mm/mprotect.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  *  mm/mprotect.c
   4  *
   5  *  (C) Copyright 1994 Linus Torvalds
   6  *  (C) Copyright 2002 Christoph Hellwig
   7  *
   8  *  Address space accounting code       <[email protected]>
   9  *  (C) Copyright 2002 Red Hat Inc, All Rights Reserved
  10  */
  11
  12 #include <linux/pagewalk.h>
  13 #include <linux/hugetlb.h>
  14 #include <linux/shm.h>
  15 #include <linux/mman.h>
  16 #include <linux/fs.h>
  17 #include <linux/highmem.h>
  18 #include <linux/security.h>
  19 #include <linux/mempolicy.h>
  20 #include <linux/personality.h>
  21 #include <linux/syscalls.h>
  22 #include <linux/swap.h>
  23 #include <linux/swapops.h>
  24 #include <linux/mmu_notifier.h>
  25 #include <linux/migrate.h>
  26 #include <linux/perf_event.h>
  27 #include <linux/pkeys.h>
  28 #include <linux/ksm.h>
  29 #include <linux/uaccess.h>
  30 #include <linux/mm_inline.h>
  31 #include <linux/pgtable.h>
  32 #include <linux/sched/sysctl.h>
  33 #include <linux/userfaultfd_k.h>
  34 #include <linux/memory-tiers.h>
  35 #include <uapi/linux/mman.h>
  36 #include <asm/cacheflush.h>
  37 #include <asm/mmu_context.h>
  38 #include <asm/tlbflush.h>
  39 #include <asm/tlb.h>
  40
  41 #include "internal.h"
  42
  43 bool can_change_pte_writable(struct vm_area_struct *vma, unsigned long addr,
  44                              pte_t pte)
  45 {
  46         struct page *page;
  47
  48         if (WARN_ON_ONCE(!(vma->vm_flags & VM_WRITE)))
  49                 return false;
  50
  51         /* Don't touch entries that are not even readable. */
  52         if (pte_protnone(pte))
  53                 return false;
  54
  55         /* Do we need write faults for softdirty tracking? */
  56         if (pte_needs_soft_dirty_wp(vma, pte))
  57                 return false;
  58
  59         /* Do we need write faults for uffd-wp tracking? */
  60         if (userfaultfd_pte_wp(vma, pte))
  61                 return false;
  62
  63         if (!(vma->vm_flags & VM_SHARED)) {
  64                 /*
  65                  * Writable MAP_PRIVATE mapping: We can only special-case on
  66                  * exclusive anonymous pages, because we know that our
  67                  * write-fault handler similarly would map them writable without
  68                  * any additional checks while holding the PT lock.
  69                  */
  70                 page = vm_normal_page(vma, addr, pte);
  71                 return page && PageAnon(page) && PageAnonExclusive(page);
  72         }
  73
  74         VM_WARN_ON_ONCE(is_zero_pfn(pte_pfn(pte)) && pte_dirty(pte));
  75
  76         /*
  77          * Writable MAP_SHARED mapping: "clean" might indicate that the FS still
  78          * needs a real write-fault for writenotify
  79          * (see vma_wants_writenotify()). If "dirty", the assumption is that the
  80          * FS was already notified and we can simply mark the PTE writable
  81          * just like the write-fault handler would do.
  82          */
  83         return pte_dirty(pte);
  84 }
  85
  86 static long change_pte_range(struct mmu_gather *tlb,
  87                 struct vm_area_struct *vma, pmd_t *pmd, unsigned long addr,
  88                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
  89 {
  90         pte_t *pte, oldpte;
  91         spinlock_t *ptl;
  92         long pages = 0;
  93         int target_node = NUMA_NO_NODE;
  94         bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
  95         bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
  96         bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
  97
  98         tlb_change_page_size(tlb, PAGE_SIZE);
  99         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
 100         if (!pte)
 101                 return -EAGAIN;
 102
 103         /* Get target node for single threaded private VMAs */
 104         if (prot_numa && !(vma->vm_flags & VM_SHARED) &&
 105             atomic_read(&vma->vm_mm->mm_users) == 1)
 106                 target_node = numa_node_id();
 107
 108         flush_tlb_batched_pending(vma->vm_mm);
 109         arch_enter_lazy_mmu_mode();
 110         do {
 111                 oldpte = ptep_get(pte);
 112                 if (pte_present(oldpte)) {
 113                         pte_t ptent;
 114
 115                         /*
 116                          * Avoid trapping faults against the zero or KSM
 117                          * pages. See similar comment in change_huge_pmd.
