mm/userfaultfd.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  *  mm/userfaultfd.c
   4  *
   5  *  Copyright (C) 2015  Red Hat, Inc.
   6  */
   7
   8 #include <linux/mm.h>
   9 #include <linux/sched/signal.h>
  10 #include <linux/pagemap.h>
  11 #include <linux/rmap.h>
  12 #include <linux/swap.h>
  13 #include <linux/swapops.h>
  14 #include <linux/userfaultfd_k.h>
  15 #include <linux/mmu_notifier.h>
  16 #include <linux/hugetlb.h>
  17 #include <linux/shmem_fs.h>
  18 #include <asm/tlbflush.h>
  19 #include "internal.h"
  20
  21 static __always_inline
  22 struct vm_area_struct *find_dst_vma(struct mm_struct *dst_mm,
  23                                     unsigned long dst_start,
  24                                     unsigned long len)
  25 {
  26         /*
  27          * Make sure that the dst range is both valid and fully within a
  28          * single existing vma.
  29          */
  30         struct vm_area_struct *dst_vma;
  31
  32         dst_vma = find_vma(dst_mm, dst_start);
  33         if (!dst_vma)
  34                 return NULL;
  35
  36         if (dst_start < dst_vma->vm_start ||
  37             dst_start + len > dst_vma->vm_end)
  38                 return NULL;
  39
  40         /*
  41          * Check the vma is registered in uffd, this is required to
  42          * enforce the VM_MAYWRITE check done at uffd registration
  43          * time.
  44          */
  45         if (!dst_vma->vm_userfaultfd_ctx.ctx)
  46                 return NULL;
  47
  48         return dst_vma;
  49 }
  50
  51 static int mcopy_atomic_pte(struct mm_struct *dst_mm,
  52                             pmd_t *dst_pmd,
  53                             struct vm_area_struct *dst_vma,
  54                             unsigned long dst_addr,
  55                             unsigned long src_addr,
  56                             struct page **pagep,
  57                             bool wp_copy)
  58 {
  59         struct mem_cgroup *memcg;
  60         pte_t _dst_pte, *dst_pte;
  61         spinlock_t *ptl;
  62         void *page_kaddr;
  63         int ret;
  64         struct page *page;
  65         pgoff_t offset, max_off;
  66         struct inode *inode;
  67
  68         if (!*pagep) {
  69                 ret = -ENOMEM;
  70                 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, dst_vma, dst_addr);
  71                 if (!page)
  72                         goto out;
  73
  74                 page_kaddr = kmap_atomic(page);
  75                 ret = copy_from_user(page_kaddr,
  76                                      (const void __user *) src_addr,
  77                                      PAGE_SIZE);
  78                 kunmap_atomic(page_kaddr);
  79
  80                 /* fallback to copy_from_user outside mmap_sem */
  81                 if (unlikely(ret)) {
  82                         ret = -ENOENT;
  83                         *pagep = page;
  84                         /* don't free the page */
  85                         goto out;
  86                 }
  87         } else {
  88                 page = *pagep;
  89                 *pagep = NULL;
  90         }
  91
  92         /*
  93          * The memory barrier inside __SetPageUptodate makes sure that
  94          * preceding stores to the page contents become visible before
  95          * the set_pte_at() write.
