mm/mprotect.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  *  mm/mprotect.c
   4  *
   5  *  (C) Copyright 1994 Linus Torvalds
   6  *  (C) Copyright 2002 Christoph Hellwig
   7  *
   8  *  Address space accounting code       <alan@lxorguk.ukuu.org.uk>
   9  *  (C) Copyright 2002 Red Hat Inc, All Rights Reserved
  10  */
  11
  12 #include <linux/pagewalk.h>
  13 #include <linux/hugetlb.h>
  14 #include <linux/shm.h>
  15 #include <linux/mman.h>
  16 #include <linux/fs.h>
  17 #include <linux/highmem.h>
  18 #include <linux/security.h>
  19 #include <linux/mempolicy.h>
  20 #include <linux/personality.h>
  21 #include <linux/syscalls.h>
  22 #include <linux/swap.h>
  23 #include <linux/swapops.h>
  24 #include <linux/mmu_notifier.h>
  25 #include <linux/migrate.h>
  26 #include <linux/perf_event.h>
  27 #include <linux/pkeys.h>
  28 #include <linux/ksm.h>
  29 #include <linux/uaccess.h>
  30 #include <linux/mm_inline.h>
  31 #include <linux/pgtable.h>
  32 #include <linux/sched/sysctl.h>
  33 #include <linux/userfaultfd_k.h>
  34 #include <linux/memory-tiers.h>
  35 #include <asm/cacheflush.h>
  36 #include <asm/mmu_context.h>
  37 #include <asm/tlbflush.h>
  38 #include <asm/tlb.h>
  39
  40 #include "internal.h"
  41
  42 bool can_change_pte_writable(struct vm_area_struct *vma, unsigned long addr,
  43                              pte_t pte)
  44 {
  45         struct page *page;
  46
  47         if (WARN_ON_ONCE(!(vma->vm_flags & VM_WRITE)))
  48                 return false;
  49
  50         /* Don't touch entries that are not even readable. */
  51         if (pte_protnone(pte))
  52                 return false;
  53
  54         /* Do we need write faults for softdirty tracking? */
  55         if (vma_soft_dirty_enabled(vma) && !pte_soft_dirty(pte))
  56                 return false;
  57
  58         /* Do we need write faults for uffd-wp tracking? */
  59         if (userfaultfd_pte_wp(vma, pte))
  60                 return false;
  61
  62         if (!(vma->vm_flags & VM_SHARED)) {
  63                 /*
  64                  * Writable MAP_PRIVATE mapping: We can only special-case on
  65                  * exclusive anonymous pages, because we know that our
  66                  * write-fault handler similarly would map them writable without
  67                  * any additional checks while holding the PT lock.
  68                  */
  69                 page = vm_normal_page(vma, addr, pte);
  70                 return page && PageAnon(page) && PageAnonExclusive(page);
  71         }
  72
  73         /*
  74          * Writable MAP_SHARED mapping: "clean" might indicate that the FS still
  75          * needs a real write-fault for writenotify
  76          * (see vma_wants_writenotify()). If "dirty", the assumption is that the
  77          * FS was already notified and we can simply mark the PTE writable
  78          * just like the write-fault handler would do.
  79          */
  80         return pte_dirty(pte);
  81 }
  82
  83 static long change_pte_range(struct mmu_gather *tlb,
  84                 struct vm_area_struct *vma, pmd_t *pmd, unsigned long addr,
  85                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
  86 {
  87         pte_t *pte, oldpte;
  88         spinlock_t *ptl;
  89         long pages = 0;
  90         int target_node = NUMA_NO_NODE;
  91         bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
  92         bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
  93         bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
  94
  95         tlb_change_page_size(tlb, PAGE_SIZE);
  96         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
  97         if (!pte)
  98                 return -EAGAIN;
  99
 100         /* Get target node for single threaded private VMAs */
 101         if (prot_numa && !(vma->vm_flags & VM_SHARED) &&
 102             atomic_read(&vma->vm_mm->mm_users) == 1)
 103                 target_node = numa_node_id();
 104
 105         flush_tlb_batched_pending(vma->vm_mm);
 106         arch_enter_lazy_mmu_mode();
 107         do {
 108                 oldpte = ptep_get(pte);
 109                 if (pte_present(oldpte)) {
 110                         pte_t ptent;
 111
 112                         /*
 113                          * Avoid trapping faults against the zero or KSM
 114                          * pages. See similar comment in change_huge_pmd.
