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 <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 static inline bool can_change_pte_writable(struct vm_area_struct *vma,
  43                                            unsigned long addr, pte_t pte)
  44 {
  45         struct page *page;
  46
  47         VM_BUG_ON(!(vma->vm_flags & VM_WRITE) || pte_write(pte));
  48
  49         if (pte_protnone(pte) || !pte_dirty(pte))
  50                 return false;
  51
  52         /* Do we need write faults for softdirty tracking? */
  53         if (vma_soft_dirty_enabled(vma) && !pte_soft_dirty(pte))
  54                 return false;
  55
  56         /* Do we need write faults for uffd-wp tracking? */
  57         if (userfaultfd_pte_wp(vma, pte))
  58                 return false;
  59
  60         if (!(vma->vm_flags & VM_SHARED)) {
  61                 /*
  62                  * We can only special-case on exclusive anonymous pages,
  63                  * because we know that our write-fault handler similarly would
  64                  * map them writable without any additional checks while holding
  65                  * the PT lock.
  66                  */
  67                 page = vm_normal_page(vma, addr, pte);
  68                 if (!page || !PageAnon(page) || !PageAnonExclusive(page))
  69                         return false;
  70         }
  71
  72         return true;
  73 }
  74
  75 static unsigned long change_pte_range(struct mmu_gather *tlb,
  76                 struct vm_area_struct *vma, pmd_t *pmd, unsigned long addr,
  77                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
  78 {
  79         pte_t *pte, oldpte;
  80         spinlock_t *ptl;
  81         unsigned long pages = 0;
  82         int target_node = NUMA_NO_NODE;
  83         bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
  84         bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
  85         bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
  86
  87         tlb_change_page_size(tlb, PAGE_SIZE);
  88
  89         /*
  90          * Can be called with only the mmap_lock for reading by
  91          * prot_numa so we must check the pmd isn't constantly
  92          * changing from under us from pmd_none to pmd_trans_huge
  93          * and/or the other way around.
  94          */
  95         if (pmd_trans_unstable(pmd))
  96                 return 0;
  97
  98         /*
  99          * The pmd points to a regular pte so the pmd can't change
 100          * from under us even if the mmap_lock is only hold for
 101          * reading.
 102          */
 103         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
 104
 105         /* Get target node for single threaded private VMAs */
 106         if (prot_numa && !(vma->vm_flags & VM_SHARED) &&
 107             atomic_read(&vma->vm_mm->mm_users) == 1)
 108                 target_node = numa_node_id();
 109
 110         flush_tlb_batched_pending(vma->vm_mm);
 111         arch_enter_lazy_mmu_mode();
 112         do {
 113                 oldpte = *pte;
 114                 if (pte_present(oldpte)) {
 115                         pte_t ptent;
 116                         bool preserve_write = prot_numa && pte_write(oldpte);
 117
 118                         /*
 119                          * Avoid trapping faults against the zero or KSM
 120                          * pages. See similar comment in change_huge_pmd.
 121                          */
 122                         if (prot_numa) {
 123                                 struct page *page;
 124                                 int nid;
 125                                 bool toptier;
 126
 127                                 /* Avoid TLB flush if possible */
 128                                 if (pte_protnone(oldpte))
 129                                         continue;
 130
 131                                 page = vm_normal_page(vma, addr, oldpte);
 132                                 if (!page || is_zone_device_page(page) || PageKsm(page))
 133                                         continue;
 134
 135                                 /* Also skip shared copy-on-write pages */
 136                                 if (is_cow_mapping(vma->vm_flags) &&
 137                                     page_count(page) != 1)
 138                                         continue;
 139
 140                                 /*
 141                                  * While migration can move some dirty pages,
 142                                  * it cannot move them all from MIGRATE_ASYNC
 143                                  * context.