  96          */
  97         __SetPageUptodate(page);
  98
  99         ret = -ENOMEM;
 100         if (mem_cgroup_try_charge(page, dst_mm, GFP_KERNEL, &memcg, false))
 101                 goto out_release;
 102
 103         _dst_pte = pte_mkdirty(mk_pte(page, dst_vma->vm_page_prot));
 104         if ((dst_vma->vm_flags & VM_WRITE) && !wp_copy)
 105                 _dst_pte = pte_mkwrite(_dst_pte);
 106
 107         dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
 108         if (dst_vma->vm_file) {
 109                 /* the shmem MAP_PRIVATE case requires checking the i_size */
 110                 inode = dst_vma->vm_file->f_inode;
 111                 offset = linear_page_index(dst_vma, dst_addr);
 112                 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
 113                 ret = -EFAULT;
 114                 if (unlikely(offset >= max_off))
 115                         goto out_release_uncharge_unlock;
 116         }
 117         ret = -EEXIST;
 118         if (!pte_none(*dst_pte))
 119                 goto out_release_uncharge_unlock;
 120
 121         inc_mm_counter(dst_mm, MM_ANONPAGES);
 122         page_add_new_anon_rmap(page, dst_vma, dst_addr, false);
 123         mem_cgroup_commit_charge(page, memcg, false, false);
 124         lru_cache_add_active_or_unevictable(page, dst_vma);
 125
 126         set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
 127
 128         /* No need to invalidate - it was non-present before */
 129         update_mmu_cache(dst_vma, dst_addr, dst_pte);
 130
 131         pte_unmap_unlock(dst_pte, ptl);
 132         ret = 0;
 133 out:
 134         return ret;
 135 out_release_uncharge_unlock:
 136         pte_unmap_unlock(dst_pte, ptl);
 137         mem_cgroup_cancel_charge(page, memcg, false);
 138 out_release:
 139         put_page(page);
 140         goto out;
 141 }
 142
 143 static int mfill_zeropage_pte(struct mm_struct *dst_mm,
 144                               pmd_t *dst_pmd,
 145                               struct vm_area_struct *dst_vma,
 146                               unsigned long dst_addr)
 147 {
 148         pte_t _dst_pte, *dst_pte;
 149         spinlock_t *ptl;
 150         int ret;
 151         pgoff_t offset, max_off;
 152         struct inode *inode;
 153
 154         _dst_pte = pte_mkspecial(pfn_pte(my_zero_pfn(dst_addr),
 155                                          dst_vma->vm_page_prot));
 156         dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
 157         if (dst_vma->vm_file) {
 158                 /* the shmem MAP_PRIVATE case requires checking the i_size */
 159                 inode = dst_vma->vm_file->f_inode;
 160                 offset = linear_page_index(dst_vma, dst_addr);
 161                 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
 162                 ret = -EFAULT;
 163                 if (unlikely(offset >= max_off))
 164                         goto out_unlock;
 165         }
 166         ret = -EEXIST;
 167         if (!pte_none(*dst_pte))
 168                 goto out_unlock;
 169         set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
 170         /* No need to invalidate - it was non-present before */
 171         update_mmu_cache(dst_vma, dst_addr, dst_pte);
 172         ret = 0;
 173 out_unlock:
 174         pte_unmap_unlock(dst_pte, ptl);
 175         return ret;
 176 }
 177
 178 static pmd_t *mm_alloc_pmd(struct mm_struct *mm, unsigned long address)
 179 {
 180         pgd_t *pgd;
 181         p4d_t *p4d;
 182         pud_t *pud;
 183
 184         pgd = pgd_offset(mm, address);
 185         p4d = p4d_alloc(mm, pgd, address);
 186         if (!p4d)
 187                 return NULL;
 188         pud = pud_alloc(mm, p4d, address);
 189         if (!pud)
 190                 return NULL;
 191         /*
 192          * Note that we didn't run this because the pmd was
 193          * missing, the *pmd may be already established and in
 194          * turn it may also be a trans_huge_pmd.
 195          */
 196         return pmd_alloc(mm, pud, address);
 197 }
 198
 199 #ifdef CONFIG_HUGETLB_PAGE
 200 /*
 201  * __mcopy_atomic processing for HUGETLB vmas.  Note that this routine is
 202  * called with mmap_sem held, it will release mmap_sem before returning.
 203  */
 204 static __always_inline ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
 205                                               struct vm_area_struct *dst_vma,
 206                                               unsigned long dst_start,
 207                                               unsigned long src_start,
 208                                               unsigned long len,
 209                                               bool zeropage)
 210 {
 211         int vm_alloc_shared = dst_vma->vm_flags & VM_SHARED;
 212         int vm_shared = dst_vma->vm_flags & VM_SHARED;
 213         ssize_t err;
 214         pte_t *dst_pte;
 215         unsigned long src_addr, dst_addr;
 216         long copied;
 217         struct page *page;
 218         unsigned long vma_hpagesize;
 219         pgoff_t idx;
 220         u32 hash;
 221         struct address_space *mapping;
 222
 223         /*
 224          * There is no default zero huge page for all huge page sizes as
 225          * supported by hugetlb.  A PMD_SIZE huge pages may exist as used
 226          * by THP.  Since we can not reliably insert a zero page, this
 227          * feature is not supported.
