mm/mprotect.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  *  mm/mprotect.c
   4  *
   5  *  (C) Copyright 1994 Linus Torvalds
   6  *  (C) Copyright 2002 Christoph Hellwig
   7  *
   8  *  Address space accounting code       <[email protected]>
   9  *  (C) Copyright 2002 Red Hat Inc, All Rights Reserved
  10  */
  11
  12 #include <linux/mm.h>
  13 #include <linux/hugetlb.h>
  14 #include <linux/shm.h>
  15 #include <linux/mman.h>
  16 #include <linux/fs.h>
  17 #include <linux/highmem.h>
  18 #include <linux/security.h>
  19 #include <linux/mempolicy.h>
  20 #include <linux/personality.h>
  21 #include <linux/syscalls.h>
  22 #include <linux/swap.h>
  23 #include <linux/swapops.h>
  24 #include <linux/mmu_notifier.h>
  25 #include <linux/migrate.h>
  26 #include <linux/perf_event.h>
  27 #include <linux/pkeys.h>
  28 #include <linux/ksm.h>
  29 #include <linux/uaccess.h>
  30 #include <linux/mm_inline.h>
  31 #include <asm/pgtable.h>
  32 #include <asm/cacheflush.h>
  33 #include <asm/mmu_context.h>
  34 #include <asm/tlbflush.h>
  35
  36 #include "internal.h"
  37
  38 static unsigned long change_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
  39                 unsigned long addr, unsigned long end, pgprot_t newprot,
  40                 int dirty_accountable, int prot_numa)
  41 {
  42         struct mm_struct *mm = vma->vm_mm;
  43         pte_t *pte, oldpte;
  44         spinlock_t *ptl;
  45         unsigned long pages = 0;
  46         int target_node = NUMA_NO_NODE;
  47
  48         /*
  49          * Can be called with only the mmap_sem for reading by
  50          * prot_numa so we must check the pmd isn't constantly
  51          * changing from under us from pmd_none to pmd_trans_huge
  52          * and/or the other way around.
  53          */
  54         if (pmd_trans_unstable(pmd))
  55                 return 0;
  56
  57         /*
  58          * The pmd points to a regular pte so the pmd can't change
  59          * from under us even if the mmap_sem is only hold for
  60          * reading.
  61          */
  62         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
  63
  64         /* Get target node for single threaded private VMAs */
  65         if (prot_numa && !(vma->vm_flags & VM_SHARED) &&
  66             atomic_read(&vma->vm_mm->mm_users) == 1)
  67                 target_node = numa_node_id();
  68
  69         flush_tlb_batched_pending(vma->vm_mm);
  70         arch_enter_lazy_mmu_mode();
  71         do {
  72                 oldpte = *pte;
  73                 if (pte_present(oldpte)) {
  74                         pte_t ptent;
  75                         bool preserve_write = prot_numa && pte_write(oldpte);
  76
  77                         /*
  78                          * Avoid trapping faults against the zero or KSM
  79                          * pages. See similar comment in change_huge_pmd.
  80                          */
  81                         if (prot_numa) {
  82                                 struct page *page;
  83
  84                                 page = vm_normal_page(vma, addr, oldpte);
  85                                 if (!page || PageKsm(page))
  86                                         continue;
  87
  88                                 /* Also skip shared copy-on-write pages */
  89                                 if (is_cow_mapping(vma->vm_flags) &&
  90                                     page_mapcount(page) != 1)
  91                                         continue;
  92
  93                                 /*
  94                                  * While migration can move some dirty pages,
  95                                  * it cannot move them all from MIGRATE_ASYNC
  96                                  * context.
  97                                  */
  98                                 if (page_is_file_cache(page) && PageDirty(page))
  99                                         continue;
 100
 101                                 /* Avoid TLB flush if possible */
 102                                 if (pte_protnone(oldpte))
 103                                         continue;
 104
 105                                 /*
 106                                  * Don't mess with PTEs if page is already on the node
 107                                  * a single-threaded process is running on.
