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 static inline bool can_change_pte_writable(struct vm_area_struct *vma,
  43                                            unsigned long addr, pte_t pte)
  44 {
  45         struct page *page;
  46
  47         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                         bool preserve_write = prot_numa && pte_write(oldpte);
 125
 126                         /*
 127                          * Avoid trapping faults against the zero or KSM
 128                          * pages. See similar comment in change_huge_pmd.
 129                          */
 130                         if (prot_numa) {
 131                                 struct page *page;
 132                                 int nid;
 133                                 bool toptier;
 134
 135                                 /* Avoid TLB flush if possible */
 136                                 if (pte_protnone(oldpte))
 137                                         continue;
 138
 139                                 page = vm_normal_page(vma, addr, oldpte);
 140                                 if (!page || is_zone_device_page(page) || PageKsm(page))
 141                                         continue;
 142
 143                                 /* Also skip shared copy-on-write pages */
 144                                 if (is_cow_mapping(vma->vm_flags) &&
 145                                     page_count(page) != 1)
 146                                         continue;
 147
 148                                 /*
 149                                  * While migration can move some dirty pages,
 150                                  * it cannot move them all from MIGRATE_ASYNC
 151                                  * context.
 152                                  */
 153                                 if (page_is_file_lru(page) && PageDirty(page))
 154                                         continue;
 155
 156                                 /*
 157                                  * Don't mess with PTEs if page is already on the node
 158                                  * a single-threaded process is running on.
 159                                  */
 160                                 nid = page_to_nid(page);
 161                                 if (target_node == nid)
 162                                         continue;
 163                                 toptier = node_is_toptier(nid);
 164
 165                                 /*
 166                                  * Skip scanning top tier node if normal numa
 167                                  * balancing is disabled
 168                                  */
 169                                 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_NORMAL) &&
 170                                     toptier)
 171                                         continue;
 172                                 if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING &&
 173                                     !toptier)
 174                                         xchg_page_access_time(page,
 175                                                 jiffies_to_msecs(jiffies));
 176                         }
 177
 178                         oldpte = ptep_modify_prot_start(vma, addr, pte);
 179                         ptent = pte_modify(oldpte, newprot);
 180                         if (preserve_write)
 181                                 ptent = pte_mk_savedwrite(ptent);
 182
 183                         if (uffd_wp) {
 184                                 ptent = pte_wrprotect(ptent);
 185                                 ptent = pte_mkuffd_wp(ptent);
 186                         } else if (uffd_wp_resolve) {
 187                                 ptent = pte_clear_uffd_wp(ptent);
 188                         }
 189
 190                         /*
 191                          * In some writable, shared mappings, we might want
 192                          * to catch actual write access -- see
 193                          * vma_wants_writenotify().
 194                          *
 195                          * In all writable, private mappings, we have to
 196                          * properly handle COW.
 197                          *
 198                          * In both cases, we can sometimes still change PTEs
 199                          * writable and avoid the write-fault handler, for
 200                          * example, if a PTE is already dirty and no other
 201                          * COW or special handling is required.
 202                          */
 203                         if ((cp_flags & MM_CP_TRY_CHANGE_WRITABLE) &&
 204                             !pte_write(ptent) &&
 205                             can_change_pte_writable(vma, addr, ptent))
 206                                 ptent = pte_mkwrite(ptent);
 207
 208                         ptep_modify_prot_commit(vma, addr, pte, oldpte, ptent);
 209                         if (pte_needs_flush(oldpte, ptent))
 210                                 tlb_flush_pte_range(tlb, addr, PAGE_SIZE);
 211                         pages++;
 212                 } else if (is_swap_pte(oldpte)) {
 213                         swp_entry_t entry = pte_to_swp_entry(oldpte);
 214                         pte_t newpte;
 215
 216                         if (is_writable_migration_entry(entry)) {
 217                                 struct page *page = pfn_swap_entry_to_page(entry);
 218
 219                                 /*
 220                                  * A protection check is difficult so
 221                                  * just be safe and disable write
 222                                  */
 223                                 if (PageAnon(page))
 224                                         entry = make_readable_exclusive_migration_entry(
 225                                                              swp_offset(entry));
 226                                 else
 227                                         entry = make_readable_migration_entry(swp_offset(entry));
 228                                 newpte = swp_entry_to_pte(entry);
 229                                 if (pte_swp_soft_dirty(oldpte))
 230                                         newpte = pte_swp_mksoft_dirty(newpte);
 231                                 if (pte_swp_uffd_wp(oldpte))
 232                                         newpte = pte_swp_mkuffd_wp(newpte);
 233                         } else if (is_writable_device_private_entry(entry)) {
 234                                 /*
 235                                  * We do not preserve soft-dirtiness. See
 236                                  * copy_one_pte() for explanation.
