mm/mprotect.c

   1 /*
   2  *  mm/mprotect.c
   3  *
   4  *  (C) Copyright 1994 Linus Torvalds
   5  *  (C) Copyright 2002 Christoph Hellwig
   6  *
   7  *  Address space accounting code       <alan@lxorguk.ukuu.org.uk>
   8  *  (C) Copyright 2002 Red Hat Inc, All Rights Reserved
   9  */
  10
  11 #include <linux/mm.h>
  12 #include <linux/hugetlb.h>
  13 #include <linux/shm.h>
  14 #include <linux/mman.h>
  15 #include <linux/fs.h>
  16 #include <linux/highmem.h>
  17 #include <linux/security.h>
  18 #include <linux/mempolicy.h>
  19 #include <linux/personality.h>
  20 #include <linux/syscalls.h>
  21 #include <linux/swap.h>
  22 #include <linux/swapops.h>
  23 #include <linux/mmu_notifier.h>
  24 #include <linux/migrate.h>
  25 #include <linux/perf_event.h>
  26 #include <linux/pkeys.h>
  27 #include <linux/ksm.h>
  28 #include <asm/uaccess.h>
  29 #include <asm/pgtable.h>
  30 #include <asm/cacheflush.h>
  31 #include <asm/mmu_context.h>
  32 #include <asm/tlbflush.h>
  33
  34 #include "internal.h"
  35
  36 /*
  37  * For a prot_numa update we only hold mmap_sem for read so there is a
  38  * potential race with faulting where a pmd was temporarily none. This
  39  * function checks for a transhuge pmd under the appropriate lock. It
  40  * returns a pte if it was successfully locked or NULL if it raced with
  41  * a transhuge insertion.
  42  */
  43 static pte_t *lock_pte_protection(struct vm_area_struct *vma, pmd_t *pmd,
  44                         unsigned long addr, int prot_numa, spinlock_t **ptl)
  45 {
  46         pte_t *pte;
  47         spinlock_t *pmdl;
  48
  49         /* !prot_numa is protected by mmap_sem held for write */
  50         if (!prot_numa)
  51                 return pte_offset_map_lock(vma->vm_mm, pmd, addr, ptl);
  52
  53         pmdl = pmd_lock(vma->vm_mm, pmd);
  54         if (unlikely(pmd_trans_huge(*pmd) || pmd_none(*pmd))) {
  55                 spin_unlock(pmdl);
  56                 return NULL;
  57         }
  58
  59         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, ptl);
  60         spin_unlock(pmdl);
  61         return pte;
  62 }
  63
  64 static unsigned long change_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
  65                 unsigned long addr, unsigned long end, pgprot_t newprot,
  66                 int dirty_accountable, int prot_numa)
  67 {
  68         struct mm_struct *mm = vma->vm_mm;
  69         pte_t *pte, oldpte;
  70         spinlock_t *ptl;
  71         unsigned long pages = 0;
  72         int target_node = NUMA_NO_NODE;
  73
  74         pte = lock_pte_protection(vma, pmd, addr, prot_numa, &ptl);
  75         if (!pte)
  76                 return 0;
  77
  78         /* Get target node for single threaded private VMAs */
  79         if (prot_numa && !(vma->vm_flags & VM_SHARED) &&
  80             atomic_read(&vma->vm_mm->mm_users) == 1)
  81                 target_node = numa_node_id();
  82
  83         arch_enter_lazy_mmu_mode();
  84         do {
  85                 oldpte = *pte;
  86                 if (pte_present(oldpte)) {
  87                         pte_t ptent;
  88                         bool preserve_write = prot_numa && pte_write(oldpte);
  89
  90                         /*
  91                          * Avoid trapping faults against the zero or KSM
  92                          * pages. See similar comment in change_huge_pmd.
  93                          */
  94                         if (prot_numa) {
  95                                 struct page *page;
  96
  97                                 page = vm_normal_page(vma, addr, oldpte);
  98                                 if (!page || PageKsm(page))
  99                                         continue;
 100
 101                                 /* Avoid TLB flush if possible */
 102                                 if (pte_protnone(oldpte))
 103                                         continue;
 104
 105                                 /*
 106                                  * Don't mess with PTEs if page is already on the node
 107                                  * a single-threaded process is running on.
