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 <linux/memory-tiers.h>
  35 #include <asm/cacheflush.h>
  36 #include <asm/mmu_context.h>
  37 #include <asm/tlbflush.h>
  38 #include <asm/tlb.h>
  39
  40 #include "internal.h"
  41
  42 bool can_change_pte_writable(struct vm_area_struct *vma, unsigned long addr,
  43                              pte_t pte)
  44 {
  45         struct page *page;
  46
  47         if (WARN_ON_ONCE(!(vma->vm_flags & VM_WRITE)))
  48                 return false;
  49
  50         /* Don't touch entries that are not even readable. */
  51         if (pte_protnone(pte))
  52                 return false;
  53
  54         /* Do we need write faults for softdirty tracking? */
  55         if (vma_soft_dirty_enabled(vma) && !pte_soft_dirty(pte))
  56                 return false;
  57
  58         /* Do we need write faults for uffd-wp tracking? */
  59         if (userfaultfd_pte_wp(vma, pte))
  60                 return false;
  61
  62         if (!(vma->vm_flags & VM_SHARED)) {
  63                 /*
  64                  * Writable MAP_PRIVATE mapping: We can only special-case on
  65                  * exclusive anonymous pages, because we know that our
  66                  * write-fault handler similarly would map them writable without
  67                  * any additional checks while holding the PT lock.
  68                  */
  69                 page = vm_normal_page(vma, addr, pte);
  70                 return page && PageAnon(page) && PageAnonExclusive(page);
  71         }
  72
  73         /*
  74          * Writable MAP_SHARED mapping: "clean" might indicate that the FS still
  75          * needs a real write-fault for writenotify
  76          * (see vma_wants_writenotify()). If "dirty", the assumption is that the
  77          * FS was already notified and we can simply mark the PTE writable
  78          * just like the write-fault handler would do.
  79          */
  80         return pte_dirty(pte);
  81 }
  82
  83 static long change_pte_range(struct mmu_gather *tlb,
  84                 struct vm_area_struct *vma, pmd_t *pmd, unsigned long addr,
  85                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
  86 {
  87         pte_t *pte, oldpte;
  88         spinlock_t *ptl;
  89         long pages = 0;
  90         int target_node = NUMA_NO_NODE;
  91         bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
  92         bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
  93         bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
  94
  95         tlb_change_page_size(tlb, PAGE_SIZE);
  96         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
  97         if (!pte)
  98                 return -EAGAIN;
  99
 100         /* Get target node for single threaded private VMAs */
 101         if (prot_numa && !(vma->vm_flags & VM_SHARED) &&
 102             atomic_read(&vma->vm_mm->mm_users) == 1)
 103                 target_node = numa_node_id();
 104
 105         flush_tlb_batched_pending(vma->vm_mm);
 106         arch_enter_lazy_mmu_mode();
 107         do {
 108                 oldpte = ptep_get(pte);
 109                 if (pte_present(oldpte)) {
 110                         pte_t ptent;
 111
 112                         /*
 113                          * Avoid trapping faults against the zero or KSM
 114                          * pages. See similar comment in change_huge_pmd.
 115                          */
 116                         if (prot_numa) {
 117                                 struct folio *folio;
 118                                 int nid;
 119                                 bool toptier;
 120
 121                                 /* Avoid TLB flush if possible */
 122                                 if (pte_protnone(oldpte))
 123                                         continue;
 124
 125                                 folio = vm_normal_folio(vma, addr, oldpte);
 126                                 if (!folio || folio_is_zone_device(folio) ||
 127                                     folio_test_ksm(folio))
 128                                         continue;
 129
 130                                 /* Also skip shared copy-on-write pages */
 131                                 if (is_cow_mapping(vma->vm_flags) &&
 132                                     folio_ref_count(folio) != 1)
 133                                         continue;
 134
 135                                 /*
 136                                  * While migration can move some dirty pages,
 137                                  * it cannot move them all from MIGRATE_ASYNC
 138                                  * context.