 144                                  */
 145                                 if (page_is_file_lru(page) && PageDirty(page))
 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 = page_to_nid(page);
 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 (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING &&
 165                                     !toptier)
 166                                         xchg_page_access_time(page,
 167                                                 jiffies_to_msecs(jiffies));
 168                         }
 169
 170                         oldpte = ptep_modify_prot_start(vma, addr, pte);
 171                         ptent = pte_modify(oldpte, newprot);
 172                         if (preserve_write)
 173                                 ptent = pte_mk_savedwrite(ptent);
 174
 175                         if (uffd_wp) {
 176                                 ptent = pte_wrprotect(ptent);
 177                                 ptent = pte_mkuffd_wp(ptent);
 178                         } else if (uffd_wp_resolve) {
 179                                 ptent = pte_clear_uffd_wp(ptent);
 180                         }
 181
 182                         /*
 183                          * In some writable, shared mappings, we might want
 184                          * to catch actual write access -- see
 185                          * vma_wants_writenotify().
 186                          *
 187                          * In all writable, private mappings, we have to
 188                          * properly handle COW.
 189                          *
 190                          * In both cases, we can sometimes still change PTEs
 191                          * writable and avoid the write-fault handler, for
 192                          * example, if a PTE is already dirty and no other
 193                          * COW or special handling is required.
 194                          */
 195                         if ((cp_flags & MM_CP_TRY_CHANGE_WRITABLE) &&
 196                             !pte_write(ptent) &&
 197                             can_change_pte_writable(vma, addr, ptent))
 198                                 ptent = pte_mkwrite(ptent);
 199
 200                         ptep_modify_prot_commit(vma, addr, pte, oldpte, ptent);
 201                         if (pte_needs_flush(oldpte, ptent))
 202                                 tlb_flush_pte_range(tlb, addr, PAGE_SIZE);
 203                         pages++;
 204                 } else if (is_swap_pte(oldpte)) {
 205                         swp_entry_t entry = pte_to_swp_entry(oldpte);
 206                         pte_t newpte;
 207
 208                         if (is_writable_migration_entry(entry)) {
 209                                 struct page *page = pfn_swap_entry_to_page(entry);
 210
 211                                 /*
 212                                  * A protection check is difficult so
 213                                  * just be safe and disable write
 214                                  */
 215                                 if (PageAnon(page))
 216                                         entry = make_readable_exclusive_migration_entry(
 217                                                              swp_offset(entry));
 218                                 else
 219                                         entry = make_readable_migration_entry(swp_offset(entry));
 220                                 newpte = swp_entry_to_pte(entry);
 221                                 if (pte_swp_soft_dirty(oldpte))
 222                                         newpte = pte_swp_mksoft_dirty(newpte);
 223                                 if (pte_swp_uffd_wp(oldpte))
 224                                         newpte = pte_swp_mkuffd_wp(newpte);
 225                         } else if (is_writable_device_private_entry(entry)) {
 226                                 /*
 227                                  * We do not preserve soft-dirtiness. See
 228                                  * copy_one_pte() for explanation.
 229                                  */
 230                                 entry = make_readable_device_private_entry(
 231                                                         swp_offset(entry));
 232                                 newpte = swp_entry_to_pte(entry);
 233                                 if (pte_swp_uffd_wp(oldpte))
 234                                         newpte = pte_swp_mkuffd_wp(newpte);
 235                         } else if (is_writable_device_exclusive_entry(entry)) {
 236                                 entry = make_readable_device_exclusive_entry(
 237                                                         swp_offset(entry));
 238                                 newpte = swp_entry_to_pte(entry);
 239                                 if (pte_swp_soft_dirty(oldpte))
 240                                         newpte = pte_swp_mksoft_dirty(newpte);
 241                                 if (pte_swp_uffd_wp(oldpte))
 242                                         newpte = pte_swp_mkuffd_wp(newpte);
 243                         } else if (pte_marker_entry_uffd_wp(entry)) {
 244                                 /*
 245                                  * If this is uffd-wp pte marker and we'd like
 246                                  * to unprotect it, drop it; the next page
 247                                  * fault will trigger without uffd trapping.
