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 inline bool can_change_pte_writable(struct vm_area_struct *vma,
  42                                            unsigned long addr, pte_t pte)
  43 {
  44         struct page *page;
  45
  46         VM_BUG_ON(!(vma->vm_flags & VM_WRITE) || pte_write(pte));
  47
  48         if (pte_protnone(pte) || !pte_dirty(pte))
  49                 return false;
  50
  51         /* Do we need write faults for softdirty tracking? */
  52         if (vma_soft_dirty_enabled(vma) && !pte_soft_dirty(pte))
  53                 return false;
  54
  55         /* Do we need write faults for uffd-wp tracking? */
  56         if (userfaultfd_pte_wp(vma, pte))
  57                 return false;
  58
  59         if (!(vma->vm_flags & VM_SHARED)) {
  60                 /*
  61                  * We can only special-case on exclusive anonymous pages,
  62                  * because we know that our write-fault handler similarly would
  63                  * map them writable without any additional checks while holding
  64                  * the PT lock.
  65                  */
  66                 page = vm_normal_page(vma, addr, pte);
  67                 if (!page || !PageAnon(page) || !PageAnonExclusive(page))
  68                         return false;
  69         }
  70
  71         return true;
  72 }
  73
  74 static unsigned long change_pte_range(struct mmu_gather *tlb,
  75                 struct vm_area_struct *vma, pmd_t *pmd, unsigned long addr,
  76                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
  77 {
  78         pte_t *pte, oldpte;
  79         spinlock_t *ptl;
  80         unsigned long pages = 0;
  81         int target_node = NUMA_NO_NODE;
  82         bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
  83         bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
  84         bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
  85
  86         tlb_change_page_size(tlb, PAGE_SIZE);
  87
  88         /*
  89          * Can be called with only the mmap_lock for reading by
  90          * prot_numa so we must check the pmd isn't constantly
  91          * changing from under us from pmd_none to pmd_trans_huge
  92          * and/or the other way around.
  93          */
  94         if (pmd_trans_unstable(pmd))
  95                 return 0;
  96
  97         /*
  98          * The pmd points to a regular pte so the pmd can't change
  99          * from under us even if the mmap_lock is only hold for
 100          * reading.
 101          */
 102         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
 103
 104         /* Get target node for single threaded private VMAs */
 105         if (prot_numa && !(vma->vm_flags & VM_SHARED) &&
 106             atomic_read(&vma->vm_mm->mm_users) == 1)
 107                 target_node = numa_node_id();
 108
 109         flush_tlb_batched_pending(vma->vm_mm);
 110         arch_enter_lazy_mmu_mode();
 111         do {
 112                 oldpte = *pte;
 113                 if (pte_present(oldpte)) {
 114                         pte_t ptent;
 115                         bool preserve_write = prot_numa && pte_write(oldpte);
 116
 117                         /*
 118                          * Avoid trapping faults against the zero or KSM
 119                          * pages. See similar comment in change_huge_pmd.
 120                          */
 121                         if (prot_numa) {
 122                                 struct page *page;
 123                                 int nid;
 124
 125                                 /* Avoid TLB flush if possible */
 126                                 if (pte_protnone(oldpte))
 127                                         continue;
 128
 129                                 page = vm_normal_page(vma, addr, oldpte);
 130                                 if (!page || is_zone_device_page(page) || PageKsm(page))
 131                                         continue;
 132
 133                                 /* Also skip shared copy-on-write pages */
 134                                 if (is_cow_mapping(vma->vm_flags) &&
 135                                     page_count(page) != 1)
 136                                         continue;
 137
 138                                 /*
 139                                  * While migration can move some dirty pages,
 140                                  * it cannot move them all from MIGRATE_ASYNC
 141                                  * context.
 142                                  */
 143                                 if (page_is_file_lru(page) && PageDirty(page))
 144                                         continue;
 145
 146                                 /*
 147                                  * Don't mess with PTEs if page is already on the node
 148                                  * a single-threaded process is running on.
