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_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                                 if (pte_swp_uffd_wp(oldpte))
 227                                         newpte = pte_swp_mkuffd_wp(newpte);
 228                         } else if (is_writable_device_private_entry(entry)) {
 229                                 /*
 230                                  * We do not preserve soft-dirtiness. See
 231                                  * copy_one_pte() for explanation.
 232                                  */
 233                                 entry = make_readable_device_private_entry(
 234                                                         swp_offset(entry));
 235                                 newpte = swp_entry_to_pte(entry);
 236                                 if (pte_swp_uffd_wp(oldpte))
 237                                         newpte = pte_swp_mkuffd_wp(newpte);
 238                         } else if (is_writable_device_exclusive_entry(entry)) {
 239                                 entry = make_readable_device_exclusive_entry(
 240                                                         swp_offset(entry));
 241                                 newpte = swp_entry_to_pte(entry);
 242                                 if (pte_swp_soft_dirty(oldpte))
 243                                         newpte = pte_swp_mksoft_dirty(newpte);
 244                                 if (pte_swp_uffd_wp(oldpte))
 245                                         newpte = pte_swp_mkuffd_wp(newpte);
 246                         } else if (is_pte_marker_entry(entry)) {
 247                                 /*
 248                                  * Ignore swapin errors unconditionally,
 249                                  * because any access should sigbus anyway.
 250                                  */
 251                                 if (is_swapin_error_entry(entry))
 252                                         continue;
 253                                 /*
 254                                  * If this is uffd-wp pte marker and we'd like
 255                                  * to unprotect it, drop it; the next page
 256                                  * fault will trigger without uffd trapping.
 257                                  */
 258                                 if (uffd_wp_resolve) {
 259                                         pte_clear(vma->vm_mm, addr, pte);
 260                                         pages++;
 261                                 }
 262                                 continue;
 263                         } else {
 264                                 newpte = oldpte;
 265                         }
 266
 267                         if (uffd_wp)
 268                                 newpte = pte_swp_mkuffd_wp(newpte);
 269                         else if (uffd_wp_resolve)
 270                                 newpte = pte_swp_clear_uffd_wp(newpte);
 271
 272                         if (!pte_same(oldpte, newpte)) {
 273                                 set_pte_at(vma->vm_mm, addr, pte, newpte);
 274                                 pages++;
 275                         }
 276                 } else {
 277                         /* It must be an none page, or what else?.. */
 278                         WARN_ON_ONCE(!pte_none(oldpte));
 279                         if (unlikely(uffd_wp && !vma_is_anonymous(vma))) {
 280                                 /*
 281                                  * For file-backed mem, we need to be able to
 282                                  * wr-protect a none pte, because even if the
 283                                  * pte is none, the page/swap cache could
 284                                  * exist.  Doing that by install a marker.
 285                                  */
 286                                 set_pte_at(vma->vm_mm, addr, pte,
 287                                            make_pte_marker(PTE_MARKER_UFFD_WP));
 288                                 pages++;
 289                         }
 290                 }
 291         } while (pte++, addr += PAGE_SIZE, addr != end);
 292         arch_leave_lazy_mmu_mode();
 293         pte_unmap_unlock(pte - 1, ptl);
 294
 295         return pages;
 296 }
 297
 298 /*
 299  * Used when setting automatic NUMA hinting protection where it is
 300  * critical that a numa hinting PMD is not confused with a bad PMD.
 301  */
 302 static inline int pmd_none_or_clear_bad_unless_trans_huge(pmd_t *pmd)
 303 {
 304         pmd_t pmdval = pmdp_get_lockless(pmd);
 305
 306         /* See pmd_none_or_trans_huge_or_clear_bad for info on barrier */
 307 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 308         barrier();
 309 #endif
 310
 311         if (pmd_none(pmdval))
 312                 return 1;
 313         if (pmd_trans_huge(pmdval))
 314                 return 0;
 315         if (unlikely(pmd_bad(pmdval))) {
 316                 pmd_clear_bad(pmd);
 317                 return 1;
 318         }
 319
 320         return 0;
 321 }
 322
 323 /* Return true if we're uffd wr-protecting file-backed memory, or false */
 324 static inline bool
 325 uffd_wp_protect_file(struct vm_area_struct *vma, unsigned long cp_flags)
 326 {
 327         return (cp_flags & MM_CP_UFFD_WP) && !vma_is_anonymous(vma);
 328 }
 329
 330 /*
 331  * If wr-protecting the range for file-backed, populate pgtable for the case
 332  * when pgtable is empty but page cache exists.  When {pte|pmd|...}_alloc()
 333  * failed it means no memory, we don't have a better option but stop.
