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(page);
  68         spinlock_t *ptl;
  69         struct folio *folio;
  70         struct inode *inode;
  71         pgoff_t offset, max_off;
  72
  73         _dst_pte = mk_pte(page, dst_vma->vm_page_prot);
  74         _dst_pte = pte_mkdirty(_dst_pte);
  75         if (page_in_cache && !vm_shared)
  76                 writable = false;
  77         if (writable)
  78                 _dst_pte = pte_mkwrite(_dst_pte);
  79         if (wp_copy)
  80                 _dst_pte = pte_mkuffd_wp(_dst_pte);
  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         /*
  96          * We allow to overwrite a pte marker: consider when both MISSING|WP
  97          * registered, we firstly wr-protect a none pte which has no page cache
  98          * page backing it, then access the page.
  99          */
 100         if (!pte_none_mostly(*dst_pte))
 101                 goto out_unlock;
 102
 103         folio = page_folio(page);
 104         if (page_in_cache) {
 105                 /* Usually, cache pages are already added to LRU */
 106                 if (newly_allocated)
 107                         folio_add_lru(folio);
 108                 page_add_file_rmap(page, dst_vma, false);
 109         } else {
 110                 page_add_new_anon_rmap(page, dst_vma, dst_addr);
 111                 folio_add_lru_vma(folio, dst_vma);
 112         }
 113
 114         /*
 115          * Must happen after rmap, as mm_counter() checks mapping (via
 116          * PageAnon()), which is set by __page_set_anon_rmap().
 117          */
 118         inc_mm_counter(dst_mm, mm_counter(page));
 119
 120         set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
 121
 122         /* No need to invalidate - it was non-present before */
 123         update_mmu_cache(dst_vma, dst_addr, dst_pte);
 124         ret = 0;
 125 out_unlock:
 126         pte_unmap_unlock(dst_pte, ptl);
 127         return ret;
 128 }
 129
 130 static int mcopy_atomic_pte(struct mm_struct *dst_mm,
 131                             pmd_t *dst_pmd,
 132                             struct vm_area_struct *dst_vma,
 133                             unsigned long dst_addr,
 134                             unsigned long src_addr,
 135                             struct page **pagep,
 136                             bool wp_copy)
 137 {
 138         void *page_kaddr;
 139         int ret;
 140         struct page *page;
 141
 142         if (!*pagep) {
 143                 ret = -ENOMEM;
 144                 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, dst_vma, dst_addr);
 145                 if (!page)
 146                         goto out;
 147
 148                 page_kaddr = kmap_local_page(page);
 149                 /*
 150                  * The read mmap_lock is held here.  Despite the
 151                  * mmap_lock being read recursive a deadlock is still
 152                  * possible if a writer has taken a lock.  For example:
 153                  *
 154                  * process A thread 1 takes read lock on own mmap_lock
 155                  * process A thread 2 calls mmap, blocks taking write lock
 156                  * process B thread 1 takes page fault, read lock on own mmap lock
 157                  * process B thread 2 calls mmap, blocks taking write lock
 158                  * process A thread 1 blocks taking read lock on process B
 159                  * process B thread 1 blocks taking read lock on process A
 160                  *
 161                  * Disable page faults to prevent potential deadlock
 162                  * and retry the copy outside the mmap_lock.
 163                  */
 164                 pagefault_disable();
 165                 ret = copy_from_user(page_kaddr,
 166                                      (const void __user *) src_addr,
 167                                      PAGE_SIZE);
 168                 pagefault_enable();
 169                 kunmap_local(page_kaddr);
 170
 171                 /* fallback to copy_from_user outside mmap_lock */
 172                 if (unlikely(ret)) {
 173                         ret = -ENOENT;
 174                         *pagep = page;
 175                         /* don't free the page */
 176                         goto out;
 177                 }
 178
 179                 flush_dcache_page(page);
 180         } else {
 181                 page = *pagep;
 182                 *pagep = NULL;
 183         }
 184
 185         /*
 186          * The memory barrier inside __SetPageUptodate makes sure that
 187          * preceding stores to the page contents become visible before
 188          * the set_pte_at() write.
