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/pagemap.h>
  20 #include <linux/shmem_fs.h>
  21 #include <asm/tlbflush.h>
  22 #include "internal.h"
  23
  24 static int mcopy_atomic_pte(struct mm_struct *dst_mm,
  25                             pmd_t *dst_pmd,
  26                             struct vm_area_struct *dst_vma,
  27                             unsigned long dst_addr,
  28                             unsigned long src_addr,
  29                             struct page **pagep)
  30 {
  31         struct mem_cgroup *memcg;
  32         pte_t _dst_pte, *dst_pte;
  33         spinlock_t *ptl;
  34         void *page_kaddr;
  35         int ret;
  36         struct page *page;
  37         pgoff_t offset, max_off;
  38         struct inode *inode;
  39
  40         if (!*pagep) {
  41                 ret = -ENOMEM;
  42                 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, dst_vma, dst_addr);
  43                 if (!page)
  44                         goto out;
  45
  46                 page_kaddr = kmap_atomic(page);
  47                 ret = copy_from_user(page_kaddr,
  48                                      (const void __user *) src_addr,
  49                                      PAGE_SIZE);
  50                 kunmap_atomic(page_kaddr);
  51
  52                 /* fallback to copy_from_user outside mmap_sem */
  53                 if (unlikely(ret)) {
  54                         ret = -ENOENT;
  55                         *pagep = page;
  56                         /* don't free the page */
  57                         goto out;
  58                 }
  59
  60                 flush_dcache_page(page);
  61         } else {
  62                 page = *pagep;
  63                 *pagep = NULL;
  64         }
  65
  66         /*
  67          * The memory barrier inside __SetPageUptodate makes sure that
  68          * preceeding stores to the page contents become visible before
  69          * the set_pte_at() write.
  70          */
  71         __SetPageUptodate(page);
  72
  73         ret = -ENOMEM;
  74         if (mem_cgroup_try_charge(page, dst_mm, GFP_KERNEL, &memcg, false))
  75                 goto out_release;
  76
  77         _dst_pte = mk_pte(page, dst_vma->vm_page_prot);
  78         if (dst_vma->vm_flags & VM_WRITE)
  79                 _dst_pte = pte_mkwrite(pte_mkdirty(_dst_pte));
  80
  81         dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
  82         if (dst_vma->vm_file) {
  83                 /* the shmem MAP_PRIVATE case requires checking the i_size */
  84                 inode = dst_vma->vm_file->f_inode;
  85                 offset = linear_page_index(dst_vma, dst_addr);
  86                 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
  87                 ret = -EFAULT;
  88                 if (unlikely(offset >= max_off))
  89                         goto out_release_uncharge_unlock;
  90         }
  91         ret = -EEXIST;
  92         if (!pte_none(*dst_pte))
  93                 goto out_release_uncharge_unlock;
  94
  95         inc_mm_counter(dst_mm, MM_ANONPAGES);
  96         page_add_new_anon_rmap(page, dst_vma, dst_addr, false);
  97         mem_cgroup_commit_charge(page, memcg, false, false);
  98         lru_cache_add_active_or_unevictable(page, dst_vma);
  99
 100         set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
 101
 102         /* No need to invalidate - it was non-present before */
 103         update_mmu_cache(dst_vma, dst_addr, dst_pte);
 104
 105         pte_unmap_unlock(dst_pte, ptl);
 106         ret = 0;
 107 out:
 108         return ret;
 109 out_release_uncharge_unlock:
 110         pte_unmap_unlock(dst_pte, ptl);
 111         mem_cgroup_cancel_charge(page, memcg, false);
 112 out_release:
 113         put_page(page);
 114         goto out;
 115 }
 116
 117 static int mfill_zeropage_pte(struct mm_struct *dst_mm,
 118                               pmd_t *dst_pmd,
 119                               struct vm_area_struct *dst_vma,
 120                               unsigned long dst_addr)
 121 {
 122         pte_t _dst_pte, *dst_pte;
 123         spinlock_t *ptl;
 124         int ret;
 125         pgoff_t offset, max_off;
 126         struct inode *inode;
 127
 128         _dst_pte = pte_mkspecial(pfn_pte(my_zero_pfn(dst_addr),
 129                                          dst_vma->vm_page_prot));
 130         dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
 131         if (dst_vma->vm_file) {
 132                 /* the shmem MAP_PRIVATE case requires checking the i_size */
 133                 inode = dst_vma->vm_file->f_inode;
 134                 offset = linear_page_index(dst_vma, dst_addr);
 135                 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
 136                 ret = -EFAULT;
 137                 if (unlikely(offset >= max_off))
 138                         goto out_unlock;
 139         }
 140         ret = -EEXIST;
 141         if (!pte_none(*dst_pte))
 142                 goto out_unlock;
 143         set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
 144         /* No need to invalidate - it was non-present before */
 145         update_mmu_cache(dst_vma, dst_addr, dst_pte);
 146         ret = 0;
 147 out_unlock:
 148         pte_unmap_unlock(dst_pte, ptl);
 149         return ret;
 150 }
 151
 152 static pmd_t *mm_alloc_pmd(struct mm_struct *mm, unsigned long address)
 153 {
 154         pgd_t *pgd;
 155         p4d_t *p4d;
 156         pud_t *pud;
 157
 158         pgd = pgd_offset(mm, address);
 159         p4d = p4d_alloc(mm, pgd, address);
 160         if (!p4d)
 161                 return NULL;
 162         pud = pud_alloc(mm, p4d, address);
 163         if (!pud)
 164                 return NULL;
 165         /*
 166          * Note that we didn't run this because the pmd was
 167          * missing, the *pmd may be already established and in
 168          * turn it may also be a trans_huge_pmd.
