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