mm/userfaultfd.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  *  mm/userfaultfd.c
   4  *
   5  *  Copyright (C) 2015  Red Hat, Inc.
   6  */
   7
   8 #include <linux/mm.h>
   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>
  19 #include "internal.h"
  20
  21 static __always_inline
  22 struct vm_area_struct *find_dst_vma(struct mm_struct *dst_mm,
  23                                     unsigned long dst_start,
  24                                     unsigned long len)
  25 {
  26         /*
  27          * Make sure that the dst range is both valid and fully within a
  28          * single existing vma.
  29          */
  30         struct vm_area_struct *dst_vma;
  31
  32         dst_vma = find_vma(dst_mm, dst_start);
  33         if (!dst_vma)
  34                 return NULL;
  35
  36         if (dst_start < dst_vma->vm_start ||
  37             dst_start + len > dst_vma->vm_end)
  38                 return NULL;
  39
  40         /*
  41          * Check the vma is registered in uffd, this is required to
  42          * enforce the VM_MAYWRITE check done at uffd registration
  43          * time.
  44          */
  45         if (!dst_vma->vm_userfaultfd_ctx.ctx)
  46                 return NULL;
  47
  48         return dst_vma;
  49 }
  50
  51 /*
  52  * Install PTEs, to map dst_addr (within dst_vma) to page.
  53  *
  54  * This function handles both MCOPY_ATOMIC_NORMAL and _CONTINUE for both shmem
  55  * and anon, and for both shared and private VMAs.
  56  */
  57 int mfill_atomic_install_pte(struct mm_struct *dst_mm, pmd_t *dst_pmd,
  58                              struct vm_area_struct *dst_vma,
  59                              unsigned long dst_addr, struct page *page,
  60                              bool newly_allocated, bool wp_copy)
  61 {
  62         int ret;
  63         pte_t _dst_pte, *dst_pte;
  64         bool writable = dst_vma->vm_flags & VM_WRITE;
  65         bool vm_shared = dst_vma->vm_flags & VM_SHARED;
  66         bool page_in_cache = page->mapping;
  67         spinlock_t *ptl;
  68         struct inode *inode;
  69         pgoff_t offset, max_off;
  70
  71         _dst_pte = mk_pte(page, dst_vma->vm_page_prot);
  72         if (page_in_cache && !vm_shared)
  73                 writable = false;
  74         if (writable || !page_in_cache)
  75                 _dst_pte = pte_mkdirty(_dst_pte);
  76         if (writable) {
  77                 if (wp_copy)
  78                         _dst_pte = pte_mkuffd_wp(_dst_pte);
  79                 else
  80                         _dst_pte = pte_mkwrite(_dst_pte);
  81         }
  82
  83         dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
  84
  85         if (vma_is_shmem(dst_vma)) {
  86                 /* serialize against truncate with the page table lock */
  87                 inode = dst_vma->vm_file->f_inode;
  88                 offset = linear_page_index(dst_vma, dst_addr);
  89                 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
  90                 ret = -EFAULT;
  91                 if (unlikely(offset >= max_off))
  92                         goto out_unlock;
  93         }
  94
  95         ret = -EEXIST;
  96         if (!pte_none(*dst_pte))
  97                 goto out_unlock;
  98
  99         if (page_in_cache)
 100                 page_add_file_rmap(page, false);
 101         else
 102                 page_add_new_anon_rmap(page, dst_vma, dst_addr, false);
 103
 104         /*
 105          * Must happen after rmap, as mm_counter() checks mapping (via
 106          * PageAnon()), which is set by __page_set_anon_rmap().
