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/pagewalk.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 <linux/pgtable.h>
  32 #include <linux/sched/sysctl.h>
  33 #include <asm/cacheflush.h>
  34 #include <asm/mmu_context.h>
  35 #include <asm/tlbflush.h>
  36 #include <asm/tlb.h>
  37
  38 #include "internal.h"
  39
  40 static unsigned long change_pte_range(struct mmu_gather *tlb,
  41                 struct vm_area_struct *vma, pmd_t *pmd, unsigned long addr,
  42                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
  43 {
  44         pte_t *pte, oldpte;
  45         spinlock_t *ptl;
  46         unsigned long pages = 0;
  47         int target_node = NUMA_NO_NODE;
  48         bool dirty_accountable = cp_flags & MM_CP_DIRTY_ACCT;
  49         bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
  50         bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
  51         bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
  52
  53         tlb_change_page_size(tlb, PAGE_SIZE);
  54
  55         /*
  56          * Can be called with only the mmap_lock for reading by
  57          * prot_numa so we must check the pmd isn't constantly
  58          * changing from under us from pmd_none to pmd_trans_huge
  59          * and/or the other way around.
  60          */
  61         if (pmd_trans_unstable(pmd))
  62                 return 0;
  63
  64         /*
  65          * The pmd points to a regular pte so the pmd can't change
  66          * from under us even if the mmap_lock is only hold for
  67          * reading.
  68          */
  69         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
  70
  71         /* Get target node for single threaded private VMAs */
  72         if (prot_numa && !(vma->vm_flags & VM_SHARED) &&
  73             atomic_read(&vma->vm_mm->mm_users) == 1)
  74                 target_node = numa_node_id();
  75
  76         flush_tlb_batched_pending(vma->vm_mm);
  77         arch_enter_lazy_mmu_mode();
  78         do {
  79                 oldpte = *pte;
  80                 if (pte_present(oldpte)) {
  81                         pte_t ptent;
  82                         bool preserve_write = prot_numa && pte_write(oldpte);
  83
  84                         /*
  85                          * Avoid trapping faults against the zero or KSM
  86                          * pages. See similar comment in change_huge_pmd.
  87                          */
  88                         if (prot_numa) {
  89                                 struct page *page;
  90                                 int nid;
  91
  92                                 /* Avoid TLB flush if possible */
  93                                 if (pte_protnone(oldpte))
  94                                         continue;
  95
  96                                 page = vm_normal_page(vma, addr, oldpte);
  97                                 if (!page || PageKsm(page))
  98                                         continue;
  99
 100                                 /* Also skip shared copy-on-write pages */
 101                                 if (is_cow_mapping(vma->vm_flags) &&
 102                                     page_count(page) != 1)
 103                                         continue;
 104
 105                                 /*
 106                                  * While migration can move some dirty pages,
 107                                  * it cannot move them all from MIGRATE_ASYNC
 108                                  * context.
 109                                  */
 110                                 if (page_is_file_lru(page) && PageDirty(page))
 111                                         continue;
 112
 113                                 /*
 114                                  * Don't mess with PTEs if page is already on the node
 115                                  * a single-threaded process is running on.
 116                                  */
 117                                 nid = page_to_nid(page);
 118                                 if (target_node == nid)
 119                                         continue;
 120
 121                                 /*
 122                                  * Skip scanning top tier node if normal numa
 123                                  * balancing is disabled
 124                                  */
 125                                 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_NORMAL) &&
 126                                     node_is_toptier(nid))
 127                                         continue;
 128                         }
 129
 130                         oldpte = ptep_modify_prot_start(vma, addr, pte);
 131                         ptent = pte_modify(oldpte, newprot);
 132                         if (preserve_write)
 133                                 ptent = pte_mk_savedwrite(ptent);
 134
 135                         if (uffd_wp) {
 136                                 ptent = pte_wrprotect(ptent);
 137                                 ptent = pte_mkuffd_wp(ptent);
 138                         } else if (uffd_wp_resolve) {
 139                                 /*
 140                                  * Leave the write bit to be handled
 141                                  * by PF interrupt handler, then
 142                                  * things like COW could be properly
 143                                  * handled.