 149                                  */
 150                                 nid = page_to_nid(page);
 151                                 if (target_node == nid)
 152                                         continue;
 153
 154                                 /*
 155                                  * Skip scanning top tier node if normal numa
 156                                  * balancing is disabled
 157                                  */
 158                                 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_NORMAL) &&
 159                                     node_is_toptier(nid))
 160                                         continue;
 161                         }
 162
 163                         oldpte = ptep_modify_prot_start(vma, addr, pte);
 164                         ptent = pte_modify(oldpte, newprot);
 165                         if (preserve_write)
 166                                 ptent = pte_mk_savedwrite(ptent);
 167
 168                         if (uffd_wp) {
 169                                 ptent = pte_wrprotect(ptent);
 170                                 ptent = pte_mkuffd_wp(ptent);
 171                         } else if (uffd_wp_resolve) {
 172                                 ptent = pte_clear_uffd_wp(ptent);
 173                         }
 174
 175                         /*
 176                          * In some writable, shared mappings, we might want
 177                          * to catch actual write access -- see
 178                          * vma_wants_writenotify().
 179                          *
 180                          * In all writable, private mappings, we have to
 181                          * properly handle COW.
 182                          *
 183                          * In both cases, we can sometimes still change PTEs
 184                          * writable and avoid the write-fault handler, for
 185                          * example, if a PTE is already dirty and no other
 186                          * COW or special handling is required.
 187                          */
 188                         if ((cp_flags & MM_CP_TRY_CHANGE_WRITABLE) &&
 189                             !pte_write(ptent) &&
 190                             can_change_pte_writable(vma, addr, ptent))
 191                                 ptent = pte_mkwrite(ptent);
 192
 193                         ptep_modify_prot_commit(vma, addr, pte, oldpte, ptent);
 194                         if (pte_needs_flush(oldpte, ptent))
 195                                 tlb_flush_pte_range(tlb, addr, PAGE_SIZE);
 196                         pages++;
 197                 } else if (is_swap_pte(oldpte)) {
 198                         swp_entry_t entry = pte_to_swp_entry(oldpte);
 199                         struct page *page = pfn_swap_entry_to_page(entry);
 200                         pte_t newpte;
 201
 202                         if (is_writable_migration_entry(entry)) {
 203                                 /*
 204                                  * A protection check is difficult so
 205                                  * just be safe and disable write
 206                                  */
 207                                 if (PageAnon(page))
 208                                         entry = make_readable_exclusive_migration_entry(
 209                                                              swp_offset(entry));
 210                                 else
 211                                         entry = make_readable_migration_entry(swp_offset(entry));
 212                                 newpte = swp_entry_to_pte(entry);
 213                                 if (pte_swp_soft_dirty(oldpte))
 214                                         newpte = pte_swp_mksoft_dirty(newpte);
 215                                 if (pte_swp_uffd_wp(oldpte))
 216                                         newpte = pte_swp_mkuffd_wp(newpte);
 217                         } else if (is_writable_device_private_entry(entry)) {
 218                                 /*
 219                                  * We do not preserve soft-dirtiness. See
 220                                  * copy_one_pte() for explanation.
 221                                  */
 222                                 entry = make_readable_device_private_entry(
 223                                                         swp_offset(entry));
 224                                 newpte = swp_entry_to_pte(entry);
 225                                 if (pte_swp_uffd_wp(oldpte))
 226                                         newpte = pte_swp_mkuffd_wp(newpte);
 227                         } else if (is_writable_device_exclusive_entry(entry)) {
 228                                 entry = make_readable_device_exclusive_entry(
 229                                                         swp_offset(entry));
 230                                 newpte = swp_entry_to_pte(entry);
 231                                 if (pte_swp_soft_dirty(oldpte))
 232                                         newpte = pte_swp_mksoft_dirty(newpte);
 233                                 if (pte_swp_uffd_wp(oldpte))
 234                                         newpte = pte_swp_mkuffd_wp(newpte);
 235                         } else if (pte_marker_entry_uffd_wp(entry)) {
 236                                 /*
 237                                  * If this is uffd-wp pte marker and we'd like
 238                                  * to unprotect it, drop it; the next page
 239                                  * fault will trigger without uffd trapping.
