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         _dst_pte = pte_mkdirty(_dst_pte);
  73         if (page_in_cache && !vm_shared)
  74                 writable = false;
  75
  76         /*
  77          * Always mark a PTE as write-protected when needed, regardless of
  78          * VM_WRITE, which the user might change.
  79          */
  80         if (wp_copy)
  81                 _dst_pte = pte_mkuffd_wp(_dst_pte);
  82         else if (writable)
  83                 _dst_pte = pte_mkwrite(_dst_pte);
  84
  85         dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
  86
  87         if (vma_is_shmem(dst_vma)) {
  88                 /* serialize against truncate with the page table lock */
  89                 inode = dst_vma->vm_file->f_inode;
  90                 offset = linear_page_index(dst_vma, dst_addr);
  91                 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
  92                 ret = -EFAULT;
  93                 if (unlikely(offset >= max_off))
  94                         goto out_unlock;
  95         }
  96
  97         ret = -EEXIST;
  98         if (!pte_none(*dst_pte))
  99                 goto out_unlock;
 100
 101         if (page_in_cache) {
 102                 /* Usually, cache pages are already added to LRU */
 103                 if (newly_allocated)
 104                         lru_cache_add(page);
 105                 page_add_file_rmap(page, dst_vma, false);
 106         } else {
 107                 page_add_new_anon_rmap(page, dst_vma, dst_addr, false);
 108                 lru_cache_add_inactive_or_unevictable(page, dst_vma);
 109         }
 110
 111         /*
 112          * Must happen after rmap, as mm_counter() checks mapping (via
 113          * PageAnon()), which is set by __page_set_anon_rmap().
 114          */
 115         inc_mm_counter(dst_mm, mm_counter(page));
 116
 117         set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
 118
 119         /* No need to invalidate - it was non-present before */
 120         update_mmu_cache(dst_vma, dst_addr, dst_pte);
 121         ret = 0;
 122 out_unlock:
 123         pte_unmap_unlock(dst_pte, ptl);
 124         return ret;
 125 }
 126
 127 static int mcopy_atomic_pte(struct mm_struct *dst_mm,
 128                             pmd_t *dst_pmd,
 129                             struct vm_area_struct *dst_vma,
 130                             unsigned long dst_addr,
 131                             unsigned long src_addr,
 132                             struct page **pagep,
 133                             bool wp_copy)
 134 {
 135         void *page_kaddr;
 136         int ret;
 137         struct page *page;
 138
 139         if (!*pagep) {
 140                 ret = -ENOMEM;
 141                 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, dst_vma, dst_addr);
 142                 if (!page)
 143                         goto out;
 144
 145                 page_kaddr = kmap_atomic(page);
 146                 ret = copy_from_user(page_kaddr,
 147                                      (const void __user *) src_addr,
 148                                      PAGE_SIZE);
 149                 kunmap_atomic(page_kaddr);
 150
 151                 /* fallback to copy_from_user outside mmap_lock */
 152                 if (unlikely(ret)) {
 153                         ret = -ENOENT;
 154                         *pagep = page;
 155                         /* don't free the page */
 156                         goto out;
 157                 }
 158
 159                 flush_dcache_page(page);
 160         } else {
 161                 page = *pagep;
 162                 *pagep = NULL;
 163         }
 164
 165         /*
 166          * The memory barrier inside __SetPageUptodate makes sure that
 167          * preceding stores to the page contents become visible before
 168          * the set_pte_at() write.
