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         pgoff_t offset, max_off;
  38         struct inode *inode;
  39
  40         if (!*pagep) {
  41                 ret = -ENOMEM;
  42                 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, dst_vma, dst_addr);
  43                 if (!page)
  44                         goto out;
  45
  46                 page_kaddr = kmap_atomic(page);
  47                 ret = copy_from_user(page_kaddr,
  48                                      (const void __user *) src_addr,
  49                                      PAGE_SIZE);
  50                 kunmap_atomic(page_kaddr);
  51
  52                 /* fallback to copy_from_user outside mmap_sem */
  53                 if (unlikely(ret)) {
  54                         ret = -ENOENT;
  55                         *pagep = page;
  56                         /* don't free the page */
  57                         goto out;
  58                 }
  59         } else {
  60                 page = *pagep;
  61                 *pagep = NULL;
  62         }
  63
  64         /*
  65          * The memory barrier inside __SetPageUptodate makes sure that
  66          * preceeding stores to the page contents become visible before
  67          * the set_pte_at() write.
  68          */
  69         __SetPageUptodate(page);
  70
  71         ret = -ENOMEM;
  72         if (mem_cgroup_try_charge(page, dst_mm, GFP_KERNEL, &memcg, false))
  73                 goto out_release;
  74
  75         _dst_pte = mk_pte(page, dst_vma->vm_page_prot);
  76         if (dst_vma->vm_flags & VM_WRITE)
  77                 _dst_pte = pte_mkwrite(pte_mkdirty(_dst_pte));
  78
  79         dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
  80         if (dst_vma->vm_file) {
  81                 /* the shmem MAP_PRIVATE case requires checking the i_size */
  82                 inode = dst_vma->vm_file->f_inode;
  83                 offset = linear_page_index(dst_vma, dst_addr);
  84                 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
  85                 ret = -EFAULT;
  86                 if (unlikely(offset >= max_off))
  87                         goto out_release_uncharge_unlock;
  88         }
  89         ret = -EEXIST;
  90         if (!pte_none(*dst_pte))
  91                 goto out_release_uncharge_unlock;
  92
  93         inc_mm_counter(dst_mm, MM_ANONPAGES);
  94         page_add_new_anon_rmap(page, dst_vma, dst_addr, false);
  95         mem_cgroup_commit_charge(page, memcg, false, false);
  96         lru_cache_add_active_or_unevictable(page, dst_vma);
  97
  98         set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
  99
 100         /* No need to invalidate - it was non-present before */
 101         update_mmu_cache(dst_vma, dst_addr, dst_pte);
 102
 103         pte_unmap_unlock(dst_pte, ptl);
 104         ret = 0;
 105 out:
 106         return ret;
 107 out_release_uncharge_unlock:
 108         pte_unmap_unlock(dst_pte, ptl);
 109         mem_cgroup_cancel_charge(page, memcg, false);
 110 out_release:
 111         put_page(page);
 112         goto out;
 113 }
 114
 115 static int mfill_zeropage_pte(struct mm_struct *dst_mm,
 116                               pmd_t *dst_pmd,
 117                               struct vm_area_struct *dst_vma,
 118                               unsigned long dst_addr)
 119 {
 120         pte_t _dst_pte, *dst_pte;
 121         spinlock_t *ptl;
 122         int ret;
 123         pgoff_t offset, max_off;
 124         struct inode *inode;
 125
 126         _dst_pte = pte_mkspecial(pfn_pte(my_zero_pfn(dst_addr),
 127                                          dst_vma->vm_page_prot));
 128         dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
 129         if (dst_vma->vm_file) {
 130                 /* the shmem MAP_PRIVATE case requires checking the i_size */
 131                 inode = dst_vma->vm_file->f_inode;
 132                 offset = linear_page_index(dst_vma, dst_addr);
 133                 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
 134                 ret = -EFAULT;
 135                 if (unlikely(offset >= max_off))
 136                         goto out_unlock;
 137         }
 138         ret = -EEXIST;
 139         if (!pte_none(*dst_pte))
 140                 goto out_unlock;
 141         set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
 142         /* No need to invalidate - it was non-present before */
 143         update_mmu_cache(dst_vma, dst_addr, dst_pte);
 144         ret = 0;
 145 out_unlock:
 146         pte_unmap_unlock(dst_pte, ptl);
 147         return ret;
 148 }
 149
 150 static pmd_t *mm_alloc_pmd(struct mm_struct *mm, unsigned long address)
 151 {
 152         pgd_t *pgd;
 153         p4d_t *p4d;
 154         pud_t *pud;
 155
 156         pgd = pgd_offset(mm, address);
 157         p4d = p4d_alloc(mm, pgd, address);
 158         if (!p4d)
 159                 return NULL;
 160         pud = pud_alloc(mm, p4d, address);
 161         if (!pud)
 162                 return NULL;
 163         /*
 164          * Note that we didn't run this because the pmd was
 165          * missing, the *pmd may be already established and in
 166          * turn it may also be a trans_huge_pmd.