 144                                  */
 145                                 ptent = pte_clear_uffd_wp(ptent);
 146                         }
 147
 148                         /* Avoid taking write faults for known dirty pages */
 149                         if (dirty_accountable && pte_dirty(ptent) &&
 150                                         (pte_soft_dirty(ptent) ||
 151                                          !(vma->vm_flags & VM_SOFTDIRTY))) {
 152                                 ptent = pte_mkwrite(ptent);
 153                         }
 154                         ptep_modify_prot_commit(vma, addr, pte, oldpte, ptent);
 155                         tlb_flush_pte_range(tlb, addr, PAGE_SIZE);
 156                         pages++;
 157                 } else if (is_swap_pte(oldpte)) {
 158                         swp_entry_t entry = pte_to_swp_entry(oldpte);
 159                         struct page *page = pfn_swap_entry_to_page(entry);
 160                         pte_t newpte;
 161
 162                         if (is_writable_migration_entry(entry)) {
 163                                 /*
 164                                  * A protection check is difficult so
 165                                  * just be safe and disable write
 166                                  */
 167                                 if (PageAnon(page))
 168                                         entry = make_readable_exclusive_migration_entry(
 169                                                              swp_offset(entry));
 170                                 else
 171                                         entry = make_readable_migration_entry(swp_offset(entry));
 172                                 newpte = swp_entry_to_pte(entry);
 173                                 if (pte_swp_soft_dirty(oldpte))
 174                                         newpte = pte_swp_mksoft_dirty(newpte);
 175                                 if (pte_swp_uffd_wp(oldpte))
 176                                         newpte = pte_swp_mkuffd_wp(newpte);
 177                         } else if (is_writable_device_private_entry(entry)) {
 178                                 /*
 179                                  * We do not preserve soft-dirtiness. See
 180                                  * copy_one_pte() for explanation.
 181                                  */
 182                                 entry = make_readable_device_private_entry(
 183                                                         swp_offset(entry));
 184                                 newpte = swp_entry_to_pte(entry);
 185                                 if (pte_swp_uffd_wp(oldpte))
 186                                         newpte = pte_swp_mkuffd_wp(newpte);
 187                         } else if (is_writable_device_exclusive_entry(entry)) {
 188                                 entry = make_readable_device_exclusive_entry(
 189                                                         swp_offset(entry));
 190                                 newpte = swp_entry_to_pte(entry);
 191                                 if (pte_swp_soft_dirty(oldpte))
 192                                         newpte = pte_swp_mksoft_dirty(newpte);
 193                                 if (pte_swp_uffd_wp(oldpte))
 194                                         newpte = pte_swp_mkuffd_wp(newpte);
 195                         } else {
 196                                 newpte = oldpte;
 197                         }
 198
 199                         if (uffd_wp)
 200                                 newpte = pte_swp_mkuffd_wp(newpte);
 201                         else if (uffd_wp_resolve)
 202                                 newpte = pte_swp_clear_uffd_wp(newpte);
 203
 204                         if (!pte_same(oldpte, newpte)) {
 205                                 set_pte_at(vma->vm_mm, addr, pte, newpte);
 206                                 pages++;
 207                         }
 208                 }
 209         } while (pte++, addr += PAGE_SIZE, addr != end);
 210         arch_leave_lazy_mmu_mode();
 211         pte_unmap_unlock(pte - 1, ptl);
 212
 213         return pages;
 214 }
 215
 216 /*
 217  * Used when setting automatic NUMA hinting protection where it is
 218  * critical that a numa hinting PMD is not confused with a bad PMD.
 219  */
 220 static inline int pmd_none_or_clear_bad_unless_trans_huge(pmd_t *pmd)
 221 {
 222         pmd_t pmdval = pmd_read_atomic(pmd);
 223
 224         /* See pmd_none_or_trans_huge_or_clear_bad for info on barrier */
 225 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 226         barrier();
 227 #endif
 228
 229         if (pmd_none(pmdval))
 230                 return 1;
 231         if (pmd_trans_huge(pmdval))
 232                 return 0;
 233         if (unlikely(pmd_bad(pmdval))) {
 234                 pmd_clear_bad(pmd);
 235                 return 1;
 236         }
 237
 238         return 0;
 239 }
 240
 241 static inline unsigned long change_pmd_range(struct mmu_gather *tlb,
 242                 struct vm_area_struct *vma, pud_t *pud, unsigned long addr,
 243                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
 244 {
 245         pmd_t *pmd;
 246         unsigned long next;
 247         unsigned long pages = 0;
 248         unsigned long nr_huge_updates = 0;
 249         struct mmu_notifier_range range;
 250
 251         range.start = 0;
 252
 253         pmd = pmd_offset(pud, addr);
 254         do {
 255                 unsigned long this_pages;
 256
 257                 next = pmd_addr_end(addr, end);
 258
 259                 /*
 260                  * Automatic NUMA balancing walks the tables with mmap_lock
 261                  * held for read. It's possible a parallel update to occur
 262                  * between pmd_trans_huge() and a pmd_none_or_clear_bad()
 263                  * check leading to a false positive and clearing.
