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;
40 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, dst_vma, dst_addr);
44 page_kaddr = kmap_atomic(page);
45 ret = copy_from_user(page_kaddr,
46 (const void __user *) src_addr,
48 kunmap_atomic(page_kaddr);
50 /* fallback to copy_from_user outside mmap_sem */
54 /* don't free the page */
63 * The memory barrier inside __SetPageUptodate makes sure that
64 * preceeding stores to the page contents become visible before
65 * the set_pte_at() write.
67 __SetPageUptodate(page);
70 if (mem_cgroup_try_charge(page, dst_mm, GFP_KERNEL, &memcg, false))
73 _dst_pte = mk_pte(page, dst_vma->vm_page_prot);
74 if (dst_vma->vm_flags & VM_WRITE)
75 _dst_pte = pte_mkwrite(pte_mkdirty(_dst_pte));
78 dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
79 if (!pte_none(*dst_pte))
80 goto out_release_uncharge_unlock;
82 inc_mm_counter(dst_mm, MM_ANONPAGES);
83 page_add_new_anon_rmap(page, dst_vma, dst_addr, false);
84 mem_cgroup_commit_charge(page, memcg, false, false);
85 lru_cache_add_active_or_unevictable(page, dst_vma);
87 set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
89 /* No need to invalidate - it was non-present before */
90 update_mmu_cache(dst_vma, dst_addr, dst_pte);
92 pte_unmap_unlock(dst_pte, ptl);
96 out_release_uncharge_unlock:
97 pte_unmap_unlock(dst_pte, ptl);
98 mem_cgroup_cancel_charge(page, memcg, false);
104 static int mfill_zeropage_pte(struct mm_struct *dst_mm,
106 struct vm_area_struct *dst_vma,
107 unsigned long dst_addr)
109 pte_t _dst_pte, *dst_pte;
113 _dst_pte = pte_mkspecial(pfn_pte(my_zero_pfn(dst_addr),
114 dst_vma->vm_page_prot));
116 dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
117 if (!pte_none(*dst_pte))
119 set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
120 /* No need to invalidate - it was non-present before */
121 update_mmu_cache(dst_vma, dst_addr, dst_pte);
124 pte_unmap_unlock(dst_pte, ptl);
128 static pmd_t *mm_alloc_pmd(struct mm_struct *mm, unsigned long address)
134 pgd = pgd_offset(mm, address);
135 pud = pud_alloc(mm, pgd, address);
138 * Note that we didn't run this because the pmd was
139 * missing, the *pmd may be already established and in
140 * turn it may also be a trans_huge_pmd.
142 pmd = pmd_alloc(mm, pud, address);
146 #ifdef CONFIG_HUGETLB_PAGE
148 * __mcopy_atomic processing for HUGETLB vmas. Note that this routine is
149 * called with mmap_sem held, it will release mmap_sem before returning.
151 static __always_inline ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
152 struct vm_area_struct *dst_vma,
153 unsigned long dst_start,
154 unsigned long src_start,
158 int vm_alloc_shared = dst_vma->vm_flags & VM_SHARED;
159 int vm_shared = dst_vma->vm_flags & VM_SHARED;
162 unsigned long src_addr, dst_addr;
166 unsigned long vma_hpagesize;
169 struct address_space *mapping;
172 * There is no default zero huge page for all huge page sizes as
173 * supported by hugetlb. A PMD_SIZE huge pages may exist as used
174 * by THP. Since we can not reliably insert a zero page, this
175 * feature is not supported.
178 up_read(&dst_mm->mmap_sem);
182 src_addr = src_start;
183 dst_addr = dst_start;
186 vma_hpagesize = vma_kernel_pagesize(dst_vma);
189 * Validate alignment based on huge page size
192 if (dst_start & (vma_hpagesize - 1) || len & (vma_hpagesize - 1))
197 * On routine entry dst_vma is set. If we had to drop mmap_sem and
198 * retry, dst_vma will be set to NULL and we must lookup again.