 115                          */
 116                         if (prot_numa) {
 117                                 struct page *page;
 118                                 int nid;
 119                                 bool toptier;
 120
 121                                 /* Avoid TLB flush if possible */
 122                                 if (pte_protnone(oldpte))
 123                                         continue;
 124
 125                                 page = vm_normal_page(vma, addr, oldpte);
 126                                 if (!page || is_zone_device_page(page) || PageKsm(page))
 127                                         continue;
 128
 129                                 /* Also skip shared copy-on-write pages */
 130                                 if (is_cow_mapping(vma->vm_flags) &&
 131                                     page_count(page) != 1)
 132                                         continue;
 133
 134                                 /*
 135                                  * While migration can move some dirty pages,
 136                                  * it cannot move them all from MIGRATE_ASYNC
 137                                  * context.
 138                                  */
 139                                 if (page_is_file_lru(page) && PageDirty(page))
 140                                         continue;
 141
 142                                 /*
 143                                  * Don't mess with PTEs if page is already on the node
 144                                  * a single-threaded process is running on.
 145                                  */
 146                                 nid = page_to_nid(page);
 147                                 if (target_node == nid)
 148                                         continue;
 149                                 toptier = node_is_toptier(nid);
 150
 151                                 /*
 152                                  * Skip scanning top tier node if normal numa
 153                                  * balancing is disabled
 154                                  */
 155                                 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_NORMAL) &&
 156                                     toptier)
 157                                         continue;
 158                                 if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING &&
 159                                     !toptier)
 160                                         xchg_page_access_time(page,
 161                                                 jiffies_to_msecs(jiffies));
 162                         }
 163
 164                         oldpte = ptep_modify_prot_start(vma, addr, pte);
 165                         ptent = pte_modify(oldpte, newprot);
 166
 167                         if (uffd_wp)
 168                                 ptent = pte_mkuffd_wp(ptent);
 169                         else if (uffd_wp_resolve)
 170                                 ptent = pte_clear_uffd_wp(ptent);
 171
 172                         /*
 173                          * In some writable, shared mappings, we might want
 174                          * to catch actual write access -- see
 175                          * vma_wants_writenotify().
 176                          *
 177                          * In all writable, private mappings, we have to
 178                          * properly handle COW.
 179                          *
 180                          * In both cases, we can sometimes still change PTEs
 181                          * writable and avoid the write-fault handler, for
 182                          * example, if a PTE is already dirty and no other
 183                          * COW or special handling is required.
 184                          */
 185                         if ((cp_flags & MM_CP_TRY_CHANGE_WRITABLE) &&
 186                             !pte_write(ptent) &&
 187                             can_change_pte_writable(vma, addr, ptent))
 188                                 ptent = pte_mkwrite(ptent);
 189
 190                         ptep_modify_prot_commit(vma, addr, pte, oldpte, ptent);
 191                         if (pte_needs_flush(oldpte, ptent))
 192                                 tlb_flush_pte_range(tlb, addr, PAGE_SIZE);
 193                         pages++;
 194                 } else if (is_swap_pte(oldpte)) {
 195                         swp_entry_t entry = pte_to_swp_entry(oldpte);
 196                         pte_t newpte;
 197
 198                         if (is_writable_migration_entry(entry)) {
 199                                 struct page *page = pfn_swap_entry_to_page(entry);
 200
 201                                 /*
 202                                  * A protection check is difficult so
 203                                  * just be safe and disable write
 204                                  */
 205                                 if (PageAnon(page))
 206                                         entry = make_readable_exclusive_migration_entry(
 207                                                              swp_offset(entry));
 208                                 else
 209                                         entry = make_readable_migration_entry(swp_offset(entry));
 210                                 newpte = swp_entry_to_pte(entry);
 211                                 if (pte_swp_soft_dirty(oldpte))
 212                                         newpte = pte_swp_mksoft_dirty(newpte);
 213                         } else if (is_writable_device_private_entry(entry)) {
 214                                 /*
 215                                  * We do not preserve soft-dirtiness. See
 216                                  * copy_one_pte() for explanation.
