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       <[email protected]>
   9  *  (C) Copyright 2002 Red Hat Inc, All Rights Reserved
  10  */
  11
  12 #include <linux/pagewalk.h>
  13 #include <linux/hugetlb.h>
  14 #include <linux/shm.h>
  15 #include <linux/mman.h>
  16 #include <linux/fs.h>
  17 #include <linux/highmem.h>
  18 #include <linux/security.h>
  19 #include <linux/mempolicy.h>
  20 #include <linux/personality.h>
  21 #include <linux/syscalls.h>
  22 #include <linux/swap.h>
  23 #include <linux/swapops.h>
  24 #include <linux/mmu_notifier.h>
  25 #include <linux/migrate.h>
  26 #include <linux/perf_event.h>
  27 #include <linux/pkeys.h>
  28 #include <linux/ksm.h>
  29 #include <linux/uaccess.h>
  30 #include <linux/mm_inline.h>
  31 #include <linux/pgtable.h>
  32 #include <linux/sched/sysctl.h>
  33 #include <linux/userfaultfd_k.h>
  34 #include <linux/memory-tiers.h>
  35 #include <asm/cacheflush.h>
  36 #include <asm/mmu_context.h>
  37 #include <asm/tlbflush.h>
  38 #include <asm/tlb.h>
  39
  40 #include "internal.h"
  41
  42 bool can_change_pte_writable(struct vm_area_struct *vma, unsigned long addr,
  43                              pte_t pte)
  44 {
  45         struct page *page;
  46
  47         if (WARN_ON_ONCE(!(vma->vm_flags & VM_WRITE)))
  48                 return false;
  49
  50         /* Don't touch entries that are not even readable. */
  51         if (pte_protnone(pte))
  52                 return false;
  53
  54         /* Do we need write faults for softdirty tracking? */
  55         if (vma_soft_dirty_enabled(vma) && !pte_soft_dirty(pte))
  56                 return false;
  57
  58         /* Do we need write faults for uffd-wp tracking? */
  59         if (userfaultfd_pte_wp(vma, pte))
  60                 return false;
  61
  62         if (!(vma->vm_flags & VM_SHARED)) {
  63                 /*
  64                  * Writable MAP_PRIVATE mapping: We can only special-case on
  65                  * exclusive anonymous pages, because we know that our
  66                  * write-fault handler similarly would map them writable without
  67                  * any additional checks while holding the PT lock.
  68                  */
  69                 page = vm_normal_page(vma, addr, pte);
  70                 return page && PageAnon(page) && PageAnonExclusive(page);
  71         }
  72
  73         /*
  74          * Writable MAP_SHARED mapping: "clean" might indicate that the FS still
  75          * needs a real write-fault for writenotify
  76          * (see vma_wants_writenotify()). If "dirty", the assumption is that the
  77          * FS was already notified and we can simply mark the PTE writable
  78          * just like the write-fault handler would do.
  79          */
  80         return pte_dirty(pte);
  81 }
  82
  83 static long change_pte_range(struct mmu_gather *tlb,
  84                 struct vm_area_struct *vma, pmd_t *pmd, unsigned long addr,
  85                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
  86 {
  87         pte_t *pte, oldpte;
  88         spinlock_t *ptl;
  89         long pages = 0;
  90         int target_node = NUMA_NO_NODE;
  91         bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
  92         bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
  93         bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
  94
  95         tlb_change_page_size(tlb, PAGE_SIZE);
  96
  97         /*
  98          * Can be called with only the mmap_lock for reading by
  99          * prot_numa so we must check the pmd isn't constantly
 100          * changing from under us from pmd_none to pmd_trans_huge
 101          * and/or the other way around.
 102          */
 103         if (pmd_trans_unstable(pmd))
 104                 return 0;
 105
 106         /*
 107          * The pmd points to a regular pte so the pmd can't change
 108          * from under us even if the mmap_lock is only hold for
 109          * reading.
 110          */
 111         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
 112
 113         /* Get target node for single threaded private VMAs */
 114         if (prot_numa && !(vma->vm_flags & VM_SHARED) &&
 115             atomic_read(&vma->vm_mm->mm_users) == 1)
 116                 target_node = numa_node_id();
 117
 118         flush_tlb_batched_pending(vma->vm_mm);
 119         arch_enter_lazy_mmu_mode();
 120         do {
 121                 oldpte = *pte;
 122                 if (pte_present(oldpte)) {
 123                         pte_t ptent;
 124
 125                         /*
 126                          * Avoid trapping faults against the zero or KSM
 127                          * pages. See similar comment in change_huge_pmd.
