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(page);
  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_NOALLOC);
 231         /* Our caller expects us to return -EFAULT if we failed to find page. */
 232         if (ret == -ENOENT)
 233                 ret = -EFAULT;
 234         if (ret)
 235                 goto out;
 236         if (!page) {
 237                 ret = -EFAULT;
 238                 goto out;
 239         }
 240
 241         if (PageHWPoison(page)) {
 242                 ret = -EIO;
 243                 goto out_release;
 244         }
 245
 246         ret = mfill_atomic_install_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
 247                                        page, false, wp_copy);
 248         if (ret)
 249                 goto out_release;
 250
 251         unlock_page(page);
 252         ret = 0;
 253 out:
 254         return ret;
 255 out_release:
 256         unlock_page(page);
 257         put_page(page);
 258         goto out;
 259 }
 260
 261 static pmd_t *mm_alloc_pmd(struct mm_struct *mm, unsigned long address)
 262 {
 263         pgd_t *pgd;
 264         p4d_t *p4d;
 265         pud_t *pud;
 266
 267         pgd = pgd_offset(mm, address);
 268         p4d = p4d_alloc(mm, pgd, address);
 269         if (!p4d)
 270                 return NULL;
 271         pud = pud_alloc(mm, p4d, address);
 272         if (!pud)
 273                 return NULL;
 274         /*
 275          * Note that we didn't run this because the pmd was
 276          * missing, the *pmd may be already established and in
 277          * turn it may also be a trans_huge_pmd.
 278          */
 279         return pmd_alloc(mm, pud, address);
 280 }
 281
 282 #ifdef CONFIG_HUGETLB_PAGE
 283 /*
 284  * __mcopy_atomic processing for HUGETLB vmas.  Note that this routine is
 285  * called with mmap_lock held, it will release mmap_lock before returning.
 286  */
 287 static __always_inline ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
 288                                               struct vm_area_struct *dst_vma,
 289                                               unsigned long dst_start,
 290                                               unsigned long src_start,
 291                                               unsigned long len,
 292                                               enum mcopy_atomic_mode mode)
 293 {
 294         int vm_shared = dst_vma->vm_flags & VM_SHARED;
 295         ssize_t err;
 296         pte_t *dst_pte;
 297         unsigned long src_addr, dst_addr;
 298         long copied;
 299         struct page *page;
 300         unsigned long vma_hpagesize;
 301         pgoff_t idx;
 302         u32 hash;
 303         struct address_space *mapping;
 304
 305         /*
 306          * There is no default zero huge page for all huge page sizes as
 307          * supported by hugetlb.  A PMD_SIZE huge pages may exist as used
 308          * by THP.  Since we can not reliably insert a zero page, this
 309          * feature is not supported.
 310          */
 311         if (mode == MCOPY_ATOMIC_ZEROPAGE) {
 312                 mmap_read_unlock(dst_mm);
 313                 return -EINVAL;
 314         }
 315
 316         src_addr = src_start;
 317         dst_addr = dst_start;
 318         copied = 0;
 319         page = NULL;
 320         vma_hpagesize = vma_kernel_pagesize(dst_vma);
 321
 322         /*
 323          * Validate alignment based on huge page size
 324          */
 325         err = -EINVAL;
 326         if (dst_start & (vma_hpagesize - 1) || len & (vma_hpagesize - 1))
 327                 goto out_unlock;
 328
 329 retry:
 330         /*
 331          * On routine entry dst_vma is set.  If we had to drop mmap_lock and
 332          * retry, dst_vma will be set to NULL and we must lookup again.
 333          */
 334         if (!dst_vma) {
 335                 err = -ENOENT;
 336                 dst_vma = find_dst_vma(dst_mm, dst_start, len);
 337                 if (!dst_vma || !is_vm_hugetlb_page(dst_vma))
 338                         goto out_unlock;
 339
 340                 err = -EINVAL;
 341                 if (vma_hpagesize != vma_kernel_pagesize(dst_vma))
 342                         goto out_unlock;
 343
 344                 vm_shared = dst_vma->vm_flags & VM_SHARED;
 345         }
 346
 347         /*
 348          * If not shared, ensure the dst_vma has a anon_vma.
