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 int mcopy_atomic_pte(struct mm_struct *dst_mm,
  22                             pmd_t *dst_pmd,
  23                             struct vm_area_struct *dst_vma,
  24                             unsigned long dst_addr,
  25                             unsigned long src_addr,
  26                             struct page **pagep)
  27 {
  28         struct mem_cgroup *memcg;
  29         pte_t _dst_pte, *dst_pte;
  30         spinlock_t *ptl;
  31         void *page_kaddr;
  32         int ret;
  33         struct page *page;
  34         pgoff_t offset, max_off;
  35         struct inode *inode;
  36
  37         if (!*pagep) {
  38                 ret = -ENOMEM;
  39                 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, dst_vma, dst_addr);
  40                 if (!page)
  41                         goto out;
  42
  43                 page_kaddr = kmap_atomic(page);
  44                 ret = copy_from_user(page_kaddr,
  45                                      (const void __user *) src_addr,
  46                                      PAGE_SIZE);
  47                 kunmap_atomic(page_kaddr);
  48
  49                 /* fallback to copy_from_user outside mmap_sem */
  50                 if (unlikely(ret)) {
  51                         ret = -ENOENT;
  52                         *pagep = page;
  53                         /* don't free the page */
  54                         goto out;
  55                 }
  56
  57                 flush_dcache_page(page);
  58         } else {
  59                 page = *pagep;
  60                 *pagep = NULL;
  61         }
  62
  63         /*
  64          * The memory barrier inside __SetPageUptodate makes sure that
  65          * preceeding stores to the page contents become visible before
  66          * the set_pte_at() write.
  67          */
  68         __SetPageUptodate(page);
  69
  70         ret = -ENOMEM;
  71         if (mem_cgroup_try_charge(page, dst_mm, GFP_KERNEL, &memcg, false))
  72                 goto out_release;
  73
  74         _dst_pte = mk_pte(page, dst_vma->vm_page_prot);
  75         if (dst_vma->vm_flags & VM_WRITE)
  76                 _dst_pte = pte_mkwrite(pte_mkdirty(_dst_pte));
  77
  78         dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
  79         if (dst_vma->vm_file) {
  80                 /* the shmem MAP_PRIVATE case requires checking the i_size */
  81                 inode = dst_vma->vm_file->f_inode;
  82                 offset = linear_page_index(dst_vma, dst_addr);
  83                 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
  84                 ret = -EFAULT;
  85                 if (unlikely(offset >= max_off))
  86                         goto out_release_uncharge_unlock;
  87         }
  88         ret = -EEXIST;
  89         if (!pte_none(*dst_pte))
  90                 goto out_release_uncharge_unlock;
  91
  92         inc_mm_counter(dst_mm, MM_ANONPAGES);
  93         page_add_new_anon_rmap(page, dst_vma, dst_addr, false);
  94         mem_cgroup_commit_charge(page, memcg, false, false);
  95         lru_cache_add_active_or_unevictable(page, dst_vma);
  96
  97         set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
  98
  99         /* No need to invalidate - it was non-present before */
 100         update_mmu_cache(dst_vma, dst_addr, dst_pte);
 101
 102         pte_unmap_unlock(dst_pte, ptl);
 103         ret = 0;
 104 out:
 105         return ret;
 106 out_release_uncharge_unlock:
 107         pte_unmap_unlock(dst_pte, ptl);
 108         mem_cgroup_cancel_charge(page, memcg, false);
 109 out_release:
 110         put_page(page);
 111         goto out;
 112 }
 113
 114 static int mfill_zeropage_pte(struct mm_struct *dst_mm,
 115                               pmd_t *dst_pmd,
 116                               struct vm_area_struct *dst_vma,
 117                               unsigned long dst_addr)
 118 {
 119         pte_t _dst_pte, *dst_pte;
 120         spinlock_t *ptl;
 121         int ret;
 122         pgoff_t offset, max_off;