 128                          */
 129                         if (prot_numa) {
 130                                 struct page *page;
 131                                 int nid;
 132                                 bool toptier;
 133
 134                                 /* Avoid TLB flush if possible */
 135                                 if (pte_protnone(oldpte))
 136                                         continue;
 137
 138                                 page = vm_normal_page(vma, addr, oldpte);
 139                                 if (!page || is_zone_device_page(page) || PageKsm(page))
 140                                         continue;
 141
 142                                 /* Also skip shared copy-on-write pages */
 143                                 if (is_cow_mapping(vma->vm_flags) &&
 144                                     page_count(page) != 1)
 145                                         continue;
 146
 147                                 /*
 148                                  * While migration can move some dirty pages,
 149                                  * it cannot move them all from MIGRATE_ASYNC
 150                                  * context.
 151                                  */
 152                                 if (page_is_file_lru(page) && PageDirty(page))
 153                                         continue;
 154
 155                                 /*
 156                                  * Don't mess with PTEs if page is already on the node
 157                                  * a single-threaded process is running on.
 158                                  */
 159                                 nid = page_to_nid(page);
 160                                 if (target_node == nid)
 161                                         continue;
 162                                 toptier = node_is_toptier(nid);
 163
 164                                 /*
 165                                  * Skip scanning top tier node if normal numa
 166                                  * balancing is disabled
 167                                  */
 168                                 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_NORMAL) &&
 169                                     toptier)
 170                                         continue;
 171                                 if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING &&
 172                                     !toptier)
 173                                         xchg_page_access_time(page,
 174                                                 jiffies_to_msecs(jiffies));
 175                         }
 176
 177                         oldpte = ptep_modify_prot_start(vma, addr, pte);
 178                         ptent = pte_modify(oldpte, newprot);
 179
 180                         if (uffd_wp)
 181                                 ptent = pte_mkuffd_wp(ptent);
 182                         else if (uffd_wp_resolve)
 183                                 ptent = pte_clear_uffd_wp(ptent);
 184
 185                         /*
 186                          * In some writable, shared mappings, we might want
 187                          * to catch actual write access -- see
 188                          * vma_wants_writenotify().
 189                          *
 190                          * In all writable, private mappings, we have to
 191                          * properly handle COW.
 192                          *
 193                          * In both cases, we can sometimes still change PTEs
 194                          * writable and avoid the write-fault handler, for
 195                          * example, if a PTE is already dirty and no other
 196                          * COW or special handling is required.
 197                          */
 198                         if ((cp_flags & MM_CP_TRY_CHANGE_WRITABLE) &&
 199                             !pte_write(ptent) &&
 200                             can_change_pte_writable(vma, addr, ptent))
 201                                 ptent = pte_mkwrite(ptent);
 202
 203                         ptep_modify_prot_commit(vma, addr, pte, oldpte, ptent);
 204                         if (pte_needs_flush(oldpte, ptent))
 205                                 tlb_flush_pte_range(tlb, addr, PAGE_SIZE);
 206                         pages++;
 207                 } else if (is_swap_pte(oldpte)) {
 208                         swp_entry_t entry = pte_to_swp_entry(oldpte);
 209                         pte_t newpte;
 210
 211                         if (is_writable_migration_entry(entry)) {
 212                                 struct page *page = pfn_swap_entry_to_page(entry);
 213
 214                                 /*
 215                                  * A protection check is difficult so
 216                                  * just be safe and disable write
 217                                  */
 218                                 if (PageAnon(page))
 219                                         entry = make_readable_exclusive_migration_entry(
 220                                                              swp_offset(entry));
 221                                 else
 222                                         entry = make_readable_migration_entry(swp_offset(entry));
 223                                 newpte = swp_entry_to_pte(entry);
 224                                 if (pte_swp_soft_dirty(oldpte))
 225                                         newpte = pte_swp_mksoft_dirty(newpte);
 226                         } else if (is_writable_device_private_entry(entry)) {
 227                                 /*
 228                                  * We do not preserve soft-dirtiness. See
 229                                  * copy_one_pte() for explanation.