 240                                  */
 241                                 if (uffd_wp_resolve) {
 242                                         pte_clear(vma->vm_mm, addr, pte);
 243                                         pages++;
 244                                 }
 245                                 continue;
 246                         } else {
 247                                 newpte = oldpte;
 248                         }
 249
 250                         if (uffd_wp)
 251                                 newpte = pte_swp_mkuffd_wp(newpte);
 252                         else if (uffd_wp_resolve)
 253                                 newpte = pte_swp_clear_uffd_wp(newpte);
 254
 255                         if (!pte_same(oldpte, newpte)) {
 256                                 set_pte_at(vma->vm_mm, addr, pte, newpte);
 257                                 pages++;
 258                         }
 259                 } else {
 260                         /* It must be an none page, or what else?.. */
 261                         WARN_ON_ONCE(!pte_none(oldpte));
 262                         if (unlikely(uffd_wp && !vma_is_anonymous(vma))) {
 263                                 /*
 264                                  * For file-backed mem, we need to be able to
 265                                  * wr-protect a none pte, because even if the
 266                                  * pte is none, the page/swap cache could
 267                                  * exist.  Doing that by install a marker.
 268                                  */
 269                                 set_pte_at(vma->vm_mm, addr, pte,
 270                                            make_pte_marker(PTE_MARKER_UFFD_WP));
 271                                 pages++;
 272                         }
 273                 }
 274         } while (pte++, addr += PAGE_SIZE, addr != end);
 275         arch_leave_lazy_mmu_mode();
 276         pte_unmap_unlock(pte - 1, ptl);
 277
 278         return pages;
 279 }
 280
 281 /*
 282  * Used when setting automatic NUMA hinting protection where it is
 283  * critical that a numa hinting PMD is not confused with a bad PMD.
 284  */
 285 static inline int pmd_none_or_clear_bad_unless_trans_huge(pmd_t *pmd)
 286 {
 287         pmd_t pmdval = pmd_read_atomic(pmd);
 288
 289         /* See pmd_none_or_trans_huge_or_clear_bad for info on barrier */
 290 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 291         barrier();
 292 #endif
 293
 294         if (pmd_none(pmdval))
 295                 return 1;
 296         if (pmd_trans_huge(pmdval))
 297                 return 0;
 298         if (unlikely(pmd_bad(pmdval))) {
 299                 pmd_clear_bad(pmd);
 300                 return 1;
 301         }
 302
 303         return 0;
 304 }
 305
 306 /* Return true if we're uffd wr-protecting file-backed memory, or false */
 307 static inline bool
 308 uffd_wp_protect_file(struct vm_area_struct *vma, unsigned long cp_flags)
 309 {
 310         return (cp_flags & MM_CP_UFFD_WP) && !vma_is_anonymous(vma);
 311 }
 312
 313 /*
 314  * If wr-protecting the range for file-backed, populate pgtable for the case
 315  * when pgtable is empty but page cache exists.  When {pte|pmd|...}_alloc()
 316  * failed it means no memory, we don't have a better option but stop.
 317  */
 318 #define  change_pmd_prepare(vma, pmd, cp_flags)                         \
 319         do {                                                            \
 320                 if (unlikely(uffd_wp_protect_file(vma, cp_flags))) {    \
 321                         if (WARN_ON_ONCE(pte_alloc(vma->vm_mm, pmd)))   \
 322                                 break;                                  \
 323                 }                                                       \
 324         } while (0)
 325 /*
 326  * This is the general pud/p4d/pgd version of change_pmd_prepare(). We need to
 327  * have separate change_pmd_prepare() because pte_alloc() returns 0 on success,
 328  * while {pmd|pud|p4d}_alloc() returns the valid pointer on success.
