1 // SPDX-License-Identifier: GPL-2.0 2 #include <linux/pagewalk.h> 3 #include <linux/highmem.h> 4 #include <linux/sched.h> 5 #include <linux/hugetlb.h> 6 #include <linux/mmu_context.h> 7 #include <linux/swap.h> 8 #include <linux/swapops.h> 9 10 #include <asm/tlbflush.h> 11 12 #include "internal.h" 13 14 /* 15 * We want to know the real level where a entry is located ignoring any 16 * folding of levels which may be happening. For example if p4d is folded then 17 * a missing entry found at level 1 (p4d) is actually at level 0 (pgd). 18 */ 19 static int real_depth(int depth) 20 { 21 if (depth == 3 && PTRS_PER_PMD == 1) 22 depth = 2; 23 if (depth == 2 && PTRS_PER_PUD == 1) 24 depth = 1; 25 if (depth == 1 && PTRS_PER_P4D == 1) 26 depth = 0; 27 return depth; 28 } 29 30 static int walk_pte_range_inner(pte_t *pte, unsigned long addr, 31 unsigned long end, struct mm_walk *walk) 32 { 33 const struct mm_walk_ops *ops = walk->ops; 34 int err = 0; 35 36 for (;;) { 37 if (ops->install_pte && pte_none(ptep_get(pte))) { 38 pte_t new_pte; 39 40 err = ops->install_pte(addr, addr + PAGE_SIZE, &new_pte, 41 walk); 42 if (err) 43 break; 44 45 set_pte_at(walk->mm, addr, pte, new_pte); 46 /* Non-present before, so for arches that need it. */ 47 if (!WARN_ON_ONCE(walk->no_vma)) 48 update_mmu_cache(walk->vma, addr, pte); 49 } else { 50 err = ops->pte_entry(pte, addr, addr + PAGE_SIZE, walk); 51 if (err) 52 break; 53 } 54 if (addr >= end - PAGE_SIZE) 55 break; 56 addr += PAGE_SIZE; 57 pte++; 58 } 59 return err; 60 } 61 62 static int walk_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end, 63 struct mm_walk *walk) 64 { 65 pte_t *pte; 66 int err = 0; 67 spinlock_t *ptl; 68 69 if (walk->no_vma) { 70 /* 71 * pte_offset_map() might apply user-specific validation. 72 * Indeed, on x86_64 the pmd entries set up by init_espfix_ap() 73 * fit its pmd_bad() check (_PAGE_NX set and _PAGE_RW clear), 74 * and CONFIG_EFI_PGT_DUMP efi_mm goes so far as to walk them. 75 */ 76 if (walk->mm == &init_mm || addr >= TASK_SIZE) 77 pte = pte_offset_kernel(pmd, addr); 78 else 79 pte = pte_offset_map(pmd, addr); 80 if (pte) { 81 err = walk_pte_range_inner(pte, addr, end, walk); 82 if (walk->mm != &init_mm && addr < TASK_SIZE) 83 pte_unmap(pte); 84 } 85 } else { 86 pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl); 87 if (pte) { 88 err = walk_pte_range_inner(pte, addr, end, walk); 89 pte_unmap_unlock(pte, ptl); 90 } 91 } 92 if (!pte) 93 walk->action = ACTION_AGAIN; 94 return err; 95 } 96 97 static int walk_pmd_range(pud_t *pud, unsigned long addr, unsigned long end, 98 struct mm_walk *walk) 99 { 100 pmd_t *pmd; 101 unsigned long next; 102 const struct mm_walk_ops *ops = walk->ops; 103 bool has_handler = ops->pte_entry; 104 bool has_install = ops->install_pte; 105 int err = 0; 106 int depth = real_depth(3); 107 108 pmd = pmd_offset(pud, addr); 109 do { 110 again: 111 next = pmd_addr_end(addr, end); 112 if (pmd_none(*pmd)) { 113 if (has_install) 114 err = __pte_alloc(walk->mm, pmd); 115 else if (ops->pte_hole) 116 err = ops->pte_hole(addr, next, depth, walk); 117 if (err) 118 break; 119 if (!has_install) 120 continue; 121 } 122 123 walk->action = ACTION_SUBTREE; 124 125 /* 126 * This implies that each ->pmd_entry() handler 127 * needs to know about pmd_trans_huge() pmds 128 */ 129 if (ops->pmd_entry) 130 err = ops->pmd_entry(pmd, addr, next, walk); 131 if (err) 132 break; 133 134 if (walk->action == ACTION_AGAIN) 135 goto again; 136 if (walk->action == ACTION_CONTINUE) 137 continue; 138 139 if (!has_handler) { /* No handlers for lower page tables. */ 140 if (!has_install) 141 continue; /* Nothing to do. */ 142 /* 143 * We are ONLY installing, so avoid unnecessarily 144 * splitting a present huge page. 145 */ 146 if (pmd_present(*pmd) && 147 (pmd_trans_huge(*pmd) || pmd_devmap(*pmd))) 148 continue; 149 } 150 