 248                                  */
 249                                 if (uffd_wp_resolve) {
 250                                         pte_clear(vma->vm_mm, addr, pte);
 251                                         pages++;
 252                                 }
 253                                 continue;
 254                         } else {
 255                                 newpte = oldpte;
 256                         }
 257
 258                         if (uffd_wp)
 259                                 newpte = pte_swp_mkuffd_wp(newpte);
 260                         else if (uffd_wp_resolve)
 261                                 newpte = pte_swp_clear_uffd_wp(newpte);
 262
 263                         if (!pte_same(oldpte, newpte)) {
 264                                 set_pte_at(vma->vm_mm, addr, pte, newpte);
 265                                 pages++;
 266                         }
 267                 } else {
 268                         /* It must be an none page, or what else?.. */
 269                         WARN_ON_ONCE(!pte_none(oldpte));
 270                         if (unlikely(uffd_wp && !vma_is_anonymous(vma))) {
 271                                 /*
 272                                  * For file-backed mem, we need to be able to
 273                                  * wr-protect a none pte, because even if the
 274                                  * pte is none, the page/swap cache could
 275                                  * exist.  Doing that by install a marker.
 276                                  */
 277                                 set_pte_at(vma->vm_mm, addr, pte,
 278                                            make_pte_marker(PTE_MARKER_UFFD_WP));
 279                                 pages++;
 280                         }
 281                 }
 282         } while (pte++, addr += PAGE_SIZE, addr != end);
 283         arch_leave_lazy_mmu_mode();
 284         pte_unmap_unlock(pte - 1, ptl);
 285
 286         return pages;
 287 }
 288
 289 /*
 290  * Used when setting automatic NUMA hinting protection where it is
 291  * critical that a numa hinting PMD is not confused with a bad PMD.
 292  */
 293 static inline int pmd_none_or_clear_bad_unless_trans_huge(pmd_t *pmd)
 294 {
 295         pmd_t pmdval = pmd_read_atomic(pmd);
 296
 297         /* See pmd_none_or_trans_huge_or_clear_bad for info on barrier */
 298 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 299         barrier();
 300 #endif
 301
 302         if (pmd_none(pmdval))
 303                 return 1;
 304         if (pmd_trans_huge(pmdval))
 305                 return 0;
 306         if (unlikely(pmd_bad(pmdval))) {
 307                 pmd_clear_bad(pmd);
 308                 return 1;
 309         }
 310
 311         return 0;
 312 }
 313
 314 /* Return true if we're uffd wr-protecting file-backed memory, or false */
 315 static inline bool
 316 uffd_wp_protect_file(struct vm_area_struct *vma, unsigned long cp_flags)
 317 {
 318         return (cp_flags & MM_CP_UFFD_WP) && !vma_is_anonymous(vma);
 319 }
 320
 321 /*
 322  * If wr-protecting the range for file-backed, populate pgtable for the case
 323  * when pgtable is empty but page cache exists.  When {pte|pmd|...}_alloc()
 324  * failed it means no memory, we don't have a better option but stop.
 325  */
 326 #define  change_pmd_prepare(vma, pmd, cp_flags)                         \
 327         do {                                                            \
 328                 if (unlikely(uffd_wp_protect_file(vma, cp_flags))) {    \
 329                         if (WARN_ON_ONCE(pte_alloc(vma->vm_mm, pmd)))   \
 330                                 break;                                  \
 331                 }                                                       \
 332         } while (0)
 333 /*
 334  * This is the general pud/p4d/pgd version of change_pmd_prepare(). We need to
 335  * have separate change_pmd_prepare() because pte_alloc() returns 0 on success,
 336  * while {pmd|pud|p4d}_alloc() returns the valid pointer on success.
