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, zeropage);
531 if (!vma_is_anonymous(dst_vma) && !vma_is_shmem(dst_vma))
535 * Ensure the dst_vma has a anon_vma or this page
536 * would get a NULL anon_vma when moved in the
540 if (!(dst_vma->vm_flags & VM_SHARED) &&
541 unlikely(anon_vma_prepare(dst_vma)))
544 while (src_addr < src_start + len) {
547 BUG_ON(dst_addr >= dst_start + len);
549 dst_pmd = mm_alloc_pmd(dst_mm, dst_addr);
550 if (unlikely(!dst_pmd)) {
555 dst_pmdval = pmd_read_atomic(dst_pmd);
557 * If the dst_pmd is mapped as THP don't
558 * override it and just be strict.
560 if (unlikely(pmd_trans_huge(dst_pmdval))) {
564 if (unlikely(pmd_none(dst_pmdval)) &&
565 unlikely(__pte_alloc(dst_mm, dst_pmd))) {
569 /* If an huge pmd materialized from under us fail */
570 if (unlikely(pmd_trans_huge(*dst_pmd))) {
575 BUG_ON(pmd_none(*dst_pmd));
576 BUG_ON(pmd_trans_huge(*dst_pmd));
578 err = mfill_atomic_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
579 src_addr, &page, zeropage, wp_copy);
582 if (unlikely(err == -ENOENT)) {
585 up_read(&dst_mm->mmap_sem);
588 page_kaddr = kmap(page);
589 err = copy_from_user(page_kaddr,
590 (const void __user *) src_addr,
602 dst_addr += PAGE_SIZE;
603 src_addr += PAGE_SIZE;
606 if (fatal_signal_pending(current))
614 up_read(&dst_mm->mmap_sem);
620 BUG_ON(!copied && !err);
621 return copied ? copied : err;
624 ssize_t mcopy_atomic(struct mm_struct *dst_mm, unsigned long dst_start,
625 unsigned long src_start, unsigned long len,
626 bool *mmap_changing, __u64 mode)
628 return __mcopy_atomic(dst_mm, dst_start, src_start, len, false,
629 mmap_changing, mode);
632 ssize_t mfill_zeropage(struct mm_struct *dst_mm, unsigned long start,
633 unsigned long len, bool *mmap_changing)
635 return __mcopy_atomic(dst_mm, start, 0, len, true, mmap_changing, 0);