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 <linux/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 static unsigned long change_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
  37                 unsigned long addr, unsigned long end, pgprot_t newprot,
  38                 int dirty_accountable, int prot_numa)
  39 {
  40         struct mm_struct *mm = vma->vm_mm;
  41         pte_t *pte, oldpte;
  42         spinlock_t *ptl;
  43         unsigned long pages = 0;
  44         int target_node = NUMA_NO_NODE;
  45
  46         /*
  47          * Can be called with only the mmap_sem for reading by
  48          * prot_numa so we must check the pmd isn't constantly
  49          * changing from under us from pmd_none to pmd_trans_huge
  50          * and/or the other way around.
  51          */
  52         if (pmd_trans_unstable(pmd))
  53                 return 0;
  54
  55         /*
  56          * The pmd points to a regular pte so the pmd can't change
  57          * from under us even if the mmap_sem is only hold for
  58          * reading.
  59          */
  60         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
  61
  62         /* Get target node for single threaded private VMAs */
  63         if (prot_numa && !(vma->vm_flags & VM_SHARED) &&
  64             atomic_read(&vma->vm_mm->mm_users) == 1)
  65                 target_node = numa_node_id();
  66
  67         arch_enter_lazy_mmu_mode();
  68         do {
  69                 oldpte = *pte;
  70                 if (pte_present(oldpte)) {
  71                         pte_t ptent;
  72                         bool preserve_write = prot_numa && pte_write(oldpte);
  73
  74                         /*
  75                          * Avoid trapping faults against the zero or KSM
  76                          * pages. See similar comment in change_huge_pmd.
  77                          */
  78                         if (prot_numa) {
  79                                 struct page *page;
  80
  81                                 page = vm_normal_page(vma, addr, oldpte);
  82                                 if (!page || PageKsm(page))
  83                                         continue;
  84
  85                                 /* Avoid TLB flush if possible */
  86                                 if (pte_protnone(oldpte))
  87                                         continue;
  88
  89                                 /*
  90                                  * Don't mess with PTEs if page is already on the node
  91                                  * a single-threaded process is running on.
  92                                  */
  93                                 if (target_node == page_to_nid(page))
  94                                         continue;
  95                         }
  96
  97                         ptent = ptep_modify_prot_start(mm, addr, pte);
  98                         ptent = pte_modify(ptent, newprot);
  99                         if (preserve_write)
 100                                 ptent = pte_mk_savedwrite(ptent);
 101
 102                         /* Avoid taking write faults for known dirty pages */
 103                         if (dirty_accountable && pte_dirty(ptent) &&
 104                                         (pte_soft_dirty(ptent) ||
 105                                          !(vma->vm_flags & VM_SOFTDIRTY))) {
 106                                 ptent = pte_mkwrite(ptent);
 107                         }
 108                         ptep_modify_prot_commit(mm, addr, pte, ptent);
 109                         pages++;
 110                 } else if (IS_ENABLED(CONFIG_MIGRATION)) {
 111                         swp_entry_t entry = pte_to_swp_entry(oldpte);
 112
 113                         if (is_write_migration_entry(entry)) {
 114                                 pte_t newpte;
 115                                 /*
 116                                  * A protection check is difficult so
 117                                  * just be safe and disable write
 118                                  */
 119                                 make_migration_entry_read(&entry);
 120                                 newpte = swp_entry_to_pte(entry);
 121                                 if (pte_swp_soft_dirty(oldpte))
 122                                         newpte = pte_swp_mksoft_dirty(newpte);
 123                                 set_pte_at(mm, addr, pte, newpte);
 124
 125                                 pages++;
