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
  97         /*
  98          * Can be called with only the mmap_lock for reading by
  99          * prot_numa so we must check the pmd isn't constantly
 100          * changing from under us from pmd_none to pmd_trans_huge
 101          * and/or the other way around.
 102          */
 103         if (pmd_trans_unstable(pmd))
 104                 return 0;
 105
 106         /*
 107          * The pmd points to a regular pte so the pmd can't change
 108          * from under us even if the mmap_lock is only hold for
 109          * reading.
 110          */
 111         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
 112
 113         /* Get target node for single threaded private VMAs */
 114         if (prot_numa && !(vma->vm_flags & VM_SHARED) &&
 115             atomic_read(&vma->vm_mm->mm_users) == 1)
 116                 target_node = numa_node_id();
 117
 118         flush_tlb_batched_pending(vma->vm_mm);
 119         arch_enter_lazy_mmu_mode();
 120         do {
 121                 oldpte = *pte;
 122                 if (pte_present(oldpte)) {
 123                         pte_t ptent;
 124
 125                         /*
 126                          * Avoid trapping faults against the zero or KSM
 127                          * pages. See similar comment in change_huge_pmd.
 128                          */
 129                         if (prot_numa) {
 130                                 struct page *page;
 131                                 int nid;
 132                                 bool toptier;
 133
 134                                 /* Avoid TLB flush if possible */
 135                                 if (pte_protnone(oldpte))
 136                                         continue;
 137
 138                                 page = vm_normal_page(vma, addr, oldpte);
 139                                 if (!page || is_zone_device_page(page) || PageKsm(page))
 140                                         continue;
 141
 142                                 /* Also skip shared copy-on-write pages */
 143                                 if (is_cow_mapping(vma->vm_flags) &&
 144                                     page_count(page) != 1)
 145                                         continue;
 146
 147                                 /*
 148                                  * While migration can move some dirty pages,
 149                                  * it cannot move them all from MIGRATE_ASYNC
 150                                  * context.
 151                                  */
 152                                 if (page_is_file_lru(page) && PageDirty(page))
 153                                         continue;
 154
 155                                 /*
 156                                  * Don't mess with PTEs if page is already on the node
 157                                  * a single-threaded process is running on.
 158                                  */
 159                                 nid = page_to_nid(page);
 160                                 if (target_node == nid)
 161                                         continue;
 162                                 toptier = node_is_toptier(nid);
 163
 164                                 /*
 165                                  * Skip scanning top tier node if normal numa
 166                                  * balancing is disabled
 167                                  */
 168                                 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_NORMAL) &&
 169                                     toptier)
 170                                         continue;
 171                                 if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING &&
 172                                     !toptier)
 173                                         xchg_page_access_time(page,
 174                                                 jiffies_to_msecs(jiffies));
 175                         }
 176
 177                         oldpte = ptep_modify_prot_start(vma, addr, pte);
 178                         ptent = pte_modify(oldpte, newprot);
 179
 180                         if (uffd_wp)
 181                                 ptent = pte_mkuffd_wp(ptent);
 182                         else if (uffd_wp_resolve)
 183                                 ptent = pte_clear_uffd_wp(ptent);
 184
 185                         /*
 186                          * In some writable, shared mappings, we might want
 187                          * to catch actual write access -- see
 188                          * vma_wants_writenotify().
 189                          *
 190                          * In all writable, private mappings, we have to
 191                          * properly handle COW.
 192                          *
 193                          * In both cases, we can sometimes still change PTEs
 194                          * writable and avoid the write-fault handler, for
 195                          * example, if a PTE is already dirty and no other
 196                          * COW or special handling is required.
