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 <linux/sched/sysctl.h>
  33 #include <linux/userfaultfd_k.h>
  34 #include <asm/cacheflush.h>
  35 #include <asm/mmu_context.h>
  36 #include <asm/tlbflush.h>
  37 #include <asm/tlb.h>
  38
  39 #include "internal.h"
  40
  41 static unsigned long change_pte_range(struct mmu_gather *tlb,
  42                 struct vm_area_struct *vma, pmd_t *pmd, unsigned long addr,
  43                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
  44 {
  45         pte_t *pte, oldpte;
  46         spinlock_t *ptl;
  47         unsigned long pages = 0;
  48         int target_node = NUMA_NO_NODE;
  49         bool dirty_accountable = cp_flags & MM_CP_DIRTY_ACCT;
  50         bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
  51         bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
  52         bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
  53
  54         tlb_change_page_size(tlb, PAGE_SIZE);
  55
  56         /*
  57          * Can be called with only the mmap_lock for reading by
  58          * prot_numa so we must check the pmd isn't constantly
  59          * changing from under us from pmd_none to pmd_trans_huge
  60          * and/or the other way around.
  61          */
  62         if (pmd_trans_unstable(pmd))
  63                 return 0;
  64
  65         /*
  66          * The pmd points to a regular pte so the pmd can't change
  67          * from under us even if the mmap_lock is only hold for
  68          * reading.
  69          */
  70         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
  71
  72         /* Get target node for single threaded private VMAs */
  73         if (prot_numa && !(vma->vm_flags & VM_SHARED) &&
  74             atomic_read(&vma->vm_mm->mm_users) == 1)
  75                 target_node = numa_node_id();
  76
  77         flush_tlb_batched_pending(vma->vm_mm);
  78         arch_enter_lazy_mmu_mode();
  79         do {
  80                 oldpte = *pte;
  81                 if (pte_present(oldpte)) {
  82                         pte_t ptent;
  83                         bool preserve_write = prot_numa && pte_write(oldpte);
  84
  85                         /*
  86                          * Avoid trapping faults against the zero or KSM
  87                          * pages. See similar comment in change_huge_pmd.
  88                          */
  89                         if (prot_numa) {
  90                                 struct page *page;
  91                                 int nid;
  92
  93                                 /* Avoid TLB flush if possible */
  94                                 if (pte_protnone(oldpte))
  95                                         continue;
  96
  97                                 page = vm_normal_page(vma, addr, oldpte);
  98                                 if (!page || PageKsm(page))
  99                                         continue;
 100
 101                                 /* Also skip shared copy-on-write pages */
 102                                 if (is_cow_mapping(vma->vm_flags) &&
 103                                     page_count(page) != 1)
 104                                         continue;
 105
 106                                 /*
 107                                  * While migration can move some dirty pages,
 108                                  * it cannot move them all from MIGRATE_ASYNC
 109                                  * context.
 110                                  */
 111                                 if (page_is_file_lru(page) && PageDirty(page))
 112                                         continue;
 113
 114                                 /*
 115                                  * Don't mess with PTEs if page is already on the node
 116                                  * a single-threaded process is running on.
 117                                  */
 118                                 nid = page_to_nid(page);
 119                                 if (target_node == nid)
 120                                         continue;
 121
 122                                 /*
 123                                  * Skip scanning top tier node if normal numa
 124                                  * balancing is disabled
 125                                  */
 126                                 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_NORMAL) &&
 127                                     node_is_toptier(nid))
 128                                         continue;
 129                         }
 130
 131                         oldpte = ptep_modify_prot_start(vma, addr, pte);
 132                         ptent = pte_modify(oldpte, newprot);
 133                         if (preserve_write)
 134                                 ptent = pte_mk_savedwrite(ptent);
 135
 136                         if (uffd_wp) {
 137                                 ptent = pte_wrprotect(ptent);
 138                                 ptent = pte_mkuffd_wp(ptent);
 139                         } else if (uffd_wp_resolve) {
 140                                 /*
 141                                  * Leave the write bit to be handled
 142                                  * by PF interrupt handler, then
 143                                  * things like COW could be properly
 144                                  * handled.
