4 * Copyright (C) 2015 Red Hat, Inc.
6 * This work is licensed under the terms of the GNU GPL, version 2. See
7 * the COPYING file in the top-level directory.
11 #include <linux/sched/signal.h>
12 #include <linux/pagemap.h>
13 #include <linux/rmap.h>
14 #include <linux/swap.h>
15 #include <linux/swapops.h>
16 #include <linux/userfaultfd_k.h>
17 #include <linux/mmu_notifier.h>
18 #include <linux/hugetlb.h>
19 #include <linux/pagemap.h>
20 #include <linux/shmem_fs.h>
21 #include <asm/tlbflush.h>
24 static int mcopy_atomic_pte(struct mm_struct *dst_mm,
26 struct vm_area_struct *dst_vma,
27 unsigned long dst_addr,
28 unsigned long src_addr,
31 struct mem_cgroup *memcg;
32 pte_t _dst_pte, *dst_pte;
37 pgoff_t offset, max_off;
42 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, dst_vma, dst_addr);
46 page_kaddr = kmap_atomic(page);
47 ret = copy_from_user(page_kaddr,
48 (const void __user *) src_addr,
50 kunmap_atomic(page_kaddr);
52 /* fallback to copy_from_user outside mmap_sem */
56 /* don't free the page */
60 flush_dcache_page(page);
67 * The memory barrier inside __SetPageUptodate makes sure that
68 * preceeding stores to the page contents become visible before
69 * the set_pte_at() write.
71 __SetPageUptodate(page);
74 if (mem_cgroup_try_charge(page, dst_mm, GFP_KERNEL, &memcg, false))
77 _dst_pte = mk_pte(page, dst_vma->vm_page_prot);
78 if (dst_vma->vm_flags & VM_WRITE)
79 _dst_pte = pte_mkwrite(pte_mkdirty(_dst_pte));
81 dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
82 if (dst_vma->vm_file) {
83 /* the shmem MAP_PRIVATE case requires checking the i_size */
84 inode = dst_vma->vm_file->f_inode;
85 offset = linear_page_index(dst_vma, dst_addr);
86 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
88 if (unlikely(offset >= max_off))
89 goto out_release_uncharge_unlock;
92 if (!pte_none(*dst_pte))
93 goto out_release_uncharge_unlock;
95 inc_mm_counter(dst_mm, MM_ANONPAGES);
96 page_add_new_anon_rmap(page, dst_vma, dst_addr, false);
97 mem_cgroup_commit_charge(page, memcg, false, false);
98 lru_cache_add_active_or_unevictable(page, dst_vma);
100 set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
102 /* No need to invalidate - it was non-present before */
103 update_mmu_cache(dst_vma, dst_addr, dst_pte);
105 pte_unmap_unlock(dst_pte, ptl);
109 out_release_uncharge_unlock:
110 pte_unmap_unlock(dst_pte, ptl);
111 mem_cgroup_cancel_charge(page, memcg, false);
117 static int mfill_zeropage_pte(struct mm_struct *dst_mm,
119 struct vm_area_struct *dst_vma,
120 unsigned long dst_addr)
122 pte_t _dst_pte, *dst_pte;
125 pgoff_t offset, max_off;
128 _dst_pte = pte_mkspecial(pfn_pte(my_zero_pfn(dst_addr),
129 dst_vma->vm_page_prot));
130 dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
131 if (dst_vma->vm_file) {
132 /* the shmem MAP_PRIVATE case requires checking the i_size */
133 inode = dst_vma->vm_file->f_inode;
134 offset = linear_page_index(dst_vma, dst_addr);
135 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
137 if (unlikely(offset >= max_off))
141 if (!pte_none(*dst_pte))
143 set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
144 /* No need to invalidate - it was non-present before */
145 update_mmu_cache(dst_vma, dst_addr, dst_pte);
148 pte_unmap_unlock(dst_pte, ptl);
152 static pmd_t *mm_alloc_pmd(struct mm_struct *mm, unsigned long address)
158 pgd = pgd_offset(mm, address);
159 p4d = p4d_alloc(mm, pgd, address);
162 pud = pud_alloc(mm, p4d, address);
166 * Note that we didn't run this because the pmd was
167 * missing, the *pmd may be already established and in
168 * turn it may also be a trans_huge_pmd.
170 return pmd_alloc(mm, pud, address);
173 #ifdef CONFIG_HUGETLB_PAGE
175 * __mcopy_atomic processing for HUGETLB vmas. Note that this routine is
176 * called with mmap_sem held, it will release mmap_sem before returning.
178 static __always_inline ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
179 struct vm_area_struct *dst_vma,
180 unsigned long dst_start,
181 unsigned long src_start,
185 int vm_alloc_shared = dst_vma->vm_flags & VM_SHARED;
186 int vm_shared = dst_vma->vm_flags & VM_SHARED;
189 unsigned long src_addr, dst_addr;
193 unsigned long vma_hpagesize;
196 struct address_space *mapping;
199 * There is no default zero huge page for all huge page sizes as
200 * supported by hugetlb. A PMD_SIZE huge pages may exist as used
201 * by THP. Since we can not reliably insert a zero page, this
202 * feature is not supported.