 108                                  */
 109                                 if (target_node == page_to_nid(page))
 110                                         continue;
 111                         }
 112
 113                         ptent = ptep_modify_prot_start(mm, addr, pte);
 114                         ptent = pte_modify(ptent, newprot);
 115                         if (preserve_write)
 116                                 ptent = pte_mkwrite(ptent);
 117
 118                         /* Avoid taking write faults for known dirty pages */
 119                         if (dirty_accountable && pte_dirty(ptent) &&
 120                                         (pte_soft_dirty(ptent) ||
 121                                          !(vma->vm_flags & VM_SOFTDIRTY))) {
 122                                 ptent = pte_mkwrite(ptent);
 123                         }
 124                         ptep_modify_prot_commit(mm, addr, pte, ptent);
 125                         pages++;
 126                 } else if (IS_ENABLED(CONFIG_MIGRATION)) {
 127                         swp_entry_t entry = pte_to_swp_entry(oldpte);
 128
 129                         if (is_write_migration_entry(entry)) {
 130                                 pte_t newpte;
 131                                 /*
 132                                  * A protection check is difficult so
 133                                  * just be safe and disable write
 134                                  */
 135                                 make_migration_entry_read(&entry);
 136                                 newpte = swp_entry_to_pte(entry);
 137                                 if (pte_swp_soft_dirty(oldpte))
 138                                         newpte = pte_swp_mksoft_dirty(newpte);
 139                                 set_pte_at(mm, addr, pte, newpte);
 140
 141                                 pages++;
 142                         }
 143                 }
 144         } while (pte++, addr += PAGE_SIZE, addr != end);
 145         arch_leave_lazy_mmu_mode();
 146         pte_unmap_unlock(pte - 1, ptl);
 147
 148         return pages;
 149 }
 150
 151 static inline unsigned long change_pmd_range(struct vm_area_struct *vma,
 152                 pud_t *pud, unsigned long addr, unsigned long end,
 153                 pgprot_t newprot, int dirty_accountable, int prot_numa)
 154 {
 155         pmd_t *pmd;
 156         struct mm_struct *mm = vma->vm_mm;
 157         unsigned long next;
 158         unsigned long pages = 0;
 159         unsigned long nr_huge_updates = 0;
 160         unsigned long mni_start = 0;
 161
 162         pmd = pmd_offset(pud, addr);
 163         do {
 164                 unsigned long this_pages;
 165
 166                 next = pmd_addr_end(addr, end);
 167                 if (!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd)
 168                                 && pmd_none_or_clear_bad(pmd))
 169                         continue;
 170
 171                 /* invoke the mmu notifier if the pmd is populated */
 172                 if (!mni_start) {
 173                         mni_start = addr;
 174                         mmu_notifier_invalidate_range_start(mm, mni_start, end);
 175                 }
 176
 177                 if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
 178                         if (next - addr != HPAGE_PMD_SIZE) {
 179                                 __split_huge_pmd(vma, pmd, addr, false, NULL);
 180                                 if (pmd_trans_unstable(pmd))
 181                                         continue;
 182                         } else {
 183                                 int nr_ptes = change_huge_pmd(vma, pmd, addr,
 184                                                 newprot, prot_numa);
 185
 186                                 if (nr_ptes) {
 187                                         if (nr_ptes == HPAGE_PMD_NR) {
 188                                                 pages += HPAGE_PMD_NR;
 189                                                 nr_huge_updates++;
 190                                         }
 191
 192                                         /* huge pmd was handled */
 193                                         continue;
 194                                 }
 195                         }
 196                         /* fall through, the trans huge pmd just split */
 197                 }
 198                 this_pages = change_pte_range(vma, pmd, addr, next, newprot,
 199                                  dirty_accountable, prot_numa);
 200                 pages += this_pages;