 237                                  */
 238                                 entry = make_readable_device_private_entry(
 239                                                         swp_offset(entry));
 240                                 newpte = swp_entry_to_pte(entry);
 241                                 if (pte_swp_uffd_wp(oldpte))
 242                                         newpte = pte_swp_mkuffd_wp(newpte);
 243                         } else if (is_writable_device_exclusive_entry(entry)) {
 244                                 entry = make_readable_device_exclusive_entry(
 245                                                         swp_offset(entry));
 246                                 newpte = swp_entry_to_pte(entry);
 247                                 if (pte_swp_soft_dirty(oldpte))
 248                                         newpte = pte_swp_mksoft_dirty(newpte);
 249                                 if (pte_swp_uffd_wp(oldpte))
 250                                         newpte = pte_swp_mkuffd_wp(newpte);
 251                         } else if (pte_marker_entry_uffd_wp(entry)) {
 252                                 /*
 253                                  * If this is uffd-wp pte marker and we'd like
 254                                  * to unprotect it, drop it; the next page
 255                                  * fault will trigger without uffd trapping.
 256                                  */
 257                                 if (uffd_wp_resolve) {
 258                                         pte_clear(vma->vm_mm, addr, pte);
 259                                         pages++;
 260                                 }
 261                                 continue;
 262                         } else {
 263                                 newpte = oldpte;
 264                         }
 265
 266                         if (uffd_wp)
 267                                 newpte = pte_swp_mkuffd_wp(newpte);
 268                         else if (uffd_wp_resolve)
 269                                 newpte = pte_swp_clear_uffd_wp(newpte);
 270
 271                         if (!pte_same(oldpte, newpte)) {
 272                                 set_pte_at(vma->vm_mm, addr, pte, newpte);
 273                                 pages++;
 274                         }
 275                 } else {
 276                         /* It must be an none page, or what else?.. */
 277                         WARN_ON_ONCE(!pte_none(oldpte));
 278                         if (unlikely(uffd_wp && !vma_is_anonymous(vma))) {
 279                                 /*
 280                                  * For file-backed mem, we need to be able to
 281                                  * wr-protect a none pte, because even if the
 282                                  * pte is none, the page/swap cache could
 283                                  * exist.  Doing that by install a marker.
 284                                  */
 285                                 set_pte_at(vma->vm_mm, addr, pte,
 286                                            make_pte_marker(PTE_MARKER_UFFD_WP));
 287                                 pages++;
 288                         }
 289                 }
 290         } while (pte++, addr += PAGE_SIZE, addr != end);
 291         arch_leave_lazy_mmu_mode();
 292         pte_unmap_unlock(pte - 1, ptl);
 293
 294         return pages;
 295 }
 296
 297 /*
 298  * Used when setting automatic NUMA hinting protection where it is
 299  * critical that a numa hinting PMD is not confused with a bad PMD.
 300  */
 301 static inline int pmd_none_or_clear_bad_unless_trans_huge(pmd_t *pmd)
 302 {
 303         pmd_t pmdval = pmd_read_atomic(pmd);
 304
 305         /* See pmd_none_or_trans_huge_or_clear_bad for info on barrier */
 306 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 307         barrier();
 308 #endif
 309
 310         if (pmd_none(pmdval))
 311                 return 1;
 312         if (pmd_trans_huge(pmdval))
 313                 return 0;
 314         if (unlikely(pmd_bad(pmdval))) {
 315                 pmd_clear_bad(pmd);
 316                 return 1;
 317         }
 318
 319         return 0;
 320 }
 321
 322 /* Return true if we're uffd wr-protecting file-backed memory, or false */
 323 static inline bool
 324 uffd_wp_protect_file(struct vm_area_struct *vma, unsigned long cp_flags)
 325 {
 326         return (cp_flags & MM_CP_UFFD_WP) && !vma_is_anonymous(vma);
 327 }
 328
 329 /*
 330  * If wr-protecting the range for file-backed, populate pgtable for the case
 331  * when pgtable is empty but page cache exists.  When {pte|pmd|...}_alloc()
 332  * failed it means no memory, we don't have a better option but stop.
