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 <linux/pgtable.h> 32 #include <linux/sched/sysctl.h> 33 #include <linux/userfaultfd_k.h> 34 #include <linux/memory-tiers.h> 35 #include <uapi/linux/mman.h> 36 #include <asm/cacheflush.h> 37 #include <asm/mmu_context.h> 38 #include <asm/tlbflush.h> 39 #include <asm/tlb.h> 40 41 #include "internal.h" 42 43 bool can_change_pte_writable(struct vm_area_struct *vma, unsigned long addr, 44 pte_t pte) 45 { 46 struct page *page; 47 48 if (WARN_ON_ONCE(!(vma->vm_flags & VM_WRITE))) 49 return false; 50 51 /* Don't touch entries that are not even readable. */ 52 if (pte_protnone(pte)) 53 return false; 54 55 /* Do we need write faults for softdirty tracking? */ 56 if (pte_needs_soft_dirty_wp(vma, pte)) 57 return false; 58 59 /* Do we need write faults for uffd-wp tracking? */ 60 if (userfaultfd_pte_wp(vma, pte)) 61 return false; 62 63 if (!(vma->vm_flags & VM_SHARED)) { 64 /* 65 * Writable MAP_PRIVATE mapping: We can only special-case on 66 * exclusive anonymous pages, because we know that our 67 * write-fault handler similarly would map them writable without 68 * any additional checks while holding the PT lock. 69 */ 70 page = vm_normal_page(vma, addr, pte); 71 return page && PageAnon(page) && PageAnonExclusive(page); 72 } 73 74 VM_WARN_ON_ONCE(is_zero_pfn(pte_pfn(pte)) && pte_dirty(pte)); 75 76 /* 77 * Writable MAP_SHARED mapping: "clean" might indicate that the FS still 78 * needs a real write-fault for writenotify 79 * (see vma_wants_writenotify()). If "dirty", the assumption is that the 80 * FS was already notified and we can simply mark the PTE writable 81 * just like the write-fault handler would do. 82 */ 83 return pte_dirty(pte); 84 } 85 86 static long change_pte_range(struct mmu_gather *tlb, 87 struct vm_area_struct *vma, pmd_t *pmd, unsigned long addr, 88 unsigned long end, pgprot_t newprot, unsigned long cp_flags) 89 { 90 pte_t *pte, oldpte; 91 spinlock_t *ptl; 92 long pages = 0; 93 int target_node = NUMA_NO_NODE; 94 bool prot_numa = cp_flags & MM_CP_PROT_NUMA; 95 bool uffd_wp = cp_flags & MM_CP_UFFD_WP; 96 bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE; 97 98 tlb_change_page_size(tlb, PAGE_SIZE); 99 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl); 100 if (!pte) 101 return -EAGAIN; 102 103 /* Get target node for single threaded private VMAs */ 104 if (prot_numa && !(vma->vm_flags & VM_SHARED) && 105 atomic_read(&vma->vm_mm->mm_users) == 1) 106 target_node = numa_node_id(); 107 108 flush_tlb_batched_pending(vma->vm_mm); 109 arch_enter_lazy_mmu_mode(); 110 do { 111 oldpte = ptep_get(pte); 112 if (pte_present(oldpte)) { 113 pte_t ptent; 114 115 /* 116 * Avoid trapping faults against the zero or KSM 117 * pages. See similar comment in change_huge_pmd. 118 */ 119 if (prot_numa) { 120 struct folio *folio; 121 int nid; 122 bool toptier; 123 124 /* Avoid TLB flush if possible */ 125 if (pte_protnone(oldpte)) 126 continue; 127 128 folio = vm_normal_folio(vma, addr, oldpte); 129 if (!folio || folio_is_zone_device(folio) || 130 folio_test_ksm(folio)) 131 continue; 132 133 /* Also skip shared copy-on-write pages */ 134 if (is_cow_mapping(vma->vm_flags) && 135 (folio_maybe_dma_pinned(folio) || 136 folio_maybe_mapped_shared(folio))) 137 continue; 138 139 /* 140 * While migration can move some dirty pages, 141 * it cannot move them all from MIGRATE_ASYNC 142 * context. 