 123         struct inode *inode;
 124
 125         _dst_pte = pte_mkspecial(pfn_pte(my_zero_pfn(dst_addr),
 126                                          dst_vma->vm_page_prot));
 127         dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
 128         if (dst_vma->vm_file) {
 129                 /* the shmem MAP_PRIVATE case requires checking the i_size */
 130                 inode = dst_vma->vm_file->f_inode;
 131                 offset = linear_page_index(dst_vma, dst_addr);
 132                 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
 133                 ret = -EFAULT;
 134                 if (unlikely(offset >= max_off))
 135                         goto out_unlock;
 136         }
 137         ret = -EEXIST;
 138         if (!pte_none(*dst_pte))
 139                 goto out_unlock;
 140         set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
 141         /* No need to invalidate - it was non-present before */
 142         update_mmu_cache(dst_vma, dst_addr, dst_pte);
 143         ret = 0;
 144 out_unlock:
 145         pte_unmap_unlock(dst_pte, ptl);
 146         return ret;
 147 }
 148
 149 static pmd_t *mm_alloc_pmd(struct mm_struct *mm, unsigned long address)
 150 {
 151         pgd_t *pgd;
 152         p4d_t *p4d;
 153         pud_t *pud;
 154
 155         pgd = pgd_offset(mm, address);
 156         p4d = p4d_alloc(mm, pgd, address);
 157         if (!p4d)
 158                 return NULL;
 159         pud = pud_alloc(mm, p4d, address);
 160         if (!pud)
 161                 return NULL;
 162         /*
 163          * Note that we didn't run this because the pmd was
 164          * missing, the *pmd may be already established and in
 165          * turn it may also be a trans_huge_pmd.
 166          */
 167         return pmd_alloc(mm, pud, address);
 168 }
 169
 170 #ifdef CONFIG_HUGETLB_PAGE
 171 /*
 172  * __mcopy_atomic processing for HUGETLB vmas.  Note that this routine is
 173  * called with mmap_sem held, it will release mmap_sem before returning.
 174  */
 175 static __always_inline ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
 176                                               struct vm_area_struct *dst_vma,
 177                                               unsigned long dst_start,
 178                                               unsigned long src_start,
 179                                               unsigned long len,
 180                                               bool *mmap_changing,
 181                                               bool zeropage)
 182 {
 183         int vm_alloc_shared = dst_vma->vm_flags & VM_SHARED;
 184         int vm_shared = dst_vma->vm_flags & VM_SHARED;
 185         ssize_t err;
 186         pte_t *dst_pte;
 187         unsigned long src_addr, dst_addr;
 188         long copied;
 189         struct page *page;
 190         struct hstate *h;
 191         unsigned long vma_hpagesize;
 192         pgoff_t idx;
 193         u32 hash;
 194         struct address_space *mapping;
 195
 196         /*
 197          * There is no default zero huge page for all huge page sizes as
 198          * supported by hugetlb.  A PMD_SIZE huge pages may exist as used
 199          * by THP.  Since we can not reliably insert a zero page, this
 200          * feature is not supported.
 201          */
 202         if (zeropage) {
 203                 up_read(&dst_mm->mmap_sem);
 204                 return -EINVAL;
 205         }
 206
 207         src_addr = src_start;
 208         dst_addr = dst_start;
 209         copied = 0;
 210         page = NULL;
 211         vma_hpagesize = vma_kernel_pagesize(dst_vma);
 212
 213         /*
 214          * Validate alignment based on huge page size
 215          */
 216         err = -EINVAL;
 217         if (dst_start & (vma_hpagesize - 1) || len & (vma_hpagesize - 1))
 218                 goto out_unlock;
 219
 220 retry:
 221         /*
 222          * On routine entry dst_vma is set.  If we had to drop mmap_sem and
 223          * retry, dst_vma will be set to NULL and we must lookup again.