 333  */
 334 #define  change_pmd_prepare(vma, pmd, cp_flags)                         \
 335         do {                                                            \
 336                 if (unlikely(uffd_wp_protect_file(vma, cp_flags))) {    \
 337                         if (WARN_ON_ONCE(pte_alloc(vma->vm_mm, pmd)))   \
 338                                 break;                                  \
 339                 }                                                       \
 340         } while (0)
 341 /*
 342  * This is the general pud/p4d/pgd version of change_pmd_prepare(). We need to
 343  * have separate change_pmd_prepare() because pte_alloc() returns 0 on success,
 344  * while {pmd|pud|p4d}_alloc() returns the valid pointer on success.
 345  */
 346 #define  change_prepare(vma, high, low, addr, cp_flags)                 \
 347         do {                                                            \
 348                 if (unlikely(uffd_wp_protect_file(vma, cp_flags))) {    \
 349                         low##_t *p = low##_alloc(vma->vm_mm, high, addr); \
 350                         if (WARN_ON_ONCE(p == NULL))                    \
 351                                 break;                                  \
 352                 }                                                       \
 353         } while (0)
 354
 355 static inline unsigned long change_pmd_range(struct mmu_gather *tlb,
 356                 struct vm_area_struct *vma, pud_t *pud, unsigned long addr,
 357                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
 358 {
 359         pmd_t *pmd;
 360         unsigned long next;
 361         unsigned long pages = 0;
 362         unsigned long nr_huge_updates = 0;
 363         struct mmu_notifier_range range;
 364
 365         range.start = 0;
 366
 367         pmd = pmd_offset(pud, addr);
 368         do {
 369                 unsigned long this_pages;
 370
 371                 next = pmd_addr_end(addr, end);
 372
 373                 change_pmd_prepare(vma, pmd, cp_flags);
 374                 /*
 375                  * Automatic NUMA balancing walks the tables with mmap_lock
 376                  * held for read. It's possible a parallel update to occur
 377                  * between pmd_trans_huge() and a pmd_none_or_clear_bad()
 378                  * check leading to a false positive and clearing.
 379                  * Hence, it's necessary to atomically read the PMD value
 380                  * for all the checks.
 381                  */
 382                 if (!is_swap_pmd(*pmd) && !pmd_devmap(*pmd) &&
 383                      pmd_none_or_clear_bad_unless_trans_huge(pmd))
 384                         goto next;
 385
 386                 /* invoke the mmu notifier if the pmd is populated */
 387                 if (!range.start) {
 388                         mmu_notifier_range_init(&range,
 389                                 MMU_NOTIFY_PROTECTION_VMA, 0,
 390                                 vma, vma->vm_mm, addr, end);
 391                         mmu_notifier_invalidate_range_start(&range);
 392                 }
 393
 394                 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
 395                         if ((next - addr != HPAGE_PMD_SIZE) ||
 396                             uffd_wp_protect_file(vma, cp_flags)) {
 397                                 __split_huge_pmd(vma, pmd, addr, false, NULL);
 398                                 /*
 399                                  * For file-backed, the pmd could have been
 400                                  * cleared; make sure pmd populated if
 401                                  * necessary, then fall-through to pte level.
 402                                  */
 403                                 change_pmd_prepare(vma, pmd, cp_flags);
 404                         } else {
 405                                 /*
 406                                  * change_huge_pmd() does not defer TLB flushes,
 407                                  * so no need to propagate the tlb argument.