202 dst_vma = find_vma(dst_mm, dst_start);
203 if (!dst_vma || !is_vm_hugetlb_page(dst_vma))
206 * Only allow __mcopy_atomic_hugetlb on userfaultfd
209 if (!dst_vma->vm_userfaultfd_ctx.ctx)
212 if (dst_start < dst_vma->vm_start ||
213 dst_start + len > dst_vma->vm_end)
217 if (vma_hpagesize != vma_kernel_pagesize(dst_vma))
220 vm_shared = dst_vma->vm_flags & VM_SHARED;
223 if (WARN_ON(dst_addr & (vma_hpagesize - 1) ||
224 (len - copied) & (vma_hpagesize - 1)))
228 * If not shared, ensure the dst_vma has a anon_vma.
232 if (unlikely(anon_vma_prepare(dst_vma)))
236 h = hstate_vma(dst_vma);
238 while (src_addr < src_start + len) {
241 BUG_ON(dst_addr >= dst_start + len);
242 VM_BUG_ON(dst_addr & ~huge_page_mask(h));
245 * Serialize via hugetlb_fault_mutex
247 idx = linear_page_index(dst_vma, dst_addr);
248 mapping = dst_vma->vm_file->f_mapping;
249 hash = hugetlb_fault_mutex_hash(h, dst_mm, dst_vma, mapping,
251 mutex_lock(&hugetlb_fault_mutex_table[hash]);
254 dst_pte = huge_pte_alloc(dst_mm, dst_addr, huge_page_size(h));
256 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
261 dst_pteval = huge_ptep_get(dst_pte);
262 if (!huge_pte_none(dst_pteval)) {
263 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
267 err = hugetlb_mcopy_atomic_pte(dst_mm, dst_pte, dst_vma,
268 dst_addr, src_addr, &page);
270 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
271 vm_alloc_shared = vm_shared;
275 if (unlikely(err == -EFAULT)) {
276 up_read(&dst_mm->mmap_sem);
279 err = copy_huge_page_from_user(page,
280 (const void __user *)src_addr,
281 pages_per_huge_page(h), true);
286 down_read(&dst_mm->mmap_sem);
294 dst_addr += vma_hpagesize;
295 src_addr += vma_hpagesize;
296 copied += vma_hpagesize;
298 if (fatal_signal_pending(current))
306 up_read(&dst_mm->mmap_sem);
310 * We encountered an error and are about to free a newly
311 * allocated huge page.
313 * Reservation handling is very subtle, and is different for
314 * private and shared mappings. See the routine
315 * restore_reserve_on_error for details. Unfortunately, we
316 * can not call restore_reserve_on_error now as it would
317 * require holding mmap_sem.
319 * If a reservation for the page existed in the reservation
320 * map of a private mapping, the map was modified to indicate
321 * the reservation was consumed when the page was allocated.
322 * We clear the PagePrivate flag now so that the global
323 * reserve count will not be incremented in free_huge_page.
324 * The reservation map will still indicate the reservation
325 * was consumed and possibly prevent later page allocation.
326 * This is better than leaking a global reservation. If no
327 * reservation existed, it is still safe to clear PagePrivate
328 * as no adjustments to reservation counts were made during
331 * The reservation map for shared mappings indicates which
332 * pages have reservations. When a huge page is allocated
333 * for an address with a reservation, no change is made to
334 * the reserve map. In this case PagePrivate will be set
335 * to indicate that the global reservation count should be
336 * incremented when the page is freed. This is the desired
337 * behavior. However, when a huge page is allocated for an
338 * address without a reservation a reservation entry is added
339 * to the reservation map, and PagePrivate will not be set.
340 * When the page is freed, the global reserve count will NOT
341 * be incremented and it will appear as though we have leaked
342 * reserved page. In this case, set PagePrivate so that the
343 * global reserve count will be incremented to match the
344 * reservation map entry which was created.
346 * Note that vm_alloc_shared is based on the flags of the vma
347 * for which the page was originally allocated. dst_vma could
348 * be different or NULL on error.
351 SetPagePrivate(page);
353 ClearPagePrivate(page);
358 BUG_ON(!copied && !err);
359 return copied ? copied : err;
361 #else /* !CONFIG_HUGETLB_PAGE */
362 /* fail at build time if gcc attempts to use this */
363 extern ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
364 struct vm_area_struct *dst_vma,
365 unsigned long dst_start,
366 unsigned long src_start,
369 #endif /* CONFIG_HUGETLB_PAGE */
371 static __always_inline ssize_t __mcopy_atomic(struct mm_struct *dst_mm,
372 unsigned long dst_start,
373 unsigned long src_start,
377 struct vm_area_struct *dst_vma;
380 unsigned long src_addr, dst_addr;
385 * Sanitize the command parameters:
387 BUG_ON(dst_start & ~PAGE_MASK);
388 BUG_ON(len & ~PAGE_MASK);
390 /* Does the address range wrap, or is the span zero-sized? */
391 BUG_ON(src_start + len <= src_start);
392 BUG_ON(dst_start + len <= dst_start);
394 src_addr = src_start;
395 dst_addr = dst_start;
399 down_read(&dst_mm->mmap_sem);
402 * Make sure the vma is not shared, that the dst range is
403 * both valid and fully within a single existing vma.