 264                  * Hence, it's necessary to atomically read the PMD value
 265                  * for all the checks.
 266                  */
 267                 if (!is_swap_pmd(*pmd) && !pmd_devmap(*pmd) &&
 268                      pmd_none_or_clear_bad_unless_trans_huge(pmd))
 269                         goto next;
 270
 271                 /* invoke the mmu notifier if the pmd is populated */
 272                 if (!range.start) {
 273                         mmu_notifier_range_init(&range,
 274                                 MMU_NOTIFY_PROTECTION_VMA, 0,
 275                                 vma, vma->vm_mm, addr, end);
 276                         mmu_notifier_invalidate_range_start(&range);
 277                 }
 278
 279                 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
 280                         if (next - addr != HPAGE_PMD_SIZE) {
 281                                 __split_huge_pmd(vma, pmd, addr, false, NULL);
 282                         } else {
 283                                 /*
 284                                  * change_huge_pmd() does not defer TLB flushes,
 285                                  * so no need to propagate the tlb argument.
 286                                  */
 287                                 int nr_ptes = change_huge_pmd(tlb, vma, pmd,
 288                                                 addr, newprot, cp_flags);
 289
 290                                 if (nr_ptes) {
 291                                         if (nr_ptes == HPAGE_PMD_NR) {
 292                                                 pages += HPAGE_PMD_NR;
 293                                                 nr_huge_updates++;
 294                                         }
 295
 296                                         /* huge pmd was handled */
 297                                         goto next;
 298                                 }
 299                         }
 300                         /* fall through, the trans huge pmd just split */
 301                 }
 302                 this_pages = change_pte_range(tlb, vma, pmd, addr, next,
 303                                               newprot, cp_flags);
 304                 pages += this_pages;
 305 next:
 306                 cond_resched();
 307         } while (pmd++, addr = next, addr != end);
 308
 309         if (range.start)
 310                 mmu_notifier_invalidate_range_end(&range);
 311
 312         if (nr_huge_updates)
 313                 count_vm_numa_events(NUMA_HUGE_PTE_UPDATES, nr_huge_updates);
 314         return pages;
 315 }
 316
 317 static inline unsigned long change_pud_range(struct mmu_gather *tlb,
 318                 struct vm_area_struct *vma, p4d_t *p4d, unsigned long addr,
 319                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
 320 {
 321         pud_t *pud;
 322         unsigned long next;
 323         unsigned long pages = 0;
 324
 325         pud = pud_offset(p4d, addr);
 326         do {
 327                 next = pud_addr_end(addr, end);
 328                 if (pud_none_or_clear_bad(pud))
 329                         continue;
 330                 pages += change_pmd_range(tlb, vma, pud, addr, next, newprot,
 331                                           cp_flags);
 332         } while (pud++, addr = next, addr != end);
 333
 334         return pages;
 335 }
 336
 337 static inline unsigned long change_p4d_range(struct mmu_gather *tlb,
 338                 struct vm_area_struct *vma, pgd_t *pgd, unsigned long addr,
 339                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
 340 {
 341         p4d_t *p4d;
 342         unsigned long next;
 343         unsigned long pages = 0;
 344
 345         p4d = p4d_offset(pgd, addr);
 346         do {
 347                 next = p4d_addr_end(addr, end);
 348                 if (p4d_none_or_clear_bad(p4d))
 349                         continue;
 350                 pages += change_pud_range(tlb, vma, p4d, addr, next, newprot,
 351                                           cp_flags);
 352         } while (p4d++, addr = next, addr != end);
 353
 354         return pages;
 355 }
 356
 357 static unsigned long change_protection_range(struct mmu_gather *tlb,
 358                 struct vm_area_struct *vma, unsigned long addr,
 359                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
 360 {
 361         struct mm_struct *mm = vma->vm_mm;
 362         pgd_t *pgd;
 363         unsigned long next;
 364         unsigned long pages = 0;
 365
 366         BUG_ON(addr >= end);
 367         pgd = pgd_offset(mm, addr);
 368         tlb_start_vma(tlb, vma);
 369         do {
 370                 next = pgd_addr_end(addr, end);
 371                 if (pgd_none_or_clear_bad(pgd))
 372                         continue;
 373                 pages += change_p4d_range(tlb, vma, pgd, addr, next, newprot,
 374                                           cp_flags);
 375         } while (pgd++, addr = next, addr != end);
 376
 377         tlb_end_vma(tlb, vma);
 378
 379         return pages;
 380 }
 381
 382 unsigned long change_protection(struct mmu_gather *tlb,
 383                        struct vm_area_struct *vma, unsigned long start,
 384                        unsigned long end, pgprot_t newprot,
 385                        unsigned long cp_flags)
 386 {
 387         unsigned long pages;
 388
 389         BUG_ON((cp_flags & MM_CP_UFFD_WP_ALL) == MM_CP_UFFD_WP_ALL);
 390
 391         if (is_vm_hugetlb_page(vma))
 392                 pages = hugetlb_change_protection(vma, start, end, newprot);
 393         else
 394                 pages = change_protection_range(tlb, vma, start, end, newprot,
 395                                                 cp_flags);
 396
 397         return pages;
 398 }
 399
 400 static int prot_none_pte_entry(pte_t *pte, unsigned long addr,
 401                                unsigned long next, struct mm_walk *walk)
 402 {
 403         return pfn_modify_allowed(pte_pfn(*pte), *(pgprot_t *)(walk->private)) ?