 217                                  */
 218                                 entry = make_readable_device_private_entry(
 219                                                         swp_offset(entry));
 220                                 newpte = swp_entry_to_pte(entry);
 221                                 if (pte_swp_uffd_wp(oldpte))
 222                                         newpte = pte_swp_mkuffd_wp(newpte);
 223                         } else if (is_writable_device_exclusive_entry(entry)) {
 224                                 entry = make_readable_device_exclusive_entry(
 225                                                         swp_offset(entry));
 226                                 newpte = swp_entry_to_pte(entry);
 227                                 if (pte_swp_soft_dirty(oldpte))
 228                                         newpte = pte_swp_mksoft_dirty(newpte);
 229                                 if (pte_swp_uffd_wp(oldpte))
 230                                         newpte = pte_swp_mkuffd_wp(newpte);
 231                         } else if (is_pte_marker_entry(entry)) {
 232                                 /*
 233                                  * Ignore swapin errors unconditionally,
 234                                  * because any access should sigbus anyway.
 235                                  */
 236                                 if (is_swapin_error_entry(entry))
 237                                         continue;
 238                                 /*
 239                                  * If this is uffd-wp pte marker and we'd like
 240                                  * to unprotect it, drop it; the next page
 241                                  * fault will trigger without uffd trapping.
 242                                  */
 243                                 if (uffd_wp_resolve) {
 244                                         pte_clear(vma->vm_mm, addr, pte);
 245                                         pages++;
 246                                 }
 247                                 continue;
 248                         } else {
 249                                 newpte = oldpte;
 250                         }
 251
 252                         if (uffd_wp)
 253                                 newpte = pte_swp_mkuffd_wp(newpte);
 254                         else if (uffd_wp_resolve)
 255                                 newpte = pte_swp_clear_uffd_wp(newpte);
 256
 257                         if (!pte_same(oldpte, newpte)) {
 258                                 set_pte_at(vma->vm_mm, addr, pte, newpte);
 259                                 pages++;
 260                         }
 261                 } else {
 262                         /* It must be an none page, or what else?.. */
 263                         WARN_ON_ONCE(!pte_none(oldpte));
 264
 265                         /*
 266                          * Nobody plays with any none ptes besides
 267                          * userfaultfd when applying the protections.
 268                          */
 269                         if (likely(!uffd_wp))
 270                                 continue;
 271
 272                         if (userfaultfd_wp_use_markers(vma)) {
 273                                 /*
 274                                  * For file-backed mem, we need to be able to
 275                                  * wr-protect a none pte, because even if the
 276                                  * pte is none, the page/swap cache could
 277                                  * exist.  Doing that by install a marker.
 278                                  */
 279                                 set_pte_at(vma->vm_mm, addr, pte,
 280                                            make_pte_marker(PTE_MARKER_UFFD_WP));
 281                                 pages++;
 282                         }
 283                 }
 284         } while (pte++, addr += PAGE_SIZE, addr != end);
 285         arch_leave_lazy_mmu_mode();
 286         pte_unmap_unlock(pte - 1, ptl);
 287
 288         return pages;
 289 }
 290
 291 /*
 292  * Return true if we want to split THPs into PTE mappings in change
 293  * protection procedure, false otherwise.
 294  */
 295 static inline bool
 296 pgtable_split_needed(struct vm_area_struct *vma, unsigned long cp_flags)
 297 {
 298         /*
 299          * pte markers only resides in pte level, if we need pte markers,
 300          * we need to split.  We cannot wr-protect shmem thp because file
 301          * thp is handled differently when split by erasing the pmd so far.
 302          */
 303         return (cp_flags & MM_CP_UFFD_WP) && !vma_is_anonymous(vma);
 304 }
 305
 306 /*
 307  * Return true if we want to populate pgtables in change protection
 308  * procedure, false otherwise
 309  */
 310 static inline bool
 311 pgtable_populate_needed(struct vm_area_struct *vma, unsigned long cp_flags)
 312 {
 313         /* If not within ioctl(UFFDIO_WRITEPROTECT), then don't bother */
 314         if (!(cp_flags & MM_CP_UFFD_WP))
 315                 return false;
 316
 317         /* Populate if the userfaultfd mode requires pte markers */
 318         return userfaultfd_wp_use_markers(vma);
 319 }
 320
 321 /*
 322  * Populate the pgtable underneath for whatever reason if requested.