 230                                  */
 231                                 entry = make_readable_device_private_entry(
 232                                                         swp_offset(entry));
 233                                 newpte = swp_entry_to_pte(entry);
 234                                 if (pte_swp_uffd_wp(oldpte))
 235                                         newpte = pte_swp_mkuffd_wp(newpte);
 236                         } else if (is_writable_device_exclusive_entry(entry)) {
 237                                 entry = make_readable_device_exclusive_entry(
 238                                                         swp_offset(entry));
 239                                 newpte = swp_entry_to_pte(entry);
 240                                 if (pte_swp_soft_dirty(oldpte))
 241                                         newpte = pte_swp_mksoft_dirty(newpte);
 242                                 if (pte_swp_uffd_wp(oldpte))
 243                                         newpte = pte_swp_mkuffd_wp(newpte);
 244                         } else if (is_pte_marker_entry(entry)) {
 245                                 /*
 246                                  * Ignore swapin errors unconditionally,
 247                                  * because any access should sigbus anyway.
 248                                  */
 249                                 if (is_swapin_error_entry(entry))
 250                                         continue;
 251                                 /*
 252                                  * If this is uffd-wp pte marker and we'd like
 253                                  * to unprotect it, drop it; the next page
 254                                  * fault will trigger without uffd trapping.
 255                                  */
 256                                 if (uffd_wp_resolve) {
 257                                         pte_clear(vma->vm_mm, addr, pte);
 258                                         pages++;
 259                                 }
 260                                 continue;
 261                         } else {
 262                                 newpte = oldpte;
 263                         }
 264
 265                         if (uffd_wp)
 266                                 newpte = pte_swp_mkuffd_wp(newpte);
 267                         else if (uffd_wp_resolve)
 268                                 newpte = pte_swp_clear_uffd_wp(newpte);
 269
 270                         if (!pte_same(oldpte, newpte)) {
 271                                 set_pte_at(vma->vm_mm, addr, pte, newpte);
 272                                 pages++;
 273                         }
 274                 } else {
 275                         /* It must be an none page, or what else?.. */
 276                         WARN_ON_ONCE(!pte_none(oldpte));
 277
 278                         /*
 279                          * Nobody plays with any none ptes besides
 280                          * userfaultfd when applying the protections.
 281                          */
 282                         if (likely(!uffd_wp))
 283                                 continue;
 284
 285                         if (userfaultfd_wp_use_markers(vma)) {
 286                                 /*
 287                                  * For file-backed mem, we need to be able to
 288                                  * wr-protect a none pte, because even if the
 289                                  * pte is none, the page/swap cache could
 290                                  * exist.  Doing that by install a marker.
 291                                  */
 292                                 set_pte_at(vma->vm_mm, addr, pte,
 293                                            make_pte_marker(PTE_MARKER_UFFD_WP));
 294                                 pages++;
 295                         }
 296                 }
 297         } while (pte++, addr += PAGE_SIZE, addr != end);
 298         arch_leave_lazy_mmu_mode();
 299         pte_unmap_unlock(pte - 1, ptl);
 300
 301         return pages;
 302 }
 303
 304 /*
 305  * Used when setting automatic NUMA hinting protection where it is
 306  * critical that a numa hinting PMD is not confused with a bad PMD.
 307  */
 308 static inline int pmd_none_or_clear_bad_unless_trans_huge(pmd_t *pmd)
 309 {
 310         pmd_t pmdval = pmdp_get_lockless(pmd);
 311
 312         if (pmd_none(pmdval))
 313                 return 1;
 314         if (pmd_trans_huge(pmdval))
 315                 return 0;
 316         if (unlikely(pmd_bad(pmdval))) {
 317                 pmd_clear_bad(pmd);
 318                 return 1;
 319         }
 320
 321         return 0;
 322 }
 323
 324 /*
 325  * Return true if we want to split THPs into PTE mappings in change
 326  * protection procedure, false otherwise.
 327  */
 328 static inline bool
 329 pgtable_split_needed(struct vm_area_struct *vma, unsigned long cp_flags)
 330 {
 331         /*
 332          * pte markers only resides in pte level, if we need pte markers,
 333          * we need to split.  We cannot wr-protect shmem thp because file
 334          * thp is handled differently when split by erasing the pmd so far.