 197                          */
 198                         if ((cp_flags & MM_CP_TRY_CHANGE_WRITABLE) &&
 199                             !pte_write(ptent) &&
 200                             can_change_pte_writable(vma, addr, ptent))
 201                                 ptent = pte_mkwrite(ptent);
 202
 203                         ptep_modify_prot_commit(vma, addr, pte, oldpte, ptent);
 204                         if (pte_needs_flush(oldpte, ptent))
 205                                 tlb_flush_pte_range(tlb, addr, PAGE_SIZE);
 206                         pages++;
 207                 } else if (is_swap_pte(oldpte)) {
 208                         swp_entry_t entry = pte_to_swp_entry(oldpte);
 209                         pte_t newpte;
 210
 211                         if (is_writable_migration_entry(entry)) {
 212                                 struct page *page = pfn_swap_entry_to_page(entry);
 213
 214                                 /*
 215                                  * A protection check is difficult so
 216                                  * just be safe and disable write
 217                                  */
 218                                 if (PageAnon(page))
 219                                         entry = make_readable_exclusive_migration_entry(
 220                                                              swp_offset(entry));
 221                                 else
 222                                         entry = make_readable_migration_entry(swp_offset(entry));
 223                                 newpte = swp_entry_to_pte(entry);
 224                                 if (pte_swp_soft_dirty(oldpte))
 225                                         newpte = pte_swp_mksoft_dirty(newpte);
 226                         } else if (is_writable_device_private_entry(entry)) {
 227                                 /*
 228                                  * We do not preserve soft-dirtiness. See
 229                                  * copy_one_pte() for explanation.
 230                                  */
 231                                 entry = make_readable_device_private_entry(
 232                                                         swp_offset(entry));
 233                                 newpte = swp_entry_to_pte(entry);
 234                                 if (pte_swp_uffd_wp(oldpte))
 235                                         newpte = pte_swp_mkuffd_wp(newpte);
 236                         } else if (is_writable_device_exclusive_entry(entry)) {
 237                                 entry = make_readable_device_exclusive_entry(
 238                                                         swp_offset(entry));
 239                                 newpte = swp_entry_to_pte(entry);
 240                                 if (pte_swp_soft_dirty(oldpte))
 241                                         newpte = pte_swp_mksoft_dirty(newpte);
 242                                 if (pte_swp_uffd_wp(oldpte))
 243                                         newpte = pte_swp_mkuffd_wp(newpte);
 244                         } else if (is_pte_marker_entry(entry)) {
 245                                 /*
 246                                  * Ignore swapin errors unconditionally,
 247                                  * because any access should sigbus anyway.
 248                                  */
 249                                 if (is_swapin_error_entry(entry))
 250                                         continue;
 251                                 /*
 252                                  * If this is uffd-wp pte marker and we'd like
 253                                  * to unprotect it, drop it; the next page
 254                                  * fault will trigger without uffd trapping.
 255                                  */
 256                                 if (uffd_wp_resolve) {
 257                                         pte_clear(vma->vm_mm, addr, pte);
 258                                         pages++;
 259                                 }
 260                                 continue;
 261                         } else {
 262                                 newpte = oldpte;
 263                         }
 264
 265                         if (uffd_wp)
 266                                 newpte = pte_swp_mkuffd_wp(newpte);
 267                         else if (uffd_wp_resolve)
 268                                 newpte = pte_swp_clear_uffd_wp(newpte);
 269
 270                         if (!pte_same(oldpte, newpte)) {
 271                                 set_pte_at(vma->vm_mm, addr, pte, newpte);
 272                                 pages++;
 273                         }
 274                 } else {
 275                         /* It must be an none page, or what else?.. */
 276                         WARN_ON_ONCE(!pte_none(oldpte));
 277
 278                         /*
 279                          * Nobody plays with any none ptes besides
 280                          * userfaultfd when applying the protections.
 281                          */
 282                         if (likely(!uffd_wp))
 283                                 continue;
 284
 285                         if (userfaultfd_wp_use_markers(vma)) {
 286                                 /*
 287                                  * For file-backed mem, we need to be able to
 288                                  * wr-protect a none pte, because even if the
 289                                  * pte is none, the page/swap cache could
 290                                  * exist.  Doing that by install a marker.