 201         } while (pmd++, addr = next, addr != end);
 202
 203         if (mni_start)
 204                 mmu_notifier_invalidate_range_end(mm, mni_start, end);
 205
 206         if (nr_huge_updates)
 207                 count_vm_numa_events(NUMA_HUGE_PTE_UPDATES, nr_huge_updates);
 208         return pages;
 209 }
 210
 211 static inline unsigned long change_pud_range(struct vm_area_struct *vma,
 212                 pgd_t *pgd, unsigned long addr, unsigned long end,
 213                 pgprot_t newprot, int dirty_accountable, int prot_numa)
 214 {
 215         pud_t *pud;
 216         unsigned long next;
 217         unsigned long pages = 0;
 218
 219         pud = pud_offset(pgd, addr);
 220         do {
 221                 next = pud_addr_end(addr, end);
 222                 if (pud_none_or_clear_bad(pud))
 223                         continue;
 224                 pages += change_pmd_range(vma, pud, addr, next, newprot,
 225                                  dirty_accountable, prot_numa);
 226         } while (pud++, addr = next, addr != end);
 227
 228         return pages;
 229 }
 230
 231 static unsigned long change_protection_range(struct vm_area_struct *vma,
 232                 unsigned long addr, unsigned long end, pgprot_t newprot,
 233                 int dirty_accountable, int prot_numa)
 234 {
 235         struct mm_struct *mm = vma->vm_mm;
 236         pgd_t *pgd;
 237         unsigned long next;
 238         unsigned long start = addr;
 239         unsigned long pages = 0;
 240
 241         BUG_ON(addr >= end);
 242         pgd = pgd_offset(mm, addr);
 243         flush_cache_range(vma, addr, end);
 244         set_tlb_flush_pending(mm);
 245         do {
 246                 next = pgd_addr_end(addr, end);
 247                 if (pgd_none_or_clear_bad(pgd))
 248                         continue;
 249                 pages += change_pud_range(vma, pgd, addr, next, newprot,
 250                                  dirty_accountable, prot_numa);
 251         } while (pgd++, addr = next, addr != end);
 252
 253         /* Only flush the TLB if we actually modified any entries: */
 254         if (pages)
 255                 flush_tlb_range(vma, start, end);
 256         clear_tlb_flush_pending(mm);
 257
 258         return pages;
 259 }
 260
 261 unsigned long change_protection(struct vm_area_struct *vma, unsigned long start,
 262                        unsigned long end, pgprot_t newprot,
 263                        int dirty_accountable, int prot_numa)
 264 {
 265         unsigned long pages;
 266
 267         if (is_vm_hugetlb_page(vma))
 268                 pages = hugetlb_change_protection(vma, start, end, newprot);
 269         else
 270                 pages = change_protection_range(vma, start, end, newprot, dirty_accountable, prot_numa);
 271
 272         return pages;
 273 }
 274
 275 int
 276 mprotect_fixup(struct vm_area_struct *vma, struct vm_area_struct **pprev,
 277         unsigned long start, unsigned long end, unsigned long newflags)
 278 {
 279         struct mm_struct *mm = vma->vm_mm;
 280         unsigned long oldflags = vma->vm_flags;
 281         long nrpages = (end - start) >> PAGE_SHIFT;
 282         unsigned long charged = 0;
 283         pgoff_t pgoff;
 284         int error;
 285         int dirty_accountable = 0;
 286
 287         if (newflags == oldflags) {
 288                 *pprev = vma;
 289                 return 0;
 290         }
 291
 292         /*
 293          * If we make a private mapping writable we increase our commit;
 294          * but (without finer accounting) cannot reduce our commit if we
 295          * make it unwritable again. hugetlb mapping were accounted for
 296          * even if read-only so there is no need to account for them here
 297          */
 298         if (newflags & VM_WRITE) {
 299                 /* Check space limits when area turns into data. */
 300                 if (!may_expand_vm(mm, newflags, nrpages) &&
 301                                 may_expand_vm(mm, oldflags, nrpages))
 302                         return -ENOMEM;
 303                 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
 304                                                 VM_SHARED|VM_NORESERVE))) {
 305                         charged = nrpages;
 306                         if (security_vm_enough_memory_mm(mm, charged))
 307                                 return -ENOMEM;
 308                         newflags |= VM_ACCOUNT;
 309                 }
 310         }
 311
 312         /*
 313          * First try to merge with previous and/or next vma.