 349          */
 350         err = -ENOMEM;
 351         if (!vm_shared) {
 352                 if (unlikely(anon_vma_prepare(dst_vma)))
 353                         goto out_unlock;
 354         }
 355
 356         while (src_addr < src_start + len) {
 357                 BUG_ON(dst_addr >= dst_start + len);
 358
 359                 /*
 360                  * Serialize via i_mmap_rwsem and hugetlb_fault_mutex.
 361                  * i_mmap_rwsem ensures the dst_pte remains valid even
 362                  * in the case of shared pmds.  fault mutex prevents
 363                  * races with other faulting threads.
 364                  */
 365                 mapping = dst_vma->vm_file->f_mapping;
 366                 i_mmap_lock_read(mapping);
 367                 idx = linear_page_index(dst_vma, dst_addr);
 368                 hash = hugetlb_fault_mutex_hash(mapping, idx);
 369                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 370
 371                 err = -ENOMEM;
 372                 dst_pte = huge_pte_alloc(dst_mm, dst_vma, dst_addr, vma_hpagesize);
 373                 if (!dst_pte) {
 374                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 375                         i_mmap_unlock_read(mapping);
 376                         goto out_unlock;
 377                 }
 378
 379                 if (mode != MCOPY_ATOMIC_CONTINUE &&
 380                     !huge_pte_none(huge_ptep_get(dst_pte))) {
 381                         err = -EEXIST;
 382                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 383                         i_mmap_unlock_read(mapping);
 384                         goto out_unlock;
 385                 }
 386
 387                 err = hugetlb_mcopy_atomic_pte(dst_mm, dst_pte, dst_vma,
 388                                                dst_addr, src_addr, mode, &page);
 389
 390                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 391                 i_mmap_unlock_read(mapping);
 392
 393                 cond_resched();
 394
 395                 if (unlikely(err == -ENOENT)) {
 396                         mmap_read_unlock(dst_mm);
 397                         BUG_ON(!page);
 398
 399                         err = copy_huge_page_from_user(page,
 400                                                 (const void __user *)src_addr,
 401                                                 vma_hpagesize / PAGE_SIZE,
 402                                                 true);
 403                         if (unlikely(err)) {
 404                                 err = -EFAULT;
 405                                 goto out;
 406                         }
 407                         mmap_read_lock(dst_mm);
 408
 409                         dst_vma = NULL;
 410                         goto retry;
 411                 } else
 412                         BUG_ON(page);
 413
 414                 if (!err) {
 415                         dst_addr += vma_hpagesize;
 416                         src_addr += vma_hpagesize;
 417                         copied += vma_hpagesize;
 418
 419                         if (fatal_signal_pending(current))
 420                                 err = -EINTR;
 421                 }
 422                 if (err)
 423                         break;
 424         }
 425
 426 out_unlock:
 427         mmap_read_unlock(dst_mm);
 428 out:
 429         if (page)
 430                 put_page(page);
 431         BUG_ON(copied < 0);
 432         BUG_ON(err > 0);
 433         BUG_ON(!copied && !err);
 434         return copied ? copied : err;
 435 }
 436 #else /* !CONFIG_HUGETLB_PAGE */
 437 /* fail at build time if gcc attempts to use this */
 438 extern ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
 439                                       struct vm_area_struct *dst_vma,
 440                                       unsigned long dst_start,
 441                                       unsigned long src_start,
 442                                       unsigned long len,
 443                                       enum mcopy_atomic_mode mode);
 444 #endif /* CONFIG_HUGETLB_PAGE */
 445
 446 static __always_inline ssize_t mfill_atomic_pte(struct mm_struct *dst_mm,
 447                                                 pmd_t *dst_pmd,
 448                                                 struct vm_area_struct *dst_vma,
 449                                                 unsigned long dst_addr,
 450                                                 unsigned long src_addr,
 451                                                 struct page **page,
 452                                                 enum mcopy_atomic_mode mode,
 453                                                 bool wp_copy)
 454 {
 455         ssize_t err;
 456
 457         if (mode == MCOPY_ATOMIC_CONTINUE) {
 458                 return mcontinue_atomic_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
 459                                             wp_copy);
 460         }
 461
 462         /*
 463          * The normal page fault path for a shmem will invoke the
 464          * fault, fill the hole in the file and COW it right away. The
 465          * result generates plain anonymous memory. So when we are
 466          * asked to fill an hole in a MAP_PRIVATE shmem mapping, we'll
 467          * generate anonymous memory directly without actually filling
 468          * the hole. For the MAP_PRIVATE case the robustness check
 469          * only happens in the pagetable (to verify it's still none)
 470          * and not in the radix tree.