 189          */
 190         __SetPageUptodate(page);
 191
 192         ret = -ENOMEM;
 193         if (mem_cgroup_charge(page_folio(page), dst_mm, GFP_KERNEL))
 194                 goto out_release;
 195
 196         ret = mfill_atomic_install_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
 197                                        page, true, wp_copy);
 198         if (ret)
 199                 goto out_release;
 200 out:
 201         return ret;
 202 out_release:
 203         put_page(page);
 204         goto out;
 205 }
 206
 207 static int mfill_zeropage_pte(struct mm_struct *dst_mm,
 208                               pmd_t *dst_pmd,
 209                               struct vm_area_struct *dst_vma,
 210                               unsigned long dst_addr)
 211 {
 212         pte_t _dst_pte, *dst_pte;
 213         spinlock_t *ptl;
 214         int ret;
 215         pgoff_t offset, max_off;
 216         struct inode *inode;
 217
 218         _dst_pte = pte_mkspecial(pfn_pte(my_zero_pfn(dst_addr),
 219                                          dst_vma->vm_page_prot));
 220         dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
 221         if (dst_vma->vm_file) {
 222                 /* the shmem MAP_PRIVATE case requires checking the i_size */
 223                 inode = dst_vma->vm_file->f_inode;
 224                 offset = linear_page_index(dst_vma, dst_addr);
 225                 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
 226                 ret = -EFAULT;
 227                 if (unlikely(offset >= max_off))
 228                         goto out_unlock;
 229         }
 230         ret = -EEXIST;
 231         if (!pte_none(*dst_pte))
 232                 goto out_unlock;
 233         set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
 234         /* No need to invalidate - it was non-present before */
 235         update_mmu_cache(dst_vma, dst_addr, dst_pte);
 236         ret = 0;
 237 out_unlock:
 238         pte_unmap_unlock(dst_pte, ptl);
 239         return ret;
 240 }
 241
 242 /* Handles UFFDIO_CONTINUE for all shmem VMAs (shared or private). */
 243 static int mcontinue_atomic_pte(struct mm_struct *dst_mm,
 244                                 pmd_t *dst_pmd,
 245                                 struct vm_area_struct *dst_vma,
 246                                 unsigned long dst_addr,
 247                                 bool wp_copy)
 248 {
 249         struct inode *inode = file_inode(dst_vma->vm_file);
 250         pgoff_t pgoff = linear_page_index(dst_vma, dst_addr);
 251         struct folio *folio;
 252         struct page *page;
 253         int ret;
 254
 255         ret = shmem_get_folio(inode, pgoff, &folio, SGP_NOALLOC);
 256         /* Our caller expects us to return -EFAULT if we failed to find folio */
 257         if (ret == -ENOENT)
 258                 ret = -EFAULT;
 259         if (ret)
 260                 goto out;
 261         if (!folio) {
 262                 ret = -EFAULT;
 263                 goto out;
 264         }
 265
 266         page = folio_file_page(folio, pgoff);
 267         if (PageHWPoison(page)) {
 268                 ret = -EIO;
 269                 goto out_release;
 270         }
 271
 272         ret = mfill_atomic_install_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
 273                                        page, false, wp_copy);
 274         if (ret)
 275                 goto out_release;
 276
 277         folio_unlock(folio);
 278         ret = 0;
 279 out:
 280         return ret;
 281 out_release:
 282         folio_unlock(folio);
 283         folio_put(folio);
 284         goto out;
 285 }
 286
 287 static pmd_t *mm_alloc_pmd(struct mm_struct *mm, unsigned long address)
 288 {
 289         pgd_t *pgd;
 290         p4d_t *p4d;
 291         pud_t *pud;
 292
 293         pgd = pgd_offset(mm, address);
 294         p4d = p4d_alloc(mm, pgd, address);
 295         if (!p4d)
 296                 return NULL;
 297         pud = pud_alloc(mm, p4d, address);
 298         if (!pud)
 299                 return NULL;
 300         /*
 301          * Note that we didn't run this because the pmd was
 302          * missing, the *pmd may be already established and in
 303          * turn it may also be a trans_huge_pmd.