 107          */
 108         inc_mm_counter(dst_mm, mm_counter(page));
 109
 110         if (newly_allocated)
 111                 lru_cache_add_inactive_or_unevictable(page, dst_vma);
 112
 113         set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
 114
 115         /* No need to invalidate - it was non-present before */
 116         update_mmu_cache(dst_vma, dst_addr, dst_pte);
 117         ret = 0;
 118 out_unlock:
 119         pte_unmap_unlock(dst_pte, ptl);
 120         return ret;
 121 }
 122
 123 static int mcopy_atomic_pte(struct mm_struct *dst_mm,
 124                             pmd_t *dst_pmd,
 125                             struct vm_area_struct *dst_vma,
 126                             unsigned long dst_addr,
 127                             unsigned long src_addr,
 128                             struct page **pagep,
 129                             bool wp_copy)
 130 {
 131         void *page_kaddr;
 132         int ret;
 133         struct page *page;
 134
 135         if (!*pagep) {
 136                 ret = -ENOMEM;
 137                 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, dst_vma, dst_addr);
 138                 if (!page)
 139                         goto out;
 140
 141                 page_kaddr = kmap_atomic(page);
 142                 ret = copy_from_user(page_kaddr,
 143                                      (const void __user *) src_addr,
 144                                      PAGE_SIZE);
 145                 kunmap_atomic(page_kaddr);
 146
 147                 /* fallback to copy_from_user outside mmap_lock */
 148                 if (unlikely(ret)) {
 149                         ret = -ENOENT;
 150                         *pagep = page;
 151                         /* don't free the page */
 152                         goto out;
 153                 }
 154
 155                 flush_dcache_page(page);
 156         } else {
 157                 page = *pagep;
 158                 *pagep = NULL;
 159         }
 160
 161         /*
 162          * The memory barrier inside __SetPageUptodate makes sure that
 163          * preceding stores to the page contents become visible before
 164          * the set_pte_at() write.
 165          */
 166         __SetPageUptodate(page);
 167
 168         ret = -ENOMEM;
 169         if (mem_cgroup_charge(page, dst_mm, GFP_KERNEL))
 170                 goto out_release;
 171
 172         ret = mfill_atomic_install_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
 173                                        page, true, wp_copy);
 174         if (ret)
 175                 goto out_release;
 176 out:
 177         return ret;
 178 out_release:
 179         put_page(page);
 180         goto out;
 181 }
 182
 183 static int mfill_zeropage_pte(struct mm_struct *dst_mm,
 184                               pmd_t *dst_pmd,
 185                               struct vm_area_struct *dst_vma,
 186                               unsigned long dst_addr)
 187 {
 188         pte_t _dst_pte, *dst_pte;
 189         spinlock_t *ptl;
 190         int ret;
 191         pgoff_t offset, max_off;
 192         struct inode *inode;
 193
 194         _dst_pte = pte_mkspecial(pfn_pte(my_zero_pfn(dst_addr),
 195                                          dst_vma->vm_page_prot));
 196         dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
 197         if (dst_vma->vm_file) {
 198                 /* the shmem MAP_PRIVATE case requires checking the i_size */
 199                 inode = dst_vma->vm_file->f_inode;
 200                 offset = linear_page_index(dst_vma, dst_addr);
 201                 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
 202                 ret = -EFAULT;
 203                 if (unlikely(offset >= max_off))
 204                         goto out_unlock;
 205         }
 206         ret = -EEXIST;
 207         if (!pte_none(*dst_pte))
 208                 goto out_unlock;
 209         set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
 210         /* No need to invalidate - it was non-present before */
 211         update_mmu_cache(dst_vma, dst_addr, dst_pte);
 212         ret = 0;
 213 out_unlock:
 214         pte_unmap_unlock(dst_pte, ptl);
 215         return ret;
 216 }
 217
 218 /* Handles UFFDIO_CONTINUE for all shmem VMAs (shared or private). */
 219 static int mcontinue_atomic_pte(struct mm_struct *dst_mm,
 220                                 pmd_t *dst_pmd,
