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 <asm/tlb.h>
  20 #include "internal.h"
  21
  22 static __always_inline
  23 struct vm_area_struct *find_dst_vma(struct mm_struct *dst_mm,
  24                                     unsigned long dst_start,
  25                                     unsigned long len)
  26 {
  27         /*
  28          * Make sure that the dst range is both valid and fully within a
  29          * single existing vma.
  30          */
  31         struct vm_area_struct *dst_vma;
  32
  33         dst_vma = find_vma(dst_mm, dst_start);
  34         if (!dst_vma)
  35                 return NULL;
  36
  37         if (dst_start < dst_vma->vm_start ||
  38             dst_start + len > dst_vma->vm_end)
  39                 return NULL;
  40
  41         /*
  42          * Check the vma is registered in uffd, this is required to
  43          * enforce the VM_MAYWRITE check done at uffd registration
  44          * time.
  45          */
  46         if (!dst_vma->vm_userfaultfd_ctx.ctx)
  47                 return NULL;
  48
  49         return dst_vma;
  50 }
  51
  52 /*
  53  * Install PTEs, to map dst_addr (within dst_vma) to page.
  54  *
  55  * This function handles both MCOPY_ATOMIC_NORMAL and _CONTINUE for both shmem
  56  * and anon, and for both shared and private VMAs.
  57  */
  58 int mfill_atomic_install_pte(struct mm_struct *dst_mm, pmd_t *dst_pmd,
  59                              struct vm_area_struct *dst_vma,
  60                              unsigned long dst_addr, struct page *page,
  61                              bool newly_allocated, bool wp_copy)
  62 {
  63         int ret;
  64         pte_t _dst_pte, *dst_pte;
  65         bool writable = dst_vma->vm_flags & VM_WRITE;
  66         bool vm_shared = dst_vma->vm_flags & VM_SHARED;
  67         bool page_in_cache = page->mapping;
  68         spinlock_t *ptl;
  69         struct inode *inode;
  70         pgoff_t offset, max_off;
  71
  72         _dst_pte = mk_pte(page, dst_vma->vm_page_prot);
  73         _dst_pte = pte_mkdirty(_dst_pte);
  74         if (page_in_cache && !vm_shared)
  75                 writable = false;
  76
  77         /*
  78          * Always mark a PTE as write-protected when needed, regardless of
  79          * VM_WRITE, which the user might change.
  80          */
  81         if (wp_copy) {
  82                 _dst_pte = pte_mkuffd_wp(_dst_pte);
  83                 writable = false;
  84         }
  85
  86         if (writable)
  87                 _dst_pte = pte_mkwrite(_dst_pte);
  88         else
  89                 /*
  90                  * We need this to make sure write bit removed; as mk_pte()
  91                  * could return a pte with write bit set.
  92                  */
  93                 _dst_pte = pte_wrprotect(_dst_pte);
  94
  95         dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
  96
  97         if (vma_is_shmem(dst_vma)) {
  98                 /* serialize against truncate with the page table lock */
  99                 inode = dst_vma->vm_file->f_inode;
 100                 offset = linear_page_index(dst_vma, dst_addr);
 101                 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
 102                 ret = -EFAULT;
 103                 if (unlikely(offset >= max_off))
 104                         goto out_unlock;
 105         }
 106
 107         ret = -EEXIST;
 108         /*
 109          * We allow to overwrite a pte marker: consider when both MISSING|WP
 110          * registered, we firstly wr-protect a none pte which has no page cache
 111          * page backing it, then access the page.
 112          */
 113         if (!pte_none_mostly(*dst_pte))
 114                 goto out_unlock;
 115
 116         if (page_in_cache) {
 117                 /* Usually, cache pages are already added to LRU */
 118                 if (newly_allocated)
 119                         lru_cache_add(page);
 120                 page_add_file_rmap(page, dst_vma, false);
 121         } else {
 122                 page_add_new_anon_rmap(page, dst_vma, dst_addr);
 123                 lru_cache_add_inactive_or_unevictable(page, dst_vma);
 124         }
 125
 126         /*
 127          * Must happen after rmap, as mm_counter() checks mapping (via
 128          * PageAnon()), which is set by __page_set_anon_rmap().