 169          */
 170         __SetPageUptodate(page);
 171
 172         ret = -ENOMEM;
 173         if (mem_cgroup_charge(page_folio(page), dst_mm, GFP_KERNEL))
 174                 goto out_release;
 175
 176         ret = mfill_atomic_install_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
 177                                        page, true, wp_copy);
 178         if (ret)
 179                 goto out_release;
 180 out:
 181         return ret;
 182 out_release:
 183         put_page(page);
 184         goto out;
 185 }
 186
 187 static int mfill_zeropage_pte(struct mm_struct *dst_mm,
 188                               pmd_t *dst_pmd,
 189                               struct vm_area_struct *dst_vma,
 190                               unsigned long dst_addr)
 191 {
 192         pte_t _dst_pte, *dst_pte;
 193         spinlock_t *ptl;
 194         int ret;
 195         pgoff_t offset, max_off;
 196         struct inode *inode;
 197
 198         _dst_pte = pte_mkspecial(pfn_pte(my_zero_pfn(dst_addr),
 199                                          dst_vma->vm_page_prot));
 200         dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
 201         if (dst_vma->vm_file) {
 202                 /* the shmem MAP_PRIVATE case requires checking the i_size */
 203                 inode = dst_vma->vm_file->f_inode;
 204                 offset = linear_page_index(dst_vma, dst_addr);
 205                 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
 206                 ret = -EFAULT;
 207                 if (unlikely(offset >= max_off))
 208                         goto out_unlock;
 209         }
 210         ret = -EEXIST;
 211         if (!pte_none(*dst_pte))
 212                 goto out_unlock;
 213         set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
 214         /* No need to invalidate - it was non-present before */
 215         update_mmu_cache(dst_vma, dst_addr, dst_pte);
 216         ret = 0;
 217 out_unlock:
 218         pte_unmap_unlock(dst_pte, ptl);
 219         return ret;
 220 }
 221
 222 /* Handles UFFDIO_CONTINUE for all shmem VMAs (shared or private). */
 223 static int mcontinue_atomic_pte(struct mm_struct *dst_mm,
 224                                 pmd_t *dst_pmd,
 225                                 struct vm_area_struct *dst_vma,
 226                                 unsigned long dst_addr,
 227                                 bool wp_copy)
 228 {
 229         struct inode *inode = file_inode(dst_vma->vm_file);
 230         pgoff_t pgoff = linear_page_index(dst_vma, dst_addr);
 231         struct page *page;
 232         int ret;
 233
 234         ret = shmem_getpage(inode, pgoff, &page, SGP_READ);
 235         if (ret)
 236                 goto out;
 237         if (!page) {
 238                 ret = -EFAULT;
 239                 goto out;
 240         }
 241
 242         if (PageHWPoison(page)) {
 243                 ret = -EIO;
 244                 goto out_release;
 245         }
 246
 247         ret = mfill_atomic_install_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
 248                                        page, false, wp_copy);
 249         if (ret)
 250                 goto out_release;
 251
 252         unlock_page(page);
 253         ret = 0;
 254 out:
 255         return ret;
 256 out_release:
 257         unlock_page(page);
 258         put_page(page);
 259         goto out;
 260 }
 261
 262 static pmd_t *mm_alloc_pmd(struct mm_struct *mm, unsigned long address)
 263 {
 264         pgd_t *pgd;
 265         p4d_t *p4d;
 266         pud_t *pud;
 267
 268         pgd = pgd_offset(mm, address);
 269         p4d = p4d_alloc(mm, pgd, address);
 270         if (!p4d)
 271                 return NULL;
 272         pud = pud_alloc(mm, p4d, address);
 273         if (!pud)
 274                 return NULL;
 275         /*
 276          * Note that we didn't run this because the pmd was
 277          * missing, the *pmd may be already established and in
 278          * turn it may also be a trans_huge_pmd.
 279          */
 280         return pmd_alloc(mm, pud, address);
 281 }
 282
 283 #ifdef CONFIG_HUGETLB_PAGE
 284 /*
 285  * __mcopy_atomic processing for HUGETLB vmas.  Note that this routine is
 286  * called with mmap_lock held, it will release mmap_lock before returning.
 287  */
 288 static __always_inline ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
 289                                               struct vm_area_struct *dst_vma,
 290                                               unsigned long dst_start,
 291                                               unsigned long src_start,
 292                                               unsigned long len,
 293                                               enum mcopy_atomic_mode mode)
 294 {
 295         int vm_shared = dst_vma->vm_flags & VM_SHARED;
 296         ssize_t err;
 297         pte_t *dst_pte;
 298         unsigned long src_addr, dst_addr;
 299         long copied;
 300         struct page *page;
 301         unsigned long vma_hpagesize;
 302         pgoff_t idx;
 303         u32 hash;
 304         struct address_space *mapping;
 305
 306         /*
 307          * There is no default zero huge page for all huge page sizes as
 308          * supported by hugetlb.  A PMD_SIZE huge pages may exist as used
 309          * by THP.  Since we can not reliably insert a zero page, this
 310          * feature is not supported.