 221                                 struct vm_area_struct *dst_vma,
 222                                 unsigned long dst_addr,
 223                                 bool wp_copy)
 224 {
 225         struct inode *inode = file_inode(dst_vma->vm_file);
 226         pgoff_t pgoff = linear_page_index(dst_vma, dst_addr);
 227         struct page *page;
 228         int ret;
 229
 230         ret = shmem_getpage(inode, pgoff, &page, SGP_READ);
 231         if (ret)
 232                 goto out;
 233         if (!page) {
 234                 ret = -EFAULT;
 235                 goto out;
 236         }
 237
 238         ret = mfill_atomic_install_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
 239                                        page, false, wp_copy);
 240         if (ret)
 241                 goto out_release;
 242
 243         unlock_page(page);
 244         ret = 0;
 245 out:
 246         return ret;
 247 out_release:
 248         unlock_page(page);
 249         put_page(page);
 250         goto out;
 251 }
 252
 253 static pmd_t *mm_alloc_pmd(struct mm_struct *mm, unsigned long address)
 254 {
 255         pgd_t *pgd;
 256         p4d_t *p4d;
 257         pud_t *pud;
 258
 259         pgd = pgd_offset(mm, address);
 260         p4d = p4d_alloc(mm, pgd, address);
 261         if (!p4d)
 262                 return NULL;
 263         pud = pud_alloc(mm, p4d, address);
 264         if (!pud)
 265                 return NULL;
 266         /*
 267          * Note that we didn't run this because the pmd was
 268          * missing, the *pmd may be already established and in
 269          * turn it may also be a trans_huge_pmd.
 270          */
 271         return pmd_alloc(mm, pud, address);
 272 }
 273
 274 #ifdef CONFIG_HUGETLB_PAGE
 275 /*
 276  * __mcopy_atomic processing for HUGETLB vmas.  Note that this routine is
 277  * called with mmap_lock held, it will release mmap_lock before returning.
 278  */
 279 static __always_inline ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
 280                                               struct vm_area_struct *dst_vma,
 281                                               unsigned long dst_start,
 282                                               unsigned long src_start,
 283                                               unsigned long len,
 284                                               enum mcopy_atomic_mode mode)
 285 {
 286         int vm_shared = dst_vma->vm_flags & VM_SHARED;
 287         ssize_t err;
 288         pte_t *dst_pte;
 289         unsigned long src_addr, dst_addr;
 290         long copied;
 291         struct page *page;
 292         unsigned long vma_hpagesize;
 293         pgoff_t idx;
 294         u32 hash;
 295         struct address_space *mapping;
 296
 297         /*
 298          * There is no default zero huge page for all huge page sizes as
 299          * supported by hugetlb.  A PMD_SIZE huge pages may exist as used
 300          * by THP.  Since we can not reliably insert a zero page, this
 301          * feature is not supported.
 302          */
 303         if (mode == MCOPY_ATOMIC_ZEROPAGE) {
 304                 mmap_read_unlock(dst_mm);
 305                 return -EINVAL;
 306         }
 307
 308         src_addr = src_start;
 309         dst_addr = dst_start;
 310         copied = 0;
 311         page = NULL;
 312         vma_hpagesize = vma_kernel_pagesize(dst_vma);
 313
 314         /*
 315          * Validate alignment based on huge page size
 316          */
 317         err = -EINVAL;
 318         if (dst_start & (vma_hpagesize - 1) || len & (vma_hpagesize - 1))
 319                 goto out_unlock;
 320
 321 retry:
 322         /*
 323          * On routine entry dst_vma is set.  If we had to drop mmap_lock and
 324          * retry, dst_vma will be set to NULL and we must lookup again.
 325          */
 326         if (!dst_vma) {
 327                 err = -ENOENT;
 328                 dst_vma = find_dst_vma(dst_mm, dst_start, len);
 329                 if (!dst_vma || !is_vm_hugetlb_page(dst_vma))
 330                         goto out_unlock;
 331
 332                 err = -EINVAL;
 333                 if (vma_hpagesize != vma_kernel_pagesize(dst_vma))
 334                         goto out_unlock;
 335
 336                 vm_shared = dst_vma->vm_flags & VM_SHARED;
 337         }
 338
 339         /*
 340          * If not shared, ensure the dst_vma has a anon_vma.