 129          */
 130         inc_mm_counter(dst_mm, mm_counter(page));
 131
 132         set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
 133
 134         /* No need to invalidate - it was non-present before */
 135         update_mmu_cache(dst_vma, dst_addr, dst_pte);
 136         ret = 0;
 137 out_unlock:
 138         pte_unmap_unlock(dst_pte, ptl);
 139         return ret;
 140 }
 141
 142 static int mcopy_atomic_pte(struct mm_struct *dst_mm,
 143                             pmd_t *dst_pmd,
 144                             struct vm_area_struct *dst_vma,
 145                             unsigned long dst_addr,
 146                             unsigned long src_addr,
 147                             struct page **pagep,
 148                             bool wp_copy)
 149 {
 150         void *page_kaddr;
 151         int ret;
 152         struct page *page;
 153
 154         if (!*pagep) {
 155                 ret = -ENOMEM;
 156                 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, dst_vma, dst_addr);
 157                 if (!page)
 158                         goto out;
 159
 160                 page_kaddr = kmap_atomic(page);
 161                 ret = copy_from_user(page_kaddr,
 162                                      (const void __user *) src_addr,
 163                                      PAGE_SIZE);
 164                 kunmap_atomic(page_kaddr);
 165
 166                 /* fallback to copy_from_user outside mmap_lock */
 167                 if (unlikely(ret)) {
 168                         ret = -ENOENT;
 169                         *pagep = page;
 170                         /* don't free the page */
 171                         goto out;
 172                 }
 173
 174                 flush_dcache_page(page);
 175         } else {
 176                 page = *pagep;
 177                 *pagep = NULL;
 178         }
 179
 180         /*
 181          * The memory barrier inside __SetPageUptodate makes sure that
 182          * preceding stores to the page contents become visible before
 183          * the set_pte_at() write.
 184          */
 185         __SetPageUptodate(page);
 186
 187         ret = -ENOMEM;
 188         if (mem_cgroup_charge(page_folio(page), dst_mm, GFP_KERNEL))
 189                 goto out_release;
 190
 191         ret = mfill_atomic_install_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
 192                                        page, true, wp_copy);
 193         if (ret)
 194                 goto out_release;
 195 out:
 196         return ret;
 197 out_release:
 198         put_page(page);
 199         goto out;
 200 }
 201
 202 static int mfill_zeropage_pte(struct mm_struct *dst_mm,
 203                               pmd_t *dst_pmd,
 204                               struct vm_area_struct *dst_vma,
 205                               unsigned long dst_addr)
 206 {
 207         pte_t _dst_pte, *dst_pte;
 208         spinlock_t *ptl;
 209         int ret;
 210         pgoff_t offset, max_off;
 211         struct inode *inode;
 212
 213         _dst_pte = pte_mkspecial(pfn_pte(my_zero_pfn(dst_addr),
 214                                          dst_vma->vm_page_prot));
 215         dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
 216         if (dst_vma->vm_file) {
 217                 /* the shmem MAP_PRIVATE case requires checking the i_size */
 218                 inode = dst_vma->vm_file->f_inode;
 219                 offset = linear_page_index(dst_vma, dst_addr);
 220                 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
 221                 ret = -EFAULT;
 222                 if (unlikely(offset >= max_off))
 223                         goto out_unlock;
 224         }
 225         ret = -EEXIST;
 226         if (!pte_none(*dst_pte))
 227                 goto out_unlock;
 228         set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
 229         /* No need to invalidate - it was non-present before */
 230         update_mmu_cache(dst_vma, dst_addr, dst_pte);
 231         ret = 0;
 232 out_unlock:
 233         pte_unmap_unlock(dst_pte, ptl);
 234         return ret;
 235 }
 236
 237 /* Handles UFFDIO_CONTINUE for all shmem VMAs (shared or private). */