 471          */
 472         if (!(dst_vma->vm_flags & VM_SHARED)) {
 473                 if (mode == MCOPY_ATOMIC_NORMAL)
 474                         err = mcopy_atomic_pte(dst_mm, dst_pmd, dst_vma,
 475                                                dst_addr, src_addr, page,
 476                                                wp_copy);
 477                 else
 478                         err = mfill_zeropage_pte(dst_mm, dst_pmd,
 479                                                  dst_vma, dst_addr);
 480         } else {
 481                 VM_WARN_ON_ONCE(wp_copy);
 482                 err = shmem_mfill_atomic_pte(dst_mm, dst_pmd, dst_vma,
 483                                              dst_addr, src_addr,
 484                                              mode != MCOPY_ATOMIC_NORMAL,
 485                                              page);
 486         }
 487
 488         return err;
 489 }
 490
 491 static __always_inline ssize_t __mcopy_atomic(struct mm_struct *dst_mm,
 492                                               unsigned long dst_start,
 493                                               unsigned long src_start,
 494                                               unsigned long len,
 495                                               enum mcopy_atomic_mode mcopy_mode,
 496                                               atomic_t *mmap_changing,
 497                                               __u64 mode)
 498 {
 499         struct vm_area_struct *dst_vma;
 500         ssize_t err;
 501         pmd_t *dst_pmd;
 502         unsigned long src_addr, dst_addr;
 503         long copied;
 504         struct page *page;
 505         bool wp_copy;
 506
 507         /*
 508          * Sanitize the command parameters:
 509          */
 510         BUG_ON(dst_start & ~PAGE_MASK);
 511         BUG_ON(len & ~PAGE_MASK);
 512
 513         /* Does the address range wrap, or is the span zero-sized? */
 514         BUG_ON(src_start + len <= src_start);
 515         BUG_ON(dst_start + len <= dst_start);
 516
 517         src_addr = src_start;
 518         dst_addr = dst_start;
 519         copied = 0;
 520         page = NULL;
 521 retry:
 522         mmap_read_lock(dst_mm);
 523
 524         /*
 525          * If memory mappings are changing because of non-cooperative
 526          * operation (e.g. mremap) running in parallel, bail out and
 527          * request the user to retry later
 528          */
 529         err = -EAGAIN;
 530         if (mmap_changing && atomic_read(mmap_changing))
 531                 goto out_unlock;
 532
 533         /*
 534          * Make sure the vma is not shared, that the dst range is
 535          * both valid and fully within a single existing vma.
 536          */
 537         err = -ENOENT;
 538         dst_vma = find_dst_vma(dst_mm, dst_start, len);
 539         if (!dst_vma)
 540                 goto out_unlock;
 541
 542         err = -EINVAL;
 543         /*
 544          * shmem_zero_setup is invoked in mmap for MAP_ANONYMOUS|MAP_SHARED but
 545          * it will overwrite vm_ops, so vma_is_anonymous must return false.