 228          */
 229         if (zeropage) {
 230                 up_read(&dst_mm->mmap_sem);
 231                 return -EINVAL;
 232         }
 233
 234         src_addr = src_start;
 235         dst_addr = dst_start;
 236         copied = 0;
 237         page = NULL;
 238         vma_hpagesize = vma_kernel_pagesize(dst_vma);
 239
 240         /*
 241          * Validate alignment based on huge page size
 242          */
 243         err = -EINVAL;
 244         if (dst_start & (vma_hpagesize - 1) || len & (vma_hpagesize - 1))
 245                 goto out_unlock;
 246
 247 retry:
 248         /*
 249          * On routine entry dst_vma is set.  If we had to drop mmap_sem and
 250          * retry, dst_vma will be set to NULL and we must lookup again.
 251          */
 252         if (!dst_vma) {
 253                 err = -ENOENT;
 254                 dst_vma = find_dst_vma(dst_mm, dst_start, len);
 255                 if (!dst_vma || !is_vm_hugetlb_page(dst_vma))
 256                         goto out_unlock;
 257
 258                 err = -EINVAL;
 259                 if (vma_hpagesize != vma_kernel_pagesize(dst_vma))
 260                         goto out_unlock;
 261
 262                 vm_shared = dst_vma->vm_flags & VM_SHARED;
 263         }
 264
 265         /*
 266          * If not shared, ensure the dst_vma has a anon_vma.
 267          */
 268         err = -ENOMEM;
 269         if (!vm_shared) {
 270                 if (unlikely(anon_vma_prepare(dst_vma)))
 271                         goto out_unlock;
 272         }
 273
 274         while (src_addr < src_start + len) {
 275                 pte_t dst_pteval;
 276
 277                 BUG_ON(dst_addr >= dst_start + len);
 278
 279                 /*
 280                  * Serialize via i_mmap_rwsem and hugetlb_fault_mutex.
 281                  * i_mmap_rwsem ensures the dst_pte remains valid even
 282                  * in the case of shared pmds.  fault mutex prevents
 283                  * races with other faulting threads.
 284                  */
 285                 mapping = dst_vma->vm_file->f_mapping;
 286                 i_mmap_lock_read(mapping);
 287                 idx = linear_page_index(dst_vma, dst_addr);
 288                 hash = hugetlb_fault_mutex_hash(mapping, idx);
 289                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 290
 291                 err = -ENOMEM;
 292                 dst_pte = huge_pte_alloc(dst_mm, dst_addr, vma_hpagesize);
 293                 if (!dst_pte) {
 294                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 295                         i_mmap_unlock_read(mapping);
 296                         goto out_unlock;
 297                 }
 298
 299                 err = -EEXIST;
 300                 dst_pteval = huge_ptep_get(dst_pte);
 301                 if (!huge_pte_none(dst_pteval)) {
 302                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 303                         i_mmap_unlock_read(mapping);
 304                         goto out_unlock;
 305                 }
 306
 307                 err = hugetlb_mcopy_atomic_pte(dst_mm, dst_pte, dst_vma,
 308                                                 dst_addr, src_addr, &page);
 309
 310                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 311                 i_mmap_unlock_read(mapping);
 312                 vm_alloc_shared = vm_shared;
 313
 314                 cond_resched();
 315
 316                 if (unlikely(err == -ENOENT)) {
 317                         up_read(&dst_mm->mmap_sem);
 318                         BUG_ON(!page);
 319
 320                         err = copy_huge_page_from_user(page,
 321                                                 (const void __user *)src_addr,
 322                                                 vma_hpagesize / PAGE_SIZE,
 323                                                 true);
 324                         if (unlikely(err)) {
 325                                 err = -EFAULT;
 326                                 goto out;
 327                         }
 328                         down_read(&dst_mm->mmap_sem);
 329
 330                         dst_vma = NULL;
 331                         goto retry;
 332                 } else
 333                         BUG_ON(page);
 334
 335                 if (!err) {
 336                         dst_addr += vma_hpagesize;
 337                         src_addr += vma_hpagesize;
 338                         copied += vma_hpagesize;
 339
 340                         if (fatal_signal_pending(current))
 341                                 err = -EINTR;
 342                 }
 343                 if (err)
 344                         break;
 345         }
 346
 347 out_unlock:
 348         up_read(&dst_mm->mmap_sem);
 349 out:
 350         if (page) {
 351                 /*
 352                  * We encountered an error and are about to free a newly
 353                  * allocated huge page.