 291                                  */
 292                                 set_pte_at(vma->vm_mm, addr, pte,
 293                                            make_pte_marker(PTE_MARKER_UFFD_WP));
 294                                 pages++;
 295                         }
 296                 }
 297         } while (pte++, addr += PAGE_SIZE, addr != end);
 298         arch_leave_lazy_mmu_mode();
 299         pte_unmap_unlock(pte - 1, ptl);
 300
 301         return pages;
 302 }
 303
 304 /*
 305  * Used when setting automatic NUMA hinting protection where it is
 306  * critical that a numa hinting PMD is not confused with a bad PMD.
 307  */
 308 static inline int pmd_none_or_clear_bad_unless_trans_huge(pmd_t *pmd)
 309 {
 310         pmd_t pmdval = pmdp_get_lockless(pmd);
 311
 312         /* See pmd_none_or_trans_huge_or_clear_bad for info on barrier */
 313 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 314         barrier();
 315 #endif
 316
 317         if (pmd_none(pmdval))
 318                 return 1;
 319         if (pmd_trans_huge(pmdval))
 320                 return 0;
 321         if (unlikely(pmd_bad(pmdval))) {
 322                 pmd_clear_bad(pmd);
 323                 return 1;
 324         }
 325
 326         return 0;
 327 }
 328
 329 /*
 330  * Return true if we want to split THPs into PTE mappings in change
 331  * protection procedure, false otherwise.
 332  */
 333 static inline bool
 334 pgtable_split_needed(struct vm_area_struct *vma, unsigned long cp_flags)
 335 {
 336         /*
 337          * pte markers only resides in pte level, if we need pte markers,
 338          * we need to split.  We cannot wr-protect shmem thp because file
 339          * thp is handled differently when split by erasing the pmd so far.
 340          */
 341         return (cp_flags & MM_CP_UFFD_WP) && !vma_is_anonymous(vma);
 342 }
 343
 344 /*
 345  * Return true if we want to populate pgtables in change protection
 346  * procedure, false otherwise
 347  */
 348 static inline bool
 349 pgtable_populate_needed(struct vm_area_struct *vma, unsigned long cp_flags)
 350 {
 351         /* If not within ioctl(UFFDIO_WRITEPROTECT), then don't bother */
 352         if (!(cp_flags & MM_CP_UFFD_WP))
 353                 return false;
 354
 355         /* Populate if the userfaultfd mode requires pte markers */
 356         return userfaultfd_wp_use_markers(vma);
 357 }
 358
 359 /*
 360  * Populate the pgtable underneath for whatever reason if requested.
 361  * When {pte|pmd|...}_alloc() failed we treat it the same way as pgtable
 362  * allocation failures during page faults by kicking OOM and returning
 363  * error.
 364  */
 365 #define  change_pmd_prepare(vma, pmd, cp_flags)                         \
 366         ({                                                              \
 367                 long err = 0;                                           \
 368                 if (unlikely(pgtable_populate_needed(vma, cp_flags))) { \
 369                         if (pte_alloc(vma->vm_mm, pmd))                 \
 370                                 err = -ENOMEM;                          \
 371                 }                                                       \
 372                 err;                                                    \
 373         })
 374
 375 /*
 376  * This is the general pud/p4d/pgd version of change_pmd_prepare(). We need to
 377  * have separate change_pmd_prepare() because pte_alloc() returns 0 on success,
 378  * while {pmd|pud|p4d}_alloc() returns the valid pointer on success.