 408                                  */
 409                                 int nr_ptes = change_huge_pmd(tlb, vma, pmd,
 410                                                 addr, newprot, cp_flags);
 411
 412                                 if (nr_ptes) {
 413                                         if (nr_ptes == HPAGE_PMD_NR) {
 414                                                 pages += HPAGE_PMD_NR;
 415                                                 nr_huge_updates++;
 416                                         }
 417
 418                                         /* huge pmd was handled */
 419                                         goto next;
 420                                 }
 421                         }
 422                         /* fall through, the trans huge pmd just split */
 423                 }
 424                 this_pages = change_pte_range(tlb, vma, pmd, addr, next,
 425                                               newprot, cp_flags);
 426                 pages += this_pages;
 427 next:
 428                 cond_resched();
 429         } while (pmd++, addr = next, addr != end);
 430
 431         if (range.start)
 432                 mmu_notifier_invalidate_range_end(&range);
 433
 434         if (nr_huge_updates)
 435                 count_vm_numa_events(NUMA_HUGE_PTE_UPDATES, nr_huge_updates);
 436         return pages;
 437 }
 438
 439 static inline unsigned long change_pud_range(struct mmu_gather *tlb,
 440                 struct vm_area_struct *vma, p4d_t *p4d, unsigned long addr,
 441                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
 442 {
 443         pud_t *pud;
 444         unsigned long next;
 445         unsigned long pages = 0;
 446
 447         pud = pud_offset(p4d, addr);
 448         do {
 449                 next = pud_addr_end(addr, end);
 450                 change_prepare(vma, pud, pmd, addr, cp_flags);
 451                 if (pud_none_or_clear_bad(pud))
 452                         continue;
 453                 pages += change_pmd_range(tlb, vma, pud, addr, next, newprot,
 454                                           cp_flags);
 455         } while (pud++, addr = next, addr != end);
 456
 457         return pages;
 458 }
 459
 460 static inline unsigned long change_p4d_range(struct mmu_gather *tlb,
 461                 struct vm_area_struct *vma, pgd_t *pgd, unsigned long addr,
 462                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
 463 {
 464         p4d_t *p4d;
 465         unsigned long next;
 466         unsigned long pages = 0;
 467
 468         p4d = p4d_offset(pgd, addr);
 469         do {
 470                 next = p4d_addr_end(addr, end);
 471                 change_prepare(vma, p4d, pud, addr, cp_flags);
 472                 if (p4d_none_or_clear_bad(p4d))
 473                         continue;
 474                 pages += change_pud_range(tlb, vma, p4d, addr, next, newprot,
 475                                           cp_flags);
 476         } while (p4d++, addr = next, addr != end);
 477
 478         return pages;
 479 }
 480
 481 static unsigned long change_protection_range(struct mmu_gather *tlb,
 482                 struct vm_area_struct *vma, unsigned long addr,
 483                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
 484 {
 485         struct mm_struct *mm = vma->vm_mm;
 486         pgd_t *pgd;
 487         unsigned long next;
 488         unsigned long pages = 0;
 489
 490         BUG_ON(addr >= end);
 491         pgd = pgd_offset(mm, addr);
 492         tlb_start_vma(tlb, vma);
 493         do {
 494                 next = pgd_addr_end(addr, end);
 495                 change_prepare(vma, pgd, p4d, addr, cp_flags);
 496                 if (pgd_none_or_clear_bad(pgd))
 497                         continue;
 498                 pages += change_p4d_range(tlb, vma, pgd, addr, next, newprot,
 499                                           cp_flags);
 500         } while (pgd++, addr = next, addr != end);
 501
 502         tlb_end_vma(tlb, vma);
 503
 504         return pages;
 505 }
 506
 507 unsigned long change_protection(struct mmu_gather *tlb,
 508                        struct vm_area_struct *vma, unsigned long start,
 509                        unsigned long end, pgprot_t newprot,
 510                        unsigned long cp_flags)
 511 {
 512         unsigned long pages;
 513
 514         BUG_ON((cp_flags & MM_CP_UFFD_WP_ALL) == MM_CP_UFFD_WP_ALL);
 515
 516         if (is_vm_hugetlb_page(vma))
 517                 pages = hugetlb_change_protection(vma, start, end, newprot,
 518                                                   cp_flags);
 519         else
 520                 pages = change_protection_range(tlb, vma, start, end, newprot,
 521                                                 cp_flags);
 522
 523         return pages;
 524 }
 525
 526 static int prot_none_pte_entry(pte_t *pte, unsigned long addr,
 527                                unsigned long next, struct mm_walk *walk)
 528 {
 529         return pfn_modify_allowed(pte_pfn(*pte), *(pgprot_t *)(walk->private)) ?