 334  */
 335 #define  change_pmd_prepare(vma, pmd, cp_flags)                         \
 336         do {                                                            \
 337                 if (unlikely(uffd_wp_protect_file(vma, cp_flags))) {    \
 338                         if (WARN_ON_ONCE(pte_alloc(vma->vm_mm, pmd)))   \
 339                                 break;                                  \
 340                 }                                                       \
 341         } while (0)
 342 /*
 343  * This is the general pud/p4d/pgd version of change_pmd_prepare(). We need to
 344  * have separate change_pmd_prepare() because pte_alloc() returns 0 on success,
 345  * while {pmd|pud|p4d}_alloc() returns the valid pointer on success.
 346  */
 347 #define  change_prepare(vma, high, low, addr, cp_flags)                 \
 348         do {                                                            \
 349                 if (unlikely(uffd_wp_protect_file(vma, cp_flags))) {    \
 350                         low##_t *p = low##_alloc(vma->vm_mm, high, addr); \
 351                         if (WARN_ON_ONCE(p == NULL))                    \
 352                                 break;                                  \
 353                 }                                                       \
 354         } while (0)
 355
 356 static inline unsigned long change_pmd_range(struct mmu_gather *tlb,
 357                 struct vm_area_struct *vma, pud_t *pud, unsigned long addr,
 358                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
 359 {
 360         pmd_t *pmd;
 361         unsigned long next;
 362         unsigned long pages = 0;
 363         unsigned long nr_huge_updates = 0;
 364         struct mmu_notifier_range range;
 365
 366         range.start = 0;
 367
 368         pmd = pmd_offset(pud, addr);
 369         do {
 370                 unsigned long this_pages;
 371
 372                 next = pmd_addr_end(addr, end);
 373
 374                 change_pmd_prepare(vma, pmd, cp_flags);
 375                 /*
 376                  * Automatic NUMA balancing walks the tables with mmap_lock
 377                  * held for read. It's possible a parallel update to occur
 378                  * between pmd_trans_huge() and a pmd_none_or_clear_bad()
 379                  * check leading to a false positive and clearing.
 380                  * Hence, it's necessary to atomically read the PMD value
 381                  * for all the checks.
 382                  */
 383                 if (!is_swap_pmd(*pmd) && !pmd_devmap(*pmd) &&
 384                      pmd_none_or_clear_bad_unless_trans_huge(pmd))
 385                         goto next;
 386
 387                 /* invoke the mmu notifier if the pmd is populated */
 388                 if (!range.start) {
 389                         mmu_notifier_range_init(&range,
 390                                 MMU_NOTIFY_PROTECTION_VMA, 0,
 391                                 vma, vma->vm_mm, addr, end);
 392                         mmu_notifier_invalidate_range_start(&range);
 393                 }
 394
 395                 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
 396                         if ((next - addr != HPAGE_PMD_SIZE) ||
 397                             uffd_wp_protect_file(vma, cp_flags)) {
 398                                 __split_huge_pmd(vma, pmd, addr, false, NULL);
 399                                 /*
 400                                  * For file-backed, the pmd could have been
 401                                  * cleared; make sure pmd populated if
 402                                  * necessary, then fall-through to pte level.
 403                                  */
 404                                 change_pmd_prepare(vma, pmd, cp_flags);
 405                         } else {
 406                                 /*
 407                                  * change_huge_pmd() does not defer TLB flushes,
 408                                  * so no need to propagate the tlb argument.