 379  */
 380 #define  change_prepare(vma, high, low, addr, cp_flags)                 \
 381           ({                                                            \
 382                 long err = 0;                                           \
 383                 if (unlikely(pgtable_populate_needed(vma, cp_flags))) { \
 384                         low##_t *p = low##_alloc(vma->vm_mm, high, addr); \
 385                         if (p == NULL)                                  \
 386                                 err = -ENOMEM;                          \
 387                 }                                                       \
 388                 err;                                                    \
 389         })
 390
 391 static inline long change_pmd_range(struct mmu_gather *tlb,
 392                 struct vm_area_struct *vma, pud_t *pud, unsigned long addr,
 393                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
 394 {
 395         pmd_t *pmd;
 396         unsigned long next;
 397         long pages = 0;
 398         unsigned long nr_huge_updates = 0;
 399         struct mmu_notifier_range range;
 400
 401         range.start = 0;
 402
 403         pmd = pmd_offset(pud, addr);
 404         do {
 405                 long ret;
 406
 407                 next = pmd_addr_end(addr, end);
 408
 409                 ret = change_pmd_prepare(vma, pmd, cp_flags);
 410                 if (ret) {
 411                         pages = ret;
 412                         break;
 413                 }
 414                 /*
 415                  * Automatic NUMA balancing walks the tables with mmap_lock
 416                  * held for read. It's possible a parallel update to occur
 417                  * between pmd_trans_huge() and a pmd_none_or_clear_bad()
 418                  * check leading to a false positive and clearing.
 419                  * Hence, it's necessary to atomically read the PMD value
 420                  * for all the checks.
 421                  */
 422                 if (!is_swap_pmd(*pmd) && !pmd_devmap(*pmd) &&
 423                      pmd_none_or_clear_bad_unless_trans_huge(pmd))
 424                         goto next;
 425
 426                 /* invoke the mmu notifier if the pmd is populated */
 427                 if (!range.start) {
 428                         mmu_notifier_range_init(&range,
 429                                 MMU_NOTIFY_PROTECTION_VMA, 0,
 430                                 vma->vm_mm, addr, end);
 431                         mmu_notifier_invalidate_range_start(&range);
 432                 }
 433
 434                 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
 435                         if ((next - addr != HPAGE_PMD_SIZE) ||
 436                             pgtable_split_needed(vma, cp_flags)) {
 437                                 __split_huge_pmd(vma, pmd, addr, false, NULL);
 438                                 /*
 439                                  * For file-backed, the pmd could have been
 440                                  * cleared; make sure pmd populated if
 441                                  * necessary, then fall-through to pte level.
 442                                  */
 443                                 ret = change_pmd_prepare(vma, pmd, cp_flags);
 444                                 if (ret) {
 445                                         pages = ret;
 446                                         break;
 447                                 }
 448                         } else {
 449                                 /*
 450                                  * change_huge_pmd() does not defer TLB flushes,
 451                                  * so no need to propagate the tlb argument.
 452                                  */
 453                                 int nr_ptes = change_huge_pmd(tlb, vma, pmd,
 454                                                 addr, newprot, cp_flags);
 455
 456                                 if (nr_ptes) {
 457                                         if (nr_ptes == HPAGE_PMD_NR) {
 458                                                 pages += HPAGE_PMD_NR;
 459                                                 nr_huge_updates++;
 460                                         }
 461
 462                                         /* huge pmd was handled */
 463                                         goto next;
 464                                 }
 465                         }
 466                         /* fall through, the trans huge pmd just split */
 467                 }
 468                 pages += change_pte_range(tlb, vma, pmd, addr, next,
 469                                           newprot, cp_flags);
 470 next:
 471                 cond_resched();
 472         } while (pmd++, addr = next, addr != end);
 473
 474         if (range.start)
 475                 mmu_notifier_invalidate_range_end(&range);
 476
