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         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                         oldpte = ptep_modify_prot_start(vma, addr, pte);
 114                         ptent = pte_modify(oldpte, 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, oldpte, 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,
 189                                 MMU_NOTIFY_PROTECTION_VMA, 0,
 190                                 vma, vma->vm_mm, addr, end);
 191                         mmu_notifier_invalidate_range_start(&range);
 192                 }
 193
 194                 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
 195                         if (next - addr != HPAGE_PMD_SIZE) {
 196                                 __split_huge_pmd(vma, pmd, addr, false, NULL);
 197                         } else {
 198                                 int nr_ptes = change_huge_pmd(vma, pmd, addr,
 199                                                 newprot, prot_numa);
 200
 201                                 if (nr_ptes) {
 202                                         if (nr_ptes == HPAGE_PMD_NR) {
 203                                                 pages += HPAGE_PMD_NR;
 204                                                 nr_huge_updates++;
 205                                         }
 206
 207                                         /* huge pmd was handled */
 208                                         goto next;
 209                                 }
 210                         }
 211                         /* fall through, the trans huge pmd just split */
 212                 }
 213                 this_pages = change_pte_range(vma, pmd, addr, next, newprot,
 214                                  dirty_accountable, prot_numa);
 215                 pages += this_pages;
 216 next:
 217                 cond_resched();
 218         } while (pmd++, addr = next, addr != end);
 219
 220         if (range.start)
 221                 mmu_notifier_invalidate_range_end(&range);
 222
 223         if (nr_huge_updates)
 224                 count_vm_numa_events(NUMA_HUGE_PTE_UPDATES, nr_huge_updates);
 225         return pages;
 226 }
 227
 228 static inline unsigned long change_pud_range(struct vm_area_struct *vma,
 229                 p4d_t *p4d, unsigned long addr, unsigned long end,
 230                 pgprot_t newprot, int dirty_accountable, int prot_numa)
 231 {
 232         pud_t *pud;
 233         unsigned long next;
 234         unsigned long pages = 0;
 235
 236         pud = pud_offset(p4d, addr);
 237         do {
 238                 next = pud_addr_end(addr, end);
 239                 if (pud_none_or_clear_bad(pud))
 240                         continue;
 241                 pages += change_pmd_range(vma, pud, addr, next, newprot,
 242                                  dirty_accountable, prot_numa);
 243         } while (pud++, addr = next, addr != end);
 244
 245         return pages;
 246 }
 247
 248 static inline unsigned long change_p4d_range(struct vm_area_struct *vma,
 249                 pgd_t *pgd, unsigned long addr, unsigned long end,
 250                 pgprot_t newprot, int dirty_accountable, int prot_numa)
 251 {
 252         p4d_t *p4d;
 253         unsigned long next;
 254         unsigned long pages = 0;
 255
 256         p4d = p4d_offset(pgd, addr);
 257         do {
 258                 next = p4d_addr_end(addr, end);
 259                 if (p4d_none_or_clear_bad(p4d))
 260                         continue;
 261                 pages += change_pud_range(vma, p4d, addr, next, newprot,
 262                                  dirty_accountable, prot_numa);
 263         } while (p4d++, addr = next, addr != end);
 264
 265         return pages;
 266 }
 267
 268 static unsigned long change_protection_range(struct vm_area_struct *vma,
 269                 unsigned long addr, unsigned long end, pgprot_t newprot,
 270                 int dirty_accountable, int prot_numa)
 271 {
 272         struct mm_struct *mm = vma->vm_mm;
 273         pgd_t *pgd;
 274         unsigned long next;
 275         unsigned long start = addr;
 276         unsigned long pages = 0;
 277
 278         BUG_ON(addr >= end);
 279         pgd = pgd_offset(mm, addr);
 280         flush_cache_range(vma, addr, end);
 281         inc_tlb_flush_pending(mm);
 282         do {
 283                 next = pgd_addr_end(addr, end);
 284                 if (pgd_none_or_clear_bad(pgd))
 285                         continue;
 286                 pages += change_p4d_range(vma, pgd, addr, next, newprot,
 287                                  dirty_accountable, prot_numa);
 288         } while (pgd++, addr = next, addr != end);
 289
 290         /* Only flush the TLB if we actually modified any entries: */
 291         if (pages)
 292                 flush_tlb_range(vma, start, end);
 293         dec_tlb_flush_pending(mm);
 294
 295         return pages;
 296 }
 297
 298 unsigned long change_protection(struct vm_area_struct *vma, unsigned long start,
 299                        unsigned long end, pgprot_t newprot,
 300                        int dirty_accountable, int prot_numa)
 301 {
 302         unsigned long pages;
 303
 304         if (is_vm_hugetlb_page(vma))
 305                 pages = hugetlb_change_protection(vma, start, end, newprot);
 306         else
 307                 pages = change_protection_range(vma, start, end, newprot, dirty_accountable, prot_numa);
 308
 309         return pages;
 310 }
 311
 312 static int prot_none_pte_entry(pte_t *pte, unsigned long addr,
 313                                unsigned long next, struct mm_walk *walk)
 314 {
 315         return pfn_modify_allowed(pte_pfn(*pte), *(pgprot_t *)(walk->private)) ?