 354                  *
 355                  * Reservation handling is very subtle, and is different for
 356                  * private and shared mappings.  See the routine
 357                  * restore_reserve_on_error for details.  Unfortunately, we
 358                  * can not call restore_reserve_on_error now as it would
 359                  * require holding mmap_sem.
 360                  *
 361                  * If a reservation for the page existed in the reservation
 362                  * map of a private mapping, the map was modified to indicate
 363                  * the reservation was consumed when the page was allocated.
 364                  * We clear the PagePrivate flag now so that the global
 365                  * reserve count will not be incremented in free_huge_page.
 366                  * The reservation map will still indicate the reservation
 367                  * was consumed and possibly prevent later page allocation.
 368                  * This is better than leaking a global reservation.  If no
 369                  * reservation existed, it is still safe to clear PagePrivate
 370                  * as no adjustments to reservation counts were made during
 371                  * allocation.
 372                  *
 373                  * The reservation map for shared mappings indicates which
 374                  * pages have reservations.  When a huge page is allocated
 375                  * for an address with a reservation, no change is made to
 376                  * the reserve map.  In this case PagePrivate will be set
 377                  * to indicate that the global reservation count should be
 378                  * incremented when the page is freed.  This is the desired
 379                  * behavior.  However, when a huge page is allocated for an
 380                  * address without a reservation a reservation entry is added
 381                  * to the reservation map, and PagePrivate will not be set.
 382                  * When the page is freed, the global reserve count will NOT
 383                  * be incremented and it will appear as though we have leaked
 384                  * reserved page.  In this case, set PagePrivate so that the
 385                  * global reserve count will be incremented to match the
 386                  * reservation map entry which was created.
 387                  *
 388                  * Note that vm_alloc_shared is based on the flags of the vma
 389                  * for which the page was originally allocated.  dst_vma could
 390                  * be different or NULL on error.
 391                  */
 392                 if (vm_alloc_shared)
 393                         SetPagePrivate(page);
 394                 else
 395                         ClearPagePrivate(page);
 396                 put_page(page);
 397         }
 398         BUG_ON(copied < 0);
 399         BUG_ON(err > 0);
 400         BUG_ON(!copied && !err);
 401         return copied ? copied : err;
 402 }
 403 #else /* !CONFIG_HUGETLB_PAGE */
 404 /* fail at build time if gcc attempts to use this */
 405 extern ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
 406                                       struct vm_area_struct *dst_vma,
 407                                       unsigned long dst_start,
 408                                       unsigned long src_start,
 409                                       unsigned long len,
 410                                       bool zeropage);
 411 #endif /* CONFIG_HUGETLB_PAGE */
 412
 413 static __always_inline ssize_t mfill_atomic_pte(struct mm_struct *dst_mm,
 414                                                 pmd_t *dst_pmd,
 415                                                 struct vm_area_struct *dst_vma,
 416                                                 unsigned long dst_addr,
 417                                                 unsigned long src_addr,
 418                                                 struct page **page,
 419                                                 bool zeropage,
 420                                                 bool wp_copy)
 421 {
 422         ssize_t err;
 423
 424         /*
 425          * The normal page fault path for a shmem will invoke the
 426          * fault, fill the hole in the file and COW it right away. The
 427          * result generates plain anonymous memory. So when we are
 428          * asked to fill an hole in a MAP_PRIVATE shmem mapping, we'll
 429          * generate anonymous memory directly without actually filling
 430          * the hole. For the MAP_PRIVATE case the robustness check
 431          * only happens in the pagetable (to verify it's still none)
 432          * and not in the radix tree.