 404                 0 : -EACCES;
 405 }
 406
 407 static int prot_none_hugetlb_entry(pte_t *pte, unsigned long hmask,
 408                                    unsigned long addr, unsigned long next,
 409                                    struct mm_walk *walk)
 410 {
 411         return pfn_modify_allowed(pte_pfn(*pte), *(pgprot_t *)(walk->private)) ?
 412                 0 : -EACCES;
 413 }
 414
 415 static int prot_none_test(unsigned long addr, unsigned long next,
 416                           struct mm_walk *walk)
 417 {
 418         return 0;
 419 }
 420
 421 static const struct mm_walk_ops prot_none_walk_ops = {
 422         .pte_entry              = prot_none_pte_entry,
 423         .hugetlb_entry          = prot_none_hugetlb_entry,
 424         .test_walk              = prot_none_test,
 425 };
 426
 427 int
 428 mprotect_fixup(struct mmu_gather *tlb, struct vm_area_struct *vma,
 429                struct vm_area_struct **pprev, unsigned long start,
 430                unsigned long end, unsigned long newflags)
 431 {
 432         struct mm_struct *mm = vma->vm_mm;
 433         unsigned long oldflags = vma->vm_flags;
 434         long nrpages = (end - start) >> PAGE_SHIFT;
 435         unsigned long charged = 0;
 436         pgoff_t pgoff;
 437         int error;
 438         int dirty_accountable = 0;
 439
 440         if (newflags == oldflags) {
 441                 *pprev = vma;
 442                 return 0;
 443         }
 444
 445         /*
 446          * Do PROT_NONE PFN permission checks here when we can still
 447          * bail out without undoing a lot of state. This is a rather
 448          * uncommon case, so doesn't need to be very optimized.
 449          */
 450         if (arch_has_pfn_modify_check() &&
 451             (vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 452             (newflags & VM_ACCESS_FLAGS) == 0) {
 453                 pgprot_t new_pgprot = vm_get_page_prot(newflags);
 454
 455                 error = walk_page_range(current->mm, start, end,
 456                                 &prot_none_walk_ops, &new_pgprot);
 457                 if (error)
 458                         return error;
 459         }
 460
 461         /*
 462          * If we make a private mapping writable we increase our commit;
 463          * but (without finer accounting) cannot reduce our commit if we
 464          * make it unwritable again. hugetlb mapping were accounted for
 465          * even if read-only so there is no need to account for them here
 466          */
 467         if (newflags & VM_WRITE) {
 468                 /* Check space limits when area turns into data. */
 469                 if (!may_expand_vm(mm, newflags, nrpages) &&
 470                                 may_expand_vm(mm, oldflags, nrpages))
 471                         return -ENOMEM;
 472                 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
 473                                                 VM_SHARED|VM_NORESERVE))) {
 474                         charged = nrpages;
 475                         if (security_vm_enough_memory_mm(mm, charged))
 476                                 return -ENOMEM;
 477                         newflags |= VM_ACCOUNT;
 478                 }
 479         }
 480
 481         /*
 482          * First try to merge with previous and/or next vma.