 316                 0 : -EACCES;
 317 }
 318
 319 static int prot_none_hugetlb_entry(pte_t *pte, unsigned long hmask,
 320                                    unsigned long addr, unsigned long next,
 321                                    struct mm_walk *walk)
 322 {
 323         return pfn_modify_allowed(pte_pfn(*pte), *(pgprot_t *)(walk->private)) ?
 324                 0 : -EACCES;
 325 }
 326
 327 static int prot_none_test(unsigned long addr, unsigned long next,
 328                           struct mm_walk *walk)
 329 {
 330         return 0;
 331 }
 332
 333 static int prot_none_walk(struct vm_area_struct *vma, unsigned long start,
 334                            unsigned long end, unsigned long newflags)
 335 {
 336         pgprot_t new_pgprot = vm_get_page_prot(newflags);
 337         struct mm_walk prot_none_walk = {
 338                 .pte_entry = prot_none_pte_entry,
 339                 .hugetlb_entry = prot_none_hugetlb_entry,
 340                 .test_walk = prot_none_test,
 341                 .mm = current->mm,
 342                 .private = &new_pgprot,
 343         };
 344
 345         return walk_page_range(start, end, &prot_none_walk);
 346 }
 347
 348 int
 349 mprotect_fixup(struct vm_area_struct *vma, struct vm_area_struct **pprev,
 350         unsigned long start, unsigned long end, unsigned long newflags)
 351 {
 352         struct mm_struct *mm = vma->vm_mm;
 353         unsigned long oldflags = vma->vm_flags;
 354         long nrpages = (end - start) >> PAGE_SHIFT;
 355         unsigned long charged = 0;
 356         pgoff_t pgoff;
 357         int error;
 358         int dirty_accountable = 0;
 359
 360         if (newflags == oldflags) {
 361                 *pprev = vma;
 362                 return 0;
 363         }
 364
 365         /*
 366          * Do PROT_NONE PFN permission checks here when we can still
 367          * bail out without undoing a lot of state. This is a rather
 368          * uncommon case, so doesn't need to be very optimized.
 369          */
 370         if (arch_has_pfn_modify_check() &&
 371             (vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 372             (newflags & (VM_READ|VM_WRITE|VM_EXEC)) == 0) {
 373                 error = prot_none_walk(vma, start, end, newflags);
 374                 if (error)
 375                         return error;
 376         }
 377
 378         /*
 379          * If we make a private mapping writable we increase our commit;
 380          * but (without finer accounting) cannot reduce our commit if we
 381          * make it unwritable again. hugetlb mapping were accounted for
 382          * even if read-only so there is no need to account for them here
 383          */
 384         if (newflags & VM_WRITE) {
 385                 /* Check space limits when area turns into data. */
 386                 if (!may_expand_vm(mm, newflags, nrpages) &&
 387                                 may_expand_vm(mm, oldflags, nrpages))
 388                         return -ENOMEM;
 389                 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
 390                                                 VM_SHARED|VM_NORESERVE))) {
 391                         charged = nrpages;
 392                         if (security_vm_enough_memory_mm(mm, charged))
 393                                 return -ENOMEM;
 394                         newflags |= VM_ACCOUNT;
 395                 }
 396         }
 397
 398         /*
 399          * First try to merge with previous and/or next vma.
