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 int mcopy_atomic_pte(struct mm_struct *dst_mm,
23 struct vm_area_struct *dst_vma,
24 unsigned long dst_addr,
25 unsigned long src_addr,
28 struct mem_cgroup *memcg;
29 pte_t _dst_pte, *dst_pte;
34 pgoff_t offset, max_off;
39 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, dst_vma, dst_addr);
43 page_kaddr = kmap_atomic(page);
44 ret = copy_from_user(page_kaddr,
45 (const void __user *) src_addr,
47 kunmap_atomic(page_kaddr);
49 /* fallback to copy_from_user outside mmap_sem */
53 /* don't free the page */
57 flush_dcache_page(page);
64 * The memory barrier inside __SetPageUptodate makes sure that
65 * preceeding stores to the page contents become visible before
66 * the set_pte_at() write.
68 __SetPageUptodate(page);
71 if (mem_cgroup_try_charge(page, dst_mm, GFP_KERNEL, &memcg, false))
74 _dst_pte = mk_pte(page, dst_vma->vm_page_prot);
75 if (dst_vma->vm_flags & VM_WRITE)
76 _dst_pte = pte_mkwrite(pte_mkdirty(_dst_pte));
78 dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
79 if (dst_vma->vm_file) {
80 /* the shmem MAP_PRIVATE case requires checking the i_size */
81 inode = dst_vma->vm_file->f_inode;
82 offset = linear_page_index(dst_vma, dst_addr);
83 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
85 if (unlikely(offset >= max_off))
86 goto out_release_uncharge_unlock;
89 if (!pte_none(*dst_pte))
90 goto out_release_uncharge_unlock;
92 inc_mm_counter(dst_mm, MM_ANONPAGES);
93 page_add_new_anon_rmap(page, dst_vma, dst_addr, false);
94 mem_cgroup_commit_charge(page, memcg, false, false);
95 lru_cache_add_active_or_unevictable(page, dst_vma);
97 set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
99 /* No need to invalidate - it was non-present before */
100 update_mmu_cache(dst_vma, dst_addr, dst_pte);
102 pte_unmap_unlock(dst_pte, ptl);
106 out_release_uncharge_unlock:
107 pte_unmap_unlock(dst_pte, ptl);
108 mem_cgroup_cancel_charge(page, memcg, false);
114 static int mfill_zeropage_pte(struct mm_struct *dst_mm,
116 struct vm_area_struct *dst_vma,
117 unsigned long dst_addr)
119 pte_t _dst_pte, *dst_pte;
122 pgoff_t offset, max_off;
125 _dst_pte = pte_mkspecial(pfn_pte(my_zero_pfn(dst_addr),
126 dst_vma->vm_page_prot));
127 dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
128 if (dst_vma->vm_file) {
129 /* the shmem MAP_PRIVATE case requires checking the i_size */
130 inode = dst_vma->vm_file->f_inode;
131 offset = linear_page_index(dst_vma, dst_addr);
132 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
134 if (unlikely(offset >= max_off))
138 if (!pte_none(*dst_pte))
140 set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
141 /* No need to invalidate - it was non-present before */
142 update_mmu_cache(dst_vma, dst_addr, dst_pte);
145 pte_unmap_unlock(dst_pte, ptl);
149 static pmd_t *mm_alloc_pmd(struct mm_struct *mm, unsigned long address)
155 pgd = pgd_offset(mm, address);
156 p4d = p4d_alloc(mm, pgd, address);
159 pud = pud_alloc(mm, p4d, address);
163 * Note that we didn't run this because the pmd was
164 * missing, the *pmd may be already established and in
165 * turn it may also be a trans_huge_pmd.
167 return pmd_alloc(mm, pud, address);
170 #ifdef CONFIG_HUGETLB_PAGE
172 * __mcopy_atomic processing for HUGETLB vmas. Note that this routine is
173 * called with mmap_sem held, it will release mmap_sem before returning.
175 static __always_inline ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
176 struct vm_area_struct *dst_vma,
177 unsigned long dst_start,
178 unsigned long src_start,
183 int vm_alloc_shared = dst_vma->vm_flags & VM_SHARED;
184 int vm_shared = dst_vma->vm_flags & VM_SHARED;
187 unsigned long src_addr, dst_addr;
191 unsigned long vma_hpagesize;
194 struct address_space *mapping;
197 * There is no default zero huge page for all huge page sizes as
198 * supported by hugetlb. A PMD_SIZE huge pages may exist as used
199 * by THP. Since we can not reliably insert a zero page, this
200 * feature is not supported.
203 up_read(&dst_mm->mmap_sem);
207 src_addr = src_start;
208 dst_addr = dst_start;
211 vma_hpagesize = vma_kernel_pagesize(dst_vma);
214 * Validate alignment based on huge page size
217 if (dst_start & (vma_hpagesize - 1) || len & (vma_hpagesize - 1))
222 * On routine entry dst_vma is set. If we had to drop mmap_sem and
223 * retry, dst_vma will be set to NULL and we must lookup again.