 139                                  */
 140                                 if (folio_is_file_lru(folio) &&
 141                                     folio_test_dirty(folio))
 142                                         continue;
 143
 144                                 /*
 145                                  * Don't mess with PTEs if page is already on the node
 146                                  * a single-threaded process is running on.
 147                                  */
 148                                 nid = folio_nid(folio);
 149                                 if (target_node == nid)
 150                                         continue;
 151                                 toptier = node_is_toptier(nid);
 152
 153                                 /*
 154                                  * Skip scanning top tier node if normal numa
 155                                  * balancing is disabled
 156                                  */
 157                                 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_NORMAL) &&
 158                                     toptier)
 159                                         continue;
 160                                 if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING &&
 161                                     !toptier)
 162                                         folio_xchg_access_time(folio,
 163                                                 jiffies_to_msecs(jiffies));
 164                         }
 165
 166                         oldpte = ptep_modify_prot_start(vma, addr, pte);
 167                         ptent = pte_modify(oldpte, newprot);
 168
 169                         if (uffd_wp)
 170                                 ptent = pte_mkuffd_wp(ptent);
 171                         else if (uffd_wp_resolve)
 172                                 ptent = pte_clear_uffd_wp(ptent);
 173
 174                         /*
 175                          * In some writable, shared mappings, we might want
 176                          * to catch actual write access -- see
 177                          * vma_wants_writenotify().
 178                          *
 179                          * In all writable, private mappings, we have to
 180                          * properly handle COW.
 181                          *
 182                          * In both cases, we can sometimes still change PTEs
 183                          * writable and avoid the write-fault handler, for
 184                          * example, if a PTE is already dirty and no other
 185                          * COW or special handling is required.
 186                          */
 187                         if ((cp_flags & MM_CP_TRY_CHANGE_WRITABLE) &&
 188                             !pte_write(ptent) &&
 189                             can_change_pte_writable(vma, addr, ptent))
 190                                 ptent = pte_mkwrite(ptent, vma);
 191
 192                         ptep_modify_prot_commit(vma, addr, pte, oldpte, ptent);
 193                         if (pte_needs_flush(oldpte, ptent))
 194                                 tlb_flush_pte_range(tlb, addr, PAGE_SIZE);
 195                         pages++;
 196                 } else if (is_swap_pte(oldpte)) {
 197                         swp_entry_t entry = pte_to_swp_entry(oldpte);
 198                         pte_t newpte;
 199
 200                         if (is_writable_migration_entry(entry)) {
 201                                 struct folio *folio = pfn_swap_entry_folio(entry);
 202
 203                                 /*
 204                                  * A protection check is difficult so
 205                                  * just be safe and disable write
 206                                  */
 207                                 if (folio_test_anon(folio))
 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                         } else if (is_writable_device_private_entry(entry)) {
 216                                 /*
 217                                  * We do not preserve soft-dirtiness. See
 218                                  * copy_nonpresent_pte() for explanation.
 219                                  */
 220                                 entry = make_readable_device_private_entry(
 221                                                         swp_offset(entry));
 222                                 newpte = swp_entry_to_pte(entry);
 223                                 if (pte_swp_uffd_wp(oldpte))
 224                                         newpte = pte_swp_mkuffd_wp(newpte);
 225                         } else if (is_writable_device_exclusive_entry(entry)) {
 226                                 entry = make_readable_device_exclusive_entry(
 227                                                         swp_offset(entry));
 228                                 newpte = swp_entry_to_pte(entry);
 229                                 if (pte_swp_soft_dirty(oldpte))
 230                                         newpte = pte_swp_mksoft_dirty(newpte);
 231                                 if (pte_swp_uffd_wp(oldpte))
 232                                         newpte = pte_swp_mkuffd_wp(newpte);
 233                         } else if (is_pte_marker_entry(entry)) {
 234                                 /*
 235                                  * Ignore error swap entries unconditionally,
 236                                  * because any access should sigbus anyway.
 237                                  */
 238                                 if (is_poisoned_swp_entry(entry))
 239                                         continue;
 240                                 /*
 241                                  * If this is uffd-wp pte marker and we'd like
 242                                  * to unprotect it, drop it; the next page
 243                                  * fault will trigger without uffd trapping.