 400          */
 401         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 402         *pprev = vma_merge(mm, *pprev, start, end, newflags,
 403                            vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
 404                            vma->vm_userfaultfd_ctx);
 405         if (*pprev) {
 406                 vma = *pprev;
 407                 VM_WARN_ON((vma->vm_flags ^ newflags) & ~VM_SOFTDIRTY);
 408                 goto success;
 409         }
 410
 411         *pprev = vma;
 412
 413         if (start != vma->vm_start) {
 414                 error = split_vma(mm, vma, start, 1);
 415                 if (error)
 416                         goto fail;
 417         }
 418
 419         if (end != vma->vm_end) {
 420                 error = split_vma(mm, vma, end, 0);
 421                 if (error)
 422                         goto fail;
 423         }
 424
 425 success:
 426         /*
 427          * vm_flags and vm_page_prot are protected by the mmap_sem
 428          * held in write mode.
 429          */
 430         vma->vm_flags = newflags;
 431         dirty_accountable = vma_wants_writenotify(vma, vma->vm_page_prot);
 432         vma_set_page_prot(vma);
 433
 434         change_protection(vma, start, end, vma->vm_page_prot,
 435                           dirty_accountable, 0);
 436
 437         /*
 438          * Private VM_LOCKED VMA becoming writable: trigger COW to avoid major
 439          * fault on access.
 440          */
 441         if ((oldflags & (VM_WRITE | VM_SHARED | VM_LOCKED)) == VM_LOCKED &&
 442                         (newflags & VM_WRITE)) {
 443                 populate_vma_page_range(vma, start, end, NULL);
 444         }
 445
 446         vm_stat_account(mm, oldflags, -nrpages);
 447         vm_stat_account(mm, newflags, nrpages);
 448         perf_event_mmap(vma);
 449         return 0;
 450
 451 fail:
 452         vm_unacct_memory(charged);
 453         return error;
 454 }
 455
 456 /*
 457  * pkey==-1 when doing a legacy mprotect()
 458  */
 459 static int do_mprotect_pkey(unsigned long start, size_t len,
 460                 unsigned long prot, int pkey)
 461 {
 462         unsigned long nstart, end, tmp, reqprot;
 463         struct vm_area_struct *vma, *prev;
 464         int error = -EINVAL;
 465         const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
 466         const bool rier = (current->personality & READ_IMPLIES_EXEC) &&
 467                                 (prot & PROT_READ);
 468
 469         prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
 470         if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
 471                 return -EINVAL;
 472
 473         if (start & ~PAGE_MASK)
 474                 return -EINVAL;
 475         if (!len)
 476                 return 0;
 477         len = PAGE_ALIGN(len);
 478         end = start + len;
 479         if (end <= start)
 480                 return -ENOMEM;
 481         if (!arch_validate_prot(prot, start))
 482                 return -EINVAL;
 483
 484         reqprot = prot;
 485
 486         if (down_write_killable(&current->mm->mmap_sem))
 487                 return -EINTR;
 488
 489         /*
 490          * If userspace did not allocate the pkey, do not let
 491          * them use it here.
 492          */
 493         error = -EINVAL;
 494         if ((pkey != -1) && !mm_pkey_is_allocated(current->mm, pkey))
 495                 goto out;
 496
 497         vma = find_vma(current->mm, start);
 498         error = -ENOMEM;
 499         if (!vma)
 500                 goto out;
 501         prev = vma->vm_prev;
 502         if (unlikely(grows & PROT_GROWSDOWN)) {
 503                 if (vma->vm_start >= end)
 504                         goto out;
 505                 start = vma->vm_start;
 506                 error = -EINVAL;
 507                 if (!(vma->vm_flags & VM_GROWSDOWN))
 508                         goto out;
 509         } else {
 510                 if (vma->vm_start > start)
 511                         goto out;
 512                 if (unlikely(grows & PROT_GROWSUP)) {
 513                         end = vma->vm_end;
 514                         error = -EINVAL;
 515                         if (!(vma->vm_flags & VM_GROWSUP))
 516                                 goto out;
 517                 }
 518         }
 519         if (start > vma->vm_start)
 520                 prev = vma;
 521
 522         for (nstart = start ; ; ) {
 523                 unsigned long mask_off_old_flags;
 524                 unsigned long newflags;
 525                 int new_vma_pkey;
 526
 527                 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
 528
 529                 /* Does the application expect PROT_READ to imply PROT_EXEC */
 530                 if (rier && (vma->vm_flags & VM_MAYEXEC))
 531                         prot |= PROT_EXEC;
 532
 533                 /*
 534                  * Each mprotect() call explicitly passes r/w/x permissions.