 477         if (nr_huge_updates)
 478                 count_vm_numa_events(NUMA_HUGE_PTE_UPDATES, nr_huge_updates);
 479         return pages;
 480 }
 481
 482 static inline long change_pud_range(struct mmu_gather *tlb,
 483                 struct vm_area_struct *vma, p4d_t *p4d, unsigned long addr,
 484                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
 485 {
 486         pud_t *pud;
 487         unsigned long next;
 488         long pages = 0, ret;
 489
 490         pud = pud_offset(p4d, addr);
 491         do {
 492                 next = pud_addr_end(addr, end);
 493                 ret = change_prepare(vma, pud, pmd, addr, cp_flags);
 494                 if (ret)
 495                         return ret;
 496                 if (pud_none_or_clear_bad(pud))
 497                         continue;
 498                 pages += change_pmd_range(tlb, vma, pud, addr, next, newprot,
 499                                           cp_flags);
 500         } while (pud++, addr = next, addr != end);
 501
 502         return pages;
 503 }
 504
 505 static inline long change_p4d_range(struct mmu_gather *tlb,
 506                 struct vm_area_struct *vma, pgd_t *pgd, unsigned long addr,
 507                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
 508 {
 509         p4d_t *p4d;
 510         unsigned long next;
 511         long pages = 0, ret;
 512
 513         p4d = p4d_offset(pgd, addr);
 514         do {
 515                 next = p4d_addr_end(addr, end);
 516                 ret = change_prepare(vma, p4d, pud, addr, cp_flags);
 517                 if (ret)
 518                         return ret;
 519                 if (p4d_none_or_clear_bad(p4d))
 520                         continue;
 521                 pages += change_pud_range(tlb, vma, p4d, addr, next, newprot,
 522                                           cp_flags);
 523         } while (p4d++, addr = next, addr != end);
 524
 525         return pages;
 526 }
 527
 528 static long change_protection_range(struct mmu_gather *tlb,
 529                 struct vm_area_struct *vma, unsigned long addr,
 530                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
 531 {
 532         struct mm_struct *mm = vma->vm_mm;
 533         pgd_t *pgd;
 534         unsigned long next;
 535         long pages = 0, ret;
 536
 537         BUG_ON(addr >= end);
 538         pgd = pgd_offset(mm, addr);
 539         tlb_start_vma(tlb, vma);
 540         do {
 541                 next = pgd_addr_end(addr, end);
 542                 ret = change_prepare(vma, pgd, p4d, addr, cp_flags);
 543                 if (ret) {
 544                         pages = ret;
 545                         break;
 546                 }
 547                 if (pgd_none_or_clear_bad(pgd))
 548                         continue;
 549                 pages += change_p4d_range(tlb, vma, pgd, addr, next, newprot,
 550                                           cp_flags);
 551         } while (pgd++, addr = next, addr != end);
 552
 553         tlb_end_vma(tlb, vma);
 554
 555         return pages;
 556 }
 557
 558 long change_protection(struct mmu_gather *tlb,
 559                        struct vm_area_struct *vma, unsigned long start,
 560                        unsigned long end, unsigned long cp_flags)
 561 {
 562         pgprot_t newprot = vma->vm_page_prot;
 563         long pages;
 564
 565         BUG_ON((cp_flags & MM_CP_UFFD_WP_ALL) == MM_CP_UFFD_WP_ALL);
 566
 567 #ifdef CONFIG_NUMA_BALANCING
 568         /*
 569          * Ordinary protection updates (mprotect, uffd-wp, softdirty tracking)
 570          * are expected to reflect their requirements via VMA flags such that
 571          * vma_set_page_prot() will adjust vma->vm_page_prot accordingly.
 572          */
 573         if (cp_flags & MM_CP_PROT_NUMA)
 574                 newprot = PAGE_NONE;
 575 #else
 576         WARN_ON_ONCE(cp_flags & MM_CP_PROT_NUMA);
 577 #endif
 578
 579         if (is_vm_hugetlb_page(vma))
 580                 pages = hugetlb_change_protection(vma, start, end, newprot,
 581                                                   cp_flags);
 582         else
 583                 pages = change_protection_range(tlb, vma, start, end, newprot,
 584                                                 cp_flags);
 585
 586         return pages;
 587 }
 588
 589 static int prot_none_pte_entry(pte_t *pte, unsigned long addr,
 590                                unsigned long next, struct mm_walk *walk)
 591 {
 592         return pfn_modify_allowed(pte_pfn(*pte), *(pgprot_t *)(walk->private)) ?
 593                 0 : -EACCES;
 594 }
 595
 596 static int prot_none_hugetlb_entry(pte_t *pte, unsigned long hmask,
 597                                    unsigned long addr, unsigned long next,
 598                                    struct mm_walk *walk)
 599 {
 600         return pfn_modify_allowed(pte_pfn(*pte), *(pgprot_t *)(walk->private)) ?
