mm/userfaultfd.c

   1 /*
   2  *  mm/userfaultfd.c
   3  *
   4  *  Copyright (C) 2015  Red Hat, Inc.
   5  *
   6  *  This work is licensed under the terms of the GNU GPL, version 2. See
   7  *  the COPYING file in the top-level directory.
   8  */
   9
  10 #include <linux/mm.h>
  11 #include <linux/pagemap.h>
  12 #include <linux/rmap.h>
  13 #include <linux/swap.h>
  14 #include <linux/swapops.h>
  15 #include <linux/userfaultfd_k.h>
  16 #include <linux/mmu_notifier.h>
  17 #include <asm/tlbflush.h>
  18 #include "internal.h"
  19
  20 static int mcopy_atomic_pte(struct mm_struct *dst_mm,
  21                             pmd_t *dst_pmd,
  22                             struct vm_area_struct *dst_vma,
  23                             unsigned long dst_addr,
  24                             unsigned long src_addr,
  25                             struct page **pagep)
  26 {
  27         struct mem_cgroup *memcg;
  28         pte_t _dst_pte, *dst_pte;
  29         spinlock_t *ptl;
  30         void *page_kaddr;
  31         int ret;
  32         struct page *page;
  33
  34         if (!*pagep) {
  35                 ret = -ENOMEM;
  36                 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, dst_vma, dst_addr);
  37                 if (!page)
  38                         goto out;
  39
  40                 page_kaddr = kmap_atomic(page);
  41                 ret = copy_from_user(page_kaddr,
  42                                      (const void __user *) src_addr,
  43                                      PAGE_SIZE);
  44                 kunmap_atomic(page_kaddr);
  45
  46                 /* fallback to copy_from_user outside mmap_sem */
  47                 if (unlikely(ret)) {
  48                         ret = -EFAULT;
  49                         *pagep = page;
  50                         /* don't free the page */
  51                         goto out;
  52                 }
  53
  54                 flush_dcache_page(page);
  55         } else {
  56                 page = *pagep;
  57                 *pagep = NULL;
  58         }
  59
  60         /*
  61          * The memory barrier inside __SetPageUptodate makes sure that
  62          * preceeding stores to the page contents become visible before
  63          * the set_pte_at() write.
  64          */
  65         __SetPageUptodate(page);
  66
  67         ret = -ENOMEM;
  68         if (mem_cgroup_try_charge(page, dst_mm, GFP_KERNEL, &memcg, false))
  69                 goto out_release;
  70
  71         _dst_pte = mk_pte(page, dst_vma->vm_page_prot);
  72         if (dst_vma->vm_flags & VM_WRITE)
  73                 _dst_pte = pte_mkwrite(pte_mkdirty(_dst_pte));
  74
  75         ret = -EEXIST;
  76         dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
  77         if (!pte_none(*dst_pte))
  78                 goto out_release_uncharge_unlock;
  79
  80         inc_mm_counter(dst_mm, MM_ANONPAGES);
  81         page_add_new_anon_rmap(page, dst_vma, dst_addr, false);
  82         mem_cgroup_commit_charge(page, memcg, false, false);
  83         lru_cache_add_active_or_unevictable(page, dst_vma);
  84
  85         set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
  86
  87         /* No need to invalidate - it was non-present before */
  88         update_mmu_cache(dst_vma, dst_addr, dst_pte);
  89
  90         pte_unmap_unlock(dst_pte, ptl);
  91         ret = 0;
  92 out:
  93         return ret;
  94 out_release_uncharge_unlock:
  95         pte_unmap_unlock(dst_pte, ptl);
  96         mem_cgroup_cancel_charge(page, memcg, false);
  97 out_release:
  98         put_page(page);
  99         goto out;
 100 }
 101
 102 static int mfill_zeropage_pte(struct mm_struct *dst_mm,
 103                               pmd_t *dst_pmd,
 104                               struct vm_area_struct *dst_vma,
 105                               unsigned long dst_addr)
 106 {
 107         pte_t _dst_pte, *dst_pte;
 108         spinlock_t *ptl;
 109         int ret;
 110
 111         _dst_pte = pte_mkspecial(pfn_pte(my_zero_pfn(dst_addr),
 112                                          dst_vma->vm_page_prot));
 113         ret = -EEXIST;
 114         dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
 115         if (!pte_none(*dst_pte))
 116                 goto out_unlock;
 117         set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
 118         /* No need to invalidate - it was non-present before */
 119         update_mmu_cache(dst_vma, dst_addr, dst_pte);
 120         ret = 0;
 121 out_unlock:
 122         pte_unmap_unlock(dst_pte, ptl);
 123         return ret;
 124 }
 125
 126 static pmd_t *mm_alloc_pmd(struct mm_struct *mm, unsigned long address)
 127 {
 128         pgd_t *pgd;
 129         pud_t *pud;
 130         pmd_t *pmd = NULL;
 131
 132         pgd = pgd_offset(mm, address);
 133         pud = pud_alloc(mm, pgd, address);
 134         if (pud)
 135                 /*
 136                  * Note that we didn't run this because the pmd was
 137                  * missing, the *pmd may be already established and in
 138                  * turn it may also be a trans_huge_pmd.