 323  * When {pte|pmd|...}_alloc() failed we treat it the same way as pgtable
 324  * allocation failures during page faults by kicking OOM and returning
 325  * error.
 326  */
 327 #define  change_pmd_prepare(vma, pmd, cp_flags)                         \
 328         ({                                                              \
 329                 long err = 0;                                           \
 330                 if (unlikely(pgtable_populate_needed(vma, cp_flags))) { \
 331                         if (pte_alloc(vma->vm_mm, pmd))                 \
 332                                 err = -ENOMEM;                          \
 333                 }                                                       \
 334                 err;                                                    \
 335         })
 336
 337 /*
 338  * This is the general pud/p4d/pgd version of change_pmd_prepare(). We need to
 339  * have separate change_pmd_prepare() because pte_alloc() returns 0 on success,
 340  * while {pmd|pud|p4d}_alloc() returns the valid pointer on success.
 341  */
 342 #define  change_prepare(vma, high, low, addr, cp_flags)                 \
 343           ({                                                            \
 344                 long err = 0;                                           \
 345                 if (unlikely(pgtable_populate_needed(vma, cp_flags))) { \
 346                         low##_t *p = low##_alloc(vma->vm_mm, high, addr); \
 347                         if (p == NULL)                                  \
 348                                 err = -ENOMEM;                          \
 349                 }                                                       \
 350                 err;                                                    \
 351         })
 352
 353 static inline long change_pmd_range(struct mmu_gather *tlb,
 354                 struct vm_area_struct *vma, pud_t *pud, unsigned long addr,
 355                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
 356 {
 357         pmd_t *pmd;
 358         unsigned long next;
 359         long pages = 0;
 360         unsigned long nr_huge_updates = 0;
 361         struct mmu_notifier_range range;
 362
 363         range.start = 0;
 364
 365         pmd = pmd_offset(pud, addr);
 366         do {
 367                 long ret;
 368                 pmd_t _pmd;
 369 again:
 370                 next = pmd_addr_end(addr, end);
 371
 372                 ret = change_pmd_prepare(vma, pmd, cp_flags);
 373                 if (ret) {
 374                         pages = ret;
 375                         break;
 376                 }
 377
 378                 if (pmd_none(*pmd))
 379                         goto next;
 380
 381                 /* invoke the mmu notifier if the pmd is populated */
 382                 if (!range.start) {
 383                         mmu_notifier_range_init(&range,
 384                                 MMU_NOTIFY_PROTECTION_VMA, 0,
 385                                 vma->vm_mm, addr, end);
 386                         mmu_notifier_invalidate_range_start(&range);
 387                 }
 388
 389                 _pmd = pmdp_get_lockless(pmd);
 390                 if (is_swap_pmd(_pmd) || pmd_trans_huge(_pmd) || pmd_devmap(_pmd)) {
 391                         if ((next - addr != HPAGE_PMD_SIZE) ||
 392                             pgtable_split_needed(vma, cp_flags)) {
 393                                 __split_huge_pmd(vma, pmd, addr, false, NULL);
 394                                 /*
 395                                  * For file-backed, the pmd could have been
 396                                  * cleared; make sure pmd populated if
 397                                  * necessary, then fall-through to pte level.