 329  */
 330 #define  change_prepare(vma, high, low, addr, cp_flags)                 \
 331         do {                                                            \
 332                 if (unlikely(uffd_wp_protect_file(vma, cp_flags))) {    \
 333                         low##_t *p = low##_alloc(vma->vm_mm, high, addr); \
 334                         if (WARN_ON_ONCE(p == NULL))                    \
 335                                 break;                                  \
 336                 }                                                       \
 337         } while (0)
 338
 339 static inline unsigned long change_pmd_range(struct mmu_gather *tlb,
 340                 struct vm_area_struct *vma, pud_t *pud, unsigned long addr,
 341                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
 342 {
 343         pmd_t *pmd;
 344         unsigned long next;
 345         unsigned long pages = 0;
 346         unsigned long nr_huge_updates = 0;
 347         struct mmu_notifier_range range;
 348
 349         range.start = 0;
 350
 351         pmd = pmd_offset(pud, addr);
 352         do {
 353                 unsigned long this_pages;
 354
 355                 next = pmd_addr_end(addr, end);
 356
 357                 change_pmd_prepare(vma, pmd, cp_flags);
 358                 /*
 359                  * Automatic NUMA balancing walks the tables with mmap_lock
 360                  * held for read. It's possible a parallel update to occur
 361                  * between pmd_trans_huge() and a pmd_none_or_clear_bad()
 362                  * check leading to a false positive and clearing.
 363                  * Hence, it's necessary to atomically read the PMD value
 364                  * for all the checks.
 365                  */
 366                 if (!is_swap_pmd(*pmd) && !pmd_devmap(*pmd) &&
 367                      pmd_none_or_clear_bad_unless_trans_huge(pmd))
 368                         goto next;
 369
 370                 /* invoke the mmu notifier if the pmd is populated */
 371                 if (!range.start) {
 372                         mmu_notifier_range_init(&range,
 373                                 MMU_NOTIFY_PROTECTION_VMA, 0,
 374                                 vma, vma->vm_mm, addr, end);
 375                         mmu_notifier_invalidate_range_start(&range);
 376                 }
 377
 378                 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
 379                         if ((next - addr != HPAGE_PMD_SIZE) ||
 380                             uffd_wp_protect_file(vma, cp_flags)) {
 381                                 __split_huge_pmd(vma, pmd, addr, false, NULL);
 382                                 /*
 383                                  * For file-backed, the pmd could have been
 384                                  * cleared; make sure pmd populated if
 385                                  * necessary, then fall-through to pte level.
 386                                  */
 387                                 change_pmd_prepare(vma, pmd, cp_flags);
 388                         } else {
 389                                 /*
 390                                  * change_huge_pmd() does not defer TLB flushes,
 391                                  * so no need to propagate the tlb argument.
 392                                  */
 393                                 int nr_ptes = change_huge_pmd(tlb, vma, pmd,
 394                                                 addr, newprot, cp_flags);
 395
 396                                 if (nr_ptes) {
 397                                         if (nr_ptes == HPAGE_PMD_NR) {
 398                                                 pages += HPAGE_PMD_NR;
 399                                                 nr_huge_updates++;
 400                                         }
 401
 402                                         /* huge pmd was handled */
 403                                         goto next;
 404                                 }
 405                         }
 406                         /* fall through, the trans huge pmd just split */
 407                 }
 408                 this_pages = change_pte_range(tlb, vma, pmd, addr, next,
 409                                               newprot, cp_flags);
 410                 pages += this_pages;
 411 next:
 412                 cond_resched();