 139                  */
 140                 pmd = pmd_alloc(mm, pud, address);
 141         return pmd;
 142 }
 143
 144 static __always_inline ssize_t __mcopy_atomic(struct mm_struct *dst_mm,
 145                                               unsigned long dst_start,
 146                                               unsigned long src_start,
 147                                               unsigned long len,
 148                                               bool zeropage)
 149 {
 150         struct vm_area_struct *dst_vma;
 151         ssize_t err;
 152         pmd_t *dst_pmd;
 153         unsigned long src_addr, dst_addr;
 154         long copied;
 155         struct page *page;
 156
 157         /*
 158          * Sanitize the command parameters:
 159          */
 160         BUG_ON(dst_start & ~PAGE_MASK);
 161         BUG_ON(len & ~PAGE_MASK);
 162
 163         /* Does the address range wrap, or is the span zero-sized? */
 164         BUG_ON(src_start + len <= src_start);
 165         BUG_ON(dst_start + len <= dst_start);
 166
 167         src_addr = src_start;
 168         dst_addr = dst_start;
 169         copied = 0;
 170         page = NULL;
 171 retry:
 172         down_read(&dst_mm->mmap_sem);
 173
 174         /*
 175          * Make sure the vma is not shared, that the dst range is
 176          * both valid and fully within a single existing vma.
 177          */
 178         err = -EINVAL;
 179         dst_vma = find_vma(dst_mm, dst_start);
 180         if (!dst_vma || (dst_vma->vm_flags & VM_SHARED))
 181                 goto out_unlock;
 182         if (dst_start < dst_vma->vm_start ||
 183             dst_start + len > dst_vma->vm_end)
 184                 goto out_unlock;
 185
 186         /*
 187          * Be strict and only allow __mcopy_atomic on userfaultfd
 188          * registered ranges to prevent userland errors going
 189          * unnoticed. As far as the VM consistency is concerned, it
 190          * would be perfectly safe to remove this check, but there's
 191          * no useful usage for __mcopy_atomic ouside of userfaultfd
 192          * registered ranges. This is after all why these are ioctls
 193          * belonging to the userfaultfd and not syscalls.
 194          */
 195         if (!dst_vma->vm_userfaultfd_ctx.ctx)
 196                 goto out_unlock;
 197
 198         /*
 199          * FIXME: only allow copying on anonymous vmas, tmpfs should
 200          * be added.
 201          */
 202         if (dst_vma->vm_ops)
 203                 goto out_unlock;
 204
 205         /*
 206          * Ensure the dst_vma has a anon_vma or this page
 207          * would get a NULL anon_vma when moved in the
 208          * dst_vma.
 209          */
 210         err = -ENOMEM;
 211         if (unlikely(anon_vma_prepare(dst_vma)))
 212                 goto out_unlock;
 213
 214         while (src_addr < src_start + len) {
 215                 pmd_t dst_pmdval;
 216
 217                 BUG_ON(dst_addr >= dst_start + len);
 218
 219                 dst_pmd = mm_alloc_pmd(dst_mm, dst_addr);
 220                 if (unlikely(!dst_pmd)) {
 221                         err = -ENOMEM;
 222                         break;
 223                 }
 224
 225                 dst_pmdval = pmd_read_atomic(dst_pmd);
 226                 /*
 227                  * If the dst_pmd is mapped as THP don't
 228                  * override it and just be strict.
 229                  */
 230                 if (unlikely(pmd_trans_huge(dst_pmdval))) {
 231                         err = -EEXIST;
 232                         break;
 233                 }
 234                 if (unlikely(pmd_none(dst_pmdval)) &&
 235                     unlikely(__pte_alloc(dst_mm, dst_pmd, dst_addr))) {
 236                         err = -ENOMEM;
 237                         break;
 238                 }
 239                 /* If an huge pmd materialized from under us fail */
 240                 if (unlikely(pmd_trans_huge(*dst_pmd))) {
 241                         err = -EFAULT;
 242                         break;
 243                 }
 244
 245                 BUG_ON(pmd_none(*dst_pmd));
 246                 BUG_ON(pmd_trans_huge(*dst_pmd));
 247
 248                 if (!zeropage)
 249                         err = mcopy_atomic_pte(dst_mm, dst_pmd, dst_vma,
 250                                                dst_addr, src_addr, &page);
 251                 else
 252                         err = mfill_zeropage_pte(dst_mm, dst_pmd, dst_vma,
 253                                                  dst_addr);
 254
 255                 cond_resched();
 256
 257                 if (unlikely(err == -EFAULT)) {
 258                         void *page_kaddr;
 259
 260                         up_read(&dst_mm->mmap_sem);
 261                         BUG_ON(!page);
 262
 263                         page_kaddr = kmap(page);
 264                         err = copy_from_user(page_kaddr,
 265                                              (const void __user *) src_addr,
 266                                              PAGE_SIZE);
 267                         kunmap(page);
 268                         if (unlikely(err)) {
 269                                 err = -EFAULT;
 270                                 goto out;
 271                         }
 272                         flush_dcache_page(page);
 273                         goto retry;
 274                 } else
 275                         BUG_ON(page);
 276
 277                 if (!err) {
 278                         dst_addr += PAGE_SIZE;
 279                         src_addr += PAGE_SIZE;
 280                         copied += PAGE_SIZE;
 281
 282                         if (fatal_signal_pending(current))
 283                                 err = -EINTR;
 284                 }
 285                 if (err)
 286                         break;
 287         }
 288
 289 out_unlock:
 290         up_read(&dst_mm->mmap_sem);
 291 out:
 292         if (page)
 293                 put_page(page);
 294         BUG_ON(copied < 0);
 295         BUG_ON(err > 0);
 296         BUG_ON(!copied && !err);
 297         return copied ? copied : err;
 298 }
 299
 300 ssize_t mcopy_atomic(struct mm_struct *dst_mm, unsigned long dst_start,
 301                      unsigned long src_start, unsigned long len)
 302 {
 303         return __mcopy_atomic(dst_mm, dst_start, src_start, len, false);
 304 }
 305
 306 ssize_t mfill_zeropage(struct mm_struct *dst_mm, unsigned long start,
 307                        unsigned long len)
 308 {
 309         return __mcopy_atomic(dst_mm, start, 0, len, true);
 310 }