 108                                  */
 109                                 if (target_node == page_to_nid(page))
 110                                         continue;
 111                         }
 112
 113                         ptent = ptep_modify_prot_start(vma, addr, pte);
 114                         ptent = pte_modify(ptent, newprot);
 115                         if (preserve_write)
 116                                 ptent = pte_mk_savedwrite(ptent);
 117
 118                         /* Avoid taking write faults for known dirty pages */
 119                         if (dirty_accountable && pte_dirty(ptent) &&
 120                                         (pte_soft_dirty(ptent) ||
 121                                          !(vma->vm_flags & VM_SOFTDIRTY))) {
 122                                 ptent = pte_mkwrite(ptent);
 123                         }
 124                         ptep_modify_prot_commit(vma, addr, pte, ptent);
 125                         pages++;
 126                 } else if (IS_ENABLED(CONFIG_MIGRATION)) {
 127                         swp_entry_t entry = pte_to_swp_entry(oldpte);
 128
 129                         if (is_write_migration_entry(entry)) {
 130                                 pte_t newpte;
 131                                 /*
 132                                  * A protection check is difficult so
 133                                  * just be safe and disable write
 134                                  */
 135                                 make_migration_entry_read(&entry);
 136                                 newpte = swp_entry_to_pte(entry);
 137                                 if (pte_swp_soft_dirty(oldpte))
 138                                         newpte = pte_swp_mksoft_dirty(newpte);
 139                                 set_pte_at(mm, addr, pte, newpte);
 140
 141                                 pages++;
 142                         }
 143
 144                         if (is_write_device_private_entry(entry)) {
 145                                 pte_t newpte;
 146
 147                                 /*
 148                                  * We do not preserve soft-dirtiness. See
 149                                  * copy_one_pte() for explanation.
 150                                  */
 151                                 make_device_private_entry_read(&entry);
 152                                 newpte = swp_entry_to_pte(entry);
 153                                 set_pte_at(mm, addr, pte, newpte);
 154
 155                                 pages++;
 156                         }
 157                 }
 158         } while (pte++, addr += PAGE_SIZE, addr != end);
 159         arch_leave_lazy_mmu_mode();
 160         pte_unmap_unlock(pte - 1, ptl);
 161
 162         return pages;
 163 }
 164
 165 static inline unsigned long change_pmd_range(struct vm_area_struct *vma,
 166                 pud_t *pud, unsigned long addr, unsigned long end,
 167                 pgprot_t newprot, int dirty_accountable, int prot_numa)
 168 {
 169         pmd_t *pmd;
 170         unsigned long next;
 171         unsigned long pages = 0;
 172         unsigned long nr_huge_updates = 0;
 173         struct mmu_notifier_range range;
 174
 175         range.start = 0;
 176
 177         pmd = pmd_offset(pud, addr);
 178         do {
 179                 unsigned long this_pages;
 180
 181                 next = pmd_addr_end(addr, end);
 182                 if (!is_swap_pmd(*pmd) && !pmd_trans_huge(*pmd) && !pmd_devmap(*pmd)
 183                                 && pmd_none_or_clear_bad(pmd))
 184                         goto next;
 185
 186                 /* invoke the mmu notifier if the pmd is populated */
 187                 if (!range.start) {
 188                         mmu_notifier_range_init(&range, vma->vm_mm, addr, end);
 189                         mmu_notifier_invalidate_range_start(&range);
 190                 }
 191
 192                 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
 193                         if (next - addr != HPAGE_PMD_SIZE) {
 194                                 __split_huge_pmd(vma, pmd, addr, false, NULL);
 195                         } else {
 196                                 int nr_ptes = change_huge_pmd(vma, pmd, addr,
 197                                                 newprot, prot_numa);
 198
 199                                 if (nr_ptes) {
 200                                         if (nr_ptes == HPAGE_PMD_NR) {
 201                                                 pages += HPAGE_PMD_NR;
