mm/userfaultfd.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  *  mm/userfaultfd.c
   4  *
   5  *  Copyright (C) 2015  Red Hat, Inc.
   6  */
   7
   8 #include <linux/mm.h>
   9 #include <linux/sched/signal.h>
  10 #include <linux/pagemap.h>
  11 #include <linux/rmap.h>
  12 #include <linux/swap.h>
  13 #include <linux/swapops.h>
  14 #include <linux/userfaultfd_k.h>
  15 #include <linux/mmu_notifier.h>
  16 #include <linux/hugetlb.h>
  17 #include <linux/shmem_fs.h>
  18 #include <asm/tlbflush.h>
  19 #include "internal.h"
  20
  21 static __always_inline
  22 struct vm_area_struct *find_dst_vma(struct mm_struct *dst_mm,
  23                                     unsigned long dst_start,
  24                                     unsigned long len)
  25 {
  26         /*
  27          * Make sure that the dst range is both valid and fully within a
  28          * single existing vma.
  29          */
  30         struct vm_area_struct *dst_vma;
  31
  32         dst_vma = find_vma(dst_mm, dst_start);
  33         if (!dst_vma)
  34                 return NULL;
  35
  36         if (dst_start < dst_vma->vm_start ||
  37             dst_start + len > dst_vma->vm_end)
  38                 return NULL;
  39
  40         /*
  41          * Check the vma is registered in uffd, this is required to
  42          * enforce the VM_MAYWRITE check done at uffd registration
  43          * time.
  44          */
  45         if (!dst_vma->vm_userfaultfd_ctx.ctx)
  46                 return NULL;
  47
  48         return dst_vma;
  49 }
  50
  51 /*
  52  * Install PTEs, to map dst_addr (within dst_vma) to page.
  53  *
  54  * This function handles both MCOPY_ATOMIC_NORMAL and _CONTINUE for both shmem
  55  * and anon, and for both shared and private VMAs.
  56  */
  57 int mfill_atomic_install_pte(struct mm_struct *dst_mm, pmd_t *dst_pmd,
  58                              struct vm_area_struct *dst_vma,
  59                              unsigned long dst_addr, struct page *page,
  60                              bool newly_allocated, bool wp_copy)
  61 {
  62         int ret;
  63         pte_t _dst_pte, *dst_pte;
  64         bool writable = dst_vma->vm_flags & VM_WRITE;
  65         bool vm_shared = dst_vma->vm_flags & VM_SHARED;
  66         bool page_in_cache = page->mapping;
  67         spinlock_t *ptl;
  68         struct inode *inode;
  69         pgoff_t offset, max_off;
  70
  71         _dst_pte = mk_pte(page, dst_vma->vm_page_prot);
  72         _dst_pte = pte_mkdirty(_dst_pte);
  73         if (page_in_cache && !vm_shared)
  74                 writable = false;
  75         if (writable) {
  76                 if (wp_copy)
  77                         _dst_pte = pte_mkuffd_wp(_dst_pte);
  78                 else
  79                         _dst_pte = pte_mkwrite(_dst_pte);
  80         }
  81
  82         dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
  83
  84         if (vma_is_shmem(dst_vma)) {
  85                 /* serialize against truncate with the page table lock */
  86                 inode = dst_vma->vm_file->f_inode;
  87                 offset = linear_page_index(dst_vma, dst_addr);
  88                 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
  89                 ret = -EFAULT;
  90                 if (unlikely(offset >= max_off))
  91                         goto out_unlock;
  92         }
  93
  94         ret = -EEXIST;
  95         if (!pte_none(*dst_pte))
  96                 goto out_unlock;
  97
  98         if (page_in_cache) {
  99                 /* Usually, cache pages are already added to LRU */
 100                 if (newly_allocated)
 101                         lru_cache_add(page);
 102                 page_add_file_rmap(page, dst_vma, false);
 103         } else {
 104                 page_add_new_anon_rmap(page, dst_vma, dst_addr, false);
 105                 lru_cache_add_inactive_or_unevictable(page, dst_vma);
 106         }
 107
 108         /*
 109          * Must happen after rmap, as mm_counter() checks mapping (via
 110          * PageAnon()), which is set by __page_set_anon_rmap().
