mm/userfaultfd.c

   1 /*
   2  *  mm/userfaultfd.c
   3  *
   4  *  Copyright (C) 2015  Red Hat, Inc.
   5  *
   6  *  This work is licensed under the terms of the GNU GPL, version 2. See
   7  *  the COPYING file in the top-level directory.
   8  */
   9
  10 #include <linux/mm.h>
  11 #include <linux/sched/signal.h>
  12 #include <linux/pagemap.h>
  13 #include <linux/rmap.h>
  14 #include <linux/swap.h>
  15 #include <linux/swapops.h>
  16 #include <linux/userfaultfd_k.h>
  17 #include <linux/mmu_notifier.h>
  18 #include <linux/hugetlb.h>
  19 #include <linux/shmem_fs.h>
  20 #include <asm/tlbflush.h>
  21 #include "internal.h"
  22
  23 static int mcopy_atomic_pte(struct mm_struct *dst_mm,
  24                             pmd_t *dst_pmd,
  25                             struct vm_area_struct *dst_vma,
  26                             unsigned long dst_addr,
  27                             unsigned long src_addr,
  28                             struct page **pagep)
  29 {
  30         struct mem_cgroup *memcg;
  31         pte_t _dst_pte, *dst_pte;
  32         spinlock_t *ptl;
  33         void *page_kaddr;
  34         int ret;
  35         struct page *page;
  36         pgoff_t offset, max_off;
  37         struct inode *inode;
  38
  39         if (!*pagep) {
  40                 ret = -ENOMEM;
  41                 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, dst_vma, dst_addr);
  42                 if (!page)
  43                         goto out;
  44
  45                 page_kaddr = kmap_atomic(page);
  46                 ret = copy_from_user(page_kaddr,
  47                                      (const void __user *) src_addr,
  48                                      PAGE_SIZE);
  49                 kunmap_atomic(page_kaddr);
  50
  51                 /* fallback to copy_from_user outside mmap_sem */
  52                 if (unlikely(ret)) {
  53                         ret = -ENOENT;
  54                         *pagep = page;
  55                         /* don't free the page */
  56                         goto out;
  57                 }
  58
  59                 flush_dcache_page(page);
  60         } else {
  61                 page = *pagep;
  62                 *pagep = NULL;
  63         }
  64
  65         /*
  66          * The memory barrier inside __SetPageUptodate makes sure that
  67          * preceeding stores to the page contents become visible before
  68          * the set_pte_at() write.
  69          */
  70         __SetPageUptodate(page);
  71
  72         ret = -ENOMEM;
  73         if (mem_cgroup_try_charge(page, dst_mm, GFP_KERNEL, &memcg, false))
  74                 goto out_release;
  75
  76         _dst_pte = mk_pte(page, dst_vma->vm_page_prot);
  77         if (dst_vma->vm_flags & VM_WRITE)
  78                 _dst_pte = pte_mkwrite(pte_mkdirty(_dst_pte));
  79
  80         dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
  81         if (dst_vma->vm_file) {
  82                 /* the shmem MAP_PRIVATE case requires checking the i_size */
  83                 inode = dst_vma->vm_file->f_inode;
  84                 offset = linear_page_index(dst_vma, dst_addr);
  85                 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
  86                 ret = -EFAULT;
  87                 if (unlikely(offset >= max_off))
  88                         goto out_release_uncharge_unlock;
  89         }
  90         ret = -EEXIST;
  91         if (!pte_none(*dst_pte))
  92                 goto out_release_uncharge_unlock;
  93
  94         inc_mm_counter(dst_mm, MM_ANONPAGES);
  95         page_add_new_anon_rmap(page, dst_vma, dst_addr, false);
  96         mem_cgroup_commit_charge(page, memcg, false, false);
  97         lru_cache_add_active_or_unevictable(page, dst_vma);
  98
  99         set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
 100
 101         /* No need to invalidate - it was non-present before */