 202                                                 nr_huge_updates++;
 203                                         }
 204
 205                                         /* huge pmd was handled */
 206                                         goto next;
 207                                 }
 208                         }
 209                         /* fall through, the trans huge pmd just split */
 210                 }
 211                 this_pages = change_pte_range(vma, pmd, addr, next, newprot,
 212                                  dirty_accountable, prot_numa);
 213                 pages += this_pages;
 214 next:
 215                 cond_resched();
 216         } while (pmd++, addr = next, addr != end);
 217
 218         if (range.start)
 219                 mmu_notifier_invalidate_range_end(&range);
 220
 221         if (nr_huge_updates)
 222                 count_vm_numa_events(NUMA_HUGE_PTE_UPDATES, nr_huge_updates);
 223         return pages;
 224 }
 225
 226 static inline unsigned long change_pud_range(struct vm_area_struct *vma,
 227                 p4d_t *p4d, unsigned long addr, unsigned long end,
 228                 pgprot_t newprot, int dirty_accountable, int prot_numa)
 229 {
 230         pud_t *pud;
 231         unsigned long next;
 232         unsigned long pages = 0;
 233
 234         pud = pud_offset(p4d, addr);
 235         do {
 236                 next = pud_addr_end(addr, end);
 237                 if (pud_none_or_clear_bad(pud))
 238                         continue;
 239                 pages += change_pmd_range(vma, pud, addr, next, newprot,
 240                                  dirty_accountable, prot_numa);
 241         } while (pud++, addr = next, addr != end);
 242
 243         return pages;
 244 }
 245
 246 static inline unsigned long change_p4d_range(struct vm_area_struct *vma,
 247                 pgd_t *pgd, unsigned long addr, unsigned long end,
 248                 pgprot_t newprot, int dirty_accountable, int prot_numa)
 249 {
 250         p4d_t *p4d;
 251         unsigned long next;
 252         unsigned long pages = 0;
 253
 254         p4d = p4d_offset(pgd, addr);
 255         do {
 256                 next = p4d_addr_end(addr, end);
 257                 if (p4d_none_or_clear_bad(p4d))
 258                         continue;
 259                 pages += change_pud_range(vma, p4d, addr, next, newprot,
 260                                  dirty_accountable, prot_numa);
 261         } while (p4d++, addr = next, addr != end);
 262
 263         return pages;
 264 }
 265
 266 static unsigned long change_protection_range(struct vm_area_struct *vma,
 267                 unsigned long addr, unsigned long end, pgprot_t newprot,
 268                 int dirty_accountable, int prot_numa)
 269 {
 270         struct mm_struct *mm = vma->vm_mm;
 271         pgd_t *pgd;
 272         unsigned long next;
 273         unsigned long start = addr;
 274         unsigned long pages = 0;
 275
 276         BUG_ON(addr >= end);
 277         pgd = pgd_offset(mm, addr);
 278         flush_cache_range(vma, addr, end);
 279         inc_tlb_flush_pending(mm);
 280         do {
 281                 next = pgd_addr_end(addr, end);
 282                 if (pgd_none_or_clear_bad(pgd))
 283                         continue;
 284                 pages += change_p4d_range(vma, pgd, addr, next, newprot,
 285                                  dirty_accountable, prot_numa);
 286         } while (pgd++, addr = next, addr != end);
 287
 288         /* Only flush the TLB if we actually modified any entries: */
 289         if (pages)
 290                 flush_tlb_range(vma, start, end);
 291         dec_tlb_flush_pending(mm);
 292
 293         return pages;
 294 }
 295
 296 unsigned long change_protection(struct vm_area_struct *vma, unsigned long start,
 297                        unsigned long end, pgprot_t newprot,
 298                        int dirty_accountable, int prot_numa)
 299 {
 300         unsigned long pages;
 301
 302         if (is_vm_hugetlb_page(vma))
 303                 pages = hugetlb_change_protection(vma, start, end, newprot);
 304         else
 305                 pages = change_protection_range(vma, start, end, newprot, dirty_accountable, prot_numa);
 306
 307         return pages;
 308 }
 309
 310 static int prot_none_pte_entry(pte_t *pte, unsigned long addr,
 311                                unsigned long next, struct mm_walk *walk)
 312 {
 313         return pfn_modify_allowed(pte_pfn(*pte), *(pgprot_t *)(walk->private)) ?