 111          */
 112         inc_mm_counter(dst_mm, mm_counter(page));
 113
 114         set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
 115
 116         /* No need to invalidate - it was non-present before */
 117         update_mmu_cache(dst_vma, dst_addr, dst_pte);
 118         ret = 0;
 119 out_unlock:
 120         pte_unmap_unlock(dst_pte, ptl);
 121         return ret;
 122 }
 123
 124 static int mcopy_atomic_pte(struct mm_struct *dst_mm,
 125                             pmd_t *dst_pmd,
 126                             struct vm_area_struct *dst_vma,
 127                             unsigned long dst_addr,
 128                             unsigned long src_addr,
 129                             struct page **pagep,
 130                             bool wp_copy)
 131 {
 132         void *page_kaddr;
 133         int ret;
 134         struct page *page;
 135
 136         if (!*pagep) {
 137                 ret = -ENOMEM;
 138                 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, dst_vma, dst_addr);
 139                 if (!page)
 140                         goto out;
 141
 142                 page_kaddr = kmap_atomic(page);
 143                 ret = copy_from_user(page_kaddr,
 144                                      (const void __user *) src_addr,
 145                                      PAGE_SIZE);
 146                 kunmap_atomic(page_kaddr);
 147
 148                 /* fallback to copy_from_user outside mmap_lock */
 149                 if (unlikely(ret)) {
 150                         ret = -ENOENT;
 151                         *pagep = page;
 152                         /* don't free the page */
 153                         goto out;
 154                 }
 155         } else {
 156                 page = *pagep;
 157                 *pagep = NULL;
 158         }
 159
 160         /*
 161          * The memory barrier inside __SetPageUptodate makes sure that
 162          * preceding stores to the page contents become visible before
 163          * the set_pte_at() write.
 164          */
 165         __SetPageUptodate(page);
 166
 167         ret = -ENOMEM;
 168         if (mem_cgroup_charge(page_folio(page), dst_mm, GFP_KERNEL))
 169                 goto out_release;
 170
 171         ret = mfill_atomic_install_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
 172                                        page, true, wp_copy);
 173         if (ret)
 174                 goto out_release;
 175 out:
 176         return ret;
 177 out_release:
 178         put_page(page);
 179         goto out;
 180 }
 181
 182 static int mfill_zeropage_pte(struct mm_struct *dst_mm,
 183                               pmd_t *dst_pmd,
 184                               struct vm_area_struct *dst_vma,
 185                               unsigned long dst_addr)
 186 {
 187         pte_t _dst_pte, *dst_pte;
 188         spinlock_t *ptl;
 189         int ret;
 190         pgoff_t offset, max_off;
 191         struct inode *inode;
 192
 193         _dst_pte = pte_mkspecial(pfn_pte(my_zero_pfn(dst_addr),
 194                                          dst_vma->vm_page_prot));
 195         dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
 196         if (dst_vma->vm_file) {
 197                 /* the shmem MAP_PRIVATE case requires checking the i_size */
 198                 inode = dst_vma->vm_file->f_inode;
 199                 offset = linear_page_index(dst_vma, dst_addr);
 200                 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
 201                 ret = -EFAULT;
 202                 if (unlikely(offset >= max_off))
 203                         goto out_unlock;
 204         }
 205         ret = -EEXIST;
 206         if (!pte_none(*dst_pte))
 207                 goto out_unlock;
 208         set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
 209         /* No need to invalidate - it was non-present before */
 210         update_mmu_cache(dst_vma, dst_addr, dst_pte);
 211         ret = 0;
 212 out_unlock:
 213         pte_unmap_unlock(dst_pte, ptl);
 214         return ret;
 215 }
 216