 126                         }
 127                 }
 128         } while (pte++, addr += PAGE_SIZE, addr != end);
 129         arch_leave_lazy_mmu_mode();
 130         pte_unmap_unlock(pte - 1, ptl);
 131
 132         return pages;
 133 }
 134
 135 static inline unsigned long change_pmd_range(struct vm_area_struct *vma,
 136                 pud_t *pud, unsigned long addr, unsigned long end,
 137                 pgprot_t newprot, int dirty_accountable, int prot_numa)
 138 {
 139         pmd_t *pmd;
 140         struct mm_struct *mm = vma->vm_mm;
 141         unsigned long next;
 142         unsigned long pages = 0;
 143         unsigned long nr_huge_updates = 0;
 144         unsigned long mni_start = 0;
 145
 146         pmd = pmd_offset(pud, addr);
 147         do {
 148                 unsigned long this_pages;
 149
 150                 next = pmd_addr_end(addr, end);
 151                 if (!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd)
 152                                 && pmd_none_or_clear_bad(pmd))
 153                         continue;
 154
 155                 /* invoke the mmu notifier if the pmd is populated */
 156                 if (!mni_start) {
 157                         mni_start = addr;
 158                         mmu_notifier_invalidate_range_start(mm, mni_start, end);
 159                 }
 160
 161                 if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
 162                         if (next - addr != HPAGE_PMD_SIZE) {
 163                                 __split_huge_pmd(vma, pmd, addr, false, NULL);
 164                         } else {
 165                                 int nr_ptes = change_huge_pmd(vma, pmd, addr,
 166                                                 newprot, prot_numa);
 167
 168                                 if (nr_ptes) {
 169                                         if (nr_ptes == HPAGE_PMD_NR) {
 170                                                 pages += HPAGE_PMD_NR;
 171                                                 nr_huge_updates++;
 172                                         }
 173
 174                                         /* huge pmd was handled */
 175                                         continue;
 176                                 }
 177                         }
 178                         /* fall through, the trans huge pmd just split */
 179                 }
 180                 this_pages = change_pte_range(vma, pmd, addr, next, newprot,
 181                                  dirty_accountable, prot_numa);
 182                 pages += this_pages;
 183         } while (pmd++, addr = next, addr != end);
 184
 185         if (mni_start)
 186                 mmu_notifier_invalidate_range_end(mm, mni_start, end);
 187
 188         if (nr_huge_updates)
 189                 count_vm_numa_events(NUMA_HUGE_PTE_UPDATES, nr_huge_updates);
 190         return pages;
 191 }
 192
 193 static inline unsigned long change_pud_range(struct vm_area_struct *vma,
 194                 p4d_t *p4d, unsigned long addr, unsigned long end,
 195                 pgprot_t newprot, int dirty_accountable, int prot_numa)
 196 {
 197         pud_t *pud;
 198         unsigned long next;
 199         unsigned long pages = 0;
 200
 201         pud = pud_offset(p4d, addr);
 202         do {
 203                 next = pud_addr_end(addr, end);
 204                 if (pud_none_or_clear_bad(pud))
 205                         continue;
 206                 pages += change_pmd_range(vma, pud, addr, next, newprot,
 207                                  dirty_accountable, prot_numa);
 208         } while (pud++, addr = next, addr != end);
 209
 210         return pages;
 211 }
 212
 213 static inline unsigned long change_p4d_range(struct vm_area_struct *vma,
 214                 pgd_t *pgd, unsigned long addr, unsigned long end,
 215                 pgprot_t newprot, int dirty_accountable, int prot_numa)
 216 {
 217         p4d_t *p4d;
 218         unsigned long next;
 219         unsigned long pages = 0;
 220
 221         p4d = p4d_offset(pgd, addr);
 222         do {
 223                 next = p4d_addr_end(addr, end);
 224                 if (p4d_none_or_clear_bad(p4d))
 225                         continue;
 226                 pages += change_pud_range(vma, p4d, addr, next, newprot,
 227                                  dirty_accountable, prot_numa);
 228         } while (p4d++, addr = next, addr != end);
 229
 230         return pages;
 231 }
 232