 413         } while (pmd++, addr = next, addr != end);
 414
 415         if (range.start)
 416                 mmu_notifier_invalidate_range_end(&range);
 417
 418         if (nr_huge_updates)
 419                 count_vm_numa_events(NUMA_HUGE_PTE_UPDATES, nr_huge_updates);
 420         return pages;
 421 }
 422
 423 static inline unsigned long change_pud_range(struct mmu_gather *tlb,
 424                 struct vm_area_struct *vma, p4d_t *p4d, unsigned long addr,
 425                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
 426 {
 427         pud_t *pud;
 428         unsigned long next;
 429         unsigned long pages = 0;
 430
 431         pud = pud_offset(p4d, addr);
 432         do {
 433                 next = pud_addr_end(addr, end);
 434                 change_prepare(vma, pud, pmd, addr, cp_flags);
 435                 if (pud_none_or_clear_bad(pud))
 436                         continue;
 437                 pages += change_pmd_range(tlb, vma, pud, addr, next, newprot,
 438                                           cp_flags);
 439         } while (pud++, addr = next, addr != end);
 440
 441         return pages;
 442 }
 443
 444 static inline unsigned long change_p4d_range(struct mmu_gather *tlb,
 445                 struct vm_area_struct *vma, pgd_t *pgd, unsigned long addr,
 446                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
 447 {
 448         p4d_t *p4d;
 449         unsigned long next;
 450         unsigned long pages = 0;
 451
 452         p4d = p4d_offset(pgd, addr);
 453         do {
 454                 next = p4d_addr_end(addr, end);
 455                 change_prepare(vma, p4d, pud, addr, cp_flags);
 456                 if (p4d_none_or_clear_bad(p4d))
 457                         continue;
 458                 pages += change_pud_range(tlb, vma, p4d, addr, next, newprot,
 459                                           cp_flags);
 460         } while (p4d++, addr = next, addr != end);
 461
 462         return pages;
 463 }
 464
 465 static unsigned long change_protection_range(struct mmu_gather *tlb,
 466                 struct vm_area_struct *vma, unsigned long addr,
 467                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
 468 {
 469         struct mm_struct *mm = vma->vm_mm;
 470         pgd_t *pgd;
 471         unsigned long next;
 472         unsigned long pages = 0;
 473
 474         BUG_ON(addr >= end);
 475         pgd = pgd_offset(mm, addr);
 476         tlb_start_vma(tlb, vma);
 477         do {
 478                 next = pgd_addr_end(addr, end);
 479                 change_prepare(vma, pgd, p4d, addr, cp_flags);
 480                 if (pgd_none_or_clear_bad(pgd))
 481                         continue;
 482                 pages += change_p4d_range(tlb, vma, pgd, addr, next, newprot,
 483                                           cp_flags);
 484         } while (pgd++, addr = next, addr != end);
 485
 486         tlb_end_vma(tlb, vma);
 487
 488         return pages;
 489 }
 490
 491 unsigned long change_protection(struct mmu_gather *tlb,
 492                        struct vm_area_struct *vma, unsigned long start,
 493                        unsigned long end, pgprot_t newprot,
 494                        unsigned long cp_flags)
 495 {
 496         unsigned long pages;
 497
 498         BUG_ON((cp_flags & MM_CP_UFFD_WP_ALL) == MM_CP_UFFD_WP_ALL);
 499
 500         if (is_vm_hugetlb_page(vma))
 501                 pages = hugetlb_change_protection(vma, start, end, newprot,
 502                                                   cp_flags);
 503         else
 504                 pages = change_protection_range(tlb, vma, start, end, newprot,
 505                                                 cp_flags);
 506
 507         return pages;
 508 }
 509
 510 static int prot_none_pte_entry(pte_t *pte, unsigned long addr,
 511                                unsigned long next, struct mm_walk *walk)
 512 {
 513         return pfn_modify_allowed(pte_pfn(*pte), *(pgprot_t *)(walk->private)) ?
 514                 0 : -EACCES;
 515 }
 516
 517 static int prot_none_hugetlb_entry(pte_t *pte, unsigned long hmask,
 518                                    unsigned long addr, unsigned long next,
 519                                    struct mm_walk *walk)
 520 {
 521         return pfn_modify_allowed(pte_pfn(*pte), *(pgprot_t *)(walk->private)) ?
