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 = -ENOENT;
  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                  * Check the vma is registered in uffd, this is
 210                  * required to enforce the VM_MAYWRITE check done at
 211                  * uffd registration time.
 212                  */
 213                 if (!dst_vma->vm_userfaultfd_ctx.ctx)
 214                         goto out_unlock;
 215
 216                 if (dst_start < dst_vma->vm_start ||
 217                     dst_start + len > dst_vma->vm_end)
 218                         goto out_unlock;
 219
 220                 err = -EINVAL;
 221                 if (vma_hpagesize != vma_kernel_pagesize(dst_vma))
 222                         goto out_unlock;
 223
 224                 vm_shared = dst_vma->vm_flags & VM_SHARED;
 225         }
 226
 227         if (WARN_ON(dst_addr & (vma_hpagesize - 1) ||
 228                     (len - copied) & (vma_hpagesize - 1)))
 229                 goto out_unlock;
 230
 231         /*
 232          * If not shared, ensure the dst_vma has a anon_vma.
 233          */
 234         err = -ENOMEM;
 235         if (!vm_shared) {
 236                 if (unlikely(anon_vma_prepare(dst_vma)))
 237                         goto out_unlock;
 238         }
 239
 240         h = hstate_vma(dst_vma);
 241
 242         while (src_addr < src_start + len) {
 243                 pte_t dst_pteval;
 244
 245                 BUG_ON(dst_addr >= dst_start + len);
 246                 VM_BUG_ON(dst_addr & ~huge_page_mask(h));
 247
 248                 /*
 249                  * Serialize via hugetlb_fault_mutex
 250                  */
 251                 idx = linear_page_index(dst_vma, dst_addr);
 252                 mapping = dst_vma->vm_file->f_mapping;
 253                 hash = hugetlb_fault_mutex_hash(h, dst_mm, dst_vma, mapping,
 254                                                                 idx, dst_addr);
 255                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 256
 257                 err = -ENOMEM;
 258                 dst_pte = huge_pte_alloc(dst_mm, dst_addr, huge_page_size(h));
 259                 if (!dst_pte) {
 260                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 261                         goto out_unlock;
 262                 }
 263
 264                 err = -EEXIST;
 265                 dst_pteval = huge_ptep_get(dst_pte);
 266                 if (!huge_pte_none(dst_pteval)) {
 267                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 268                         goto out_unlock;
 269                 }
 270
 271                 err = hugetlb_mcopy_atomic_pte(dst_mm, dst_pte, dst_vma,
 272                                                 dst_addr, src_addr, &page);
 273
 274                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 275                 vm_alloc_shared = vm_shared;
 276
 277                 cond_resched();
 278
 279                 if (unlikely(err == -ENOENT)) {
 280                         up_read(&dst_mm->mmap_sem);
 281                         BUG_ON(!page);
 282
 283                         err = copy_huge_page_from_user(page,
 284                                                 (const void __user *)src_addr,
 285                                                 pages_per_huge_page(h), true);
 286                         if (unlikely(err)) {
 287                                 err = -EFAULT;
 288                                 goto out;
 289                         }
 290                         down_read(&dst_mm->mmap_sem);
 291
 292                         dst_vma = NULL;
 293                         goto retry;
 294                 } else
 295                         BUG_ON(page);
 296
 297                 if (!err) {
 298                         dst_addr += vma_hpagesize;
 299                         src_addr += vma_hpagesize;
 300                         copied += vma_hpagesize;
 301
 302                         if (fatal_signal_pending(current))
 303                                 err = -EINTR;
 304                 }
 305                 if (err)
 306                         break;
 307         }
 308
 309 out_unlock:
 310         up_read(&dst_mm->mmap_sem);
 311 out:
 312         if (page) {
 313                 /*
 314                  * We encountered an error and are about to free a newly
 315                  * allocated huge page.
 316                  *
 317                  * Reservation handling is very subtle, and is different for
 318                  * private and shared mappings.  See the routine
 319                  * restore_reserve_on_error for details.  Unfortunately, we
 320                  * can not call restore_reserve_on_error now as it would
 321                  * require holding mmap_sem.
 322                  *
 323                  * If a reservation for the page existed in the reservation
 324                  * map of a private mapping, the map was modified to indicate
 325                  * the reservation was consumed when the page was allocated.
 326                  * We clear the PagePrivate flag now so that the global
 327                  * reserve count will not be incremented in free_huge_page.
 328                  * The reservation map will still indicate the reservation
 329                  * was consumed and possibly prevent later page allocation.
 330                  * This is better than leaking a global reservation.  If no
 331                  * reservation existed, it is still safe to clear PagePrivate
 332                  * as no adjustments to reservation counts were made during
 333                  * allocation.