 483          */
 484         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 485         *pprev = vma_merge(mm, *pprev, start, end, newflags,
 486                            vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
 487                            vma->vm_userfaultfd_ctx, anon_vma_name(vma));
 488         if (*pprev) {
 489                 vma = *pprev;
 490                 VM_WARN_ON((vma->vm_flags ^ newflags) & ~VM_SOFTDIRTY);
 491                 goto success;
 492         }
 493
 494         *pprev = vma;
 495
 496         if (start != vma->vm_start) {
 497                 error = split_vma(mm, vma, start, 1);
 498                 if (error)
 499                         goto fail;
 500         }
 501
 502         if (end != vma->vm_end) {
 503                 error = split_vma(mm, vma, end, 0);
 504                 if (error)
 505                         goto fail;
 506         }
 507
 508 success:
 509         /*
 510          * vm_flags and vm_page_prot are protected by the mmap_lock
 511          * held in write mode.
 512          */
 513         vma->vm_flags = newflags;
 514         dirty_accountable = vma_wants_writenotify(vma, vma->vm_page_prot);
 515         vma_set_page_prot(vma);
 516
 517         change_protection(tlb, vma, start, end, vma->vm_page_prot,
 518                           dirty_accountable ? MM_CP_DIRTY_ACCT : 0);
 519
 520         /*
 521          * Private VM_LOCKED VMA becoming writable: trigger COW to avoid major
 522          * fault on access.
 523          */
 524         if ((oldflags & (VM_WRITE | VM_SHARED | VM_LOCKED)) == VM_LOCKED &&
 525                         (newflags & VM_WRITE)) {
 526                 populate_vma_page_range(vma, start, end, NULL);
 527         }
 528
 529         vm_stat_account(mm, oldflags, -nrpages);
 530         vm_stat_account(mm, newflags, nrpages);
 531         perf_event_mmap(vma);
 532         return 0;
 533
 534 fail:
 535         vm_unacct_memory(charged);
 536         return error;
 537 }
 538
 539 /*
 540  * pkey==-1 when doing a legacy mprotect()
 541  */
 542 static int do_mprotect_pkey(unsigned long start, size_t len,
 543                 unsigned long prot, int pkey)
 544 {
 545         unsigned long nstart, end, tmp, reqprot;
 546         struct vm_area_struct *vma, *prev;
 547         int error = -EINVAL;
 548         const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
 549         const bool rier = (current->personality & READ_IMPLIES_EXEC) &&
 550                                 (prot & PROT_READ);
 551         struct mmu_gather tlb;
 552
 553         start = untagged_addr(start);
 554
 555         prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
 556         if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
 557                 return -EINVAL;
 558
 559         if (start & ~PAGE_MASK)
 560                 return -EINVAL;
 561         if (!len)
 562                 return 0;
 563         len = PAGE_ALIGN(len);
 564         end = start + len;
 565         if (end <= start)
 566                 return -ENOMEM;
 567         if (!arch_validate_prot(prot, start))
 568                 return -EINVAL;
 569
 570         reqprot = prot;
 571
 572         if (mmap_write_lock_killable(current->mm))
 573                 return -EINTR;
 574
 575         /*
 576          * If userspace did not allocate the pkey, do not let
 577          * them use it here.
 578          */
 579         error = -EINVAL;
 580         if ((pkey != -1) && !mm_pkey_is_allocated(current->mm, pkey))
 581                 goto out;
 582
 583         vma = find_vma(current->mm, start);
 584         error = -ENOMEM;
 585         if (!vma)
 586                 goto out;
 587
 588         if (unlikely(grows & PROT_GROWSDOWN)) {
 589                 if (vma->vm_start >= end)
 590                         goto out;
 591                 start = vma->vm_start;
 592                 error = -EINVAL;
 593                 if (!(vma->vm_flags & VM_GROWSDOWN))
 594                         goto out;
 595         } else {
 596                 if (vma->vm_start > start)
 597                         goto out;
 598                 if (unlikely(grows & PROT_GROWSUP)) {
 599                         end = vma->vm_end;
 600                         error = -EINVAL;
 601                         if (!(vma->vm_flags & VM_GROWSUP))
 602                                 goto out;
 603                 }
 604         }
 605
 606         if (start > vma->vm_start)
 607                 prev = vma;
 608         else
 609                 prev = vma->vm_prev;
 610
 611         tlb_gather_mmu(&tlb, current->mm);
 612         for (nstart = start ; ; ) {
 613                 unsigned long mask_off_old_flags;
 614                 unsigned long newflags;
 615                 int new_vma_pkey;
 616
 617                 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
 618
 619                 /* Does the application expect PROT_READ to imply PROT_EXEC */
 620                 if (rier && (vma->vm_flags & VM_MAYEXEC))
 621                         prot |= PROT_EXEC;
 622
 623                 /*
 624                  * Each mprotect() call explicitly passes r/w/x permissions.