 118                          */
 119                         if (prot_numa) {
 120                                 struct folio *folio;
 121                                 int nid;
 122                                 bool toptier;
 123
 124                                 /* Avoid TLB flush if possible */
 125                                 if (pte_protnone(oldpte))
 126                                         continue;
 127
 128                                 folio = vm_normal_folio(vma, addr, oldpte);
 129                                 if (!folio || folio_is_zone_device(folio) ||
 130                                     folio_test_ksm(folio))
 131                                         continue;
 132
 133                                 /* Also skip shared copy-on-write pages */
 134                                 if (is_cow_mapping(vma->vm_flags) &&
 135                                     (folio_maybe_dma_pinned(folio) ||
 136                                      folio_likely_mapped_shared(folio)))
 137                                         continue;
 138
 139                                 /*
 140                                  * While migration can move some dirty pages,
 141                                  * it cannot move them all from MIGRATE_ASYNC
 142                                  * context.
 143                                  */
 144                                 if (folio_is_file_lru(folio) &&
 145                                     folio_test_dirty(folio))
 146                                         continue;
 147
 148                                 /*
 149                                  * Don't mess with PTEs if page is already on the node
 150                                  * a single-threaded process is running on.
 151                                  */
 152                                 nid = folio_nid(folio);
 153                                 if (target_node == nid)
 154                                         continue;
 155                                 toptier = node_is_toptier(nid);
 156
 157                                 /*
 158                                  * Skip scanning top tier node if normal numa
 159                                  * balancing is disabled
 160                                  */
 161                                 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_NORMAL) &&
 162                                     toptier)
 163                                         continue;
 164                                 if (folio_use_access_time(folio))
 165                                         folio_xchg_access_time(folio,
 166                                                 jiffies_to_msecs(jiffies));
 167                         }
 168
 169                         oldpte = ptep_modify_prot_start(vma, addr, pte);
 170                         ptent = pte_modify(oldpte, newprot);
 171
 172                         if (uffd_wp)
 173                                 ptent = pte_mkuffd_wp(ptent);
 174                         else if (uffd_wp_resolve)
 175                                 ptent = pte_clear_uffd_wp(ptent);
 176
 177                         /*
 178                          * In some writable, shared mappings, we might want
 179                          * to catch actual write access -- see
 180                          * vma_wants_writenotify().
 181                          *
 182                          * In all writable, private mappings, we have to
 183                          * properly handle COW.
 184                          *
 185                          * In both cases, we can sometimes still change PTEs
 186                          * writable and avoid the write-fault handler, for
 187                          * example, if a PTE is already dirty and no other
 188                          * COW or special handling is required.
 189                          */
 190                         if ((cp_flags & MM_CP_TRY_CHANGE_WRITABLE) &&
 191                             !pte_write(ptent) &&
 192                             can_change_pte_writable(vma, addr, ptent))
 193                                 ptent = pte_mkwrite(ptent, vma);
 194
 195                         ptep_modify_prot_commit(vma, addr, pte, oldpte, ptent);
 196                         if (pte_needs_flush(oldpte, ptent))
 197                                 tlb_flush_pte_range(tlb, addr, PAGE_SIZE);
 198                         pages++;
 199                 } else if (is_swap_pte(oldpte)) {
 200                         swp_entry_t entry = pte_to_swp_entry(oldpte);
 201                         pte_t newpte;
 202
 203                         if (is_writable_migration_entry(entry)) {
 204                                 struct folio *folio = pfn_swap_entry_folio(entry);
 205
 206                                 /*
 207                                  * A protection check is difficult so
 208                                  * just be safe and disable write
 209                                  */
 210                                 if (folio_test_anon(folio))
 211                                         entry = make_readable_exclusive_migration_entry(
 212                                                              swp_offset(entry));
 213                                 else
 214                                         entry = make_readable_migration_entry(swp_offset(entry));
 215                                 newpte = swp_entry_to_pte(entry);
 216                                 if (pte_swp_soft_dirty(oldpte))
 217                                         newpte = pte_swp_mksoft_dirty(newpte);
 218                         } else if (is_writable_device_private_entry(entry)) {
 219                                 /*
 220                                  * We do not preserve soft-dirtiness. See
 221                                  * copy_nonpresent_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 (is_pte_marker_entry(entry)) {
 237                                 /*
 238                                  * Ignore error swap entries unconditionally,
 239                                  * because any access should sigbus anyway.