 535                  * If a permission is not passed to mprotect(), it must be
 536                  * cleared from the VMA.
 537                  */
 538                 mask_off_old_flags = VM_READ | VM_WRITE | VM_EXEC |
 539                                         VM_FLAGS_CLEAR;
 540
 541                 new_vma_pkey = arch_override_mprotect_pkey(vma, prot, pkey);
 542                 newflags = calc_vm_prot_bits(prot, new_vma_pkey);
 543                 newflags |= (vma->vm_flags & ~mask_off_old_flags);
 544
 545                 /* newflags >> 4 shift VM_MAY% in place of VM_% */
 546                 if ((newflags & ~(newflags >> 4)) & (VM_READ | VM_WRITE | VM_EXEC)) {
 547                         error = -EACCES;
 548                         goto out;
 549                 }
 550
 551                 error = security_file_mprotect(vma, reqprot, prot);
 552                 if (error)
 553                         goto out;
 554
 555                 tmp = vma->vm_end;
 556                 if (tmp > end)
 557                         tmp = end;
 558                 error = mprotect_fixup(vma, &prev, nstart, tmp, newflags);
 559                 if (error)
 560                         goto out;
 561                 nstart = tmp;
 562
 563                 if (nstart < prev->vm_end)
 564                         nstart = prev->vm_end;
 565                 if (nstart >= end)
 566                         goto out;
 567
 568                 vma = prev->vm_next;
 569                 if (!vma || vma->vm_start != nstart) {
 570                         error = -ENOMEM;
 571                         goto out;
 572                 }
 573                 prot = reqprot;
 574         }
 575 out:
 576         up_write(&current->mm->mmap_sem);
 577         return error;
 578 }
 579
 580 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
 581                 unsigned long, prot)
 582 {
 583         return do_mprotect_pkey(start, len, prot, -1);
 584 }
 585
 586 #ifdef CONFIG_ARCH_HAS_PKEYS
 587
 588 SYSCALL_DEFINE4(pkey_mprotect, unsigned long, start, size_t, len,
 589                 unsigned long, prot, int, pkey)
 590 {
 591         return do_mprotect_pkey(start, len, prot, pkey);
 592 }
 593
 594 SYSCALL_DEFINE2(pkey_alloc, unsigned long, flags, unsigned long, init_val)
 595 {
 596         int pkey;
 597         int ret;
 598
 599         /* No flags supported yet. */
 600         if (flags)
 601                 return -EINVAL;
 602         /* check for unsupported init values */
 603         if (init_val & ~PKEY_ACCESS_MASK)
 604                 return -EINVAL;
 605
 606         down_write(&current->mm->mmap_sem);
 607         pkey = mm_pkey_alloc(current->mm);
 608
 609         ret = -ENOSPC;
 610         if (pkey == -1)
 611                 goto out;
 612
 613         ret = arch_set_user_pkey_access(current, pkey, init_val);
 614         if (ret) {
 615                 mm_pkey_free(current->mm, pkey);
 616                 goto out;
 617         }
 618         ret = pkey;
 619 out:
 620         up_write(&current->mm->mmap_sem);
 621         return ret;
 622 }
 623
 624 SYSCALL_DEFINE1(pkey_free, int, pkey)
 625 {
 626         int ret;
 627
 628         down_write(&current->mm->mmap_sem);
 629         ret = mm_pkey_free(current->mm, pkey);
 630         up_write(&current->mm->mmap_sem);
 631
 632         /*
 633          * We could provie warnings or errors if any VMA still
 634          * has the pkey set here.
 635          */
 636         return ret;
 637 }
 638
 639 #endif /* CONFIG_ARCH_HAS_PKEYS */