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 #include "internal.h" 33 34 /* 35 * For a prot_numa update we only hold mmap_sem for read so there is a 36 * potential race with faulting where a pmd was temporarily none. This 37 * function checks for a transhuge pmd under the appropriate lock. It 38 * returns a pte if it was successfully locked or NULL if it raced with 39 * a transhuge insertion. 40 */ 41 static pte_t *lock_pte_protection(struct vm_area_struct *vma, pmd_t *pmd, 42 unsigned long addr, int prot_numa, spinlock_t **ptl) 43 { 44 pte_t *pte; 45 spinlock_t *pmdl; 46 47 /* !prot_numa is protected by mmap_sem held for write */ 48 if (!prot_numa) 49 return pte_offset_map_lock(vma->vm_mm, pmd, addr, ptl); 50 51 pmdl = pmd_lock(vma->vm_mm, pmd); 52 if (unlikely(pmd_trans_huge(*pmd) || pmd_none(*pmd))) { 53 spin_unlock(pmdl); 54 return NULL; 55 } 56 57 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, ptl); 58 spin_unlock(pmdl); 59 return pte; 60 } 61 62 static unsigned long change_pte_range(struct vm_area_struct *vma, pmd_t *pmd, 63 unsigned long addr, unsigned long end, pgprot_t newprot, 64 int dirty_accountable, int prot_numa) 65 { 66 struct mm_struct *mm = vma->vm_mm; 67 pte_t *pte, oldpte; 68 spinlock_t *ptl; 69 unsigned long pages = 0; 70 71 pte = lock_pte_protection(vma, pmd, addr, prot_numa, &ptl); 72 if (!pte) 73 return 0; 74 75 arch_enter_lazy_mmu_mode(); 76 do { 77 oldpte = *pte; 78 if (pte_present(oldpte)) { 79 pte_t ptent; 80 bool preserve_write = prot_numa && pte_write(oldpte); 81 82 /* 83 * Avoid trapping faults against the zero or KSM 84 * pages. See similar comment in change_huge_pmd. 85 */ 86 if (prot_numa) { 87 struct page *page; 88 89 page = vm_normal_page(vma, addr, oldpte); 90 if (!page || PageKsm(page)) 91 continue; 92 93 /* Avoid TLB flush if possible */ 94 if (pte_protnone(oldpte)) 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_mkwrite(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); 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 pgd_t *pgd, 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(pgd, 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 unsigned long change_protection_range(struct vm_area_struct *vma, 215 unsigned long addr, unsigned long end, pgprot_t newprot, 216 int dirty_accountable, int prot_numa) 217 { 218 struct mm_struct *mm = vma->vm_mm; 219 pgd_t *pgd; 220 unsigned long next; 221 unsigned long start = addr; 222 unsigned long pages = 0; 223 224 BUG_ON(addr >= end); 225 pgd = pgd_offset(mm, addr); 226 flush_cache_range(vma, addr, end); 227 set_tlb_flush_pending(mm); 228 do { 229 next = pgd_addr_end(addr, end); 230 if (pgd_none_or_clear_bad(pgd)) 231 continue; 232 pages += change_pud_range(vma, pgd, addr, next, newprot, 233 dirty_accountable, prot_numa); 234 } while (pgd++, addr = next, addr != end); 235 236 /* Only flush the TLB if we actually modified any entries: */ 237 if (pages) 238 flush_tlb_range(vma, start, end); 239 clear_tlb_flush_pending(mm); 240 241 return pages; 242 } 243 244 unsigned long change_protection(struct vm_area_struct *vma, unsigned long start, 245 unsigned long end, pgprot_t newprot, 246 int dirty_accountable, int prot_numa) 247 { 248 unsigned long pages; 249 250 if (is_vm_hugetlb_page(vma)) 251 pages = hugetlb_change_protection(vma, start, end, newprot); 252 else 253 pages = change_protection_range(vma, start, end, newprot, dirty_accountable, prot_numa); 254 255 return pages; 256 } 257 258 int 259 mprotect_fixup(struct vm_area_struct *vma, struct vm_area_struct **pprev, 260 unsigned long start, unsigned long end, unsigned long newflags) 261 { 262 struct mm_struct *mm = vma->vm_mm; 263 unsigned long oldflags = vma->vm_flags; 264 long nrpages = (end - start) >> PAGE_SHIFT; 265 unsigned long charged = 0; 266 pgoff_t pgoff; 267 int error; 268 int dirty_accountable = 0; 269 270 if (newflags == oldflags) { 271 *pprev = vma; 272 return 0; 273 } 274 275 /* 276 * If we make a private mapping writable we increase our commit; 277 * but (without finer accounting) cannot reduce our commit if we 278 * make it unwritable again. hugetlb mapping were accounted for 279 * even if read-only so there is no need to account for them here 280 */ 281 if (newflags & VM_WRITE) { 282 /* Check space limits when area turns into data. */ 283 if (!may_expand_vm(mm, newflags, nrpages) && 284 may_expand_vm(mm, oldflags, nrpages)) 285 return -ENOMEM; 286 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB| 287 VM_SHARED|VM_NORESERVE))) { 288 charged = nrpages; 289 if (security_vm_enough_memory_mm(mm, charged)) 290 return -ENOMEM; 291 newflags |= VM_ACCOUNT; 292 } 293 } 294 295 /* 296 * First try to merge with previous and/or next vma. 297 */ 298 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT); 299 *pprev = vma_merge(mm, *pprev, start, end, newflags, 300 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma), 301 vma->vm_userfaultfd_ctx); 302 if (*pprev) { 303 vma = *pprev; 304 goto success; 305 } 306 307 *pprev = vma; 308 309 if (start != vma->vm_start) { 310 error = split_vma(mm, vma, start, 1); 311 if (error) 312 goto fail; 313 } 314 315 if (end != vma->vm_end) { 316 error = split_vma(mm, vma, end, 0); 317 if (error) 318 goto fail; 319 } 320 321 success: 322 /* 323 * vm_flags and vm_page_prot are protected by the mmap_sem 324 * held in write mode. 325 */ 326 vma->vm_flags = newflags; 327 dirty_accountable = vma_wants_writenotify(vma); 328 vma_set_page_prot(vma); 329 330 change_protection(vma, start, end, vma->vm_page_prot, 331 dirty_accountable, 0); 332 333 /* 334 * Private VM_LOCKED VMA becoming writable: trigger COW to avoid major 335 * fault on access. 336 */ 337 if ((oldflags & (VM_WRITE | VM_SHARED | VM_LOCKED)) == VM_LOCKED && 338 (newflags & VM_WRITE)) { 339 populate_vma_page_range(vma, start, end, NULL); 340 } 341 342 vm_stat_account(mm, oldflags, -nrpages); 343 vm_stat_account(mm, newflags, nrpages); 344 perf_event_mmap(vma); 345 return 0; 346 347 fail: 348 vm_unacct_memory(charged); 349 return error; 350 } 351 352 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len, 353 unsigned long, prot) 354 { 355 unsigned long vm_flags, nstart, end, tmp, reqprot; 356 struct vm_area_struct *vma, *prev; 357 int error = -EINVAL; 358 const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP); 359 prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP); 360 if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */ 361 return -EINVAL; 362 363 if (start & ~PAGE_MASK) 364 return -EINVAL; 365 if (!len) 366 return 0; 367 len = PAGE_ALIGN(len); 368 end = start + len; 369 if (end <= start) 370 return -ENOMEM; 371 if (!arch_validate_prot(prot)) 372 return -EINVAL; 373 374 reqprot = prot; 375 /* 376 * Does the application expect PROT_READ to imply PROT_EXEC: 377 */ 378 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC)) 379 prot |= PROT_EXEC; 380 381 vm_flags = calc_vm_prot_bits(prot); 382 383 down_write(¤t->mm->mmap_sem); 384 385 vma = find_vma(current->mm, start); 386 error = -ENOMEM; 387 if (!vma) 388 goto out; 389 prev = vma->vm_prev; 390 if (unlikely(grows & PROT_GROWSDOWN)) { 391 if (vma->vm_start >= end) 392 goto out; 393 start = vma->vm_start; 394 error = -EINVAL; 395 if (!(vma->vm_flags & VM_GROWSDOWN)) 396 goto out; 397 } else { 398 if (vma->vm_start > start) 399 goto out; 400 if (unlikely(grows & PROT_GROWSUP)) { 401 end = vma->vm_end; 402 error = -EINVAL; 403 if (!(vma->vm_flags & VM_GROWSUP)) 404 goto out; 405 } 406 } 407 if (start > vma->vm_start) 408 prev = vma; 409 410 for (nstart = start ; ; ) { 411 unsigned long newflags; 412 413 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */ 414 415 newflags = vm_flags; 416 newflags |= (vma->vm_flags & ~(VM_READ | VM_WRITE | VM_EXEC)); 417 418 /* newflags >> 4 shift VM_MAY% in place of VM_% */ 419 if ((newflags & ~(newflags >> 4)) & (VM_READ | VM_WRITE | VM_EXEC)) { 420 error = -EACCES; 421 goto out; 422 } 423 424 error = security_file_mprotect(vma, reqprot, prot); 425 if (error) 426 goto out; 427 428 tmp = vma->vm_end; 429 if (tmp > end) 430 tmp = end; 431 error = mprotect_fixup(vma, &prev, nstart, tmp, newflags); 432 if (error) 433 goto out; 434 nstart = tmp; 435 436 if (nstart < prev->vm_end) 437 nstart = prev->vm_end; 438 if (nstart >= end) 439 goto out; 440 441 vma = prev->vm_next; 442 if (!vma || vma->vm_start != nstart) { 443 error = -ENOMEM; 444 goto out; 445 } 446 } 447 out: 448 up_write(¤t->mm->mmap_sem); 449 return error; 450 } 451