 244                                  */
 245                                 if (uffd_wp_resolve) {
 246                                         pte_clear(vma->vm_mm, addr, pte);
 247                                         pages++;
 248                                 }
 249                                 continue;
 250                         } else {
 251                                 newpte = oldpte;
 252                         }
 253
 254                         if (uffd_wp)
 255                                 newpte = pte_swp_mkuffd_wp(newpte);
 256                         else if (uffd_wp_resolve)
 257                                 newpte = pte_swp_clear_uffd_wp(newpte);
 258
 259                         if (!pte_same(oldpte, newpte)) {
 260                                 set_pte_at(vma->vm_mm, addr, pte, newpte);
 261                                 pages++;
 262                         }
 263                 } else {
 264                         /* It must be an none page, or what else?.. */
 265                         WARN_ON_ONCE(!pte_none(oldpte));
 266
 267                         /*
 268                          * Nobody plays with any none ptes besides
 269                          * userfaultfd when applying the protections.
 270                          */
 271                         if (likely(!uffd_wp))
 272                                 continue;
 273
 274                         if (userfaultfd_wp_use_markers(vma)) {
 275                                 /*
 276                                  * For file-backed mem, we need to be able to
 277                                  * wr-protect a none pte, because even if the
 278                                  * pte is none, the page/swap cache could
 279                                  * exist.  Doing that by install a marker.
 280                                  */
 281                                 set_pte_at(vma->vm_mm, addr, pte,
 282                                            make_pte_marker(PTE_MARKER_UFFD_WP));
 283                                 pages++;
 284                         }
 285                 }
 286         } while (pte++, addr += PAGE_SIZE, addr != end);
 287         arch_leave_lazy_mmu_mode();
 288         pte_unmap_unlock(pte - 1, ptl);
 289
 290         return pages;
 291 }
 292
 293 /*
 294  * Return true if we want to split THPs into PTE mappings in change
 295  * protection procedure, false otherwise.
 296  */
 297 static inline bool
 298 pgtable_split_needed(struct vm_area_struct *vma, unsigned long cp_flags)
 299 {
 300         /*
 301          * pte markers only resides in pte level, if we need pte markers,
 302          * we need to split.  We cannot wr-protect shmem thp because file
 303          * thp is handled differently when split by erasing the pmd so far.
 304          */
 305         return (cp_flags & MM_CP_UFFD_WP) && !vma_is_anonymous(vma);
 306 }
 307
 308 /*
 309  * Return true if we want to populate pgtables in change protection
 310  * procedure, false otherwise
 311  */
 312 static inline bool
 313 pgtable_populate_needed(struct vm_area_struct *vma, unsigned long cp_flags)
 314 {
 315         /* If not within ioctl(UFFDIO_WRITEPROTECT), then don't bother */
 316         if (!(cp_flags & MM_CP_UFFD_WP))
 317                 return false;
 318
 319         /* Populate if the userfaultfd mode requires pte markers */
 320         return userfaultfd_wp_use_markers(vma);
 321 }
 322
 323 /*
 324  * Populate the pgtable underneath for whatever reason if requested.
 325  * When {pte|pmd|...}_alloc() failed we treat it the same way as pgtable
 326  * allocation failures during page faults by kicking OOM and returning
 327  * error.
 328  */
 329 #define  change_pmd_prepare(vma, pmd, cp_flags)                         \
 330         ({                                                              \
 331                 long err = 0;                                           \
 332                 if (unlikely(pgtable_populate_needed(vma, cp_flags))) { \
 333                         if (pte_alloc(vma->vm_mm, pmd))                 \
 334                                 err = -ENOMEM;                          \
 335                 }                                                       \
 336                 err;                                                    \
 337         })
 338
 339 /*
 340  * This is the general pud/p4d/pgd version of change_pmd_prepare(). We need to
 341  * have separate change_pmd_prepare() because pte_alloc() returns 0 on success,
 342  * while {pmd|pud|p4d}_alloc() returns the valid pointer on success.
