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