 240                                  */
 241                                 if (is_poisoned_swp_entry(entry))
 242                                         continue;
 243                                 /*
 244                                  * If this is uffd-wp pte marker and we'd like
 245                                  * to unprotect it, drop it; the next page
 246                                  * fault will trigger without uffd trapping.
 247                                  */
 248                                 if (uffd_wp_resolve) {
 249                                         pte_clear(vma->vm_mm, addr, pte);
 250                                         pages++;
 251                                 }
 252                                 continue;
 253                         } else {
 254                                 newpte = oldpte;
 255                         }
 256
 257                         if (uffd_wp)
 258                                 newpte = pte_swp_mkuffd_wp(newpte);
 259                         else if (uffd_wp_resolve)
 260                                 newpte = pte_swp_clear_uffd_wp(newpte);
 261
 262                         if (!pte_same(oldpte, newpte)) {
 263                                 set_pte_at(vma->vm_mm, addr, pte, newpte);
 264                                 pages++;
 265                         }
 266                 } else {
 267                         /* It must be an none page, or what else?.. */
 268                         WARN_ON_ONCE(!pte_none(oldpte));
 269
 270                         /*
 271                          * Nobody plays with any none ptes besides
 272                          * userfaultfd when applying the protections.
 273                          */
 274                         if (likely(!uffd_wp))
 275                                 continue;
 276
 277                         if (userfaultfd_wp_use_markers(vma)) {
 278                                 /*
 279                                  * For file-backed mem, we need to be able to
 280                                  * wr-protect a none pte, because even if the
 281                                  * pte is none, the page/swap cache could
 282                                  * exist.  Doing that by install a marker.
 283                                  */
 284                                 set_pte_at(vma->vm_mm, addr, pte,
 285                                            make_pte_marker(PTE_MARKER_UFFD_WP));
 286                                 pages++;
 287                         }
 288                 }
 289         } while (pte++, addr += PAGE_SIZE, addr != end);
 290         arch_leave_lazy_mmu_mode();
 291         pte_unmap_unlock(pte - 1, ptl);
 292
 293         return pages;
 294 }
 295
 296 /*
 297  * Return true if we want to split THPs into PTE mappings in change
 298  * protection procedure, false otherwise.
 299  */
 300 static inline bool
 301 pgtable_split_needed(struct vm_area_struct *vma, unsigned long cp_flags)
 302 {
 303         /*
 304          * pte markers only resides in pte level, if we need pte markers,
 305          * we need to split.  For example, we cannot wr-protect a file thp
 306          * (e.g. 2M shmem) because file thp is handled differently when
 307          * split by erasing the pmd so far.
 308          */
 309         return (cp_flags & MM_CP_UFFD_WP) && !vma_is_anonymous(vma);
 310 }
 311
 312 /*
 313  * Return true if we want to populate pgtables in change protection
 314  * procedure, false otherwise
 315  */
 316 static inline bool
 317 pgtable_populate_needed(struct vm_area_struct *vma, unsigned long cp_flags)
 318 {
 319         /* If not within ioctl(UFFDIO_WRITEPROTECT), then don't bother */
 320         if (!(cp_flags & MM_CP_UFFD_WP))
 321                 return false;
 322
 323         /* Populate if the userfaultfd mode requires pte markers */
 324         return userfaultfd_wp_use_markers(vma);
 325 }
 326
 327 /*
 328  * Populate the pgtable underneath for whatever reason if requested.
 329  * When {pte|pmd|...}_alloc() failed we treat it the same way as pgtable
 330  * allocation failures during page faults by kicking OOM and returning
 331  * error.
 332  */
 333 #define  change_pmd_prepare(vma, pmd, cp_flags)                         \
 334         ({                                                              \
 335                 long err = 0;                                           \
 336                 if (unlikely(pgtable_populate_needed(vma, cp_flags))) { \
 337                         if (pte_alloc(vma->vm_mm, pmd))                 \
 338                                 err = -ENOMEM;                          \
 339                 }                                                       \
 340                 err;                                                    \
 341         })
 342
 343 /*
 344  * This is the general pud/p4d/pgd version of change_pmd_prepare(). We need to
 345  * have separate change_pmd_prepare() because pte_alloc() returns 0 on success,
 346  * while {pmd|pud|p4d}_alloc() returns the valid pointer on success.