 224          */
 225         if (!dst_vma) {
 226                 err = -ENOENT;
 227                 dst_vma = find_vma(dst_mm, dst_start);
 228                 if (!dst_vma || !is_vm_hugetlb_page(dst_vma))
 229                         goto out_unlock;
 230                 /*
 231                  * Check the vma is registered in uffd, this is
 232                  * required to enforce the VM_MAYWRITE check done at
 233                  * uffd registration time.
 234                  */
 235                 if (!dst_vma->vm_userfaultfd_ctx.ctx)
 236                         goto out_unlock;
 237
 238                 if (dst_start < dst_vma->vm_start ||
 239                     dst_start + len > dst_vma->vm_end)
 240                         goto out_unlock;
 241
 242                 err = -EINVAL;
 243                 if (vma_hpagesize != vma_kernel_pagesize(dst_vma))
 244                         goto out_unlock;
 245
 246                 vm_shared = dst_vma->vm_flags & VM_SHARED;
 247         }
 248
 249         if (WARN_ON(dst_addr & (vma_hpagesize - 1) ||
 250                     (len - copied) & (vma_hpagesize - 1)))
 251                 goto out_unlock;
 252
 253         /*
 254          * If not shared, ensure the dst_vma has a anon_vma.
 255          */
 256         err = -ENOMEM;
 257         if (!vm_shared) {
 258                 if (unlikely(anon_vma_prepare(dst_vma)))
 259                         goto out_unlock;
 260         }
 261
 262         h = hstate_vma(dst_vma);
 263
 264         while (src_addr < src_start + len) {
 265                 pte_t dst_pteval;
 266
 267                 BUG_ON(dst_addr >= dst_start + len);
 268                 VM_BUG_ON(dst_addr & ~huge_page_mask(h));
 269
 270                 /*
 271                  * Serialize via hugetlb_fault_mutex
 272                  */
 273                 idx = linear_page_index(dst_vma, dst_addr);
 274                 mapping = dst_vma->vm_file->f_mapping;
 275                 hash = hugetlb_fault_mutex_hash(h, mapping, idx);
 276                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 277
 278                 err = -ENOMEM;
 279                 dst_pte = huge_pte_alloc(dst_mm, dst_addr, huge_page_size(h));
 280                 if (!dst_pte) {
 281                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 282                         goto out_unlock;
 283                 }
 284
 285                 err = -EEXIST;
 286                 dst_pteval = huge_ptep_get(dst_pte);
 287                 if (!huge_pte_none(dst_pteval)) {
 288                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 289                         goto out_unlock;
 290                 }
 291
 292                 err = hugetlb_mcopy_atomic_pte(dst_mm, dst_pte, dst_vma,
 293                                                 dst_addr, src_addr, &page);
 294
 295                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 296                 vm_alloc_shared = vm_shared;
 297
 298                 cond_resched();
 299
 300                 if (unlikely(err == -ENOENT)) {
 301                         up_read(&dst_mm->mmap_sem);
 302                         BUG_ON(!page);
 303
 304                         err = copy_huge_page_from_user(page,
 305                                                 (const void __user *)src_addr,
 306                                                 pages_per_huge_page(h), true);
 307                         if (unlikely(err)) {
 308                                 err = -EFAULT;
 309                                 goto out;
 310                         }
 311                         down_read(&dst_mm->mmap_sem);
 312                         /*
 313                          * If memory mappings are changing because of non-cooperative
 314                          * operation (e.g. mremap) running in parallel, bail out and
 315                          * request the user to retry later
 316                          */
 317                         if (mmap_changing && READ_ONCE(*mmap_changing)) {
 318                                 err = -EAGAIN;
 319                                 break;
 320                         }
 321
 322                         dst_vma = NULL;
 323                         goto retry;
 324                 } else
 325                         BUG_ON(page);
 326
 327                 if (!err) {
 328                         dst_addr += vma_hpagesize;
 329                         src_addr += vma_hpagesize;
 330                         copied += vma_hpagesize;
 331
 332                         if (fatal_signal_pending(current))
 333                                 err = -EINTR;
 334                 }
 335                 if (err)
 336                         break;
 337         }
 338
 339 out_unlock:
 340         up_read(&dst_mm->mmap_sem);
 341 out:
 342         if (page) {
 343                 /*
 344                  * We encountered an error and are about to free a newly
 345                  * allocated huge page.