 601                 0 : -EACCES;
 602 }
 603
 604 static int prot_none_test(unsigned long addr, unsigned long next,
 605                           struct mm_walk *walk)
 606 {
 607         return 0;
 608 }
 609
 610 static const struct mm_walk_ops prot_none_walk_ops = {
 611         .pte_entry              = prot_none_pte_entry,
 612         .hugetlb_entry          = prot_none_hugetlb_entry,
 613         .test_walk              = prot_none_test,
 614 };
 615
 616 int
 617 mprotect_fixup(struct vma_iterator *vmi, struct mmu_gather *tlb,
 618                struct vm_area_struct *vma, struct vm_area_struct **pprev,
 619                unsigned long start, unsigned long end, unsigned long newflags)
 620 {
 621         struct mm_struct *mm = vma->vm_mm;
 622         unsigned long oldflags = vma->vm_flags;
 623         long nrpages = (end - start) >> PAGE_SHIFT;
 624         unsigned int mm_cp_flags = 0;
 625         unsigned long charged = 0;
 626         pgoff_t pgoff;
 627         int error;
 628
 629         if (newflags == oldflags) {
 630                 *pprev = vma;
 631                 return 0;
 632         }
 633
 634         /*
 635          * Do PROT_NONE PFN permission checks here when we can still
 636          * bail out without undoing a lot of state. This is a rather
 637          * uncommon case, so doesn't need to be very optimized.
 638          */
 639         if (arch_has_pfn_modify_check() &&
 640             (vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 641             (newflags & VM_ACCESS_FLAGS) == 0) {
 642                 pgprot_t new_pgprot = vm_get_page_prot(newflags);
 643
 644                 error = walk_page_range(current->mm, start, end,
 645                                 &prot_none_walk_ops, &new_pgprot);
 646                 if (error)
 647                         return error;
 648         }
 649
 650         /*
 651          * If we make a private mapping writable we increase our commit;
 652          * but (without finer accounting) cannot reduce our commit if we
 653          * make it unwritable again. hugetlb mapping were accounted for
 654          * even if read-only so there is no need to account for them here
 655          */
 656         if (newflags & VM_WRITE) {
 657                 /* Check space limits when area turns into data. */
 658                 if (!may_expand_vm(mm, newflags, nrpages) &&
 659                                 may_expand_vm(mm, oldflags, nrpages))
 660                         return -ENOMEM;
 661                 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
 662                                                 VM_SHARED|VM_NORESERVE))) {
 663                         charged = nrpages;
 664                         if (security_vm_enough_memory_mm(mm, charged))
 665                                 return -ENOMEM;
 666                         newflags |= VM_ACCOUNT;
 667                 }
 668         }
 669
 670         /*
 671          * First try to merge with previous and/or next vma.
 672          */
 673         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 674         *pprev = vma_merge(vmi, mm, *pprev, start, end, newflags,
 675                            vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
 676                            vma->vm_userfaultfd_ctx, anon_vma_name(vma));
 677         if (*pprev) {
 678                 vma = *pprev;
 679                 VM_WARN_ON((vma->vm_flags ^ newflags) & ~VM_SOFTDIRTY);
 680                 goto success;
 681         }
 682
 683         *pprev = vma;
 684
 685         if (start != vma->vm_start) {
 686                 error = split_vma(vmi, vma, start, 1);
 687                 if (error)
 688                         goto fail;
 689         }
 690
 691         if (end != vma->vm_end) {
 692                 error = split_vma(vmi, vma, end, 0);
 693                 if (error)
 694                         goto fail;
 695         }
 696
 697 success:
 698         /*
 699          * vm_flags and vm_page_prot are protected by the mmap_lock
 700          * held in write mode.