 522                 0 : -EACCES;
 523 }
 524
 525 static int prot_none_test(unsigned long addr, unsigned long next,
 526                           struct mm_walk *walk)
 527 {
 528         return 0;
 529 }
 530
 531 static const struct mm_walk_ops prot_none_walk_ops = {
 532         .pte_entry              = prot_none_pte_entry,
 533         .hugetlb_entry          = prot_none_hugetlb_entry,
 534         .test_walk              = prot_none_test,
 535 };
 536
 537 int
 538 mprotect_fixup(struct mmu_gather *tlb, struct vm_area_struct *vma,
 539                struct vm_area_struct **pprev, unsigned long start,
 540                unsigned long end, unsigned long newflags)
 541 {
 542         struct mm_struct *mm = vma->vm_mm;
 543         unsigned long oldflags = vma->vm_flags;
 544         long nrpages = (end - start) >> PAGE_SHIFT;
 545         unsigned long charged = 0;
 546         bool try_change_writable;
 547         pgoff_t pgoff;
 548         int error;
 549
 550         if (newflags == oldflags) {
 551                 *pprev = vma;
 552                 return 0;
 553         }
 554
 555         /*
 556          * Do PROT_NONE PFN permission checks here when we can still
 557          * bail out without undoing a lot of state. This is a rather
 558          * uncommon case, so doesn't need to be very optimized.
 559          */
 560         if (arch_has_pfn_modify_check() &&
 561             (vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 562             (newflags & VM_ACCESS_FLAGS) == 0) {
 563                 pgprot_t new_pgprot = vm_get_page_prot(newflags);
 564
 565                 error = walk_page_range(current->mm, start, end,
 566                                 &prot_none_walk_ops, &new_pgprot);
 567                 if (error)
 568                         return error;
 569         }
 570
 571         /*
 572          * If we make a private mapping writable we increase our commit;
 573          * but (without finer accounting) cannot reduce our commit if we
 574          * make it unwritable again. hugetlb mapping were accounted for
 575          * even if read-only so there is no need to account for them here
 576          */
 577         if (newflags & VM_WRITE) {
 578                 /* Check space limits when area turns into data. */
 579                 if (!may_expand_vm(mm, newflags, nrpages) &&
 580                                 may_expand_vm(mm, oldflags, nrpages))
 581                         return -ENOMEM;
 582                 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
 583                                                 VM_SHARED|VM_NORESERVE))) {
 584                         charged = nrpages;
 585                         if (security_vm_enough_memory_mm(mm, charged))
 586                                 return -ENOMEM;
 587                         newflags |= VM_ACCOUNT;
 588                 }
 589         }
 590
 591         /*
 592          * First try to merge with previous and/or next vma.
 593          */
 594         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 595         *pprev = vma_merge(mm, *pprev, start, end, newflags,
 596                            vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
 597                            vma->vm_userfaultfd_ctx, anon_vma_name(vma));
 598         if (*pprev) {
 599                 vma = *pprev;
 600                 VM_WARN_ON((vma->vm_flags ^ newflags) & ~VM_SOFTDIRTY);
 601                 goto success;
 602         }
 603
 604         *pprev = vma;
 605
 606         if (start != vma->vm_start) {
 607                 error = split_vma(mm, vma, start, 1);
 608                 if (error)
 609                         goto fail;
 610         }
 611
 612         if (end != vma->vm_end) {
 613                 error = split_vma(mm, vma, end, 0);
 614                 if (error)
 615                         goto fail;
 616         }
 617
 618 success:
 619         /*
 620          * vm_flags and vm_page_prot are protected by the mmap_lock
 621          * held in write mode.
 622          */
 623         vma->vm_flags = newflags;
 624         /*
 625          * We want to check manually if we can change individual PTEs writable
 626          * if we can't do that automatically for all PTEs in a mapping. For
 627          * private mappings, that's always the case when we have write
 628          * permissions as we properly have to handle COW.