205 up_read(&dst_mm->mmap_sem);
209 src_addr = src_start;
210 dst_addr = dst_start;
213 vma_hpagesize = vma_kernel_pagesize(dst_vma);
216 * Validate alignment based on huge page size
219 if (dst_start & (vma_hpagesize - 1) || len & (vma_hpagesize - 1))
224 * On routine entry dst_vma is set. If we had to drop mmap_sem and
225 * retry, dst_vma will be set to NULL and we must lookup again.
229 dst_vma = find_vma(dst_mm, dst_start);
230 if (!dst_vma || !is_vm_hugetlb_page(dst_vma))
233 * Check the vma is registered in uffd, this is
234 * required to enforce the VM_MAYWRITE check done at
235 * uffd registration time.
237 if (!dst_vma->vm_userfaultfd_ctx.ctx)
240 if (dst_start < dst_vma->vm_start ||
241 dst_start + len > dst_vma->vm_end)
245 if (vma_hpagesize != vma_kernel_pagesize(dst_vma))
248 vm_shared = dst_vma->vm_flags & VM_SHARED;
251 if (WARN_ON(dst_addr & (vma_hpagesize - 1) ||
252 (len - copied) & (vma_hpagesize - 1)))
256 * If not shared, ensure the dst_vma has a anon_vma.
260 if (unlikely(anon_vma_prepare(dst_vma)))
264 h = hstate_vma(dst_vma);
266 while (src_addr < src_start + len) {
269 BUG_ON(dst_addr >= dst_start + len);
270 VM_BUG_ON(dst_addr & ~huge_page_mask(h));
273 * Serialize via hugetlb_fault_mutex
275 idx = linear_page_index(dst_vma, dst_addr);
276 mapping = dst_vma->vm_file->f_mapping;
277 hash = hugetlb_fault_mutex_hash(h, mapping, idx);
278 mutex_lock(&hugetlb_fault_mutex_table[hash]);
281 dst_pte = huge_pte_alloc(dst_mm, dst_addr, huge_page_size(h));
283 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
288 dst_pteval = huge_ptep_get(dst_pte);
289 if (!huge_pte_none(dst_pteval)) {
290 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
294 err = hugetlb_mcopy_atomic_pte(dst_mm, dst_pte, dst_vma,
295 dst_addr, src_addr, &page);
297 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
298 vm_alloc_shared = vm_shared;
302 if (unlikely(err == -ENOENT)) {
303 up_read(&dst_mm->mmap_sem);
306 err = copy_huge_page_from_user(page,
307 (const void __user *)src_addr,
308 pages_per_huge_page(h), true);
313 down_read(&dst_mm->mmap_sem);
321 dst_addr += vma_hpagesize;
322 src_addr += vma_hpagesize;
323 copied += vma_hpagesize;
325 if (fatal_signal_pending(current))
333 up_read(&dst_mm->mmap_sem);
337 * We encountered an error and are about to free a newly
338 * allocated huge page.
340 * Reservation handling is very subtle, and is different for
341 * private and shared mappings. See the routine
342 * restore_reserve_on_error for details. Unfortunately, we
343 * can not call restore_reserve_on_error now as it would
344 * require holding mmap_sem.
346 * If a reservation for the page existed in the reservation
347 * map of a private mapping, the map was modified to indicate
348 * the reservation was consumed when the page was allocated.
349 * We clear the PagePrivate flag now so that the global
350 * reserve count will not be incremented in free_huge_page.
351 * The reservation map will still indicate the reservation
352 * was consumed and possibly prevent later page allocation.
353 * This is better than leaking a global reservation. If no
354 * reservation existed, it is still safe to clear PagePrivate
355 * as no adjustments to reservation counts were made during
358 * The reservation map for shared mappings indicates which
359 * pages have reservations. When a huge page is allocated
360 * for an address with a reservation, no change is made to
361 * the reserve map. In this case PagePrivate will be set
362 * to indicate that the global reservation count should be
363 * incremented when the page is freed. This is the desired
364 * behavior. However, when a huge page is allocated for an
365 * address without a reservation a reservation entry is added
366 * to the reservation map, and PagePrivate will not be set.
367 * When the page is freed, the global reserve count will NOT
368 * be incremented and it will appear as though we have leaked
369 * reserved page. In this case, set PagePrivate so that the
370 * global reserve count will be incremented to match the
371 * reservation map entry which was created.
373 * Note that vm_alloc_shared is based on the flags of the vma
374 * for which the page was originally allocated. dst_vma could
375 * be different or NULL on error.