143 */ 144 if (folio_is_file_lru(folio) && 145 folio_test_dirty(folio)) 146 continue; 147 148 /* 149 * Don't mess with PTEs if page is already on the node 150 * a single-threaded process is running on. 151 */ 152 nid = folio_nid(folio); 153 if (target_node == nid) 154 continue; 155 toptier = node_is_toptier(nid); 156 157 /* 158 * Skip scanning top tier node if normal numa 159 * balancing is disabled 160 */ 161 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_NORMAL) && 162 toptier) 163 continue; 164 if (folio_use_access_time(folio)) 165 folio_xchg_access_time(folio, 166 jiffies_to_msecs(jiffies)); 167 } 168 169 oldpte = ptep_modify_prot_start(vma, addr, pte); 170 ptent = pte_modify(oldpte, newprot); 171 172 if (uffd_wp) 173 ptent = pte_mkuffd_wp(ptent); 174 else if (uffd_wp_resolve) 175 ptent = pte_clear_uffd_wp(ptent); 176 177 /* 178 * In some writable, shared mappings, we might want 179 * to catch actual write access -- see 180 * vma_wants_writenotify(). 181 * 182 * In all writable, private mappings, we have to 183 * properly handle COW. 184 * 185 * In both cases, we can sometimes still change PTEs 186 * writable and avoid the write-fault handler, for 187 * example, if a PTE is already dirty and no other 188 * COW or special handling is required. 189 */ 190 if ((cp_flags & MM_CP_TRY_CHANGE_WRITABLE) && 191 !pte_write(ptent) && 192 can_change_pte_writable(vma, addr, ptent)) 193 ptent = pte_mkwrite(ptent, vma); 194 195 ptep_modify_prot_commit(vma, addr, pte, oldpte, ptent); 196 if (pte_needs_flush(oldpte, ptent)) 197 tlb_flush_pte_range(tlb, addr, PAGE_SIZE); 198 pages++; 199 } else if (is_swap_pte(oldpte)) { 200 swp_entry_t entry = pte_to_swp_entry(oldpte); 201 pte_t newpte; 202 203 if (is_writable_migration_entry(entry)) { 204 struct folio *folio = pfn_swap_entry_folio(entry); 205 206 /* 207 * A protection check is difficult so 208 * just be safe and disable write 209 */ 210 if (folio_test_anon(folio)) 211 entry = make_readable_exclusive_migration_entry( 212 swp_offset(entry)); 213 else 214 entry = make_readable_migration_entry(swp_offset(entry)); 215 newpte = swp_entry_to_pte(entry); 216 if (pte_swp_soft_dirty(oldpte)) 217 newpte = pte_swp_mksoft_dirty(newpte); 218 } else if (is_writable_device_private_entry(entry)) { 219 /* 220 * We do not preserve soft-dirtiness. See 221 * copy_nonpresent_pte() for explanation. 222 */ 223 entry = make_readable_device_private_entry( 224 swp_offset(entry)); 225 newpte = swp_entry_to_pte(entry); 226 if (pte_swp_uffd_wp(oldpte)) 227 newpte = pte_swp_mkuffd_wp(newpte); 228 } else if (is_pte_marker_entry(entry)) { 229 /* 230 * Ignore error swap entries unconditionally, 231 * because any access should sigbus/sigsegv 232 * anyway. 233 */ 234 if (is_poisoned_swp_entry(entry) || 235 is_guard_swp_entry(entry)) 236 continue; 237 /* 238 * If this is uffd-wp pte marker and we'd like 239 * to unprotect it, drop it; the next page 240 * fault will trigger without uffd trapping. 241 */ 242 if (uffd_wp_resolve) { 243 pte_clear(vma->vm_mm, addr, pte); 244 pages++; 245 } 246 continue; 247 } else { 248 newpte = oldpte; 249 } 250 251 if (uffd_wp) 252 newpte = pte_swp_mkuffd_wp(newpte); 253 else if (uffd_wp_resolve) 254 newpte = pte_swp_clear_uffd_wp(newpte); 255 256 if (!pte_same(oldpte, newpte)) { 257 set_pte_at(vma->vm_mm, addr, pte, newpte); 258 pages++; 259 } 260 } else { 261 /* It must be an none page, or what else?.. */ 262 WARN_ON_ONCE(!pte_none(oldpte)); 263 264 /* 265 * Nobody plays with any none ptes besides 266 * userfaultfd when applying the protections. 