 343  */
 344 #define  change_prepare(vma, high, low, addr, cp_flags)                 \
 345           ({                                                            \
 346                 long err = 0;                                           \
 347                 if (unlikely(pgtable_populate_needed(vma, cp_flags))) { \
 348                         low##_t *p = low##_alloc(vma->vm_mm, high, addr); \
 349                         if (p == NULL)                                  \
 350                                 err = -ENOMEM;                          \
 351                 }                                                       \
 352                 err;                                                    \
 353         })
 354
 355 static inline long change_pmd_range(struct mmu_gather *tlb,
 356                 struct vm_area_struct *vma, pud_t *pud, unsigned long addr,
 357                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
 358 {
 359         pmd_t *pmd;
 360         unsigned long next;
 361         long pages = 0;
 362         unsigned long nr_huge_updates = 0;
 363         struct mmu_notifier_range range;
 364
 365         range.start = 0;
 366
 367         pmd = pmd_offset(pud, addr);
 368         do {
 369                 long ret;
 370                 pmd_t _pmd;
 371 again:
 372                 next = pmd_addr_end(addr, end);
 373
 374                 ret = change_pmd_prepare(vma, pmd, cp_flags);
 375                 if (ret) {
 376                         pages = ret;
 377                         break;
 378                 }
 379
 380                 if (pmd_none(*pmd))
 381                         goto next;
 382
 383                 /* invoke the mmu notifier if the pmd is populated */
 384                 if (!range.start) {
 385                         mmu_notifier_range_init(&range,
 386                                 MMU_NOTIFY_PROTECTION_VMA, 0,
 387                                 vma->vm_mm, addr, end);
 388                         mmu_notifier_invalidate_range_start(&range);
 389                 }
 390
 391                 _pmd = pmdp_get_lockless(pmd);
 392                 if (is_swap_pmd(_pmd) || pmd_trans_huge(_pmd) || pmd_devmap(_pmd)) {
 393                         if ((next - addr != HPAGE_PMD_SIZE) ||
 394                             pgtable_split_needed(vma, cp_flags)) {
 395                                 __split_huge_pmd(vma, pmd, addr, false, NULL);
 396                                 /*
 397                                  * For file-backed, the pmd could have been
 398                                  * cleared; make sure pmd populated if
 399                                  * necessary, then fall-through to pte level.
 400                                  */
 401                                 ret = change_pmd_prepare(vma, pmd, cp_flags);
 402                                 if (ret) {
 403                                         pages = ret;
 404                                         break;
 405                                 }
 406                         } else {
 407                                 ret = change_huge_pmd(tlb, vma, pmd,
 408                                                 addr, newprot, cp_flags);
 409                                 if (ret) {
 410                                         if (ret == HPAGE_PMD_NR) {
 411                                                 pages += HPAGE_PMD_NR;
 412                                                 nr_huge_updates++;
 413                                         }
 414
 415                                         /* huge pmd was handled */
 416                                         goto next;
 417                                 }
 418                         }
 419                         /* fall through, the trans huge pmd just split */
 420                 }
 421
 422                 ret = change_pte_range(tlb, vma, pmd, addr, next, newprot,
 423                                        cp_flags);
 424                 if (ret < 0)
 425                         goto again;
 426                 pages += ret;
 427 next:
 428                 cond_resched();
 429         } while (pmd++, addr = next, addr != end);
 430
 431         if (range.start)
 432                 mmu_notifier_invalidate_range_end(&range);
 433
 434         if (nr_huge_updates)
 435                 count_vm_numa_events(NUMA_HUGE_PTE_UPDATES, nr_huge_updates);
 436         return pages;
 437 }
 438
