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         pgoff_t offset, max_off;
  37         struct inode *inode;
  38
  39         if (!*pagep) {
  40                 ret = -ENOMEM;
  41                 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, dst_vma, dst_addr);
  42                 if (!page)
  43                         goto out;
  44
  45                 page_kaddr = kmap_atomic(page);
  46                 ret = copy_from_user(page_kaddr,
  47                                      (const void __user *) src_addr,
  48                                      PAGE_SIZE);
  49                 kunmap_atomic(page_kaddr);
  50
  51                 /* fallback to copy_from_user outside mmap_sem */
  52                 if (unlikely(ret)) {
  53                         ret = -ENOENT;
  54                         *pagep = page;
  55                         /* don't free the page */
  56                         goto out;
  57                 }
  58
  59                 flush_dcache_page(page);
  60         } else {
  61                 page = *pagep;
  62                 *pagep = NULL;
  63         }
  64
  65         /*
  66          * The memory barrier inside __SetPageUptodate makes sure that
  67          * preceeding stores to the page contents become visible before
  68          * the set_pte_at() write.
  69          */
  70         __SetPageUptodate(page);
  71
  72         ret = -ENOMEM;
  73         if (mem_cgroup_try_charge(page, dst_mm, GFP_KERNEL, &memcg, false))
  74                 goto out_release;
  75
  76         _dst_pte = mk_pte(page, dst_vma->vm_page_prot);
  77         if (dst_vma->vm_flags & VM_WRITE)
  78                 _dst_pte = pte_mkwrite(pte_mkdirty(_dst_pte));
  79
  80         dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
  81         if (dst_vma->vm_file) {
  82                 /* the shmem MAP_PRIVATE case requires checking the i_size */
  83                 inode = dst_vma->vm_file->f_inode;
  84                 offset = linear_page_index(dst_vma, dst_addr);
  85                 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
  86                 ret = -EFAULT;
  87                 if (unlikely(offset >= max_off))
  88                         goto out_release_uncharge_unlock;
  89         }
  90         ret = -EEXIST;
  91         if (!pte_none(*dst_pte))
  92                 goto out_release_uncharge_unlock;
  93
  94         inc_mm_counter(dst_mm, MM_ANONPAGES);
  95         page_add_new_anon_rmap(page, dst_vma, dst_addr, false);
  96         mem_cgroup_commit_charge(page, memcg, false, false);
  97         lru_cache_add_active_or_unevictable(page, dst_vma);
  98
  99         set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
 100
 101         /* No need to invalidate - it was non-present before */
 102         update_mmu_cache(dst_vma, dst_addr, dst_pte);
 103
 104         pte_unmap_unlock(dst_pte, ptl);
 105         ret = 0;
 106 out:
 107         return ret;
 108 out_release_uncharge_unlock:
 109         pte_unmap_unlock(dst_pte, ptl);
 110         mem_cgroup_cancel_charge(page, memcg, false);
 111 out_release:
 112         put_page(page);
 113         goto out;
 114 }
 115
 116 static int mfill_zeropage_pte(struct mm_struct *dst_mm,
 117                               pmd_t *dst_pmd,
 118                               struct vm_area_struct *dst_vma,
 119                               unsigned long dst_addr)
 120 {
 121         pte_t _dst_pte, *dst_pte;
 122         spinlock_t *ptl;
 123         int ret;
 124         pgoff_t offset, max_off;
 125         struct inode *inode;
 126
 127         _dst_pte = pte_mkspecial(pfn_pte(my_zero_pfn(dst_addr),
 128                                          dst_vma->vm_page_prot));
 129         dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
 130         if (dst_vma->vm_file) {
 131                 /* the shmem MAP_PRIVATE case requires checking the i_size */
 132                 inode = dst_vma->vm_file->f_inode;
 133                 offset = linear_page_index(dst_vma, dst_addr);
 134                 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
 135                 ret = -EFAULT;
 136                 if (unlikely(offset >= max_off))
 137                         goto out_unlock;
 138         }
 139         ret = -EEXIST;
 140         if (!pte_none(*dst_pte))
 141                 goto out_unlock;
 142         set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
 143         /* No need to invalidate - it was non-present before */
 144         update_mmu_cache(dst_vma, dst_addr, dst_pte);
 145         ret = 0;
 146 out_unlock:
 147         pte_unmap_unlock(dst_pte, ptl);
 148         return ret;
 149 }
 150
 151 static pmd_t *mm_alloc_pmd(struct mm_struct *mm, unsigned long address)
 152 {
 153         pgd_t *pgd;
 154         p4d_t *p4d;
 155         pud_t *pud;
 156
 157         pgd = pgd_offset(mm, address);
 158         p4d = p4d_alloc(mm, pgd, address);
 159         if (!p4d)
 160                 return NULL;
 161         pud = pud_alloc(mm, p4d, address);
 162         if (!pud)
 163                 return NULL;
 164         /*
 165          * Note that we didn't run this because the pmd was
 166          * missing, the *pmd may be already established and in
 167          * turn it may also be a trans_huge_pmd.