 102         update_mmu_cache(dst_vma, dst_addr, dst_pte);
 103
 104         pte_unmap_unlock(dst_pte, ptl);
 105         ret = 0;
 106 out:
 107         return ret;
 108 out_release_uncharge_unlock:
 109         pte_unmap_unlock(dst_pte, ptl);
 110         mem_cgroup_cancel_charge(page, memcg, false);
 111 out_release:
 112         put_page(page);
 113         goto out;
 114 }
 115
 116 static int mfill_zeropage_pte(struct mm_struct *dst_mm,
 117                               pmd_t *dst_pmd,
 118                               struct vm_area_struct *dst_vma,
 119                               unsigned long dst_addr)
 120 {
 121         pte_t _dst_pte, *dst_pte;
 122         spinlock_t *ptl;
 123         int ret;
 124         pgoff_t offset, max_off;
 125         struct inode *inode;
 126
 127         _dst_pte = pte_mkspecial(pfn_pte(my_zero_pfn(dst_addr),
 128                                          dst_vma->vm_page_prot));
 129         dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
 130         if (dst_vma->vm_file) {
 131                 /* the shmem MAP_PRIVATE case requires checking the i_size */
 132                 inode = dst_vma->vm_file->f_inode;
 133                 offset = linear_page_index(dst_vma, dst_addr);
 134                 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
 135                 ret = -EFAULT;
 136                 if (unlikely(offset >= max_off))
 137                         goto out_unlock;
 138         }
 139         ret = -EEXIST;
 140         if (!pte_none(*dst_pte))
 141                 goto out_unlock;
 142         set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
 143         /* No need to invalidate - it was non-present before */
 144         update_mmu_cache(dst_vma, dst_addr, dst_pte);
 145         ret = 0;
 146 out_unlock:
 147         pte_unmap_unlock(dst_pte, ptl);
 148         return ret;
 149 }
 150
 151 static pmd_t *mm_alloc_pmd(struct mm_struct *mm, unsigned long address)
 152 {
 153         pgd_t *pgd;
 154         p4d_t *p4d;
 155         pud_t *pud;
 156
 157         pgd = pgd_offset(mm, address);
 158         p4d = p4d_alloc(mm, pgd, address);
 159         if (!p4d)
 160                 return NULL;
 161         pud = pud_alloc(mm, p4d, address);
 162         if (!pud)
 163                 return NULL;
 164         /*
 165          * Note that we didn't run this because the pmd was
 166          * missing, the *pmd may be already established and in
 167          * turn it may also be a trans_huge_pmd.
 168          */
 169         return pmd_alloc(mm, pud, address);
 170 }
 171
 172 #ifdef CONFIG_HUGETLB_PAGE
 173 /*
 174  * __mcopy_atomic processing for HUGETLB vmas.  Note that this routine is
 175  * called with mmap_sem held, it will release mmap_sem before returning.
 176  */
 177 static __always_inline ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
 178                                               struct vm_area_struct *dst_vma,
 179                                               unsigned long dst_start,
 180                                               unsigned long src_start,
 181                                               unsigned long len,
 182                                               bool zeropage)
 183 {
 184         int vm_alloc_shared = dst_vma->vm_flags & VM_SHARED;
 185         int vm_shared = dst_vma->vm_flags & VM_SHARED;
 186         ssize_t err;
 187         pte_t *dst_pte;
 188         unsigned long src_addr, dst_addr;
 189         long copied;
 190         struct page *page;
 191         struct hstate *h;
 192         unsigned long vma_hpagesize;
 193         pgoff_t idx;
 194         u32 hash;
 195         struct address_space *mapping;
 196
 197         /*
 198          * There is no default zero huge page for all huge page sizes as
 199          * supported by hugetlb.  A PMD_SIZE huge pages may exist as used
 200          * by THP.  Since we can not reliably insert a zero page, this
 201          * feature is not supported.