 335          */
 336         return (cp_flags & MM_CP_UFFD_WP) && !vma_is_anonymous(vma);
 337 }
 338
 339 /*
 340  * Return true if we want to populate pgtables in change protection
 341  * procedure, false otherwise
 342  */
 343 static inline bool
 344 pgtable_populate_needed(struct vm_area_struct *vma, unsigned long cp_flags)
 345 {
 346         /* If not within ioctl(UFFDIO_WRITEPROTECT), then don't bother */
 347         if (!(cp_flags & MM_CP_UFFD_WP))
 348                 return false;
 349
 350         /* Populate if the userfaultfd mode requires pte markers */
 351         return userfaultfd_wp_use_markers(vma);
 352 }
 353
 354 /*
 355  * Populate the pgtable underneath for whatever reason if requested.
 356  * When {pte|pmd|...}_alloc() failed we treat it the same way as pgtable
 357  * allocation failures during page faults by kicking OOM and returning
 358  * error.
 359  */
 360 #define  change_pmd_prepare(vma, pmd, cp_flags)                         \
 361         ({                                                              \
 362                 long err = 0;                                           \
 363                 if (unlikely(pgtable_populate_needed(vma, cp_flags))) { \
 364                         if (pte_alloc(vma->vm_mm, pmd))                 \
 365                                 err = -ENOMEM;                          \
 366                 }                                                       \
 367                 err;                                                    \
 368         })
 369
 370 /*
 371  * This is the general pud/p4d/pgd version of change_pmd_prepare(). We need to
 372  * have separate change_pmd_prepare() because pte_alloc() returns 0 on success,
 373  * while {pmd|pud|p4d}_alloc() returns the valid pointer on success.
 374  */
 375 #define  change_prepare(vma, high, low, addr, cp_flags)                 \
 376           ({                                                            \
 377                 long err = 0;                                           \
 378                 if (unlikely(pgtable_populate_needed(vma, cp_flags))) { \
 379                         low##_t *p = low##_alloc(vma->vm_mm, high, addr); \
 380                         if (p == NULL)                                  \
 381                                 err = -ENOMEM;                          \
 382                 }                                                       \
 383                 err;                                                    \
 384         })
 385
 386 static inline long change_pmd_range(struct mmu_gather *tlb,
 387                 struct vm_area_struct *vma, pud_t *pud, unsigned long addr,
 388                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
 389 {
 390         pmd_t *pmd;
 391         unsigned long next;
 392         long pages = 0;
 393         unsigned long nr_huge_updates = 0;
 394         struct mmu_notifier_range range;
 395
 396         range.start = 0;
 397
 398         pmd = pmd_offset(pud, addr);
 399         do {
 400                 long ret;
 401
 402                 next = pmd_addr_end(addr, end);
 403
 404                 ret = change_pmd_prepare(vma, pmd, cp_flags);
 405                 if (ret) {
 406                         pages = ret;
 407                         break;
 408                 }
 409                 /*
 410                  * Automatic NUMA balancing walks the tables with mmap_lock
 411                  * held for read. It's possible a parallel update to occur
 412                  * between pmd_trans_huge() and a pmd_none_or_clear_bad()
 413                  * check leading to a false positive and clearing.
 414                  * Hence, it's necessary to atomically read the PMD value
 415                  * for all the checks.
 416                  */
 417                 if (!is_swap_pmd(*pmd) && !pmd_devmap(*pmd) &&
 418                      pmd_none_or_clear_bad_unless_trans_huge(pmd))
 419                         goto next;
 420
 421                 /* invoke the mmu notifier if the pmd is populated */
 422                 if (!range.start) {
 423                         mmu_notifier_range_init(&range,
 424                                 MMU_NOTIFY_PROTECTION_VMA, 0,
 425                                 vma->vm_mm, addr, end);
 426                         mmu_notifier_invalidate_range_start(&range);
 427                 }
 428
 429                 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
 430                         if ((next - addr != HPAGE_PMD_SIZE) ||
 431                             pgtable_split_needed(vma, cp_flags)) {
 432                                 __split_huge_pmd(vma, pmd, addr, false, NULL);
 433                                 /*
 434                                  * For file-backed, the pmd could have been
 435                                  * cleared; make sure pmd populated if
 436                                  * necessary, then fall-through to pte level.
 437                                  */
 438                                 ret = change_pmd_prepare(vma, pmd, cp_flags);
 439                                 if (ret) {
 440                                         pages = ret;
 441                                         break;
 442                                 }
 443                         } else {
 444                                 /*
 445                                  * change_huge_pmd() does not defer TLB flushes,
 446                                  * so no need to propagate the tlb argument.