 629          */
 630         if (vma->vm_flags & VM_SHARED)
 631                 try_change_writable = vma_wants_writenotify(vma, vma->vm_page_prot);
 632         else
 633                 try_change_writable = !!(vma->vm_flags & VM_WRITE);
 634         vma_set_page_prot(vma);
 635
 636         change_protection(tlb, vma, start, end, vma->vm_page_prot,
 637                           try_change_writable ? MM_CP_TRY_CHANGE_WRITABLE : 0);
 638
 639         /*
 640          * Private VM_LOCKED VMA becoming writable: trigger COW to avoid major
 641          * fault on access.
 642          */
 643         if ((oldflags & (VM_WRITE | VM_SHARED | VM_LOCKED)) == VM_LOCKED &&
 644                         (newflags & VM_WRITE)) {
 645                 populate_vma_page_range(vma, start, end, NULL);
 646         }
 647
 648         vm_stat_account(mm, oldflags, -nrpages);
 649         vm_stat_account(mm, newflags, nrpages);
 650         perf_event_mmap(vma);
 651         return 0;
 652
 653 fail:
 654         vm_unacct_memory(charged);
 655         return error;
 656 }
 657
 658 /*
 659  * pkey==-1 when doing a legacy mprotect()
 660  */
 661 static int do_mprotect_pkey(unsigned long start, size_t len,
 662                 unsigned long prot, int pkey)
 663 {
 664         unsigned long nstart, end, tmp, reqprot;
 665         struct vm_area_struct *vma, *prev;
 666         int error;
 667         const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
 668         const bool rier = (current->personality & READ_IMPLIES_EXEC) &&
 669                                 (prot & PROT_READ);
 670         struct mmu_gather tlb;
 671
 672         start = untagged_addr(start);
 673
 674         prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
 675         if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
 676                 return -EINVAL;
 677
 678         if (start & ~PAGE_MASK)
 679                 return -EINVAL;
 680         if (!len)
 681                 return 0;
 682         len = PAGE_ALIGN(len);
 683         end = start + len;
 684         if (end <= start)
 685                 return -ENOMEM;
 686         if (!arch_validate_prot(prot, start))
 687                 return -EINVAL;
 688
 689         reqprot = prot;
 690
 691         if (mmap_write_lock_killable(current->mm))
 692                 return -EINTR;
 693
 694         /*
 695          * If userspace did not allocate the pkey, do not let
 696          * them use it here.
 697          */
 698         error = -EINVAL;
 699         if ((pkey != -1) && !mm_pkey_is_allocated(current->mm, pkey))
 700                 goto out;
 701
 702         vma = find_vma(current->mm, start);
 703         error = -ENOMEM;
 704         if (!vma)
 705                 goto out;
 706
 707         if (unlikely(grows & PROT_GROWSDOWN)) {
 708                 if (vma->vm_start >= end)
 709                         goto out;
 710                 start = vma->vm_start;
 711                 error = -EINVAL;
 712                 if (!(vma->vm_flags & VM_GROWSDOWN))
 713                         goto out;
 714         } else {
 715                 if (vma->vm_start > start)
 716                         goto out;
 717                 if (unlikely(grows & PROT_GROWSUP)) {
 718                         end = vma->vm_end;
 719                         error = -EINVAL;
 720                         if (!(vma->vm_flags & VM_GROWSUP))
 721                                 goto out;
 722                 }
 723         }
 724
 725         if (start > vma->vm_start)
 726                 prev = vma;
 727         else
 728                 prev = vma->vm_prev;
 729
 730         tlb_gather_mmu(&tlb, current->mm);
 731         for (nstart = start ; ; ) {
 732                 unsigned long mask_off_old_flags;
 733                 unsigned long newflags;
 734                 int new_vma_pkey;
 735
 736                 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
 737
 738                 /* Does the application expect PROT_READ to imply PROT_EXEC */
 739                 if (rier && (vma->vm_flags & VM_MAYEXEC))
 740                         prot |= PROT_EXEC;
 741
 742                 /*
 743                  * Each mprotect() call explicitly passes r/w/x permissions.
 744                  * If a permission is not passed to mprotect(), it must be
 745                  * cleared from the VMA.