 202          */
 203         if (zeropage) {
 204                 up_read(&dst_mm->mmap_sem);
 205                 return -EINVAL;
 206         }
 207
 208         src_addr = src_start;
 209         dst_addr = dst_start;
 210         copied = 0;
 211         page = NULL;
 212         vma_hpagesize = vma_kernel_pagesize(dst_vma);
 213
 214         /*
 215          * Validate alignment based on huge page size
 216          */
 217         err = -EINVAL;
 218         if (dst_start & (vma_hpagesize - 1) || len & (vma_hpagesize - 1))
 219                 goto out_unlock;
 220
 221 retry:
 222         /*
 223          * On routine entry dst_vma is set.  If we had to drop mmap_sem and
 224          * retry, dst_vma will be set to NULL and we must lookup again.
 225          */
 226         if (!dst_vma) {
 227                 err = -ENOENT;
 228                 dst_vma = find_vma(dst_mm, dst_start);
 229                 if (!dst_vma || !is_vm_hugetlb_page(dst_vma))
 230                         goto out_unlock;
 231                 /*
 232                  * Check the vma is registered in uffd, this is
 233                  * required to enforce the VM_MAYWRITE check done at
 234                  * uffd registration time.
 235                  */
 236                 if (!dst_vma->vm_userfaultfd_ctx.ctx)
 237                         goto out_unlock;
 238
 239                 if (dst_start < dst_vma->vm_start ||
 240                     dst_start + len > dst_vma->vm_end)
 241                         goto out_unlock;
 242
 243                 err = -EINVAL;
 244                 if (vma_hpagesize != vma_kernel_pagesize(dst_vma))
 245                         goto out_unlock;
 246
 247                 vm_shared = dst_vma->vm_flags & VM_SHARED;
 248         }
 249
 250         if (WARN_ON(dst_addr & (vma_hpagesize - 1) ||
 251                     (len - copied) & (vma_hpagesize - 1)))
 252                 goto out_unlock;
 253
 254         /*
 255          * If not shared, ensure the dst_vma has a anon_vma.
 256          */
 257         err = -ENOMEM;
 258         if (!vm_shared) {
 259                 if (unlikely(anon_vma_prepare(dst_vma)))
 260                         goto out_unlock;
 261         }
 262
 263         h = hstate_vma(dst_vma);
 264
 265         while (src_addr < src_start + len) {
 266                 pte_t dst_pteval;
 267
 268                 BUG_ON(dst_addr >= dst_start + len);
 269                 VM_BUG_ON(dst_addr & ~huge_page_mask(h));
 270
 271                 /*
 272                  * Serialize via hugetlb_fault_mutex
 273                  */
 274                 idx = linear_page_index(dst_vma, dst_addr);
 275                 mapping = dst_vma->vm_file->f_mapping;
 276                 hash = hugetlb_fault_mutex_hash(h, mapping, idx);
 277                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 278
 279                 err = -ENOMEM;
 280                 dst_pte = huge_pte_alloc(dst_mm, dst_addr, huge_page_size(h));
 281                 if (!dst_pte) {
 282                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 283                         goto out_unlock;
 284                 }
 285
 286                 err = -EEXIST;
 287                 dst_pteval = huge_ptep_get(dst_pte);
 288                 if (!huge_pte_none(dst_pteval)) {
 289                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 290                         goto out_unlock;
 291                 }
 292
 293                 err = hugetlb_mcopy_atomic_pte(dst_mm, dst_pte, dst_vma,
 294                                                 dst_addr, src_addr, &page);
 295
 296                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 297                 vm_alloc_shared = vm_shared;
 298
 299                 cond_resched();
 300
 301                 if (unlikely(err == -ENOENT)) {
 302                         up_read(&dst_mm->mmap_sem);
 303                         BUG_ON(!page);
 304
 305                         err = copy_huge_page_from_user(page,
 306                                                 (const void __user *)src_addr,
 307                                                 pages_per_huge_page(h), true);
 308                         if (unlikely(err)) {
 309                                 err = -EFAULT;
 310                                 goto out;
 311                         }
 312                         down_read(&dst_mm->mmap_sem);
 313
 314                         dst_vma = NULL;
 315                         goto retry;
 316                 } else
 317                         BUG_ON(page);
 318
 319                 if (!err) {
 320                         dst_addr += vma_hpagesize;
 321                         src_addr += vma_hpagesize;
 322                         copied += vma_hpagesize;
 323
 324                         if (fatal_signal_pending(current))
 325                                 err = -EINTR;
 326                 }
 327                 if (err)
 328                         break;
 329         }
 330
 331 out_unlock:
 332         up_read(&dst_mm->mmap_sem);
 333 out:
 334         if (page) {
 335                 /*
 336                  * We encountered an error and are about to free a newly
 337                  * allocated huge page.