 314                 0 : -EACCES;
 315 }
 316
 317 static int prot_none_hugetlb_entry(pte_t *pte, unsigned long hmask,
 318                                    unsigned long addr, unsigned long next,
 319                                    struct mm_walk *walk)
 320 {
 321         return pfn_modify_allowed(pte_pfn(*pte), *(pgprot_t *)(walk->private)) ?
 322                 0 : -EACCES;
 323 }
 324
 325 static int prot_none_test(unsigned long addr, unsigned long next,
 326                           struct mm_walk *walk)
 327 {
 328         return 0;
 329 }
 330
 331 static int prot_none_walk(struct vm_area_struct *vma, unsigned long start,
 332                            unsigned long end, unsigned long newflags)
 333 {
 334         pgprot_t new_pgprot = vm_get_page_prot(newflags);
 335         struct mm_walk prot_none_walk = {
 336                 .pte_entry = prot_none_pte_entry,
 337                 .hugetlb_entry = prot_none_hugetlb_entry,
 338                 .test_walk = prot_none_test,
 339                 .mm = current->mm,
 340                 .private = &new_pgprot,
 341         };
 342
 343         return walk_page_range(start, end, &prot_none_walk);
 344 }
 345
 346 int
 347 mprotect_fixup(struct vm_area_struct *vma, struct vm_area_struct **pprev,
 348         unsigned long start, unsigned long end, unsigned long newflags)
 349 {
 350         struct mm_struct *mm = vma->vm_mm;
 351         unsigned long oldflags = vma->vm_flags;
 352         long nrpages = (end - start) >> PAGE_SHIFT;
 353         unsigned long charged = 0;
 354         pgoff_t pgoff;
 355         int error;
 356         int dirty_accountable = 0;
 357
 358         if (newflags == oldflags) {
 359                 *pprev = vma;
 360                 return 0;
 361         }
 362
 363         /*
 364          * Do PROT_NONE PFN permission checks here when we can still
 365          * bail out without undoing a lot of state. This is a rather
 366          * uncommon case, so doesn't need to be very optimized.
 367          */
 368         if (arch_has_pfn_modify_check() &&
 369             (vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 370             (newflags & (VM_READ|VM_WRITE|VM_EXEC)) == 0) {
 371                 error = prot_none_walk(vma, start, end, newflags);
 372                 if (error)
 373                         return error;
 374         }
 375
 376         /*
 377          * If we make a private mapping writable we increase our commit;
 378          * but (without finer accounting) cannot reduce our commit if we
 379          * make it unwritable again. hugetlb mapping were accounted for
 380          * even if read-only so there is no need to account for them here
 381          */
 382         if (newflags & VM_WRITE) {
 383                 /* Check space limits when area turns into data. */
 384                 if (!may_expand_vm(mm, newflags, nrpages) &&
 385                                 may_expand_vm(mm, oldflags, nrpages))
 386                         return -ENOMEM;
 387                 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
 388                                                 VM_SHARED|VM_NORESERVE))) {
 389                         charged = nrpages;
 390                         if (security_vm_enough_memory_mm(mm, charged))
 391                                 return -ENOMEM;
 392                         newflags |= VM_ACCOUNT;
 393                 }
 394         }
 395
 396         /*
 397          * First try to merge with previous and/or next vma.