 625                  * If a permission is not passed to mprotect(), it must be
 626                  * cleared from the VMA.
 627                  */
 628                 mask_off_old_flags = VM_READ | VM_WRITE | VM_EXEC |
 629                                         VM_FLAGS_CLEAR;
 630
 631                 new_vma_pkey = arch_override_mprotect_pkey(vma, prot, pkey);
 632                 newflags = calc_vm_prot_bits(prot, new_vma_pkey);
 633                 newflags |= (vma->vm_flags & ~mask_off_old_flags);
 634
 635                 /* newflags >> 4 shift VM_MAY% in place of VM_% */
 636                 if ((newflags & ~(newflags >> 4)) & VM_ACCESS_FLAGS) {
 637                         error = -EACCES;
 638                         break;
 639                 }
 640
 641                 /* Allow architectures to sanity-check the new flags */
 642                 if (!arch_validate_flags(newflags)) {
 643                         error = -EINVAL;
 644                         break;
 645                 }
 646
 647                 error = security_file_mprotect(vma, reqprot, prot);
 648                 if (error)
 649                         break;
 650
 651                 tmp = vma->vm_end;
 652                 if (tmp > end)
 653                         tmp = end;
 654
 655                 if (vma->vm_ops && vma->vm_ops->mprotect) {
 656                         error = vma->vm_ops->mprotect(vma, nstart, tmp, newflags);
 657                         if (error)
 658                                 break;
 659                 }
 660
 661                 error = mprotect_fixup(&tlb, vma, &prev, nstart, tmp, newflags);
 662                 if (error)
 663                         break;
 664
 665                 nstart = tmp;
 666
 667                 if (nstart < prev->vm_end)
 668                         nstart = prev->vm_end;
 669                 if (nstart >= end)
 670                         break;
 671
 672                 vma = prev->vm_next;
 673                 if (!vma || vma->vm_start != nstart) {
 674                         error = -ENOMEM;
 675                         break;
 676                 }
 677                 prot = reqprot;
 678         }
 679         tlb_finish_mmu(&tlb);
 680 out:
 681         mmap_write_unlock(current->mm);
 682         return error;
 683 }
 684
 685 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
 686                 unsigned long, prot)
 687 {
 688         return do_mprotect_pkey(start, len, prot, -1);
 689 }
 690
 691 #ifdef CONFIG_ARCH_HAS_PKEYS
 692
 693 SYSCALL_DEFINE4(pkey_mprotect, unsigned long, start, size_t, len,
 694                 unsigned long, prot, int, pkey)
 695 {
 696         return do_mprotect_pkey(start, len, prot, pkey);
 697 }
 698
 699 SYSCALL_DEFINE2(pkey_alloc, unsigned long, flags, unsigned long, init_val)
 700 {
 701         int pkey;
 702         int ret;
 703
 704         /* No flags supported yet. */
 705         if (flags)
 706                 return -EINVAL;
 707         /* check for unsupported init values */
 708         if (init_val & ~PKEY_ACCESS_MASK)
 709                 return -EINVAL;
 710
 711         mmap_write_lock(current->mm);
 712         pkey = mm_pkey_alloc(current->mm);
 713
 714         ret = -ENOSPC;
 715         if (pkey == -1)
 716                 goto out;
 717
 718         ret = arch_set_user_pkey_access(current, pkey, init_val);
 719         if (ret) {
 720                 mm_pkey_free(current->mm, pkey);
 721                 goto out;
 722         }
 723         ret = pkey;
 724 out:
 725         mmap_write_unlock(current->mm);
 726         return ret;
 727 }
 728
 729 SYSCALL_DEFINE1(pkey_free, int, pkey)
 730 {
 731         int ret;
 732
 733         mmap_write_lock(current->mm);
 734         ret = mm_pkey_free(current->mm, pkey);
 735         mmap_write_unlock(current->mm);
 736
 737         /*
 738          * We could provide warnings or errors if any VMA still
 739          * has the pkey set here.
 740          */
 741         return ret;
 742 }
 743
 744 #endif /* CONFIG_ARCH_HAS_PKEYS */