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/ksm.h>
  27 #include <asm/uaccess.h>
  28 #include <asm/pgtable.h>
  29 #include <asm/cacheflush.h>
  30 #include <asm/tlbflush.h>
  31
  32 #ifndef pgprot_modify
  33 static inline pgprot_t pgprot_modify(pgprot_t oldprot, pgprot_t newprot)
  34 {
  35         return newprot;
  36 }
  37 #endif
  38
  39 /*
  40  * For a prot_numa update we only hold mmap_sem for read so there is a
  41  * potential race with faulting where a pmd was temporarily none. This
  42  * function checks for a transhuge pmd under the appropriate lock. It
  43  * returns a pte if it was successfully locked or NULL if it raced with
  44  * a transhuge insertion.
  45  */
  46 static pte_t *lock_pte_protection(struct vm_area_struct *vma, pmd_t *pmd,
  47                         unsigned long addr, int prot_numa, spinlock_t **ptl)
  48 {
  49         pte_t *pte;
  50         spinlock_t *pmdl;
  51
  52         /* !prot_numa is protected by mmap_sem held for write */
  53         if (!prot_numa)
  54                 return pte_offset_map_lock(vma->vm_mm, pmd, addr, ptl);
  55
  56         pmdl = pmd_lock(vma->vm_mm, pmd);
  57         if (unlikely(pmd_trans_huge(*pmd) || pmd_none(*pmd))) {
  58                 spin_unlock(pmdl);
  59                 return NULL;
  60         }
  61
  62         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, ptl);
  63         spin_unlock(pmdl);
  64         return pte;
  65 }
  66
  67 static unsigned long change_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
  68                 unsigned long addr, unsigned long end, pgprot_t newprot,
  69                 int dirty_accountable, int prot_numa)
  70 {
  71         struct mm_struct *mm = vma->vm_mm;
  72         pte_t *pte, oldpte;
  73         spinlock_t *ptl;
  74         unsigned long pages = 0;
  75
  76         pte = lock_pte_protection(vma, pmd, addr, prot_numa, &ptl);
  77         if (!pte)
  78                 return 0;
  79
  80         arch_enter_lazy_mmu_mode();
  81         do {
  82                 oldpte = *pte;
  83                 if (pte_present(oldpte)) {
  84                         pte_t ptent;
  85                         bool updated = false;
  86
  87                         if (!prot_numa) {
  88                                 ptent = ptep_modify_prot_start(mm, addr, pte);
  89                                 if (pte_numa(ptent))
  90                                         ptent = pte_mknonnuma(ptent);
  91                                 ptent = pte_modify(ptent, newprot);
  92                                 /*
  93                                  * Avoid taking write faults for pages we
  94                                  * know to be dirty.
  95                                  */
  96                                 if (dirty_accountable && pte_dirty(ptent))
  97                                         ptent = pte_mkwrite(ptent);
  98                                 ptep_modify_prot_commit(mm, addr, pte, ptent);
  99                                 updated = true;
 100                         } else {
 101                                 struct page *page;
 102
 103                                 page = vm_normal_page(vma, addr, oldpte);
 104                                 if (page && !PageKsm(page)) {
 105                                         if (!pte_numa(oldpte)) {
 106                                                 ptep_set_numa(mm, addr, pte);
 107                                                 updated = true;
 108                                         }
 109                                 }
 110                         }
 111                         if (updated)
 112                                 pages++;
 113                 } else if (IS_ENABLED(CONFIG_MIGRATION)) {
 114                         swp_entry_t entry = pte_to_swp_entry(oldpte);
 115
 116                         if (is_write_migration_entry(entry)) {
 117                                 pte_t newpte;
 118                                 /*
 119                                  * A protection check is difficult so
 120                                  * just be safe and disable write
 121                                  */
 122                                 make_migration_entry_read(&entry);
 123                                 newpte = swp_entry_to_pte(entry);
 124                                 if (pte_swp_soft_dirty(oldpte))
 125                                         newpte = pte_swp_mksoft_dirty(newpte);