 337  */
 338 #define  change_prepare(vma, high, low, addr, cp_flags)                 \
 339         do {                                                            \
 340                 if (unlikely(uffd_wp_protect_file(vma, cp_flags))) {    \
 341                         low##_t *p = low##_alloc(vma->vm_mm, high, addr); \
 342                         if (WARN_ON_ONCE(p == NULL))                    \
 343                                 break;                                  \
 344                 }                                                       \
 345         } while (0)
 346
 347 static inline unsigned long change_pmd_range(struct mmu_gather *tlb,
 348                 struct vm_area_struct *vma, pud_t *pud, unsigned long addr,
 349                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
 350 {
 351         pmd_t *pmd;
 352         unsigned long next;
 353         unsigned long pages = 0;
 354         unsigned long nr_huge_updates = 0;
 355         struct mmu_notifier_range range;
 356
 357         range.start = 0;
 358
 359         pmd = pmd_offset(pud, addr);
 360         do {
 361                 unsigned long this_pages;
 362
 363                 next = pmd_addr_end(addr, end);
 364
 365                 change_pmd_prepare(vma, pmd, cp_flags);
 366                 /*
 367                  * Automatic NUMA balancing walks the tables with mmap_lock
 368                  * held for read. It's possible a parallel update to occur
 369                  * between pmd_trans_huge() and a pmd_none_or_clear_bad()
 370                  * check leading to a false positive and clearing.
 371                  * Hence, it's necessary to atomically read the PMD value
 372                  * for all the checks.
 373                  */
 374                 if (!is_swap_pmd(*pmd) && !pmd_devmap(*pmd) &&
 375                      pmd_none_or_clear_bad_unless_trans_huge(pmd))
 376                         goto next;
 377
 378                 /* invoke the mmu notifier if the pmd is populated */
 379                 if (!range.start) {
 380                         mmu_notifier_range_init(&range,
 381                                 MMU_NOTIFY_PROTECTION_VMA, 0,
 382                                 vma, vma->vm_mm, addr, end);
 383                         mmu_notifier_invalidate_range_start(&range);
 384                 }
 385
 386                 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
 387                         if ((next - addr != HPAGE_PMD_SIZE) ||
 388                             uffd_wp_protect_file(vma, cp_flags)) {
 389                                 __split_huge_pmd(vma, pmd, addr, false, NULL);
 390                                 /*
 391                                  * For file-backed, the pmd could have been
 392                                  * cleared; make sure pmd populated if
 393                                  * necessary, then fall-through to pte level.
 394                                  */
 395                                 change_pmd_prepare(vma, pmd, cp_flags);
 396                         } else {
 397                                 /*
 398                                  * change_huge_pmd() does not defer TLB flushes,
 399                                  * so no need to propagate the tlb argument.
 400                                  */
 401                                 int nr_ptes = change_huge_pmd(tlb, vma, pmd,
 402                                                 addr, newprot, cp_flags);
 403
 404                                 if (nr_ptes) {
 405                                         if (nr_ptes == HPAGE_PMD_NR) {
 406                                                 pages += HPAGE_PMD_NR;
 407                                                 nr_huge_updates++;
 408                                         }
 409
 410                                         /* huge pmd was handled */
 411                                         goto next;
 412                                 }
 413                         }
 414                         /* fall through, the trans huge pmd just split */
 415                 }
 416                 this_pages = change_pte_range(tlb, vma, pmd, addr, next,
 417                                               newprot, cp_flags);
 418                 pages += this_pages;
 419 next:
 420                 cond_resched();