 447                                  */
 448                                 int nr_ptes = change_huge_pmd(tlb, vma, pmd,
 449                                                 addr, newprot, cp_flags);
 450
 451                                 if (nr_ptes) {
 452                                         if (nr_ptes == HPAGE_PMD_NR) {
 453                                                 pages += HPAGE_PMD_NR;
 454                                                 nr_huge_updates++;
 455                                         }
 456
 457                                         /* huge pmd was handled */
 458                                         goto next;
 459                                 }
 460                         }
 461                         /* fall through, the trans huge pmd just split */
 462                 }
 463                 pages += change_pte_range(tlb, vma, pmd, addr, next,
 464                                           newprot, cp_flags);
 465 next:
 466                 cond_resched();
 467         } while (pmd++, addr = next, addr != end);
 468
 469         if (range.start)
 470                 mmu_notifier_invalidate_range_end(&range);
 471
 472         if (nr_huge_updates)
 473                 count_vm_numa_events(NUMA_HUGE_PTE_UPDATES, nr_huge_updates);
 474         return pages;
 475 }
 476
 477 static inline long change_pud_range(struct mmu_gather *tlb,
 478                 struct vm_area_struct *vma, p4d_t *p4d, unsigned long addr,
 479                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
 480 {
 481         pud_t *pud;
 482         unsigned long next;
 483         long pages = 0, ret;
 484
 485         pud = pud_offset(p4d, addr);
 486         do {
 487                 next = pud_addr_end(addr, end);
 488                 ret = change_prepare(vma, pud, pmd, addr, cp_flags);
 489                 if (ret)
 490                         return ret;
 491                 if (pud_none_or_clear_bad(pud))
 492                         continue;
 493                 pages += change_pmd_range(tlb, vma, pud, addr, next, newprot,
 494                                           cp_flags);
 495         } while (pud++, addr = next, addr != end);
 496
 497         return pages;
 498 }
 499
 500 static inline long change_p4d_range(struct mmu_gather *tlb,
 501                 struct vm_area_struct *vma, pgd_t *pgd, unsigned long addr,
 502                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
 503 {
 504         p4d_t *p4d;
 505         unsigned long next;
 506         long pages = 0, ret;
 507
 508         p4d = p4d_offset(pgd, addr);
 509         do {
 510                 next = p4d_addr_end(addr, end);
 511                 ret = change_prepare(vma, p4d, pud, addr, cp_flags);
 512                 if (ret)
 513                         return ret;
 514                 if (p4d_none_or_clear_bad(p4d))
 515                         continue;
 516                 pages += change_pud_range(tlb, vma, p4d, addr, next, newprot,
 517                                           cp_flags);
 518         } while (p4d++, addr = next, addr != end);
 519
 520         return pages;
 521 }
 522
 523 static long change_protection_range(struct mmu_gather *tlb,
 524                 struct vm_area_struct *vma, unsigned long addr,
 525                 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
 526 {
 527         struct mm_struct *mm = vma->vm_mm;
 528         pgd_t *pgd;
 529         unsigned long next;
 530         long pages = 0, ret;
 531
 532         BUG_ON(addr >= end);
 533         pgd = pgd_offset(mm, addr);
 534         tlb_start_vma(tlb, vma);
 535         do {
 536                 next = pgd_addr_end(addr, end);
 537                 ret = change_prepare(vma, pgd, p4d, addr, cp_flags);
 538                 if (ret) {
 539                         pages = ret;
 540                         break;
 541                 }
 542                 if (pgd_none_or_clear_bad(pgd))
 543                         continue;
 544                 pages += change_p4d_range(tlb, vma, pgd, addr, next, newprot,
 545                                           cp_flags);
 546         } while (pgd++, addr = next, addr != end);
 547
 548         tlb_end_vma(tlb, vma);
 549
 550         return pages;
 551 }
 552
 553 long change_protection(struct mmu_gather *tlb,
 554                        struct vm_area_struct *vma, unsigned long start,
 555                        unsigned long end, unsigned long cp_flags)
 556 {
 557         pgprot_t newprot = vma->vm_page_prot;
 558         long pages;
 559
 560         BUG_ON((cp_flags & MM_CP_UFFD_WP_ALL) == MM_CP_UFFD_WP_ALL);
 561
 562 #ifdef CONFIG_NUMA_BALANCING
 563         /*
 564          * Ordinary protection updates (mprotect, uffd-wp, softdirty tracking)
 565          * are expected to reflect their requirements via VMA flags such that
 566          * vma_set_page_prot() will adjust vma->vm_page_prot accordingly.