 304          */
 305         return pmd_alloc(mm, pud, address);
 306 }
 307
 308 #ifdef CONFIG_HUGETLB_PAGE
 309 /*
 310  * __mcopy_atomic processing for HUGETLB vmas.  Note that this routine is
 311  * called with mmap_lock held, it will release mmap_lock before returning.
 312  */
 313 static __always_inline ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
 314                                               struct vm_area_struct *dst_vma,
 315                                               unsigned long dst_start,
 316                                               unsigned long src_start,
 317                                               unsigned long len,
 318                                               enum mcopy_atomic_mode mode,
 319                                               bool wp_copy)
 320 {
 321         int vm_shared = dst_vma->vm_flags & VM_SHARED;
 322         ssize_t err;
 323         pte_t *dst_pte;
 324         unsigned long src_addr, dst_addr;
 325         long copied;
 326         struct page *page;
 327         unsigned long vma_hpagesize;
 328         pgoff_t idx;
 329         u32 hash;
 330         struct address_space *mapping;
 331
 332         /*
 333          * There is no default zero huge page for all huge page sizes as
 334          * supported by hugetlb.  A PMD_SIZE huge pages may exist as used
 335          * by THP.  Since we can not reliably insert a zero page, this
 336          * feature is not supported.
 337          */
 338         if (mode == MCOPY_ATOMIC_ZEROPAGE) {
 339                 mmap_read_unlock(dst_mm);
 340                 return -EINVAL;
 341         }
 342
 343         src_addr = src_start;
 344         dst_addr = dst_start;
 345         copied = 0;
 346         page = NULL;
 347         vma_hpagesize = vma_kernel_pagesize(dst_vma);
 348
 349         /*
 350          * Validate alignment based on huge page size
 351          */
 352         err = -EINVAL;
 353         if (dst_start & (vma_hpagesize - 1) || len & (vma_hpagesize - 1))
 354                 goto out_unlock;
 355
 356 retry:
 357         /*
 358          * On routine entry dst_vma is set.  If we had to drop mmap_lock and
 359          * retry, dst_vma will be set to NULL and we must lookup again.
 360          */
 361         if (!dst_vma) {
 362                 err = -ENOENT;
 363                 dst_vma = find_dst_vma(dst_mm, dst_start, len);
 364                 if (!dst_vma || !is_vm_hugetlb_page(dst_vma))
 365                         goto out_unlock;
 366
 367                 err = -EINVAL;
 368                 if (vma_hpagesize != vma_kernel_pagesize(dst_vma))
 369                         goto out_unlock;
 370
 371                 vm_shared = dst_vma->vm_flags & VM_SHARED;
 372         }
 373
 374         /*
 375          * If not shared, ensure the dst_vma has a anon_vma.
 376          */
 377         err = -ENOMEM;
 378         if (!vm_shared) {
 379                 if (unlikely(anon_vma_prepare(dst_vma)))
 380                         goto out_unlock;
 381         }
 382
 383         while (src_addr < src_start + len) {
 384                 BUG_ON(dst_addr >= dst_start + len);
 385
 386                 /*
 387                  * Serialize via vma_lock and hugetlb_fault_mutex.
 388                  * vma_lock ensures the dst_pte remains valid even
 389                  * in the case of shared pmds.  fault mutex prevents
 390                  * races with other faulting threads.