267 */ 268 if (likely(!uffd_wp)) 269 continue; 270 271 if (userfaultfd_wp_use_markers(vma)) { 272 /* 273 * For file-backed mem, we need to be able to 274 * wr-protect a none pte, because even if the 275 * pte is none, the page/swap cache could 276 * exist. Doing that by install a marker. 277 */ 278 set_pte_at(vma->vm_mm, addr, pte, 279 make_pte_marker(PTE_MARKER_UFFD_WP)); 280 pages++; 281 } 282 } 283 } while (pte++, addr += PAGE_SIZE, addr != end); 284 arch_leave_lazy_mmu_mode(); 285 pte_unmap_unlock(pte - 1, ptl); 286 287 return pages; 288 } 289 290 /* 291 * Return true if we want to split THPs into PTE mappings in change 292 * protection procedure, false otherwise. 293 */ 294 static inline bool 295 pgtable_split_needed(struct vm_area_struct *vma, unsigned long cp_flags) 296 { 297 /* 298 * pte markers only resides in pte level, if we need pte markers, 299 * we need to split. For example, we cannot wr-protect a file thp 300 * (e.g. 2M shmem) because file thp is handled differently when 301 * split by erasing the pmd so far. 302 */ 303 return (cp_flags & MM_CP_UFFD_WP) && !vma_is_anonymous(vma); 304 } 305 306 /* 307 * Return true if we want to populate pgtables in change protection 308 * procedure, false otherwise 309 */ 310 static inline bool 311 pgtable_populate_needed(struct vm_area_struct *vma, unsigned long cp_flags) 312 { 313 /* If not within ioctl(UFFDIO_WRITEPROTECT), then don't bother */ 314 if (!(cp_flags & MM_CP_UFFD_WP)) 315 return false; 316 317 /* Populate if the userfaultfd mode requires pte markers */ 318 return userfaultfd_wp_use_markers(vma); 319 } 320 321 /* 322 * Populate the pgtable underneath for whatever reason if requested. 323 * When {pte|pmd|...}_alloc() failed we treat it the same way as pgtable 324 * allocation failures during page faults by kicking OOM and returning 325 * error. 326 */ 327 #define change_pmd_prepare(vma, pmd, cp_flags) \ 328 ({ \ 329 long err = 0; \ 330 if (unlikely(pgtable_populate_needed(vma, cp_flags))) { \ 331 if (pte_alloc(vma->vm_mm, pmd)) \ 332 err = -ENOMEM; \ 333 } \ 334 err; \ 335 }) 336 337 /* 338 * This is the general pud/p4d/pgd version of change_pmd_prepare(). We need to 339 * have separate change_pmd_prepare() because pte_alloc() returns 0 on success, 340 * while {pmd|pud|p4d}_alloc() returns the valid pointer on success. 341 */ 342 #define change_prepare(vma, high, low, addr, cp_flags) \ 343 ({ \ 344 long err = 0; \ 345 if (unlikely(pgtable_populate_needed(vma, cp_flags))) { \ 346 low##_t *p = low##_alloc(vma->vm_mm, high, addr); \ 347 if (p == NULL) \ 348 err = -ENOMEM; \ 349 } \ 350 err; \ 351 }) 352 353 static inline long change_pmd_range(struct mmu_gather *tlb, 354 struct vm_area_struct *vma, pud_t *pud, unsigned long addr, 355 unsigned long end, pgprot_t newprot, unsigned long cp_flags) 356 { 357 pmd_t *pmd; 358 unsigned long next; 359 long pages = 0; 360 unsigned long nr_huge_updates = 0; 361 362 pmd = pmd_offset(pud, addr); 363 do { 364 long ret; 365 pmd_t _pmd; 366 again: 367 next = pmd_addr_end(addr, end); 368 369 ret = change_pmd_prepare(vma, pmd, cp_flags); 370 if (ret) { 371 pages = ret; 372 break; 373 } 374 375 if (pmd_none(*pmd)) 376 goto next; 377 378 _pmd = pmdp_get_lockless(pmd); 379 if (is_swap_pmd(_pmd) || pmd_trans_huge(_pmd) || pmd_devmap(_pmd)) { 380 if ((next - addr != HPAGE_PMD_SIZE) || 381 pgtable_split_needed(vma, cp_flags)) { 382 __split_huge_pmd(vma, pmd, addr, false); 383 /* 384 * For file-backed, the pmd could have been 385 * cleared; make sure pmd populated if 386 * necessary, then fall-through to pte level. 