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