 233 static unsigned long change_protection_range(struct vm_area_struct *vma,
 234                 unsigned long addr, unsigned long end, pgprot_t newprot,
 235                 int dirty_accountable, int prot_numa)
 236 {
 237         struct mm_struct *mm = vma->vm_mm;
 238         pgd_t *pgd;
 239         unsigned long next;
 240         unsigned long start = addr;
 241         unsigned long pages = 0;
 242
 243         BUG_ON(addr >= end);
 244         pgd = pgd_offset(mm, addr);
 245         flush_cache_range(vma, addr, end);
 246         set_tlb_flush_pending(mm);
 247         do {
 248                 next = pgd_addr_end(addr, end);
 249                 if (pgd_none_or_clear_bad(pgd))
 250                         continue;
 251                 pages += change_p4d_range(vma, pgd, addr, next, newprot,
 252                                  dirty_accountable, prot_numa);
 253         } while (pgd++, addr = next, addr != end);
 254
 255         /* Only flush the TLB if we actually modified any entries: */
 256         if (pages)
 257                 flush_tlb_range(vma, start, end);
 258         clear_tlb_flush_pending(mm);
 259
 260         return pages;
 261 }
 262
 263 unsigned long change_protection(struct vm_area_struct *vma, unsigned long start,
 264                        unsigned long end, pgprot_t newprot,
 265                        int dirty_accountable, int prot_numa)
 266 {
 267         unsigned long pages;
 268
 269         if (is_vm_hugetlb_page(vma))
 270                 pages = hugetlb_change_protection(vma, start, end, newprot);
 271         else
 272                 pages = change_protection_range(vma, start, end, newprot, dirty_accountable, prot_numa);
 273
 274         return pages;
 275 }
 276
 277 int
 278 mprotect_fixup(struct vm_area_struct *vma, struct vm_area_struct **pprev,
 279         unsigned long start, unsigned long end, unsigned long newflags)
 280 {
 281         struct mm_struct *mm = vma->vm_mm;
 282         unsigned long oldflags = vma->vm_flags;
 283         long nrpages = (end - start) >> PAGE_SHIFT;
 284         unsigned long charged = 0;
 285         pgoff_t pgoff;
 286         int error;
 287         int dirty_accountable = 0;
 288
 289         if (newflags == oldflags) {
 290                 *pprev = vma;
 291                 return 0;
 292         }
 293
 294         /*
 295          * If we make a private mapping writable we increase our commit;
 296          * but (without finer accounting) cannot reduce our commit if we
 297          * make it unwritable again. hugetlb mapping were accounted for
 298          * even if read-only so there is no need to account for them here
 299          */
 300         if (newflags & VM_WRITE) {
 301                 /* Check space limits when area turns into data. */
 302                 if (!may_expand_vm(mm, newflags, nrpages) &&
 303                                 may_expand_vm(mm, oldflags, nrpages))
 304                         return -ENOMEM;
 305                 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
 306                                                 VM_SHARED|VM_NORESERVE))) {
 307                         charged = nrpages;
 308                         if (security_vm_enough_memory_mm(mm, charged))
 309                                 return -ENOMEM;
 310                         newflags |= VM_ACCOUNT;
 311                 }
 312         }
 313
 314         /*
 315          * First try to merge with previous and/or next vma.
 316          */
 317         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 318         *pprev = vma_merge(mm, *pprev, start, end, newflags,
 319                            vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
 320                            vma->vm_userfaultfd_ctx);
 321         if (*pprev) {
 322                 vma = *pprev;
 323                 VM_WARN_ON((vma->vm_flags ^ newflags) & ~VM_SOFTDIRTY);
 324                 goto success;
 325         }
 326
 327         *pprev = vma;
 328
 329         if (start != vma->vm_start) {
 330                 error = split_vma(mm, vma, start, 1);
 331                 if (error)
 332                         goto fail;
 333         }
 334
 335         if (end != vma->vm_end) {
 336                 error = split_vma(mm, vma, end, 0);
 337                 if (error)
 338                         goto fail;
 339         }
 340
 341 success:
 342         /*
 343          * vm_flags and vm_page_prot are protected by the mmap_sem
 344          * held in write mode.