 347  */
 348 #define  change_prepare(vma, high, low, addr, cp_flags)                 \
 349           ({                                                            \
 350                 long err = 0;                                           \
 351                 if (unlikely(pgtable_populate_needed(vma, cp_flags))) { \
 352                         low##_t *p = low##_alloc(vma->vm_mm, high, addr); \
 353                         if (p == NULL)                                  \
 354                                 err = -ENOMEM;                          \
 355                 }                                                       \
 356                 err;                                                    \
 357         })
 358
 359 static inline long change_pmd_range(struct mmu_gather *tlb,
 360                 struct vm_area_struct *vma, pud_t *pud, unsigned long addr,
 361                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
 362 {
 363         pmd_t *pmd;
 364         unsigned long next;
 365         long pages = 0;
 366         unsigned long nr_huge_updates = 0;
 367
 368         pmd = pmd_offset(pud, addr);
 369         do {
 370                 long ret;
 371                 pmd_t _pmd;
 372 again:
 373                 next = pmd_addr_end(addr, end);
 374
 375                 ret = change_pmd_prepare(vma, pmd, cp_flags);
 376                 if (ret) {
 377                         pages = ret;
 378                         break;
 379                 }
 380
 381                 if (pmd_none(*pmd))
 382                         goto next;
 383
 384                 _pmd = pmdp_get_lockless(pmd);
 385                 if (is_swap_pmd(_pmd) || pmd_trans_huge(_pmd) || pmd_devmap(_pmd)) {
 386                         if ((next - addr != HPAGE_PMD_SIZE) ||
 387                             pgtable_split_needed(vma, cp_flags)) {
 388                                 __split_huge_pmd(vma, pmd, addr, false, NULL);
 389                                 /*
 390                                  * For file-backed, the pmd could have been
 391                                  * cleared; make sure pmd populated if
 392                                  * necessary, then fall-through to pte level.
 393                                  */
 394                                 ret = change_pmd_prepare(vma, pmd, cp_flags);
 395                                 if (ret) {
 396                                         pages = ret;
 397                                         break;
 398                                 }
 399                         } else {
 400                                 ret = change_huge_pmd(tlb, vma, pmd,
 401                                                 addr, newprot, cp_flags);
 402                                 if (ret) {
 403                                         if (ret == HPAGE_PMD_NR) {
 404                                                 pages += HPAGE_PMD_NR;
 405                                                 nr_huge_updates++;
 406                                         }
 407
 408                                         /* huge pmd was handled */
 409                                         goto next;
 410                                 }
 411                         }
 412                         /* fall through, the trans huge pmd just split */
 413                 }
 414
 415                 ret = change_pte_range(tlb, vma, pmd, addr, next, newprot,
 416                                        cp_flags);
 417                 if (ret < 0)
 418                         goto again;
 419                 pages += ret;
 420 next:
 421                 cond_resched();
 422         } while (pmd++, addr = next, addr != end);
 423
 424         if (nr_huge_updates)
 425                 count_vm_numa_events(NUMA_HUGE_PTE_UPDATES, nr_huge_updates);
 426         return pages;
 427 }
 428
 429 static inline long change_pud_range(struct mmu_gather *tlb,
 430                 struct vm_area_struct *vma, p4d_t *p4d, unsigned long addr,
 431                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
 432 {
 433         struct mmu_notifier_range range;
 434         pud_t *pudp, pud;
 435         unsigned long next;
 436         long pages = 0, ret;
 437
 438         range.start = 0;
 439
 440         pudp = pud_offset(p4d, addr);
 441         do {
 442 again:
 443                 next = pud_addr_end(addr, end);
 444                 ret = change_prepare(vma, pudp, pmd, addr, cp_flags);
 445                 if (ret) {
 446                         pages = ret;
 447                         break;
 448                 }
 449
 450                 pud = READ_ONCE(*pudp);
 451                 if (pud_none(pud))
 452                         continue;
 453
 454                 if (!range.start) {
 455                         mmu_notifier_range_init(&range,
 456                                                 MMU_NOTIFY_PROTECTION_VMA, 0,
