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