 217 /* Handles UFFDIO_CONTINUE for all shmem VMAs (shared or private). */
 218 static int mcontinue_atomic_pte(struct mm_struct *dst_mm,
 219                                 pmd_t *dst_pmd,
 220                                 struct vm_area_struct *dst_vma,
 221                                 unsigned long dst_addr,
 222                                 bool wp_copy)
 223 {
 224         struct inode *inode = file_inode(dst_vma->vm_file);
 225         pgoff_t pgoff = linear_page_index(dst_vma, dst_addr);
 226         struct page *page;
 227         int ret;
 228
 229         ret = shmem_getpage(inode, pgoff, &page, SGP_READ);
 230         if (ret)
 231                 goto out;
 232         if (!page) {
 233                 ret = -EFAULT;
 234                 goto out;
 235         }
 236
 237         if (PageHWPoison(page)) {
 238                 ret = -EIO;
 239                 goto out_release;
 240         }
 241
 242         ret = mfill_atomic_install_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
 243                                        page, false, wp_copy);
 244         if (ret)
 245                 goto out_release;
 246
 247         unlock_page(page);
 248         ret = 0;
 249 out:
 250         return ret;
 251 out_release:
 252         unlock_page(page);
 253         put_page(page);
 254         goto out;
 255 }
 256
 257 static pmd_t *mm_alloc_pmd(struct mm_struct *mm, unsigned long address)
 258 {
 259         pgd_t *pgd;
 260         p4d_t *p4d;
 261         pud_t *pud;
 262
 263         pgd = pgd_offset(mm, address);
 264         p4d = p4d_alloc(mm, pgd, address);
 265         if (!p4d)
 266                 return NULL;
 267         pud = pud_alloc(mm, p4d, address);
 268         if (!pud)
 269                 return NULL;
 270         /*
 271          * Note that we didn't run this because the pmd was
 272          * missing, the *pmd may be already established and in
 273          * turn it may also be a trans_huge_pmd.
 274          */
 275         return pmd_alloc(mm, pud, address);
 276 }
 277
 278 #ifdef CONFIG_HUGETLB_PAGE
 279 /*
 280  * __mcopy_atomic processing for HUGETLB vmas.  Note that this routine is
 281  * called with mmap_lock held, it will release mmap_lock before returning.
 282  */
 283 static __always_inline ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
 284                                               struct vm_area_struct *dst_vma,
 285                                               unsigned long dst_start,
 286                                               unsigned long src_start,
 287                                               unsigned long len,
 288                                               enum mcopy_atomic_mode mode)
 289 {
 290         int vm_shared = dst_vma->vm_flags & VM_SHARED;
 291         ssize_t err;
 292         pte_t *dst_pte;
 293         unsigned long src_addr, dst_addr;
 294         long copied;
 295         struct page *page;
 296         unsigned long vma_hpagesize;
 297         pgoff_t idx;
 298         u32 hash;
 299         struct address_space *mapping;
 300
 301         /*
 302          * There is no default zero huge page for all huge page sizes as
 303          * supported by hugetlb.  A PMD_SIZE huge pages may exist as used
 304          * by THP.  Since we can not reliably insert a zero page, this
 305          * feature is not supported.
 306          */
 307         if (mode == MCOPY_ATOMIC_ZEROPAGE) {
 308                 mmap_read_unlock(dst_mm);
 309                 return -EINVAL;
 310         }
 311
 312         src_addr = src_start;
 313         dst_addr = dst_start;
 314         copied = 0;
 315         page = NULL;
 316         vma_hpagesize = vma_kernel_pagesize(dst_vma);
 317
 318         /*
 319          * Validate alignment based on huge page size
 320          */
 321         err = -EINVAL;
 322         if (dst_start & (vma_hpagesize - 1) || len & (vma_hpagesize - 1))
 323                 goto out_unlock;
 324
 325 retry:
 326         /*
 327          * On routine entry dst_vma is set.  If we had to drop mmap_lock and
 328          * retry, dst_vma will be set to NULL and we must lookup again.