 398                                  */
 399                                 ret = change_pmd_prepare(vma, pmd, cp_flags);
 400                                 if (ret) {
 401                                         pages = ret;
 402                                         break;
 403                                 }
 404                         } else {
 405                                 ret = change_huge_pmd(tlb, vma, pmd,
 406                                                 addr, newprot, cp_flags);
 407                                 if (ret) {
 408                                         if (ret == HPAGE_PMD_NR) {
 409                                                 pages += HPAGE_PMD_NR;
 410                                                 nr_huge_updates++;
 411                                         }
 412
 413                                         /* huge pmd was handled */
 414                                         goto next;
 415                                 }
 416                         }
 417                         /* fall through, the trans huge pmd just split */
 418                 }
 419
 420                 ret = change_pte_range(tlb, vma, pmd, addr, next, newprot,
 421                                        cp_flags);
 422                 if (ret < 0)
 423                         goto again;
 424                 pages += ret;
 425 next:
 426                 cond_resched();
 427         } while (pmd++, addr = next, addr != end);
 428
 429         if (range.start)
 430                 mmu_notifier_invalidate_range_end(&range);
 431
 432         if (nr_huge_updates)
 433                 count_vm_numa_events(NUMA_HUGE_PTE_UPDATES, nr_huge_updates);
 434         return pages;
 435 }
 436
 437 static inline long change_pud_range(struct mmu_gather *tlb,
 438                 struct vm_area_struct *vma, p4d_t *p4d, unsigned long addr,
 439                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
 440 {
 441         pud_t *pud;
 442         unsigned long next;
 443         long pages = 0, ret;
 444
 445         pud = pud_offset(p4d, addr);
 446         do {
 447                 next = pud_addr_end(addr, end);
 448                 ret = change_prepare(vma, pud, pmd, addr, cp_flags);
 449                 if (ret)
 450                         return ret;
 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 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         long pages = 0, ret;
 467
 468         p4d = p4d_offset(pgd, addr);
 469         do {
 470                 next = p4d_addr_end(addr, end);
 471                 ret = change_prepare(vma, p4d, pud, addr, cp_flags);
 472                 if (ret)
 473                         return ret;
 474                 if (p4d_none_or_clear_bad(p4d))
 475                         continue;
 476                 pages += change_pud_range(tlb, vma, p4d, addr, next, newprot,
 477                                           cp_flags);
 478         } while (p4d++, addr = next, addr != end);
 479
 480         return pages;
 481 }
 482
 483 static long change_protection_range(struct mmu_gather *tlb,
 484                 struct vm_area_struct *vma, unsigned long addr,
 485                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
 486 {
 487         struct mm_struct *mm = vma->vm_mm;
 488         pgd_t *pgd;
 489         unsigned long next;
 490         long pages = 0, ret;
 491
 492         BUG_ON(addr >= end);
 493         pgd = pgd_offset(mm, addr);
 494         tlb_start_vma(tlb, vma);
 495         do {
 496                 next = pgd_addr_end(addr, end);
 497                 ret = change_prepare(vma, pgd, p4d, addr, cp_flags);
 498                 if (ret) {
 499                         pages = ret;
 500                         break;
 501                 }
 502                 if (pgd_none_or_clear_bad(pgd))
 503                         continue;
 504                 pages += change_p4d_range(tlb, vma, pgd, addr, next, newprot,
 505                                           cp_flags);
 506         } while (pgd++, addr = next, addr != end);
 507
 508         tlb_end_vma(tlb, vma);
 509
 510         return pages;
 511 }
 512
 513 long change_protection(struct mmu_gather *tlb,
 514                        struct vm_area_struct *vma, unsigned long start,
 515                        unsigned long end, unsigned long cp_flags)
 516 {
 517         pgprot_t newprot = vma->vm_page_prot;
 518         long pages;
 519
 520         BUG_ON((cp_flags & MM_CP_UFFD_WP_ALL) == MM_CP_UFFD_WP_ALL);
 521
 522 #ifdef CONFIG_NUMA_BALANCING
 523         /*
 524          * Ordinary protection updates (mprotect, uffd-wp, softdirty tracking)
 525          * are expected to reflect their requirements via VMA flags such that
 526          * vma_set_page_prot() will adjust vma->vm_page_prot accordingly.
 527          */
 528         if (cp_flags & MM_CP_PROT_NUMA)
 529                 newprot = PAGE_NONE;
 530 #else
 531         WARN_ON_ONCE(cp_flags & MM_CP_PROT_NUMA);
 532 #endif
 533
 534         if (is_vm_hugetlb_page(vma))
 535                 pages = hugetlb_change_protection(vma, start, end, newprot,
 536                                                   cp_flags);
 537         else
 538                 pages = change_protection_range(tlb, vma, start, end, newprot,
 539                                                 cp_flags);
 540
 541         return pages;
 542 }
 543
 544 static int prot_none_pte_entry(pte_t *pte, unsigned long addr,
 545                                unsigned long next, struct mm_walk *walk)
 546 {
 547         return pfn_modify_allowed(pte_pfn(ptep_get(pte)),
 548                                   *(pgprot_t *)(walk->private)) ?
