1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * mm/mprotect.c 4 * 5 * (C) Copyright 1994 Linus Torvalds 6 * (C) Copyright 2002 Christoph Hellwig 7 * 8 * Address space accounting code <alan@lxorguk.ukuu.org.uk> 9 * (C) Copyright 2002 Red Hat Inc, All Rights Reserved 10 */ 11 12 #include <linux/pagewalk.h> 13 #include <linux/hugetlb.h> 14 #include <linux/shm.h> 15 #include <linux/mman.h> 16 #include <linux/fs.h> 17 #include <linux/highmem.h> 18 #include <linux/security.h> 19 #include <linux/mempolicy.h> 20 #include <linux/personality.h> 21 #include <linux/syscalls.h> 22 #include <linux/swap.h> 23 #include <linux/swapops.h> 24 #include <linux/mmu_notifier.h> 25 #include <linux/migrate.h> 26 #include <linux/perf_event.h> 27 #include <linux/pkeys.h> 28 #include <linux/ksm.h> 29 #include <linux/uaccess.h> 30 #include <linux/mm_inline.h> 31 #include <asm/pgtable.h> 32 #include <asm/cacheflush.h> 33 #include <asm/mmu_context.h> 34 #include <asm/tlbflush.h> 35 36 #include "internal.h" 37 38 static unsigned long change_pte_range(struct vm_area_struct *vma, pmd_t *pmd, 39 unsigned long addr, unsigned long end, pgprot_t newprot, 40 int dirty_accountable, int prot_numa) 41 { 42 pte_t *pte, oldpte; 43 spinlock_t *ptl; 44 unsigned long pages = 0; 45 int target_node = NUMA_NO_NODE; 46 47 /* 48 * Can be called with only the mmap_sem for reading by 49 * prot_numa so we must check the pmd isn't constantly 50 * changing from under us from pmd_none to pmd_trans_huge 51 * and/or the other way around. 52 */ 53 if (pmd_trans_unstable(pmd)) 54 return 0; 55 56 /* 57 * The pmd points to a regular pte so the pmd can't change 58 * from under us even if the mmap_sem is only hold for 59 * reading. 60 */ 61 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl); 62 63 /* Get target node for single threaded private VMAs */ 64 if (prot_numa && !(vma->vm_flags & VM_SHARED) && 65 atomic_read(&vma->vm_mm->mm_users) == 1) 66 target_node = numa_node_id(); 67 68 flush_tlb_batched_pending(vma->vm_mm); 69 arch_enter_lazy_mmu_mode(); 70 do { 71 oldpte = *pte; 72 if (pte_present(oldpte)) { 73 pte_t ptent; 74 bool preserve_write = prot_numa && pte_write(oldpte); 75 76 /* 77 * Avoid trapping faults against the zero or KSM 78 * pages. See similar comment in change_huge_pmd. 79 */ 80 if (prot_numa) { 81 struct page *page; 82 83 page = vm_normal_page(vma, addr, oldpte); 84 if (!page || PageKsm(page)) 85 continue; 86 87 /* Also skip shared copy-on-write pages */ 88 if (is_cow_mapping(vma->vm_flags) && 89 page_mapcount(page) != 1) 90 continue; 91 92 /* 93 * While migration can move some dirty pages, 94 * it cannot move them all from MIGRATE_ASYNC 95 * context. 96 */ 97 if (page_is_file_cache(page) && PageDirty(page)) 98 continue; 99 100 /* Avoid TLB flush if possible */ 101 if (pte_protnone(oldpte)) 102 continue; 103 104 /* 105 * Don't mess with PTEs if page is already on the node 106 * a single-threaded process is running on. 107 */ 108 if (target_node == page_to_nid(page)) 109 continue; 110 } 111 112 oldpte = ptep_modify_prot_start(vma, addr, pte); 113 ptent = pte_modify(oldpte, newprot); 114 if (preserve_write) 115 ptent = pte_mk_savedwrite(ptent); 116 117 /* Avoid taking write faults for known dirty pages */ 118 if (dirty_accountable && pte_dirty(ptent) && 119 (pte_soft_dirty(ptent) || 120 !