 391                  */
 392                 idx = linear_page_index(dst_vma, dst_addr);
 393                 mapping = dst_vma->vm_file->f_mapping;
 394                 hash = hugetlb_fault_mutex_hash(mapping, idx);
 395                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 396                 hugetlb_vma_lock_read(dst_vma);
 397
 398                 err = -ENOMEM;
 399                 dst_pte = huge_pte_alloc(dst_mm, dst_vma, dst_addr, vma_hpagesize);
 400                 if (!dst_pte) {
 401                         hugetlb_vma_unlock_read(dst_vma);
 402                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 403                         goto out_unlock;
 404                 }
 405
 406                 if (mode != MCOPY_ATOMIC_CONTINUE &&
 407                     !huge_pte_none_mostly(huge_ptep_get(dst_pte))) {
 408                         err = -EEXIST;
 409                         hugetlb_vma_unlock_read(dst_vma);
 410                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 411                         goto out_unlock;
 412                 }
 413
 414                 err = hugetlb_mcopy_atomic_pte(dst_mm, dst_pte, dst_vma,
 415                                                dst_addr, src_addr, mode, &page,
 416                                                wp_copy);
 417
 418                 hugetlb_vma_unlock_read(dst_vma);
 419                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 420
 421                 cond_resched();
 422
 423                 if (unlikely(err == -ENOENT)) {
 424                         mmap_read_unlock(dst_mm);
 425                         BUG_ON(!page);
 426
 427                         err = copy_huge_page_from_user(page,
 428                                                 (const void __user *)src_addr,
 429                                                 vma_hpagesize / PAGE_SIZE,
 430                                                 true);
 431                         if (unlikely(err)) {
 432                                 err = -EFAULT;
 433                                 goto out;
 434                         }
 435                         mmap_read_lock(dst_mm);
 436
 437                         dst_vma = NULL;
 438                         goto retry;
 439                 } else
 440                         BUG_ON(page);
 441
 442                 if (!err) {
 443                         dst_addr += vma_hpagesize;
 444                         src_addr += vma_hpagesize;
 445                         copied += vma_hpagesize;
 446
 447                         if (fatal_signal_pending(current))
 448                                 err = -EINTR;
 449                 }
 450                 if (err)
 451                         break;
 452         }
 453
 454 out_unlock:
 455         mmap_read_unlock(dst_mm);
 456 out:
 457         if (page)
 458                 put_page(page);
 459         BUG_ON(copied < 0);
 460         BUG_ON(err > 0);
 461         BUG_ON(!copied && !err);
 462         return copied ? copied : err;
 463 }
 464 #else /* !CONFIG_HUGETLB_PAGE */
 465 /* fail at build time if gcc attempts to use this */
 466 extern ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
 467                                       struct vm_area_struct *dst_vma,
 468                                       unsigned long dst_start,
 469                                       unsigned long src_start,
 470                                       unsigned long len,
 471                                       enum mcopy_atomic_mode mode,
 472                                       bool wp_copy);
 473 #endif /* CONFIG_HUGETLB_PAGE */
 474
 475 static __always_inline ssize_t mfill_atomic_pte(struct mm_struct *dst_mm,
 476                                                 pmd_t *dst_pmd,
 477                                                 struct vm_area_struct *dst_vma,
 478                                                 unsigned long dst_addr,
 479                                                 unsigned long src_addr,
 480                                                 struct page **page,
 481                                                 enum mcopy_atomic_mode mode,
 482                                                 bool wp_copy)
 483 {
 484         ssize_t err;
 485
 486         if (mode == MCOPY_ATOMIC_CONTINUE) {
 487                 return mcontinue_atomic_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
 488                                             wp_copy);
 489         }
 490
 491         /*
 492          * The normal page fault path for a shmem will invoke the
 493          * fault, fill the hole in the file and COW it right away. The
 494          * result generates plain anonymous memory. So when we are
 495          * asked to fill an hole in a MAP_PRIVATE shmem mapping, we'll
 496          * generate anonymous memory directly without actually filling
 497          * the hole. For the MAP_PRIVATE case the robustness check
 498          * only happens in the pagetable (to verify it's still none)
 499          * and not in the radix tree.