 439 static inline long change_pud_range(struct mmu_gather *tlb,
 440                 struct vm_area_struct *vma, p4d_t *p4d, unsigned long addr,
 441                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
 442 {
 443         pud_t *pud;
 444         unsigned long next;
 445         long pages = 0, ret;
 446
 447         pud = pud_offset(p4d, addr);
 448         do {
 449                 next = pud_addr_end(addr, end);
 450                 ret = change_prepare(vma, pud, pmd, addr, cp_flags);
 451                 if (ret)
 452                         return ret;
 453                 if (pud_none_or_clear_bad(pud))
 454                         continue;
 455                 pages += change_pmd_range(tlb, vma, pud, addr, next, newprot,
 456                                           cp_flags);
 457         } while (pud++, addr = next, addr != end);
 458
 459         return pages;
 460 }
 461
 462 static inline long change_p4d_range(struct mmu_gather *tlb,
 463                 struct vm_area_struct *vma, pgd_t *pgd, unsigned long addr,
 464                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
 465 {
 466         p4d_t *p4d;
 467         unsigned long next;
 468         long pages = 0, ret;
 469
 470         p4d = p4d_offset(pgd, addr);
 471         do {
 472                 next = p4d_addr_end(addr, end);
 473                 ret = change_prepare(vma, p4d, pud, addr, cp_flags);
 474                 if (ret)
 475                         return ret;
 476                 if (p4d_none_or_clear_bad(p4d))
 477                         continue;
 478                 pages += change_pud_range(tlb, vma, p4d, addr, next, newprot,
 479                                           cp_flags);
 480         } while (p4d++, addr = next, addr != end);
 481
 482         return pages;
 483 }
 484
 485 static long change_protection_range(struct mmu_gather *tlb,
 486                 struct vm_area_struct *vma, unsigned long addr,
 487                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
 488 {
 489         struct mm_struct *mm = vma->vm_mm;
 490         pgd_t *pgd;
 491         unsigned long next;
 492         long pages = 0, ret;
 493
 494         BUG_ON(addr >= end);
 495         pgd = pgd_offset(mm, addr);
 496         tlb_start_vma(tlb, vma);
 497         do {
 498                 next = pgd_addr_end(addr, end);
 499                 ret = change_prepare(vma, pgd, p4d, addr, cp_flags);
 500                 if (ret) {
 501                         pages = ret;
 502                         break;
 503                 }
 504                 if (pgd_none_or_clear_bad(pgd))
 505                         continue;
 506                 pages += change_p4d_range(tlb, vma, pgd, addr, next, newprot,
 507                                           cp_flags);
 508         } while (pgd++, addr = next, addr != end);
 509
 510         tlb_end_vma(tlb, vma);
 511
 512         return pages;
 513 }
 514
 515 long change_protection(struct mmu_gather *tlb,
 516                        struct vm_area_struct *vma, unsigned long start,
 517                        unsigned long end, unsigned long cp_flags)
 518 {
 519         pgprot_t newprot = vma->vm_page_prot;
 520         long pages;
 521
 522         BUG_ON((cp_flags & MM_CP_UFFD_WP_ALL) == MM_CP_UFFD_WP_ALL);
 523
 524 #ifdef CONFIG_NUMA_BALANCING
 525         /*
 526          * Ordinary protection updates (mprotect, uffd-wp, softdirty tracking)
 527          * are expected to reflect their requirements via VMA flags such that
 528          * vma_set_page_prot() will adjust vma->vm_page_prot accordingly.
 529          */
 530         if (cp_flags & MM_CP_PROT_NUMA)
 531                 newprot = PAGE_NONE;
 532 #else
 533         WARN_ON_ONCE(cp_flags & MM_CP_PROT_NUMA);
 534 #endif
 535
 536         if (is_vm_hugetlb_page(vma))
 537                 pages = hugetlb_change_protection(vma, start, end, newprot,
 538                                                   cp_flags);
 539         else
 540                 pages = change_protection_range(tlb, vma, start, end, newprot,
 541                                                 cp_flags);
 542
 543         return pages;
 544 }
 545
 546 static int prot_none_pte_entry(pte_t *pte, unsigned long addr,
 547                                unsigned long next, struct mm_walk *walk)
 548 {
 549         return pfn_modify_allowed(pte_pfn(ptep_get(pte)),
 550                                   *(pgprot_t *)(walk->private)) ?