 238 static int mcontinue_atomic_pte(struct mm_struct *dst_mm,
 239                                 pmd_t *dst_pmd,
 240                                 struct vm_area_struct *dst_vma,
 241                                 unsigned long dst_addr,
 242                                 bool wp_copy)
 243 {
 244         struct inode *inode = file_inode(dst_vma->vm_file);
 245         pgoff_t pgoff = linear_page_index(dst_vma, dst_addr);
 246         struct page *page;
 247         int ret;
 248
 249         ret = shmem_getpage(inode, pgoff, &page, SGP_READ);
 250         if (ret)
 251                 goto out;
 252         if (!page) {
 253                 ret = -EFAULT;
 254                 goto out;
 255         }
 256
 257         if (PageHWPoison(page)) {
 258                 ret = -EIO;
 259                 goto out_release;
 260         }
 261
 262         ret = mfill_atomic_install_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
 263                                        page, false, wp_copy);
 264         if (ret)
 265                 goto out_release;
 266
 267         unlock_page(page);
 268         ret = 0;
 269 out:
 270         return ret;
 271 out_release:
 272         unlock_page(page);
 273         put_page(page);
 274         goto out;
 275 }
 276
 277 static pmd_t *mm_alloc_pmd(struct mm_struct *mm, unsigned long address)
 278 {
 279         pgd_t *pgd;
 280         p4d_t *p4d;
 281         pud_t *pud;
 282
 283         pgd = pgd_offset(mm, address);
 284         p4d = p4d_alloc(mm, pgd, address);
 285         if (!p4d)
 286                 return NULL;
 287         pud = pud_alloc(mm, p4d, address);
 288         if (!pud)
 289                 return NULL;
 290         /*
 291          * Note that we didn't run this because the pmd was
 292          * missing, the *pmd may be already established and in
 293          * turn it may also be a trans_huge_pmd.
 294          */
 295         return pmd_alloc(mm, pud, address);
 296 }
 297
 298 #ifdef CONFIG_HUGETLB_PAGE
 299 /*
 300  * __mcopy_atomic processing for HUGETLB vmas.  Note that this routine is
 301  * called with mmap_lock held, it will release mmap_lock before returning.
 302  */
 303 static __always_inline ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
 304                                               struct vm_area_struct *dst_vma,
 305                                               unsigned long dst_start,
 306                                               unsigned long src_start,
 307                                               unsigned long len,
 308                                               enum mcopy_atomic_mode mode,
 309                                               bool wp_copy)
 310 {
 311         int vm_shared = dst_vma->vm_flags & VM_SHARED;
 312         ssize_t err;
 313         pte_t *dst_pte;
 314         unsigned long src_addr, dst_addr;
 315         long copied;
 316         struct page *page;
 317         unsigned long vma_hpagesize;
 318         pgoff_t idx;
 319         u32 hash;
 320         struct address_space *mapping;
 321
 322         /*
 323          * There is no default zero huge page for all huge page sizes as
 324          * supported by hugetlb.  A PMD_SIZE huge pages may exist as used
 325          * by THP.  Since we can not reliably insert a zero page, this
 326          * feature is not supported.
 327          */
 328         if (mode == MCOPY_ATOMIC_ZEROPAGE) {
 329                 mmap_read_unlock(dst_mm);
 330                 return -EINVAL;
 331         }
 332
 333         src_addr = src_start;
 334         dst_addr = dst_start;
 335         copied = 0;
 336         page = NULL;
 337         vma_hpagesize = vma_kernel_pagesize(dst_vma);
 338
 339         /*
 340          * Validate alignment based on huge page size
 341          */
 342         err = -EINVAL;
 343         if (dst_start & (vma_hpagesize - 1) || len & (vma_hpagesize - 1))
 344                 goto out_unlock;
 345
 346 retry:
 347         /*
 348          * On routine entry dst_vma is set.  If we had to drop mmap_lock and
 349          * retry, dst_vma will be set to NULL and we must lookup again.