 338                  *
 339                  * Reservation handling is very subtle, and is different for
 340                  * private and shared mappings.  See the routine
 341                  * restore_reserve_on_error for details.  Unfortunately, we
 342                  * can not call restore_reserve_on_error now as it would
 343                  * require holding mmap_sem.
 344                  *
 345                  * If a reservation for the page existed in the reservation
 346                  * map of a private mapping, the map was modified to indicate
 347                  * the reservation was consumed when the page was allocated.
 348                  * We clear the PagePrivate flag now so that the global
 349                  * reserve count will not be incremented in free_huge_page.
 350                  * The reservation map will still indicate the reservation
 351                  * was consumed and possibly prevent later page allocation.
 352                  * This is better than leaking a global reservation.  If no
 353                  * reservation existed, it is still safe to clear PagePrivate
 354                  * as no adjustments to reservation counts were made during
 355                  * allocation.
 356                  *
 357                  * The reservation map for shared mappings indicates which
 358                  * pages have reservations.  When a huge page is allocated
 359                  * for an address with a reservation, no change is made to
 360                  * the reserve map.  In this case PagePrivate will be set
 361                  * to indicate that the global reservation count should be
 362                  * incremented when the page is freed.  This is the desired
 363                  * behavior.  However, when a huge page is allocated for an
 364                  * address without a reservation a reservation entry is added
 365                  * to the reservation map, and PagePrivate will not be set.
 366                  * When the page is freed, the global reserve count will NOT
 367                  * be incremented and it will appear as though we have leaked
 368                  * reserved page.  In this case, set PagePrivate so that the
 369                  * global reserve count will be incremented to match the
 370                  * reservation map entry which was created.
 371                  *
 372                  * Note that vm_alloc_shared is based on the flags of the vma
 373                  * for which the page was originally allocated.  dst_vma could
 374                  * be different or NULL on error.
 375                  */
 376                 if (vm_alloc_shared)
 377                         SetPagePrivate(page);
 378                 else
 379                         ClearPagePrivate(page);
 380                 put_page(page);
 381         }
 382         BUG_ON(copied < 0);
 383         BUG_ON(err > 0);
 384         BUG_ON(!copied && !err);
 385         return copied ? copied : err;
 386 }
 387 #else /* !CONFIG_HUGETLB_PAGE */
 388 /* fail at build time if gcc attempts to use this */
 389 extern ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
 390                                       struct vm_area_struct *dst_vma,
 391                                       unsigned long dst_start,
 392                                       unsigned long src_start,
 393                                       unsigned long len,
 394                                       bool zeropage);
 395 #endif /* CONFIG_HUGETLB_PAGE */
 396
 397 static __always_inline ssize_t mfill_atomic_pte(struct mm_struct *dst_mm,
 398                                                 pmd_t *dst_pmd,
 399                                                 struct vm_area_struct *dst_vma,
 400                                                 unsigned long dst_addr,
 401                                                 unsigned long src_addr,
 402                                                 struct page **page,
 403                                                 bool zeropage)
 404 {
 405         ssize_t err;
 406
 407         /*
 408          * The normal page fault path for a shmem will invoke the
 409          * fault, fill the hole in the file and COW it right away. The
 410          * result generates plain anonymous memory. So when we are
 411          * asked to fill an hole in a MAP_PRIVATE shmem mapping, we'll
 412          * generate anonymous memory directly without actually filling
 413          * the hole. For the MAP_PRIVATE case the robustness check
 414          * only happens in the pagetable (to verify it's still none)
 415          * and not in the radix tree.