 145                                  */
 146                                 ptent = pte_clear_uffd_wp(ptent);
 147                         }
 148
 149                         /* Avoid taking write faults for known dirty pages */
 150                         if (dirty_accountable && pte_dirty(ptent) &&
 151                                         (pte_soft_dirty(ptent) ||
 152                                          !(vma->vm_flags & VM_SOFTDIRTY))) {
 153                                 ptent = pte_mkwrite(ptent);
 154                         }
 155                         ptep_modify_prot_commit(vma, addr, pte, oldpte, ptent);
 156                         if (pte_needs_flush(oldpte, ptent))
 157                                 tlb_flush_pte_range(tlb, addr, PAGE_SIZE);
 158                         pages++;
 159                 } else if (is_swap_pte(oldpte)) {
 160                         swp_entry_t entry = pte_to_swp_entry(oldpte);
 161                         struct page *page = pfn_swap_entry_to_page(entry);
 162                         pte_t newpte;
 163
 164                         if (is_writable_migration_entry(entry)) {
 165                                 /*
 166                                  * A protection check is difficult so
 167                                  * just be safe and disable write
 168                                  */
 169                                 if (PageAnon(page))
 170                                         entry = make_readable_exclusive_migration_entry(
 171                                                              swp_offset(entry));
 172                                 else
 173                                         entry = make_readable_migration_entry(swp_offset(entry));
 174                                 newpte = swp_entry_to_pte(entry);
 175                                 if (pte_swp_soft_dirty(oldpte))
 176                                         newpte = pte_swp_mksoft_dirty(newpte);
 177                                 if (pte_swp_uffd_wp(oldpte))
 178                                         newpte = pte_swp_mkuffd_wp(newpte);
 179                         } else if (is_writable_device_private_entry(entry)) {
 180                                 /*
 181                                  * We do not preserve soft-dirtiness. See
 182                                  * copy_one_pte() for explanation.
 183                                  */
 184                                 entry = make_readable_device_private_entry(
 185                                                         swp_offset(entry));
 186                                 newpte = swp_entry_to_pte(entry);
 187                                 if (pte_swp_uffd_wp(oldpte))
 188                                         newpte = pte_swp_mkuffd_wp(newpte);
 189                         } else if (is_writable_device_exclusive_entry(entry)) {
 190                                 entry = make_readable_device_exclusive_entry(
 191                                                         swp_offset(entry));
 192                                 newpte = swp_entry_to_pte(entry);
 193                                 if (pte_swp_soft_dirty(oldpte))
 194                                         newpte = pte_swp_mksoft_dirty(newpte);
 195                                 if (pte_swp_uffd_wp(oldpte))
 196                                         newpte = pte_swp_mkuffd_wp(newpte);
 197                         } else if (pte_marker_entry_uffd_wp(entry)) {
 198                                 /*
 199                                  * If this is uffd-wp pte marker and we'd like
 200                                  * to unprotect it, drop it; the next page
 201                                  * fault will trigger without uffd trapping.
 202                                  */
 203                                 if (uffd_wp_resolve) {
 204                                         pte_clear(vma->vm_mm, addr, pte);
 205                                         pages++;
 206                                 }
 207                                 continue;
 208                         } else {
 209                                 newpte = oldpte;
 210                         }
 211
 212                         if (uffd_wp)
 213                                 newpte = pte_swp_mkuffd_wp(newpte);
 214                         else if (uffd_wp_resolve)
 215                                 newpte = pte_swp_clear_uffd_wp(newpte);
 216
 217                         if (!pte_same(oldpte, newpte)) {
 218                                 set_pte_at(vma->vm_mm, addr, pte, newpte);
 219                                 pages++;
 220                         }
 221                 } else {
 222                         /* It must be an none page, or what else?.. */
 223                         WARN_ON_ONCE(!pte_none(oldpte));
 224                         if (unlikely(uffd_wp && !vma_is_anonymous(vma))) {
 225                                 /*
 226                                  * For file-backed mem, we need to be able to
 227                                  * wr-protect a none pte, because even if the
 228                                  * pte is none, the page/swap cache could
 229                                  * exist.  Doing that by install a marker.
 230                                  */
 231                                 set_pte_at(vma->vm_mm, addr, pte,
 232                                            make_pte_marker(PTE_MARKER_UFFD_WP));
 233                                 pages++;
 234                         }
 235                 }
 236         } while (pte++, addr += PAGE_SIZE, addr != end);
 237         arch_leave_lazy_mmu_mode();
 238         pte_unmap_unlock(pte - 1, ptl);
 239
 240         return pages;
 241 }
 242
 243 /*
 244  * Used when setting automatic NUMA hinting protection where it is
 245  * critical that a numa hinting PMD is not confused with a bad PMD.