 409                                  */
 410                                 int nr_ptes = change_huge_pmd(tlb, vma, pmd,
 411                                                 addr, newprot, cp_flags);
 412
 413                                 if (nr_ptes) {
 414                                         if (nr_ptes == HPAGE_PMD_NR) {
 415                                                 pages += HPAGE_PMD_NR;
 416                                                 nr_huge_updates++;
 417                                         }
 418
 419                                         /* huge pmd was handled */
 420                                         goto next;
 421                                 }
 422                         }
 423                         /* fall through, the trans huge pmd just split */
 424                 }
 425                 this_pages = change_pte_range(tlb, vma, pmd, addr, next,
 426                                               newprot, cp_flags);
 427                 pages += this_pages;
 428 next:
 429                 cond_resched();
 430         } while (pmd++, addr = next, addr != end);
 431
 432         if (range.start)
 433                 mmu_notifier_invalidate_range_end(&range);
 434
 435         if (nr_huge_updates)
 436                 count_vm_numa_events(NUMA_HUGE_PTE_UPDATES, nr_huge_updates);
 437         return pages;
 438 }
 439
 440 static inline unsigned long change_pud_range(struct mmu_gather *tlb,
 441                 struct vm_area_struct *vma, p4d_t *p4d, unsigned long addr,
 442                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
 443 {
 444         pud_t *pud;
 445         unsigned long next;
 446         unsigned long pages = 0;
 447
 448         pud = pud_offset(p4d, addr);
 449         do {
 450                 next = pud_addr_end(addr, end);
 451                 change_prepare(vma, pud, pmd, addr, cp_flags);
 452                 if (pud_none_or_clear_bad(pud))
 453                         continue;
 454                 pages += change_pmd_range(tlb, vma, pud, addr, next, newprot,
 455                                           cp_flags);
 456         } while (pud++, addr = next, addr != end);
 457
 458         return pages;
 459 }
 460
 461 static inline unsigned long change_p4d_range(struct mmu_gather *tlb,
 462                 struct vm_area_struct *vma, pgd_t *pgd, unsigned long addr,
 463                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
 464 {
 465         p4d_t *p4d;
 466         unsigned long next;
 467         unsigned long pages = 0;
 468
 469         p4d = p4d_offset(pgd, addr);
 470         do {
 471                 next = p4d_addr_end(addr, end);
 472                 change_prepare(vma, p4d, pud, addr, cp_flags);
 473                 if (p4d_none_or_clear_bad(p4d))
 474                         continue;
 475                 pages += change_pud_range(tlb, vma, p4d, addr, next, newprot,
 476                                           cp_flags);
 477         } while (p4d++, addr = next, addr != end);
 478
 479         return pages;
 480 }
 481
 482 static unsigned long change_protection_range(struct mmu_gather *tlb,
 483                 struct vm_area_struct *vma, unsigned long addr,
 484                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
 485 {
 486         struct mm_struct *mm = vma->vm_mm;
 487         pgd_t *pgd;
 488         unsigned long next;
 489         unsigned long pages = 0;
 490
 491         BUG_ON(addr >= end);
 492         pgd = pgd_offset(mm, addr);
 493         tlb_start_vma(tlb, vma);
 494         do {
 495                 next = pgd_addr_end(addr, end);
 496                 change_prepare(vma, pgd, p4d, addr, cp_flags);
 497                 if (pgd_none_or_clear_bad(pgd))
 498                         continue;
 499                 pages += change_p4d_range(tlb, vma, pgd, addr, next, newprot,
 500                                           cp_flags);
 501         } while (pgd++, addr = next, addr != end);
 502
 503         tlb_end_vma(tlb, vma);
 504
 505         return pages;
 506 }
 507
 508 unsigned long change_protection(struct mmu_gather *tlb,
 509                        struct vm_area_struct *vma, unsigned long start,
 510                        unsigned long end, pgprot_t newprot,
 511                        unsigned long cp_flags)
 512 {
 513         unsigned long pages;
 514
 515         BUG_ON((cp_flags & MM_CP_UFFD_WP_ALL) == MM_CP_UFFD_WP_ALL);
 516
 517         if (is_vm_hugetlb_page(vma))
 518                 pages = hugetlb_change_protection(vma, start, end, newprot,
 519                                                   cp_flags);
 520         else
 521                 pages = change_protection_range(tlb, vma, start, end, newprot,
 522                                                 cp_flags);
 523
 524         return pages;
 525 }
 526
 527 static int prot_none_pte_entry(pte_t *pte, unsigned long addr,
 528                                unsigned long next, struct mm_walk *walk)
 529 {
 530         return pfn_modify_allowed(pte_pfn(*pte), *(pgprot_t *)(walk->private)) ?
