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