 746                  */
 747                 mask_off_old_flags = VM_READ | VM_WRITE | VM_EXEC |
 748                                         VM_FLAGS_CLEAR;
 749
 750                 new_vma_pkey = arch_override_mprotect_pkey(vma, prot, pkey);
 751                 newflags = calc_vm_prot_bits(prot, new_vma_pkey);
 752                 newflags |= (vma->vm_flags & ~mask_off_old_flags);
 753
 754                 /* newflags >> 4 shift VM_MAY% in place of VM_% */
 755                 if ((newflags & ~(newflags >> 4)) & VM_ACCESS_FLAGS) {
 756                         error = -EACCES;
 757                         break;
 758                 }
 759
 760                 /* Allow architectures to sanity-check the new flags */
 761                 if (!arch_validate_flags(newflags)) {
 762                         error = -EINVAL;
 763                         break;
 764                 }
 765
 766                 error = security_file_mprotect(vma, reqprot, prot);
 767                 if (error)
 768                         break;
 769
 770                 tmp = vma->vm_end;
 771                 if (tmp > end)
 772                         tmp = end;
 773
 774                 if (vma->vm_ops && vma->vm_ops->mprotect) {
 775                         error = vma->vm_ops->mprotect(vma, nstart, tmp, newflags);
 776                         if (error)
 777                                 break;
 778                 }
 779
 780                 error = mprotect_fixup(&tlb, vma, &prev, nstart, tmp, newflags);
 781                 if (error)
 782                         break;
 783
 784                 nstart = tmp;
 785
 786                 if (nstart < prev->vm_end)
 787                         nstart = prev->vm_end;
 788                 if (nstart >= end)
 789                         break;
 790
 791                 vma = prev->vm_next;
 792                 if (!vma || vma->vm_start != nstart) {
 793                         error = -ENOMEM;
 794                         break;
 795                 }
 796                 prot = reqprot;
 797         }
 798         tlb_finish_mmu(&tlb);
 799 out:
 800         mmap_write_unlock(current->mm);
 801         return error;
 802 }
 803
 804 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
 805                 unsigned long, prot)
 806 {
 807         return do_mprotect_pkey(start, len, prot, -1);
 808 }
 809
 810 #ifdef CONFIG_ARCH_HAS_PKEYS
 811
 812 SYSCALL_DEFINE4(pkey_mprotect, unsigned long, start, size_t, len,
 813                 unsigned long, prot, int, pkey)
 814 {
 815         return do_mprotect_pkey(start, len, prot, pkey);
 816 }
 817
 818 SYSCALL_DEFINE2(pkey_alloc, unsigned long, flags, unsigned long, init_val)
 819 {
 820         int pkey;
 821         int ret;
 822
 823         /* No flags supported yet. */
 824         if (flags)
 825                 return -EINVAL;
 826         /* check for unsupported init values */
 827         if (init_val & ~PKEY_ACCESS_MASK)
 828                 return -EINVAL;
 829
 830         mmap_write_lock(current->mm);
 831         pkey = mm_pkey_alloc(current->mm);
 832
 833         ret = -ENOSPC;
 834         if (pkey == -1)
 835                 goto out;
 836
 837         ret = arch_set_user_pkey_access(current, pkey, init_val);
 838         if (ret) {
 839                 mm_pkey_free(current->mm, pkey);
 840                 goto out;
 841         }
 842         ret = pkey;
 843 out:
 844         mmap_write_unlock(current->mm);
 845         return ret;
 846 }
 847
 848 SYSCALL_DEFINE1(pkey_free, int, pkey)
 849 {
 850         int ret;
 851
 852         mmap_write_lock(current->mm);
 853         ret = mm_pkey_free(current->mm, pkey);
 854         mmap_write_unlock(current->mm);
 855
 856         /*
 857          * We could provide warnings or errors if any VMA still
 858          * has the pkey set here.
 859          */
 860         return ret;
 861 }
 862
 863 #endif /* CONFIG_ARCH_HAS_PKEYS */