 168          */
 169         return pmd_alloc(mm, pud, address);
 170 }
 171
 172 #ifdef CONFIG_HUGETLB_PAGE
 173 /*
 174  * __mcopy_atomic processing for HUGETLB vmas.  Note that this routine is
 175  * called with mmap_sem held, it will release mmap_sem before returning.
 176  */
 177 static __always_inline ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
 178                                               struct vm_area_struct *dst_vma,
 179                                               unsigned long dst_start,
 180                                               unsigned long src_start,
 181                                               unsigned long len,
 182                                               bool *mmap_changing,
 183                                               bool zeropage)
 184 {
 185         int vm_alloc_shared = dst_vma->vm_flags & VM_SHARED;
 186         int vm_shared = dst_vma->vm_flags & VM_SHARED;
 187         ssize_t err;
 188         pte_t *dst_pte;
 189         unsigned long src_addr, dst_addr;
 190         long copied;
 191         struct page *page;
 192         struct hstate *h;
 193         unsigned long vma_hpagesize;
 194         pgoff_t idx;
 195         u32 hash;
 196         struct address_space *mapping;
 197
 198         /*
 199          * There is no default zero huge page for all huge page sizes as
 200          * supported by hugetlb.  A PMD_SIZE huge pages may exist as used
 201          * by THP.  Since we can not reliably insert a zero page, this
 202          * feature is not supported.
 203          */
 204         if (zeropage) {
 205                 up_read(&dst_mm->mmap_sem);
 206                 return -EINVAL;
 207         }
 208
 209         src_addr = src_start;
 210         dst_addr = dst_start;
 211         copied = 0;
 212         page = NULL;
 213         vma_hpagesize = vma_kernel_pagesize(dst_vma);
 214
 215         /*
 216          * Validate alignment based on huge page size
 217          */
 218         err = -EINVAL;
 219         if (dst_start & (vma_hpagesize - 1) || len & (vma_hpagesize - 1))
 220                 goto out_unlock;
 221
 222 retry:
 223         /*
 224          * On routine entry dst_vma is set.  If we had to drop mmap_sem and
 225          * retry, dst_vma will be set to NULL and we must lookup again.
 226          */
 227         if (!dst_vma) {
 228                 err = -ENOENT;
 229                 dst_vma = find_vma(dst_mm, dst_start);
 230                 if (!dst_vma || !is_vm_hugetlb_page(dst_vma))
 231                         goto out_unlock;
 232                 /*
 233                  * Check the vma is registered in uffd, this is
 234                  * required to enforce the VM_MAYWRITE check done at
 235                  * uffd registration time.
 236                  */
 237                 if (!dst_vma->vm_userfaultfd_ctx.ctx)
 238                         goto out_unlock;
 239
 240                 if (dst_start < dst_vma->vm_start ||
 241                     dst_start + len > dst_vma->vm_end)
 242                         goto out_unlock;
 243
 244                 err = -EINVAL;
 245                 if (vma_hpagesize != vma_kernel_pagesize(dst_vma))
 246                         goto out_unlock;
 247
 248                 vm_shared = dst_vma->vm_flags & VM_SHARED;
 249         }
 250
 251         if (WARN_ON(dst_addr & (vma_hpagesize - 1) ||
 252                     (len - copied) & (vma_hpagesize - 1)))
 253                 goto out_unlock;
 254
 255         /*
 256          * If not shared, ensure the dst_vma has a anon_vma.