 126                                 set_pte_at(mm, addr, pte, newpte);
 127
 128                                 pages++;
 129                         }
 130                 }
 131         } while (pte++, addr += PAGE_SIZE, addr != end);
 132         arch_leave_lazy_mmu_mode();
 133         pte_unmap_unlock(pte - 1, ptl);
 134
 135         return pages;
 136 }
 137
 138 static inline unsigned long change_pmd_range(struct vm_area_struct *vma,
 139                 pud_t *pud, unsigned long addr, unsigned long end,
 140                 pgprot_t newprot, int dirty_accountable, int prot_numa)
 141 {
 142         pmd_t *pmd;
 143         struct mm_struct *mm = vma->vm_mm;
 144         unsigned long next;
 145         unsigned long pages = 0;
 146         unsigned long nr_huge_updates = 0;
 147         unsigned long mni_start = 0;
 148
 149         pmd = pmd_offset(pud, addr);
 150         do {
 151                 unsigned long this_pages;
 152
 153                 next = pmd_addr_end(addr, end);
 154                 if (!pmd_trans_huge(*pmd) && pmd_none_or_clear_bad(pmd))
 155                         goto next;
 156
 157                 /* invoke the mmu notifier if the pmd is populated */
 158                 if (!mni_start) {
 159                         mni_start = addr;
 160                         mmu_notifier_invalidate_range_start(mm, mni_start, end);
 161                 }
 162
 163                 if (pmd_trans_huge(*pmd)) {
 164                         if (next - addr != HPAGE_PMD_SIZE)
 165                                 split_huge_page_pmd(vma, addr, pmd);
 166                         else {
 167                                 int nr_ptes = change_huge_pmd(vma, pmd, addr,
 168                                                 newprot, prot_numa);
 169
 170                                 if (nr_ptes) {
 171                                         if (nr_ptes == HPAGE_PMD_NR) {
 172                                                 pages += HPAGE_PMD_NR;
 173                                                 nr_huge_updates++;
 174                                         }
 175
 176                                         /* huge pmd was handled */
 177                                         goto next;
 178                                 }
 179                         }
 180                         /* fall through, the trans huge pmd just split */
 181                 }
 182                 this_pages = change_pte_range(vma, pmd, addr, next, newprot,
 183                                  dirty_accountable, prot_numa);
 184                 pages += this_pages;
 185 next:
 186                 cond_resched();
 187         } while (pmd++, addr = next, addr != end);
 188
 189         if (mni_start)
 190                 mmu_notifier_invalidate_range_end(mm, mni_start, end);
 191
 192         if (nr_huge_updates)
 193                 count_vm_numa_events(NUMA_HUGE_PTE_UPDATES, nr_huge_updates);
 194         return pages;
 195 }
 196
 197 static inline unsigned long change_pud_range(struct vm_area_struct *vma,
 198                 pgd_t *pgd, unsigned long addr, unsigned long end,
 199                 pgprot_t newprot, int dirty_accountable, int prot_numa)
 200 {
 201         pud_t *pud;
 202         unsigned long next;
 203         unsigned long pages = 0;
 204
 205         pud = pud_offset(pgd, addr);
 206         do {
 207                 next = pud_addr_end(addr, end);
 208                 if (pud_none_or_clear_bad(pud))
 209                         continue;
 210                 pages += change_pmd_range(vma, pud, addr, next, newprot,
 211                                  dirty_accountable, prot_numa);
 212         } while (pud++, addr = next, addr != end);
 213
 214         return pages;
 215 }
 216
 217 static unsigned long change_protection_range(struct vm_area_struct *vma,
 218                 unsigned long addr, unsigned long end, pgprot_t newprot,
 219                 int dirty_accountable, int prot_numa)
 220 {
 221         struct mm_struct *mm = vma->vm_mm;
 222         pgd_t *pgd;
 223         unsigned long next;
 224         unsigned long start = addr;
 225         unsigned long pages = 0;
 226
 227         BUG_ON(addr >= end);
 228         pgd = pgd_offset(mm, addr);
 229         flush_cache_range(vma, addr, end);
 230         inc_tlb_flush_pending(mm);
 231         do {
 232                 next = pgd_addr_end(addr, end);
 233                 if (pgd_none_or_clear_bad(pgd))
 234                         continue;