 546          */
 547         if (WARN_ON_ONCE(vma_is_anonymous(dst_vma) &&
 548             dst_vma->vm_flags & VM_SHARED))
 549                 goto out_unlock;
 550
 551         /*
 552          * validate 'mode' now that we know the dst_vma: don't allow
 553          * a wrprotect copy if the userfaultfd didn't register as WP.
 554          */
 555         wp_copy = mode & UFFDIO_COPY_MODE_WP;
 556         if (wp_copy && !(dst_vma->vm_flags & VM_UFFD_WP))
 557                 goto out_unlock;
 558
 559         /*
 560          * If this is a HUGETLB vma, pass off to appropriate routine
 561          */
 562         if (is_vm_hugetlb_page(dst_vma))
 563                 return  __mcopy_atomic_hugetlb(dst_mm, dst_vma, dst_start,
 564                                                 src_start, len, mcopy_mode);
 565
 566         if (!vma_is_anonymous(dst_vma) && !vma_is_shmem(dst_vma))
 567                 goto out_unlock;
 568         if (!vma_is_shmem(dst_vma) && mcopy_mode == MCOPY_ATOMIC_CONTINUE)
 569                 goto out_unlock;
 570
 571         /*
 572          * Ensure the dst_vma has a anon_vma or this page
 573          * would get a NULL anon_vma when moved in the
 574          * dst_vma.
 575          */
 576         err = -ENOMEM;
 577         if (!(dst_vma->vm_flags & VM_SHARED) &&
 578             unlikely(anon_vma_prepare(dst_vma)))
 579                 goto out_unlock;
 580
 581         while (src_addr < src_start + len) {
 582                 pmd_t dst_pmdval;
 583
 584                 BUG_ON(dst_addr >= dst_start + len);
 585
 586                 dst_pmd = mm_alloc_pmd(dst_mm, dst_addr);
 587                 if (unlikely(!dst_pmd)) {
 588                         err = -ENOMEM;
 589                         break;
 590                 }
 591
 592                 dst_pmdval = pmd_read_atomic(dst_pmd);
 593                 /*
 594                  * If the dst_pmd is mapped as THP don't
 595                  * override it and just be strict.
 596                  */
 597                 if (unlikely(pmd_trans_huge(dst_pmdval))) {
 598                         err = -EEXIST;
 599                         break;
 600                 }
 601                 if (unlikely(pmd_none(dst_pmdval)) &&
 602                     unlikely(__pte_alloc(dst_mm, dst_pmd))) {
 603                         err = -ENOMEM;
 604                         break;
 605                 }
 606                 /* If an huge pmd materialized from under us fail */
 607                 if (unlikely(pmd_trans_huge(*dst_pmd))) {
 608                         err = -EFAULT;
 609                         break;
 610                 }
 611
 612                 BUG_ON(pmd_none(*dst_pmd));
 613                 BUG_ON(pmd_trans_huge(*dst_pmd));
 614
 615                 err = mfill_atomic_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
 616                                        src_addr, &page, mcopy_mode, wp_copy);
 617                 cond_resched();
 618
 619                 if (unlikely(err == -ENOENT)) {
 620                         void *page_kaddr;
 621
 622                         mmap_read_unlock(dst_mm);
 623                         BUG_ON(!page);
 624
 625                         page_kaddr = kmap(page);
 626                         err = copy_from_user(page_kaddr,
 627                                              (const void __user *) src_addr,
 628                                              PAGE_SIZE);
 629                         kunmap(page);
 630                         if (unlikely(err)) {
 631                                 err = -EFAULT;
 632                                 goto out;
 633                         }
 634                         flush_dcache_page(page);
 635                         goto retry;