227 dst_vma = find_vma(dst_mm, dst_start);
228 if (!dst_vma || !is_vm_hugetlb_page(dst_vma))
231 * Check the vma is registered in uffd, this is
232 * required to enforce the VM_MAYWRITE check done at
233 * uffd registration time.
235 if (!dst_vma->vm_userfaultfd_ctx.ctx)
238 if (dst_start < dst_vma->vm_start ||
239 dst_start + len > dst_vma->vm_end)
243 if (vma_hpagesize != vma_kernel_pagesize(dst_vma))
246 vm_shared = dst_vma->vm_flags & VM_SHARED;
249 if (WARN_ON(dst_addr & (vma_hpagesize - 1) ||
250 (len - copied) & (vma_hpagesize - 1)))
254 * If not shared, ensure the dst_vma has a anon_vma.
258 if (unlikely(anon_vma_prepare(dst_vma)))
262 h = hstate_vma(dst_vma);
264 while (src_addr < src_start + len) {
267 BUG_ON(dst_addr >= dst_start + len);
268 VM_BUG_ON(dst_addr & ~huge_page_mask(h));
271 * Serialize via hugetlb_fault_mutex
273 idx = linear_page_index(dst_vma, dst_addr);
274 mapping = dst_vma->vm_file->f_mapping;
275 hash = hugetlb_fault_mutex_hash(h, mapping, idx);
276 mutex_lock(&hugetlb_fault_mutex_table[hash]);
279 dst_pte = huge_pte_alloc(dst_mm, dst_addr, huge_page_size(h));
281 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
286 dst_pteval = huge_ptep_get(dst_pte);
287 if (!huge_pte_none(dst_pteval)) {
288 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
292 err = hugetlb_mcopy_atomic_pte(dst_mm, dst_pte, dst_vma,
293 dst_addr, src_addr, &page);
295 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
296 vm_alloc_shared = vm_shared;
300 if (unlikely(err == -ENOENT)) {
301 up_read(&dst_mm->mmap_sem);
304 err = copy_huge_page_from_user(page,
305 (const void __user *)src_addr,
306 pages_per_huge_page(h), true);
311 down_read(&dst_mm->mmap_sem);
313 * If memory mappings are changing because of non-cooperative
314 * operation (e.g. mremap) running in parallel, bail out and
315 * request the user to retry later
317 if (mmap_changing && READ_ONCE(*mmap_changing)) {
328 dst_addr += vma_hpagesize;
329 src_addr += vma_hpagesize;
330 copied += vma_hpagesize;
332 if (fatal_signal_pending(current))
340 up_read(&dst_mm->mmap_sem);
344 * We encountered an error and are about to free a newly
345 * allocated huge page.
347 * Reservation handling is very subtle, and is different for
348 * private and shared mappings. See the routine
349 * restore_reserve_on_error for details. Unfortunately, we
350 * can not call restore_reserve_on_error now as it would
351 * require holding mmap_sem.
353 * If a reservation for the page existed in the reservation
354 * map of a private mapping, the map was modified to indicate
355 * the reservation was consumed when the page was allocated.
356 * We clear the PagePrivate flag now so that the global
357 * reserve count will not be incremented in free_huge_page.
358 * The reservation map will still indicate the reservation
359 * was consumed and possibly prevent later page allocation.
360 * This is better than leaking a global reservation. If no
361 * reservation existed, it is still safe to clear PagePrivate
362 * as no adjustments to reservation counts were made during
365 * The reservation map for shared mappings indicates which
366 * pages have reservations. When a huge page is allocated
367 * for an address with a reservation, no change is made to
368 * the reserve map. In this case PagePrivate will be set
369 * to indicate that the global reservation count should be
370 * incremented when the page is freed. This is the desired
371 * behavior. However, when a huge page is allocated for an
372 * address without a reservation a reservation entry is added
373 * to the reservation map, and PagePrivate will not be set.
374 * When the page is freed, the global reserve count will NOT
375 * be incremented and it will appear as though we have leaked
376 * reserved page. In this case, set PagePrivate so that the
377 * global reserve count will be incremented to match the
378 * reservation map entry which was created.
380 * Note that vm_alloc_shared is based on the flags of the vma
381 * for which the page was originally allocated. dst_vma could
382 * be different or NULL on error.
385 SetPagePrivate(page);
387 ClearPagePrivate(page);
392 BUG_ON(!copied && !err);
393 return copied ? copied : err;
395 #else /* !CONFIG_HUGETLB_PAGE */
396 /* fail at build time if gcc attempts to use this */
397 extern ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
398 struct vm_area_struct *dst_vma,
399 unsigned long dst_start,
400 unsigned long src_start,
404 #endif /* CONFIG_HUGETLB_PAGE */
406 static __always_inline ssize_t mfill_atomic_pte(struct mm_struct *dst_mm,
408 struct vm_area_struct *dst_vma,
409 unsigned long dst_addr,
410 unsigned long src_addr,
417 * The normal page fault path for a shmem will invoke the
418 * fault, fill the hole in the file and COW it right away. The
419 * result generates plain anonymous memory. So when we are
420 * asked to fill an hole in a MAP_PRIVATE shmem mapping, we'll
421 * generate anonymous memory directly without actually filling
422 * the hole. For the MAP_PRIVATE case the robustness check
423 * only happens in the pagetable (to verify it's still none)
424 * and not in the radix tree.