 549                 0 : -EACCES;
 550 }
 551
 552 static int prot_none_hugetlb_entry(pte_t *pte, unsigned long hmask,
 553                                    unsigned long addr, unsigned long next,
 554                                    struct mm_walk *walk)
 555 {
 556         return pfn_modify_allowed(pte_pfn(ptep_get(pte)),
 557                                   *(pgprot_t *)(walk->private)) ?
 558                 0 : -EACCES;
 559 }
 560
 561 static int prot_none_test(unsigned long addr, unsigned long next,
 562                           struct mm_walk *walk)
 563 {
 564         return 0;
 565 }
 566
 567 static const struct mm_walk_ops prot_none_walk_ops = {
 568         .pte_entry              = prot_none_pte_entry,
 569         .hugetlb_entry          = prot_none_hugetlb_entry,
 570         .test_walk              = prot_none_test,
 571 };
 572
 573 int
 574 mprotect_fixup(struct vma_iterator *vmi, struct mmu_gather *tlb,
 575                struct vm_area_struct *vma, struct vm_area_struct **pprev,
 576                unsigned long start, unsigned long end, unsigned long newflags)
 577 {
 578         struct mm_struct *mm = vma->vm_mm;
 579         unsigned long oldflags = vma->vm_flags;
 580         long nrpages = (end - start) >> PAGE_SHIFT;
 581         unsigned int mm_cp_flags = 0;
 582         unsigned long charged = 0;
 583         pgoff_t pgoff;
 584         int error;
 585
 586         if (newflags == oldflags) {
 587                 *pprev = vma;
 588                 return 0;
 589         }
 590
 591         /*
 592          * Do PROT_NONE PFN permission checks here when we can still
 593          * bail out without undoing a lot of state. This is a rather
 594          * uncommon case, so doesn't need to be very optimized.
 595          */
 596         if (arch_has_pfn_modify_check() &&
 597             (vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 598             (newflags & VM_ACCESS_FLAGS) == 0) {
 599                 pgprot_t new_pgprot = vm_get_page_prot(newflags);
 600
 601                 error = walk_page_range(current->mm, start, end,
 602                                 &prot_none_walk_ops, &new_pgprot);
 603                 if (error)
 604                         return error;
 605         }
 606
 607         /*
 608          * If we make a private mapping writable we increase our commit;
 609          * but (without finer accounting) cannot reduce our commit if we
 610          * make it unwritable again. hugetlb mapping were accounted for
 611          * even if read-only so there is no need to account for them here
 612          */
 613         if (newflags & VM_WRITE) {
 614                 /* Check space limits when area turns into data. */
 615                 if (!may_expand_vm(mm, newflags, nrpages) &&
 616                                 may_expand_vm(mm, oldflags, nrpages))
 617                         return -ENOMEM;
 618                 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
 619                                                 VM_SHARED|VM_NORESERVE))) {
 620                         charged = nrpages;
 621                         if (security_vm_enough_memory_mm(mm, charged))
 622                                 return -ENOMEM;
 623                         newflags |= VM_ACCOUNT;
 624                 }
 625         }
 626
 627         /*
 628          * First try to merge with previous and/or next vma.
 629          */
 630         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 631         *pprev = vma_merge(vmi, mm, *pprev, start, end, newflags,
 632                            vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
 633                            vma->vm_userfaultfd_ctx, anon_vma_name(vma));
 634         if (*pprev) {
 635                 vma = *pprev;
 636                 VM_WARN_ON((vma->vm_flags ^ newflags) & ~VM_SOFTDIRTY);
 637                 goto success;
 638         }
 639
 640         *pprev = vma;
 641
 642         if (start != vma->vm_start) {
 643                 error = split_vma(vmi, vma, start, 1);
 644                 if (error)
 645                         goto fail;
 646         }
 647
 648         if (end != vma->vm_end) {
 649                 error = split_vma(vmi, vma, end, 0);
 650                 if (error)
 651                         goto fail;
 652         }
 653
 654 success:
 655         /*
 656          * vm_flags and vm_page_prot are protected by the mmap_lock
 657          * held in write mode.