 246  */
 247 static inline int pmd_none_or_clear_bad_unless_trans_huge(pmd_t *pmd)
 248 {
 249         pmd_t pmdval = pmd_read_atomic(pmd);
 250
 251         /* See pmd_none_or_trans_huge_or_clear_bad for info on barrier */
 252 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 253         barrier();
 254 #endif
 255
 256         if (pmd_none(pmdval))
 257                 return 1;
 258         if (pmd_trans_huge(pmdval))
 259                 return 0;
 260         if (unlikely(pmd_bad(pmdval))) {
 261                 pmd_clear_bad(pmd);
 262                 return 1;
 263         }
 264
 265         return 0;
 266 }
 267
 268 /* Return true if we're uffd wr-protecting file-backed memory, or false */
 269 static inline bool
 270 uffd_wp_protect_file(struct vm_area_struct *vma, unsigned long cp_flags)
 271 {
 272         return (cp_flags & MM_CP_UFFD_WP) && !vma_is_anonymous(vma);
 273 }
 274
 275 /*
 276  * If wr-protecting the range for file-backed, populate pgtable for the case
 277  * when pgtable is empty but page cache exists.  When {pte|pmd|...}_alloc()
 278  * failed it means no memory, we don't have a better option but stop.
 279  */
 280 #define  change_pmd_prepare(vma, pmd, cp_flags)                         \
 281         do {                                                            \
 282                 if (unlikely(uffd_wp_protect_file(vma, cp_flags))) {    \
 283                         if (WARN_ON_ONCE(pte_alloc(vma->vm_mm, pmd)))   \
 284                                 break;                                  \
 285                 }                                                       \
 286         } while (0)
 287 /*
 288  * This is the general pud/p4d/pgd version of change_pmd_prepare(). We need to
 289  * have separate change_pmd_prepare() because pte_alloc() returns 0 on success,
 290  * while {pmd|pud|p4d}_alloc() returns the valid pointer on success.
 291  */
 292 #define  change_prepare(vma, high, low, addr, cp_flags)                 \
 293         do {                                                            \
 294                 if (unlikely(uffd_wp_protect_file(vma, cp_flags))) {    \
 295                         low##_t *p = low##_alloc(vma->vm_mm, high, addr); \
 296                         if (WARN_ON_ONCE(p == NULL))                    \
 297                                 break;                                  \
 298                 }                                                       \
 299         } while (0)
 300
 301 static inline unsigned long change_pmd_range(struct mmu_gather *tlb,
 302                 struct vm_area_struct *vma, pud_t *pud, unsigned long addr,
 303                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
 304 {
 305         pmd_t *pmd;
 306         unsigned long next;
 307         unsigned long pages = 0;
 308         unsigned long nr_huge_updates = 0;
 309         struct mmu_notifier_range range;
 310
 311         range.start = 0;
 312
 313         pmd = pmd_offset(pud, addr);
 314         do {
 315                 unsigned long this_pages;
 316
 317                 next = pmd_addr_end(addr, end);
 318
 319                 change_pmd_prepare(vma, pmd, cp_flags);
 320                 /*
 321                  * Automatic NUMA balancing walks the tables with mmap_lock
 322                  * held for read. It's possible a parallel update to occur
 323                  * between pmd_trans_huge() and a pmd_none_or_clear_bad()
 324                  * check leading to a false positive and clearing.
 325                  * Hence, it's necessary to atomically read the PMD value
 326                  * for all the checks.
 327                  */
 328                 if (!is_swap_pmd(*pmd) && !pmd_devmap(*pmd) &&
 329                      pmd_none_or_clear_bad_unless_trans_huge(pmd))
 330                         goto next;
 331
 332                 /* invoke the mmu notifier if the pmd is populated */
 333                 if (!range.start) {
 334                         mmu_notifier_range_init(&range,
 335                                 MMU_NOTIFY_PROTECTION_VMA, 0,
 336                                 vma, vma->vm_mm, addr, end);
 337                         mmu_notifier_invalidate_range_start(&range);
 338                 }
 339
 340                 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
 341                         if ((next - addr != HPAGE_PMD_SIZE) ||
 342                             uffd_wp_protect_file(vma, cp_flags)) {
 343                                 __split_huge_pmd(vma, pmd, addr, false, NULL);
 344                                 /*
 345                                  * For file-backed, the pmd could have been
 346                                  * cleared; make sure pmd populated if
 347                                  * necessary, then fall-through to pte level.