 567          */
 568         if (cp_flags & MM_CP_PROT_NUMA)
 569                 newprot = PAGE_NONE;
 570 #else
 571         WARN_ON_ONCE(cp_flags & MM_CP_PROT_NUMA);
 572 #endif
 573
 574         if (is_vm_hugetlb_page(vma))
 575                 pages = hugetlb_change_protection(vma, start, end, newprot,
 576                                                   cp_flags);
 577         else
 578                 pages = change_protection_range(tlb, vma, start, end, newprot,
 579                                                 cp_flags);
 580
 581         return pages;
 582 }
 583
 584 static int prot_none_pte_entry(pte_t *pte, unsigned long addr,
 585                                unsigned long next, struct mm_walk *walk)
 586 {
 587         return pfn_modify_allowed(pte_pfn(*pte), *(pgprot_t *)(walk->private)) ?
 588                 0 : -EACCES;
 589 }
 590
 591 static int prot_none_hugetlb_entry(pte_t *pte, unsigned long hmask,
 592                                    unsigned long addr, unsigned long next,
 593                                    struct mm_walk *walk)
 594 {
 595         return pfn_modify_allowed(pte_pfn(*pte), *(pgprot_t *)(walk->private)) ?
 596                 0 : -EACCES;
 597 }
 598
 599 static int prot_none_test(unsigned long addr, unsigned long next,
 600                           struct mm_walk *walk)
 601 {
 602         return 0;
 603 }
 604
 605 static const struct mm_walk_ops prot_none_walk_ops = {
 606         .pte_entry              = prot_none_pte_entry,
 607         .hugetlb_entry          = prot_none_hugetlb_entry,
 608         .test_walk              = prot_none_test,
 609 };
 610
 611 int
 612 mprotect_fixup(struct vma_iterator *vmi, struct mmu_gather *tlb,
 613                struct vm_area_struct *vma, struct vm_area_struct **pprev,
 614                unsigned long start, unsigned long end, unsigned long newflags)
 615 {
 616         struct mm_struct *mm = vma->vm_mm;
 617         unsigned long oldflags = vma->vm_flags;
 618         long nrpages = (end - start) >> PAGE_SHIFT;
 619         unsigned int mm_cp_flags = 0;
 620         unsigned long charged = 0;
 621         pgoff_t pgoff;
 622         int error;
 623
 624         if (newflags == oldflags) {
 625                 *pprev = vma;
 626                 return 0;
 627         }
 628
 629         /*
 630          * Do PROT_NONE PFN permission checks here when we can still
 631          * bail out without undoing a lot of state. This is a rather
 632          * uncommon case, so doesn't need to be very optimized.
 633          */
 634         if (arch_has_pfn_modify_check() &&
 635             (vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 636             (newflags & VM_ACCESS_FLAGS) == 0) {
 637                 pgprot_t new_pgprot = vm_get_page_prot(newflags);
 638
 639                 error = walk_page_range(current->mm, start, end,
 640                                 &prot_none_walk_ops, &new_pgprot);
 641                 if (error)
 642                         return error;
 643         }
 644
 645         /*
 646          * If we make a private mapping writable we increase our commit;
 647          * but (without finer accounting) cannot reduce our commit if we
 648          * make it unwritable again. hugetlb mapping were accounted for
 649          * even if read-only so there is no need to account for them here
 650          */
 651         if (newflags & VM_WRITE) {
 652                 /* Check space limits when area turns into data. */
 653                 if (!may_expand_vm(mm, newflags, nrpages) &&
 654                                 may_expand_vm(mm, oldflags, nrpages))
 655                         return -ENOMEM;
 656                 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
 657                                                 VM_SHARED|VM_NORESERVE))) {
 658                         charged = nrpages;
 659                         if (security_vm_enough_memory_mm(mm, charged))
 660                                 return -ENOMEM;
 661                         newflags |= VM_ACCOUNT;
 662                 }
 663         }
 664
 665         /*
 666          * First try to merge with previous and/or next vma.