 329          */
 330         if (!dst_vma) {
 331                 err = -ENOENT;
 332                 dst_vma = find_dst_vma(dst_mm, dst_start, len);
 333                 if (!dst_vma || !is_vm_hugetlb_page(dst_vma))
 334                         goto out_unlock;
 335
 336                 err = -EINVAL;
 337                 if (vma_hpagesize != vma_kernel_pagesize(dst_vma))
 338                         goto out_unlock;
 339
 340                 vm_shared = dst_vma->vm_flags & VM_SHARED;
 341         }
 342
 343         /*
 344          * If not shared, ensure the dst_vma has a anon_vma.
 345          */
 346         err = -ENOMEM;
 347         if (!vm_shared) {
 348                 if (unlikely(anon_vma_prepare(dst_vma)))
 349                         goto out_unlock;
 350         }
 351
 352         while (src_addr < src_start + len) {
 353                 BUG_ON(dst_addr >= dst_start + len);
 354
 355                 /*
 356                  * Serialize via i_mmap_rwsem and hugetlb_fault_mutex.
 357                  * i_mmap_rwsem ensures the dst_pte remains valid even
 358                  * in the case of shared pmds.  fault mutex prevents
 359                  * races with other faulting threads.
 360                  */
 361                 mapping = dst_vma->vm_file->f_mapping;
 362                 i_mmap_lock_read(mapping);
 363                 idx = linear_page_index(dst_vma, dst_addr);
 364                 hash = hugetlb_fault_mutex_hash(mapping, idx);
 365                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 366
 367                 err = -ENOMEM;
 368                 dst_pte = huge_pte_alloc(dst_mm, dst_vma, dst_addr, vma_hpagesize);
 369                 if (!dst_pte) {
 370                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 371                         i_mmap_unlock_read(mapping);
 372                         goto out_unlock;
 373                 }
 374
 375                 if (mode != MCOPY_ATOMIC_CONTINUE &&
 376                     !huge_pte_none(huge_ptep_get(dst_pte))) {
 377                         err = -EEXIST;
 378                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 379                         i_mmap_unlock_read(mapping);
 380                         goto out_unlock;
 381                 }
 382
 383                 err = hugetlb_mcopy_atomic_pte(dst_mm, dst_pte, dst_vma,
 384                                                dst_addr, src_addr, mode, &page);
 385
 386                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 387                 i_mmap_unlock_read(mapping);
 388
 389                 cond_resched();
 390
 391                 if (unlikely(err == -ENOENT)) {
 392                         mmap_read_unlock(dst_mm);
 393                         BUG_ON(!page);
 394
 395                         err = copy_huge_page_from_user(page,
 396                                                 (const void __user *)src_addr,
 397                                                 vma_hpagesize / PAGE_SIZE,
 398                                                 true);
 399                         if (unlikely(err)) {
 400                                 err = -EFAULT;
 401                                 goto out;
 402                         }
 403                         mmap_read_lock(dst_mm);
 404
 405                         dst_vma = NULL;
 406                         goto retry;
 407                 } else
 408                         BUG_ON(page);
 409
 410                 if (!err) {
 411                         dst_addr += vma_hpagesize;
 412                         src_addr += vma_hpagesize;
 413                         copied += vma_hpagesize;
 414
 415                         if (fatal_signal_pending(current))
 416                                 err = -EINTR;
 417                 }
 418                 if (err)
 419                         break;
 420         }
 421
 422 out_unlock:
 423         mmap_read_unlock(dst_mm);
 424 out:
 425         if (page)
 426                 put_page(page);
 427         BUG_ON(copied < 0);
 428         BUG_ON(err > 0);
 429         BUG_ON(!copied && !err);
 430         return copied ? copied : err;
 431 }
 432 #else /* !CONFIG_HUGETLB_PAGE */
 433 /* fail at build time if gcc attempts to use this */
 434 extern ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
 435                                       struct vm_area_struct *dst_vma,