151 if (walk->vma) 152 split_huge_pmd(walk->vma, pmd, addr); 153 else if (pmd_leaf(*pmd) || !pmd_present(*pmd)) 154 continue; /* Nothing to do. */ 155 156 err = walk_pte_range(pmd, addr, next, walk); 157 if (err) 158 break; 159 160 if (walk->action == ACTION_AGAIN) 161 goto again; 162 163 } while (pmd++, addr = next, addr != end); 164 165 return err; 166 } 167 168 static int walk_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end, 169 struct mm_walk *walk) 170 { 171 pud_t *pud; 172 unsigned long next; 173 const struct mm_walk_ops *ops = walk->ops; 174 bool has_handler = ops->pmd_entry || ops->pte_entry; 175 bool has_install = ops->install_pte; 176 int err = 0; 177 int depth = real_depth(2); 178 179 pud = pud_offset(p4d, addr); 180 do { 181 again: 182 next = pud_addr_end(addr, end); 183 if (pud_none(*pud)) { 184 if (has_install) 185 err = __pmd_alloc(walk->mm, pud, addr); 186 else if (ops->pte_hole) 187 err = ops->pte_hole(addr, next, depth, walk); 188 if (err) 189 break; 190 if (!has_install) 191 continue; 192 } 193 194 walk->action = ACTION_SUBTREE; 195 196 if (ops->pud_entry) 197 err = ops->pud_entry(pud, addr, next, walk); 198 if (err) 199 break; 200 201 if (walk->action == ACTION_AGAIN) 202 goto again; 203 if (walk->action == ACTION_CONTINUE) 204 continue; 205 206 if (!has_handler) { /* No handlers for lower page tables. */ 207 if (!has_install) 208 continue; /* Nothing to do. */ 209 /* 210 * We are ONLY installing, so avoid unnecessarily 211 * splitting a present huge page. 212 */ 213 if (pud_present(*pud) && 214 (pud_trans_huge(*pud) || pud_devmap(*pud))) 215 continue; 216 } 217 218 if (walk->vma) 219 split_huge_pud(walk->vma, pud, addr); 220 else if (pud_leaf(*pud) || !pud_present(*pud)) 221 continue; /* Nothing to do. */ 222 223 if (pud_none(*pud)) 224 goto again; 225 226 err = walk_pmd_range(pud, addr, next, walk); 227 if (err) 228 break; 229 } while (pud++, addr = next, addr != end); 230 231 return err; 232 } 233 234 static int walk_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end, 235 struct mm_walk *walk) 236 { 237 p4d_t *p4d; 238 unsigned long next; 239 const struct mm_walk_ops *ops = walk->ops; 240 bool has_handler = ops->pud_entry || ops->pmd_entry || ops->pte_entry; 241 bool has_install = ops->install_pte; 242 int err = 0; 243 int depth = real_depth(1); 244 245 p4d = p4d_offset(pgd, addr); 246 do { 247 next = p4d_addr_end(addr, end); 248 if (p4d_none_or_clear_bad(p4d)) { 249 if (has_install) 250 err = __pud_alloc(walk->mm, p4d, addr); 251 else if (ops->pte_hole) 252 err = ops->pte_hole(addr, next, depth, walk); 253 if (err) 254 break; 255 if (!has_install) 256 continue; 257 } 258 if (ops->p4d_entry) { 259 err = ops->p4d_entry(p4d, addr, next, walk); 260 if (err) 261 break; 262 } 263 if (has_handler || has_install) 264 err = walk_pud_range(p4d, addr, next, walk); 265 if (err) 266 break; 267 } while (p4d++, addr = next, addr != end); 268 269 return err; 270 } 271 272 static int walk_pgd_range(unsigned long addr, unsigned long end, 273 struct mm_walk *walk) 274 { 275 pgd_t *pgd; 276 unsigned long next; 277 const struct mm_walk_ops *ops = walk->ops; 278 bool has_handler = ops->p4d_entry || ops->pud_entry || ops->pmd_entry || 279 ops->pte_entry; 280 bool has_install = ops->install_pte; 281 int err = 0; 282 283 if (walk->pgd) 284 pgd = walk->pgd + pgd_index(addr); 285 else 286 pgd = pgd_offset(walk->mm, addr); 287 do { 288 next = pgd_addr_end(addr, end); 289 if (pgd_none_or_clear_bad(pgd)) { 290 if (has_install) 291 err = __p4d_alloc(walk->mm, pgd, addr); 292 else if (ops->pte_hole) 293 err = ops->pte_hole(addr, next, 0, walk); 294 if (err) 295 break; 296 if (!has_install) 297 continue; 298 } 299 if (ops->pgd_entry) { 300 err = ops->pgd_entry(pgd, addr, next, walk); 301 if (err) 302 break; 303 } 304 if (has_handler || has_install) 305 err = walk_p4d_range(pgd, addr, next, walk); 306 if (err) 307 break; 308 } while (pgd++, addr = next, addr != end); 309 310 return err; 311 } 312 313 #ifdef CONFIG_HUGETLB_PAGE 314 static unsigned long hugetlb_entry_end(struct hstate *h, unsigned long addr, 315 unsigned long end) 316 { 317 unsigned long boundary = (addr & huge_page_mask(h)) + huge_page_size(h); 318 return boundary < end ? boundary : end; 319 } 320 321 static int walk_hugetlb_range(unsigned long addr, unsigned long end, 322 struct mm_walk *walk) 323 { 324 struct vm_area_struct *vma = walk->vma; 325 struct hstate *h = hstate_vma(vma); 326 unsigned long next; 327 unsigned long hmask = huge_page_mask(h); 328 unsigned long sz = huge_page_size(h); 329 pte_t *pte; 330 const struct mm_walk_ops *ops = walk->ops; 331 int err = 0; 332 333 hugetlb_vma_lock_read(vma); 334 do { 335 next = hugetlb_entry_end(h, addr, end); 336 pte = hugetlb_walk(vma, addr & hmask, sz); 337 if (pte) 338 err = ops->hugetlb_entry(pte, hmask, addr, next, walk); 339 else if (ops->pte_hole) 340 err = ops->pte_hole(addr, next, -1, walk); 341 if (err) 342 break; 343 } while (addr = next, addr != end); 344 hugetlb_vma_unlock_read(vma); 345 346 return err; 347 } 348 349 #else /* CONFIG_HUGETLB_PAGE */ 350 static int walk_hugetlb_range(unsigned long addr, unsigned long end, 351 struct mm_walk *walk) 352 { 353 return 0; 354 } 355 356 #endif /* CONFIG_HUGETLB_PAGE */ 357 358 /* 359 * Decide whether we really walk over the current vma on [@start, @end) 360 * or skip it via the returned value. Return 0 if we do walk over the 361 * current vma, and return 1 if we skip the vma. Negative values means 362 * error, where we abort the current walk. 363 */ 364 static int walk_page_test(unsigned long start, unsigned long end, 365 struct mm_walk *walk) 366 { 367 struct vm_area_struct *vma = walk->vma; 368 const struct mm_walk_ops *ops = walk->ops; 369 370 if (ops->test_walk) 371 return ops->test_walk(start, end, walk); 372 373 /* 374 * vma(VM_PFNMAP) doesn't have any valid struct pages behind VM_PFNMAP 375 * range, so we don't walk over it as we do for normal vmas. However, 376 * Some callers are interested in handling hole range and they don't 377 * want to just ignore any single address range. Such users certainly 378 * define their ->pte_hole() callbacks, so let's delegate them to handle 379 * vma(VM_PFNMAP). 380 */ 381 if (vma->vm_flags & VM_PFNMAP) { 382 int err = 1; 383 if (ops->pte_hole) 384 err = ops->pte_hole(start, end, -1, walk); 385 return err ? err : 1; 386 } 387 return 0; 388 } 389 390 static int __walk_page_range(unsigned long start, unsigned long end, 391 struct mm_walk *walk) 392 { 393 int err = 0; 394 struct vm_area_struct *vma = walk->vma; 395 const struct mm_walk_ops *ops = walk->ops; 396 bool is_hugetlb = is_vm_hugetlb_page(vma); 397 398 /* We do not support hugetlb PTE installation. */ 399 if (ops->install_pte && is_hugetlb) 400 return -EINVAL; 401 402 if (ops->pre_vma) { 403 err = ops->pre_vma(start, end, walk); 404 if (err) 405 return err; 406 } 407 408 if (is_hugetlb) { 409 if (ops->hugetlb_entry) 410 err = walk_hugetlb_range(start, end, walk); 411 } else 412 err = walk_pgd_range(start, end, walk); 413 414 if (ops->post_vma) 415 ops->post_vma(walk); 416 417 return err; 418 } 419 420 static inline void process_mm_walk_lock(struct mm_struct *mm, 421 enum page_walk_lock walk_lock) 422 { 423 if (walk_lock == PGWALK_RDLOCK) 424 mmap_assert_locked(mm); 425 else if (walk_lock != PGWALK_VMA_RDLOCK_VERIFY) 426 mmap_assert_write_locked(mm); 427 } 428 429 static inline void process_vma_walk_lock(struct vm_area_struct *vma, 430 enum page_walk_lock walk_lock) 431 { 432 #ifdef CONFIG_PER_VMA_LOCK 433 switch (walk_lock) { 434 case PGWALK_WRLOCK: 435 vma_start_write(vma); 436 break; 437 case PGWALK_WRLOCK_VERIFY: 438 vma_assert_write_locked(vma); 439 break; 440 case PGWALK_VMA_RDLOCK_VERIFY: 441 vma_assert_locked(vma); 442 break; 443 case PGWALK_RDLOCK: 444 /* PGWALK_RDLOCK is handled by process_mm_walk_lock */ 445 break; 446 } 447 #endif 448 } 449 450 /* 451 * See the comment for walk_page_range(), this performs the heavy lifting of the 452 * operation, only sets no restrictions on how the walk proceeds. 