 658          */
 659         vm_flags_reset(vma, newflags);
 660         if (vma_wants_manual_pte_write_upgrade(vma))
 661                 mm_cp_flags |= MM_CP_TRY_CHANGE_WRITABLE;
 662         vma_set_page_prot(vma);
 663
 664         change_protection(tlb, vma, start, end, mm_cp_flags);
 665
 666         /*
 667          * Private VM_LOCKED VMA becoming writable: trigger COW to avoid major
 668          * fault on access.
 669          */
 670         if ((oldflags & (VM_WRITE | VM_SHARED | VM_LOCKED)) == VM_LOCKED &&
 671                         (newflags & VM_WRITE)) {
 672                 populate_vma_page_range(vma, start, end, NULL);
 673         }
 674
 675         vm_stat_account(mm, oldflags, -nrpages);
 676         vm_stat_account(mm, newflags, nrpages);
 677         perf_event_mmap(vma);
 678         return 0;
 679
 680 fail:
 681         vm_unacct_memory(charged);
 682         return error;
 683 }
 684
 685 /*
 686  * pkey==-1 when doing a legacy mprotect()
 687  */
 688 static int do_mprotect_pkey(unsigned long start, size_t len,
 689                 unsigned long prot, int pkey)
 690 {
 691         unsigned long nstart, end, tmp, reqprot;
 692         struct vm_area_struct *vma, *prev;
 693         int error;
 694         const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
 695         const bool rier = (current->personality & READ_IMPLIES_EXEC) &&
 696                                 (prot & PROT_READ);
 697         struct mmu_gather tlb;
 698         struct vma_iterator vmi;
 699
 700         start = untagged_addr(start);
 701
 702         prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
 703         if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
 704                 return -EINVAL;
 705
 706         if (start & ~PAGE_MASK)
 707                 return -EINVAL;
 708         if (!len)
 709                 return 0;
 710         len = PAGE_ALIGN(len);
 711         end = start + len;
 712         if (end <= start)
 713                 return -ENOMEM;
 714         if (!arch_validate_prot(prot, start))
 715                 return -EINVAL;
 716
 717         reqprot = prot;
 718
 719         if (mmap_write_lock_killable(current->mm))
 720                 return -EINTR;
 721
 722         /*
 723          * If userspace did not allocate the pkey, do not let
 724          * them use it here.
 725          */
 726         error = -EINVAL;
 727         if ((pkey != -1) && !mm_pkey_is_allocated(current->mm, pkey))
 728                 goto out;
 729
 730         vma_iter_init(&vmi, current->mm, start);
 731         vma = vma_find(&vmi, end);
 732         error = -ENOMEM;
 733         if (!vma)
 734                 goto out;
 735
 736         if (unlikely(grows & PROT_GROWSDOWN)) {
 737                 if (vma->vm_start >= end)
 738                         goto out;
 739                 start = vma->vm_start;
 740                 error = -EINVAL;
 741                 if (!(vma->vm_flags & VM_GROWSDOWN))
 742                         goto out;
 743         } else {
 744                 if (vma->vm_start > start)
 745                         goto out;
 746                 if (unlikely(grows & PROT_GROWSUP)) {
 747                         end = vma->vm_end;
 748                         error = -EINVAL;
 749                         if (!(vma->vm_flags & VM_GROWSUP))
 750                                 goto out;
 751                 }
 752         }
 753
 754         prev = vma_prev(&vmi);
 755         if (start > vma->vm_start)
 756                 prev = vma;
 757
 758         tlb_gather_mmu(&tlb, current->mm);
 759         nstart = start;
 760         tmp = vma->vm_start;
 761         for_each_vma_range(vmi, vma, end) {
 762                 unsigned long mask_off_old_flags;
 763                 unsigned long newflags;
 764                 int new_vma_pkey;
 765
 766                 if (vma->vm_start != tmp) {
 767                         error = -ENOMEM;
 768                         break;
 769                 }
 770
 771                 /* Does the application expect PROT_READ to imply PROT_EXEC */
 772                 if (rier && (vma->vm_flags & VM_MAYEXEC))
 773                         prot |= PROT_EXEC;
 774
 775                 /*
 776                  * Each mprotect() call explicitly passes r/w/x permissions.