 341          */
 342         err = -ENOMEM;
 343         if (!vm_shared) {
 344                 if (unlikely(anon_vma_prepare(dst_vma)))
 345                         goto out_unlock;
 346         }
 347
 348         while (src_addr < src_start + len) {
 349                 BUG_ON(dst_addr >= dst_start + len);
 350
 351                 /*
 352                  * Serialize via i_mmap_rwsem and hugetlb_fault_mutex.
 353                  * i_mmap_rwsem ensures the dst_pte remains valid even
 354                  * in the case of shared pmds.  fault mutex prevents
 355                  * races with other faulting threads.
 356                  */
 357                 mapping = dst_vma->vm_file->f_mapping;
 358                 i_mmap_lock_read(mapping);
 359                 idx = linear_page_index(dst_vma, dst_addr);
 360                 hash = hugetlb_fault_mutex_hash(mapping, idx);
 361                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 362
 363                 err = -ENOMEM;
 364                 dst_pte = huge_pte_alloc(dst_mm, dst_vma, dst_addr, vma_hpagesize);
 365                 if (!dst_pte) {
 366                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 367                         i_mmap_unlock_read(mapping);
 368                         goto out_unlock;
 369                 }
 370
 371                 if (mode != MCOPY_ATOMIC_CONTINUE &&
 372                     !huge_pte_none(huge_ptep_get(dst_pte))) {
 373                         err = -EEXIST;
 374                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 375                         i_mmap_unlock_read(mapping);
 376                         goto out_unlock;
 377                 }
 378
 379                 err = hugetlb_mcopy_atomic_pte(dst_mm, dst_pte, dst_vma,
 380                                                dst_addr, src_addr, mode, &page);
 381
 382                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 383                 i_mmap_unlock_read(mapping);
 384
 385                 cond_resched();
 386
 387                 if (unlikely(err == -ENOENT)) {
 388                         mmap_read_unlock(dst_mm);
 389                         BUG_ON(!page);
 390
 391                         err = copy_huge_page_from_user(page,
 392                                                 (const void __user *)src_addr,
 393                                                 vma_hpagesize / PAGE_SIZE,
 394                                                 true);
 395                         if (unlikely(err)) {
 396                                 err = -EFAULT;
 397                                 goto out;
 398                         }
 399                         mmap_read_lock(dst_mm);
 400
 401                         dst_vma = NULL;
 402                         goto retry;
 403                 } else
 404                         BUG_ON(page);
 405
 406                 if (!err) {
 407                         dst_addr += vma_hpagesize;
 408                         src_addr += vma_hpagesize;
 409                         copied += vma_hpagesize;
 410
 411                         if (fatal_signal_pending(current))
 412                                 err = -EINTR;
 413                 }
 414                 if (err)
 415                         break;
 416         }
 417
 418 out_unlock:
 419         mmap_read_unlock(dst_mm);
 420 out:
 421         if (page)
 422                 put_page(page);
 423         BUG_ON(copied < 0);
 424         BUG_ON(err > 0);
 425         BUG_ON(!copied && !err);
 426         return copied ? copied : err;
 427 }
 428 #else /* !CONFIG_HUGETLB_PAGE */
 429 /* fail at build time if gcc attempts to use this */
 430 extern ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
 431                                       struct vm_area_struct *dst_vma,
 432                                       unsigned long dst_start,
 433                                       unsigned long src_start,
 434                                       unsigned long len,
 435                                       enum mcopy_atomic_mode mode);
 436 #endif /* CONFIG_HUGETLB_PAGE */
 437
 438 static __always_inline ssize_t mfill_atomic_pte(struct mm_struct *dst_mm,
 439                                                 pmd_t *dst_pmd,
 440                                                 struct vm_area_struct *dst_vma,
 441                                                 unsigned long dst_addr,
 442                                                 unsigned long src_addr,
 443                                                 struct page **page,
 444                                                 enum mcopy_atomic_mode mode,
 445                                                 bool wp_copy)
 446 {
 447         ssize_t err;
 448
 449         if (mode == MCOPY_ATOMIC_CONTINUE) {
 450                 return mcontinue_atomic_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
 451                                             wp_copy);
 452         }
 453
 454         /*
 455          * The normal page fault path for a shmem will invoke the
 456          * fault, fill the hole in the file and COW it right away. The
 457          * result generates plain anonymous memory. So when we are
 458          * asked to fill an hole in a MAP_PRIVATE shmem mapping, we'll
 459          * generate anonymous memory directly without actually filling
 460          * the hole. For the MAP_PRIVATE case the robustness check
 461          * only happens in the pagetable (to verify it's still none)
 462          * and not in the radix tree.