 348                                  */
 349                                 change_pmd_prepare(vma, pmd, cp_flags);
 350                         } else {
 351                                 /*
 352                                  * change_huge_pmd() does not defer TLB flushes,
 353                                  * so no need to propagate the tlb argument.
 354                                  */
 355                                 int nr_ptes = change_huge_pmd(tlb, vma, pmd,
 356                                                 addr, newprot, cp_flags);
 357
 358                                 if (nr_ptes) {
 359                                         if (nr_ptes == HPAGE_PMD_NR) {
 360                                                 pages += HPAGE_PMD_NR;
 361                                                 nr_huge_updates++;
 362                                         }
 363
 364                                         /* huge pmd was handled */
 365                                         goto next;
 366                                 }
 367                         }
 368                         /* fall through, the trans huge pmd just split */
 369                 }
 370                 this_pages = change_pte_range(tlb, vma, pmd, addr, next,
 371                                               newprot, cp_flags);
 372                 pages += this_pages;
 373 next:
 374                 cond_resched();
 375         } while (pmd++, addr = next, addr != end);
 376
 377         if (range.start)
 378                 mmu_notifier_invalidate_range_end(&range);
 379
 380         if (nr_huge_updates)
 381                 count_vm_numa_events(NUMA_HUGE_PTE_UPDATES, nr_huge_updates);
 382         return pages;
 383 }
 384
 385 static inline unsigned long change_pud_range(struct mmu_gather *tlb,
 386                 struct vm_area_struct *vma, p4d_t *p4d, unsigned long addr,
 387                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
 388 {
 389         pud_t *pud;
 390         unsigned long next;
 391         unsigned long pages = 0;
 392
 393         pud = pud_offset(p4d, addr);
 394         do {
 395                 next = pud_addr_end(addr, end);
 396                 change_prepare(vma, pud, pmd, addr, cp_flags);
 397                 if (pud_none_or_clear_bad(pud))
 398                         continue;
 399                 pages += change_pmd_range(tlb, vma, pud, addr, next, newprot,
 400                                           cp_flags);
 401         } while (pud++, addr = next, addr != end);
 402
 403         return pages;
 404 }
 405
 406 static inline unsigned long change_p4d_range(struct mmu_gather *tlb,
 407                 struct vm_area_struct *vma, pgd_t *pgd, unsigned long addr,
 408                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
 409 {
 410         p4d_t *p4d;
 411         unsigned long next;
 412         unsigned long pages = 0;
 413
 414         p4d = p4d_offset(pgd, addr);
 415         do {
 416                 next = p4d_addr_end(addr, end);
 417                 change_prepare(vma, p4d, pud, addr, cp_flags);
 418                 if (p4d_none_or_clear_bad(p4d))
 419                         continue;
 420                 pages += change_pud_range(tlb, vma, p4d, addr, next, newprot,
 421                                           cp_flags);
 422         } while (p4d++, addr = next, addr != end);
 423
 424         return pages;
 425 }
 426
 427 static unsigned long change_protection_range(struct mmu_gather *tlb,
 428                 struct vm_area_struct *vma, unsigned long addr,
 429                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
 430 {
 431         struct mm_struct *mm = vma->vm_mm;
 432         pgd_t *pgd;
 433         unsigned long next;
 434         unsigned long pages = 0;
 435
 436         BUG_ON(addr >= end);
 437         pgd = pgd_offset(mm, addr);
 438         tlb_start_vma(tlb, vma);
 439         do {
 440                 next = pgd_addr_end(addr, end);
 441                 change_prepare(vma, pgd, p4d, addr, cp_flags);
 442                 if (pgd_none_or_clear_bad(pgd))
 443                         continue;
 444                 pages += change_p4d_range(tlb, vma, pgd, addr, next, newprot,
 445                                           cp_flags);
 446         } while (pgd++, addr = next, addr != end);
 447
 448         tlb_end_vma(tlb, vma);
 449
 450         return pages;
 451 }
 452
 453 unsigned long change_protection(struct mmu_gather *tlb,
 454                        struct vm_area_struct *vma, unsigned long start,
 455                        unsigned long end, pgprot_t newprot,
 456                        unsigned long cp_flags)
 457 {
 458         unsigned long pages;
 459
 460         BUG_ON((cp_flags & MM_CP_UFFD_WP_ALL) == MM_CP_UFFD_WP_ALL);
 461
 462         if (is_vm_hugetlb_page(vma))
 463                 pages = hugetlb_change_protection(vma, start, end, newprot,
 464                                                   cp_flags);
 465         else
 466                 pages = change_protection_range(tlb, vma, start, end, newprot,
 467                                                 cp_flags);
 468
 469         return pages;
 470 }
 471
 472 static int prot_none_pte_entry(pte_t *pte, unsigned long addr,
 473                                unsigned long next, struct mm_walk *walk)
 474 {
 475         return pfn_modify_allowed(pte_pfn(*pte), *(pgprot_t *)(walk->private)) ?