 167          */
 168         return pmd_alloc(mm, pud, address);
 169 }
 170
 171 #ifdef CONFIG_HUGETLB_PAGE
 172 /*
 173  * __mcopy_atomic processing for HUGETLB vmas.  Note that this routine is
 174  * called with mmap_sem held, it will release mmap_sem before returning.
 175  */
 176 static __always_inline ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
 177                                               struct vm_area_struct *dst_vma,
 178                                               unsigned long dst_start,
 179                                               unsigned long src_start,
 180                                               unsigned long len,
 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                         dst_vma = NULL;
 314                         goto retry;
 315                 } else
 316                         BUG_ON(page);
 317
 318                 if (!err) {
 319                         dst_addr += vma_hpagesize;
 320                         src_addr += vma_hpagesize;
 321                         copied += vma_hpagesize;
 322
 323                         if (fatal_signal_pending(current))
 324                                 err = -EINTR;
 325                 }
 326                 if (err)
 327                         break;
 328         }
 329
 330 out_unlock:
 331         up_read(&dst_mm->mmap_sem);
 332 out:
 333         if (page) {
 334                 /*
 335                  * We encountered an error and are about to free a newly
 336                  * allocated huge page.
 337                  *
 338                  * Reservation handling is very subtle, and is different for
 339                  * private and shared mappings.  See the routine
 340                  * restore_reserve_on_error for details.  Unfortunately, we
 341                  * can not call restore_reserve_on_error now as it would
 342                  * require holding mmap_sem.
 343                  *
 344                  * If a reservation for the page existed in the reservation
 345                  * map of a private mapping, the map was modified to indicate
 346                  * the reservation was consumed when the page was allocated.
 347                  * We clear the PagePrivate flag now so that the global
 348                  * reserve count will not be incremented in free_huge_page.
 349                  * The reservation map will still indicate the reservation
 350                  * was consumed and possibly prevent later page allocation.
 351                  * This is better than leaking a global reservation.  If no
 352                  * reservation existed, it is still safe to clear PagePrivate
 353                  * as no adjustments to reservation counts were made during
 354                  * allocation.
 355                  *
 356                  * The reservation map for shared mappings indicates which
 357                  * pages have reservations.  When a huge page is allocated
 358                  * for an address with a reservation, no change is made to
 359                  * the reserve map.  In this case PagePrivate will be set
 360                  * to indicate that the global reservation count should be
 361                  * incremented when the page is freed.  This is the desired
 362                  * behavior.  However, when a huge page is allocated for an
 363                  * address without a reservation a reservation entry is added
 364                  * to the reservation map, and PagePrivate will not be set.
 365                  * When the page is freed, the global reserve count will NOT
 366                  * be incremented and it will appear as though we have leaked
 367                  * reserved page.  In this case, set PagePrivate so that the
 368                  * global reserve count will be incremented to match the
 369                  * reservation map entry which was created.
 370                  *
 371                  * Note that vm_alloc_shared is based on the flags of the vma
 372                  * for which the page was originally allocated.  dst_vma could
 373                  * be different or NULL on error.