 235                 pages += change_pud_range(vma, pgd, addr, next, newprot,
 236                                  dirty_accountable, prot_numa);
 237         } while (pgd++, addr = next, addr != end);
 238
 239         /* Only flush the TLB if we actually modified any entries: */
 240         if (pages)
 241                 flush_tlb_range(vma, start, end);
 242         dec_tlb_flush_pending(mm);
 243
 244         return pages;
 245 }
 246
 247 unsigned long change_protection(struct vm_area_struct *vma, unsigned long start,
 248                        unsigned long end, pgprot_t newprot,
 249                        int dirty_accountable, int prot_numa)
 250 {
 251         unsigned long pages;
 252
 253         if (is_vm_hugetlb_page(vma))
 254                 pages = hugetlb_change_protection(vma, start, end, newprot);
 255         else
 256                 pages = change_protection_range(vma, start, end, newprot, dirty_accountable, prot_numa);
 257
 258         return pages;
 259 }
 260
 261 static int prot_none_pte_entry(pte_t *pte, unsigned long addr,
 262                                unsigned long next, struct mm_walk *walk)
 263 {
 264         return pfn_modify_allowed(pte_pfn(*pte), *(pgprot_t *)(walk->private)) ?
 265                 0 : -EACCES;
 266 }
 267
 268 static int prot_none_hugetlb_entry(pte_t *pte, unsigned long hmask,
 269                                    unsigned long addr, unsigned long next,
 270                                    struct mm_walk *walk)
 271 {
 272         return pfn_modify_allowed(pte_pfn(*pte), *(pgprot_t *)(walk->private)) ?
 273                 0 : -EACCES;
 274 }
 275
 276 static int prot_none_test(unsigned long addr, unsigned long next,
 277                           struct mm_walk *walk)
 278 {
 279         return 0;
 280 }
 281
 282 static int prot_none_walk(struct vm_area_struct *vma, unsigned long start,
 283                            unsigned long end, unsigned long newflags)
 284 {
 285         pgprot_t new_pgprot = vm_get_page_prot(newflags);
 286         struct mm_walk prot_none_walk = {
 287                 .pte_entry = prot_none_pte_entry,
 288                 .hugetlb_entry = prot_none_hugetlb_entry,
 289                 .test_walk = prot_none_test,
 290                 .mm = current->mm,
 291                 .private = &new_pgprot,
 292         };
 293
 294         return walk_page_range(start, end, &prot_none_walk);
 295 }
 296
 297 int
 298 mprotect_fixup(struct vm_area_struct *vma, struct vm_area_struct **pprev,
 299         unsigned long start, unsigned long end, unsigned long newflags)
 300 {
 301         struct mm_struct *mm = vma->vm_mm;
 302         unsigned long oldflags = vma->vm_flags;
 303         long nrpages = (end - start) >> PAGE_SHIFT;
 304         unsigned long charged = 0;
 305         pgoff_t pgoff;
 306         int error;
 307         int dirty_accountable = 0;
 308
 309         if (newflags == oldflags) {
 310                 *pprev = vma;
 311                 return 0;
 312         }
 313
 314         /*
 315          * Do PROT_NONE PFN permission checks here when we can still
 316          * bail out without undoing a lot of state. This is a rather
 317          * uncommon case, so doesn't need to be very optimized.
 318          */
 319         if (arch_has_pfn_modify_check() &&
 320             (vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 321             (newflags & (VM_READ|VM_WRITE|VM_EXEC)) == 0) {
 322                 error = prot_none_walk(vma, start, end, newflags);
 323                 if (error)
 324                         return error;
 325         }
 326
 327         /*
 328          * If we make a private mapping writable we increase our commit;
 329          * but (without finer accounting) cannot reduce our commit if we
 330          * make it unwritable again. hugetlb mapping were accounted for
 331          * even if read-only so there is no need to account for them here
 332          */
 333         if (newflags & VM_WRITE) {
 334                 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
 335                                                 VM_SHARED|VM_NORESERVE))) {
 336                         charged = nrpages;
 337                         if (security_vm_enough_memory_mm(mm, charged))
 338                                 return -ENOMEM;
 339                         newflags |= VM_ACCOUNT;
 340                 }
 341         }
 342
 343         /*
 344          * First try to merge with previous and/or next vma.