(vma->vm_flags & VM_SOFTDIRTY))) { 121 ptent = pte_mkwrite(ptent); 122 } 123 ptep_modify_prot_commit(vma, addr, pte, oldpte, ptent); 124 pages++; 125 } else if (IS_ENABLED(CONFIG_MIGRATION)) { 126 swp_entry_t entry = pte_to_swp_entry(oldpte); 127 128 if (is_write_migration_entry(entry)) { 129 pte_t newpte; 130 /* 131 * A protection check is difficult so 132 * just be safe and disable write 133 */ 134 make_migration_entry_read(&entry); 135 newpte = swp_entry_to_pte(entry); 136 if (pte_swp_soft_dirty(oldpte)) 137 newpte = pte_swp_mksoft_dirty(newpte); 138 set_pte_at(vma->vm_mm, addr, pte, newpte); 139 140 pages++; 141 } 142 143 if (is_write_device_private_entry(entry)) { 144 pte_t newpte; 145 146 /* 147 * We do not preserve soft-dirtiness. See 148 * copy_one_pte() for explanation. 149 */ 150 make_device_private_entry_read(&entry); 151 newpte = swp_entry_to_pte(entry); 152 set_pte_at(vma->vm_mm, addr, pte, newpte); 153 154 pages++; 155 } 156 } 157 } while (pte++, addr += PAGE_SIZE, addr != end); 158 arch_leave_lazy_mmu_mode(); 159 pte_unmap_unlock(pte - 1, ptl); 160 161 return pages; 162 } 163 164 static inline unsigned long change_pmd_range(struct vm_area_struct *vma, 165 pud_t *pud, unsigned long addr, unsigned long end, 166 pgprot_t newprot, int dirty_accountable, int prot_numa) 167 { 168 pmd_t *pmd; 169 unsigned long next; 170 unsigned long pages = 0; 171 unsigned long nr_huge_updates = 0; 172 struct mmu_notifier_range range; 173 174 range.start = 0; 175 176 pmd = pmd_offset(pud, addr); 177 do { 178 unsigned long this_pages; 179 180 next = pmd_addr_end(addr, end); 181 if (!is_swap_pmd(*pmd) && !pmd_trans_huge(*pmd) && !pmd_devmap(*pmd) 182 && pmd_none_or_clear_bad(pmd)) 183 goto next; 184 185 /* invoke the mmu notifier if the pmd is populated */ 186 if (!range.start) { 187 mmu_notifier_range_init(&range, 188 MMU_NOTIFY_PROTECTION_VMA, 0, 189 vma, vma->vm_mm, addr, end); 190 mmu_notifier_invalidate_range_start(&range); 191 } 192 193 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) { 194 if (next - addr != HPAGE_PMD_SIZE) { 195 __split_huge_pmd(vma, pmd, addr, false, NULL); 196 } else { 197 int nr_ptes = change_huge_pmd(vma, pmd, addr, 198 newprot, prot_numa); 199 200 if (nr_ptes) { 201 if (nr_ptes == HPAGE_PMD_NR) { 202 pages += HPAGE_PMD_NR; 203 nr_huge_updates++; 204 } 205 206 /* huge pmd was handled */ 207 goto next; 208 } 209 } 210 /* fall through, the trans huge pmd just split */ 211 } 212 this_pages = change_pte_range(vma, pmd, addr, next, newprot, 213 dirty_accountable, prot_numa); 214 pages += this_pages; 215 next: 216 cond_resched(); 217 } while (pmd++, addr = next, addr != end); 218 219 if (range.start) 220 mmu_notifier_invalidate_range_end(&range); 221 222 if (nr_huge_updates) 223 count_vm_numa_events(NUMA_HUGE_PTE_UPDATES, nr_huge_updates); 224 return pages; 225 } 226 227 static inline unsigned long change_pud_range(struct vm_area_struct *vma, 228 p4d_t *p4d, unsigned long addr, unsigned long end, 229 pgprot_t newprot, int dirty_accountable, int prot_numa) 230 { 231 pud_t *pud; 232 unsigned long next; 233 unsigned long pages = 0; 234 235 pud = pud_offset(p4d, addr); 236 do { 237 next = pud_addr_end(addr, end); 238 if (pud_none_or_clear_bad(pud)) 239 continue; 240 pages += change_pmd_range(vma, pud, addr, next, newprot, 241 dirty_accountable, prot_numa); 242 } while (pud++, addr = next, addr != end); 243 244 return pages; 245 } 246 247 static inline unsigned long change_p4d_range(struct vm_area_struct *vma, 248 pgd_t *pgd, unsigned long addr, unsigned