 433          */
 434         if (!(dst_vma->vm_flags & VM_SHARED)) {
 435                 if (!zeropage)
 436                         err = mcopy_atomic_pte(dst_mm, dst_pmd, dst_vma,
 437                                                dst_addr, src_addr, page,
 438                                                wp_copy);
 439                 else
 440                         err = mfill_zeropage_pte(dst_mm, dst_pmd,
 441                                                  dst_vma, dst_addr);
 442         } else {
 443                 VM_WARN_ON_ONCE(wp_copy);
 444                 if (!zeropage)
 445                         err = shmem_mcopy_atomic_pte(dst_mm, dst_pmd,
 446                                                      dst_vma, dst_addr,
 447                                                      src_addr, page);
 448                 else
 449                         err = shmem_mfill_zeropage_pte(dst_mm, dst_pmd,
 450                                                        dst_vma, dst_addr);
 451         }
 452
 453         return err;
 454 }
 455
 456 static __always_inline ssize_t __mcopy_atomic(struct mm_struct *dst_mm,
 457                                               unsigned long dst_start,
 458                                               unsigned long src_start,
 459                                               unsigned long len,
 460                                               bool zeropage,
 461                                               bool *mmap_changing,
 462                                               __u64 mode)
 463 {
 464         struct vm_area_struct *dst_vma;
 465         ssize_t err;
 466         pmd_t *dst_pmd;
 467         unsigned long src_addr, dst_addr;
 468         long copied;
 469         struct page *page;
 470         bool wp_copy;
 471
 472         /*
 473          * Sanitize the command parameters:
 474          */
 475         BUG_ON(dst_start & ~PAGE_MASK);
 476         BUG_ON(len & ~PAGE_MASK);
 477
 478         /* Does the address range wrap, or is the span zero-sized? */
 479         BUG_ON(src_start + len <= src_start);
 480         BUG_ON(dst_start + len <= dst_start);
 481
 482         src_addr = src_start;
 483         dst_addr = dst_start;
 484         copied = 0;
 485         page = NULL;
 486 retry:
 487         down_read(&dst_mm->mmap_sem);
 488
 489         /*
 490          * If memory mappings are changing because of non-cooperative
 491          * operation (e.g. mremap) running in parallel, bail out and
 492          * request the user to retry later
 493          */
 494         err = -EAGAIN;
 495         if (mmap_changing && READ_ONCE(*mmap_changing))
 496                 goto out_unlock;
 497
 498         /*
 499          * Make sure the vma is not shared, that the dst range is
 500          * both valid and fully within a single existing vma.
 501          */
 502         err = -ENOENT;
 503         dst_vma = find_dst_vma(dst_mm, dst_start, len);
 504         if (!dst_vma)
 505                 goto out_unlock;
 506
 507         err = -EINVAL;
 508         /*
 509          * shmem_zero_setup is invoked in mmap for MAP_ANONYMOUS|MAP_SHARED but
 510          * it will overwrite vm_ops, so vma_is_anonymous must return false.
 511          */
 512         if (WARN_ON_ONCE(vma_is_anonymous(dst_vma) &&
 513             dst_vma->vm_flags & VM_SHARED))
 514                 goto out_unlock;
 515
 516         /*
 517          * validate 'mode' now that we know the dst_vma: don't allow
 518          * a wrprotect copy if the userfaultfd didn't register as WP.