 436                                       unsigned long dst_start,
 437                                       unsigned long src_start,
 438                                       unsigned long len,
 439                                       enum mcopy_atomic_mode mode);
 440 #endif /* CONFIG_HUGETLB_PAGE */
 441
 442 static __always_inline ssize_t mfill_atomic_pte(struct mm_struct *dst_mm,
 443                                                 pmd_t *dst_pmd,
 444                                                 struct vm_area_struct *dst_vma,
 445                                                 unsigned long dst_addr,
 446                                                 unsigned long src_addr,
 447                                                 struct page **page,
 448                                                 enum mcopy_atomic_mode mode,
 449                                                 bool wp_copy)
 450 {
 451         ssize_t err;
 452
 453         if (mode == MCOPY_ATOMIC_CONTINUE) {
 454                 return mcontinue_atomic_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
 455                                             wp_copy);
 456         }
 457
 458         /*
 459          * The normal page fault path for a shmem will invoke the
 460          * fault, fill the hole in the file and COW it right away. The
 461          * result generates plain anonymous memory. So when we are
 462          * asked to fill an hole in a MAP_PRIVATE shmem mapping, we'll
 463          * generate anonymous memory directly without actually filling
 464          * the hole. For the MAP_PRIVATE case the robustness check
 465          * only happens in the pagetable (to verify it's still none)
 466          * and not in the radix tree.
 467          */
 468         if (!(dst_vma->vm_flags & VM_SHARED)) {
 469                 if (mode == MCOPY_ATOMIC_NORMAL)
 470                         err = mcopy_atomic_pte(dst_mm, dst_pmd, dst_vma,
 471                                                dst_addr, src_addr, page,
 472                                                wp_copy);
 473                 else
 474                         err = mfill_zeropage_pte(dst_mm, dst_pmd,
 475                                                  dst_vma, dst_addr);
 476         } else {
 477                 VM_WARN_ON_ONCE(wp_copy);
 478                 err = shmem_mfill_atomic_pte(dst_mm, dst_pmd, dst_vma,
 479                                              dst_addr, src_addr,
 480                                              mode != MCOPY_ATOMIC_NORMAL,
 481                                              page);
 482         }
 483
 484         return err;
 485 }
 486
 487 static __always_inline ssize_t __mcopy_atomic(struct mm_struct *dst_mm,
 488                                               unsigned long dst_start,
 489                                               unsigned long src_start,
 490                                               unsigned long len,
 491                                               enum mcopy_atomic_mode mcopy_mode,
 492                                               atomic_t *mmap_changing,
 493                                               __u64 mode)
 494 {
 495         struct vm_area_struct *dst_vma;
 496         ssize_t err;
 497         pmd_t *dst_pmd;
 498         unsigned long src_addr, dst_addr;
 499         long copied;
 500         struct page *page;
 501         bool wp_copy;
 502
 503         /*
 504          * Sanitize the command parameters:
 505          */
 506         BUG_ON(dst_start & ~PAGE_MASK);
 507         BUG_ON(len & ~PAGE_MASK);
 508
 509         /* Does the address range wrap, or is the span zero-sized? */
 510         BUG_ON(src_start + len <= src_start);
 511         BUG_ON(dst_start + len <= dst_start);
 512
 513         src_addr = src_start;
 514         dst_addr = dst_start;
 515         copied = 0;
 516         page = NULL;
 517 retry:
 518         mmap_read_lock(dst_mm);
 519
 520         /*
 521          * If memory mappings are changing because of non-cooperative
 522          * operation (e.g. mremap) running in parallel, bail out and
 523          * request the user to retry later
 524          */
 525         err = -EAGAIN;
 526         if (mmap_changing && atomic_read(mmap_changing))
 527                 goto out_unlock;
 528
 529         /*
 530          * Make sure the vma is not shared, that the dst range is
 531          * both valid and fully within a single existing vma.