long end, 249 pgprot_t newprot, int dirty_accountable, int prot_numa) 250 { 251 p4d_t *p4d; 252 unsigned long next; 253 unsigned long pages = 0; 254 255 p4d = p4d_offset(pgd, addr); 256 do { 257 next = p4d_addr_end(addr, end); 258 if (p4d_none_or_clear_bad(p4d)) 259 continue; 260 pages += change_pud_range(vma, p4d, addr, next, newprot, 261 dirty_accountable, prot_numa); 262 } while (p4d++, addr = next, addr != end); 263 264 return pages; 265 } 266 267 static unsigned long change_protection_range(struct vm_area_struct *vma, 268 unsigned long addr, unsigned long end, pgprot_t newprot, 269 int dirty_accountable, int prot_numa) 270 { 271 struct mm_struct *mm = vma->vm_mm; 272 pgd_t *pgd; 273 unsigned long next; 274 unsigned long start = addr; 275 unsigned long pages = 0; 276 277 BUG_ON(addr >= end); 278 pgd = pgd_offset(mm, addr); 279 flush_cache_range(vma, addr, end); 280 inc_tlb_flush_pending(mm); 281 do { 282 next = pgd_addr_end(addr, end); 283 if (pgd_none_or_clear_bad(pgd)) 284 continue; 285 pages += change_p4d_range(vma, pgd, addr, next, newprot, 286 dirty_accountable, prot_numa); 287 } while (pgd++, addr = next, addr != end); 288 289 /* Only flush the TLB if we actually modified any entries: */ 290 if (pages) 291 flush_tlb_range(vma, start, end); 292 dec_tlb_flush_pending(mm); 293 294 return pages; 295 } 296 297 unsigned long change_protection(struct vm_area_struct *vma, unsigned long start, 298 unsigned long end, pgprot_t newprot, 299 int dirty_accountable, int prot_numa) 300 { 301 unsigned long pages; 302 303 if (is_vm_hugetlb_page(vma)) 304 pages = hugetlb_change_protection(vma, start, end, newprot); 305 else 306 pages = change_protection_range(vma, start, end, newprot, dirty_accountable, prot_numa); 307 308 return pages; 309 } 310 311 static int prot_none_pte_entry(pte_t *pte, unsigned long addr, 312 unsigned long next, struct mm_walk *walk) 313 { 314 return pfn_modify_allowed(pte_pfn(*pte), *(pgprot_t *)(walk->private)) ? 315 0 : -EACCES; 316 } 317 318 static int prot_none_hugetlb_entry(pte_t *pte, unsigned long hmask, 319 unsigned long addr, unsigned long next, 320 struct mm_walk *walk) 321 { 322 return pfn_modify_allowed(pte_pfn(*pte), *(pgprot_t *)(walk->private)) ? 323 0 : -EACCES; 324 } 325 326 static int prot_none_test(unsigned long addr, unsigned long next, 327 struct mm_walk *walk) 328 { 329 return 0; 330 } 331 332 static const struct mm_walk_ops prot_none_walk_ops = { 333 .pte_entry = prot_none_pte_entry, 334 .hugetlb_entry = prot_none_hugetlb_entry, 335 .test_walk = prot_none_test, 336 }; 337 338 int 339 mprotect_fixup(struct vm_area_struct *vma, struct vm_area_struct **pprev, 340 unsigned long start, unsigned long end, unsigned long newflags) 341 { 342 struct mm_struct *mm = vma->vm_mm; 343 unsigned long oldflags = vma->vm_flags; 344 long nrpages = (end - start) >> PAGE_SHIFT; 345 unsigned long charged = 0; 346 pgoff_t pgoff; 347 int error; 348 int dirty_accountable = 0; 349 350 if (newflags == oldflags) { 351 *pprev = vma; 352 return 0; 353 } 354 355 /* 356 * Do PROT_NONE PFN permission checks here when we can still 357 * bail out without undoing a lot of state. This is a rather 358 * uncommon case, so doesn't need to be very optimized. 