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