 350          */
 351         if (!dst_vma) {
 352                 err = -ENOENT;
 353                 dst_vma = find_dst_vma(dst_mm, dst_start, len);
 354                 if (!dst_vma || !is_vm_hugetlb_page(dst_vma))
 355                         goto out_unlock;
 356
 357                 err = -EINVAL;
 358                 if (vma_hpagesize != vma_kernel_pagesize(dst_vma))
 359                         goto out_unlock;
 360
 361                 vm_shared = dst_vma->vm_flags & VM_SHARED;
 362         }
 363
 364         /*
 365          * If not shared, ensure the dst_vma has a anon_vma.
 366          */
 367         err = -ENOMEM;
 368         if (!vm_shared) {
 369                 if (unlikely(anon_vma_prepare(dst_vma)))
 370                         goto out_unlock;
 371         }
 372
 373         while (src_addr < src_start + len) {
 374                 BUG_ON(dst_addr >= dst_start + len);
 375
 376                 /*
 377                  * Serialize via i_mmap_rwsem and hugetlb_fault_mutex.
 378                  * i_mmap_rwsem ensures the dst_pte remains valid even
 379                  * in the case of shared pmds.  fault mutex prevents
 380                  * races with other faulting threads.
 381                  */
 382                 mapping = dst_vma->vm_file->f_mapping;
 383                 i_mmap_lock_read(mapping);
 384                 idx = linear_page_index(dst_vma, dst_addr);
 385                 hash = hugetlb_fault_mutex_hash(mapping, idx);
 386                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 387
 388                 err = -ENOMEM;
 389                 dst_pte = huge_pte_alloc(dst_mm, dst_vma, dst_addr, vma_hpagesize);
 390                 if (!dst_pte) {
 391                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 392                         i_mmap_unlock_read(mapping);
 393                         goto out_unlock;
 394                 }
 395
 396                 if (mode != MCOPY_ATOMIC_CONTINUE &&
 397                     !huge_pte_none_mostly(huge_ptep_get(dst_pte))) {
 398                         err = -EEXIST;
 399                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 400                         i_mmap_unlock_read(mapping);
 401                         goto out_unlock;
 402                 }
 403
 404                 err = hugetlb_mcopy_atomic_pte(dst_mm, dst_pte, dst_vma,
 405                                                dst_addr, src_addr, mode, &page,
 406                                                wp_copy);
 407
 408                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 409                 i_mmap_unlock_read(mapping);
 410
 411                 cond_resched();
 412
 413                 if (unlikely(err == -ENOENT)) {
 414                         mmap_read_unlock(dst_mm);
 415                         BUG_ON(!page);
 416
 417                         err = copy_huge_page_from_user(page,
 418                                                 (const void __user *)src_addr,
 419                                                 vma_hpagesize / PAGE_SIZE,
 420                                                 true);
 421                         if (unlikely(err)) {
 422                                 err = -EFAULT;
 423                                 goto out;
 424                         }
 425                         mmap_read_lock(dst_mm);
 426
 427                         dst_vma = NULL;
 428                         goto retry;
 429                 } else
 430                         BUG_ON(page);
 431
 432                 if (!err) {
 433                         dst_addr += vma_hpagesize;
 434                         src_addr += vma_hpagesize;
 435                         copied += vma_hpagesize;
 436
 437                         if (fatal_signal_pending(current))
 438                                 err = -EINTR;
 439                 }
 440                 if (err)
 441                         break;
 442         }
 443
 444 out_unlock:
 445         mmap_read_unlock(dst_mm);
 446 out:
 447         if (page)
 448                 put_page(page);
 449         BUG_ON(copied < 0);
 450         BUG_ON(err > 0);
 451         BUG_ON(!copied && !err);
 452         return copied ? copied : err;
 453 }
 454 #else /* !CONFIG_HUGETLB_PAGE */
 455 /* fail at build time if gcc attempts to use this */
 456 extern ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
 457                                       struct vm_area_struct *dst_vma,
 458                                       unsigned long dst_start,
 459                                       unsigned long src_start,
 460                                       unsigned long len,