 346                  *
 347                  * Reservation handling is very subtle, and is different for
 348                  * private and shared mappings.  See the routine
 349                  * restore_reserve_on_error for details.  Unfortunately, we
 350                  * can not call restore_reserve_on_error now as it would
 351                  * require holding mmap_sem.
 352                  *
 353                  * If a reservation for the page existed in the reservation
 354                  * map of a private mapping, the map was modified to indicate
 355                  * the reservation was consumed when the page was allocated.
 356                  * We clear the PagePrivate flag now so that the global
 357                  * reserve count will not be incremented in free_huge_page.
 358                  * The reservation map will still indicate the reservation
 359                  * was consumed and possibly prevent later page allocation.
 360                  * This is better than leaking a global reservation.  If no
 361                  * reservation existed, it is still safe to clear PagePrivate
 362                  * as no adjustments to reservation counts were made during
 363                  * allocation.
 364                  *
 365                  * The reservation map for shared mappings indicates which
 366                  * pages have reservations.  When a huge page is allocated
 367                  * for an address with a reservation, no change is made to
 368                  * the reserve map.  In this case PagePrivate will be set
 369                  * to indicate that the global reservation count should be
 370                  * incremented when the page is freed.  This is the desired
 371                  * behavior.  However, when a huge page is allocated for an
 372                  * address without a reservation a reservation entry is added
 373                  * to the reservation map, and PagePrivate will not be set.
 374                  * When the page is freed, the global reserve count will NOT
 375                  * be incremented and it will appear as though we have leaked
 376                  * reserved page.  In this case, set PagePrivate so that the
 377                  * global reserve count will be incremented to match the
 378                  * reservation map entry which was created.
 379                  *
 380                  * Note that vm_alloc_shared is based on the flags of the vma
 381                  * for which the page was originally allocated.  dst_vma could
 382                  * be different or NULL on error.
 383                  */
 384                 if (vm_alloc_shared)
 385                         SetPagePrivate(page);
 386                 else
 387                         ClearPagePrivate(page);
 388                 put_page(page);
 389         }
 390         BUG_ON(copied < 0);
 391         BUG_ON(err > 0);
 392         BUG_ON(!copied && !err);
 393         return copied ? copied : err;
 394 }
 395 #else /* !CONFIG_HUGETLB_PAGE */
 396 /* fail at build time if gcc attempts to use this */
 397 extern ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
 398                                       struct vm_area_struct *dst_vma,
 399                                       unsigned long dst_start,
 400                                       unsigned long src_start,
 401                                       unsigned long len,
 402                                       bool *mmap_changing,
 403                                       bool zeropage);
 404 #endif /* CONFIG_HUGETLB_PAGE */
 405
 406 static __always_inline ssize_t mfill_atomic_pte(struct mm_struct *dst_mm,
 407                                                 pmd_t *dst_pmd,
 408                                                 struct vm_area_struct *dst_vma,
 409                                                 unsigned long dst_addr,
 410                                                 unsigned long src_addr,
 411                                                 struct page **page,
 412                                                 bool zeropage)
 413 {
 414         ssize_t err;
 415
 416         /*
 417          * The normal page fault path for a shmem will invoke the
 418          * fault, fill the hole in the file and COW it right away. The
 419          * result generates plain anonymous memory. So when we are
 420          * asked to fill an hole in a MAP_PRIVATE shmem mapping, we'll
 421          * generate anonymous memory directly without actually filling
 422          * the hole. For the MAP_PRIVATE case the robustness check
 423          * only happens in the pagetable (to verify it's still none)
 424          * and not in the radix tree.