 463          */
 464         if (!(dst_vma->vm_flags & VM_SHARED)) {
 465                 if (mode == MCOPY_ATOMIC_NORMAL)
 466                         err = mcopy_atomic_pte(dst_mm, dst_pmd, dst_vma,
 467                                                dst_addr, src_addr, page,
 468                                                wp_copy);
 469                 else
 470                         err = mfill_zeropage_pte(dst_mm, dst_pmd,
 471                                                  dst_vma, dst_addr);
 472         } else {
 473                 VM_WARN_ON_ONCE(wp_copy);
 474                 err = shmem_mfill_atomic_pte(dst_mm, dst_pmd, dst_vma,
 475                                              dst_addr, src_addr,
 476                                              mode != MCOPY_ATOMIC_NORMAL,
 477                                              page);
 478         }
 479
 480         return err;
 481 }
 482
 483 static __always_inline ssize_t __mcopy_atomic(struct mm_struct *dst_mm,
 484                                               unsigned long dst_start,
 485                                               unsigned long src_start,
 486                                               unsigned long len,
 487                                               enum mcopy_atomic_mode mcopy_mode,
 488                                               atomic_t *mmap_changing,
 489                                               __u64 mode)
 490 {
 491         struct vm_area_struct *dst_vma;
 492         ssize_t err;
 493         pmd_t *dst_pmd;
 494         unsigned long src_addr, dst_addr;
 495         long copied;
 496         struct page *page;
 497         bool wp_copy;
 498
 499         /*
 500          * Sanitize the command parameters:
 501          */
 502         BUG_ON(dst_start & ~PAGE_MASK);
 503         BUG_ON(len & ~PAGE_MASK);
 504
 505         /* Does the address range wrap, or is the span zero-sized? */
 506         BUG_ON(src_start + len <= src_start);
 507         BUG_ON(dst_start + len <= dst_start);
 508
 509         src_addr = src_start;
 510         dst_addr = dst_start;
 511         copied = 0;
 512         page = NULL;
 513 retry:
 514         mmap_read_lock(dst_mm);
 515
 516         /*
 517          * If memory mappings are changing because of non-cooperative
 518          * operation (e.g. mremap) running in parallel, bail out and
 519          * request the user to retry later
 520          */
 521         err = -EAGAIN;
 522         if (mmap_changing && atomic_read(mmap_changing))
 523                 goto out_unlock;
 524
 525         /*
 526          * Make sure the vma is not shared, that the dst range is
 527          * both valid and fully within a single existing vma.
 528          */
 529         err = -ENOENT;
 530         dst_vma = find_dst_vma(dst_mm, dst_start, len);
 531         if (!dst_vma)
 532                 goto out_unlock;
 533
 534         err = -EINVAL;
 535         /*
 536          * shmem_zero_setup is invoked in mmap for MAP_ANONYMOUS|MAP_SHARED but
 537          * it will overwrite vm_ops, so vma_is_anonymous must return false.