 461                                       enum mcopy_atomic_mode mode,
 462                                       bool wp_copy);
 463 #endif /* CONFIG_HUGETLB_PAGE */
 464
 465 static __always_inline ssize_t mfill_atomic_pte(struct mm_struct *dst_mm,
 466                                                 pmd_t *dst_pmd,
 467                                                 struct vm_area_struct *dst_vma,
 468                                                 unsigned long dst_addr,
 469                                                 unsigned long src_addr,
 470                                                 struct page **page,
 471                                                 enum mcopy_atomic_mode mode,
 472                                                 bool wp_copy)
 473 {
 474         ssize_t err;
 475
 476         if (mode == MCOPY_ATOMIC_CONTINUE) {
 477                 return mcontinue_atomic_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
 478                                             wp_copy);
 479         }
 480
 481         /*
 482          * The normal page fault path for a shmem will invoke the
 483          * fault, fill the hole in the file and COW it right away. The
 484          * result generates plain anonymous memory. So when we are
 485          * asked to fill an hole in a MAP_PRIVATE shmem mapping, we'll
 486          * generate anonymous memory directly without actually filling
 487          * the hole. For the MAP_PRIVATE case the robustness check
 488          * only happens in the pagetable (to verify it's still none)
 489          * and not in the radix tree.
 490          */
 491         if (!(dst_vma->vm_flags & VM_SHARED)) {
 492                 if (mode == MCOPY_ATOMIC_NORMAL)
 493                         err = mcopy_atomic_pte(dst_mm, dst_pmd, dst_vma,
 494                                                dst_addr, src_addr, page,
 495                                                wp_copy);
 496                 else
 497                         err = mfill_zeropage_pte(dst_mm, dst_pmd,
 498                                                  dst_vma, dst_addr);
 499         } else {
 500                 err = shmem_mfill_atomic_pte(dst_mm, dst_pmd, dst_vma,
 501                                              dst_addr, src_addr,
 502                                              mode != MCOPY_ATOMIC_NORMAL,
 503                                              wp_copy, page);
 504         }
 505
 506         return err;
 507 }
 508
 509 static __always_inline ssize_t __mcopy_atomic(struct mm_struct *dst_mm,
 510                                               unsigned long dst_start,
 511                                               unsigned long src_start,
 512                                               unsigned long len,
 513                                               enum mcopy_atomic_mode mcopy_mode,
 514                                               atomic_t *mmap_changing,
 515                                               __u64 mode)
 516 {
 517         struct vm_area_struct *dst_vma;
 518         ssize_t err;
 519         pmd_t *dst_pmd;
 520         unsigned long src_addr, dst_addr;
 521         long copied;
 522         struct page *page;
 523         bool wp_copy;
 524
 525         /*
 526          * Sanitize the command parameters:
 527          */
 528         BUG_ON(dst_start & ~PAGE_MASK);
 529         BUG_ON(len & ~PAGE_MASK);
 530
 531         /* Does the address range wrap, or is the span zero-sized? */
 532         BUG_ON(src_start + len <= src_start);
 533         BUG_ON(dst_start + len <= dst_start);
 534
 535         src_addr = src_start;
 536         dst_addr = dst_start;
 537         copied = 0;
 538         page = NULL;
 539 retry:
 540         mmap_read_lock(dst_mm);
 541
 542         /*
 543          * If memory mappings are changing because of non-cooperative
 544          * operation (e.g. mremap) running in parallel, bail out and
 545          * request the user to retry later
 546          */
 547         err = -EAGAIN;
 548         if (mmap_changing && atomic_read(mmap_changing))
 549                 goto out_unlock;
 550
 551         /*
 552          * Make sure the vma is not shared, that the dst range is
 553          * both valid and fully within a single existing vma.
 554          */
 555         err = -ENOENT;
 556         dst_vma = find_dst_vma(dst_mm, dst_start, len);
 557         if (!dst_vma)
 558                 goto out_unlock;
 559
 560         err = -EINVAL;
 561         /*
 562          * shmem_zero_setup is invoked in mmap for MAP_ANONYMOUS|MAP_SHARED but
 563          * it will overwrite vm_ops, so vma_is_anonymous must return false.