 334                  *
 335                  * The reservation map for shared mappings indicates which
 336                  * pages have reservations.  When a huge page is allocated
 337                  * for an address with a reservation, no change is made to
 338                  * the reserve map.  In this case PagePrivate will be set
 339                  * to indicate that the global reservation count should be
 340                  * incremented when the page is freed.  This is the desired
 341                  * behavior.  However, when a huge page is allocated for an
 342                  * address without a reservation a reservation entry is added
 343                  * to the reservation map, and PagePrivate will not be set.
 344                  * When the page is freed, the global reserve count will NOT
 345                  * be incremented and it will appear as though we have leaked
 346                  * reserved page.  In this case, set PagePrivate so that the
 347                  * global reserve count will be incremented to match the
 348                  * reservation map entry which was created.
 349                  *
 350                  * Note that vm_alloc_shared is based on the flags of the vma
 351                  * for which the page was originally allocated.  dst_vma could
 352                  * be different or NULL on error.
 353                  */
 354                 if (vm_alloc_shared)
 355                         SetPagePrivate(page);
 356                 else
 357                         ClearPagePrivate(page);
 358                 put_page(page);
 359         }
 360         BUG_ON(copied < 0);
 361         BUG_ON(err > 0);
 362         BUG_ON(!copied && !err);
 363         return copied ? copied : err;
 364 }
 365 #else /* !CONFIG_HUGETLB_PAGE */
 366 /* fail at build time if gcc attempts to use this */
 367 extern ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
 368                                       struct vm_area_struct *dst_vma,
 369                                       unsigned long dst_start,
 370                                       unsigned long src_start,
 371                                       unsigned long len,
 372                                       bool zeropage);
 373 #endif /* CONFIG_HUGETLB_PAGE */
 374
 375 static __always_inline ssize_t mfill_atomic_pte(struct mm_struct *dst_mm,
 376                                                 pmd_t *dst_pmd,
 377                                                 struct vm_area_struct *dst_vma,
 378                                                 unsigned long dst_addr,
 379                                                 unsigned long src_addr,
 380                                                 struct page **page,
 381                                                 bool zeropage)
 382 {
 383         ssize_t err;
 384
 385         /*
 386          * The normal page fault path for a shmem will invoke the
 387          * fault, fill the hole in the file and COW it right away. The
 388          * result generates plain anonymous memory. So when we are
 389          * asked to fill an hole in a MAP_PRIVATE shmem mapping, we'll
 390          * generate anonymous memory directly without actually filling
 391          * the hole. For the MAP_PRIVATE case the robustness check
 392          * only happens in the pagetable (to verify it's still none)
 393          * and not in the radix tree.
 394          */
 395         if (!(dst_vma->vm_flags & VM_SHARED)) {
 396                 if (!zeropage)
 397                         err = mcopy_atomic_pte(dst_mm, dst_pmd, dst_vma,
 398                                                dst_addr, src_addr, page);
 399                 else
 400                         err = mfill_zeropage_pte(dst_mm, dst_pmd,
 401                                                  dst_vma, dst_addr);
 402         } else {
 403                 if (!zeropage)
 404                         err = shmem_mcopy_atomic_pte(dst_mm, dst_pmd,
 405                                                      dst_vma, dst_addr,
 406                                                      src_addr, page);
 407                 else
 408                         err = shmem_mfill_zeropage_pte(dst_mm, dst_pmd,
 409                                                        dst_vma, dst_addr);
 410         }
 411
 412         return err;
 413 }
 414
 415 static __always_inline ssize_t __mcopy_atomic(struct mm_struct *dst_mm,
 416                                               unsigned long dst_start,
 417                                               unsigned long src_start,
 418                                               unsigned long len,
 419                                               bool zeropage)
 420 {
 421         struct vm_area_struct *dst_vma;
 422         ssize_t err;
 423         pmd_t *dst_pmd;
 424         unsigned long src_addr, dst_addr;
 425         long copied;
 426         struct page *page;
 427
 428         /*
 429          * Sanitize the command parameters:
 430          */
 431         BUG_ON(dst_start & ~PAGE_MASK);
 432         BUG_ON(len & ~PAGE_MASK);
 433
 434         /* Does the address range wrap, or is the span zero-sized? */
 435         BUG_ON(src_start + len <= src_start);
 436         BUG_ON(dst_start + len <= dst_start);
 437
 438         src_addr = src_start;
 439         dst_addr = dst_start;
 440         copied = 0;
 441         page = NULL;
 442 retry:
 443         down_read(&dst_mm->mmap_sem);
 444
 445         /*
 446          * Make sure the vma is not shared, that the dst range is
 447          * both valid and fully within a single existing vma.
 448          */
 449         err = -ENOENT;
 450         dst_vma = find_vma(dst_mm, dst_start);
 451         if (!dst_vma)
 452                 goto out_unlock;
 453         /*
 454          * Check the vma is registered in uffd, this is required to
 455          * enforce the VM_MAYWRITE check done at uffd registration
 456          * time.