 551                 0 : -EACCES;
 552 }
 553
 554 static int prot_none_hugetlb_entry(pte_t *pte, unsigned long hmask,
 555                                    unsigned long addr, unsigned long next,
 556                                    struct mm_walk *walk)
 557 {
 558         return pfn_modify_allowed(pte_pfn(ptep_get(pte)),
 559                                   *(pgprot_t *)(walk->private)) ?
 560                 0 : -EACCES;
 561 }
 562
 563 static int prot_none_test(unsigned long addr, unsigned long next,
 564                           struct mm_walk *walk)
 565 {
 566         return 0;
 567 }
 568
 569 static const struct mm_walk_ops prot_none_walk_ops = {
 570         .pte_entry              = prot_none_pte_entry,
 571         .hugetlb_entry          = prot_none_hugetlb_entry,
 572         .test_walk              = prot_none_test,
 573         .walk_lock              = PGWALK_WRLOCK,
 574 };
 575
 576 int
 577 mprotect_fixup(struct vma_iterator *vmi, struct mmu_gather *tlb,
 578                struct vm_area_struct *vma, struct vm_area_struct **pprev,
 579                unsigned long start, unsigned long end, unsigned long newflags)
 580 {
 581         struct mm_struct *mm = vma->vm_mm;
 582         unsigned long oldflags = vma->vm_flags;
 583         long nrpages = (end - start) >> PAGE_SHIFT;
 584         unsigned int mm_cp_flags = 0;
 585         unsigned long charged = 0;
 586         int error;
 587
 588         if (newflags == oldflags) {
 589                 *pprev = vma;
 590                 return 0;
 591         }
 592
 593         /*
 594          * Do PROT_NONE PFN permission checks here when we can still
 595          * bail out without undoing a lot of state. This is a rather
 596          * uncommon case, so doesn't need to be very optimized.
 597          */
 598         if (arch_has_pfn_modify_check() &&
 599             (vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 600             (newflags & VM_ACCESS_FLAGS) == 0) {
 601                 pgprot_t new_pgprot = vm_get_page_prot(newflags);
 602
 603                 error = walk_page_range(current->mm, start, end,
 604                                 &prot_none_walk_ops, &new_pgprot);
 605                 if (error)
 606                         return error;
 607         }
 608
 609         /*
 610          * If we make a private mapping writable we increase our commit;
 611          * but (without finer accounting) cannot reduce our commit if we
 612          * make it unwritable again except in the anonymous case where no
 613          * anon_vma has yet to be assigned.
 614          *
 615          * hugetlb mapping were accounted for even if read-only so there is
 616          * no need to account for them here.
 617          */
 618         if (newflags & VM_WRITE) {
 619                 /* Check space limits when area turns into data. */
 620                 if (!may_expand_vm(mm, newflags, nrpages) &&
 621                                 may_expand_vm(mm, oldflags, nrpages))
 622                         return -ENOMEM;
 623                 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
 624                                                 VM_SHARED|VM_NORESERVE))) {
 625                         charged = nrpages;
 626                         if (security_vm_enough_memory_mm(mm, charged))
 627                                 return -ENOMEM;
 628                         newflags |= VM_ACCOUNT;
 629                 }
 630         } else if ((oldflags & VM_ACCOUNT) && vma_is_anonymous(vma) &&
 631                    !vma->anon_vma) {
 632                 newflags &= ~VM_ACCOUNT;
 633         }
 634
 635         vma = vma_modify_flags(vmi, *pprev, vma, start, end, newflags);
 636         if (IS_ERR(vma)) {
 637                 error = PTR_ERR(vma);
 638                 goto fail;
 639         }
 640
 641         *pprev = vma;
 642
 643         /*
 644          * vm_flags and vm_page_prot are protected by the mmap_lock
 645          * held in write mode.
