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