359 */ 360 if (arch_has_pfn_modify_check() && 361 (vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) && 362 (newflags & (VM_READ|VM_WRITE|VM_EXEC)) == 0) { 363 pgprot_t new_pgprot = vm_get_page_prot(newflags); 364 365 error = walk_page_range(current->mm, start, end, 366 &prot_none_walk_ops, &new_pgprot); 367 if (error) 368 return error; 369 } 370 371 /* 372 * If we make a private mapping writable we increase our commit; 373 * but (without finer accounting) cannot reduce our commit if we 374 * make it unwritable again. hugetlb mapping were accounted for 375 * even if read-only so there is no need to account for them here 376 */ 377 if (newflags & VM_WRITE) { 378 /* Check space limits when area turns into data. */ 379 if (!may_expand_vm(mm, newflags, nrpages) && 380 may_expand_vm(mm, oldflags, nrpages)) 381 return -ENOMEM; 382 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB| 383 VM_SHARED|VM_NORESERVE))) { 384 charged = nrpages; 385 if (security_vm_enough_memory_mm(mm, charged)) 386 return -ENOMEM; 387 newflags |= VM_ACCOUNT; 388 } 389 } 390 391 /* 392 * First try to merge with previous and/or next vma. 393 */ 394 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT); 395 *pprev = vma_merge(mm, *pprev, start, end, newflags, 396 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma), 397 vma->vm_userfaultfd_ctx); 398 if (*pprev) { 399 vma = *pprev; 400 VM_WARN_ON((vma->vm_flags ^ newflags) & ~VM_SOFTDIRTY); 401 goto success; 402 } 403 404 *pprev = vma; 405 406 if (start != vma->vm_start) { 407 error = split_vma(mm, vma, start, 1); 408 if (error) 409 goto fail; 410 } 411 412 if (end != vma->vm_end) { 413 error = split_vma(mm, vma, end, 0); 414 if (error) 415 goto fail; 416 } 417 418 success: 419 /* 420 * vm_flags and vm_page_prot are protected by the mmap_sem 421 * held in write mode. 422 */ 423 vma->vm_flags = newflags; 424 dirty_accountable = vma_wants_writenotify(vma, vma->vm_page_prot); 425 vma_set_page_prot(vma); 426 427 change_protection(vma, start, end, vma->vm_page_prot, 428 dirty_accountable, 0); 429 430 /* 431 * Private VM_LOCKED VMA becoming writable: trigger COW to avoid major 432 * fault on access. 433 */ 434 if ((oldflags & (VM_WRITE | VM_SHARED | VM_LOCKED)) == VM_LOCKED && 435 (newflags & VM_WRITE)) { 436 populate_vma_page_range(vma, start, end, NULL); 437 } 438 439 vm_stat_account(mm, oldflags, -nrpages); 440 vm_stat_account(mm, newflags, nrpages); 441 perf_event_mmap(vma); 442 return 0; 443 444 fail: 445 vm_unacct_memory(charged); 446 return error; 447 } 448 449 /* 450 * pkey==-1 when doing a legacy mprotect() 451 */ 452 static int do_mprotect_pkey(unsigned long start, size_t len, 453 unsigned long prot, int pkey) 454 { 455 unsigned long nstart, end, tmp, reqprot; 456 struct vm_area_struct *vma, *prev; 457 int error = -EINVAL; 458 const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP); 459 const bool rier = (current->personality & READ_IMPLIES_EXEC) && 460 (prot & PROT_READ); 461 462 prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP); 463 if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */ 464 return -EINVAL; 465 466 if (start & ~PAGE_MASK) 467 return -EINVAL; 468 if (!len) 469 return 0; 470 len = PAGE_ALIGN(len); 471 end = start + len; 472 if (end <= start) 473 return -ENOMEM; 474 if (!arch_validate_prot(prot, start)) 475 return -EINVAL; 476 477 reqprot = prot; 478 479 if (down_write_killable(¤t->mm->mmap_sem)) 480 return -EINTR; 481 482 /* 483 * If userspace did not allocate the pkey, do not let 484 * them use it here. 485 */ 486 error = -EINVAL; 487 if ((pkey != -1) && !mm_pkey_is_allocated(current->mm, pkey)) 488 goto out; 489 490 vma = find_vma(current->mm, start); 491 error = -ENOMEM; 492 if (!vma) 493 goto out; 494 prev = vma->vm_prev; 495 if (unlikely(grows & PROT_GROWSDOWN)) { 496 if (vma->vm_start >= end) 497 goto out; 498 start = vma->vm_start; 499 error = -EINVAL; 500 if (!