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