 169          */
 170         return pmd_alloc(mm, pud, address);
 171 }
 172
 173 #ifdef CONFIG_HUGETLB_PAGE
 174 /*
 175  * __mcopy_atomic processing for HUGETLB vmas.  Note that this routine is
 176  * called with mmap_sem held, it will release mmap_sem before returning.
 177  */
 178 static __always_inline ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
 179                                               struct vm_area_struct *dst_vma,
 180                                               unsigned long dst_start,
 181                                               unsigned long src_start,
 182                                               unsigned long len,
 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                         dst_vma = NULL;
 316                         goto retry;
 317                 } else
 318                         BUG_ON(page);
 319
 320                 if (!err) {
 321                         dst_addr += vma_hpagesize;
 322                         src_addr += vma_hpagesize;
 323                         copied += vma_hpagesize;
 324
 325                         if (fatal_signal_pending(current))
 326                                 err = -EINTR;
 327                 }
 328                 if (err)
 329                         break;
 330         }
 331
 332 out_unlock:
 333         up_read(&dst_mm->mmap_sem);
 334 out:
 335         if (page) {
 336                 /*
 337                  * We encountered an error and are about to free a newly
 338                  * allocated huge page.
 339                  *
 340                  * Reservation handling is very subtle, and is different for
 341                  * private and shared mappings.  See the routine
 342                  * restore_reserve_on_error for details.  Unfortunately, we
 343                  * can not call restore_reserve_on_error now as it would
 344                  * require holding mmap_sem.
 345                  *
 346                  * If a reservation for the page existed in the reservation
 347                  * map of a private mapping, the map was modified to indicate
 348                  * the reservation was consumed when the page was allocated.
 349                  * We clear the PagePrivate flag now so that the global
 350                  * reserve count will not be incremented in free_huge_page.
 351                  * The reservation map will still indicate the reservation
 352                  * was consumed and possibly prevent later page allocation.
 353                  * This is better than leaking a global reservation.  If no
 354                  * reservation existed, it is still safe to clear PagePrivate
 355                  * as no adjustments to reservation counts were made during
 356                  * allocation.
 357                  *
 358                  * The reservation map for shared mappings indicates which
 359                  * pages have reservations.  When a huge page is allocated
 360                  * for an address with a reservation, no change is made to
 361                  * the reserve map.  In this case PagePrivate will be set
 362                  * to indicate that the global reservation count should be
 363                  * incremented when the page is freed.  This is the desired
 364                  * behavior.  However, when a huge page is allocated for an
 365                  * address without a reservation a reservation entry is added
 366                  * to the reservation map, and PagePrivate will not be set.
 367                  * When the page is freed, the global reserve count will NOT
 368                  * be incremented and it will appear as though we have leaked
 369                  * reserved page.  In this case, set PagePrivate so that the
 370                  * global reserve count will be incremented to match the
 371                  * reservation map entry which was created.
 372                  *
 373                  * Note that vm_alloc_shared is based on the flags of the vma
 374                  * for which the page was originally allocated.  dst_vma could
 375                  * be different or NULL on error.