 530                 0 : -EACCES;
 531 }
 532
 533 static int prot_none_hugetlb_entry(pte_t *pte, unsigned long hmask,
 534                                    unsigned long addr, unsigned long next,
 535                                    struct mm_walk *walk)
 536 {
 537         return pfn_modify_allowed(pte_pfn(*pte), *(pgprot_t *)(walk->private)) ?
 538                 0 : -EACCES;
 539 }
 540
 541 static int prot_none_test(unsigned long addr, unsigned long next,
 542                           struct mm_walk *walk)
 543 {
 544         return 0;
 545 }
 546
 547 static const struct mm_walk_ops prot_none_walk_ops = {
 548         .pte_entry              = prot_none_pte_entry,
 549         .hugetlb_entry          = prot_none_hugetlb_entry,
 550         .test_walk              = prot_none_test,
 551 };
 552
 553 int
 554 mprotect_fixup(struct mmu_gather *tlb, struct vm_area_struct *vma,
 555                struct vm_area_struct **pprev, unsigned long start,
 556                unsigned long end, unsigned long newflags)
 557 {
 558         struct mm_struct *mm = vma->vm_mm;
 559         unsigned long oldflags = vma->vm_flags;
 560         long nrpages = (end - start) >> PAGE_SHIFT;
 561         unsigned long charged = 0;
 562         bool try_change_writable;
 563         pgoff_t pgoff;
 564         int error;
 565
 566         if (newflags == oldflags) {
 567                 *pprev = vma;
 568                 return 0;
 569         }
 570
 571         /*
 572          * Do PROT_NONE PFN permission checks here when we can still
 573          * bail out without undoing a lot of state. This is a rather
 574          * uncommon case, so doesn't need to be very optimized.
 575          */
 576         if (arch_has_pfn_modify_check() &&
 577             (vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 578             (newflags & VM_ACCESS_FLAGS) == 0) {
 579                 pgprot_t new_pgprot = vm_get_page_prot(newflags);
 580
 581                 error = walk_page_range(current->mm, start, end,
 582                                 &prot_none_walk_ops, &new_pgprot);
 583                 if (error)
 584                         return error;
 585         }
 586
 587         /*
 588          * If we make a private mapping writable we increase our commit;
 589          * but (without finer accounting) cannot reduce our commit if we
 590          * make it unwritable again. hugetlb mapping were accounted for
 591          * even if read-only so there is no need to account for them here
 592          */
 593         if (newflags & VM_WRITE) {
 594                 /* Check space limits when area turns into data. */
 595                 if (!may_expand_vm(mm, newflags, nrpages) &&
 596                                 may_expand_vm(mm, oldflags, nrpages))
 597                         return -ENOMEM;
 598                 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
 599                                                 VM_SHARED|VM_NORESERVE))) {
 600                         charged = nrpages;
 601                         if (security_vm_enough_memory_mm(mm, charged))
 602                                 return -ENOMEM;
 603                         newflags |= VM_ACCOUNT;
 604                 }
 605         }
 606
 607         /*
 608          * First try to merge with previous and/or next vma.
 609          */
 610         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 611         *pprev = vma_merge(mm, *pprev, start, end, newflags,
 612                            vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
 613                            vma->vm_userfaultfd_ctx, anon_vma_name(vma));
 614         if (*pprev) {
 615                 vma = *pprev;
 616                 VM_WARN_ON((vma->vm_flags ^ newflags) & ~VM_SOFTDIRTY);
 617                 goto success;
 618         }
 619
 620         *pprev = vma;
 621
 622         if (start != vma->vm_start) {
 623                 error = split_vma(mm, vma, start, 1);
 624                 if (error)
 625                         goto fail;
 626         }
 627
 628         if (end != vma->vm_end) {
 629                 error = split_vma(mm, vma, end, 0);
 630                 if (error)
 631                         goto fail;
 632         }
 633
 634 success:
 635         /*
 636          * vm_flags and vm_page_prot are protected by the mmap_lock
 637          * held in write mode.