387 */ 388 ret = change_pmd_prepare(vma, pmd, cp_flags); 389 if (ret) { 390 pages = ret; 391 break; 392 } 393 } else { 394 ret = change_huge_pmd(tlb, vma, pmd, 395 addr, newprot, cp_flags); 396 if (ret) { 397 if (ret == HPAGE_PMD_NR) { 398 pages += HPAGE_PMD_NR; 399 nr_huge_updates++; 400 } 401 402 /* huge pmd was handled */ 403 goto next; 404 } 405 } 406 /* fall through, the trans huge pmd just split */ 407 } 408 409 ret = change_pte_range(tlb, vma, pmd, addr, next, newprot, 410 cp_flags); 411 if (ret < 0) 412 goto again; 413 pages += ret; 414 next: 415 cond_resched(); 416 } while (pmd++, addr = next, addr != end); 417 418 if (nr_huge_updates) 419 count_vm_numa_events(NUMA_HUGE_PTE_UPDATES, nr_huge_updates); 420 return pages; 421 } 422 423 static inline long change_pud_range(struct mmu_gather *tlb, 424 struct vm_area_struct *vma, p4d_t *p4d, unsigned long addr, 425 unsigned long end, pgprot_t newprot, unsigned long cp_flags) 426 { 427 struct mmu_notifier_range range; 428 pud_t *pudp, pud; 429 unsigned long next; 430 long pages = 0, ret; 431 432 range.start = 0; 433 434 pudp = pud_offset(p4d, addr); 435 do { 436 again: 437 next = pud_addr_end(addr, end); 438 ret = change_prepare(vma, pudp, pmd, addr, cp_flags); 439 if (ret) { 440 pages = ret; 441 break; 442 } 443 444 pud = READ_ONCE(*pudp); 445 if (pud_none(pud)) 446 continue; 447 448 if (!range.start) { 449 mmu_notifier_range_init(&range, 450 MMU_NOTIFY_PROTECTION_VMA, 0, 451 vma->vm_mm, addr, end); 452 mmu_notifier_invalidate_range_start(&range); 453 } 454 455 if (pud_leaf(pud)) { 456 if ((next - addr != PUD_SIZE) || 457 pgtable_split_needed(vma, cp_flags)) { 458 __split_huge_pud(vma, pudp, addr); 459 goto again; 460 } else { 461 ret = change_huge_pud(tlb, vma, pudp, 462 addr, newprot, cp_flags); 463 if (ret == 0) 464 goto again; 465 /* huge pud was handled */ 466 if (ret == HPAGE_PUD_NR) 467 pages += HPAGE_PUD_NR; 468 continue; 469 } 470 } 471 472 pages += change_pmd_range(tlb, vma, pudp, addr, next, newprot, 473 cp_flags); 474 } while (pudp++, addr = next, addr != end); 475 476 if (range.start) 477 mmu_notifier_invalidate_range_end(&range); 478 479 return pages; 480 } 481 482 static inline long change_p4d_range(struct mmu_gather *tlb, 483 struct vm_area_struct *vma, pgd_t *pgd, unsigned long addr, 484 unsigned long end, pgprot_t newprot, unsigned long cp_flags) 485 { 486 p4d_t *p4d; 487 unsigned long next; 488 long pages = 0, ret; 489 490 p4d = p4d_offset(pgd, addr); 491 do { 492 next = p4d_addr_end(addr, end); 493 ret = change_prepare(vma, p4d, pud, addr, cp_flags); 494 if (ret) 495 return ret; 496 if (p4d_none_or_clear_bad(p4d)) 497 continue; 498 pages += change_pud_range(tlb, vma, p4d, addr, next, newprot, 499 cp_flags); 500 } while (p4d++, addr = next, addr != end); 501 502 return pages; 503 } 504 505 static long change_protection_range(struct mmu_gather *tlb, 506 struct vm_area_struct *vma, unsigned long addr, 507 unsigned long end, pgprot_t newprot, unsigned long cp_flags) 508 { 509 struct mm_struct *mm = vma->vm_mm; 510 pgd_t *pgd; 511 unsigned long next; 512 long pages = 0, ret; 513 514 BUG_ON(addr >= end); 515 pgd = pgd_offset(mm, addr); 516 tlb_start_vma(tlb, vma); 517 do { 518 next = pgd_addr_end(addr, end); 519 ret = change_prepare(vma, pgd, p4d, addr, cp_flags); 520 if (ret) { 521 pages = ret; 522 break; 523 } 524 if (pgd_none_or_clear_bad(pgd)) 525 continue; 526 pages += change_p4d_range(tlb, vma, pgd, addr, next, newprot, 527 cp_flags); 528 } while (pgd++, addr = next, addr != end); 529 530 tlb_end_vma(tlb, vma); 531 532 return pages; 533 } 534 535 long change_protection(struct mmu_gather *tlb, 536 struct vm_area_struct *vma, unsigned long start, 537 unsigned long end, unsigned long cp_flags) 538 { 539 pgprot_t newprot = vma->vm_page_prot; 540 long pages; 541 542 BUG_ON((cp_flags & MM_CP_UFFD_WP_ALL) == MM_CP_UFFD_WP_ALL); 543 544 #ifdef CONFIG_NUMA_BALANCING 545 /* 546 * Ordinary protection updates (mprotect, uffd-wp, softdirty tracking) 547 * are expected to reflect their requirements via VMA flags such that 548 * vma_set_page_prot() will adjust vma->vm_page_prot accordingly. 549 */ 550 if (cp_flags & MM_CP_PROT_NUMA) 551 newprot = PAGE_NONE; 552 #else 553 WARN_ON_ONCE(cp_flags & MM_CP_PROT_NUMA); 554 #endif 555 556 if (is_vm_hugetlb_page(vma)) 557 pages = hugetlb_change_protection(vma, start, end, newprot, 558 cp_flags); 559 else 560 pages = change_protection_range(tlb, vma, start, end, newprot, 561 cp_flags); 562 563 return pages; 564 } 565 566 static int prot_none_pte_entry(pte_t *pte, unsigned long addr, 567 unsigned long next, struct mm_walk *walk) 568 { 569 return pfn_modify_allowed(pte_pfn(ptep_get(pte)), 570 *(pgprot_t *)(walk->private)) ? 571 0 : -EACCES; 572 } 573 574 static int prot_none_hugetlb_entry(pte_t *pte, unsigned long hmask, 575 unsigned long addr, unsigned long next, 576 struct mm_walk *walk) 577 { 578 return pfn_modify_allowed(pte_pfn(ptep_get(pte)), 579 *(pgprot_t *)(walk->private)) ? 580 0 : -EACCES; 581 } 582 583 static int prot_none_test(unsigned long addr, unsigned long next, 584 struct mm_walk *walk) 585 { 586 return 0; 587 } 588 589 static const struct mm_walk_ops prot_none_walk_ops = { 590 .pte_entry = prot_none_pte_entry, 591 .hugetlb_entry = prot_none_hugetlb_entry, 592 .test_walk = prot_none_test, 593 .walk_lock = PGWALK_WRLOCK, 594 }; 595 596 int 597 mprotect_fixup(struct vma_iterator *vmi, struct mmu_gather *tlb, 598 struct vm_area_struct *vma, struct vm_area_struct **pprev, 599 unsigned long start, unsigned long end, unsigned long newflags) 600 { 601 struct mm_struct *mm = vma->vm_mm; 602 unsigned long oldflags = READ_ONCE(vma->vm_flags); 603 long nrpages = (end - start) >> PAGE_SHIFT; 604 unsigned int mm_cp_flags = 0; 605 unsigned long charged = 0; 606 int error; 607 608 if (!can_modify_vma(vma)) 609 return -EPERM; 610 611 if (newflags == oldflags) { 612 *pprev = vma; 613 return 0; 614 } 615 616 /* 617 * Do PROT_NONE PFN permission checks here when we can still 618 * bail out without undoing a lot of state. This is a rather 619 * uncommon case, so doesn't need to be very optimized. 620 */ 621 if (arch_has_pfn_modify_check() && 622 (oldflags & (VM_PFNMAP|VM_MIXEDMAP)) && 623 (newflags & VM_ACCESS_FLAGS) == 0) { 624 pgprot_t new_pgprot = vm_get_page_prot(newflags); 625 626 error = walk_page_range(current->mm, start, end, 627 &prot_none_walk_ops, &new_pgprot); 628 if (error) 629 return error; 630 } 631 632 /* 633 * If we make a private mapping writable we increase our commit; 634 * but (without finer accounting) cannot reduce our commit if we 635 * make it unwritable again except in the anonymous case where no 636 * anon_vma has yet to be assigned. 