 532          */
 533         err = -ENOENT;
 534         dst_vma = find_dst_vma(dst_mm, dst_start, len);
 535         if (!dst_vma)
 536                 goto out_unlock;
 537
 538         err = -EINVAL;
 539         /*
 540          * shmem_zero_setup is invoked in mmap for MAP_ANONYMOUS|MAP_SHARED but
 541          * it will overwrite vm_ops, so vma_is_anonymous must return false.
 542          */
 543         if (WARN_ON_ONCE(vma_is_anonymous(dst_vma) &&
 544             dst_vma->vm_flags & VM_SHARED))
 545                 goto out_unlock;
 546
 547         /*
 548          * validate 'mode' now that we know the dst_vma: don't allow
 549          * a wrprotect copy if the userfaultfd didn't register as WP.
 550          */
 551         wp_copy = mode & UFFDIO_COPY_MODE_WP;
 552         if (wp_copy && !(dst_vma->vm_flags & VM_UFFD_WP))
 553                 goto out_unlock;
 554
 555         /*
 556          * If this is a HUGETLB vma, pass off to appropriate routine
 557          */
 558         if (is_vm_hugetlb_page(dst_vma))
 559                 return  __mcopy_atomic_hugetlb(dst_mm, dst_vma, dst_start,
 560                                                 src_start, len, mcopy_mode);
 561
 562         if (!vma_is_anonymous(dst_vma) && !vma_is_shmem(dst_vma))
 563                 goto out_unlock;
 564         if (!vma_is_shmem(dst_vma) && mcopy_mode == MCOPY_ATOMIC_CONTINUE)
 565                 goto out_unlock;
 566
 567         /*
 568          * Ensure the dst_vma has a anon_vma or this page
 569          * would get a NULL anon_vma when moved in the
 570          * dst_vma.
 571          */
 572         err = -ENOMEM;
 573         if (!(dst_vma->vm_flags & VM_SHARED) &&
 574             unlikely(anon_vma_prepare(dst_vma)))
 575                 goto out_unlock;
 576
 577         while (src_addr < src_start + len) {
 578                 pmd_t dst_pmdval;
 579
 580                 BUG_ON(dst_addr >= dst_start + len);
 581
 582                 dst_pmd = mm_alloc_pmd(dst_mm, dst_addr);
 583                 if (unlikely(!dst_pmd)) {
 584                         err = -ENOMEM;
 585                         break;
 586                 }
 587
 588                 dst_pmdval = pmd_read_atomic(dst_pmd);
 589                 /*
 590                  * If the dst_pmd is mapped as THP don't
 591                  * override it and just be strict.
 592                  */
 593                 if (unlikely(pmd_trans_huge(dst_pmdval))) {
 594                         err = -EEXIST;
 595                         break;
 596                 }
 597                 if (unlikely(pmd_none(dst_pmdval)) &&
 598                     unlikely(__pte_alloc(dst_mm, dst_pmd))) {
 599                         err = -ENOMEM;
 600                         break;
 601                 }
 602                 /* If an huge pmd materialized from under us fail */
 603                 if (unlikely(pmd_trans_huge(*dst_pmd))) {
 604                         err = -EFAULT;
 605                         break;
 606                 }
 607
 608                 BUG_ON(pmd_none(*dst_pmd));
 609                 BUG_ON(pmd_trans_huge(*dst_pmd));
 610
 611                 err = mfill_atomic_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
 612                                        src_addr, &page, mcopy_mode, wp_copy);
 613                 cond_resched();
 614
 615                 if (unlikely(err == -ENOENT)) {
 616                         void *page_kaddr;
 617
 618                         mmap_read_unlock(dst_mm);
 619                         BUG_ON(!page);
 620
 621                         page_kaddr = kmap(page);
 622                         err = copy_from_user(page_kaddr,
 623                                              (const void __user *) src_addr,
 624                                              PAGE_SIZE);