 257          */
 258         err = -ENOMEM;
 259         if (!vm_shared) {
 260                 if (unlikely(anon_vma_prepare(dst_vma)))
 261                         goto out_unlock;
 262         }
 263
 264         h = hstate_vma(dst_vma);
 265
 266         while (src_addr < src_start + len) {
 267                 pte_t dst_pteval;
 268
 269                 BUG_ON(dst_addr >= dst_start + len);
 270                 VM_BUG_ON(dst_addr & ~huge_page_mask(h));
 271
 272                 /*
 273                  * Serialize via hugetlb_fault_mutex
 274                  */
 275                 idx = linear_page_index(dst_vma, dst_addr);
 276                 mapping = dst_vma->vm_file->f_mapping;
 277                 hash = hugetlb_fault_mutex_hash(h, mapping, idx);
 278                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 279
 280                 err = -ENOMEM;
 281                 dst_pte = huge_pte_alloc(dst_mm, dst_addr, huge_page_size(h));
 282                 if (!dst_pte) {
 283                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 284                         goto out_unlock;
 285                 }
 286
 287                 err = -EEXIST;
 288                 dst_pteval = huge_ptep_get(dst_pte);
 289                 if (!huge_pte_none(dst_pteval)) {
 290                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 291                         goto out_unlock;
 292                 }
 293
 294                 err = hugetlb_mcopy_atomic_pte(dst_mm, dst_pte, dst_vma,
 295                                                 dst_addr, src_addr, &page);
 296
 297                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 298                 vm_alloc_shared = vm_shared;
 299
 300                 cond_resched();
 301
 302                 if (unlikely(err == -ENOENT)) {
 303                         up_read(&dst_mm->mmap_sem);
 304                         BUG_ON(!page);
 305
 306                         err = copy_huge_page_from_user(page,
 307                                                 (const void __user *)src_addr,
 308                                                 pages_per_huge_page(h), true);
 309                         if (unlikely(err)) {
 310                                 err = -EFAULT;
 311                                 goto out;
 312                         }
 313                         down_read(&dst_mm->mmap_sem);
 314                         /*
 315                          * If memory mappings are changing because of non-cooperative
 316                          * operation (e.g. mremap) running in parallel, bail out and
 317                          * request the user to retry later
 318                          */
 319                         if (mmap_changing && READ_ONCE(*mmap_changing)) {
 320                                 err = -EAGAIN;
 321                                 break;
 322                         }
 323
 324                         dst_vma = NULL;
 325                         goto retry;
 326                 } else
 327                         BUG_ON(page);
 328
 329                 if (!err) {
 330                         dst_addr += vma_hpagesize;
 331                         src_addr += vma_hpagesize;
 332                         copied += vma_hpagesize;
 333
 334                         if (fatal_signal_pending(current))
 335                                 err = -EINTR;
 336                 }
 337                 if (err)
 338                         break;
 339         }
 340
 341 out_unlock:
 342         up_read(&dst_mm->mmap_sem);
 343 out:
 344         if (page) {
 345                 /*
 346                  * We encountered an error and are about to free a newly
 347                  * allocated huge page.
 348                  *
 349                  * Reservation handling is very subtle, and is different for
 350                  * private and shared mappings.  See the routine
 351                  * restore_reserve_on_error for details.  Unfortunately, we
 352                  * can not call restore_reserve_on_error now as it would
 353                  * require holding mmap_sem.
 354                  *
 355                  * If a reservation for the page existed in the reservation
 356                  * map of a private mapping, the map was modified to indicate
 357                  * the reservation was consumed when the page was allocated.
 358                  * We clear the PagePrivate flag now so that the global
 359                  * reserve count will not be incremented in free_huge_page.
 360                  * The reservation map will still indicate the reservation
 361                  * was consumed and possibly prevent later page allocation.
 362                  * This is better than leaking a global reservation.  If no
 363                  * reservation existed, it is still safe to clear PagePrivate
 364                  * as no adjustments to reservation counts were made during
 365                  * allocation.
 366                  *
 367                  * The reservation map for shared mappings indicates which
 368                  * pages have reservations.  When a huge page is allocated
 369                  * for an address with a reservation, no change is made to
 370                  * the reserve map.  In this case PagePrivate will be set
 371                  * to indicate that the global reservation count should be
 372                  * incremented when the page is freed.  This is the desired
 373                  * behavior.  However, when a huge page is allocated for an
 374                  * address without a reservation a reservation entry is added
 375                  * to the reservation map, and PagePrivate will not be set.
 376                  * When the page is freed, the global reserve count will NOT
 377                  * be incremented and it will appear as though we have leaked
 378                  * reserved page.  In this case, set PagePrivate so that the
 379                  * global reserve count will be incremented to match the
 380                  * reservation map entry which was created.
 381                  *
 382                  * Note that vm_alloc_shared is based on the flags of the vma
 383                  * for which the page was originally allocated.  dst_vma could
 384                  * be different or NULL on error.