 421         } while (pmd++, addr = next, addr != end);
 422
 423         if (range.start)
 424                 mmu_notifier_invalidate_range_end(&range);
 425
 426         if (nr_huge_updates)
 427                 count_vm_numa_events(NUMA_HUGE_PTE_UPDATES, nr_huge_updates);
 428         return pages;
 429 }
 430
 431 static inline unsigned long change_pud_range(struct mmu_gather *tlb,
 432                 struct vm_area_struct *vma, p4d_t *p4d, unsigned long addr,
 433                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
 434 {
 435         pud_t *pud;
 436         unsigned long next;
 437         unsigned long pages = 0;
 438
 439         pud = pud_offset(p4d, addr);
 440         do {
 441                 next = pud_addr_end(addr, end);
 442                 change_prepare(vma, pud, pmd, addr, cp_flags);
 443                 if (pud_none_or_clear_bad(pud))
 444                         continue;
 445                 pages += change_pmd_range(tlb, vma, pud, addr, next, newprot,
 446                                           cp_flags);
 447         } while (pud++, addr = next, addr != end);
 448
 449         return pages;
 450 }
 451
 452 static inline unsigned long change_p4d_range(struct mmu_gather *tlb,
 453                 struct vm_area_struct *vma, pgd_t *pgd, unsigned long addr,
 454                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
 455 {
 456         p4d_t *p4d;
 457         unsigned long next;
 458         unsigned long pages = 0;
 459
 460         p4d = p4d_offset(pgd, addr);
 461         do {
 462                 next = p4d_addr_end(addr, end);
 463                 change_prepare(vma, p4d, pud, addr, cp_flags);
 464                 if (p4d_none_or_clear_bad(p4d))
 465                         continue;
 466                 pages += change_pud_range(tlb, vma, p4d, addr, next, newprot,
 467                                           cp_flags);
 468         } while (p4d++, addr = next, addr != end);
 469
 470         return pages;
 471 }
 472
 473 static unsigned long change_protection_range(struct mmu_gather *tlb,
 474                 struct vm_area_struct *vma, unsigned long addr,
 475                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
 476 {
 477         struct mm_struct *mm = vma->vm_mm;
 478         pgd_t *pgd;
 479         unsigned long next;
 480         unsigned long pages = 0;
 481
 482         BUG_ON(addr >= end);
 483         pgd = pgd_offset(mm, addr);
 484         tlb_start_vma(tlb, vma);
 485         do {
 486                 next = pgd_addr_end(addr, end);
 487                 change_prepare(vma, pgd, p4d, addr, cp_flags);
 488                 if (pgd_none_or_clear_bad(pgd))
 489                         continue;
 490                 pages += change_p4d_range(tlb, vma, pgd, addr, next, newprot,
 491                                           cp_flags);
 492         } while (pgd++, addr = next, addr != end);
 493
 494         tlb_end_vma(tlb, vma);
 495
 496         return pages;
 497 }
 498
 499 unsigned long change_protection(struct mmu_gather *tlb,
 500                        struct vm_area_struct *vma, unsigned long start,
 501                        unsigned long end, pgprot_t newprot,
 502                        unsigned long cp_flags)
 503 {
 504         unsigned long pages;
 505
 506         BUG_ON((cp_flags & MM_CP_UFFD_WP_ALL) == MM_CP_UFFD_WP_ALL);
 507
 508         if (is_vm_hugetlb_page(vma))
 509                 pages = hugetlb_change_protection(vma, start, end, newprot,
 510                                                   cp_flags);
 511         else
 512                 pages = change_protection_range(tlb, vma, start, end, newprot,
 513                                                 cp_flags);
 514
 515         return pages;
 516 }
 517
 518 static int prot_none_pte_entry(pte_t *pte, unsigned long addr,
 519                                unsigned long next, struct mm_walk *walk)
 520 {
 521         return pfn_modify_allowed(pte_pfn(*pte), *(pgprot_t *)(walk->private)) ?
 522                 0 : -EACCES;
 523 }
 524
 525 static int prot_none_hugetlb_entry(pte_t *pte, unsigned long hmask,
 526                                    unsigned long addr, unsigned long next,
 527                                    struct mm_walk *walk)
 528 {
 529         return pfn_modify_allowed(pte_pfn(*pte), *(pgprot_t *)(walk->private)) ?