 457                                                 vma->vm_mm, addr, end);
 458                         mmu_notifier_invalidate_range_start(&range);
 459                 }
 460
 461                 if (pud_leaf(pud)) {
 462                         if ((next - addr != PUD_SIZE) ||
 463                             pgtable_split_needed(vma, cp_flags)) {
 464                                 __split_huge_pud(vma, pudp, addr);
 465                                 goto again;
 466                         } else {
 467                                 ret = change_huge_pud(tlb, vma, pudp,
 468                                                       addr, newprot, cp_flags);
 469                                 if (ret == 0)
 470                                         goto again;
 471                                 /* huge pud was handled */
 472                                 if (ret == HPAGE_PUD_NR)
 473                                         pages += HPAGE_PUD_NR;
 474                                 continue;
 475                         }
 476                 }
 477
 478                 pages += change_pmd_range(tlb, vma, pudp, addr, next, newprot,
 479                                           cp_flags);
 480         } while (pudp++, addr = next, addr != end);
 481
 482         if (range.start)
 483                 mmu_notifier_invalidate_range_end(&range);
 484
 485         return pages;
 486 }
 487
 488 static inline long change_p4d_range(struct mmu_gather *tlb,
 489                 struct vm_area_struct *vma, pgd_t *pgd, unsigned long addr,
 490                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
 491 {
 492         p4d_t *p4d;
 493         unsigned long next;
 494         long pages = 0, ret;
 495
 496         p4d = p4d_offset(pgd, addr);
 497         do {
 498                 next = p4d_addr_end(addr, end);
 499                 ret = change_prepare(vma, p4d, pud, addr, cp_flags);
 500                 if (ret)
 501                         return ret;
 502                 if (p4d_none_or_clear_bad(p4d))
 503                         continue;
 504                 pages += change_pud_range(tlb, vma, p4d, addr, next, newprot,
 505                                           cp_flags);
 506         } while (p4d++, addr = next, addr != end);
 507
 508         return pages;
 509 }
 510
 511 static long change_protection_range(struct mmu_gather *tlb,
 512                 struct vm_area_struct *vma, unsigned long addr,
 513                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
 514 {
 515         struct mm_struct *mm = vma->vm_mm;
 516         pgd_t *pgd;
 517         unsigned long next;
 518         long pages = 0, ret;
 519
 520         BUG_ON(addr >= end);
 521         pgd = pgd_offset(mm, addr);
 522         tlb_start_vma(tlb, vma);
 523         do {
 524                 next = pgd_addr_end(addr, end);
 525                 ret = change_prepare(vma, pgd, p4d, addr, cp_flags);
 526                 if (ret) {
 527                         pages = ret;
 528                         break;
 529                 }
 530                 if (pgd_none_or_clear_bad(pgd))
 531                         continue;
 532                 pages += change_p4d_range(tlb, vma, pgd, addr, next, newprot,
 533                                           cp_flags);
 534         } while (pgd++, addr = next, addr != end);
 535
 536         tlb_end_vma(tlb, vma);
 537
 538         return pages;
 539 }
 540
 541 long change_protection(struct mmu_gather *tlb,
 542                        struct vm_area_struct *vma, unsigned long start,
 543                        unsigned long end, unsigned long cp_flags)
 544 {
 545         pgprot_t newprot = vma->vm_page_prot;
 546         long pages;
 547
 548         BUG_ON((cp_flags & MM_CP_UFFD_WP_ALL) == MM_CP_UFFD_WP_ALL);
 549
 550 #ifdef CONFIG_NUMA_BALANCING
 551         /*
 552          * Ordinary protection updates (mprotect, uffd-wp, softdirty tracking)
 553          * are expected to reflect their requirements via VMA flags such that
 554          * vma_set_page_prot() will adjust vma->vm_page_prot accordingly.
 555          */
 556         if (cp_flags & MM_CP_PROT_NUMA)
 557                 newprot = PAGE_NONE;
 558 #else
 559         WARN_ON_ONCE(cp_flags & MM_CP_PROT_NUMA);
 560 #endif
 561
 562         if (is_vm_hugetlb_page(vma))
 563                 pages = hugetlb_change_protection(vma, start, end, newprot,
 564                                                   cp_flags);
 565         else
 566                 pages = change_protection_range(tlb, vma, start, end, newprot,
 567                                                 cp_flags);
 568
 569         return pages;
 570 }
 571
 572 static int prot_none_pte_entry(pte_t *pte, unsigned long addr,
 573                                unsigned long next, struct mm_walk *walk)
 574 {
 575         return pfn_modify_allowed(pte_pfn(ptep_get(pte)),
 576                                   *(pgprot_t *)(walk->private)) ?