 500          */
 501         if (!(dst_vma->vm_flags & VM_SHARED)) {
 502                 if (mode == MCOPY_ATOMIC_NORMAL)
 503                         err = mcopy_atomic_pte(dst_mm, dst_pmd, dst_vma,
 504                                                dst_addr, src_addr, page,
 505                                                wp_copy);
 506                 else
 507                         err = mfill_zeropage_pte(dst_mm, dst_pmd,
 508                                                  dst_vma, dst_addr);
 509         } else {
 510                 err = shmem_mfill_atomic_pte(dst_mm, dst_pmd, dst_vma,
 511                                              dst_addr, src_addr,
 512                                              mode != MCOPY_ATOMIC_NORMAL,
 513                                              wp_copy, page);
 514         }
 515
 516         return err;
 517 }
 518
 519 static __always_inline ssize_t __mcopy_atomic(struct mm_struct *dst_mm,
 520                                               unsigned long dst_start,
 521                                               unsigned long src_start,
 522                                               unsigned long len,
 523                                               enum mcopy_atomic_mode mcopy_mode,
 524                                               atomic_t *mmap_changing,
 525                                               __u64 mode)
 526 {
 527         struct vm_area_struct *dst_vma;
 528         ssize_t err;
 529         pmd_t *dst_pmd;
 530         unsigned long src_addr, dst_addr;
 531         long copied;
 532         struct page *page;
 533         bool wp_copy;
 534
 535         /*
 536          * Sanitize the command parameters:
 537          */
 538         BUG_ON(dst_start & ~PAGE_MASK);
 539         BUG_ON(len & ~PAGE_MASK);
 540
 541         /* Does the address range wrap, or is the span zero-sized? */
 542         BUG_ON(src_start + len <= src_start);
 543         BUG_ON(dst_start + len <= dst_start);
 544
 545         src_addr = src_start;
 546         dst_addr = dst_start;
 547         copied = 0;
 548         page = NULL;
 549 retry:
 550         mmap_read_lock(dst_mm);
 551
 552         /*
 553          * If memory mappings are changing because of non-cooperative
 554          * operation (e.g. mremap) running in parallel, bail out and
 555          * request the user to retry later
 556          */
 557         err = -EAGAIN;
 558         if (mmap_changing && atomic_read(mmap_changing))
 559                 goto out_unlock;
 560
 561         /*
 562          * Make sure the vma is not shared, that the dst range is
 563          * both valid and fully within a single existing vma.
 564          */
 565         err = -ENOENT;
 566         dst_vma = find_dst_vma(dst_mm, dst_start, len);
 567         if (!dst_vma)
 568                 goto out_unlock;
 569
 570         err = -EINVAL;
 571         /*
 572          * shmem_zero_setup is invoked in mmap for MAP_ANONYMOUS|MAP_SHARED but
 573          * it will overwrite vm_ops, so vma_is_anonymous must return false.
 574          */
 575         if (WARN_ON_ONCE(vma_is_anonymous(dst_vma) &&
 576             dst_vma->vm_flags & VM_SHARED))
 577                 goto out_unlock;
 578
 579         /*
 580          * validate 'mode' now that we know the dst_vma: don't allow
 581          * a wrprotect copy if the userfaultfd didn't register as WP.
 582          */
 583         wp_copy = mode & UFFDIO_COPY_MODE_WP;
 584         if (wp_copy && !(dst_vma->vm_flags & VM_UFFD_WP))
 585                 goto out_unlock;
 586
 587         /*
 588          * If this is a HUGETLB vma, pass off to appropriate routine
 589          */
 590         if (is_vm_hugetlb_page(dst_vma))
 591                 return  __mcopy_atomic_hugetlb(dst_mm, dst_vma, dst_start,
 592                                                src_start, len, mcopy_mode,
 593                                                wp_copy);
 594
 595         if (!vma_is_anonymous(dst_vma) && !vma_is_shmem(dst_vma))
 596                 goto out_unlock;
 597         if (!vma_is_shmem(dst_vma) && mcopy_mode == MCOPY_ATOMIC_CONTINUE)
 598                 goto out_unlock;
 599
 600         /*
 601          * Ensure the dst_vma has a anon_vma or this page
 602          * would get a NULL anon_vma when moved in the
 603          * dst_vma.