453 * 454 * We usually restrict the ability to install PTEs, but this functionality is 455 * available to internal memory management code and provided in mm/internal.h. 456 */ 457 int walk_page_range_mm(struct mm_struct *mm, unsigned long start, 458 unsigned long end, const struct mm_walk_ops *ops, 459 void *private) 460 { 461 int err = 0; 462 unsigned long next; 463 struct vm_area_struct *vma; 464 struct mm_walk walk = { 465 .ops = ops, 466 .mm = mm, 467 .private = private, 468 }; 469 470 if (start >= end) 471 return -EINVAL; 472 473 if (!walk.mm) 474 return -EINVAL; 475 476 process_mm_walk_lock(walk.mm, ops->walk_lock); 477 478 vma = find_vma(walk.mm, start); 479 do { 480 if (!vma) { /* after the last vma */ 481 walk.vma = NULL; 482 next = end; 483 if (ops->pte_hole) 484 err = ops->pte_hole(start, next, -1, &walk); 485 } else if (start < vma->vm_start) { /* outside vma */ 486 walk.vma = NULL; 487 next = min(end, vma->vm_start); 488 if (ops->pte_hole) 489 err = ops->pte_hole(start, next, -1, &walk); 490 } else { /* inside vma */ 491 process_vma_walk_lock(vma, ops->walk_lock); 492 walk.vma = vma; 493 next = min(end, vma->vm_end); 494 vma = find_vma(mm, vma->vm_end); 495 496 err = walk_page_test(start, next, &walk); 497 if (err > 0) { 498 /* 499 * positive return values are purely for 500 * controlling the pagewalk, so should never 501 * be passed to the callers. 502 */ 503 err = 0; 504 continue; 505 } 506 if (err < 0) 507 break; 508 err = __walk_page_range(start, next, &walk); 509 } 510 if (err) 511 break; 512 } while (start = next, start < end); 513 return err; 514 } 515 516 /* 517 * Determine if the walk operations specified are permitted to be used for a 518 * page table walk. 519 * 520 * This check is performed on all functions which are parameterised by walk 521 * operations and exposed in include/linux/pagewalk.h. 522 * 523 * Internal memory management code can use the walk_page_range_mm() function to 524 * be able to use all page walking operations. 525 */ 526 static bool check_ops_valid(const struct mm_walk_ops *ops) 527 { 528 /* 529 * The installation of PTEs is solely under the control of memory 530 * management logic and subject to many subtle locking, security and 531 * cache considerations so we cannot permit other users to do so, and 532 * certainly not for exported symbols. 533 */ 534 if (ops->install_pte) 535 return false; 536 537 return true; 538 } 539 540 /** 541 * walk_page_range - walk page table with caller specific callbacks 542 * @mm: mm_struct representing the target process of page table walk 543 * @start: start address of the virtual address range 544 * @end: end address of the virtual address range 545 * @ops: operation to call during the walk 546 * @private: private data for callbacks' usage 547 * 548 * Recursively walk the page table tree of the process represented by @mm 549 * within the virtual address range [@start, @end). During walking, we can do 550 * some caller-specific works for each entry, by setting up pmd_entry(), 551 * pte_entry(), and/or hugetlb_entry(). If you don't set up for some of these 552 * callbacks, the associated entries/pages are just ignored. 553 * The return values of these callbacks are commonly defined like below: 554 * 555 * - 0 : succeeded to handle the current entry, and if you don't reach the 556 * end address yet, continue to walk. 557 * - >0 : succeeded to handle the current entry, and return to the caller 558 * with caller specific value. 