(vma->vm_flags & VM_GROWSDOWN)) 501 goto out; 502 } else { 503 if (vma->vm_start > start) 504 goto out; 505 if (unlikely(grows & PROT_GROWSUP)) { 506 end = vma->vm_end; 507 error = -EINVAL; 508 if (!(vma->vm_flags & VM_GROWSUP)) 509 goto out; 510 } 511 } 512 if (start > vma->vm_start) 513 prev = vma; 514 515 for (nstart = start ; ; ) { 516 unsigned long mask_off_old_flags; 517 unsigned long newflags; 518 int new_vma_pkey; 519 520 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */ 521 522 /* Does the application expect PROT_READ to imply PROT_EXEC */ 523 if (rier && (vma->vm_flags & VM_MAYEXEC)) 524 prot |= PROT_EXEC; 525 526 /* 527 * Each mprotect() call explicitly passes r/w/x permissions. 528 * If a permission is not passed to mprotect(), it must be 529 * cleared from the VMA. 530 */ 531 mask_off_old_flags = VM_READ | VM_WRITE | VM_EXEC | 532 VM_FLAGS_CLEAR; 533 534 new_vma_pkey = arch_override_mprotect_pkey(vma, prot, pkey); 535 newflags = calc_vm_prot_bits(prot, new_vma_pkey); 536 newflags |= (vma->vm_flags & ~mask_off_old_flags); 537 538 /* newflags >> 4 shift VM_MAY% in place of VM_% */ 539 if ((newflags & ~(newflags >> 4)) & (VM_READ | VM_WRITE | VM_EXEC)) { 540 error = -EACCES; 541 goto out; 542 } 543 544 error = security_file_mprotect(vma, reqprot, prot); 545 if (error) 546 goto out; 547 548 tmp = vma->vm_end; 549 if (tmp > end) 550 tmp = end; 551 error = mprotect_fixup(vma, &prev, nstart, tmp, newflags); 552 if (error) 553 goto out; 554 nstart = tmp; 555 556 if (nstart < prev->vm_end) 557 nstart = prev->vm_end; 558 if (nstart >= end) 559 goto out; 560 561 vma = prev->vm_next; 562 if (!vma || vma->vm_start != nstart) { 563 error = -ENOMEM; 564 goto out; 565 } 566 prot = reqprot; 567 } 568 out: 569 up_write(¤t->mm->mmap_sem); 570 return error; 571 } 572 573 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len, 574 unsigned long, prot) 575 { 576 return do_mprotect_pkey(start, len, prot, -1); 577 } 578 579 #ifdef CONFIG_ARCH_HAS_PKEYS 580 581 SYSCALL_DEFINE4(pkey_mprotect, unsigned long, start, size_t, len, 582 unsigned long, prot, int, pkey) 583 { 584 return do_mprotect_pkey(start, len, prot, pkey); 585 } 586 587 SYSCALL_DEFINE2(pkey_alloc, unsigned long, flags, unsigned long, init_val) 588 { 589 int pkey; 590 int ret; 591 592 /* No flags supported yet. */ 593 if (flags) 594 return -EINVAL; 595 /* check for unsupported init values */ 596 if (init_val & ~PKEY_ACCESS_MASK) 597 return -EINVAL; 598 599 down_write(¤t->mm->mmap_sem); 600 pkey = mm_pkey_alloc(current->mm); 601 602 ret = -ENOSPC; 603 if (pkey == -1) 604 goto out; 605 606 ret = arch_set_user_pkey_access(current, pkey, init_val); 607 if (ret) { 608 mm_pkey_free(current->mm, pkey); 609 goto out; 610 } 611 ret = pkey; 612 out: 613 up_write(¤t->mm->mmap_sem); 614 return ret; 615 } 616 617 SYSCALL_DEFINE1(pkey_free, int, pkey) 618 { 619 int ret; 620 621 down_write(¤t->mm->mmap_sem); 622 ret = mm_pkey_free(current->mm, pkey); 623 up_write(¤t->mm->mmap_sem); 624 625 /* 626 * We could provie warnings or errors if any VMA still 627 * has the pkey set here. 628 */ 629 return ret; 630 } 631 632 #endif /* CONFIG_ARCH_HAS_PKEYS */ 633