 646          */
 647         vma_start_write(vma);
 648         vm_flags_reset(vma, newflags);
 649         if (vma_wants_manual_pte_write_upgrade(vma))
 650                 mm_cp_flags |= MM_CP_TRY_CHANGE_WRITABLE;
 651         vma_set_page_prot(vma);
 652
 653         change_protection(tlb, vma, start, end, mm_cp_flags);
 654
 655         if ((oldflags & VM_ACCOUNT) && !(newflags & VM_ACCOUNT))
 656                 vm_unacct_memory(nrpages);
 657
 658         /*
 659          * Private VM_LOCKED VMA becoming writable: trigger COW to avoid major
 660          * fault on access.
 661          */
 662         if ((oldflags & (VM_WRITE | VM_SHARED | VM_LOCKED)) == VM_LOCKED &&
 663                         (newflags & VM_WRITE)) {
 664                 populate_vma_page_range(vma, start, end, NULL);
 665         }
 666
 667         vm_stat_account(mm, oldflags, -nrpages);
 668         vm_stat_account(mm, newflags, nrpages);
 669         perf_event_mmap(vma);
 670         return 0;
 671
 672 fail:
 673         vm_unacct_memory(charged);
 674         return error;
 675 }
 676
 677 /*
 678  * pkey==-1 when doing a legacy mprotect()
 679  */
 680 static int do_mprotect_pkey(unsigned long start, size_t len,
 681                 unsigned long prot, int pkey)
 682 {
 683         unsigned long nstart, end, tmp, reqprot;
 684         struct vm_area_struct *vma, *prev;
 685         int error;
 686         const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
 687         const bool rier = (current->personality & READ_IMPLIES_EXEC) &&
 688                                 (prot & PROT_READ);
 689         struct mmu_gather tlb;
 690         struct vma_iterator vmi;
 691
 692         start = untagged_addr(start);
 693
 694         prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
 695         if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
 696                 return -EINVAL;
 697
 698         if (start & ~PAGE_MASK)
 699                 return -EINVAL;
 700         if (!len)
 701                 return 0;
 702         len = PAGE_ALIGN(len);
 703         end = start + len;
 704         if (end <= start)
 705                 return -ENOMEM;
 706         if (!arch_validate_prot(prot, start))
 707                 return -EINVAL;
 708
 709         reqprot = prot;
 710
 711         if (mmap_write_lock_killable(current->mm))
 712                 return -EINTR;
 713
 714         /*
 715          * If userspace did not allocate the pkey, do not let
 716          * them use it here.
 717          */
 718         error = -EINVAL;
 719         if ((pkey != -1) && !mm_pkey_is_allocated(current->mm, pkey))
 720                 goto out;
 721
 722         vma_iter_init(&vmi, current->mm, start);
 723         vma = vma_find(&vmi, end);
 724         error = -ENOMEM;
 725         if (!vma)
 726                 goto out;
 727
 728         if (unlikely(grows & PROT_GROWSDOWN)) {
 729                 if (vma->vm_start >= end)
 730                         goto out;
 731                 start = vma->vm_start;
 732                 error = -EINVAL;
 733                 if (!(vma->vm_flags & VM_GROWSDOWN))
 734                         goto out;
 735         } else {
 736                 if (vma->vm_start > start)
 737                         goto out;
 738                 if (unlikely(grows & PROT_GROWSUP)) {
 739                         end = vma->vm_end;
 740                         error = -EINVAL;
 741                         if (!(vma->vm_flags & VM_GROWSUP))
 742                                 goto out;
 743                 }
 744         }
 745
 746         prev = vma_prev(&vmi);
 747         if (start > vma->vm_start)
 748                 prev = vma;
 749
 750         tlb_gather_mmu(&tlb, current->mm);
 751         nstart = start;
 752         tmp = vma->vm_start;
 753         for_each_vma_range(vmi, vma, end) {
 754                 unsigned long mask_off_old_flags;
 755                 unsigned long newflags;
 756                 int new_vma_pkey;
 757
 758                 if (vma->vm_start != tmp) {
 759                         error = -ENOMEM;
 760                         break;
 761                 }
 762
 763                 /* Does the application expect PROT_READ to imply PROT_EXEC */
 764                 if (rier && (vma->vm_flags & VM_MAYEXEC))
 765                         prot |= PROT_EXEC;
 766
 767                 /*
 768                  * Each mprotect() call explicitly passes r/w/x permissions.