 531                 0 : -EACCES;
 532 }
 533
 534 static int prot_none_hugetlb_entry(pte_t *pte, unsigned long hmask,
 535                                    unsigned long addr, unsigned long next,
 536                                    struct mm_walk *walk)
 537 {
 538         return pfn_modify_allowed(pte_pfn(*pte), *(pgprot_t *)(walk->private)) ?
 539                 0 : -EACCES;
 540 }
 541
 542 static int prot_none_test(unsigned long addr, unsigned long next,
 543                           struct mm_walk *walk)
 544 {
 545         return 0;
 546 }
 547
 548 static const struct mm_walk_ops prot_none_walk_ops = {
 549         .pte_entry              = prot_none_pte_entry,
 550         .hugetlb_entry          = prot_none_hugetlb_entry,
 551         .test_walk              = prot_none_test,
 552 };
 553
 554 int
 555 mprotect_fixup(struct mmu_gather *tlb, struct vm_area_struct *vma,
 556                struct vm_area_struct **pprev, unsigned long start,
 557                unsigned long end, unsigned long newflags)
 558 {
 559         struct mm_struct *mm = vma->vm_mm;
 560         unsigned long oldflags = vma->vm_flags;
 561         long nrpages = (end - start) >> PAGE_SHIFT;
 562         unsigned int mm_cp_flags = 0;
 563         unsigned long charged = 0;
 564         pgoff_t pgoff;
 565         int error;
 566
 567         if (newflags == oldflags) {
 568                 *pprev = vma;
 569                 return 0;
 570         }
 571
 572         /*
 573          * Do PROT_NONE PFN permission checks here when we can still
 574          * bail out without undoing a lot of state. This is a rather
 575          * uncommon case, so doesn't need to be very optimized.
 576          */
 577         if (arch_has_pfn_modify_check() &&
 578             (vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 579             (newflags & VM_ACCESS_FLAGS) == 0) {
 580                 pgprot_t new_pgprot = vm_get_page_prot(newflags);
 581
 582                 error = walk_page_range(current->mm, start, end,
 583                                 &prot_none_walk_ops, &new_pgprot);
 584                 if (error)
 585                         return error;
 586         }
 587
 588         /*
 589          * If we make a private mapping writable we increase our commit;
 590          * but (without finer accounting) cannot reduce our commit if we
 591          * make it unwritable again. hugetlb mapping were accounted for
 592          * even if read-only so there is no need to account for them here
 593          */
 594         if (newflags & VM_WRITE) {
 595                 /* Check space limits when area turns into data. */
 596                 if (!may_expand_vm(mm, newflags, nrpages) &&
 597                                 may_expand_vm(mm, oldflags, nrpages))
 598                         return -ENOMEM;
 599                 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
 600                                                 VM_SHARED|VM_NORESERVE))) {
 601                         charged = nrpages;
 602                         if (security_vm_enough_memory_mm(mm, charged))
 603                                 return -ENOMEM;
 604                         newflags |= VM_ACCOUNT;
 605                 }
 606         }
 607
 608         /*
 609          * First try to merge with previous and/or next vma.
 610          */
 611         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 612         *pprev = vma_merge(mm, *pprev, start, end, newflags,
 613                            vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
 614                            vma->vm_userfaultfd_ctx, anon_vma_name(vma));
 615         if (*pprev) {
 616                 vma = *pprev;
 617                 VM_WARN_ON((vma->vm_flags ^ newflags) & ~VM_SOFTDIRTY);
 618                 goto success;
 619         }
 620
 621         *pprev = vma;
 622
 623         if (start != vma->vm_start) {
 624                 error = split_vma(mm, vma, start, 1);
 625                 if (error)
 626                         goto fail;
 627         }
 628
 629         if (end != vma->vm_end) {
 630                 error = split_vma(mm, vma, end, 0);
 631                 if (error)
 632                         goto fail;
 633         }
 634
 635 success:
 636         /*
 637          * vm_flags and vm_page_prot are protected by the mmap_lock
 638          * held in write mode.