559 * - <0 : failed to handle the current entry, and return to the caller 560 * with error code. 561 * 562 * Before starting to walk page table, some callers want to check whether 563 * they really want to walk over the current vma, typically by checking 564 * its vm_flags. walk_page_test() and @ops->test_walk() are used for this 565 * purpose. 566 * 567 * If operations need to be staged before and committed after a vma is walked, 568 * there are two callbacks, pre_vma() and post_vma(). Note that post_vma(), 569 * since it is intended to handle commit-type operations, can't return any 570 * errors. 571 * 572 * struct mm_walk keeps current values of some common data like vma and pmd, 573 * which are useful for the access from callbacks. If you want to pass some 574 * caller-specific data to callbacks, @private should be helpful. 575 * 576 * Locking: 577 * Callers of walk_page_range() and walk_page_vma() should hold @mm->mmap_lock, 578 * because these function traverse vma list and/or access to vma's data. 579 */ 580 int walk_page_range(struct mm_struct *mm, unsigned long start, 581 unsigned long end, const struct mm_walk_ops *ops, 582 void *private) 583 { 584 if (!check_ops_valid(ops)) 585 return -EINVAL; 586 587 return walk_page_range_mm(mm, start, end, ops, private); 588 } 589 590 /** 591 * walk_kernel_page_table_range - walk a range of kernel pagetables. 592 * @start: start address of the virtual address range 593 * @end: end address of the virtual address range 594 * @ops: operation to call during the walk 595 * @pgd: pgd to walk if different from mm->pgd 596 * @private: private data for callbacks' usage 597 * 598 * Similar to walk_page_range() but can walk any page tables even if they are 599 * not backed by VMAs. Because 'unusual' entries may be walked this function 600 * will also not lock the PTEs for the pte_entry() callback. This is useful for 601 * walking kernel pages tables or page tables for firmware. 602 * 603 * Note: Be careful to walk the kernel pages tables, the caller may be need to 604 * take other effective approaches (mmap lock may be insufficient) to prevent 605 * the intermediate kernel page tables belonging to the specified address range 606 * from being freed (e.g. memory hot-remove). 607 */ 608 int walk_kernel_page_table_range(unsigned long start, unsigned long end, 609 const struct mm_walk_ops *ops, pgd_t *pgd, void *private) 610 { 611 struct mm_struct *mm = &init_mm; 612 struct mm_walk walk = { 613 .ops = ops, 614 .mm = mm, 615 .pgd = pgd, 616 .private = private, 617 .no_vma = true 618 }; 619 620 if (start >= end) 621 return -EINVAL; 622 if (!check_ops_valid(ops)) 623 return -EINVAL; 624 625 /* 626 * Kernel intermediate page tables are usually not freed, so the mmap 627 * read lock is sufficient. But there are some exceptions. 628 * E.g. memory hot-remove. In which case, the mmap lock is insufficient 629 * to prevent the intermediate kernel pages tables belonging to the 630 * specified address range from being freed. The caller should take 631 * other actions to prevent this race. 632 */ 633 mmap_assert_locked(mm); 634 635 return walk_pgd_range(start, end, &walk); 636 } 637 638 /** 639 * walk_page_range_debug - walk a range of pagetables not backed by a vma 640 * @mm: mm_struct representing the target process of page table walk 641 * @start: start address of the virtual address range 642 * @end: end address of the virtual address range 643 * @ops: operation to call during the walk 644 * @pgd: pgd to walk if different from mm->pgd 645 * @private: private data for callbacks' usage 646 * 647 * Similar to walk_page_range() but can walk any page tables even if they are 648 * not backed by VMAs. Because 'unusual' entries may be walked this function 649 * will also not lock the PTEs for the pte_entry() callback. 