 311          */
 312         if (mode == MCOPY_ATOMIC_ZEROPAGE) {
 313                 mmap_read_unlock(dst_mm);
 314                 return -EINVAL;
 315         }
 316
 317         src_addr = src_start;
 318         dst_addr = dst_start;
 319         copied = 0;
 320         page = NULL;
 321         vma_hpagesize = vma_kernel_pagesize(dst_vma);
 322
 323         /*
 324          * Validate alignment based on huge page size
 325          */
 326         err = -EINVAL;
 327         if (dst_start & (vma_hpagesize - 1) || len & (vma_hpagesize - 1))
 328                 goto out_unlock;
 329
 330 retry:
 331         /*
 332          * On routine entry dst_vma is set.  If we had to drop mmap_lock and
 333          * retry, dst_vma will be set to NULL and we must lookup again.
 334          */
 335         if (!dst_vma) {
 336                 err = -ENOENT;
 337                 dst_vma = find_dst_vma(dst_mm, dst_start, len);
 338                 if (!dst_vma || !is_vm_hugetlb_page(dst_vma))
 339                         goto out_unlock;
 340
 341                 err = -EINVAL;
 342                 if (vma_hpagesize != vma_kernel_pagesize(dst_vma))
 343                         goto out_unlock;
 344
 345                 vm_shared = dst_vma->vm_flags & VM_SHARED;
 346         }
 347
 348         /*
 349          * If not shared, ensure the dst_vma has a anon_vma.
 350          */
 351         err = -ENOMEM;
 352         if (!vm_shared) {
 353                 if (unlikely(anon_vma_prepare(dst_vma)))
 354                         goto out_unlock;
 355         }
 356
 357         while (src_addr < src_start + len) {
 358                 BUG_ON(dst_addr >= dst_start + len);
 359
 360                 /*
 361                  * Serialize via i_mmap_rwsem and hugetlb_fault_mutex.
 362                  * i_mmap_rwsem ensures the dst_pte remains valid even
 363                  * in the case of shared pmds.  fault mutex prevents
 364                  * races with other faulting threads.
 365                  */
 366                 mapping = dst_vma->vm_file->f_mapping;
 367                 i_mmap_lock_read(mapping);
 368                 idx = linear_page_index(dst_vma, dst_addr);
 369                 hash = hugetlb_fault_mutex_hash(mapping, idx);
 370                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 371
 372                 err = -ENOMEM;
 373                 dst_pte = huge_pte_alloc(dst_mm, dst_vma, dst_addr, vma_hpagesize);
 374                 if (!dst_pte) {
 375                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 376                         i_mmap_unlock_read(mapping);
 377                         goto out_unlock;
 378                 }
 379
 380                 if (mode != MCOPY_ATOMIC_CONTINUE &&
 381                     !huge_pte_none(huge_ptep_get(dst_pte))) {
 382                         err = -EEXIST;
 383                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 384                         i_mmap_unlock_read(mapping);
 385                         goto out_unlock;
 386                 }
 387
 388                 err = hugetlb_mcopy_atomic_pte(dst_mm, dst_pte, dst_vma,
 389                                                dst_addr, src_addr, mode, &page);
 390
 391                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 392                 i_mmap_unlock_read(mapping);
 393
 394                 cond_resched();
 395
 396                 if (unlikely(err == -ENOENT)) {
 397                         mmap_read_unlock(dst_mm);
 398                         BUG_ON(!page);
 399
 400                         err = copy_huge_page_from_user(page,
 401                                                 (const void __user *)src_addr,
 402                                                 vma_hpagesize / PAGE_SIZE,
 403                                                 true);
 404                         if (unlikely(err)) {
 405                                 err = -EFAULT;
 406                                 goto out;
 407                         }
 408                         mmap_read_lock(dst_mm);
 409
 410                         dst_vma = NULL;
 411                         goto retry;
 412                 } else
 413                         BUG_ON(page);
 414
 415                 if (!err) {
 416                         dst_addr += vma_hpagesize;
 417                         src_addr += vma_hpagesize;
 418                         copied += vma_hpagesize;
 419
 420                         if (fatal_signal_pending(current))
 421                                 err = -EINTR;
 422                 }
 423                 if (err)
 424                         break;
 425         }
 426
 427 out_unlock:
 428         mmap_read_unlock(dst_mm);
 429 out:
 430         if (page)
 431                 put_page(page);
 432         BUG_ON(copied < 0);
