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