 376                  */
 377                 if (vm_alloc_shared)
 378                         SetPagePrivate(page);
 379                 else
 380                         ClearPagePrivate(page);
 381                 put_page(page);
 382         }
 383         BUG_ON(copied < 0);
 384         BUG_ON(err > 0);
 385         BUG_ON(!copied && !err);
 386         return copied ? copied : err;
 387 }
 388 #else /* !CONFIG_HUGETLB_PAGE */
 389 /* fail at build time if gcc attempts to use this */
 390 extern ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
 391                                       struct vm_area_struct *dst_vma,
 392                                       unsigned long dst_start,
 393                                       unsigned long src_start,
 394                                       unsigned long len,
 395                                       bool zeropage);
 396 #endif /* CONFIG_HUGETLB_PAGE */
 397
 398 static __always_inline ssize_t mfill_atomic_pte(struct mm_struct *dst_mm,
 399                                                 pmd_t *dst_pmd,
 400                                                 struct vm_area_struct *dst_vma,
 401                                                 unsigned long dst_addr,
 402                                                 unsigned long src_addr,
 403                                                 struct page **page,
 404                                                 bool zeropage)
 405 {
 406         ssize_t err;
 407
 408         /*
 409          * The normal page fault path for a shmem will invoke the
 410          * fault, fill the hole in the file and COW it right away. The
 411          * result generates plain anonymous memory. So when we are
 412          * asked to fill an hole in a MAP_PRIVATE shmem mapping, we'll
 413          * generate anonymous memory directly without actually filling
 414          * the hole. For the MAP_PRIVATE case the robustness check
 415          * only happens in the pagetable (to verify it's still none)
 416          * and not in the radix tree.
 417          */
 418         if (!(dst_vma->vm_flags & VM_SHARED)) {
 419                 if (!zeropage)
 420                         err = mcopy_atomic_pte(dst_mm, dst_pmd, dst_vma,
 421                                                dst_addr, src_addr, page);
 422                 else
 423                         err = mfill_zeropage_pte(dst_mm, dst_pmd,
 424                                                  dst_vma, dst_addr);
 425         } else {
 426                 if (!zeropage)
 427                         err = shmem_mcopy_atomic_pte(dst_mm, dst_pmd,
 428                                                      dst_vma, dst_addr,
 429                                                      src_addr, page);
 430                 else
 431                         err = shmem_mfill_zeropage_pte(dst_mm, dst_pmd,
 432                                                        dst_vma, dst_addr);
 433         }
 434
 435         return err;
 436 }
 437
 438 static __always_inline ssize_t __mcopy_atomic(struct mm_struct *dst_mm,
 439                                               unsigned long dst_start,
 440                                               unsigned long src_start,
 441                                               unsigned long len,
 442                                               bool zeropage)
 443 {
 444         struct vm_area_struct *dst_vma;
 445         ssize_t err;
 446         pmd_t *dst_pmd;
 447         unsigned long src_addr, dst_addr;
 448         long copied;
 449         struct page *page;
 450
 451         /*
 452          * Sanitize the command parameters:
 453          */
 454         BUG_ON(dst_start & ~PAGE_MASK);
 455         BUG_ON(len & ~PAGE_MASK);
 456
 457         /* Does the address range wrap, or is the span zero-sized? */
 458         BUG_ON(src_start + len <= src_start);
 459         BUG_ON(dst_start + len <= dst_start);
 460
 461         src_addr = src_start;
 462         dst_addr = dst_start;
 463         copied = 0;
 464         page = NULL;
 465 retry:
 466         down_read(&dst_mm->mmap_sem);
 467
 468         /*
 469          * Make sure the vma is not shared, that the dst range is
 470          * both valid and fully within a single existing vma.
 471          */
 472         err = -ENOENT;
 473         dst_vma = find_vma(dst_mm, dst_start);
 474         if (!dst_vma)
 475                 goto out_unlock;
 476         /*
 477          * Check the vma is registered in uffd, this is required to
 478          * enforce the VM_MAYWRITE check done at uffd registration
 479          * time.
 480          */
 481         if (!dst_vma->vm_userfaultfd_ctx.ctx)
 482                 goto out_unlock;
 483
 484         if (dst_start < dst_vma->vm_start ||
 485             dst_start + len > dst_vma->vm_end)
 486                 goto out_unlock;
 487
 488         err = -EINVAL;
 489         /*
 490          * shmem_zero_setup is invoked in mmap for MAP_ANONYMOUS|MAP_SHARED but
 491          * it will overwrite vm_ops, so vma_is_anonymous must return false.