 777                  * If a permission is not passed to mprotect(), it must be
 778                  * cleared from the VMA.
 779                  */
 780                 mask_off_old_flags = VM_ACCESS_FLAGS | VM_FLAGS_CLEAR;
 781
 782                 new_vma_pkey = arch_override_mprotect_pkey(vma, prot, pkey);
 783                 newflags = calc_vm_prot_bits(prot, new_vma_pkey);
 784                 newflags |= (vma->vm_flags & ~mask_off_old_flags);
 785
 786                 /* newflags >> 4 shift VM_MAY% in place of VM_% */
 787                 if ((newflags & ~(newflags >> 4)) & VM_ACCESS_FLAGS) {
 788                         error = -EACCES;
 789                         break;
 790                 }
 791
 792                 if (map_deny_write_exec(vma, newflags)) {
 793                         error = -EACCES;
 794                         break;
 795                 }
 796
 797                 /* Allow architectures to sanity-check the new flags */
 798                 if (!arch_validate_flags(newflags)) {
 799                         error = -EINVAL;
 800                         break;
 801                 }
 802
 803                 error = security_file_mprotect(vma, reqprot, prot);
 804                 if (error)
 805                         break;
 806
 807                 tmp = vma->vm_end;
 808                 if (tmp > end)
 809                         tmp = end;
 810
 811                 if (vma->vm_ops && vma->vm_ops->mprotect) {
 812                         error = vma->vm_ops->mprotect(vma, nstart, tmp, newflags);
 813                         if (error)
 814                                 break;
 815                 }
 816
 817                 error = mprotect_fixup(&vmi, &tlb, vma, &prev, nstart, tmp, newflags);
 818                 if (error)
 819                         break;
 820
 821                 tmp = vma_iter_end(&vmi);
 822                 nstart = tmp;
 823                 prot = reqprot;
 824         }
 825         tlb_finish_mmu(&tlb);
 826
 827         if (!error && tmp < end)
 828                 error = -ENOMEM;
 829
 830 out:
 831         mmap_write_unlock(current->mm);
 832         return error;
 833 }
 834
 835 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
 836                 unsigned long, prot)
 837 {
 838         return do_mprotect_pkey(start, len, prot, -1);
 839 }
 840
 841 #ifdef CONFIG_ARCH_HAS_PKEYS
 842
 843 SYSCALL_DEFINE4(pkey_mprotect, unsigned long, start, size_t, len,
 844                 unsigned long, prot, int, pkey)
 845 {
 846         return do_mprotect_pkey(start, len, prot, pkey);
 847 }
 848
 849 SYSCALL_DEFINE2(pkey_alloc, unsigned long, flags, unsigned long, init_val)
 850 {
 851         int pkey;
 852         int ret;
 853
 854         /* No flags supported yet. */
 855         if (flags)
 856                 return -EINVAL;
 857         /* check for unsupported init values */
 858         if (init_val & ~PKEY_ACCESS_MASK)
 859                 return -EINVAL;
 860
 861         mmap_write_lock(current->mm);
 862         pkey = mm_pkey_alloc(current->mm);
 863
 864         ret = -ENOSPC;
 865         if (pkey == -1)
 866                 goto out;
 867
 868         ret = arch_set_user_pkey_access(current, pkey, init_val);
 869         if (ret) {
 870                 mm_pkey_free(current->mm, pkey);
 871                 goto out;
 872         }
 873         ret = pkey;
 874 out:
 875         mmap_write_unlock(current->mm);
 876         return ret;
 877 }
 878
 879 SYSCALL_DEFINE1(pkey_free, int, pkey)
 880 {
 881         int ret;
 882
 883         mmap_write_lock(current->mm);
 884         ret = mm_pkey_free(current->mm, pkey);
 885         mmap_write_unlock(current->mm);
 886
 887         /*
 888          * We could provide warnings or errors if any VMA still
 889          * has the pkey set here.
 890          */
 891         return ret;
 892 }
 893
 894 #endif /* CONFIG_ARCH_HAS_PKEYS */