 345          */
 346         vma->vm_flags = newflags;
 347         dirty_accountable = vma_wants_writenotify(vma, vma->vm_page_prot);
 348         vma_set_page_prot(vma);
 349
 350         change_protection(vma, start, end, vma->vm_page_prot,
 351                           dirty_accountable, 0);
 352
 353         /*
 354          * Private VM_LOCKED VMA becoming writable: trigger COW to avoid major
 355          * fault on access.
 356          */
 357         if ((oldflags & (VM_WRITE | VM_SHARED | VM_LOCKED)) == VM_LOCKED &&
 358                         (newflags & VM_WRITE)) {
 359                 populate_vma_page_range(vma, start, end, NULL);
 360         }
 361
 362         vm_stat_account(mm, oldflags, -nrpages);
 363         vm_stat_account(mm, newflags, nrpages);
 364         perf_event_mmap(vma);
 365         return 0;
 366
 367 fail:
 368         vm_unacct_memory(charged);
 369         return error;
 370 }
 371
 372 /*
 373  * pkey==-1 when doing a legacy mprotect()
 374  */
 375 static int do_mprotect_pkey(unsigned long start, size_t len,
 376                 unsigned long prot, int pkey)
 377 {
 378         unsigned long nstart, end, tmp, reqprot;
 379         struct vm_area_struct *vma, *prev;
 380         int error = -EINVAL;
 381         const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
 382         const bool rier = (current->personality & READ_IMPLIES_EXEC) &&
 383                                 (prot & PROT_READ);
 384
 385         prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
 386         if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
 387                 return -EINVAL;
 388
 389         if (start & ~PAGE_MASK)
 390                 return -EINVAL;
 391         if (!len)
 392                 return 0;
 393         len = PAGE_ALIGN(len);
 394         end = start + len;
 395         if (end <= start)
 396                 return -ENOMEM;
 397         if (!arch_validate_prot(prot))
 398                 return -EINVAL;
 399
 400         reqprot = prot;
 401
 402         if (down_write_killable(&current->mm->mmap_sem))
 403                 return -EINTR;
 404
 405         /*
 406          * If userspace did not allocate the pkey, do not let
 407          * them use it here.
 408          */
 409         error = -EINVAL;
 410         if ((pkey != -1) && !mm_pkey_is_allocated(current->mm, pkey))
 411                 goto out;
 412
 413         vma = find_vma(current->mm, start);
 414         error = -ENOMEM;
 415         if (!vma)
 416                 goto out;
 417         prev = vma->vm_prev;
 418         if (unlikely(grows & PROT_GROWSDOWN)) {
 419                 if (vma->vm_start >= end)
 420                         goto out;
 421                 start = vma->vm_start;
 422                 error = -EINVAL;
 423                 if (!(vma->vm_flags & VM_GROWSDOWN))
 424                         goto out;
 425         } else {
 426                 if (vma->vm_start > start)
 427                         goto out;
 428                 if (unlikely(grows & PROT_GROWSUP)) {
 429                         end = vma->vm_end;
 430                         error = -EINVAL;
 431                         if (!(vma->vm_flags & VM_GROWSUP))
 432                                 goto out;
 433                 }
 434         }
 435         if (start > vma->vm_start)
 436                 prev = vma;
 437
 438         for (nstart = start ; ; ) {
 439                 unsigned long mask_off_old_flags;
 440                 unsigned long newflags;
 441                 int new_vma_pkey;
 442
 443                 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
 444
 445                 /* Does the application expect PROT_READ to imply PROT_EXEC */
 446                 if (rier && (vma->vm_flags & VM_MAYEXEC))
 447                         prot |= PROT_EXEC;
 448
 449                 /*
 450                  * Each mprotect() call explicitly passes r/w/x permissions.