 564          */
 565         if (WARN_ON_ONCE(vma_is_anonymous(dst_vma) &&
 566             dst_vma->vm_flags & VM_SHARED))
 567                 goto out_unlock;
 568
 569         /*
 570          * validate 'mode' now that we know the dst_vma: don't allow
 571          * a wrprotect copy if the userfaultfd didn't register as WP.
 572          */
 573         wp_copy = mode & UFFDIO_COPY_MODE_WP;
 574         if (wp_copy && !(dst_vma->vm_flags & VM_UFFD_WP))
 575                 goto out_unlock;
 576
 577         /*
 578          * If this is a HUGETLB vma, pass off to appropriate routine
 579          */
 580         if (is_vm_hugetlb_page(dst_vma))
 581                 return  __mcopy_atomic_hugetlb(dst_mm, dst_vma, dst_start,
 582                                                src_start, len, mcopy_mode,
 583                                                wp_copy);
 584
 585         if (!vma_is_anonymous(dst_vma) && !vma_is_shmem(dst_vma))
 586                 goto out_unlock;
 587         if (!vma_is_shmem(dst_vma) && mcopy_mode == MCOPY_ATOMIC_CONTINUE)
 588                 goto out_unlock;
 589
 590         /*
 591          * Ensure the dst_vma has a anon_vma or this page
 592          * would get a NULL anon_vma when moved in the
 593          * dst_vma.
 594          */
 595         err = -ENOMEM;
 596         if (!(dst_vma->vm_flags & VM_SHARED) &&
 597             unlikely(anon_vma_prepare(dst_vma)))
 598                 goto out_unlock;
 599
 600         while (src_addr < src_start + len) {
 601                 pmd_t dst_pmdval;
 602
 603                 BUG_ON(dst_addr >= dst_start + len);
 604
 605                 dst_pmd = mm_alloc_pmd(dst_mm, dst_addr);
 606                 if (unlikely(!dst_pmd)) {
 607                         err = -ENOMEM;
 608                         break;
 609                 }
 610
 611                 dst_pmdval = pmd_read_atomic(dst_pmd);
 612                 /*
 613                  * If the dst_pmd is mapped as THP don't
 614                  * override it and just be strict.
 615                  */
 616                 if (unlikely(pmd_trans_huge(dst_pmdval))) {
 617                         err = -EEXIST;
 618                         break;
 619                 }
 620                 if (unlikely(pmd_none(dst_pmdval)) &&
 621                     unlikely(__pte_alloc(dst_mm, dst_pmd))) {
 622                         err = -ENOMEM;
 623                         break;
 624                 }
 625                 /* If an huge pmd materialized from under us fail */
 626                 if (unlikely(pmd_trans_huge(*dst_pmd))) {
 627                         err = -EFAULT;
 628                         break;
 629                 }
 630
 631                 BUG_ON(pmd_none(*dst_pmd));
 632                 BUG_ON(pmd_trans_huge(*dst_pmd));
 633
 634                 err = mfill_atomic_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
 635                                        src_addr, &page, mcopy_mode, wp_copy);
 636                 cond_resched();
 637
 638                 if (unlikely(err == -ENOENT)) {
 639                         void *page_kaddr;
 640
 641                         mmap_read_unlock(dst_mm);
 642                         BUG_ON(!page);
 643
 644                         page_kaddr = kmap(page);
 645                         err = copy_from_user(page_kaddr,
 646                                              (const void __user *) src_addr,
 647                                              PAGE_SIZE);
 648                         kunmap(page);
 649                         if (unlikely(err)) {
 650                                 err = -EFAULT;
 651                                 goto out;
 652                         }
 653                         flush_dcache_page(page);
 654                         goto retry;
 655                 } else
 656                         BUG_ON(page);
 657