 639          */
 640         vma->vm_flags = newflags;
 641         if (vma_wants_manual_pte_write_upgrade(vma))
 642                 mm_cp_flags |= MM_CP_TRY_CHANGE_WRITABLE;
 643         vma_set_page_prot(vma);
 644
 645         change_protection(tlb, vma, start, end, vma->vm_page_prot, mm_cp_flags);
 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_ACCESS_FLAGS | VM_FLAGS_CLEAR;
 758
 759                 new_vma_pkey = arch_override_mprotect_pkey(vma, prot, pkey);
 760                 newflags = calc_vm_prot_bits(prot, new_vma_pkey);
 761                 newflags |= (vma->vm_flags & ~mask_off_old_flags);
 762
 763                 /* newflags >> 4 shift VM_MAY% in place of VM_% */
 764                 if ((newflags & ~(newflags >> 4)) & VM_ACCESS_FLAGS) {
 765                         error = -EACCES;
 766                         break;
 767                 }
 768
 769                 /* Allow architectures to sanity-check the new flags */
 770                 if (!arch_validate_flags(newflags)) {
 771                         error = -EINVAL;
 772                         break;
 773                 }
 774
 775                 error = security_file_mprotect(vma, reqprot, prot);
 776                 if (error)
 777                         break;
 778
 779                 tmp = vma->vm_end;
 780                 if (tmp > end)
 781                         tmp = end;
 782
 783                 if (vma->vm_ops && vma->vm_ops->mprotect) {
 784                         error = vma->vm_ops->mprotect(vma, nstart, tmp, newflags);
 785                         if (error)
 786                                 break;
 787                 }
 788
 789                 error = mprotect_fixup(&tlb, vma, &prev, nstart, tmp, newflags);
 790                 if (error)
 791                         break;
 792
 793                 nstart = tmp;
 794
 795                 if (nstart < prev->vm_end)
 796                         nstart = prev->vm_end;
 797                 if (nstart >= end)
 798                         break;
 799
 800                 vma = find_vma(current->mm, prev->vm_end);
 801                 if (!vma || vma->vm_start != nstart) {
 802                         error = -ENOMEM;
 803                         break;
 804                 }
 805                 prot = reqprot;
 806         }
 807         tlb_finish_mmu(&tlb);
 808 out:
 809         mmap_write_unlock(current->mm);
 810         return error;
 811 }
 812
 813 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
 814                 unsigned long, prot)
 815 {
 816         return do_mprotect_pkey(start, len, prot, -1);
 817 }
 818
 819 #ifdef CONFIG_ARCH_HAS_PKEYS
 820
 821 SYSCALL_DEFINE4(pkey_mprotect, unsigned long, start, size_t, len,
 822                 unsigned long, prot, int, pkey)
 823 {
 824         return do_mprotect_pkey(start, len, prot, pkey);
 825 }
 826
 827 SYSCALL_DEFINE2(pkey_alloc, unsigned long, flags, unsigned long, init_val)
 828 {
 829         int pkey;
 830         int ret;
 831
 832         /* No flags supported yet. */
 833         if (flags)
 834                 return -EINVAL;
 835         /* check for unsupported init values */
 836         if (init_val & ~PKEY_ACCESS_MASK)
 837                 return -EINVAL;
 838
 839         mmap_write_lock(current->mm);
 840         pkey = mm_pkey_alloc(current->mm);
 841
 842         ret = -ENOSPC;
 843         if (pkey == -1)
 844                 goto out;
 845
 846         ret = arch_set_user_pkey_access(current, pkey, init_val);
 847         if (ret) {
 848                 mm_pkey_free(current->mm, pkey);
 849                 goto out;
 850         }
 851         ret = pkey;
 852 out:
 853         mmap_write_unlock(current->mm);
 854         return ret;
 855 }
 856
 857 SYSCALL_DEFINE1(pkey_free, int, pkey)
 858 {
 859         int ret;
 860
 861         mmap_write_lock(current->mm);
 862         ret = mm_pkey_free(current->mm, pkey);
 863         mmap_write_unlock(current->mm);
 864
 865         /*
 866          * We could provide warnings or errors if any VMA still
 867          * has the pkey set here.
 868          */
 869         return ret;
 870 }
 871
 872 #endif /* CONFIG_ARCH_HAS_PKEYS */