 345          */
 346         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 347         *pprev = vma_merge(mm, *pprev, start, end, newflags,
 348                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
 349         if (*pprev) {
 350                 vma = *pprev;
 351                 goto success;
 352         }
 353
 354         *pprev = vma;
 355
 356         if (start != vma->vm_start) {
 357                 error = split_vma(mm, vma, start, 1);
 358                 if (error)
 359                         goto fail;
 360         }
 361
 362         if (end != vma->vm_end) {
 363                 error = split_vma(mm, vma, end, 0);
 364                 if (error)
 365                         goto fail;
 366         }
 367
 368 success:
 369         /*
 370          * vm_flags and vm_page_prot are protected by the mmap_sem
 371          * held in write mode.
 372          */
 373         vma->vm_flags = newflags;
 374         vma->vm_page_prot = pgprot_modify(vma->vm_page_prot,
 375                                           vm_get_page_prot(newflags));
 376
 377         if (vma_wants_writenotify(vma)) {
 378                 vma->vm_page_prot = vm_get_page_prot(newflags & ~VM_SHARED);
 379                 dirty_accountable = 1;
 380         }
 381
 382         change_protection(vma, start, end, vma->vm_page_prot,
 383                           dirty_accountable, 0);
 384
 385         vm_stat_account(mm, oldflags, vma->vm_file, -nrpages);
 386         vm_stat_account(mm, newflags, vma->vm_file, nrpages);
 387         perf_event_mmap(vma);
 388         return 0;
 389
 390 fail:
 391         vm_unacct_memory(charged);
 392         return error;
 393 }
 394
 395 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
 396                 unsigned long, prot)
 397 {
 398         unsigned long vm_flags, nstart, end, tmp, reqprot;
 399         struct vm_area_struct *vma, *prev;
 400         int error = -EINVAL;
 401         const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
 402         prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
 403         if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
 404                 return -EINVAL;
 405
 406         if (start & ~PAGE_MASK)
 407                 return -EINVAL;
 408         if (!len)
 409                 return 0;
 410         len = PAGE_ALIGN(len);
 411         end = start + len;
 412         if (end <= start)
 413                 return -ENOMEM;
 414         if (!arch_validate_prot(prot))
 415                 return -EINVAL;
 416
 417         reqprot = prot;
 418         /*
 419          * Does the application expect PROT_READ to imply PROT_EXEC:
 420          */
 421         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
 422                 prot |= PROT_EXEC;
 423
 424         vm_flags = calc_vm_prot_bits(prot);
 425
 426         down_write(&current->mm->mmap_sem);
 427
 428         vma = find_vma(current->mm, start);
 429         error = -ENOMEM;
 430         if (!vma)
 431                 goto out;
 432         prev = vma->vm_prev;
 433         if (unlikely(grows & PROT_GROWSDOWN)) {
 434                 if (vma->vm_start >= end)
 435                         goto out;
 436                 start = vma->vm_start;
 437                 error = -EINVAL;
 438                 if (!(vma->vm_flags & VM_GROWSDOWN))
 439                         goto out;
 440         } else {
 441                 if (vma->vm_start > start)
 442                         goto out;
 443                 if (unlikely(grows & PROT_GROWSUP)) {
 444                         end = vma->vm_end;
 445                         error = -EINVAL;
 446                         if (!(vma->vm_flags & VM_GROWSUP))
 447                                 goto out;
 448                 }
 449         }
 450         if (start > vma->vm_start)
 451                 prev = vma;
 452
 453         for (nstart = start ; ; ) {
 454                 unsigned long newflags;
 455
 456                 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
 457
 458                 newflags = vm_flags;
 459                 newflags |= (vma->vm_flags & ~(VM_READ | VM_WRITE | VM_EXEC));
 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         }
 490 out:
 491         up_write(&current->mm->mmap_sem);
 492         return error;
 493 }