 769                  * If a permission is not passed to mprotect(), it must be
 770                  * cleared from the VMA.
 771                  */
 772                 mask_off_old_flags = VM_ACCESS_FLAGS | VM_FLAGS_CLEAR;
 773
 774                 new_vma_pkey = arch_override_mprotect_pkey(vma, prot, pkey);
 775                 newflags = calc_vm_prot_bits(prot, new_vma_pkey);
 776                 newflags |= (vma->vm_flags & ~mask_off_old_flags);
 777
 778                 /* newflags >> 4 shift VM_MAY% in place of VM_% */
 779                 if ((newflags & ~(newflags >> 4)) & VM_ACCESS_FLAGS) {
 780                         error = -EACCES;
 781                         break;
 782                 }
 783
 784                 if (map_deny_write_exec(vma, newflags)) {
 785                         error = -EACCES;
 786                         break;
 787                 }
 788
 789                 /* Allow architectures to sanity-check the new flags */
 790                 if (!arch_validate_flags(newflags)) {
 791                         error = -EINVAL;
 792                         break;
 793                 }
 794
 795                 error = security_file_mprotect(vma, reqprot, prot);
 796                 if (error)
 797                         break;
 798
 799                 tmp = vma->vm_end;
 800                 if (tmp > end)
 801                         tmp = end;
 802
 803                 if (vma->vm_ops && vma->vm_ops->mprotect) {
 804                         error = vma->vm_ops->mprotect(vma, nstart, tmp, newflags);
 805                         if (error)
 806                                 break;
 807                 }
 808
 809                 error = mprotect_fixup(&vmi, &tlb, vma, &prev, nstart, tmp, newflags);
 810                 if (error)
 811                         break;
 812
 813                 tmp = vma_iter_end(&vmi);
 814                 nstart = tmp;
 815                 prot = reqprot;
 816         }
 817         tlb_finish_mmu(&tlb);
 818
 819         if (!error && tmp < end)
 820                 error = -ENOMEM;
 821
 822 out:
 823         mmap_write_unlock(current->mm);
 824         return error;
 825 }
 826
 827 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
 828                 unsigned long, prot)
 829 {
 830         return do_mprotect_pkey(start, len, prot, -1);
 831 }
 832
 833 #ifdef CONFIG_ARCH_HAS_PKEYS
 834
 835 SYSCALL_DEFINE4(pkey_mprotect, unsigned long, start, size_t, len,
 836                 unsigned long, prot, int, pkey)
 837 {
 838         return do_mprotect_pkey(start, len, prot, pkey);
 839 }
 840
 841 SYSCALL_DEFINE2(pkey_alloc, unsigned long, flags, unsigned long, init_val)
 842 {
 843         int pkey;
 844         int ret;
 845
 846         /* No flags supported yet. */
 847         if (flags)
 848                 return -EINVAL;
 849         /* check for unsupported init values */
 850         if (init_val & ~PKEY_ACCESS_MASK)
 851                 return -EINVAL;
 852
 853         mmap_write_lock(current->mm);
 854         pkey = mm_pkey_alloc(current->mm);
 855
 856         ret = -ENOSPC;
 857         if (pkey == -1)
 858                 goto out;
 859
 860         ret = arch_set_user_pkey_access(current, pkey, init_val);
 861         if (ret) {
 862                 mm_pkey_free(current->mm, pkey);
 863                 goto out;
 864         }
 865         ret = pkey;
 866 out:
 867         mmap_write_unlock(current->mm);
 868         return ret;
 869 }
 870
 871 SYSCALL_DEFINE1(pkey_free, int, pkey)
 872 {
 873         int ret;
 874
 875         mmap_write_lock(current->mm);
 876         ret = mm_pkey_free(current->mm, pkey);
 877         mmap_write_unlock(current->mm);
 878
 879         /*
 880          * We could provide warnings or errors if any VMA still
 881          * has the pkey set here.
 882          */
 883         return ret;
 884 }
 885
 886 #endif /* CONFIG_ARCH_HAS_PKEYS */