 604          */
 605         err = -ENOMEM;
 606         if (!(dst_vma->vm_flags & VM_SHARED) &&
 607             unlikely(anon_vma_prepare(dst_vma)))
 608                 goto out_unlock;
 609
 610         while (src_addr < src_start + len) {
 611                 pmd_t dst_pmdval;
 612
 613                 BUG_ON(dst_addr >= dst_start + len);
 614
 615                 dst_pmd = mm_alloc_pmd(dst_mm, dst_addr);
 616                 if (unlikely(!dst_pmd)) {
 617                         err = -ENOMEM;
 618                         break;
 619                 }
 620
 621                 dst_pmdval = pmdp_get_lockless(dst_pmd);
 622                 /*
 623                  * If the dst_pmd is mapped as THP don't
 624                  * override it and just be strict.
 625                  */
 626                 if (unlikely(pmd_trans_huge(dst_pmdval))) {
 627                         err = -EEXIST;
 628                         break;
 629                 }
 630                 if (unlikely(pmd_none(dst_pmdval)) &&
 631                     unlikely(__pte_alloc(dst_mm, dst_pmd))) {
 632                         err = -ENOMEM;
 633                         break;
 634                 }
 635                 /* If an huge pmd materialized from under us fail */
 636                 if (unlikely(pmd_trans_huge(*dst_pmd))) {
 637                         err = -EFAULT;
 638                         break;
 639                 }
 640
 641                 BUG_ON(pmd_none(*dst_pmd));
 642                 BUG_ON(pmd_trans_huge(*dst_pmd));
 643
 644                 err = mfill_atomic_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
 645                                        src_addr, &page, mcopy_mode, wp_copy);
 646                 cond_resched();
 647
 648                 if (unlikely(err == -ENOENT)) {
 649                         void *page_kaddr;
 650
 651                         mmap_read_unlock(dst_mm);
 652                         BUG_ON(!page);
 653
 654                         page_kaddr = kmap_local_page(page);
 655                         err = copy_from_user(page_kaddr,
 656                                              (const void __user *) src_addr,
 657                                              PAGE_SIZE);
 658                         kunmap_local(page_kaddr);
 659                         if (unlikely(err)) {
 660                                 err = -EFAULT;
 661                                 goto out;
 662                         }
 663                         flush_dcache_page(page);
 664                         goto retry;
 665                 } else
 666                         BUG_ON(page);
 667
 668                 if (!err) {
 669                         dst_addr += PAGE_SIZE;
 670                         src_addr += PAGE_SIZE;
 671                         copied += PAGE_SIZE;
 672
 673                         if (fatal_signal_pending(current))
 674                                 err = -EINTR;
 675                 }
 676                 if (err)
 677                         break;
 678         }
 679
 680 out_unlock:
 681         mmap_read_unlock(dst_mm);
 682 out:
 683         if (page)
 684                 put_page(page);
 685         BUG_ON(copied < 0);
 686         BUG_ON(err > 0);
 687         BUG_ON(!copied && !err);
 688         return copied ? copied : err;
 689 }
 690
 691 ssize_t mcopy_atomic(struct mm_struct *dst_mm, unsigned long dst_start,
 692                      unsigned long src_start, unsigned long len,
 693                      atomic_t *mmap_changing, __u64 mode)
 694 {
 695         return __mcopy_atomic(dst_mm, dst_start, src_start, len,
 696                               MCOPY_ATOMIC_NORMAL, mmap_changing, mode);