 476                 0 : -EACCES;
 477 }
 478
 479 static int prot_none_hugetlb_entry(pte_t *pte, unsigned long hmask,
 480                                    unsigned long addr, unsigned long next,
 481                                    struct mm_walk *walk)
 482 {
 483         return pfn_modify_allowed(pte_pfn(*pte), *(pgprot_t *)(walk->private)) ?
 484                 0 : -EACCES;
 485 }
 486
 487 static int prot_none_test(unsigned long addr, unsigned long next,
 488                           struct mm_walk *walk)
 489 {
 490         return 0;
 491 }
 492
 493 static const struct mm_walk_ops prot_none_walk_ops = {
 494         .pte_entry              = prot_none_pte_entry,
 495         .hugetlb_entry          = prot_none_hugetlb_entry,
 496         .test_walk              = prot_none_test,
 497 };
 498
 499 int
 500 mprotect_fixup(struct mmu_gather *tlb, struct vm_area_struct *vma,
 501                struct vm_area_struct **pprev, unsigned long start,
 502                unsigned long end, unsigned long newflags)
 503 {
 504         struct mm_struct *mm = vma->vm_mm;
 505         unsigned long oldflags = vma->vm_flags;
 506         long nrpages = (end - start) >> PAGE_SHIFT;
 507         unsigned long charged = 0;
 508         pgoff_t pgoff;
 509         int error;
 510         int dirty_accountable = 0;
 511
 512         if (newflags == oldflags) {
 513                 *pprev = vma;
 514                 return 0;
 515         }
 516
 517         /*
 518          * Do PROT_NONE PFN permission checks here when we can still
 519          * bail out without undoing a lot of state. This is a rather
 520          * uncommon case, so doesn't need to be very optimized.
 521          */
 522         if (arch_has_pfn_modify_check() &&
 523             (vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 524             (newflags & VM_ACCESS_FLAGS) == 0) {
 525                 pgprot_t new_pgprot = vm_get_page_prot(newflags);
 526
 527                 error = walk_page_range(current->mm, start, end,
 528                                 &prot_none_walk_ops, &new_pgprot);
 529                 if (error)
 530                         return error;
 531         }
 532
 533         /*
 534          * If we make a private mapping writable we increase our commit;
 535          * but (without finer accounting) cannot reduce our commit if we
 536          * make it unwritable again. hugetlb mapping were accounted for
 537          * even if read-only so there is no need to account for them here
 538          */
 539         if (newflags & VM_WRITE) {
 540                 /* Check space limits when area turns into data. */
 541                 if (!may_expand_vm(mm, newflags, nrpages) &&
 542                                 may_expand_vm(mm, oldflags, nrpages))
 543                         return -ENOMEM;
 544                 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
 545                                                 VM_SHARED|VM_NORESERVE))) {
 546                         charged = nrpages;
 547                         if (security_vm_enough_memory_mm(mm, charged))
 548                                 return -ENOMEM;
 549                         newflags |= VM_ACCOUNT;
 550                 }
 551         }
 552
 553         /*
 554          * First try to merge with previous and/or next vma.
