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