 577                 0 : -EACCES;
 578 }
 579
 580 static int prot_none_hugetlb_entry(pte_t *pte, unsigned long hmask,
 581                                    unsigned long addr, unsigned long next,
 582                                    struct mm_walk *walk)
 583 {
 584         return pfn_modify_allowed(pte_pfn(ptep_get(pte)),
 585                                   *(pgprot_t *)(walk->private)) ?
 586                 0 : -EACCES;
 587 }
 588
 589 static int prot_none_test(unsigned long addr, unsigned long next,
 590                           struct mm_walk *walk)
 591 {
 592         return 0;
 593 }
 594
 595 static const struct mm_walk_ops prot_none_walk_ops = {
 596         .pte_entry              = prot_none_pte_entry,
 597         .hugetlb_entry          = prot_none_hugetlb_entry,
 598         .test_walk              = prot_none_test,
 599         .walk_lock              = PGWALK_WRLOCK,
 600 };
 601
 602 int
 603 mprotect_fixup(struct vma_iterator *vmi, struct mmu_gather *tlb,
 604                struct vm_area_struct *vma, struct vm_area_struct **pprev,
 605                unsigned long start, unsigned long end, unsigned long newflags)
 606 {
 607         struct mm_struct *mm = vma->vm_mm;
 608         unsigned long oldflags = vma->vm_flags;
 609         long nrpages = (end - start) >> PAGE_SHIFT;
 610         unsigned int mm_cp_flags = 0;
 611         unsigned long charged = 0;
 612         int error;
 613
 614         if (newflags == oldflags) {
 615                 *pprev = vma;
 616                 return 0;
 617         }
 618
 619         /*
 620          * Do PROT_NONE PFN permission checks here when we can still
 621          * bail out without undoing a lot of state. This is a rather
 622          * uncommon case, so doesn't need to be very optimized.
 623          */
 624         if (arch_has_pfn_modify_check() &&
 625             (vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 626             (newflags & VM_ACCESS_FLAGS) == 0) {
 627                 pgprot_t new_pgprot = vm_get_page_prot(newflags);
 628
 629                 error = walk_page_range(current->mm, start, end,
 630                                 &prot_none_walk_ops, &new_pgprot);
 631                 if (error)
 632                         return error;
 633         }
 634
 635         /*
 636          * If we make a private mapping writable we increase our commit;
 637          * but (without finer accounting) cannot reduce our commit if we
 638          * make it unwritable again except in the anonymous case where no
 639          * anon_vma has yet to be assigned.
 640          *
 641          * hugetlb mapping were accounted for even if read-only so there is
 642          * no need to account for them here.
 643          */
 644         if (newflags & VM_WRITE) {
 645                 /* Check space limits when area turns into data. */
 646                 if (!may_expand_vm(mm, newflags, nrpages) &&
 647                                 may_expand_vm(mm, oldflags, nrpages))
 648                         return -ENOMEM;
 649                 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
 650                                                 VM_SHARED|VM_NORESERVE))) {
 651                         charged = nrpages;
 652                         if (security_vm_enough_memory_mm(mm, charged))
 653                                 return -ENOMEM;
 654                         newflags |= VM_ACCOUNT;
 655                 }
 656         } else if ((oldflags & VM_ACCOUNT) && vma_is_anonymous(vma) &&
 657                    !vma->anon_vma) {
 658                 newflags &= ~VM_ACCOUNT;
 659         }
 660
 661         vma = vma_modify_flags(vmi, *pprev, vma, start, end, newflags);
 662         if (IS_ERR(vma)) {
 663                 error = PTR_ERR(vma);
 664                 goto fail;
 665         }
 666
 667         *pprev = vma;
 668
 669         /*
 670          * vm_flags and vm_page_prot are protected by the mmap_lock
 671          * held in write mode.