 701          */
 702         vm_flags_reset(vma, newflags);
 703         if (vma_wants_manual_pte_write_upgrade(vma))
 704                 mm_cp_flags |= MM_CP_TRY_CHANGE_WRITABLE;
 705         vma_set_page_prot(vma);
 706
 707         change_protection(tlb, vma, start, end, mm_cp_flags);
 708
 709         /*
 710          * Private VM_LOCKED VMA becoming writable: trigger COW to avoid major
 711          * fault on access.
 712          */
 713         if ((oldflags & (VM_WRITE | VM_SHARED | VM_LOCKED)) == VM_LOCKED &&
 714                         (newflags & VM_WRITE)) {
 715                 populate_vma_page_range(vma, start, end, NULL);
 716         }
 717
 718         vm_stat_account(mm, oldflags, -nrpages);
 719         vm_stat_account(mm, newflags, nrpages);
 720         perf_event_mmap(vma);
 721         return 0;
 722
 723 fail:
 724         vm_unacct_memory(charged);
 725         return error;
 726 }
 727
 728 /*
 729  * pkey==-1 when doing a legacy mprotect()
 730  */
 731 static int do_mprotect_pkey(unsigned long start, size_t len,
 732                 unsigned long prot, int pkey)
 733 {
 734         unsigned long nstart, end, tmp, reqprot;
 735         struct vm_area_struct *vma, *prev;
 736         int error;
 737         const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
 738         const bool rier = (current->personality & READ_IMPLIES_EXEC) &&
 739                                 (prot & PROT_READ);
 740         struct mmu_gather tlb;
 741         struct vma_iterator vmi;
 742
 743         start = untagged_addr(start);
 744
 745         prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
 746         if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
 747                 return -EINVAL;
 748
 749         if (start & ~PAGE_MASK)
 750                 return -EINVAL;
 751         if (!len)
 752                 return 0;
 753         len = PAGE_ALIGN(len);
 754         end = start + len;
 755         if (end <= start)
 756                 return -ENOMEM;
 757         if (!arch_validate_prot(prot, start))
 758                 return -EINVAL;
 759
 760         reqprot = prot;
 761
 762         if (mmap_write_lock_killable(current->mm))
 763                 return -EINTR;
 764
 765         /*
 766          * If userspace did not allocate the pkey, do not let
 767          * them use it here.
 768          */
 769         error = -EINVAL;
 770         if ((pkey != -1) && !mm_pkey_is_allocated(current->mm, pkey))
 771                 goto out;
 772
 773         vma_iter_init(&vmi, current->mm, start);
 774         vma = vma_find(&vmi, end);
 775         error = -ENOMEM;
 776         if (!vma)
 777                 goto out;
 778
 779         if (unlikely(grows & PROT_GROWSDOWN)) {
 780                 if (vma->vm_start >= end)
 781                         goto out;
 782                 start = vma->vm_start;
 783                 error = -EINVAL;
 784                 if (!(vma->vm_flags & VM_GROWSDOWN))
 785                         goto out;
 786         } else {
 787                 if (vma->vm_start > start)
 788                         goto out;
 789                 if (unlikely(grows & PROT_GROWSUP)) {
 790                         end = vma->vm_end;
 791                         error = -EINVAL;
 792                         if (!(vma->vm_flags & VM_GROWSUP))
 793                                 goto out;
 794                 }
 795         }
 796
 797         prev = vma_prev(&vmi);
 798         if (start > vma->vm_start)
 799                 prev = vma;
 800
 801         tlb_gather_mmu(&tlb, current->mm);
 802         nstart = start;
 803         tmp = vma->vm_start;
 804         for_each_vma_range(vmi, vma, end) {
 805                 unsigned long mask_off_old_flags;
 806                 unsigned long newflags;
 807                 int new_vma_pkey;
 808
 809                 if (vma->vm_start != tmp) {
 810                         error = -ENOMEM;
 811                         break;
 812                 }
 813
 814                 /* Does the application expect PROT_READ to imply PROT_EXEC */
 815                 if (rier && (vma->vm_flags & VM_MAYEXEC))
 816                         prot |= PROT_EXEC;
 817
 818                 /*
 819                  * Each mprotect() call explicitly passes r/w/x permissions.