 398          */
 399         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 400         *pprev = vma_merge(mm, *pprev, start, end, newflags,
 401                            vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
 402                            vma->vm_userfaultfd_ctx);
 403         if (*pprev) {
 404                 vma = *pprev;
 405                 VM_WARN_ON((vma->vm_flags ^ newflags) & ~VM_SOFTDIRTY);
 406                 goto success;
 407         }
 408
 409         *pprev = vma;
 410
 411         if (start != vma->vm_start) {
 412                 error = split_vma(mm, vma, start, 1);
 413                 if (error)
 414                         goto fail;
 415         }
 416
 417         if (end != vma->vm_end) {
 418                 error = split_vma(mm, vma, end, 0);
 419                 if (error)
 420                         goto fail;
 421         }
 422
 423 success:
 424         /*
 425          * vm_flags and vm_page_prot are protected by the mmap_sem
 426          * held in write mode.
 427          */
 428         vma->vm_flags = newflags;
 429         dirty_accountable = vma_wants_writenotify(vma, vma->vm_page_prot);
 430         vma_set_page_prot(vma);
 431
 432         change_protection(vma, start, end, vma->vm_page_prot,
 433                           dirty_accountable, 0);
 434
 435         /*
 436          * Private VM_LOCKED VMA becoming writable: trigger COW to avoid major
 437          * fault on access.
 438          */
 439         if ((oldflags & (VM_WRITE | VM_SHARED | VM_LOCKED)) == VM_LOCKED &&
 440                         (newflags & VM_WRITE)) {
 441                 populate_vma_page_range(vma, start, end, NULL);
 442         }
 443
 444         vm_stat_account(mm, oldflags, -nrpages);
 445         vm_stat_account(mm, newflags, nrpages);
 446         perf_event_mmap(vma);
 447         return 0;
 448
 449 fail:
 450         vm_unacct_memory(charged);
 451         return error;
 452 }
 453
 454 /*
 455  * pkey==-1 when doing a legacy mprotect()
 456  */
 457 static int do_mprotect_pkey(unsigned long start, size_t len,
 458                 unsigned long prot, int pkey)
 459 {
 460         unsigned long nstart, end, tmp, reqprot;
 461         struct vm_area_struct *vma, *prev;
 462         int error = -EINVAL;
 463         const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
 464         const bool rier = (current->personality & READ_IMPLIES_EXEC) &&
 465                                 (prot & PROT_READ);
 466
 467         prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
 468         if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
 469                 return -EINVAL;
 470
 471         if (start & ~PAGE_MASK)
 472                 return -EINVAL;
 473         if (!len)
 474                 return 0;
 475         len = PAGE_ALIGN(len);
 476         end = start + len;
 477         if (end <= start)
 478                 return -ENOMEM;
 479         if (!arch_validate_prot(prot, start))
 480                 return -EINVAL;
 481
 482         reqprot = prot;
 483
 484         if (down_write_killable(&current->mm->mmap_sem))
 485                 return -EINTR;
 486
 487         /*
 488          * If userspace did not allocate the pkey, do not let
 489          * them use it here.
 490          */
 491         error = -EINVAL;
 492         if ((pkey != -1) && !mm_pkey_is_allocated(current->mm, pkey))
 493                 goto out;
 494
 495         vma = find_vma(current->mm, start);
 496         error = -ENOMEM;
 497         if (!vma)
 498                 goto out;
 499         prev = vma->vm_prev;
 500         if (unlikely(grows & PROT_GROWSDOWN)) {
 501                 if (vma->vm_start >= end)
 502                         goto out;
 503                 start = vma->vm_start;
 504                 error = -EINVAL;
 505                 if (!(vma->vm_flags & VM_GROWSDOWN))
 506                         goto out;
 507         } else {
 508                 if (vma->vm_start > start)
 509                         goto out;
 510                 if (unlikely(grows & PROT_GROWSUP)) {
 511                         end = vma->vm_end;
 512                         error = -EINVAL;
 513                         if (!(vma->vm_flags & VM_GROWSUP))
 514                                 goto out;
 515                 }
 516         }
 517         if (start > vma->vm_start)
 518                 prev = vma;
 519
 520         for (nstart = start ; ; ) {
 521                 unsigned long mask_off_old_flags;
 522                 unsigned long newflags;
 523                 int new_vma_pkey;
 524
 525                 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
 526
 527                 /* Does the application expect PROT_READ to imply PROT_EXEC */
 528                 if (rier && (vma->vm_flags & VM_MAYEXEC))
 529                         prot |= PROT_EXEC;
 530
 531                 /*
 532                  * Each mprotect() call explicitly passes r/w/x permissions.