 672          */
 673         vma_start_write(vma);
 674         vm_flags_reset(vma, newflags);
 675         if (vma_wants_manual_pte_write_upgrade(vma))
 676                 mm_cp_flags |= MM_CP_TRY_CHANGE_WRITABLE;
 677         vma_set_page_prot(vma);
 678
 679         change_protection(tlb, vma, start, end, mm_cp_flags);
 680
 681         if ((oldflags & VM_ACCOUNT) && !(newflags & VM_ACCOUNT))
 682                 vm_unacct_memory(nrpages);
 683
 684         /*
 685          * Private VM_LOCKED VMA becoming writable: trigger COW to avoid major
 686          * fault on access.
 687          */
 688         if ((oldflags & (VM_WRITE | VM_SHARED | VM_LOCKED)) == VM_LOCKED &&
 689                         (newflags & VM_WRITE)) {
 690                 populate_vma_page_range(vma, start, end, NULL);
 691         }
 692
 693         vm_stat_account(mm, oldflags, -nrpages);
 694         vm_stat_account(mm, newflags, nrpages);
 695         perf_event_mmap(vma);
 696         return 0;
 697
 698 fail:
 699         vm_unacct_memory(charged);
 700         return error;
 701 }
 702
 703 /*
 704  * pkey==-1 when doing a legacy mprotect()
 705  */
 706 static int do_mprotect_pkey(unsigned long start, size_t len,
 707                 unsigned long prot, int pkey)
 708 {
 709         unsigned long nstart, end, tmp, reqprot;
 710         struct vm_area_struct *vma, *prev;
 711         int error;
 712         const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
 713         const bool rier = (current->personality & READ_IMPLIES_EXEC) &&
 714                                 (prot & PROT_READ);
 715         struct mmu_gather tlb;
 716         struct vma_iterator vmi;
 717
 718         start = untagged_addr(start);
 719
 720         prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
 721         if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
 722                 return -EINVAL;
 723
 724         if (start & ~PAGE_MASK)
 725                 return -EINVAL;
 726         if (!len)
 727                 return 0;
 728         len = PAGE_ALIGN(len);
 729         end = start + len;
 730         if (end <= start)
 731                 return -ENOMEM;
 732         if (!arch_validate_prot(prot, start))
 733                 return -EINVAL;
 734
 735         reqprot = prot;
 736
 737         if (mmap_write_lock_killable(current->mm))
 738                 return -EINTR;
 739
 740         /*
 741          * If userspace did not allocate the pkey, do not let
 742          * them use it here.
 743          */
 744         error = -EINVAL;
 745         if ((pkey != -1) && !mm_pkey_is_allocated(current->mm, pkey))
 746                 goto out;
 747
 748         vma_iter_init(&vmi, current->mm, start);
 749         vma = vma_find(&vmi, end);
 750         error = -ENOMEM;
 751         if (!vma)
 752                 goto out;
 753
 754         if (unlikely(grows & PROT_GROWSDOWN)) {
 755                 if (vma->vm_start >= end)
 756                         goto out;
 757                 start = vma->vm_start;
 758                 error = -EINVAL;
 759                 if (!(vma->vm_flags & VM_GROWSDOWN))
 760                         goto out;
 761         } else {
 762                 if (vma->vm_start > start)
 763                         goto out;
 764                 if (unlikely(grows & PROT_GROWSUP)) {
 765                         end = vma->vm_end;
 766                         error = -EINVAL;
 767                         if (!(vma->vm_flags & VM_GROWSUP))
 768                                 goto out;
 769                 }
 770         }
 771
 772         /*
 773          * checking if memory is sealed.
 774          * can_modify_mm assumes we have acquired the lock on MM.
 775          */
 776         if (unlikely(!can_modify_mm(current->mm, start, end))) {
 777                 error = -EPERM;
 778                 goto out;
 779         }
 780
 781         prev = vma_prev(&vmi);
 782         if (start > vma->vm_start)
 783                 prev = vma;
 784
 785         tlb_gather_mmu(&tlb, current->mm);
 786         nstart = start;
 787         tmp = vma->vm_start;
 788         for_each_vma_range(vmi, vma, end) {
 789                 unsigned long mask_off_old_flags;
 790                 unsigned long newflags;
 791                 int new_vma_pkey;
 792
 793                 if (vma->vm_start != tmp) {
 794                         error = -ENOMEM;
 795                         break;
 796                 }
 797
 798                 /* Does the application expect PROT_READ to imply PROT_EXEC */
 799                 if (rier && (vma->vm_flags & VM_MAYEXEC))
 800                         prot |= PROT_EXEC;
 801
 802                 /*
 803                  * Each mprotect() call explicitly passes r/w/x permissions.
