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