 638          */
 639         vma->vm_flags = newflags;
 640         /*
 641          * We want to check manually if we can change individual PTEs writable
 642          * if we can't do that automatically for all PTEs in a mapping. For
 643          * private mappings, that's always the case when we have write
 644          * permissions as we properly have to handle COW.
 645          */
 646         if (vma->vm_flags & VM_SHARED)
 647                 try_change_writable = vma_wants_writenotify(vma, vma->vm_page_prot);
 648         else
 649                 try_change_writable = !!(vma->vm_flags & VM_WRITE);
 650         vma_set_page_prot(vma);
 651
 652         change_protection(tlb, vma, start, end, vma->vm_page_prot,
 653                           try_change_writable ? MM_CP_TRY_CHANGE_WRITABLE : 0);
 654
 655         /*
 656          * Private VM_LOCKED VMA becoming writable: trigger COW to avoid major
 657          * fault on access.
 658          */
 659         if ((oldflags & (VM_WRITE | VM_SHARED | VM_LOCKED)) == VM_LOCKED &&
 660                         (newflags & VM_WRITE)) {
 661                 populate_vma_page_range(vma, start, end, NULL);
 662         }
 663
 664         vm_stat_account(mm, oldflags, -nrpages);
 665         vm_stat_account(mm, newflags, nrpages);
 666         perf_event_mmap(vma);
 667         return 0;
 668
 669 fail:
 670         vm_unacct_memory(charged);
 671         return error;
 672 }
 673
 674 /*
 675  * pkey==-1 when doing a legacy mprotect()
 676  */
 677 static int do_mprotect_pkey(unsigned long start, size_t len,
 678                 unsigned long prot, int pkey)
 679 {
 680         unsigned long nstart, end, tmp, reqprot;
 681         struct vm_area_struct *vma, *prev;
 682         int error;
 683         const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
 684         const bool rier = (current->personality & READ_IMPLIES_EXEC) &&
 685                                 (prot & PROT_READ);
 686         struct mmu_gather tlb;
 687         MA_STATE(mas, &current->mm->mm_mt, 0, 0);
 688
 689         start = untagged_addr(start);
 690
 691         prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
 692         if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
 693                 return -EINVAL;
 694
 695         if (start & ~PAGE_MASK)
 696                 return -EINVAL;
 697         if (!len)
 698                 return 0;
 699         len = PAGE_ALIGN(len);
 700         end = start + len;
 701         if (end <= start)
 702                 return -ENOMEM;
 703         if (!arch_validate_prot(prot, start))
 704                 return -EINVAL;
 705
 706         reqprot = prot;
 707
 708         if (mmap_write_lock_killable(current->mm))
 709                 return -EINTR;
 710
 711         /*
 712          * If userspace did not allocate the pkey, do not let
 713          * them use it here.
 714          */
 715         error = -EINVAL;
 716         if ((pkey != -1) && !mm_pkey_is_allocated(current->mm, pkey))
 717                 goto out;
 718
 719         mas_set(&mas, start);
 720         vma = mas_find(&mas, ULONG_MAX);
 721         error = -ENOMEM;
 722         if (!vma)
 723                 goto out;
 724
 725         if (unlikely(grows & PROT_GROWSDOWN)) {
 726                 if (vma->vm_start >= end)
 727                         goto out;
 728                 start = vma->vm_start;
 729                 error = -EINVAL;
 730                 if (!(vma->vm_flags & VM_GROWSDOWN))
 731                         goto out;
 732         } else {
 733                 if (vma->vm_start > start)
 734                         goto out;
 735                 if (unlikely(grows & PROT_GROWSUP)) {
 736                         end = vma->vm_end;
 737                         error = -EINVAL;
 738                         if (!(vma->vm_flags & VM_GROWSUP))
 739                                 goto out;
 740                 }
 741         }
 742
 743         if (start > vma->vm_start)
 744                 prev = vma;
 745         else
 746                 prev = mas_prev(&mas, 0);
 747
 748         tlb_gather_mmu(&tlb, current->mm);
 749         for (nstart = start ; ; ) {
 750                 unsigned long mask_off_old_flags;
 751                 unsigned long newflags;
 752                 int new_vma_pkey;
 753
 754                 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
 755
 756                 /* Does the application expect PROT_READ to imply PROT_EXEC */
 757                 if (rier && (vma->vm_flags & VM_MAYEXEC))
 758                         prot |= PROT_EXEC;
 759
 760                 /*
 761                  * Each mprotect() call explicitly passes r/w/x permissions.