 451                  * If a permission is not passed to mprotect(), it must be
 452                  * cleared from the VMA.
 453                  */
 454                 mask_off_old_flags = VM_READ | VM_WRITE | VM_EXEC |
 455                                         ARCH_VM_PKEY_FLAGS;
 456
 457                 new_vma_pkey = arch_override_mprotect_pkey(vma, prot, pkey);
 458                 newflags = calc_vm_prot_bits(prot, new_vma_pkey);
 459                 newflags |= (vma->vm_flags & ~mask_off_old_flags);
 460
 461                 /* newflags >> 4 shift VM_MAY% in place of VM_% */
 462                 if ((newflags & ~(newflags >> 4)) & (VM_READ | VM_WRITE | VM_EXEC)) {
 463                         error = -EACCES;
 464                         goto out;
 465                 }
 466
 467                 error = security_file_mprotect(vma, reqprot, prot);
 468                 if (error)
 469                         goto out;
 470
 471                 tmp = vma->vm_end;
 472                 if (tmp > end)
 473                         tmp = end;
 474                 error = mprotect_fixup(vma, &prev, nstart, tmp, newflags);
 475                 if (error)
 476                         goto out;
 477                 nstart = tmp;
 478
 479                 if (nstart < prev->vm_end)
 480                         nstart = prev->vm_end;
 481                 if (nstart >= end)
 482                         goto out;
 483
 484                 vma = prev->vm_next;
 485                 if (!vma || vma->vm_start != nstart) {
 486                         error = -ENOMEM;
 487                         goto out;
 488                 }
 489                 prot = reqprot;
 490         }
 491 out:
 492         up_write(&current->mm->mmap_sem);
 493         return error;
 494 }
 495
 496 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
 497                 unsigned long, prot)
 498 {
 499         return do_mprotect_pkey(start, len, prot, -1);
 500 }
 501
 502 #ifdef CONFIG_ARCH_HAS_PKEYS
 503
 504 SYSCALL_DEFINE4(pkey_mprotect, unsigned long, start, size_t, len,
 505                 unsigned long, prot, int, pkey)
 506 {
 507         return do_mprotect_pkey(start, len, prot, pkey);
 508 }
 509
 510 SYSCALL_DEFINE2(pkey_alloc, unsigned long, flags, unsigned long, init_val)
 511 {
 512         int pkey;
 513         int ret;
 514
 515         /* No flags supported yet. */
 516         if (flags)
 517                 return -EINVAL;
 518         /* check for unsupported init values */
 519         if (init_val & ~PKEY_ACCESS_MASK)
 520                 return -EINVAL;
 521
 522         down_write(&current->mm->mmap_sem);
 523         pkey = mm_pkey_alloc(current->mm);
 524
 525         ret = -ENOSPC;
 526         if (pkey == -1)
 527                 goto out;
 528
 529         ret = arch_set_user_pkey_access(current, pkey, init_val);
 530         if (ret) {
 531                 mm_pkey_free(current->mm, pkey);
 532                 goto out;
 533         }
 534         ret = pkey;
 535 out:
 536         up_write(&current->mm->mmap_sem);
 537         return ret;
 538 }
 539
 540 SYSCALL_DEFINE1(pkey_free, int, pkey)
 541 {
 542         int ret;
 543
 544         down_write(&current->mm->mmap_sem);
 545         ret = mm_pkey_free(current->mm, pkey);
 546         up_write(&current->mm->mmap_sem);
 547
 548         /*
 549          * We could provie warnings or errors if any VMA still
 550          * has the pkey set here.
 551          */
 552         return ret;
 553 }
 554
 555 #endif /* CONFIG_ARCH_HAS_PKEYS */