 658                 if (!err) {
 659                         dst_addr += PAGE_SIZE;
 660                         src_addr += PAGE_SIZE;
 661                         copied += PAGE_SIZE;
 662
 663                         if (fatal_signal_pending(current))
 664                                 err = -EINTR;
 665                 }
 666                 if (err)
 667                         break;
 668         }
 669
 670 out_unlock:
 671         mmap_read_unlock(dst_mm);
 672 out:
 673         if (page)
 674                 put_page(page);
 675         BUG_ON(copied < 0);
 676         BUG_ON(err > 0);
 677         BUG_ON(!copied && !err);
 678         return copied ? copied : err;
 679 }
 680
 681 ssize_t mcopy_atomic(struct mm_struct *dst_mm, unsigned long dst_start,
 682                      unsigned long src_start, unsigned long len,
 683                      atomic_t *mmap_changing, __u64 mode)
 684 {
 685         return __mcopy_atomic(dst_mm, dst_start, src_start, len,
 686                               MCOPY_ATOMIC_NORMAL, mmap_changing, mode);
 687 }
 688
 689 ssize_t mfill_zeropage(struct mm_struct *dst_mm, unsigned long start,
 690                        unsigned long len, atomic_t *mmap_changing)
 691 {
 692         return __mcopy_atomic(dst_mm, start, 0, len, MCOPY_ATOMIC_ZEROPAGE,
 693                               mmap_changing, 0);
 694 }
 695
 696 ssize_t mcopy_continue(struct mm_struct *dst_mm, unsigned long start,
 697                        unsigned long len, atomic_t *mmap_changing)
 698 {
 699         return __mcopy_atomic(dst_mm, start, 0, len, MCOPY_ATOMIC_CONTINUE,
 700                               mmap_changing, 0);
 701 }
 702
 703 int mwriteprotect_range(struct mm_struct *dst_mm, unsigned long start,
 704                         unsigned long len, bool enable_wp,
 705                         atomic_t *mmap_changing)
 706 {
 707         struct vm_area_struct *dst_vma;
 708         unsigned long page_mask;
 709         struct mmu_gather tlb;
 710         pgprot_t newprot;
 711         int err;
 712
 713         /*
 714          * Sanitize the command parameters:
 715          */
 716         BUG_ON(start & ~PAGE_MASK);
 717         BUG_ON(len & ~PAGE_MASK);
 718
 719         /* Does the address range wrap, or is the span zero-sized? */
 720         BUG_ON(start + len <= start);
 721
 722         mmap_read_lock(dst_mm);
 723
 724         /*
 725          * If memory mappings are changing because of non-cooperative
 726          * operation (e.g. mremap) running in parallel, bail out and
 727          * request the user to retry later
 728          */
 729         err = -EAGAIN;
 730         if (mmap_changing && atomic_read(mmap_changing))
 731                 goto out_unlock;
 732
 733         err = -ENOENT;
 734         dst_vma = find_dst_vma(dst_mm, start, len);
 735
 736         if (!dst_vma)
 737                 goto out_unlock;
 738         if (!userfaultfd_wp(dst_vma))
 739                 goto out_unlock;
 740         if (!vma_can_userfault(dst_vma, dst_vma->vm_flags))
 741                 goto out_unlock;
 742
 743         if (is_vm_hugetlb_page(dst_vma)) {
 744                 err = -EINVAL;
 745                 page_mask = vma_kernel_pagesize(dst_vma) - 1;
 746                 if ((start & page_mask) || (len & page_mask))
 747                         goto out_unlock;
 748         }
 749
 750         if (enable_wp)
 751                 newprot = vm_get_page_prot(dst_vma->vm_flags & ~(VM_WRITE));
 752         else
 753                 newprot = vm_get_page_prot(dst_vma->vm_flags);
 754
 755         tlb_gather_mmu(&tlb, dst_mm);
 756         change_protection(&tlb, dst_vma, start, start + len, newprot,
 757                           enable_wp ? MM_CP_UFFD_WP : MM_CP_UFFD_WP_RESOLVE);
 758         tlb_finish_mmu(&tlb);
 759
 760         err = 0;
 761 out_unlock:
 762         mmap_read_unlock(dst_mm);
 763         return err;
 764 }