 314          */
 315         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 316         *pprev = vma_merge(mm, *pprev, start, end, newflags,
 317                            vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
 318                            vma->vm_userfaultfd_ctx);
 319         if (*pprev) {
 320                 vma = *pprev;
 321                 VM_WARN_ON((vma->vm_flags ^ newflags) & ~VM_SOFTDIRTY);
 322                 goto success;
 323         }
 324
 325         *pprev = vma;
 326
 327         if (start != vma->vm_start) {
 328                 error = split_vma(mm, vma, start, 1);
 329                 if (error)
 330                         goto fail;
 331         }
 332
 333         if (end != vma->vm_end) {
 334                 error = split_vma(mm, vma, end, 0);
 335                 if (error)
 336                         goto fail;
 337         }
 338
 339 success:
 340         /*
 341          * vm_flags and vm_page_prot are protected by the mmap_sem
 342          * held in write mode.
 343          */
 344         vma->vm_flags = newflags;
 345         dirty_accountable = vma_wants_writenotify(vma, vma->vm_page_prot);
 346         vma_set_page_prot(vma);
 347
 348         change_protection(vma, start, end, vma->vm_page_prot,
 349                           dirty_accountable, 0);
 350
 351         /*
 352          * Private VM_LOCKED VMA becoming writable: trigger COW to avoid major
 353          * fault on access.
 354          */
 355         if ((oldflags & (VM_WRITE | VM_SHARED | VM_LOCKED)) == VM_LOCKED &&
 356                         (newflags & VM_WRITE)) {
 357                 populate_vma_page_range(vma, start, end, NULL);
 358         }
 359
 360         vm_stat_account(mm, oldflags, -nrpages);
 361         vm_stat_account(mm, newflags, nrpages);
 362         perf_event_mmap(vma);
 363         return 0;
 364
 365 fail:
 366         vm_unacct_memory(charged);
 367         return error;
 368 }
 369
 370 /*
 371  * pkey==-1 when doing a legacy mprotect()
 372  */
 373 static int do_mprotect_pkey(unsigned long start, size_t len,
 374                 unsigned long prot, int pkey)
 375 {
 376         unsigned long nstart, end, tmp, reqprot;
 377         struct vm_area_struct *vma, *prev;
 378         int error = -EINVAL;
 379         const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
 380         const bool rier = (current->personality & READ_IMPLIES_EXEC) &&
 381                                 (prot & PROT_READ);
 382
 383         prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
 384         if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
 385                 return -EINVAL;
 386
 387         if (start & ~PAGE_MASK)
 388                 return -EINVAL;
 389         if (!len)
 390                 return 0;
 391         len = PAGE_ALIGN(len);
 392         end = start + len;
 393         if (end <= start)
 394                 return -ENOMEM;
 395         if (!arch_validate_prot(prot))
 396                 return -EINVAL;
 397
 398         reqprot = prot;
 399
 400         if (down_write_killable(&current->mm->mmap_sem))
 401                 return -EINTR;
 402
 403         /*
 404          * If userspace did not allocate the pkey, do not let
 405          * them use it here.
 406          */
 407         error = -EINVAL;
 408         if ((pkey != -1) && !mm_pkey_is_allocated(current->mm, pkey))
 409                 goto out;
 410
 411         vma = find_vma(current->mm, start);
 412         error = -ENOMEM;
 413         if (!vma)
 414                 goto out;
 415         prev = vma->vm_prev;
 416         if (unlikely(grows & PROT_GROWSDOWN)) {
 417                 if (vma->vm_start >= end)
 418                         goto out;
 419                 start = vma->vm_start;
 420                 error = -EINVAL;
 421                 if (!(vma->vm_flags & VM_GROWSDOWN))
 422                         goto out;
 423         } else {
 424                 if (vma->vm_start > start)
 425                         goto out;
 426                 if (unlikely(grows & PROT_GROWSUP)) {
 427                         end = vma->vm_end;
 428                         error = -EINVAL;
 429                         if (!(vma->vm_flags & VM_GROWSUP))
 430                                 goto out;
 431                 }
 432         }
 433         if (start > vma->vm_start)
 434                 prev = vma;
 435
 436         for (nstart = start ; ; ) {
 437                 unsigned long mask_off_old_flags;
 438                 unsigned long newflags;
 439                 int new_vma_pkey;
 440
 441                 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
 442
 443                 /* Does the application expect PROT_READ to imply PROT_EXEC */
 444                 if (rier && (vma->vm_flags & VM_MAYEXEC))
 445                         prot |= PROT_EXEC;
 446
 447                 /*
 448                  * Each mprotect() call explicitly passes r/w/x permissions.