 530                 0 : -EACCES;
 531 }
 532
 533 static int prot_none_test(unsigned long addr, unsigned long next,
 534                           struct mm_walk *walk)
 535 {
 536         return 0;
 537 }
 538
 539 static const struct mm_walk_ops prot_none_walk_ops = {
 540         .pte_entry              = prot_none_pte_entry,
 541         .hugetlb_entry          = prot_none_hugetlb_entry,
 542         .test_walk              = prot_none_test,
 543 };
 544
 545 int
 546 mprotect_fixup(struct mmu_gather *tlb, struct vm_area_struct *vma,
 547                struct vm_area_struct **pprev, unsigned long start,
 548                unsigned long end, unsigned long newflags)
 549 {
 550         struct mm_struct *mm = vma->vm_mm;
 551         unsigned long oldflags = vma->vm_flags;
 552         long nrpages = (end - start) >> PAGE_SHIFT;
 553         unsigned long charged = 0;
 554         bool try_change_writable;
 555         pgoff_t pgoff;
 556         int error;
 557
 558         if (newflags == oldflags) {
 559                 *pprev = vma;
 560                 return 0;
 561         }
 562
 563         /*
 564          * Do PROT_NONE PFN permission checks here when we can still
 565          * bail out without undoing a lot of state. This is a rather
 566          * uncommon case, so doesn't need to be very optimized.
 567          */
 568         if (arch_has_pfn_modify_check() &&
 569             (vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 570             (newflags & VM_ACCESS_FLAGS) == 0) {
 571                 pgprot_t new_pgprot = vm_get_page_prot(newflags);
 572
 573                 error = walk_page_range(current->mm, start, end,
 574                                 &prot_none_walk_ops, &new_pgprot);
 575                 if (error)
 576                         return error;
 577         }
 578
 579         /*
 580          * If we make a private mapping writable we increase our commit;
 581          * but (without finer accounting) cannot reduce our commit if we
 582          * make it unwritable again. hugetlb mapping were accounted for
 583          * even if read-only so there is no need to account for them here
 584          */
 585         if (newflags & VM_WRITE) {
 586                 /* Check space limits when area turns into data. */
 587                 if (!may_expand_vm(mm, newflags, nrpages) &&
 588                                 may_expand_vm(mm, oldflags, nrpages))
 589                         return -ENOMEM;
 590                 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
 591                                                 VM_SHARED|VM_NORESERVE))) {
 592                         charged = nrpages;
 593                         if (security_vm_enough_memory_mm(mm, charged))
 594                                 return -ENOMEM;
 595                         newflags |= VM_ACCOUNT;
 596                 }
 597         }
 598
 599         /*
 600          * First try to merge with previous and/or next vma.
 601          */
 602         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 603         *pprev = vma_merge(mm, *pprev, start, end, newflags,
 604                            vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
 605                            vma->vm_userfaultfd_ctx, anon_vma_name(vma));
 606         if (*pprev) {
 607                 vma = *pprev;
 608                 VM_WARN_ON((vma->vm_flags ^ newflags) & ~VM_SOFTDIRTY);
 609                 goto success;
 610         }
 611
 612         *pprev = vma;
 613
 614         if (start != vma->vm_start) {
 615                 error = split_vma(mm, vma, start, 1);
 616                 if (error)
 617                         goto fail;
 618         }
 619
 620         if (end != vma->vm_end) {
 621                 error = split_vma(mm, vma, end, 0);
 622                 if (error)
 623                         goto fail;
 624         }
 625
 626 success:
 627         /*
 628          * vm_flags and vm_page_prot are protected by the mmap_lock
 629          * held in write mode.
 630          */
 631         vma->vm_flags = newflags;
 632         /*
 633          * We want to check manually if we can change individual PTEs writable
 634          * if we can't do that automatically for all PTEs in a mapping. For
 635          * private mappings, that's always the case when we have write
 636          * permissions as we properly have to handle COW.
 637          */
 638         if (vma->vm_flags & VM_SHARED)
 639                 try_change_writable = vma_wants_writenotify(vma, vma->vm_page_prot);
 640         else
 641                 try_change_writable = !!(vma->vm_flags & VM_WRITE);
 642         vma_set_page_prot(vma);
 643
 644         change_protection(tlb, vma, start, end, vma->vm_page_prot,
 645                           try_change_writable ? MM_CP_TRY_CHANGE_WRITABLE : 0);
 646
 647         /*
 648          * Private VM_LOCKED VMA becoming writable: trigger COW to avoid major
 649          * fault on access.
