mm/userfaultfd.c

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