 667          */
 668         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 669         *pprev = vma_merge(vmi, mm, *pprev, start, end, newflags,
 670                            vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
 671                            vma->vm_userfaultfd_ctx, anon_vma_name(vma));
 672         if (*pprev) {
 673                 vma = *pprev;
 674                 VM_WARN_ON((vma->vm_flags ^ newflags) & ~VM_SOFTDIRTY);
 675                 goto success;
 676         }
 677
 678         *pprev = vma;
 679
 680         if (start != vma->vm_start) {
 681                 error = split_vma(vmi, vma, start, 1);
 682                 if (error)
 683                         goto fail;
 684         }
 685
 686         if (end != vma->vm_end) {
 687                 error = split_vma(vmi, vma, end, 0);
 688                 if (error)
 689                         goto fail;
 690         }
 691
 692 success:
 693         /*
 694          * vm_flags and vm_page_prot are protected by the mmap_lock
 695          * held in write mode.
 696          */
 697         vm_flags_reset(vma, newflags);
 698         if (vma_wants_manual_pte_write_upgrade(vma))
 699                 mm_cp_flags |= MM_CP_TRY_CHANGE_WRITABLE;
 700         vma_set_page_prot(vma);
 701
 702         change_protection(tlb, vma, start, end, mm_cp_flags);
 703
 704         /*
 705          * Private VM_LOCKED VMA becoming writable: trigger COW to avoid major
 706          * fault on access.
 707          */
 708         if ((oldflags & (VM_WRITE | VM_SHARED | VM_LOCKED)) == VM_LOCKED &&
 709                         (newflags & VM_WRITE)) {
 710                 populate_vma_page_range(vma, start, end, NULL);
 711         }
 712
 713         vm_stat_account(mm, oldflags, -nrpages);
 714         vm_stat_account(mm, newflags, nrpages);
 715         perf_event_mmap(vma);
 716         return 0;
 717
 718 fail:
 719         vm_unacct_memory(charged);
 720         return error;
 721 }
 722
 723 /*
 724  * pkey==-1 when doing a legacy mprotect()
 725  */
 726 static int do_mprotect_pkey(unsigned long start, size_t len,
 727                 unsigned long prot, int pkey)
 728 {
 729         unsigned long nstart, end, tmp, reqprot;
 730         struct vm_area_struct *vma, *prev;
 731         int error;
 732         const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
 733         const bool rier = (current->personality & READ_IMPLIES_EXEC) &&
 734                                 (prot & PROT_READ);
 735         struct mmu_gather tlb;
 736         struct vma_iterator vmi;
 737
 738         start = untagged_addr(start);
 739
 740         prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
 741         if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
 742                 return -EINVAL;
 743
 744         if (start & ~PAGE_MASK)
 745                 return -EINVAL;
 746         if (!len)
 747                 return 0;
 748         len = PAGE_ALIGN(len);
 749         end = start + len;
 750         if (end <= start)
 751                 return -ENOMEM;
 752         if (!arch_validate_prot(prot, start))
 753                 return -EINVAL;
 754
 755         reqprot = prot;
 756
 757         if (mmap_write_lock_killable(current->mm))
 758                 return -EINTR;
 759
 760         /*
 761          * If userspace did not allocate the pkey, do not let
 762          * them use it here.
 763          */
 764         error = -EINVAL;
 765         if ((pkey != -1) && !mm_pkey_is_allocated(current->mm, pkey))
 766                 goto out;
 767
 768         vma_iter_init(&vmi, current->mm, start);
 769         vma = vma_find(&vmi, end);
 770         error = -ENOMEM;
 771         if (!vma)
 772                 goto out;
 773
 774         if (unlikely(grows & PROT_GROWSDOWN)) {
 775                 if (vma->vm_start >= end)
 776                         goto out;
 777                 start = vma->vm_start;
 778                 error = -EINVAL;
 779                 if (!(vma->vm_flags & VM_GROWSDOWN))
 780                         goto out;
 781         } else {
 782                 if (vma->vm_start > start)
 783                         goto out;
 784                 if (unlikely(grows & PROT_GROWSUP)) {
 785                         end = vma->vm_end;
 786                         error = -EINVAL;
 787                         if (!(vma->vm_flags & VM_GROWSUP))
 788                                 goto out;
 789                 }
 790         }
 791
 792         prev = vma_prev(&vmi);
 793         if (start > vma->vm_start)
 794                 prev = vma;
 795
 796         tlb_gather_mmu(&tlb, current->mm);
 797         nstart = start;
 798         tmp = vma->vm_start;
 799         for_each_vma_range(vmi, vma, end) {
 800                 unsigned long mask_off_old_flags;
 801                 unsigned long newflags;
 802                 int new_vma_pkey;
 803
 804                 if (vma->vm_start != tmp) {
 805                         error = -ENOMEM;
 806                         break;
 807                 }
 808
 809                 /* Does the application expect PROT_READ to imply PROT_EXEC */
 810                 if (rier && (vma->vm_flags & VM_MAYEXEC))
 811                         prot |= PROT_EXEC;
 812
 813                 /*
 814                  * Each mprotect() call explicitly passes r/w/x permissions.