 650          */
 651         if ((oldflags & (VM_WRITE | VM_SHARED | VM_LOCKED)) == VM_LOCKED &&
 652                         (newflags & VM_WRITE)) {
 653                 populate_vma_page_range(vma, start, end, NULL);
 654         }
 655
 656         vm_stat_account(mm, oldflags, -nrpages);
 657         vm_stat_account(mm, newflags, nrpages);
 658         perf_event_mmap(vma);
 659         return 0;
 660
 661 fail:
 662         vm_unacct_memory(charged);
 663         return error;
 664 }
 665
 666 /*
 667  * pkey==-1 when doing a legacy mprotect()
 668  */
 669 static int do_mprotect_pkey(unsigned long start, size_t len,
 670                 unsigned long prot, int pkey)
 671 {
 672         unsigned long nstart, end, tmp, reqprot;
 673         struct vm_area_struct *vma, *prev;
 674         int error;
 675         const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
 676         const bool rier = (current->personality & READ_IMPLIES_EXEC) &&
 677                                 (prot & PROT_READ);
 678         struct mmu_gather tlb;
 679         MA_STATE(mas, &current->mm->mm_mt, 0, 0);
 680
 681         start = untagged_addr(start);
 682
 683         prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
 684         if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
 685                 return -EINVAL;
 686
 687         if (start & ~PAGE_MASK)
 688                 return -EINVAL;
 689         if (!len)
 690                 return 0;
 691         len = PAGE_ALIGN(len);
 692         end = start + len;
 693         if (end <= start)
 694                 return -ENOMEM;
 695         if (!arch_validate_prot(prot, start))
 696                 return -EINVAL;
 697
 698         reqprot = prot;
 699
 700         if (mmap_write_lock_killable(current->mm))
 701                 return -EINTR;
 702
 703         /*
 704          * If userspace did not allocate the pkey, do not let
 705          * them use it here.
 706          */
 707         error = -EINVAL;
 708         if ((pkey != -1) && !mm_pkey_is_allocated(current->mm, pkey))
 709                 goto out;
 710
 711         mas_set(&mas, start);
 712         vma = mas_find(&mas, ULONG_MAX);
 713         error = -ENOMEM;
 714         if (!vma)
 715                 goto out;
 716
 717         if (unlikely(grows & PROT_GROWSDOWN)) {
 718                 if (vma->vm_start >= end)
 719                         goto out;
 720                 start = vma->vm_start;
 721                 error = -EINVAL;
 722                 if (!(vma->vm_flags & VM_GROWSDOWN))
 723                         goto out;
 724         } else {
 725                 if (vma->vm_start > start)
 726                         goto out;
 727                 if (unlikely(grows & PROT_GROWSUP)) {
 728                         end = vma->vm_end;
 729                         error = -EINVAL;
 730                         if (!(vma->vm_flags & VM_GROWSUP))
 731                                 goto out;
 732                 }
 733         }
 734
 735         if (start > vma->vm_start)
 736                 prev = vma;
 737         else
 738                 prev = mas_prev(&mas, 0);
 739
 740         tlb_gather_mmu(&tlb, current->mm);
 741         for (nstart = start ; ; ) {
 742                 unsigned long mask_off_old_flags;
 743                 unsigned long newflags;
 744                 int new_vma_pkey;
 745
 746                 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
 747
 748                 /* Does the application expect PROT_READ to imply PROT_EXEC */
 749                 if (rier && (vma->vm_flags & VM_MAYEXEC))
 750                         prot |= PROT_EXEC;
 751
 752                 /*
 753                  * Each mprotect() call explicitly passes r/w/x permissions.
 754                  * If a permission is not passed to mprotect(), it must be
 755                  * cleared from the VMA.