 492          */
 493         if (WARN_ON_ONCE(vma_is_anonymous(dst_vma) &&
 494             dst_vma->vm_flags & VM_SHARED))
 495                 goto out_unlock;
 496
 497         /*
 498          * If this is a HUGETLB vma, pass off to appropriate routine
 499          */
 500         if (is_vm_hugetlb_page(dst_vma))
 501                 return  __mcopy_atomic_hugetlb(dst_mm, dst_vma, dst_start,
 502                                                 src_start, len, zeropage);
 503
 504         if (!vma_is_anonymous(dst_vma) && !vma_is_shmem(dst_vma))
 505                 goto out_unlock;
 506
 507         /*
 508          * Ensure the dst_vma has a anon_vma or this page
 509          * would get a NULL anon_vma when moved in the
 510          * dst_vma.
 511          */
 512         err = -ENOMEM;
 513         if (!(dst_vma->vm_flags & VM_SHARED) &&
 514             unlikely(anon_vma_prepare(dst_vma)))
 515                 goto out_unlock;
 516
 517         while (src_addr < src_start + len) {
 518                 pmd_t dst_pmdval;
 519
 520                 BUG_ON(dst_addr >= dst_start + len);
 521
 522                 dst_pmd = mm_alloc_pmd(dst_mm, dst_addr);
 523                 if (unlikely(!dst_pmd)) {
 524                         err = -ENOMEM;
 525                         break;
 526                 }
 527
 528                 dst_pmdval = pmd_read_atomic(dst_pmd);
 529                 /*
 530                  * If the dst_pmd is mapped as THP don't
 531                  * override it and just be strict.
 532                  */
 533                 if (unlikely(pmd_trans_huge(dst_pmdval))) {
 534                         err = -EEXIST;
 535                         break;
 536                 }
 537                 if (unlikely(pmd_none(dst_pmdval)) &&
 538                     unlikely(__pte_alloc(dst_mm, dst_pmd, dst_addr))) {
 539                         err = -ENOMEM;
 540                         break;
 541                 }
 542                 /* If an huge pmd materialized from under us fail */
 543                 if (unlikely(pmd_trans_huge(*dst_pmd))) {
 544                         err = -EFAULT;
 545                         break;
 546                 }
 547
 548                 BUG_ON(pmd_none(*dst_pmd));
 549                 BUG_ON(pmd_trans_huge(*dst_pmd));
 550
 551                 err = mfill_atomic_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
 552                                        src_addr, &page, zeropage);
 553                 cond_resched();
 554
 555                 if (unlikely(err == -ENOENT)) {
 556                         void *page_kaddr;
 557
 558                         up_read(&dst_mm->mmap_sem);
 559                         BUG_ON(!page);
 560
 561                         page_kaddr = kmap(page);
 562                         err = copy_from_user(page_kaddr,
 563                                              (const void __user *) src_addr,
 564                                              PAGE_SIZE);
 565                         kunmap(page);
 566                         if (unlikely(err)) {
 567                                 err = -EFAULT;
 568                                 goto out;
 569                         }
 570                         flush_dcache_page(page);
 571                         goto retry;
 572                 } else
 573                         BUG_ON(page);
 574
 575                 if (!err) {
 576                         dst_addr += PAGE_SIZE;
 577                         src_addr += PAGE_SIZE;
 578                         copied += PAGE_SIZE;
 579
 580                         if (fatal_signal_pending(current))
 581                                 err = -EINTR;
 582                 }
 583                 if (err)
 584                         break;
 585         }
 586
 587 out_unlock:
 588         up_read(&dst_mm->mmap_sem);
 589 out:
 590         if (page)
 591                 put_page(page);
 592         BUG_ON(copied < 0);
 593         BUG_ON(err > 0);
 594         BUG_ON(!copied && !err);
 595         return copied ? copied : err;
 596 }
 597
 598 ssize_t mcopy_atomic(struct mm_struct *dst_mm, unsigned long dst_start,
 599                      unsigned long src_start, unsigned long len)
 600 {
 601         return __mcopy_atomic(dst_mm, dst_start, src_start, len, false);
 602 }
 603
 604 ssize_t mfill_zeropage(struct mm_struct *dst_mm, unsigned long start,
 605                        unsigned long len)
 606 {
 607         return __mcopy_atomic(dst_mm, start, 0, len, true);
 608 }