 457          */
 458         if (!dst_vma->vm_userfaultfd_ctx.ctx)
 459                 goto out_unlock;
 460
 461         if (dst_start < dst_vma->vm_start ||
 462             dst_start + len > dst_vma->vm_end)
 463                 goto out_unlock;
 464
 465         err = -EINVAL;
 466         /*
 467          * shmem_zero_setup is invoked in mmap for MAP_ANONYMOUS|MAP_SHARED but
 468          * it will overwrite vm_ops, so vma_is_anonymous must return false.
 469          */
 470         if (WARN_ON_ONCE(vma_is_anonymous(dst_vma) &&
 471             dst_vma->vm_flags & VM_SHARED))
 472                 goto out_unlock;
 473
 474         /*
 475          * If this is a HUGETLB vma, pass off to appropriate routine
 476          */
 477         if (is_vm_hugetlb_page(dst_vma))
 478                 return  __mcopy_atomic_hugetlb(dst_mm, dst_vma, dst_start,
 479                                                 src_start, len, zeropage);
 480
 481         if (!vma_is_anonymous(dst_vma) && !vma_is_shmem(dst_vma))
 482                 goto out_unlock;
 483
 484         /*
 485          * Ensure the dst_vma has a anon_vma or this page
 486          * would get a NULL anon_vma when moved in the
 487          * dst_vma.
 488          */
 489         err = -ENOMEM;
 490         if (!(dst_vma->vm_flags & VM_SHARED) &&
 491             unlikely(anon_vma_prepare(dst_vma)))
 492                 goto out_unlock;
 493
 494         while (src_addr < src_start + len) {
 495                 pmd_t dst_pmdval;
 496
 497                 BUG_ON(dst_addr >= dst_start + len);
 498
 499                 dst_pmd = mm_alloc_pmd(dst_mm, dst_addr);
 500                 if (unlikely(!dst_pmd)) {
 501                         err = -ENOMEM;
 502                         break;
 503                 }
 504
 505                 dst_pmdval = pmd_read_atomic(dst_pmd);
 506                 /*
 507                  * If the dst_pmd is mapped as THP don't
 508                  * override it and just be strict.
 509                  */
 510                 if (unlikely(pmd_trans_huge(dst_pmdval))) {
 511                         err = -EEXIST;
 512                         break;
 513                 }
 514                 if (unlikely(pmd_none(dst_pmdval)) &&
 515                     unlikely(__pte_alloc(dst_mm, dst_pmd, dst_addr))) {
 516                         err = -ENOMEM;
 517                         break;
 518                 }
 519                 /* If an huge pmd materialized from under us fail */
 520                 if (unlikely(pmd_trans_huge(*dst_pmd))) {
 521                         err = -EFAULT;
 522                         break;
 523                 }
 524
 525                 BUG_ON(pmd_none(*dst_pmd));
 526                 BUG_ON(pmd_trans_huge(*dst_pmd));
 527
 528                 err = mfill_atomic_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
 529                                        src_addr, &page, zeropage);
 530                 cond_resched();
 531
 532                 if (unlikely(err == -ENOENT)) {
 533                         void *page_kaddr;
 534
 535                         up_read(&dst_mm->mmap_sem);
 536                         BUG_ON(!page);
 537
 538                         page_kaddr = kmap(page);
 539                         err = copy_from_user(page_kaddr,
 540                                              (const void __user *) src_addr,
 541                                              PAGE_SIZE);
 542                         kunmap(page);
 543                         if (unlikely(err)) {
 544                                 err = -EFAULT;
 545                                 goto out;
 546                         }
 547                         goto retry;
 548                 } else
 549                         BUG_ON(page);
 550
 551                 if (!err) {
 552                         dst_addr += PAGE_SIZE;
 553                         src_addr += PAGE_SIZE;
 554                         copied += PAGE_SIZE;
 555
 556                         if (fatal_signal_pending(current))
 557                                 err = -EINTR;
 558                 }
 559                 if (err)
 560                         break;
 561         }
 562
 563 out_unlock:
 564         up_read(&dst_mm->mmap_sem);
 565 out:
 566         if (page)
 567                 put_page(page);
 568         BUG_ON(copied < 0);
 569         BUG_ON(err > 0);
 570         BUG_ON(!copied && !err);
 571         return copied ? copied : err;
 572 }
 573
 574 ssize_t mcopy_atomic(struct mm_struct *dst_mm, unsigned long dst_start,
 575                      unsigned long src_start, unsigned long len)
 576 {
 577         return __mcopy_atomic(dst_mm, dst_start, src_start, len, false);
 578 }
 579
 580 ssize_t mfill_zeropage(struct mm_struct *dst_mm, unsigned long start,
 581                        unsigned long len)
 582 {
 583         return __mcopy_atomic(dst_mm, start, 0, len, true);
 584 }