 374                  */
 375                 if (vm_alloc_shared)
 376                         SetPagePrivate(page);
 377                 else
 378                         ClearPagePrivate(page);
 379                 put_page(page);
 380         }
 381         BUG_ON(copied < 0);
 382         BUG_ON(err > 0);
 383         BUG_ON(!copied && !err);
 384         return copied ? copied : err;
 385 }
 386 #else /* !CONFIG_HUGETLB_PAGE */
 387 /* fail at build time if gcc attempts to use this */
 388 extern ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
 389                                       struct vm_area_struct *dst_vma,
 390                                       unsigned long dst_start,
 391                                       unsigned long src_start,
 392                                       unsigned long len,
 393                                       bool zeropage);
 394 #endif /* CONFIG_HUGETLB_PAGE */
 395
 396 static __always_inline ssize_t mfill_atomic_pte(struct mm_struct *dst_mm,
 397                                                 pmd_t *dst_pmd,
 398                                                 struct vm_area_struct *dst_vma,
 399                                                 unsigned long dst_addr,
 400                                                 unsigned long src_addr,
 401                                                 struct page **page,
 402                                                 bool zeropage)
 403 {
 404         ssize_t err;
 405
 406         /*
 407          * The normal page fault path for a shmem will invoke the
 408          * fault, fill the hole in the file and COW it right away. The
 409          * result generates plain anonymous memory. So when we are
 410          * asked to fill an hole in a MAP_PRIVATE shmem mapping, we'll
 411          * generate anonymous memory directly without actually filling
 412          * the hole. For the MAP_PRIVATE case the robustness check
 413          * only happens in the pagetable (to verify it's still none)
 414          * and not in the radix tree.
 415          */
 416         if (!(dst_vma->vm_flags & VM_SHARED)) {
 417                 if (!zeropage)
 418                         err = mcopy_atomic_pte(dst_mm, dst_pmd, dst_vma,
 419                                                dst_addr, src_addr, page);
 420                 else
 421                         err = mfill_zeropage_pte(dst_mm, dst_pmd,
 422                                                  dst_vma, dst_addr);
 423         } else {
 424                 if (!zeropage)
 425                         err = shmem_mcopy_atomic_pte(dst_mm, dst_pmd,
 426                                                      dst_vma, dst_addr,
 427                                                      src_addr, page);
 428                 else
 429                         err = shmem_mfill_zeropage_pte(dst_mm, dst_pmd,
 430                                                        dst_vma, dst_addr);
 431         }
 432
 433         return err;
 434 }
 435
 436 static __always_inline ssize_t __mcopy_atomic(struct mm_struct *dst_mm,
 437                                               unsigned long dst_start,
 438                                               unsigned long src_start,
 439                                               unsigned long len,
 440                                               bool zeropage)
 441 {
 442         struct vm_area_struct *dst_vma;
 443         ssize_t err;
 444         pmd_t *dst_pmd;
 445         unsigned long src_addr, dst_addr;
 446         long copied;
 447         struct page *page;
 448
 449         /*
 450          * Sanitize the command parameters:
 451          */
 452         BUG_ON(dst_start & ~PAGE_MASK);
 453         BUG_ON(len & ~PAGE_MASK);
 454
 455         /* Does the address range wrap, or is the span zero-sized? */
 456         BUG_ON(src_start + len <= src_start);
 457         BUG_ON(dst_start + len <= dst_start);
 458
 459         src_addr = src_start;
 460         dst_addr = dst_start;
 461         copied = 0;
 462         page = NULL;
 463 retry:
 464         down_read(&dst_mm->mmap_sem);
 465
 466         /*
 467          * Make sure the vma is not shared, that the dst range is
 468          * both valid and fully within a single existing vma.
 469          */
 470         err = -ENOENT;
 471         dst_vma = find_vma(dst_mm, dst_start);
 472         if (!dst_vma)
 473                 goto out_unlock;
 474         /*
 475          * Check the vma is registered in uffd, this is required to
 476          * enforce the VM_MAYWRITE check done at uffd registration
 477          * time.
 478          */
 479         if (!dst_vma->vm_userfaultfd_ctx.ctx)
 480                 goto out_unlock;
 481
 482         if (dst_start < dst_vma->vm_start ||
 483             dst_start + len > dst_vma->vm_end)
 484                 goto out_unlock;
 485
 486         err = -EINVAL;
 487         /*
 488          * shmem_zero_setup is invoked in mmap for MAP_ANONYMOUS|MAP_SHARED but
 489          * it will overwrite vm_ops, so vma_is_anonymous must return false.