 804                  * If a permission is not passed to mprotect(), it must be
 805                  * cleared from the VMA.
 806                  */
 807                 mask_off_old_flags = VM_ACCESS_FLAGS | VM_FLAGS_CLEAR;
 808
 809                 new_vma_pkey = arch_override_mprotect_pkey(vma, prot, pkey);
 810                 newflags = calc_vm_prot_bits(prot, new_vma_pkey);
 811                 newflags |= (vma->vm_flags & ~mask_off_old_flags);
 812
 813                 /* newflags >> 4 shift VM_MAY% in place of VM_% */
 814                 if ((newflags & ~(newflags >> 4)) & VM_ACCESS_FLAGS) {
 815                         error = -EACCES;
 816                         break;
 817                 }
 818
 819                 if (map_deny_write_exec(vma, newflags)) {
 820                         error = -EACCES;
 821                         break;
 822                 }
 823
 824                 /* Allow architectures to sanity-check the new flags */
 825                 if (!arch_validate_flags(newflags)) {
 826                         error = -EINVAL;
 827                         break;
 828                 }
 829
 830                 error = security_file_mprotect(vma, reqprot, prot);
 831                 if (error)
 832                         break;
 833
 834                 tmp = vma->vm_end;
 835                 if (tmp > end)
 836                         tmp = end;
 837
 838                 if (vma->vm_ops && vma->vm_ops->mprotect) {
 839                         error = vma->vm_ops->mprotect(vma, nstart, tmp, newflags);
 840                         if (error)
 841                                 break;
 842                 }
 843
 844                 error = mprotect_fixup(&vmi, &tlb, vma, &prev, nstart, tmp, newflags);
 845                 if (error)
 846                         break;
 847
 848                 tmp = vma_iter_end(&vmi);
 849                 nstart = tmp;
 850                 prot = reqprot;
 851         }
 852         tlb_finish_mmu(&tlb);
 853
 854         if (!error && tmp < end)
 855                 error = -ENOMEM;
 856
 857 out:
 858         mmap_write_unlock(current->mm);
 859         return error;
 860 }
 861
 862 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
 863                 unsigned long, prot)
 864 {
 865         return do_mprotect_pkey(start, len, prot, -1);
 866 }
 867
 868 #ifdef CONFIG_ARCH_HAS_PKEYS
 869
 870 SYSCALL_DEFINE4(pkey_mprotect, unsigned long, start, size_t, len,
 871                 unsigned long, prot, int, pkey)
 872 {
 873         return do_mprotect_pkey(start, len, prot, pkey);
 874 }
 875
 876 SYSCALL_DEFINE2(pkey_alloc, unsigned long, flags, unsigned long, init_val)
 877 {
 878         int pkey;
 879         int ret;
 880
 881         /* No flags supported yet. */
 882         if (flags)
 883                 return -EINVAL;
 884         /* check for unsupported init values */
 885         if (init_val & ~PKEY_ACCESS_MASK)
 886                 return -EINVAL;
 887
 888         mmap_write_lock(current->mm);
 889         pkey = mm_pkey_alloc(current->mm);
 890
 891         ret = -ENOSPC;
 892         if (pkey == -1)
 893                 goto out;
 894
 895         ret = arch_set_user_pkey_access(current, pkey, init_val);
 896         if (ret) {
 897                 mm_pkey_free(current->mm, pkey);
 898                 goto out;
 899         }
 900         ret = pkey;
 901 out:
 902         mmap_write_unlock(current->mm);
 903         return ret;
 904 }
 905
 906 SYSCALL_DEFINE1(pkey_free, int, pkey)
 907 {
 908         int ret;
 909
 910         mmap_write_lock(current->mm);
 911         ret = mm_pkey_free(current->mm, pkey);
 912         mmap_write_unlock(current->mm);
 913
 914         /*
 915          * We could provide warnings or errors if any VMA still
 916          * has the pkey set here.
 917          */
 918         return ret;
 919 }
 920
 921 #endif /* CONFIG_ARCH_HAS_PKEYS */