 385                  */
 386                 if (vm_alloc_shared)
 387                         SetPagePrivate(page);
 388                 else
 389                         ClearPagePrivate(page);
 390                 put_page(page);
 391         }
 392         BUG_ON(copied < 0);
 393         BUG_ON(err > 0);
 394         BUG_ON(!copied && !err);
 395         return copied ? copied : err;
 396 }
 397 #else /* !CONFIG_HUGETLB_PAGE */
 398 /* fail at build time if gcc attempts to use this */
 399 extern ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
 400                                       struct vm_area_struct *dst_vma,
 401                                       unsigned long dst_start,
 402                                       unsigned long src_start,
 403                                       unsigned long len,
 404                                       bool *mmap_changing,
 405                                       bool zeropage);
 406 #endif /* CONFIG_HUGETLB_PAGE */
 407
 408 static __always_inline ssize_t mfill_atomic_pte(struct mm_struct *dst_mm,
 409                                                 pmd_t *dst_pmd,
 410                                                 struct vm_area_struct *dst_vma,
 411                                                 unsigned long dst_addr,
 412                                                 unsigned long src_addr,
 413                                                 struct page **page,
 414                                                 bool zeropage)
 415 {
 416         ssize_t err;
 417
 418         /*
 419          * The normal page fault path for a shmem will invoke the
 420          * fault, fill the hole in the file and COW it right away. The
 421          * result generates plain anonymous memory. So when we are
 422          * asked to fill an hole in a MAP_PRIVATE shmem mapping, we'll
 423          * generate anonymous memory directly without actually filling
 424          * the hole. For the MAP_PRIVATE case the robustness check
 425          * only happens in the pagetable (to verify it's still none)
 426          * and not in the radix tree.
 427          */
 428         if (!(dst_vma->vm_flags & VM_SHARED)) {
 429                 if (!zeropage)
 430                         err = mcopy_atomic_pte(dst_mm, dst_pmd, dst_vma,
 431                                                dst_addr, src_addr, page);
 432                 else
 433                         err = mfill_zeropage_pte(dst_mm, dst_pmd,
 434                                                  dst_vma, dst_addr);
 435         } else {
 436                 if (!zeropage)
 437                         err = shmem_mcopy_atomic_pte(dst_mm, dst_pmd,
 438                                                      dst_vma, dst_addr,
 439                                                      src_addr, page);
 440                 else
 441                         err = shmem_mfill_zeropage_pte(dst_mm, dst_pmd,
 442                                                        dst_vma, dst_addr);
 443         }
 444
 445         return err;
 446 }
 447
 448 static __always_inline ssize_t __mcopy_atomic(struct mm_struct *dst_mm,
 449                                               unsigned long dst_start,
 450                                               unsigned long src_start,
 451                                               unsigned long len,
 452                                               bool zeropage,
 453                                               bool *mmap_changing)
 454 {
 455         struct vm_area_struct *dst_vma;
 456         ssize_t err;
 457         pmd_t *dst_pmd;
 458         unsigned long src_addr, dst_addr;
 459         long copied;
 460         struct page *page;
 461
 462         /*
 463          * Sanitize the command parameters:
 464          */
 465         BUG_ON(dst_start & ~PAGE_MASK);
 466         BUG_ON(len & ~PAGE_MASK);
 467
 468         /* Does the address range wrap, or is the span zero-sized? */
 469         BUG_ON(src_start + len <= src_start);
 470         BUG_ON(dst_start + len <= dst_start);
 471
 472         src_addr = src_start;
 473         dst_addr = dst_start;
 474         copied = 0;
 475         page = NULL;
 476 retry:
 477         down_read(&dst_mm->mmap_sem);
 478
 479         /*
 480          * If memory mappings are changing because of non-cooperative
 481          * operation (e.g. mremap) running in parallel, bail out and
 482          * request the user to retry later
 483          */
 484         err = -EAGAIN;
 485         if (mmap_changing && READ_ONCE(*mmap_changing))
 486                 goto out_unlock;
 487
 488         /*
 489          * Make sure the vma is not shared, that the dst range is
 490          * both valid and fully within a single existing vma.
 491          */
 492         err = -ENOENT;
 493         dst_vma = find_vma(dst_mm, dst_start);
 494         if (!dst_vma)
 495                 goto out_unlock;
 496         /*
 497          * Check the vma is registered in uffd, this is required to
 498          * enforce the VM_MAYWRITE check done at uffd registration
 499          * time.
 500          */
 501         if (!dst_vma->vm_userfaultfd_ctx.ctx)
 502                 goto out_unlock;
 503
 504         if (dst_start < dst_vma->vm_start ||
 505             dst_start + len > dst_vma->vm_end)
 506                 goto out_unlock;
 507
 508         err = -EINVAL;
 509         /*
 510          * shmem_zero_setup is invoked in mmap for MAP_ANONYMOUS|MAP_SHARED but
 511          * it will overwrite vm_ops, so vma_is_anonymous must return false.