 538          */
 539         if (WARN_ON_ONCE(vma_is_anonymous(dst_vma) &&
 540             dst_vma->vm_flags & VM_SHARED))
 541                 goto out_unlock;
 542
 543         /*
 544          * validate 'mode' now that we know the dst_vma: don't allow
 545          * a wrprotect copy if the userfaultfd didn't register as WP.
 546          */
 547         wp_copy = mode & UFFDIO_COPY_MODE_WP;
 548         if (wp_copy && !(dst_vma->vm_flags & VM_UFFD_WP))
 549                 goto out_unlock;
 550
 551         /*
 552          * If this is a HUGETLB vma, pass off to appropriate routine
 553          */
 554         if (is_vm_hugetlb_page(dst_vma))
 555                 return  __mcopy_atomic_hugetlb(dst_mm, dst_vma, dst_start,
 556                                                 src_start, len, mcopy_mode);
 557
 558         if (!vma_is_anonymous(dst_vma) && !vma_is_shmem(dst_vma))
 559                 goto out_unlock;
 560         if (!vma_is_shmem(dst_vma) && mcopy_mode == MCOPY_ATOMIC_CONTINUE)
 561                 goto out_unlock;
 562
 563         /*
 564          * Ensure the dst_vma has a anon_vma or this page
 565          * would get a NULL anon_vma when moved in the
 566          * dst_vma.
 567          */
 568         err = -ENOMEM;
 569         if (!(dst_vma->vm_flags & VM_SHARED) &&
 570             unlikely(anon_vma_prepare(dst_vma)))
 571                 goto out_unlock;
 572
 573         while (src_addr < src_start + len) {
 574                 pmd_t dst_pmdval;
 575
 576                 BUG_ON(dst_addr >= dst_start + len);
 577
 578                 dst_pmd = mm_alloc_pmd(dst_mm, dst_addr);
 579                 if (unlikely(!dst_pmd)) {
 580                         err = -ENOMEM;
 581                         break;
 582                 }
 583
 584                 dst_pmdval = pmd_read_atomic(dst_pmd);
 585                 /*
 586                  * If the dst_pmd is mapped as THP don't
 587                  * override it and just be strict.
 588                  */
 589                 if (unlikely(pmd_trans_huge(dst_pmdval))) {
 590                         err = -EEXIST;
 591                         break;
 592                 }
 593                 if (unlikely(pmd_none(dst_pmdval)) &&
 594                     unlikely(__pte_alloc(dst_mm, dst_pmd))) {
 595                         err = -ENOMEM;
 596                         break;
 597                 }
 598                 /* If an huge pmd materialized from under us fail */
 599                 if (unlikely(pmd_trans_huge(*dst_pmd))) {
 600                         err = -EFAULT;
 601                         break;
 602                 }
 603
 604                 BUG_ON(pmd_none(*dst_pmd));
 605                 BUG_ON(pmd_trans_huge(*dst_pmd));
 606
 607                 err = mfill_atomic_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
 608                                        src_addr, &page, mcopy_mode, wp_copy);
 609                 cond_resched();
 610
 611                 if (unlikely(err == -ENOENT)) {
 612                         void *page_kaddr;
 613
 614                         mmap_read_unlock(dst_mm);
 615                         BUG_ON(!page);
 616
 617                         page_kaddr = kmap(page);
 618                         err = copy_from_user(page_kaddr,
 619                                              (const void __user *) src_addr,
 620                                              PAGE_SIZE);
 621                         kunmap(page);
 622                         if (unlikely(err)) {
 623                                 err = -EFAULT;
 624                                 goto out;
 625                         }
 626                         flush_dcache_page(page);
 627                         goto retry;