650 * 651 * This is for debugging purposes ONLY. 652 */ 653 int walk_page_range_debug(struct mm_struct *mm, unsigned long start, 654 unsigned long end, const struct mm_walk_ops *ops, 655 pgd_t *pgd, void *private) 656 { 657 struct mm_walk walk = { 658 .ops = ops, 659 .mm = mm, 660 .pgd = pgd, 661 .private = private, 662 .no_vma = true 663 }; 664 665 /* For convenience, we allow traversal of kernel mappings. */ 666 if (mm == &init_mm) 667 return walk_kernel_page_table_range(start, end, ops, 668 pgd, private); 669 if (start >= end || !walk.mm) 670 return -EINVAL; 671 if (!check_ops_valid(ops)) 672 return -EINVAL; 673 674 /* 675 * The mmap lock protects the page walker from changes to the page 676 * tables during the walk. However a read lock is insufficient to 677 * protect those areas which don't have a VMA as munmap() detaches 678 * the VMAs before downgrading to a read lock and actually tearing 679 * down PTEs/page tables. In which case, the mmap write lock should 680 * be held. 681 */ 682 mmap_assert_write_locked(mm); 683 684 return walk_pgd_range(start, end, &walk); 685 } 686 687 int walk_page_range_vma(struct vm_area_struct *vma, unsigned long start, 688 unsigned long end, const struct mm_walk_ops *ops, 689 void *private) 690 { 691 struct mm_walk walk = { 692 .ops = ops, 693 .mm = vma->vm_mm, 694 .vma = vma, 695 .private = private, 696 }; 697 698 if (start >= end || !walk.mm) 699 return -EINVAL; 700 if (start < vma->vm_start || end > vma->vm_end) 701 return -EINVAL; 702 if (!check_ops_valid(ops)) 703 return -EINVAL; 704 705 process_mm_walk_lock(walk.mm, ops->walk_lock); 706 process_vma_walk_lock(vma, ops->walk_lock); 707 return __walk_page_range(start, end, &walk); 708 } 709 710 int walk_page_vma(struct vm_area_struct *vma, const struct mm_walk_ops *ops, 711 void *private) 712 { 713 struct mm_walk walk = { 714 .ops = ops, 715 .mm = vma->vm_mm, 716 .vma = vma, 717 .private = private, 718 }; 719 720 if (!walk.mm) 721 return -EINVAL; 722 if (!check_ops_valid(ops)) 723 return -EINVAL; 724 725 process_mm_walk_lock(walk.mm, ops->walk_lock); 726 process_vma_walk_lock(vma, ops->walk_lock); 727 return __walk_page_range(vma->vm_start, vma->vm_end, &walk); 728 } 729 730 /** 731 * walk_page_mapping - walk all memory areas mapped into a struct address_space. 732 * @mapping: Pointer to the struct address_space 733 * @first_index: First page offset in the address_space 734 * @nr: Number of incremental page offsets to cover 735 * @ops: operation to call during the walk 736 * @private: private data for callbacks' usage 737 * 738 * This function walks all memory areas mapped into a struct address_space. 739 * The walk is limited to only the given page-size index range, but if 740 * the index boundaries cross a huge page-table entry, that entry will be 741 * included. 742 * 743 * Also see walk_page_range() for additional information. 744 * 745 * Locking: 746 * This function can't require that the struct mm_struct::mmap_lock is held, 747 * since @mapping may be mapped by multiple processes. Instead 748 * @mapping->i_mmap_rwsem must be held. This might have implications in the 749 * callbacks, and it's up tho the caller to ensure that the 750 * struct mm_struct::mmap_lock is not needed. 751 * 752 * Also this means that a caller can't rely on the struct 753 * vm_area_struct::vm_flags to be constant across a call, 754 * except for immutable flags. Callers requiring this shouldn't use 755 * this function. 756 * 757 * Return: 0 on success, negative error code on failure, positive number on 758 * caller defined premature termination. 