 519          */
 520         wp_copy = mode & UFFDIO_COPY_MODE_WP;
 521         if (wp_copy && !(dst_vma->vm_flags & VM_UFFD_WP))
 522                 goto out_unlock;
 523
 524         /*
 525          * If this is a HUGETLB vma, pass off to appropriate routine
 526          */
 527         if (is_vm_hugetlb_page(dst_vma))
 528                 return  __mcopy_atomic_hugetlb(dst_mm, dst_vma, dst_start,
 529                                                 src_start, len, zeropage);
 530
 531         if (!vma_is_anonymous(dst_vma) && !vma_is_shmem(dst_vma))
 532                 goto out_unlock;
 533
 534         /*
 535          * Ensure the dst_vma has a anon_vma or this page
 536          * would get a NULL anon_vma when moved in the
 537          * dst_vma.
 538          */
 539         err = -ENOMEM;
 540         if (!(dst_vma->vm_flags & VM_SHARED) &&
 541             unlikely(anon_vma_prepare(dst_vma)))
 542                 goto out_unlock;
 543
 544         while (src_addr < src_start + len) {
 545                 pmd_t dst_pmdval;
 546
 547                 BUG_ON(dst_addr >= dst_start + len);
 548
 549                 dst_pmd = mm_alloc_pmd(dst_mm, dst_addr);
 550                 if (unlikely(!dst_pmd)) {
 551                         err = -ENOMEM;
 552                         break;
 553                 }
 554
 555                 dst_pmdval = pmd_read_atomic(dst_pmd);
 556                 /*
 557                  * If the dst_pmd is mapped as THP don't
 558                  * override it and just be strict.
 559                  */
 560                 if (unlikely(pmd_trans_huge(dst_pmdval))) {
 561                         err = -EEXIST;
 562                         break;
 563                 }
 564                 if (unlikely(pmd_none(dst_pmdval)) &&
 565                     unlikely(__pte_alloc(dst_mm, dst_pmd))) {
 566                         err = -ENOMEM;
 567                         break;
 568                 }
 569                 /* If an huge pmd materialized from under us fail */
 570                 if (unlikely(pmd_trans_huge(*dst_pmd))) {
 571                         err = -EFAULT;
 572                         break;
 573                 }
 574
 575                 BUG_ON(pmd_none(*dst_pmd));
 576                 BUG_ON(pmd_trans_huge(*dst_pmd));
 577
 578                 err = mfill_atomic_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
 579                                        src_addr, &page, zeropage, wp_copy);
 580                 cond_resched();
 581
 582                 if (unlikely(err == -ENOENT)) {
 583                         void *page_kaddr;
 584
 585                         up_read(&dst_mm->mmap_sem);
 586                         BUG_ON(!page);
 587
 588                         page_kaddr = kmap(page);
 589                         err = copy_from_user(page_kaddr,
 590                                              (const void __user *) src_addr,
 591                                              PAGE_SIZE);
 592                         kunmap(page);
 593                         if (unlikely(err)) {
 594                                 err = -EFAULT;
 595                                 goto out;
 596                         }
 597                         goto retry;
 598                 } else
 599                         BUG_ON(page);
 600
 601                 if (!err) {
 602                         dst_addr += PAGE_SIZE;
 603                         src_addr += PAGE_SIZE;
 604                         copied += PAGE_SIZE;
 605
 606                         if (fatal_signal_pending(current))
 607                                 err = -EINTR;
 608                 }
 609                 if (err)
 610                         break;
 611         }
 612
 613 out_unlock:
 614         up_read(&dst_mm->mmap_sem);
 615 out:
 616         if (page)
 617                 put_page(page);
 618         BUG_ON(copied < 0);
 619         BUG_ON(err > 0);
 620         BUG_ON(!copied && !err);
 621         return copied ? copied : err;
 622 }
 623
 624 ssize_t mcopy_atomic(struct mm_struct *dst_mm, unsigned long dst_start,
 625                      unsigned long src_start, unsigned long len,
 626                      bool *mmap_changing, __u64 mode)
 627 {
 628         return __mcopy_atomic(dst_mm, dst_start, src_start, len, false,
 629                               mmap_changing, mode);
 630 }
 631
 632 ssize_t mfill_zeropage(struct mm_struct *dst_mm, unsigned long start,
 633                        unsigned long len, bool *mmap_changing)
 634 {
 635         return __mcopy_atomic(dst_mm, start, 0, len, true, mmap_changing, 0);
 636 }