 697 }
 698
 699 ssize_t mfill_zeropage(struct mm_struct *dst_mm, unsigned long start,
 700                        unsigned long len, atomic_t *mmap_changing)
 701 {
 702         return __mcopy_atomic(dst_mm, start, 0, len, MCOPY_ATOMIC_ZEROPAGE,
 703                               mmap_changing, 0);
 704 }
 705
 706 ssize_t mcopy_continue(struct mm_struct *dst_mm, unsigned long start,
 707                        unsigned long len, atomic_t *mmap_changing)
 708 {
 709         return __mcopy_atomic(dst_mm, start, 0, len, MCOPY_ATOMIC_CONTINUE,
 710                               mmap_changing, 0);
 711 }
 712
 713 long uffd_wp_range(struct mm_struct *dst_mm, struct vm_area_struct *dst_vma,
 714                    unsigned long start, unsigned long len, bool enable_wp)
 715 {
 716         unsigned int mm_cp_flags;
 717         struct mmu_gather tlb;
 718         long ret;
 719
 720         if (enable_wp)
 721                 mm_cp_flags = MM_CP_UFFD_WP;
 722         else
 723                 mm_cp_flags = MM_CP_UFFD_WP_RESOLVE;
 724
 725         /*
 726          * vma->vm_page_prot already reflects that uffd-wp is enabled for this
 727          * VMA (see userfaultfd_set_vm_flags()) and that all PTEs are supposed
 728          * to be write-protected as default whenever protection changes.
 729          * Try upgrading write permissions manually.
 730          */
 731         if (!enable_wp && vma_wants_manual_pte_write_upgrade(dst_vma))
 732                 mm_cp_flags |= MM_CP_TRY_CHANGE_WRITABLE;
 733         tlb_gather_mmu(&tlb, dst_mm);
 734         ret = change_protection(&tlb, dst_vma, start, start + len, mm_cp_flags);
 735         tlb_finish_mmu(&tlb);
 736
 737         return ret;
 738 }
 739
 740 int mwriteprotect_range(struct mm_struct *dst_mm, unsigned long start,
 741                         unsigned long len, bool enable_wp,
 742                         atomic_t *mmap_changing)
 743 {
 744         struct vm_area_struct *dst_vma;
 745         unsigned long page_mask;
 746         long err;
 747
 748         /*
 749          * Sanitize the command parameters:
 750          */
 751         BUG_ON(start & ~PAGE_MASK);
 752         BUG_ON(len & ~PAGE_MASK);
 753
 754         /* Does the address range wrap, or is the span zero-sized? */
 755         BUG_ON(start + len <= start);
 756
 757         mmap_read_lock(dst_mm);
 758
 759         /*
 760          * If memory mappings are changing because of non-cooperative
 761          * operation (e.g. mremap) running in parallel, bail out and
 762          * request the user to retry later
 763          */
 764         err = -EAGAIN;
 765         if (mmap_changing && atomic_read(mmap_changing))
 766                 goto out_unlock;
 767
 768         err = -ENOENT;
 769         dst_vma = find_dst_vma(dst_mm, start, len);
 770
 771         if (!dst_vma)
 772                 goto out_unlock;
 773         if (!userfaultfd_wp(dst_vma))
 774                 goto out_unlock;
 775         if (!vma_can_userfault(dst_vma, dst_vma->vm_flags))
 776                 goto out_unlock;
 777
 778         if (is_vm_hugetlb_page(dst_vma)) {
 779                 err = -EINVAL;
 780                 page_mask = vma_kernel_pagesize(dst_vma) - 1;
 781                 if ((start & page_mask) || (len & page_mask))
 782                         goto out_unlock;
 783         }
 784
 785         err = uffd_wp_range(dst_mm, dst_vma, start, len, enable_wp);
 786
 787         /* Return 0 on success, <0 on failures */
 788         if (err > 0)
 789                 err = 0;
 790
 791 out_unlock:
 792         mmap_read_unlock(dst_mm);
 793         return err;
 794 }