 416          */
 417         if (!(dst_vma->vm_flags & VM_SHARED)) {
 418                 if (!zeropage)
 419                         err = mcopy_atomic_pte(dst_mm, dst_pmd, dst_vma,
 420                                                dst_addr, src_addr, page);
 421                 else
 422                         err = mfill_zeropage_pte(dst_mm, dst_pmd,
 423                                                  dst_vma, dst_addr);
 424         } else {
 425                 if (!zeropage)
 426                         err = shmem_mcopy_atomic_pte(dst_mm, dst_pmd,
 427                                                      dst_vma, dst_addr,
 428                                                      src_addr, page);
 429                 else
 430                         err = shmem_mfill_zeropage_pte(dst_mm, dst_pmd,
 431                                                        dst_vma, dst_addr);
 432         }
 433
 434         return err;
 435 }
 436
 437 static __always_inline ssize_t __mcopy_atomic(struct mm_struct *dst_mm,
 438                                               unsigned long dst_start,
 439                                               unsigned long src_start,
 440                                               unsigned long len,
 441                                               bool zeropage,
 442                                               bool *mmap_changing)
 443 {
 444         struct vm_area_struct *dst_vma;
 445         ssize_t err;
 446         pmd_t *dst_pmd;
 447         unsigned long src_addr, dst_addr;
 448         long copied;
 449         struct page *page;
 450
 451         /*
 452          * Sanitize the command parameters:
 453          */
 454         BUG_ON(dst_start & ~PAGE_MASK);
 455         BUG_ON(len & ~PAGE_MASK);
 456
 457         /* Does the address range wrap, or is the span zero-sized? */
 458         BUG_ON(src_start + len <= src_start);
 459         BUG_ON(dst_start + len <= dst_start);
 460
 461         src_addr = src_start;
 462         dst_addr = dst_start;
 463         copied = 0;
 464         page = NULL;
 465 retry:
 466         down_read(&dst_mm->mmap_sem);
 467
 468         /*
 469          * If memory mappings are changing because of non-cooperative
 470          * operation (e.g. mremap) running in parallel, bail out and
 471          * request the user to retry later
 472          */
 473         err = -EAGAIN;
 474         if (mmap_changing && READ_ONCE(*mmap_changing))
 475                 goto out_unlock;
 476
 477         /*
 478          * Make sure the vma is not shared, that the dst range is
 479          * both valid and fully within a single existing vma.
 480          */
 481         err = -ENOENT;
 482         dst_vma = find_vma(dst_mm, dst_start);
 483         if (!dst_vma)
 484                 goto out_unlock;
 485         /*
 486          * Check the vma is registered in uffd, this is required to
 487          * enforce the VM_MAYWRITE check done at uffd registration
 488          * time.
 489          */
 490         if (!dst_vma->vm_userfaultfd_ctx.ctx)
 491                 goto out_unlock;
 492
 493         if (dst_start < dst_vma->vm_start ||
 494             dst_start + len > dst_vma->vm_end)
 495                 goto out_unlock;
 496
 497         err = -EINVAL;
 498         /*
 499          * shmem_zero_setup is invoked in mmap for MAP_ANONYMOUS|MAP_SHARED but
 500          * it will overwrite vm_ops, so vma_is_anonymous must return false.
 501          */
 502         if (WARN_ON_ONCE(vma_is_anonymous(dst_vma) &&
 503             dst_vma->vm_flags & VM_SHARED))
 504                 goto out_unlock;
 505
 506         /*
 507          * If this is a HUGETLB vma, pass off to appropriate routine
 508          */
 509         if (is_vm_hugetlb_page(dst_vma))
 510                 return  __mcopy_atomic_hugetlb(dst_mm, dst_vma, dst_start,
 511                                                 src_start, len, zeropage);
 512
 513         if (!vma_is_anonymous(dst_vma) && !vma_is_shmem(dst_vma))
 514                 goto out_unlock;
 515
 516         /*
 517          * Ensure the dst_vma has a anon_vma or this page
 518          * would get a NULL anon_vma when moved in the
 519          * dst_vma.