 425          */
 426         if (!(dst_vma->vm_flags & VM_SHARED)) {
 427                 if (!zeropage)
 428                         err = mcopy_atomic_pte(dst_mm, dst_pmd, dst_vma,
 429                                                dst_addr, src_addr, page);
 430                 else
 431                         err = mfill_zeropage_pte(dst_mm, dst_pmd,
 432                                                  dst_vma, dst_addr);
 433         } else {
 434                 if (!zeropage)
 435                         err = shmem_mcopy_atomic_pte(dst_mm, dst_pmd,
 436                                                      dst_vma, dst_addr,
 437                                                      src_addr, page);
 438                 else
 439                         err = shmem_mfill_zeropage_pte(dst_mm, dst_pmd,
 440                                                        dst_vma, dst_addr);
 441         }
 442
 443         return err;
 444 }
 445
 446 static __always_inline ssize_t __mcopy_atomic(struct mm_struct *dst_mm,
 447                                               unsigned long dst_start,
 448                                               unsigned long src_start,
 449                                               unsigned long len,
 450                                               bool zeropage,
 451                                               bool *mmap_changing)
 452 {
 453         struct vm_area_struct *dst_vma;
 454         ssize_t err;
 455         pmd_t *dst_pmd;
 456         unsigned long src_addr, dst_addr;
 457         long copied;
 458         struct page *page;
 459
 460         /*
 461          * Sanitize the command parameters:
 462          */
 463         BUG_ON(dst_start & ~PAGE_MASK);
 464         BUG_ON(len & ~PAGE_MASK);
 465
 466         /* Does the address range wrap, or is the span zero-sized? */
 467         BUG_ON(src_start + len <= src_start);
 468         BUG_ON(dst_start + len <= dst_start);
 469
 470         src_addr = src_start;
 471         dst_addr = dst_start;
 472         copied = 0;
 473         page = NULL;
 474 retry:
 475         down_read(&dst_mm->mmap_sem);
 476
 477         /*
 478          * If memory mappings are changing because of non-cooperative
 479          * operation (e.g. mremap) running in parallel, bail out and
 480          * request the user to retry later
 481          */
 482         err = -EAGAIN;
 483         if (mmap_changing && READ_ONCE(*mmap_changing))
 484                 goto out_unlock;
 485
 486         /*
 487          * Make sure the vma is not shared, that the dst range is
 488          * both valid and fully within a single existing vma.
 489          */
 490         err = -ENOENT;
 491         dst_vma = find_vma(dst_mm, dst_start);
 492         if (!dst_vma)
 493                 goto out_unlock;
 494         /*
 495          * Check the vma is registered in uffd, this is required to
 496          * enforce the VM_MAYWRITE check done at uffd registration
 497          * time.
 498          */
 499         if (!dst_vma->vm_userfaultfd_ctx.ctx)
 500                 goto out_unlock;
 501
 502         if (dst_start < dst_vma->vm_start ||
 503             dst_start + len > dst_vma->vm_end)
 504                 goto out_unlock;
 505
 506         err = -EINVAL;
 507         /*
 508          * shmem_zero_setup is invoked in mmap for MAP_ANONYMOUS|MAP_SHARED but
 509          * it will overwrite vm_ops, so vma_is_anonymous must return false.
 510          */
 511         if (WARN_ON_ONCE(vma_is_anonymous(dst_vma) &&
 512             dst_vma->vm_flags & VM_SHARED))
 513                 goto out_unlock;
 514
 515         /*
 516          * If this is a HUGETLB vma, pass off to appropriate routine
 517          */
 518         if (is_vm_hugetlb_page(dst_vma))
 519                 return  __mcopy_atomic_hugetlb(dst_mm, dst_vma, dst_start,
 520                                                src_start, len, mmap_changing,
 521                                                zeropage);
 522
 523         if (!vma_is_anonymous(dst_vma) && !vma_is_shmem(dst_vma))
 524                 goto out_unlock;
 525
 526         /*
 527          * Ensure the dst_vma has a anon_vma or this page
 528          * would get a NULL anon_vma when moved in the
 529          * dst_vma.