 512          */
 513         if (WARN_ON_ONCE(vma_is_anonymous(dst_vma) &&
 514             dst_vma->vm_flags & VM_SHARED))
 515                 goto out_unlock;
 516
 517         /*
 518          * If this is a HUGETLB vma, pass off to appropriate routine
 519          */
 520         if (is_vm_hugetlb_page(dst_vma))
 521                 return  __mcopy_atomic_hugetlb(dst_mm, dst_vma, dst_start,
 522                                                src_start, len, mmap_changing,
 523                                                zeropage);
 524
 525         if (!vma_is_anonymous(dst_vma) && !vma_is_shmem(dst_vma))
 526                 goto out_unlock;
 527
 528         /*
 529          * Ensure the dst_vma has a anon_vma or this page
 530          * would get a NULL anon_vma when moved in the
 531          * dst_vma.
 532          */
 533         err = -ENOMEM;
 534         if (!(dst_vma->vm_flags & VM_SHARED) &&
 535             unlikely(anon_vma_prepare(dst_vma)))
 536                 goto out_unlock;
 537
 538         while (src_addr < src_start + len) {
 539                 pmd_t dst_pmdval;
 540
 541                 BUG_ON(dst_addr >= dst_start + len);
 542
 543                 dst_pmd = mm_alloc_pmd(dst_mm, dst_addr);
 544                 if (unlikely(!dst_pmd)) {
 545                         err = -ENOMEM;
 546                         break;
 547                 }
 548
 549                 dst_pmdval = pmd_read_atomic(dst_pmd);
 550                 /*
 551                  * If the dst_pmd is mapped as THP don't
 552                  * override it and just be strict.
 553                  */
 554                 if (unlikely(pmd_trans_huge(dst_pmdval))) {
 555                         err = -EEXIST;
 556                         break;
 557                 }
 558                 if (unlikely(pmd_none(dst_pmdval)) &&
 559                     unlikely(__pte_alloc(dst_mm, dst_pmd, dst_addr))) {
 560                         err = -ENOMEM;
 561                         break;
 562                 }
 563                 /* If an huge pmd materialized from under us fail */
 564                 if (unlikely(pmd_trans_huge(*dst_pmd))) {
 565                         err = -EFAULT;
 566                         break;
 567                 }
 568
 569                 BUG_ON(pmd_none(*dst_pmd));
 570                 BUG_ON(pmd_trans_huge(*dst_pmd));
 571
 572                 err = mfill_atomic_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
 573                                        src_addr, &page, zeropage);
 574                 cond_resched();
 575
 576                 if (unlikely(err == -ENOENT)) {
 577                         void *page_kaddr;
 578
 579                         up_read(&dst_mm->mmap_sem);
 580                         BUG_ON(!page);
 581
 582                         page_kaddr = kmap(page);
 583                         err = copy_from_user(page_kaddr,
 584                                              (const void __user *) src_addr,
 585                                              PAGE_SIZE);
 586                         kunmap(page);
 587                         if (unlikely(err)) {
 588                                 err = -EFAULT;
 589                                 goto out;
 590                         }
 591                         flush_dcache_page(page);
 592                         goto retry;
 593                 } else
 594                         BUG_ON(page);
 595
 596                 if (!err) {
 597                         dst_addr += PAGE_SIZE;
 598                         src_addr += PAGE_SIZE;
 599                         copied += PAGE_SIZE;
 600
 601                         if (fatal_signal_pending(current))
 602                                 err = -EINTR;
 603                 }
 604                 if (err)
 605                         break;
 606         }
 607
 608 out_unlock:
 609         up_read(&dst_mm->mmap_sem);
 610 out:
 611         if (page)
 612                 put_page(page);
 613         BUG_ON(copied < 0);
 614         BUG_ON(err > 0);
 615         BUG_ON(!copied && !err);
 616         return copied ? copied : err;
 617 }
 618
 619 ssize_t mcopy_atomic(struct mm_struct *dst_mm, unsigned long dst_start,
 620                      unsigned long src_start, unsigned long len,
 621                      bool *mmap_changing)
 622 {
 623         return __mcopy_atomic(dst_mm, dst_start, src_start, len, false,
 624                               mmap_changing);
 625 }
 626
 627 ssize_t mfill_zeropage(struct mm_struct *dst_mm, unsigned long start,
 628                        unsigned long len, bool *mmap_changing)
 629 {
 630         return __mcopy_atomic(dst_mm, start, 0, len, true, mmap_changing);
 631 }