 628                 } else
 629                         BUG_ON(page);
 630
 631                 if (!err) {
 632                         dst_addr += PAGE_SIZE;
 633                         src_addr += PAGE_SIZE;
 634                         copied += PAGE_SIZE;
 635
 636                         if (fatal_signal_pending(current))
 637                                 err = -EINTR;
 638                 }
 639                 if (err)
 640                         break;
 641         }
 642
 643 out_unlock:
 644         mmap_read_unlock(dst_mm);
 645 out:
 646         if (page)
 647                 put_page(page);
 648         BUG_ON(copied < 0);
 649         BUG_ON(err > 0);
 650         BUG_ON(!copied && !err);
 651         return copied ? copied : err;
 652 }
 653
 654 ssize_t mcopy_atomic(struct mm_struct *dst_mm, unsigned long dst_start,
 655                      unsigned long src_start, unsigned long len,
 656                      atomic_t *mmap_changing, __u64 mode)
 657 {
 658         return __mcopy_atomic(dst_mm, dst_start, src_start, len,
 659                               MCOPY_ATOMIC_NORMAL, mmap_changing, mode);
 660 }
 661
 662 ssize_t mfill_zeropage(struct mm_struct *dst_mm, unsigned long start,
 663                        unsigned long len, atomic_t *mmap_changing)
 664 {
 665         return __mcopy_atomic(dst_mm, start, 0, len, MCOPY_ATOMIC_ZEROPAGE,
 666                               mmap_changing, 0);
 667 }
 668
 669 ssize_t mcopy_continue(struct mm_struct *dst_mm, unsigned long start,
 670                        unsigned long len, atomic_t *mmap_changing)
 671 {
 672         return __mcopy_atomic(dst_mm, start, 0, len, MCOPY_ATOMIC_CONTINUE,
 673                               mmap_changing, 0);
 674 }
 675
 676 int mwriteprotect_range(struct mm_struct *dst_mm, unsigned long start,
 677                         unsigned long len, bool enable_wp,
 678                         atomic_t *mmap_changing)
 679 {
 680         struct vm_area_struct *dst_vma;
 681         pgprot_t newprot;
 682         int err;
 683
 684         /*
 685          * Sanitize the command parameters:
 686          */
 687         BUG_ON(start & ~PAGE_MASK);
 688         BUG_ON(len & ~PAGE_MASK);
 689
 690         /* Does the address range wrap, or is the span zero-sized? */
 691         BUG_ON(start + len <= start);
 692
 693         mmap_read_lock(dst_mm);
 694
 695         /*
 696          * If memory mappings are changing because of non-cooperative
 697          * operation (e.g. mremap) running in parallel, bail out and
 698          * request the user to retry later
 699          */
 700         err = -EAGAIN;
 701         if (mmap_changing && atomic_read(mmap_changing))
 702                 goto out_unlock;
 703
 704         err = -ENOENT;
 705         dst_vma = find_dst_vma(dst_mm, start, len);
 706         /*
 707          * Make sure the vma is not shared, that the dst range is
 708          * both valid and fully within a single existing vma.
 709          */
 710         if (!dst_vma || (dst_vma->vm_flags & VM_SHARED))
 711                 goto out_unlock;
 712         if (!userfaultfd_wp(dst_vma))
 713                 goto out_unlock;
 714         if (!vma_is_anonymous(dst_vma))
 715                 goto out_unlock;
 716
 717         if (enable_wp)
 718                 newprot = vm_get_page_prot(dst_vma->vm_flags & ~(VM_WRITE));
 719         else
 720                 newprot = vm_get_page_prot(dst_vma->vm_flags);
 721
 722         change_protection(dst_vma, start, start + len, newprot,
 723                           enable_wp ? MM_CP_UFFD_WP : MM_CP_UFFD_WP_RESOLVE);
 724
 725         err = 0;
 726 out_unlock:
 727         mmap_read_unlock(dst_mm);
 728         return err;
 729 }