406 dst_vma = find_vma(dst_mm, dst_start);
410 * Be strict and only allow __mcopy_atomic on userfaultfd
411 * registered ranges to prevent userland errors going
412 * unnoticed. As far as the VM consistency is concerned, it
413 * would be perfectly safe to remove this check, but there's
414 * no useful usage for __mcopy_atomic ouside of userfaultfd
415 * registered ranges. This is after all why these are ioctls
416 * belonging to the userfaultfd and not syscalls.
418 if (!dst_vma->vm_userfaultfd_ctx.ctx)
421 if (dst_start < dst_vma->vm_start ||
422 dst_start + len > dst_vma->vm_end)
427 * shmem_zero_setup is invoked in mmap for MAP_ANONYMOUS|MAP_SHARED but
428 * it will overwrite vm_ops, so vma_is_anonymous must return false.
430 if (WARN_ON_ONCE(vma_is_anonymous(dst_vma) &&
431 dst_vma->vm_flags & VM_SHARED))
435 * If this is a HUGETLB vma, pass off to appropriate routine
437 if (is_vm_hugetlb_page(dst_vma))
438 return __mcopy_atomic_hugetlb(dst_mm, dst_vma, dst_start,
439 src_start, len, zeropage);
441 if (!vma_is_anonymous(dst_vma) && !vma_is_shmem(dst_vma))
445 * Ensure the dst_vma has a anon_vma or this page
446 * would get a NULL anon_vma when moved in the
450 if (vma_is_anonymous(dst_vma) && unlikely(anon_vma_prepare(dst_vma)))
453 while (src_addr < src_start + len) {
456 BUG_ON(dst_addr >= dst_start + len);
458 dst_pmd = mm_alloc_pmd(dst_mm, dst_addr);
459 if (unlikely(!dst_pmd)) {
464 dst_pmdval = pmd_read_atomic(dst_pmd);
466 * If the dst_pmd is mapped as THP don't
467 * override it and just be strict.
469 if (unlikely(pmd_trans_huge(dst_pmdval))) {
473 if (unlikely(pmd_none(dst_pmdval)) &&
474 unlikely(__pte_alloc(dst_mm, dst_pmd, dst_addr))) {
478 /* If an huge pmd materialized from under us fail */
479 if (unlikely(pmd_trans_huge(*dst_pmd))) {
484 BUG_ON(pmd_none(*dst_pmd));
485 BUG_ON(pmd_trans_huge(*dst_pmd));
487 if (vma_is_anonymous(dst_vma)) {
489 err = mcopy_atomic_pte(dst_mm, dst_pmd, dst_vma,
493 err = mfill_zeropage_pte(dst_mm, dst_pmd,
496 err = -EINVAL; /* if zeropage is true return -EINVAL */
497 if (likely(!zeropage))
498 err = shmem_mcopy_atomic_pte(dst_mm, dst_pmd,
505 if (unlikely(err == -EFAULT)) {
508 up_read(&dst_mm->mmap_sem);
511 page_kaddr = kmap(page);
512 err = copy_from_user(page_kaddr,
513 (const void __user *) src_addr,
525 dst_addr += PAGE_SIZE;
526 src_addr += PAGE_SIZE;
529 if (fatal_signal_pending(current))
537 up_read(&dst_mm->mmap_sem);
543 BUG_ON(!copied && !err);
544 return copied ? copied : err;
547 ssize_t mcopy_atomic(struct mm_struct *dst_mm, unsigned long dst_start,
548 unsigned long src_start, unsigned long len)
550 return __mcopy_atomic(dst_mm, dst_start, src_start, len, false);
553 ssize_t mfill_zeropage(struct mm_struct *dst_mm, unsigned long start,
556 return __mcopy_atomic(dst_mm, start, 0, len, true);