 533                  * If a permission is not passed to mprotect(), it must be
 534                  * cleared from the VMA.
 535                  */
 536                 mask_off_old_flags = VM_READ | VM_WRITE | VM_EXEC |
 537                                         VM_FLAGS_CLEAR;
 538
 539                 new_vma_pkey = arch_override_mprotect_pkey(vma, prot, pkey);
 540                 newflags = calc_vm_prot_bits(prot, new_vma_pkey);
 541                 newflags |= (vma->vm_flags & ~mask_off_old_flags);
 542
 543                 /* newflags >> 4 shift VM_MAY% in place of VM_% */
 544                 if ((newflags & ~(newflags >> 4)) & (VM_READ | VM_WRITE | VM_EXEC)) {
 545                         error = -EACCES;
 546                         goto out;
 547                 }
 548
 549                 error = security_file_mprotect(vma, reqprot, prot);
 550                 if (error)
 551                         goto out;
 552
 553                 tmp = vma->vm_end;
 554                 if (tmp > end)
 555                         tmp = end;
 556                 error = mprotect_fixup(vma, &prev, nstart, tmp, newflags);
 557                 if (error)
 558                         goto out;
 559                 nstart = tmp;
 560
 561                 if (nstart < prev->vm_end)
 562                         nstart = prev->vm_end;
 563                 if (nstart >= end)
 564                         goto out;
 565
 566                 vma = prev->vm_next;
 567                 if (!vma || vma->vm_start != nstart) {
 568                         error = -ENOMEM;
 569                         goto out;
 570                 }
 571                 prot = reqprot;
 572         }
 573 out:
 574         up_write(&current->mm->mmap_sem);
 575         return error;
 576 }
 577
 578 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
 579                 unsigned long, prot)
 580 {
 581         return do_mprotect_pkey(start, len, prot, -1);
 582 }
 583
 584 #ifdef CONFIG_ARCH_HAS_PKEYS
 585
 586 SYSCALL_DEFINE4(pkey_mprotect, unsigned long, start, size_t, len,
 587                 unsigned long, prot, int, pkey)
 588 {
 589         return do_mprotect_pkey(start, len, prot, pkey);
 590 }
 591
 592 SYSCALL_DEFINE2(pkey_alloc, unsigned long, flags, unsigned long, init_val)
 593 {
 594         int pkey;
 595         int ret;
 596
 597         /* No flags supported yet. */
 598         if (flags)
 599                 return -EINVAL;
 600         /* check for unsupported init values */
 601         if (init_val & ~PKEY_ACCESS_MASK)
 602                 return -EINVAL;
 603
 604         down_write(&current->mm->mmap_sem);
 605         pkey = mm_pkey_alloc(current->mm);
 606
 607         ret = -ENOSPC;
 608         if (pkey == -1)
 609                 goto out;
 610
 611         ret = arch_set_user_pkey_access(current, pkey, init_val);
 612         if (ret) {
 613                 mm_pkey_free(current->mm, pkey);
 614                 goto out;
 615         }
 616         ret = pkey;
 617 out:
 618         up_write(&current->mm->mmap_sem);
 619         return ret;
 620 }
 621
 622 SYSCALL_DEFINE1(pkey_free, int, pkey)
 623 {
 624         int ret;
 625
 626         down_write(&current->mm->mmap_sem);
 627         ret = mm_pkey_free(current->mm, pkey);
 628         up_write(&current->mm->mmap_sem);
 629
 630         /*
 631          * We could provie warnings or errors if any VMA still
 632          * has the pkey set here.
 633          */
 634         return ret;
 635 }
 636
 637 #endif /* CONFIG_ARCH_HAS_PKEYS */