 636                 } else
 637                         BUG_ON(page);
 638
 639                 if (!err) {
 640                         dst_addr += PAGE_SIZE;
 641                         src_addr += PAGE_SIZE;
 642                         copied += PAGE_SIZE;
 643
 644                         if (fatal_signal_pending(current))
 645                                 err = -EINTR;
 646                 }
 647                 if (err)
 648                         break;
 649         }
 650
 651 out_unlock:
 652         mmap_read_unlock(dst_mm);
 653 out:
 654         if (page)
 655                 put_page(page);
 656         BUG_ON(copied < 0);
 657         BUG_ON(err > 0);
 658         BUG_ON(!copied && !err);
 659         return copied ? copied : err;
 660 }
 661
 662 ssize_t mcopy_atomic(struct mm_struct *dst_mm, unsigned long dst_start,
 663                      unsigned long src_start, unsigned long len,
 664                      atomic_t *mmap_changing, __u64 mode)
 665 {
 666         return __mcopy_atomic(dst_mm, dst_start, src_start, len,
 667                               MCOPY_ATOMIC_NORMAL, mmap_changing, mode);
 668 }
 669
 670 ssize_t mfill_zeropage(struct mm_struct *dst_mm, unsigned long start,
 671                        unsigned long len, atomic_t *mmap_changing)
 672 {
 673         return __mcopy_atomic(dst_mm, start, 0, len, MCOPY_ATOMIC_ZEROPAGE,
 674                               mmap_changing, 0);
 675 }
 676
 677 ssize_t mcopy_continue(struct mm_struct *dst_mm, unsigned long start,
 678                        unsigned long len, atomic_t *mmap_changing)
 679 {
 680         return __mcopy_atomic(dst_mm, start, 0, len, MCOPY_ATOMIC_CONTINUE,
 681                               mmap_changing, 0);
 682 }
 683
 684 int mwriteprotect_range(struct mm_struct *dst_mm, unsigned long start,
 685                         unsigned long len, bool enable_wp,
 686                         atomic_t *mmap_changing)
 687 {
 688         struct vm_area_struct *dst_vma;
 689         pgprot_t newprot;
 690         int err;
 691
 692         /*
 693          * Sanitize the command parameters:
 694          */
 695         BUG_ON(start & ~PAGE_MASK);
 696         BUG_ON(len & ~PAGE_MASK);
 697
 698         /* Does the address range wrap, or is the span zero-sized? */
 699         BUG_ON(start + len <= start);
 700
 701         mmap_read_lock(dst_mm);
 702
 703         /*
 704          * If memory mappings are changing because of non-cooperative
 705          * operation (e.g. mremap) running in parallel, bail out and
 706          * request the user to retry later
 707          */
 708         err = -EAGAIN;
 709         if (mmap_changing && atomic_read(mmap_changing))
 710                 goto out_unlock;
 711
 712         err = -ENOENT;
 713         dst_vma = find_dst_vma(dst_mm, start, len);
 714         /*
 715          * Make sure the vma is not shared, that the dst range is
 716          * both valid and fully within a single existing vma.
 717          */
 718         if (!dst_vma || (dst_vma->vm_flags & VM_SHARED))
 719                 goto out_unlock;
 720         if (!userfaultfd_wp(dst_vma))
 721                 goto out_unlock;
 722         if (!vma_is_anonymous(dst_vma))
 723                 goto out_unlock;
 724
 725         if (enable_wp)
 726                 newprot = vm_get_page_prot(dst_vma->vm_flags & ~(VM_WRITE));
 727         else
 728                 newprot = vm_get_page_prot(dst_vma->vm_flags);
 729
 730         change_protection(dst_vma, start, start + len, newprot,
 731                           enable_wp ? MM_CP_UFFD_WP : MM_CP_UFFD_WP_RESOLVE);
 732
 733         err = 0;
 734 out_unlock:
 735         mmap_read_unlock(dst_mm);
 736         return err;
 737 }