 490          */
 491         if (WARN_ON_ONCE(vma_is_anonymous(dst_vma) &&
 492             dst_vma->vm_flags & VM_SHARED))
 493                 goto out_unlock;
 494
 495         /*
 496          * If this is a HUGETLB vma, pass off to appropriate routine
 497          */
 498         if (is_vm_hugetlb_page(dst_vma))
 499                 return  __mcopy_atomic_hugetlb(dst_mm, dst_vma, dst_start,
 500                                                 src_start, len, zeropage);
 501
 502         if (!vma_is_anonymous(dst_vma) && !vma_is_shmem(dst_vma))
 503                 goto out_unlock;
 504
 505         /*
 506          * Ensure the dst_vma has a anon_vma or this page
 507          * would get a NULL anon_vma when moved in the
 508          * dst_vma.
 509          */
 510         err = -ENOMEM;
 511         if (!(dst_vma->vm_flags & VM_SHARED) &&
 512             unlikely(anon_vma_prepare(dst_vma)))
 513                 goto out_unlock;
 514
 515         while (src_addr < src_start + len) {
 516                 pmd_t dst_pmdval;
 517
 518                 BUG_ON(dst_addr >= dst_start + len);
 519
 520                 dst_pmd = mm_alloc_pmd(dst_mm, dst_addr);
 521                 if (unlikely(!dst_pmd)) {
 522                         err = -ENOMEM;
 523                         break;
 524                 }
 525
 526                 dst_pmdval = pmd_read_atomic(dst_pmd);
 527                 /*
 528                  * If the dst_pmd is mapped as THP don't
 529                  * override it and just be strict.
 530                  */
 531                 if (unlikely(pmd_trans_huge(dst_pmdval))) {
 532                         err = -EEXIST;
 533                         break;
 534                 }
 535                 if (unlikely(pmd_none(dst_pmdval)) &&
 536                     unlikely(__pte_alloc(dst_mm, dst_pmd, dst_addr))) {
 537                         err = -ENOMEM;
 538                         break;
 539                 }
 540                 /* If an huge pmd materialized from under us fail */
 541                 if (unlikely(pmd_trans_huge(*dst_pmd))) {
 542                         err = -EFAULT;
 543                         break;
 544                 }
 545
 546                 BUG_ON(pmd_none(*dst_pmd));
 547                 BUG_ON(pmd_trans_huge(*dst_pmd));
 548
 549                 err = mfill_atomic_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
 550                                        src_addr, &page, zeropage);
 551                 cond_resched();
 552
 553                 if (unlikely(err == -ENOENT)) {
 554                         void *page_kaddr;
 555
 556                         up_read(&dst_mm->mmap_sem);
 557                         BUG_ON(!page);
 558
 559                         page_kaddr = kmap(page);
 560                         err = copy_from_user(page_kaddr,
 561                                              (const void __user *) src_addr,
 562                                              PAGE_SIZE);
 563                         kunmap(page);
 564                         if (unlikely(err)) {
 565                                 err = -EFAULT;
 566                                 goto out;
 567                         }
 568                         goto retry;
 569                 } else
 570                         BUG_ON(page);
 571
 572                 if (!err) {
 573                         dst_addr += PAGE_SIZE;
 574                         src_addr += PAGE_SIZE;
 575                         copied += PAGE_SIZE;
 576
 577                         if (fatal_signal_pending(current))
 578                                 err = -EINTR;
 579                 }
 580                 if (err)
 581                         break;
 582         }
 583
 584 out_unlock:
 585         up_read(&dst_mm->mmap_sem);
 586 out:
 587         if (page)
 588                 put_page(page);
 589         BUG_ON(copied < 0);
 590         BUG_ON(err > 0);
 591         BUG_ON(!copied && !err);
 592         return copied ? copied : err;
 593 }
 594
 595 ssize_t mcopy_atomic(struct mm_struct *dst_mm, unsigned long dst_start,
 596                      unsigned long src_start, unsigned long len)
 597 {
 598         return __mcopy_atomic(dst_mm, dst_start, src_start, len, false);
 599 }
 600
 601 ssize_t mfill_zeropage(struct mm_struct *dst_mm, unsigned long start,
 602                        unsigned long len)
 603 {
 604         return __mcopy_atomic(dst_mm, start, 0, len, true);
 605 }