 449                  * If a permission is not passed to mprotect(), it must be
 450                  * cleared from the VMA.
 451                  */
 452                 mask_off_old_flags = VM_READ | VM_WRITE | VM_EXEC |
 453                                         ARCH_VM_PKEY_FLAGS;
 454
 455                 new_vma_pkey = arch_override_mprotect_pkey(vma, prot, pkey);
 456                 newflags = calc_vm_prot_bits(prot, new_vma_pkey);
 457                 newflags |= (vma->vm_flags & ~mask_off_old_flags);
 458
 459                 /* newflags >> 4 shift VM_MAY% in place of VM_% */
 460                 if ((newflags & ~(newflags >> 4)) & (VM_READ | VM_WRITE | VM_EXEC)) {
 461                         error = -EACCES;
 462                         goto out;
 463                 }
 464
 465                 error = security_file_mprotect(vma, reqprot, prot);
 466                 if (error)
 467                         goto out;
 468
 469                 tmp = vma->vm_end;
 470                 if (tmp > end)
 471                         tmp = end;
 472                 error = mprotect_fixup(vma, &prev, nstart, tmp, newflags);
 473                 if (error)
 474                         goto out;
 475                 nstart = tmp;
 476
 477                 if (nstart < prev->vm_end)
 478                         nstart = prev->vm_end;
 479                 if (nstart >= end)
 480                         goto out;
 481
 482                 vma = prev->vm_next;
 483                 if (!vma || vma->vm_start != nstart) {
 484                         error = -ENOMEM;
 485                         goto out;
 486                 }
 487                 prot = reqprot;
 488         }
 489 out:
 490         up_write(&current->mm->mmap_sem);
 491         return error;
 492 }
 493
 494 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
 495                 unsigned long, prot)
 496 {
 497         return do_mprotect_pkey(start, len, prot, -1);
 498 }
 499
 500 #ifdef CONFIG_ARCH_HAS_PKEYS
 501
 502 SYSCALL_DEFINE4(pkey_mprotect, unsigned long, start, size_t, len,
 503                 unsigned long, prot, int, pkey)
 504 {
 505         return do_mprotect_pkey(start, len, prot, pkey);
 506 }
 507
 508 SYSCALL_DEFINE2(pkey_alloc, unsigned long, flags, unsigned long, init_val)
 509 {
 510         int pkey;
 511         int ret;
 512
 513         /* No flags supported yet. */
 514         if (flags)
 515                 return -EINVAL;
 516         /* check for unsupported init values */
 517         if (init_val & ~PKEY_ACCESS_MASK)
 518                 return -EINVAL;
 519
 520         down_write(&current->mm->mmap_sem);
 521         pkey = mm_pkey_alloc(current->mm);
 522
 523         ret = -ENOSPC;
 524         if (pkey == -1)
 525                 goto out;
 526
 527         ret = arch_set_user_pkey_access(current, pkey, init_val);
 528         if (ret) {
 529                 mm_pkey_free(current->mm, pkey);
 530                 goto out;
 531         }
 532         ret = pkey;
 533 out:
 534         up_write(&current->mm->mmap_sem);
 535         return ret;
 536 }
 537
 538 SYSCALL_DEFINE1(pkey_free, int, pkey)
 539 {
 540         int ret;
 541
 542         down_write(&current->mm->mmap_sem);
 543         ret = mm_pkey_free(current->mm, pkey);
 544         up_write(&current->mm->mmap_sem);
 545
 546         /*
 547          * We could provie warnings or errors if any VMA still
 548          * has the pkey set here.
 549          */
 550         return ret;
 551 }
 552
 553 #endif /* CONFIG_ARCH_HAS_PKEYS */