 530          */
 531         err = -ENOMEM;
 532         if (!(dst_vma->vm_flags & VM_SHARED) &&
 533             unlikely(anon_vma_prepare(dst_vma)))
 534                 goto out_unlock;
 535
 536         while (src_addr < src_start + len) {
 537                 pmd_t dst_pmdval;
 538
 539                 BUG_ON(dst_addr >= dst_start + len);
 540
 541                 dst_pmd = mm_alloc_pmd(dst_mm, dst_addr);
 542                 if (unlikely(!dst_pmd)) {
 543                         err = -ENOMEM;
 544                         break;
 545                 }
 546
 547                 dst_pmdval = pmd_read_atomic(dst_pmd);
 548                 /*
 549                  * If the dst_pmd is mapped as THP don't
 550                  * override it and just be strict.
 551                  */
 552                 if (unlikely(pmd_trans_huge(dst_pmdval))) {
 553                         err = -EEXIST;
 554                         break;
 555                 }
 556                 if (unlikely(pmd_none(dst_pmdval)) &&
 557                     unlikely(__pte_alloc(dst_mm, dst_pmd))) {
 558                         err = -ENOMEM;
 559                         break;
 560                 }
 561                 /* If an huge pmd materialized from under us fail */
 562                 if (unlikely(pmd_trans_huge(*dst_pmd))) {
 563                         err = -EFAULT;
 564                         break;
 565                 }
 566
 567                 BUG_ON(pmd_none(*dst_pmd));
 568                 BUG_ON(pmd_trans_huge(*dst_pmd));
 569
 570                 err = mfill_atomic_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
 571                                        src_addr, &page, zeropage);
 572                 cond_resched();
 573
 574                 if (unlikely(err == -ENOENT)) {
 575                         void *page_kaddr;
 576
 577                         up_read(&dst_mm->mmap_sem);
 578                         BUG_ON(!page);
 579
 580                         page_kaddr = kmap(page);
 581                         err = copy_from_user(page_kaddr,
 582                                              (const void __user *) src_addr,
 583                                              PAGE_SIZE);
 584                         kunmap(page);
 585                         if (unlikely(err)) {
 586                                 err = -EFAULT;
 587                                 goto out;
 588                         }
 589                         flush_dcache_page(page);
 590                         goto retry;
 591                 } else
 592                         BUG_ON(page);
 593
 594                 if (!err) {
 595                         dst_addr += PAGE_SIZE;
 596                         src_addr += PAGE_SIZE;
 597                         copied += PAGE_SIZE;
 598
 599                         if (fatal_signal_pending(current))
 600                                 err = -EINTR;
 601                 }
 602                 if (err)
 603                         break;
 604         }
 605
 606 out_unlock:
 607         up_read(&dst_mm->mmap_sem);
 608 out:
 609         if (page)
 610                 put_page(page);
 611         BUG_ON(copied < 0);
 612         BUG_ON(err > 0);
 613         BUG_ON(!copied && !err);
 614         return copied ? copied : err;
 615 }
 616
 617 ssize_t mcopy_atomic(struct mm_struct *dst_mm, unsigned long dst_start,
 618                      unsigned long src_start, unsigned long len,
 619                      bool *mmap_changing)
 620 {
 621         return __mcopy_atomic(dst_mm, dst_start, src_start, len, false,
 622                               mmap_changing);
 623 }
 624
 625 ssize_t mfill_zeropage(struct mm_struct *dst_mm, unsigned long start,
 626                        unsigned long len, bool *mmap_changing)
 627 {
 628         return __mcopy_atomic(dst_mm, start, 0, len, true, mmap_changing);
 629 }