 555          */
 556         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 557         *pprev = vma_merge(mm, *pprev, start, end, newflags,
 558                            vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
 559                            vma->vm_userfaultfd_ctx, anon_vma_name(vma));
 560         if (*pprev) {
 561                 vma = *pprev;
 562                 VM_WARN_ON((vma->vm_flags ^ newflags) & ~VM_SOFTDIRTY);
 563                 goto success;
 564         }
 565
 566         *pprev = vma;
 567
 568         if (start != vma->vm_start) {
 569                 error = split_vma(mm, vma, start, 1);
 570                 if (error)
 571                         goto fail;
 572         }
 573
 574         if (end != vma->vm_end) {
 575                 error = split_vma(mm, vma, end, 0);
 576                 if (error)
 577                         goto fail;
 578         }
 579
 580 success:
 581         /*
 582          * vm_flags and vm_page_prot are protected by the mmap_lock
 583          * held in write mode.
 584          */
 585         vma->vm_flags = newflags;
 586         dirty_accountable = vma_wants_writenotify(vma, vma->vm_page_prot);
 587         vma_set_page_prot(vma);
 588
 589         change_protection(tlb, vma, start, end, vma->vm_page_prot,
 590                           dirty_accountable ? MM_CP_DIRTY_ACCT : 0);
 591
 592         /*
 593          * Private VM_LOCKED VMA becoming writable: trigger COW to avoid major
 594          * fault on access.
 595          */
 596         if ((oldflags & (VM_WRITE | VM_SHARED | VM_LOCKED)) == VM_LOCKED &&
 597                         (newflags & VM_WRITE)) {
 598                 populate_vma_page_range(vma, start, end, NULL);
 599         }
 600
 601         vm_stat_account(mm, oldflags, -nrpages);
 602         vm_stat_account(mm, newflags, nrpages);
 603         perf_event_mmap(vma);
 604         return 0;
 605
 606 fail:
 607         vm_unacct_memory(charged);
 608         return error;
 609 }
 610
 611 /*
 612  * pkey==-1 when doing a legacy mprotect()
 613  */
 614 static int do_mprotect_pkey(unsigned long start, size_t len,
 615                 unsigned long prot, int pkey)
 616 {
 617         unsigned long nstart, end, tmp, reqprot;
 618         struct vm_area_struct *vma, *prev;
 619         int error = -EINVAL;
 620         const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
 621         const bool rier = (current->personality & READ_IMPLIES_EXEC) &&
 622                                 (prot & PROT_READ);
 623         struct mmu_gather tlb;
 624
 625         start = untagged_addr(start);
 626
 627         prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
 628         if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
 629                 return -EINVAL;
 630
 631         if (start & ~PAGE_MASK)
 632                 return -EINVAL;
 633         if (!len)
 634                 return 0;
 635         len = PAGE_ALIGN(len);
 636         end = start + len;
 637         if (end <= start)
 638                 return -ENOMEM;
 639         if (!arch_validate_prot(prot, start))
 640                 return -EINVAL;
 641
 642         reqprot = prot;
 643
 644         if (mmap_write_lock_killable(current->mm))
 645                 return -EINTR;
 646
 647         /*
 648          * If userspace did not allocate the pkey, do not let
 649          * them use it here.
 650          */
 651         error = -EINVAL;
 652         if ((pkey != -1) && !mm_pkey_is_allocated(current->mm, pkey))
 653                 goto out;
 654
 655         vma = find_vma(current->mm, start);
 656         error = -ENOMEM;
 657         if (!vma)
 658                 goto out;
 659
 660         if (unlikely(grows & PROT_GROWSDOWN)) {
 661                 if (vma->vm_start >= end)
 662                         goto out;
 663                 start = vma->vm_start;
 664                 error = -EINVAL;
 665                 if (!(vma->vm_flags & VM_GROWSDOWN))
 666                         goto out;
 667         } else {
 668                 if (vma->vm_start > start)
 669                         goto out;
 670                 if (unlikely(grows & PROT_GROWSUP)) {
 671                         end = vma->vm_end;
 672                         error = -EINVAL;
 673                         if (!(vma->vm_flags & VM_GROWSUP))
 674                                 goto out;
 675                 }
 676         }
 677
 678         if (start > vma->vm_start)
 679                 prev = vma;
 680         else
 681                 prev = vma->vm_prev;
 682
 683         tlb_gather_mmu(&tlb, current->mm);
 684         for (nstart = start ; ; ) {
 685                 unsigned long mask_off_old_flags;
 686                 unsigned long newflags;
 687                 int new_vma_pkey;
 688
 689                 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
 690
 691                 /* Does the application expect PROT_READ to imply PROT_EXEC */
 692                 if (rier && (vma->vm_flags & VM_MAYEXEC))
 693                         prot |= PROT_EXEC;
 694
 695                 /*
 696                  * Each mprotect() call explicitly passes r/w/x permissions.