 756                  */
 757                 mask_off_old_flags = VM_READ | VM_WRITE | VM_EXEC |
 758                                         VM_FLAGS_CLEAR;
 759
 760                 new_vma_pkey = arch_override_mprotect_pkey(vma, prot, pkey);
 761                 newflags = calc_vm_prot_bits(prot, new_vma_pkey);
 762                 newflags |= (vma->vm_flags & ~mask_off_old_flags);
 763
 764                 /* newflags >> 4 shift VM_MAY% in place of VM_% */
 765                 if ((newflags & ~(newflags >> 4)) & VM_ACCESS_FLAGS) {
 766                         error = -EACCES;
 767                         break;
 768                 }
 769
 770                 /* Allow architectures to sanity-check the new flags */
 771                 if (!arch_validate_flags(newflags)) {
 772                         error = -EINVAL;
 773                         break;
 774                 }
 775
 776                 error = security_file_mprotect(vma, reqprot, prot);
 777                 if (error)
 778                         break;
 779
 780                 tmp = vma->vm_end;
 781                 if (tmp > end)
 782                         tmp = end;
 783
 784                 if (vma->vm_ops && vma->vm_ops->mprotect) {
 785                         error = vma->vm_ops->mprotect(vma, nstart, tmp, newflags);
 786                         if (error)
 787                                 break;
 788                 }
 789
 790                 error = mprotect_fixup(&tlb, vma, &prev, nstart, tmp, newflags);
 791                 if (error)
 792                         break;
 793
 794                 nstart = tmp;
 795
 796                 if (nstart < prev->vm_end)
 797                         nstart = prev->vm_end;
 798                 if (nstart >= end)
 799                         break;
 800
 801                 vma = find_vma(current->mm, prev->vm_end);
 802                 if (!vma || vma->vm_start != nstart) {
 803                         error = -ENOMEM;
 804                         break;
 805                 }
 806                 prot = reqprot;
 807         }
 808         tlb_finish_mmu(&tlb);
 809 out:
 810         mmap_write_unlock(current->mm);
 811         return error;
 812 }
 813
 814 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
 815                 unsigned long, prot)
 816 {
 817         return do_mprotect_pkey(start, len, prot, -1);
 818 }
 819
 820 #ifdef CONFIG_ARCH_HAS_PKEYS
 821
 822 SYSCALL_DEFINE4(pkey_mprotect, unsigned long, start, size_t, len,
 823                 unsigned long, prot, int, pkey)
 824 {
 825         return do_mprotect_pkey(start, len, prot, pkey);
 826 }
 827
 828 SYSCALL_DEFINE2(pkey_alloc, unsigned long, flags, unsigned long, init_val)
 829 {
 830         int pkey;
 831         int ret;
 832
 833         /* No flags supported yet. */
 834         if (flags)
 835                 return -EINVAL;
 836         /* check for unsupported init values */
 837         if (init_val & ~PKEY_ACCESS_MASK)
 838                 return -EINVAL;
 839
 840         mmap_write_lock(current->mm);
 841         pkey = mm_pkey_alloc(current->mm);
 842
 843         ret = -ENOSPC;
 844         if (pkey == -1)
 845                 goto out;
 846
 847         ret = arch_set_user_pkey_access(current, pkey, init_val);
 848         if (ret) {
 849                 mm_pkey_free(current->mm, pkey);
 850                 goto out;
 851         }
 852         ret = pkey;
 853 out:
 854         mmap_write_unlock(current->mm);
 855         return ret;
 856 }
 857
 858 SYSCALL_DEFINE1(pkey_free, int, pkey)
 859 {
 860         int ret;
 861
 862         mmap_write_lock(current->mm);
 863         ret = mm_pkey_free(current->mm, pkey);
 864         mmap_write_unlock(current->mm);
 865
 866         /*
 867          * We could provide warnings or errors if any VMA still
 868          * has the pkey set here.
 869          */
 870         return ret;
 871 }
 872
 873 #endif /* CONFIG_ARCH_HAS_PKEYS */