 520          */
 521         err = -ENOMEM;
 522         if (!(dst_vma->vm_flags & VM_SHARED) &&
 523             unlikely(anon_vma_prepare(dst_vma)))
 524                 goto out_unlock;
 525
 526         while (src_addr < src_start + len) {
 527                 pmd_t dst_pmdval;
 528
 529                 BUG_ON(dst_addr >= dst_start + len);
 530
 531                 dst_pmd = mm_alloc_pmd(dst_mm, dst_addr);
 532                 if (unlikely(!dst_pmd)) {
 533                         err = -ENOMEM;
 534                         break;
 535                 }
 536
 537                 dst_pmdval = pmd_read_atomic(dst_pmd);
 538                 /*
 539                  * If the dst_pmd is mapped as THP don't
 540                  * override it and just be strict.
 541                  */
 542                 if (unlikely(pmd_trans_huge(dst_pmdval))) {
 543                         err = -EEXIST;
 544                         break;
 545                 }
 546                 if (unlikely(pmd_none(dst_pmdval)) &&
 547                     unlikely(__pte_alloc(dst_mm, dst_pmd, dst_addr))) {
 548                         err = -ENOMEM;
 549                         break;
 550                 }
 551                 /* If an huge pmd materialized from under us fail */
 552                 if (unlikely(pmd_trans_huge(*dst_pmd))) {
 553                         err = -EFAULT;
 554                         break;
 555                 }
 556
 557                 BUG_ON(pmd_none(*dst_pmd));
 558                 BUG_ON(pmd_trans_huge(*dst_pmd));
 559
 560                 err = mfill_atomic_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
 561                                        src_addr, &page, zeropage);
 562                 cond_resched();
 563
 564                 if (unlikely(err == -ENOENT)) {
 565                         void *page_kaddr;
 566
 567                         up_read(&dst_mm->mmap_sem);
 568                         BUG_ON(!page);
 569
 570                         page_kaddr = kmap(page);
 571                         err = copy_from_user(page_kaddr,
 572                                              (const void __user *) src_addr,
 573                                              PAGE_SIZE);
 574                         kunmap(page);
 575                         if (unlikely(err)) {
 576                                 err = -EFAULT;
 577                                 goto out;
 578                         }
 579                         flush_dcache_page(page);
 580                         goto retry;
 581                 } else
 582                         BUG_ON(page);
 583
 584                 if (!err) {
 585                         dst_addr += PAGE_SIZE;
 586                         src_addr += PAGE_SIZE;
 587                         copied += PAGE_SIZE;
 588
 589                         if (fatal_signal_pending(current))
 590                                 err = -EINTR;
 591                 }
 592                 if (err)
 593                         break;
 594         }
 595
 596 out_unlock:
 597         up_read(&dst_mm->mmap_sem);
 598 out:
 599         if (page)
 600                 put_page(page);
 601         BUG_ON(copied < 0);
 602         BUG_ON(err > 0);
 603         BUG_ON(!copied && !err);
 604         return copied ? copied : err;
 605 }
 606
 607 ssize_t mcopy_atomic(struct mm_struct *dst_mm, unsigned long dst_start,
 608                      unsigned long src_start, unsigned long len,
 609                      bool *mmap_changing)
 610 {
 611         return __mcopy_atomic(dst_mm, dst_start, src_start, len, false,
 612                               mmap_changing);
 613 }
 614
 615 ssize_t mfill_zeropage(struct mm_struct *dst_mm, unsigned long start,
 616                        unsigned long len, bool *mmap_changing)
 617 {
 618         return __mcopy_atomic(dst_mm, start, 0, len, true, mmap_changing);
 619 }