 433         BUG_ON(err > 0);
 434         BUG_ON(!copied && !err);
 435         return copied ? copied : err;
 436 }
 437 #else /* !CONFIG_HUGETLB_PAGE */
 438 /* fail at build time if gcc attempts to use this */
 439 extern ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
 440                                       struct vm_area_struct *dst_vma,
 441                                       unsigned long dst_start,
 442                                       unsigned long src_start,
 443                                       unsigned long len,
 444                                       enum mcopy_atomic_mode mode);
 445 #endif /* CONFIG_HUGETLB_PAGE */
 446
 447 static __always_inline ssize_t mfill_atomic_pte(struct mm_struct *dst_mm,
 448                                                 pmd_t *dst_pmd,
 449                                                 struct vm_area_struct *dst_vma,
 450                                                 unsigned long dst_addr,
 451                                                 unsigned long src_addr,
 452                                                 struct page **page,
 453                                                 enum mcopy_atomic_mode mode,
 454                                                 bool wp_copy)
 455 {
 456         ssize_t err;
 457
 458         if (mode == MCOPY_ATOMIC_CONTINUE) {
 459                 return mcontinue_atomic_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
 460                                             wp_copy);
 461         }
 462
 463         /*
 464          * The normal page fault path for a shmem will invoke the
 465          * fault, fill the hole in the file and COW it right away. The
 466          * result generates plain anonymous memory. So when we are
 467          * asked to fill an hole in a MAP_PRIVATE shmem mapping, we'll
 468          * generate anonymous memory directly without actually filling
 469          * the hole. For the MAP_PRIVATE case the robustness check
 470          * only happens in the pagetable (to verify it's still none)
 471          * and not in the radix tree.
 472          */
 473         if (!(dst_vma->vm_flags & VM_SHARED)) {
 474                 if (mode == MCOPY_ATOMIC_NORMAL)
 475                         err = mcopy_atomic_pte(dst_mm, dst_pmd, dst_vma,
 476                                                dst_addr, src_addr, page,
 477                                                wp_copy);
 478                 else
 479                         err = mfill_zeropage_pte(dst_mm, dst_pmd,
 480                                                  dst_vma, dst_addr);
 481         } else {
 482                 VM_WARN_ON_ONCE(wp_copy);
 483                 err = shmem_mfill_atomic_pte(dst_mm, dst_pmd, dst_vma,
 484                                              dst_addr, src_addr,
 485                                              mode != MCOPY_ATOMIC_NORMAL,
 486                                              page);
 487         }
 488
 489         return err;
 490 }
 491
 492 static __always_inline ssize_t __mcopy_atomic(struct mm_struct *dst_mm,
 493                                               unsigned long dst_start,
 494                                               unsigned long src_start,
 495                                               unsigned long len,
 496                                               enum mcopy_atomic_mode mcopy_mode,
 497                                               atomic_t *mmap_changing,
 498                                               __u64 mode)
 499 {
 500         struct vm_area_struct *dst_vma;
 501         ssize_t err;
 502         pmd_t *dst_pmd;
 503         unsigned long src_addr, dst_addr;
 504         long copied;
 505         struct page *page;
 506         bool wp_copy;
 507
 508         /*
 509          * Sanitize the command parameters:
 510          */
 511         BUG_ON(dst_start & ~PAGE_MASK);
 512         BUG_ON(len & ~PAGE_MASK);
 513
 514         /* Does the address range wrap, or is the span zero-sized? */
 515         BUG_ON(src_start + len <= src_start);
 516         BUG_ON(dst_start + len <= dst_start);
 517
 518         src_addr = src_start;
 519         dst_addr = dst_start;
 520         copied = 0;
 521         page = NULL;
 522 retry:
 523         mmap_read_lock(dst_mm);
 524
 525         /*
 526          * If memory mappings are changing because of non-cooperative
 527          * operation (e.g. mremap) running in parallel, bail out and
 528          * request the user to retry later
 529          */
 530         err = -EAGAIN;
 531         if (mmap_changing && atomic_read(mmap_changing))
 532                 goto out_unlock;
 533
 534         /*
 535          * Make sure the vma is not shared, that the dst range is
 536          * both valid and fully within a single existing vma.