 815                  * If a permission is not passed to mprotect(), it must be
 816                  * cleared from the VMA.
 817                  */
 818                 mask_off_old_flags = VM_ACCESS_FLAGS | VM_FLAGS_CLEAR;
 819
 820                 new_vma_pkey = arch_override_mprotect_pkey(vma, prot, pkey);
 821                 newflags = calc_vm_prot_bits(prot, new_vma_pkey);
 822                 newflags |= (vma->vm_flags & ~mask_off_old_flags);
 823
 824                 /* newflags >> 4 shift VM_MAY% in place of VM_% */
 825                 if ((newflags & ~(newflags >> 4)) & VM_ACCESS_FLAGS) {
 826                         error = -EACCES;
 827                         break;
 828                 }
 829
 830                 if (map_deny_write_exec(vma, newflags)) {
 831                         error = -EACCES;
 832                         break;
 833                 }
 834
 835                 /* Allow architectures to sanity-check the new flags */
 836                 if (!arch_validate_flags(newflags)) {
 837                         error = -EINVAL;
 838                         break;
 839                 }
 840
 841                 error = security_file_mprotect(vma, reqprot, prot);
 842                 if (error)
 843                         break;
 844
 845                 tmp = vma->vm_end;
 846                 if (tmp > end)
 847                         tmp = end;
 848
 849                 if (vma->vm_ops && vma->vm_ops->mprotect) {
 850                         error = vma->vm_ops->mprotect(vma, nstart, tmp, newflags);
 851                         if (error)
 852                                 break;
 853                 }
 854
 855                 error = mprotect_fixup(&vmi, &tlb, vma, &prev, nstart, tmp, newflags);
 856                 if (error)
 857                         break;
 858
 859                 tmp = vma_iter_end(&vmi);
 860                 nstart = tmp;
 861                 prot = reqprot;
 862         }
 863         tlb_finish_mmu(&tlb);
 864
 865         if (!error && vma_iter_end(&vmi) < end)
 866                 error = -ENOMEM;
 867
 868 out:
 869         mmap_write_unlock(current->mm);
 870         return error;
 871 }
 872
 873 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
 874                 unsigned long, prot)
 875 {
 876         return do_mprotect_pkey(start, len, prot, -1);
 877 }
 878
 879 #ifdef CONFIG_ARCH_HAS_PKEYS
 880
 881 SYSCALL_DEFINE4(pkey_mprotect, unsigned long, start, size_t, len,
 882                 unsigned long, prot, int, pkey)
 883 {
 884         return do_mprotect_pkey(start, len, prot, pkey);
 885 }
 886
 887 SYSCALL_DEFINE2(pkey_alloc, unsigned long, flags, unsigned long, init_val)
 888 {
 889         int pkey;
 890         int ret;
 891
 892         /* No flags supported yet. */
 893         if (flags)
 894                 return -EINVAL;
 895         /* check for unsupported init values */
 896         if (init_val & ~PKEY_ACCESS_MASK)
 897                 return -EINVAL;
 898
 899         mmap_write_lock(current->mm);
 900         pkey = mm_pkey_alloc(current->mm);
 901
 902         ret = -ENOSPC;
 903         if (pkey == -1)
 904                 goto out;
 905
 906         ret = arch_set_user_pkey_access(current, pkey, init_val);
 907         if (ret) {
 908                 mm_pkey_free(current->mm, pkey);
 909                 goto out;
 910         }
 911         ret = pkey;
 912 out:
 913         mmap_write_unlock(current->mm);
 914         return ret;
 915 }
 916
 917 SYSCALL_DEFINE1(pkey_free, int, pkey)
 918 {
 919         int ret;
 920
 921         mmap_write_lock(current->mm);
 922         ret = mm_pkey_free(current->mm, pkey);
 923         mmap_write_unlock(current->mm);
 924
 925         /*
 926          * We could provide warnings or errors if any VMA still
 927          * has the pkey set here.
 928          */
 929         return ret;
 930 }
 931
 932 #endif /* CONFIG_ARCH_HAS_PKEYS */