 762                  * If a permission is not passed to mprotect(), it must be
 763                  * cleared from the VMA.
 764                  */
 765                 mask_off_old_flags = VM_ACCESS_FLAGS | VM_FLAGS_CLEAR;
 766
 767                 new_vma_pkey = arch_override_mprotect_pkey(vma, prot, pkey);
 768                 newflags = calc_vm_prot_bits(prot, new_vma_pkey);
 769                 newflags |= (vma->vm_flags & ~mask_off_old_flags);
 770
 771                 /* newflags >> 4 shift VM_MAY% in place of VM_% */
 772                 if ((newflags & ~(newflags >> 4)) & VM_ACCESS_FLAGS) {
 773                         error = -EACCES;
 774                         break;
 775                 }
 776
 777                 /* Allow architectures to sanity-check the new flags */
 778                 if (!arch_validate_flags(newflags)) {
 779                         error = -EINVAL;
 780                         break;
 781                 }
 782
 783                 error = security_file_mprotect(vma, reqprot, prot);
 784                 if (error)
 785                         break;
 786
 787                 tmp = vma->vm_end;
 788                 if (tmp > end)
 789                         tmp = end;
 790
 791                 if (vma->vm_ops && vma->vm_ops->mprotect) {
 792                         error = vma->vm_ops->mprotect(vma, nstart, tmp, newflags);
 793                         if (error)
 794                                 break;
 795                 }
 796
 797                 error = mprotect_fixup(&tlb, vma, &prev, nstart, tmp, newflags);
 798                 if (error)
 799                         break;
 800
 801                 nstart = tmp;
 802
 803                 if (nstart < prev->vm_end)
 804                         nstart = prev->vm_end;
 805                 if (nstart >= end)
 806                         break;
 807
 808                 vma = find_vma(current->mm, prev->vm_end);
 809                 if (!vma || vma->vm_start != nstart) {
 810                         error = -ENOMEM;
 811                         break;
 812                 }
 813                 prot = reqprot;
 814         }
 815         tlb_finish_mmu(&tlb);
 816 out:
 817         mmap_write_unlock(current->mm);
 818         return error;
 819 }
 820
 821 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
 822                 unsigned long, prot)
 823 {
 824         return do_mprotect_pkey(start, len, prot, -1);
 825 }
 826
 827 #ifdef CONFIG_ARCH_HAS_PKEYS
 828
 829 SYSCALL_DEFINE4(pkey_mprotect, unsigned long, start, size_t, len,
 830                 unsigned long, prot, int, pkey)
 831 {
 832         return do_mprotect_pkey(start, len, prot, pkey);
 833 }
 834
 835 SYSCALL_DEFINE2(pkey_alloc, unsigned long, flags, unsigned long, init_val)
 836 {
 837         int pkey;
 838         int ret;
 839
 840         /* No flags supported yet. */
 841         if (flags)
 842                 return -EINVAL;
 843         /* check for unsupported init values */
 844         if (init_val & ~PKEY_ACCESS_MASK)
 845                 return -EINVAL;
 846
 847         mmap_write_lock(current->mm);
 848         pkey = mm_pkey_alloc(current->mm);
 849
 850         ret = -ENOSPC;
 851         if (pkey == -1)
 852                 goto out;
 853
 854         ret = arch_set_user_pkey_access(current, pkey, init_val);
 855         if (ret) {
 856                 mm_pkey_free(current->mm, pkey);
 857                 goto out;
 858         }
 859         ret = pkey;
 860 out:
 861         mmap_write_unlock(current->mm);
 862         return ret;
 863 }
 864
 865 SYSCALL_DEFINE1(pkey_free, int, pkey)
 866 {
 867         int ret;
 868
 869         mmap_write_lock(current->mm);
 870         ret = mm_pkey_free(current->mm, pkey);
 871         mmap_write_unlock(current->mm);
 872
 873         /*
 874          * We could provide warnings or errors if any VMA still
 875          * has the pkey set here.
 876          */
 877         return ret;
 878 }
 879
 880 #endif /* CONFIG_ARCH_HAS_PKEYS */