759 */ 760 int walk_page_mapping(struct address_space *mapping, pgoff_t first_index, 761 pgoff_t nr, const struct mm_walk_ops *ops, 762 void *private) 763 { 764 struct mm_walk walk = { 765 .ops = ops, 766 .private = private, 767 }; 768 struct vm_area_struct *vma; 769 pgoff_t vba, vea, cba, cea; 770 unsigned long start_addr, end_addr; 771 int err = 0; 772 773 if (!check_ops_valid(ops)) 774 return -EINVAL; 775 776 lockdep_assert_held(&mapping->i_mmap_rwsem); 777 vma_interval_tree_foreach(vma, &mapping->i_mmap, first_index, 778 first_index + nr - 1) { 779 /* Clip to the vma */ 780 vba = vma->vm_pgoff; 781 vea = vba + vma_pages(vma); 782 cba = first_index; 783 cba = max(cba, vba); 784 cea = first_index + nr; 785 cea = min(cea, vea); 786 787 start_addr = ((cba - vba) << PAGE_SHIFT) + vma->vm_start; 788 end_addr = ((cea - vba) << PAGE_SHIFT) + vma->vm_start; 789 if (start_addr >= end_addr) 790 continue; 791 792 walk.vma = vma; 793 walk.mm = vma->vm_mm; 794 795 err = walk_page_test(vma->vm_start, vma->vm_end, &walk); 796 if (err > 0) { 797 err = 0; 798 break; 799 } else if (err < 0) 800 break; 801 802 err = __walk_page_range(start_addr, end_addr, &walk); 803 if (err) 804 break; 805 } 806 807 return err; 808 } 809 810 /** 811 * folio_walk_start - walk the page tables to a folio 812 * @fw: filled with information on success. 813 * @vma: the VMA. 814 * @addr: the virtual address to use for the page table walk. 815 * @flags: flags modifying which folios to walk to. 816 * 817 * Walk the page tables using @addr in a given @vma to a mapped folio and 818 * return the folio, making sure that the page table entry referenced by 819 * @addr cannot change until folio_walk_end() was called. 820 * 821 * As default, this function returns only folios that are not special (e.g., not 822 * the zeropage) and never returns folios that are supposed to be ignored by the 823 * VM as documented by vm_normal_page(). If requested, zeropages will be 824 * returned as well. 825 * 826 * As default, this function only considers present page table entries. 827 * If requested, it will also consider migration entries. 828 * 829 * If this function returns NULL it might either indicate "there is nothing" or 830 * "there is nothing suitable". 831 * 832 * On success, @fw is filled and the function returns the folio while the PTL 833 * is still held and folio_walk_end() must be called to clean up, 834 * releasing any held locks. The returned folio must *not* be used after the 835 * call to folio_walk_end(), unless a short-term folio reference is taken before 836 * that call. 837 * 838 * @fw->page will correspond to the page that is effectively referenced by 839 * @addr. However, for migration entries and shared zeropages @fw->page is 840 * set to NULL. Note that large folios might be mapped by multiple page table 841 * entries, and this function will always only lookup a single entry as 842 * specified by @addr, which might or might not cover more than a single page of 843 * the returned folio. 844 * 845 * This function must *not* be used as a naive replacement for 846 * get_user_pages() / pin_user_pages(), especially not to perform DMA or 847 * to carelessly modify page content. This function may *only* be used to grab 848 * short-term folio references, never to grab long-term folio references. 849 * 850 * Using the page table entry pointers in @fw for reading or modifying the 851 * entry should be avoided where possible: however, there might be valid 852 * use cases. 853 * 854 * WARNING: Modifying page table entries in hugetlb VMAs requires a lot of care. 855 * For example, PMD page table sharing might require prior unsharing. Also, 856 * logical hugetlb entries might span multiple physical page table entries, 857 * which *must* be modified in a single operation (set_huge_pte_at(), 858 * huge_ptep_set_*, ...). Note that the page table entry stored in @fw might 859 * not correspond to the first physical entry of a logical hugetlb entry. 