 537          */
 538         err = -ENOENT;
 539         dst_vma = find_dst_vma(dst_mm, dst_start, len);
 540         if (!dst_vma)
 541                 goto out_unlock;
 542
 543         err = -EINVAL;
 544         /*
 545          * shmem_zero_setup is invoked in mmap for MAP_ANONYMOUS|MAP_SHARED but
 546          * it will overwrite vm_ops, so vma_is_anonymous must return false.
 547          */
 548         if (WARN_ON_ONCE(vma_is_anonymous(dst_vma) &&
 549             dst_vma->vm_flags & VM_SHARED))
 550                 goto out_unlock;
 551
 552         /*
 553          * validate 'mode' now that we know the dst_vma: don't allow
 554          * a wrprotect copy if the userfaultfd didn't register as WP.
 555          */
 556         wp_copy = mode & UFFDIO_COPY_MODE_WP;
 557         if (wp_copy && !(dst_vma->vm_flags & VM_UFFD_WP))
 558                 goto out_unlock;
 559
 560         /*
 561          * If this is a HUGETLB vma, pass off to appropriate routine
 562          */
 563         if (is_vm_hugetlb_page(dst_vma))
 564                 return  __mcopy_atomic_hugetlb(dst_mm, dst_vma, dst_start,
 565                                                 src_start, len, mcopy_mode);
 566
 567         if (!vma_is_anonymous(dst_vma) && !vma_is_shmem(dst_vma))
 568                 goto out_unlock;
 569         if (!vma_is_shmem(dst_vma) && mcopy_mode == MCOPY_ATOMIC_CONTINUE)
 570                 goto out_unlock;
 571
 572         /*
 573          * Ensure the dst_vma has a anon_vma or this page
 574          * would get a NULL anon_vma when moved in the
 575          * dst_vma.
 576          */
 577         err = -ENOMEM;
 578         if (!(dst_vma->vm_flags & VM_SHARED) &&
 579             unlikely(anon_vma_prepare(dst_vma)))
 580                 goto out_unlock;
 581
 582         while (src_addr < src_start + len) {
 583                 pmd_t dst_pmdval;
 584
 585                 BUG_ON(dst_addr >= dst_start + len);
 586
 587                 dst_pmd = mm_alloc_pmd(dst_mm, dst_addr);
 588                 if (unlikely(!dst_pmd)) {
 589                         err = -ENOMEM;
 590                         break;
 591                 }
 592
 593                 dst_pmdval = pmd_read_atomic(dst_pmd);
 594                 /*
 595                  * If the dst_pmd is mapped as THP don't
 596                  * override it and just be strict.