426 if (!(dst_vma->vm_flags & VM_SHARED)) {
428 err = mcopy_atomic_pte(dst_mm, dst_pmd, dst_vma,
429 dst_addr, src_addr, page);
431 err = mfill_zeropage_pte(dst_mm, dst_pmd,
435 err = shmem_mcopy_atomic_pte(dst_mm, dst_pmd,
439 err = shmem_mfill_zeropage_pte(dst_mm, dst_pmd,
446 static __always_inline ssize_t __mcopy_atomic(struct mm_struct *dst_mm,
447 unsigned long dst_start,
448 unsigned long src_start,
453 struct vm_area_struct *dst_vma;
456 unsigned long src_addr, dst_addr;
461 * Sanitize the command parameters:
463 BUG_ON(dst_start & ~PAGE_MASK);
464 BUG_ON(len & ~PAGE_MASK);
466 /* Does the address range wrap, or is the span zero-sized? */
467 BUG_ON(src_start + len <= src_start);
468 BUG_ON(dst_start + len <= dst_start);
470 src_addr = src_start;
471 dst_addr = dst_start;
475 down_read(&dst_mm->mmap_sem);
478 * If memory mappings are changing because of non-cooperative
479 * operation (e.g. mremap) running in parallel, bail out and
480 * request the user to retry later
483 if (mmap_changing && READ_ONCE(*mmap_changing))
487 * Make sure the vma is not shared, that the dst range is
488 * both valid and fully within a single existing vma.
491 dst_vma = find_vma(dst_mm, dst_start);
495 * Check the vma is registered in uffd, this is required to
496 * enforce the VM_MAYWRITE check done at uffd registration
499 if (!dst_vma->vm_userfaultfd_ctx.ctx)
502 if (dst_start < dst_vma->vm_start ||
503 dst_start + len > dst_vma->vm_end)
508 * shmem_zero_setup is invoked in mmap for MAP_ANONYMOUS|MAP_SHARED but
509 * it will overwrite vm_ops, so vma_is_anonymous must return false.
511 if (WARN_ON_ONCE(vma_is_anonymous(dst_vma) &&
512 dst_vma->vm_flags & VM_SHARED))
516 * If this is a HUGETLB vma, pass off to appropriate routine
518 if (is_vm_hugetlb_page(dst_vma))
519 return __mcopy_atomic_hugetlb(dst_mm, dst_vma, dst_start,
520 src_start, len, mmap_changing,
523 if (!vma_is_anonymous(dst_vma) && !vma_is_shmem(dst_vma))
527 * Ensure the dst_vma has a anon_vma or this page
528 * would get a NULL anon_vma when moved in the
532 if (!(dst_vma->vm_flags & VM_SHARED) &&
533 unlikely(anon_vma_prepare(dst_vma)))
536 while (src_addr < src_start + len) {
539 BUG_ON(dst_addr >= dst_start + len);
541 dst_pmd = mm_alloc_pmd(dst_mm, dst_addr);
542 if (unlikely(!dst_pmd)) {
547 dst_pmdval = pmd_read_atomic(dst_pmd);
549 * If the dst_pmd is mapped as THP don't
550 * override it and just be strict.
552 if (unlikely(pmd_trans_huge(dst_pmdval))) {
556 if (unlikely(pmd_none(dst_pmdval)) &&
557 unlikely(__pte_alloc(dst_mm, dst_pmd))) {
561 /* If an huge pmd materialized from under us fail */
562 if (unlikely(pmd_trans_huge(*dst_pmd))) {
567 BUG_ON(pmd_none(*dst_pmd));
568 BUG_ON(pmd_trans_huge(*dst_pmd));
570 err = mfill_atomic_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
571 src_addr, &page, zeropage);
574 if (unlikely(err == -ENOENT)) {
577 up_read(&dst_mm->mmap_sem);
580 page_kaddr = kmap(page);
581 err = copy_from_user(page_kaddr,
582 (const void __user *) src_addr,
589 flush_dcache_page(page);
595 dst_addr += PAGE_SIZE;
596 src_addr += PAGE_SIZE;
599 if (fatal_signal_pending(current))
607 up_read(&dst_mm->mmap_sem);
613 BUG_ON(!copied && !err);
614 return copied ? copied : err;
617 ssize_t mcopy_atomic(struct mm_struct *dst_mm, unsigned long dst_start,
618 unsigned long src_start, unsigned long len,
621 return __mcopy_atomic(dst_mm, dst_start, src_start, len, false,
625 ssize_t mfill_zeropage(struct mm_struct *dst_mm, unsigned long start,
626 unsigned long len, bool *mmap_changing)
628 return __mcopy_atomic(dst_mm, start, 0, len, true, mmap_changing);