378 SetPagePrivate(page);
380 ClearPagePrivate(page);
385 BUG_ON(!copied && !err);
386 return copied ? copied : err;
388 #else /* !CONFIG_HUGETLB_PAGE */
389 /* fail at build time if gcc attempts to use this */
390 extern ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
391 struct vm_area_struct *dst_vma,
392 unsigned long dst_start,
393 unsigned long src_start,
396 #endif /* CONFIG_HUGETLB_PAGE */
398 static __always_inline ssize_t mfill_atomic_pte(struct mm_struct *dst_mm,
400 struct vm_area_struct *dst_vma,
401 unsigned long dst_addr,
402 unsigned long src_addr,
409 * The normal page fault path for a shmem will invoke the
410 * fault, fill the hole in the file and COW it right away. The
411 * result generates plain anonymous memory. So when we are
412 * asked to fill an hole in a MAP_PRIVATE shmem mapping, we'll
413 * generate anonymous memory directly without actually filling
414 * the hole. For the MAP_PRIVATE case the robustness check
415 * only happens in the pagetable (to verify it's still none)
416 * and not in the radix tree.
418 if (!(dst_vma->vm_flags & VM_SHARED)) {
420 err = mcopy_atomic_pte(dst_mm, dst_pmd, dst_vma,
421 dst_addr, src_addr, page);
423 err = mfill_zeropage_pte(dst_mm, dst_pmd,
427 err = shmem_mcopy_atomic_pte(dst_mm, dst_pmd,
431 err = shmem_mfill_zeropage_pte(dst_mm, dst_pmd,
438 static __always_inline ssize_t __mcopy_atomic(struct mm_struct *dst_mm,
439 unsigned long dst_start,
440 unsigned long src_start,
444 struct vm_area_struct *dst_vma;
447 unsigned long src_addr, dst_addr;
452 * Sanitize the command parameters:
454 BUG_ON(dst_start & ~PAGE_MASK);
455 BUG_ON(len & ~PAGE_MASK);
457 /* Does the address range wrap, or is the span zero-sized? */
458 BUG_ON(src_start + len <= src_start);
459 BUG_ON(dst_start + len <= dst_start);
461 src_addr = src_start;
462 dst_addr = dst_start;
466 down_read(&dst_mm->mmap_sem);
469 * Make sure the vma is not shared, that the dst range is
470 * both valid and fully within a single existing vma.
473 dst_vma = find_vma(dst_mm, dst_start);
477 * Check the vma is registered in uffd, this is required to
478 * enforce the VM_MAYWRITE check done at uffd registration
481 if (!dst_vma->vm_userfaultfd_ctx.ctx)
484 if (dst_start < dst_vma->vm_start ||
485 dst_start + len > dst_vma->vm_end)
490 * shmem_zero_setup is invoked in mmap for MAP_ANONYMOUS|MAP_SHARED but
491 * it will overwrite vm_ops, so vma_is_anonymous must return false.
493 if (WARN_ON_ONCE(vma_is_anonymous(dst_vma) &&
494 dst_vma->vm_flags & VM_SHARED))
498 * If this is a HUGETLB vma, pass off to appropriate routine
500 if (is_vm_hugetlb_page(dst_vma))
501 return __mcopy_atomic_hugetlb(dst_mm, dst_vma, dst_start,
502 src_start, len, zeropage);
504 if (!vma_is_anonymous(dst_vma) && !vma_is_shmem(dst_vma))
508 * Ensure the dst_vma has a anon_vma or this page
509 * would get a NULL anon_vma when moved in the
513 if (!(dst_vma->vm_flags & VM_SHARED) &&
514 unlikely(anon_vma_prepare(dst_vma)))
517 while (src_addr < src_start + len) {
520 BUG_ON(dst_addr >= dst_start + len);
522 dst_pmd = mm_alloc_pmd(dst_mm, dst_addr);
523 if (unlikely(!dst_pmd)) {
528 dst_pmdval = pmd_read_atomic(dst_pmd);
530 * If the dst_pmd is mapped as THP don't
531 * override it and just be strict.
533 if (unlikely(pmd_trans_huge(dst_pmdval))) {
537 if (unlikely(pmd_none(dst_pmdval)) &&
538 unlikely(__pte_alloc(dst_mm, dst_pmd, dst_addr))) {
542 /* If an huge pmd materialized from under us fail */
543 if (unlikely(pmd_trans_huge(*dst_pmd))) {
548 BUG_ON(pmd_none(*dst_pmd));
549 BUG_ON(pmd_trans_huge(*dst_pmd));
551 err = mfill_atomic_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
552 src_addr, &page, zeropage);
555 if (unlikely(err == -ENOENT)) {
558 up_read(&dst_mm->mmap_sem);
561 page_kaddr = kmap(page);
562 err = copy_from_user(page_kaddr,
563 (const void __user *) src_addr,
570 flush_dcache_page(page);
576 dst_addr += PAGE_SIZE;
577 src_addr += PAGE_SIZE;
580 if (fatal_signal_pending(current))
588 up_read(&dst_mm->mmap_sem);
594 BUG_ON(!copied && !err);
595 return copied ? copied : err;
598 ssize_t mcopy_atomic(struct mm_struct *dst_mm, unsigned long dst_start,
599 unsigned long src_start, unsigned long len)
601 return __mcopy_atomic(dst_mm, dst_start, src_start, len, false);
604 ssize_t mfill_zeropage(struct mm_struct *dst_mm, unsigned long start,
607 return __mcopy_atomic(dst_mm, start, 0, len, true);