 597                  */
 598                 if (unlikely(pmd_trans_huge(dst_pmdval))) {
 599                         err = -EEXIST;
 600                         break;
 601                 }
 602                 if (unlikely(pmd_none(dst_pmdval)) &&
 603                     unlikely(__pte_alloc(dst_mm, dst_pmd))) {
 604                         err = -ENOMEM;
 605                         break;
 606                 }
 607                 /* If an huge pmd materialized from under us fail */
 608                 if (unlikely(pmd_trans_huge(*dst_pmd))) {
 609                         err = -EFAULT;
 610                         break;
 611                 }
 612
 613                 BUG_ON(pmd_none(*dst_pmd));
 614                 BUG_ON(pmd_trans_huge(*dst_pmd));
 615
 616                 err = mfill_atomic_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
 617                                        src_addr, &page, mcopy_mode, wp_copy);
 618                 cond_resched();
 619
 620                 if (unlikely(err == -ENOENT)) {
 621                         void *page_kaddr;
 622
 623                         mmap_read_unlock(dst_mm);
 624                         BUG_ON(!page);
 625
 626                         page_kaddr = kmap(page);
 627                         err = copy_from_user(page_kaddr,
 628                                              (const void __user *) src_addr,
 629                                              PAGE_SIZE);
 630                         kunmap(page);
 631                         if (unlikely(err)) {
 632                                 err = -EFAULT;
 633                                 goto out;
 634                         }
 635                         flush_dcache_page(page);
 636                         goto retry;
 637                 } else
 638                         BUG_ON(page);
 639
 640                 if (!err) {
 641                         dst_addr += PAGE_SIZE;
 642                         src_addr += PAGE_SIZE;
 643                         copied += PAGE_SIZE;
 644
 645                         if (fatal_signal_pending(current))
 646                                 err = -EINTR;
 647                 }
 648                 if (err)
 649                         break;
 650         }
 651
 652 out_unlock:
 653         mmap_read_unlock(dst_mm);
 654 out:
 655         if (page)
 656                 put_page(page);
 657         BUG_ON(copied < 0);
 658         BUG_ON(err > 0);
 659         BUG_ON(!copied && !err);
 660         return copied ? copied : err;
 661 }
 662
 663 ssize_t mcopy_atomic(struct mm_struct *dst_mm, unsigned long dst_start,
 664                      unsigned long src_start, unsigned long len,
 665                      atomic_t *mmap_changing, __u64 mode)
 666 {
 667         return __mcopy_atomic(dst_mm, dst_start, src_start, len,
 668                               MCOPY_ATOMIC_NORMAL, mmap_changing, mode);
 669 }
 670
 671 ssize_t mfill_zeropage(struct mm_struct *dst_mm, unsigned long start,
 672                        unsigned long len, atomic_t *mmap_changing)
 673 {
 674         return __mcopy_atomic(dst_mm, start, 0, len, MCOPY_ATOMIC_ZEROPAGE,
 675                               mmap_changing, 0);
 676 }
 677
 678 ssize_t mcopy_continue(struct mm_struct *dst_mm, unsigned long start,
 679                        unsigned long len, atomic_t *mmap_changing)
 680 {
 681         return __mcopy_atomic(dst_mm, start, 0, len, MCOPY_ATOMIC_CONTINUE,
 682                               mmap_changing, 0);
 683 }
 684
 685 int mwriteprotect_range(struct mm_struct *dst_mm, unsigned long start,
 686                         unsigned long len, bool enable_wp,
 687                         atomic_t *mmap_changing)
 688 {
 689         struct vm_area_struct *dst_vma;
 690         pgprot_t newprot;
 691         int err;
 692
 693         /*
 694          * Sanitize the command parameters:
 695          */
 696         BUG_ON(start & ~PAGE_MASK);
 697         BUG_ON(len & ~PAGE_MASK);
 698
 699         /* Does the address range wrap, or is the span zero-sized? */
 700         BUG_ON(start + len <= start);
 701
 702         mmap_read_lock(dst_mm);
 703
 704         /*
 705          * If memory mappings are changing because of non-cooperative
 706          * operation (e.g. mremap) running in parallel, bail out and
 707          * request the user to retry later
 708          */
 709         err = -EAGAIN;
 710         if (mmap_changing && atomic_read(mmap_changing))
 711                 goto out_unlock;
 712
 713         err = -ENOENT;
 714         dst_vma = find_dst_vma(dst_mm, start, len);
 715         /*
 716          * Make sure the vma is not shared, that the dst range is
 717          * both valid and fully within a single existing vma.
 718          */
 719         if (!dst_vma || (dst_vma->vm_flags & VM_SHARED))
 720                 goto out_unlock;
 721         if (!userfaultfd_wp(dst_vma))
 722                 goto out_unlock;
 723         if (!vma_is_anonymous(dst_vma))
 724                 goto out_unlock;
 725
 726         if (enable_wp)
 727                 newprot = vm_get_page_prot(dst_vma->vm_flags & ~(VM_WRITE));
 728         else
 729                 newprot = vm_get_page_prot(dst_vma->vm_flags);
 730
 731         change_protection(dst_vma, start, start + len, newprot,
 732                           enable_wp ? MM_CP_UFFD_WP : MM_CP_UFFD_WP_RESOLVE);
 733
 734         err = 0;
 735 out_unlock:
 736         mmap_read_unlock(dst_mm);
 737         return err;
 738 }