 697                  * If a permission is not passed to mprotect(), it must be
 698                  * cleared from the VMA.
 699                  */
 700                 mask_off_old_flags = VM_READ | VM_WRITE | VM_EXEC |
 701                                         VM_FLAGS_CLEAR;
 702
 703                 new_vma_pkey = arch_override_mprotect_pkey(vma, prot, pkey);
 704                 newflags = calc_vm_prot_bits(prot, new_vma_pkey);
 705                 newflags |= (vma->vm_flags & ~mask_off_old_flags);
 706
 707                 /* newflags >> 4 shift VM_MAY% in place of VM_% */
 708                 if ((newflags & ~(newflags >> 4)) & VM_ACCESS_FLAGS) {
 709                         error = -EACCES;
 710                         break;
 711                 }
 712
 713                 /* Allow architectures to sanity-check the new flags */
 714                 if (!arch_validate_flags(newflags)) {
 715                         error = -EINVAL;
 716                         break;
 717                 }
 718
 719                 error = security_file_mprotect(vma, reqprot, prot);
 720                 if (error)
 721                         break;
 722
 723                 tmp = vma->vm_end;
 724                 if (tmp > end)
 725                         tmp = end;
 726
 727                 if (vma->vm_ops && vma->vm_ops->mprotect) {
 728                         error = vma->vm_ops->mprotect(vma, nstart, tmp, newflags);
 729                         if (error)
 730                                 break;
 731                 }
 732
 733                 error = mprotect_fixup(&tlb, vma, &prev, nstart, tmp, newflags);
 734                 if (error)
 735                         break;
 736
 737                 nstart = tmp;
 738
 739                 if (nstart < prev->vm_end)
 740                         nstart = prev->vm_end;
 741                 if (nstart >= end)
 742                         break;
 743
 744                 vma = prev->vm_next;
 745                 if (!vma || vma->vm_start != nstart) {
 746                         error = -ENOMEM;
 747                         break;
 748                 }
 749                 prot = reqprot;
 750         }
 751         tlb_finish_mmu(&tlb);
 752 out:
 753         mmap_write_unlock(current->mm);
 754         return error;
 755 }
 756
 757 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
 758                 unsigned long, prot)
 759 {
 760         return do_mprotect_pkey(start, len, prot, -1);
 761 }
 762
 763 #ifdef CONFIG_ARCH_HAS_PKEYS
 764
 765 SYSCALL_DEFINE4(pkey_mprotect, unsigned long, start, size_t, len,
 766                 unsigned long, prot, int, pkey)
 767 {
 768         return do_mprotect_pkey(start, len, prot, pkey);
 769 }
 770
 771 SYSCALL_DEFINE2(pkey_alloc, unsigned long, flags, unsigned long, init_val)
 772 {
 773         int pkey;
 774         int ret;
 775
 776         /* No flags supported yet. */
 777         if (flags)
 778                 return -EINVAL;
 779         /* check for unsupported init values */
 780         if (init_val & ~PKEY_ACCESS_MASK)
 781                 return -EINVAL;
 782
 783         mmap_write_lock(current->mm);
 784         pkey = mm_pkey_alloc(current->mm);
 785
 786         ret = -ENOSPC;
 787         if (pkey == -1)
 788                 goto out;
 789
 790         ret = arch_set_user_pkey_access(current, pkey, init_val);
 791         if (ret) {
 792                 mm_pkey_free(current->mm, pkey);
 793                 goto out;
 794         }
 795         ret = pkey;
 796 out:
 797         mmap_write_unlock(current->mm);
 798         return ret;
 799 }
 800
 801 SYSCALL_DEFINE1(pkey_free, int, pkey)
 802 {
 803         int ret;
 804
 805         mmap_write_lock(current->mm);
 806         ret = mm_pkey_free(current->mm, pkey);
 807         mmap_write_unlock(current->mm);
 808
 809         /*
 810          * We could provide warnings or errors if any VMA still
 811          * has the pkey set here.
 812          */
 813         return ret;
 814 }
 815
 816 #endif /* CONFIG_ARCH_HAS_PKEYS */