637 * 638 * hugetlb mapping were accounted for even if read-only so there is 639 * no need to account for them here. 640 */ 641 if (newflags & VM_WRITE) { 642 /* Check space limits when area turns into data. */ 643 if (!may_expand_vm(mm, newflags, nrpages) && 644 may_expand_vm(mm, oldflags, nrpages)) 645 return -ENOMEM; 646 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB| 647 VM_SHARED|VM_NORESERVE))) { 648 charged = nrpages; 649 if (security_vm_enough_memory_mm(mm, charged)) 650 return -ENOMEM; 651 newflags |= VM_ACCOUNT; 652 } 653 } else if ((oldflags & VM_ACCOUNT) && vma_is_anonymous(vma) && 654 !vma->anon_vma) { 655 newflags &= ~VM_ACCOUNT; 656 } 657 658 vma = vma_modify_flags(vmi, *pprev, vma, start, end, newflags); 659 if (IS_ERR(vma)) { 660 error = PTR_ERR(vma); 661 goto fail; 662 } 663 664 *pprev = vma; 665 666 /* 667 * vm_flags and vm_page_prot are protected by the mmap_lock 668 * held in write mode. 669 */ 670 vma_start_write(vma); 671 vm_flags_reset_once(vma, newflags); 672 if (vma_wants_manual_pte_write_upgrade(vma)) 673 mm_cp_flags |= MM_CP_TRY_CHANGE_WRITABLE; 674 vma_set_page_prot(vma); 675 676 change_protection(tlb, vma, start, end, mm_cp_flags); 677 678 if ((oldflags & VM_ACCOUNT) && !(newflags & VM_ACCOUNT)) 679 vm_unacct_memory(nrpages); 680 681 /* 682 * Private VM_LOCKED VMA becoming writable: trigger COW to avoid major 683 * fault on access. 684 */ 685 if ((oldflags & (VM_WRITE | VM_SHARED | VM_LOCKED)) == VM_LOCKED && 686 (newflags & VM_WRITE)) { 687 populate_vma_page_range(vma, start, end, NULL); 688 } 689 690 vm_stat_account(mm, oldflags, -nrpages); 691 vm_stat_account(mm, newflags, nrpages); 692 perf_event_mmap(vma); 693 return 0; 694 695 fail: 696 vm_unacct_memory(charged); 697 return error; 698 } 699 700 /* 701 * pkey==-1 when doing a legacy mprotect() 702 */ 703 static int do_mprotect_pkey(unsigned long start, size_t len, 704 unsigned long prot, int pkey) 705 { 706 unsigned long nstart, end, tmp, reqprot; 707 struct vm_area_struct *vma, *prev; 708 int error; 709 const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP); 710 const bool rier = (current->personality & READ_IMPLIES_EXEC) && 711 (prot & PROT_READ); 712 struct mmu_gather tlb; 713 struct vma_iterator vmi; 714 715 start = untagged_addr(start); 716 717 prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP); 718 if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */ 719 return -EINVAL; 720 721 if (start & ~PAGE_MASK) 722 return -EINVAL; 723 if (!len) 724 return 0; 725 len = PAGE_ALIGN(len); 726 end = start + len; 727 if (end <= start) 728 return -ENOMEM; 729 if (!arch_validate_prot(prot, start)) 730 return -EINVAL; 731 732 reqprot = prot; 733 734 if (mmap_write_lock_killable(current->mm)) 735 return -EINTR; 736 737 /* 738 * If userspace did not allocate the pkey, do not let 739 * them use it here. 740 */ 741 error = -EINVAL; 742 if ((pkey != -1) && !mm_pkey_is_allocated(current->mm, pkey)) 743 goto out; 744 745 vma_iter_init(&vmi, current->mm, start); 746 vma = vma_find(&vmi, end); 747 error = -ENOMEM; 748 if (!vma) 749 goto out; 750 751 if (unlikely(grows & PROT_GROWSDOWN)) { 752 if (vma->vm_start >= end) 753 goto out; 754 start = vma->vm_start; 755 error = -EINVAL; 756 if (!(vma->vm_flags & VM_GROWSDOWN)) 757 goto out; 758 } else { 759 if (vma->vm_start > start) 760 goto out; 761 if (unlikely(grows & PROT_GROWSUP)) { 762 end = vma->vm_end; 763 error = -EINVAL; 764 if (!