860 * 861 * The mmap lock must be held in read mode. 862 * 863 * Return: folio pointer on success, otherwise NULL. 864 */ 865 struct folio *folio_walk_start(struct folio_walk *fw, 866 struct vm_area_struct *vma, unsigned long addr, 867 folio_walk_flags_t flags) 868 { 869 unsigned long entry_size; 870 bool expose_page = true; 871 struct page *page; 872 pud_t *pudp, pud; 873 pmd_t *pmdp, pmd; 874 pte_t *ptep, pte; 875 spinlock_t *ptl; 876 pgd_t *pgdp; 877 p4d_t *p4dp; 878 879 mmap_assert_locked(vma->vm_mm); 880 vma_pgtable_walk_begin(vma); 881 882 if (WARN_ON_ONCE(addr < vma->vm_start || addr >= vma->vm_end)) 883 goto not_found; 884 885 pgdp = pgd_offset(vma->vm_mm, addr); 886 if (pgd_none_or_clear_bad(pgdp)) 887 goto not_found; 888 889 p4dp = p4d_offset(pgdp, addr); 890 if (p4d_none_or_clear_bad(p4dp)) 891 goto not_found; 892 893 pudp = pud_offset(p4dp, addr); 894 pud = pudp_get(pudp); 895 if (pud_none(pud)) 896 goto not_found; 897 if (IS_ENABLED(CONFIG_PGTABLE_HAS_HUGE_LEAVES) && 898 (!pud_present(pud) || pud_leaf(pud))) { 899 ptl = pud_lock(vma->vm_mm, pudp); 900 pud = pudp_get(pudp); 901 902 entry_size = PUD_SIZE; 903 fw->level = FW_LEVEL_PUD; 904 fw->pudp = pudp; 905 fw->pud = pud; 906 907 /* 908 * TODO: FW_MIGRATION support for PUD migration entries 909 * once there are relevant users. 910 */ 911 if (!pud_present(pud) || pud_devmap(pud) || pud_special(pud)) { 912 spin_unlock(ptl); 913 goto not_found; 914 } else if (!pud_leaf(pud)) { 915 spin_unlock(ptl); 916 goto pmd_table; 917 } 918 /* 919 * TODO: vm_normal_page_pud() will be handy once we want to 920 * support PUD mappings in VM_PFNMAP|VM_MIXEDMAP VMAs. 921 */ 922 page = pud_page(pud); 923 goto found; 924 } 925 926 pmd_table: 927 VM_WARN_ON_ONCE(!pud_present(pud) || pud_leaf(pud)); 928 pmdp = pmd_offset(pudp, addr); 929 pmd = pmdp_get_lockless(pmdp); 930 if (pmd_none(pmd)) 931 goto not_found; 932 if (IS_ENABLED(CONFIG_PGTABLE_HAS_HUGE_LEAVES) && 933 (!pmd_present(pmd) || pmd_leaf(pmd))) { 934 ptl = pmd_lock(vma->vm_mm, pmdp); 935 pmd = pmdp_get(pmdp); 936 937 entry_size = PMD_SIZE; 938 fw->level = FW_LEVEL_PMD; 939 fw->pmdp = pmdp; 940 fw->pmd = pmd; 941 942 if (pmd_none(pmd)) { 943 spin_unlock(ptl); 944 goto not_found; 945 } else if (pmd_present(pmd) && !pmd_leaf(pmd)) { 946 spin_unlock(ptl); 947 goto pte_table; 948 } else if (pmd_present(pmd)) { 949 page = vm_normal_page_pmd(vma, addr, pmd); 950 if (page) { 951 goto found; 952 } else if ((flags & FW_ZEROPAGE) && 953 is_huge_zero_pmd(pmd)) { 954 page = pfn_to_page(pmd_pfn(pmd)); 955 expose_page = false; 956 goto found; 957 } 958 } else if ((flags & FW_MIGRATION) && 959 is_pmd_migration_entry(pmd)) { 960 swp_entry_t entry = pmd_to_swp_entry(pmd); 961 962 page = pfn_swap_entry_to_page(entry); 963 expose_page = false; 964 goto found; 965 } 966 spin_unlock(ptl); 967 goto not_found; 968 } 969 970 pte_table: 971 VM_WARN_ON_ONCE(!pmd_present(pmd) || pmd_leaf(pmd)); 972 ptep = pte_offset_map_lock(vma->vm_mm, pmdp, addr, &ptl); 973 if (!ptep) 974 goto not_found; 975 pte = ptep_get(ptep); 976 977 entry_size = PAGE_SIZE; 978 fw->level = FW_LEVEL_PTE; 979 fw->ptep = ptep; 980 fw->pte = pte; 981 982 if (pte_present(pte)) { 983 page = vm_normal_page(vma, addr, pte); 984 if (page) 985 goto found; 986 if ((flags & FW_ZEROPAGE) && 987 is_zero_pfn(pte_pfn(pte))) { 988 page = pfn_to_page(pte_pfn(pte)); 989 expose_page = false; 990 goto found; 991 } 992 } else if (!pte_none(pte)) { 993 swp_entry_t entry = pte_to_swp_entry(pte); 994 995 if ((flags & FW_MIGRATION) && 996 is_migration_entry(entry)) { 997 page = pfn_swap_entry_to_page(entry); 998 expose_page = false; 999 goto found; 1000 } 1001 } 1002 pte_unmap_unlock(ptep, ptl); 1003 not_found: 1004 vma_pgtable_walk_end(vma); 1005 return NULL; 1006 found: 1007 if (expose_page) 1008 /* Note: Offset from the mapped page, not the folio start. */ 1009 fw->page = nth_page(page, (addr & (entry_size - 1)) >> PAGE_SHIFT); 1010 else 1011 fw->page = NULL; 1012 fw->ptl = ptl; 1013 return page_folio(page); 1014 } 1015