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