 820                  * If a permission is not passed to mprotect(), it must be
 821                  * cleared from the VMA.
 822                  */
 823                 mask_off_old_flags = VM_ACCESS_FLAGS | VM_FLAGS_CLEAR;
 824
 825                 new_vma_pkey = arch_override_mprotect_pkey(vma, prot, pkey);
 826                 newflags = calc_vm_prot_bits(prot, new_vma_pkey);
 827                 newflags |= (vma->vm_flags & ~mask_off_old_flags);
 828
 829                 /* newflags >> 4 shift VM_MAY% in place of VM_% */
 830                 if ((newflags & ~(newflags >> 4)) & VM_ACCESS_FLAGS) {
 831                         error = -EACCES;
 832                         break;
 833                 }
 834
 835                 if (map_deny_write_exec(vma, newflags)) {
 836                         error = -EACCES;
 837                         break;
 838                 }
 839
 840                 /* Allow architectures to sanity-check the new flags */
 841                 if (!arch_validate_flags(newflags)) {
 842                         error = -EINVAL;
 843                         break;
 844                 }
 845
 846                 error = security_file_mprotect(vma, reqprot, prot);
 847                 if (error)
 848                         break;
 849
 850                 tmp = vma->vm_end;
 851                 if (tmp > end)
 852                         tmp = end;
 853
 854                 if (vma->vm_ops && vma->vm_ops->mprotect) {
 855                         error = vma->vm_ops->mprotect(vma, nstart, tmp, newflags);
 856                         if (error)
 857                                 break;
 858                 }
 859
 860                 error = mprotect_fixup(&vmi, &tlb, vma, &prev, nstart, tmp, newflags);
 861                 if (error)
 862                         break;
 863
 864                 tmp = vma_iter_end(&vmi);
 865                 nstart = tmp;
 866                 prot = reqprot;
 867         }
 868         tlb_finish_mmu(&tlb);
 869
 870         if (!error && tmp < end)
 871                 error = -ENOMEM;
 872
 873 out:
 874         mmap_write_unlock(current->mm);
 875         return error;
 876 }
 877
 878 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
 879                 unsigned long, prot)
 880 {
 881         return do_mprotect_pkey(start, len, prot, -1);
 882 }
 883
 884 #ifdef CONFIG_ARCH_HAS_PKEYS
 885
 886 SYSCALL_DEFINE4(pkey_mprotect, unsigned long, start, size_t, len,
 887                 unsigned long, prot, int, pkey)
 888 {
 889         return do_mprotect_pkey(start, len, prot, pkey);
 890 }
 891
 892 SYSCALL_DEFINE2(pkey_alloc, unsigned long, flags, unsigned long, init_val)
 893 {
 894         int pkey;
 895         int ret;
 896
 897         /* No flags supported yet. */
 898         if (flags)
 899                 return -EINVAL;
 900         /* check for unsupported init values */
 901         if (init_val & ~PKEY_ACCESS_MASK)
 902                 return -EINVAL;
 903
 904         mmap_write_lock(current->mm);
 905         pkey = mm_pkey_alloc(current->mm);
 906
 907         ret = -ENOSPC;
 908         if (pkey == -1)
 909                 goto out;
 910
 911         ret = arch_set_user_pkey_access(current, pkey, init_val);
 912         if (ret) {
 913                 mm_pkey_free(current->mm, pkey);
 914                 goto out;
 915         }
 916         ret = pkey;
 917 out:
 918         mmap_write_unlock(current->mm);
 919         return ret;
 920 }
 921
 922 SYSCALL_DEFINE1(pkey_free, int, pkey)
 923 {
 924         int ret;
 925
 926         mmap_write_lock(current->mm);
 927         ret = mm_pkey_free(current->mm, pkey);
 928         mmap_write_unlock(current->mm);
 929
 930         /*
 931          * We could provide warnings or errors if any VMA still
 932          * has the pkey set here.
 933          */
 934         return ret;
 935 }
 936
 937 #endif /* CONFIG_ARCH_HAS_PKEYS */