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                  * Check the vma is registered in uffd, this is
 209                  * required to enforce the VM_MAYWRITE check done at
 210                  * uffd registration time.
 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                                               bool *mmap_changing)
 410 {
 411         struct vm_area_struct *dst_vma;
 412         ssize_t err;
 413         pmd_t *dst_pmd;
 414         unsigned long src_addr, dst_addr;
 415         long copied;
 416         struct page *page;
 417
 418         /*
 419          * Sanitize the command parameters:
 420          */
 421         BUG_ON(dst_start & ~PAGE_MASK);
 422         BUG_ON(len & ~PAGE_MASK);
 423
 424         /* Does the address range wrap, or is the span zero-sized? */
 425         BUG_ON(src_start + len <= src_start);
 426         BUG_ON(dst_start + len <= dst_start);
 427
 428         src_addr = src_start;
 429         dst_addr = dst_start;
 430         copied = 0;
 431         page = NULL;
 432 retry:
 433         down_read(&dst_mm->mmap_sem);
 434
 435         /*
 436          * If memory mappings are changing because of non-cooperative
 437          * operation (e.g. mremap) running in parallel, bail out and
 438          * request the user to retry later
 439          */
 440         err = -EAGAIN;
 441         if (mmap_changing && READ_ONCE(*mmap_changing))
 442                 goto out_unlock;
 443
 444         /*
 445          * Make sure the vma is not shared, that the dst range is
 446          * both valid and fully within a single existing vma.
 447          */
 448         err = -ENOENT;
 449         dst_vma = find_vma(dst_mm, dst_start);
 450         if (!dst_vma)
 451                 goto out_unlock;
 452         /*
 453          * Check the vma is registered in uffd, this is required to
 454          * enforce the VM_MAYWRITE check done at uffd registration
 455          * time.
 456          */
 457         if (!dst_vma->vm_userfaultfd_ctx.ctx)
 458                 goto out_unlock;
 459
 460         if (dst_start < dst_vma->vm_start ||
 461             dst_start + len > dst_vma->vm_end)
 462                 goto out_unlock;
 463
 464         err = -EINVAL;
 465         /*
 466          * shmem_zero_setup is invoked in mmap for MAP_ANONYMOUS|MAP_SHARED but
 467          * it will overwrite vm_ops, so vma_is_anonymous must return false.
 468          */
 469         if (WARN_ON_ONCE(vma_is_anonymous(dst_vma) &&
 470             dst_vma->vm_flags & VM_SHARED))
 471                 goto out_unlock;
 472
 473         /*
 474          * If this is a HUGETLB vma, pass off to appropriate routine
 475          */
 476         if (is_vm_hugetlb_page(dst_vma))
 477                 return  __mcopy_atomic_hugetlb(dst_mm, dst_vma, dst_start,
 478                                                 src_start, len, zeropage);
 479
 480         if (!vma_is_anonymous(dst_vma) && !vma_is_shmem(dst_vma))
 481                 goto out_unlock;
 482
 483         /*
 484          * Ensure the dst_vma has a anon_vma or this page
 485          * would get a NULL anon_vma when moved in the
 486          * dst_vma.
 487          */
 488         err = -ENOMEM;
 489         if (vma_is_anonymous(dst_vma) && unlikely(anon_vma_prepare(dst_vma)))
 490                 goto out_unlock;
 491
 492         while (src_addr < src_start + len) {
 493                 pmd_t dst_pmdval;
 494
 495                 BUG_ON(dst_addr >= dst_start + len);
 496
 497                 dst_pmd = mm_alloc_pmd(dst_mm, dst_addr);
 498                 if (unlikely(!dst_pmd)) {
 499                         err = -ENOMEM;
 500                         break;
 501                 }
 502
 503                 dst_pmdval = pmd_read_atomic(dst_pmd);
 504                 /*
 505                  * If the dst_pmd is mapped as THP don't
 506                  * override it and just be strict.
 507                  */
 508                 if (unlikely(pmd_trans_huge(dst_pmdval))) {
 509                         err = -EEXIST;
 510                         break;
 511                 }
 512                 if (unlikely(pmd_none(dst_pmdval)) &&
 513                     unlikely(__pte_alloc(dst_mm, dst_pmd, dst_addr))) {
 514                         err = -ENOMEM;
 515                         break;
 516                 }
 517                 /* If an huge pmd materialized from under us fail */
 518                 if (unlikely(pmd_trans_huge(*dst_pmd))) {
 519                         err = -EFAULT;
 520                         break;
 521                 }
 522
 523                 BUG_ON(pmd_none(*dst_pmd));
 524                 BUG_ON(pmd_trans_huge(*dst_pmd));
 525
 526                 err = mfill_atomic_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
 527                                        src_addr, &page, zeropage);
 528                 cond_resched();
 529
 530                 if (unlikely(err == -EFAULT)) {
 531                         void *page_kaddr;
 532
 533                         up_read(&dst_mm->mmap_sem);
 534                         BUG_ON(!page);
 535
 536                         page_kaddr = kmap(page);
 537                         err = copy_from_user(page_kaddr,
 538                                              (const void __user *) src_addr,
 539                                              PAGE_SIZE);
 540                         kunmap(page);
 541                         if (unlikely(err)) {
 542                                 err = -EFAULT;
 543                                 goto out;
 544                         }
 545                         goto retry;
 546                 } else
 547                         BUG_ON(page);
 548
 549                 if (!err) {
 550                         dst_addr += PAGE_SIZE;
 551                         src_addr += PAGE_SIZE;
 552                         copied += PAGE_SIZE;
 553
 554                         if (fatal_signal_pending(current))
 555                                 err = -EINTR;
 556                 }
 557                 if (err)
 558                         break;
 559         }
 560
 561 out_unlock:
 562         up_read(&dst_mm->mmap_sem);
 563 out:
 564         if (page)
 565                 put_page(page);
 566         BUG_ON(copied < 0);
 567         BUG_ON(err > 0);
 568         BUG_ON(!copied && !err);
 569         return copied ? copied : err;
 570 }
 571
 572 ssize_t mcopy_atomic(struct mm_struct *dst_mm, unsigned long dst_start,
 573                      unsigned long src_start, unsigned long len,
 574                      bool *mmap_changing)
 575 {
 576         return __mcopy_atomic(dst_mm, dst_start, src_start, len, false,
 577                               mmap_changing);
 578 }
 579
 580 ssize_t mfill_zeropage(struct mm_struct *dst_mm, unsigned long start,
 581                        unsigned long len, bool *mmap_changing)
 582 {
 583         return __mcopy_atomic(dst_mm, start, 0, len, true, mmap_changing);
 584 }