(vma->vm_flags & VM_GROWSUP)) 765 goto out; 766 } 767 } 768 769 prev = vma_prev(&vmi); 770 if (start > vma->vm_start) 771 prev = vma; 772 773 tlb_gather_mmu(&tlb, current->mm); 774 nstart = start; 775 tmp = vma->vm_start; 776 for_each_vma_range(vmi, vma, end) { 777 unsigned long mask_off_old_flags; 778 unsigned long newflags; 779 int new_vma_pkey; 780 781 if (vma->vm_start != tmp) { 782 error = -ENOMEM; 783 break; 784 } 785 786 /* Does the application expect PROT_READ to imply PROT_EXEC */ 787 if (rier && (vma->vm_flags & VM_MAYEXEC)) 788 prot |= PROT_EXEC; 789 790 /* 791 * Each mprotect() call explicitly passes r/w/x permissions. 792 * If a permission is not passed to mprotect(), it must be 793 * cleared from the VMA. 794 */ 795 mask_off_old_flags = VM_ACCESS_FLAGS | VM_FLAGS_CLEAR; 796 797 new_vma_pkey = arch_override_mprotect_pkey(vma, prot, pkey); 798 newflags = calc_vm_prot_bits(prot, new_vma_pkey); 799 newflags |= (vma->vm_flags & ~mask_off_old_flags); 800 801 /* newflags >> 4 shift VM_MAY% in place of VM_% */ 802 if ((newflags & ~(newflags >> 4)) & VM_ACCESS_FLAGS) { 803 error = -EACCES; 804 break; 805 } 806 807 if (map_deny_write_exec(vma->vm_flags, newflags)) { 808 error = -EACCES; 809 break; 810 } 811 812 /* Allow architectures to sanity-check the new flags */ 813 if (!arch_validate_flags(newflags)) { 814 error = -EINVAL; 815 break; 816 } 817 818 error = security_file_mprotect(vma, reqprot, prot); 819 if (error) 820 break; 821 822 tmp = vma->vm_end; 823 if (tmp > end) 824 tmp = end; 825 826 if (vma->vm_ops && vma->vm_ops->mprotect) { 827 error = vma->vm_ops->mprotect(vma, nstart, tmp, newflags); 828 if (error) 829 break; 830 } 831 832 error = mprotect_fixup(&vmi, &tlb, vma, &prev, nstart, tmp, newflags); 833 if (error) 834 break; 835 836 tmp = vma_iter_end(&vmi); 837 nstart = tmp; 838 prot = reqprot; 839 } 840 tlb_finish_mmu(&tlb); 841 842 if (!error && tmp < end) 843 error = -ENOMEM; 844 845 out: 846 mmap_write_unlock(current->mm); 847 return error; 848 } 849 850 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len, 851 unsigned long, prot) 852 { 853 return do_mprotect_pkey(start, len, prot, -1); 854 } 855 856 #ifdef CONFIG_ARCH_HAS_PKEYS 857 858 SYSCALL_DEFINE4(pkey_mprotect, unsigned long, start, size_t, len, 859 unsigned long, prot, int, pkey) 860 { 861 return do_mprotect_pkey(start, len, prot, pkey); 862 } 863 864 SYSCALL_DEFINE2(pkey_alloc, unsigned long, flags, unsigned long, init_val) 865 { 866 int pkey; 867 int ret; 868 869 /* No flags supported yet. */ 870 if (flags) 871 return -EINVAL; 872 /* check for unsupported init values */ 873 if (init_val & ~PKEY_ACCESS_MASK) 874 return -EINVAL; 875 876 mmap_write_lock(current->mm); 877 pkey = mm_pkey_alloc(current->mm); 878 879 ret = -ENOSPC; 880 if (pkey == -1) 881 goto out; 882 883 ret = arch_set_user_pkey_access(current, pkey, init_val); 884 if (ret) { 885 mm_pkey_free(current->mm, pkey); 886 goto out; 887 } 888 ret = pkey; 889 out: 890 mmap_write_unlock(current->mm); 891 return ret; 892 } 893 894 SYSCALL_DEFINE1(pkey_free, int, pkey) 895 { 896 int ret; 897 898 mmap_write_lock(current->mm); 899 ret = mm_pkey_free(current->mm, pkey); 900 mmap_write_unlock(current->mm); 901 902 /* 903 * We could provide warnings or errors if any VMA still 904 * has the pkey set here. 905 */ 906 return ret; 907 } 908 909 #endif /* CONFIG_ARCH_HAS_PKEYS */ 910