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