 625                         kunmap(page);
 626                         if (unlikely(err)) {
 627                                 err = -EFAULT;
 628                                 goto out;
 629                         }
 630                         goto retry;
 631                 } else
 632                         BUG_ON(page);
 633
 634                 if (!err) {
 635                         dst_addr += PAGE_SIZE;
 636                         src_addr += PAGE_SIZE;
 637                         copied += PAGE_SIZE;
 638
 639                         if (fatal_signal_pending(current))
 640                                 err = -EINTR;
 641                 }
 642                 if (err)
 643                         break;
 644         }
 645
 646 out_unlock:
 647         mmap_read_unlock(dst_mm);
 648 out:
 649         if (page)
 650                 put_page(page);
 651         BUG_ON(copied < 0);
 652         BUG_ON(err > 0);
 653         BUG_ON(!copied && !err);
 654         return copied ? copied : err;
 655 }
 656
 657 ssize_t mcopy_atomic(struct mm_struct *dst_mm, unsigned long dst_start,
 658                      unsigned long src_start, unsigned long len,
 659                      atomic_t *mmap_changing, __u64 mode)
 660 {
 661         return __mcopy_atomic(dst_mm, dst_start, src_start, len,
 662                               MCOPY_ATOMIC_NORMAL, mmap_changing, mode);
 663 }
 664
 665 ssize_t mfill_zeropage(struct mm_struct *dst_mm, unsigned long start,
 666                        unsigned long len, atomic_t *mmap_changing)
 667 {
 668         return __mcopy_atomic(dst_mm, start, 0, len, MCOPY_ATOMIC_ZEROPAGE,
 669                               mmap_changing, 0);
 670 }
 671
 672 ssize_t mcopy_continue(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_CONTINUE,
 676                               mmap_changing, 0);
 677 }
 678
 679 int mwriteprotect_range(struct mm_struct *dst_mm, unsigned long start,
 680                         unsigned long len, bool enable_wp,
 681                         atomic_t *mmap_changing)
 682 {
 683         struct vm_area_struct *dst_vma;
 684         pgprot_t newprot;
 685         int err;
 686
 687         /*
 688          * Sanitize the command parameters:
 689          */
 690         BUG_ON(start & ~PAGE_MASK);
 691         BUG_ON(len & ~PAGE_MASK);
 692
 693         /* Does the address range wrap, or is the span zero-sized? */
 694         BUG_ON(start + len <= start);
 695
 696         mmap_read_lock(dst_mm);
 697
 698         /*
 699          * If memory mappings are changing because of non-cooperative
 700          * operation (e.g. mremap) running in parallel, bail out and
 701          * request the user to retry later
 702          */
 703         err = -EAGAIN;
 704         if (mmap_changing && atomic_read(mmap_changing))
 705                 goto out_unlock;
 706
 707         err = -ENOENT;
 708         dst_vma = find_dst_vma(dst_mm, start, len);
 709         /*
 710          * Make sure the vma is not shared, that the dst range is
 711          * both valid and fully within a single existing vma.
 712          */
 713         if (!dst_vma || (dst_vma->vm_flags & VM_SHARED))
 714                 goto out_unlock;
 715         if (!userfaultfd_wp(dst_vma))
 716                 goto out_unlock;
 717         if (!vma_is_anonymous(dst_vma))
 718                 goto out_unlock;
 719
 720         if (enable_wp)
 721                 newprot = vm_get_page_prot(dst_vma->vm_flags & ~(VM_WRITE));
 722         else
 723                 newprot = vm_get_page_prot(dst_vma->vm_flags);
 724
 725         change_protection(dst_vma, start, start + len, newprot,
 726                           enable_wp ? MM_CP_UFFD_WP : MM_CP_UFFD_WP_RESOLVE);
 727
 728         err = 0;
 729 out_unlock:
 730         mmap_read_unlock(dst_mm);
 731         return err;
 732 }