xref: /linux/mm/pagewalk.c (revision 7f71507851fc7764b36a3221839607d3a45c2025)
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
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_RDLOCK:
441 		/* PGWALK_RDLOCK is handled by process_mm_walk_lock */
442 		break;
443 	}
444 #endif
445 }
446 
447 /*
448  * See the comment for walk_page_range(), this performs the heavy lifting of the
449  * operation, only sets no restrictions on how the walk proceeds.
450  *
451  * We usually restrict the ability to install PTEs, but this functionality is
452  * available to internal memory management code and provided in mm/internal.h.
453  */
454 int walk_page_range_mm(struct mm_struct *mm, unsigned long start,
455 		unsigned long end, const struct mm_walk_ops *ops,
456 		void *private)
457 {
458 	int err = 0;
459 	unsigned long next;
460 	struct vm_area_struct *vma;
461 	struct mm_walk walk = {
462 		.ops		= ops,
463 		.mm		= mm,
464 		.private	= private,
465 	};
466 
467 	if (start >= end)
468 		return -EINVAL;
469 
470 	if (!walk.mm)
471 		return -EINVAL;
472 
473 	process_mm_walk_lock(walk.mm, ops->walk_lock);
474 
475 	vma = find_vma(walk.mm, start);
476 	do {
477 		if (!vma) { /* after the last vma */
478 			walk.vma = NULL;
479 			next = end;
480 			if (ops->pte_hole)
481 				err = ops->pte_hole(start, next, -1, &walk);
482 		} else if (start < vma->vm_start) { /* outside vma */
483 			walk.vma = NULL;
484 			next = min(end, vma->vm_start);
485 			if (ops->pte_hole)
486 				err = ops->pte_hole(start, next, -1, &walk);
487 		} else { /* inside vma */
488 			process_vma_walk_lock(vma, ops->walk_lock);
489 			walk.vma = vma;
490 			next = min(end, vma->vm_end);
491 			vma = find_vma(mm, vma->vm_end);
492 
493 			err = walk_page_test(start, next, &walk);
494 			if (err > 0) {
495 				/*
496 				 * positive return values are purely for
497 				 * controlling the pagewalk, so should never
498 				 * be passed to the callers.
499 				 */
500 				err = 0;
501 				continue;
502 			}
503 			if (err < 0)
504 				break;
505 			err = __walk_page_range(start, next, &walk);
506 		}
507 		if (err)
508 			break;
509 	} while (start = next, start < end);
510 	return err;
511 }
512 
513 /*
514  * Determine if the walk operations specified are permitted to be used for a
515  * page table walk.
516  *
517  * This check is performed on all functions which are parameterised by walk
518  * operations and exposed in include/linux/pagewalk.h.
519  *
520  * Internal memory management code can use the walk_page_range_mm() function to
521  * be able to use all page walking operations.
522  */
523 static bool check_ops_valid(const struct mm_walk_ops *ops)
524 {
525 	/*
526 	 * The installation of PTEs is solely under the control of memory
527 	 * management logic and subject to many subtle locking, security and
528 	 * cache considerations so we cannot permit other users to do so, and
529 	 * certainly not for exported symbols.
530 	 */
531 	if (ops->install_pte)
532 		return false;
533 
534 	return true;
535 }
536 
537 /**
538  * walk_page_range - walk page table with caller specific callbacks
539  * @mm:		mm_struct representing the target process of page table walk
540  * @start:	start address of the virtual address range
541  * @end:	end address of the virtual address range
542  * @ops:	operation to call during the walk
543  * @private:	private data for callbacks' usage
544  *
545  * Recursively walk the page table tree of the process represented by @mm
546  * within the virtual address range [@start, @end). During walking, we can do
547  * some caller-specific works for each entry, by setting up pmd_entry(),
548  * pte_entry(), and/or hugetlb_entry(). If you don't set up for some of these
549  * callbacks, the associated entries/pages are just ignored.
550  * The return values of these callbacks are commonly defined like below:
551  *
552  *  - 0  : succeeded to handle the current entry, and if you don't reach the
553  *         end address yet, continue to walk.
554  *  - >0 : succeeded to handle the current entry, and return to the caller
555  *         with caller specific value.
556  *  - <0 : failed to handle the current entry, and return to the caller
557  *         with error code.
558  *
559  * Before starting to walk page table, some callers want to check whether
560  * they really want to walk over the current vma, typically by checking
561  * its vm_flags. walk_page_test() and @ops->test_walk() are used for this
562  * purpose.
563  *
564  * If operations need to be staged before and committed after a vma is walked,
565  * there are two callbacks, pre_vma() and post_vma(). Note that post_vma(),
566  * since it is intended to handle commit-type operations, can't return any
567  * errors.
568  *
569  * struct mm_walk keeps current values of some common data like vma and pmd,
570  * which are useful for the access from callbacks. If you want to pass some
571  * caller-specific data to callbacks, @private should be helpful.
572  *
573  * Locking:
574  *   Callers of walk_page_range() and walk_page_vma() should hold @mm->mmap_lock,
575  *   because these function traverse vma list and/or access to vma's data.
576  */
577 int walk_page_range(struct mm_struct *mm, unsigned long start,
578 		unsigned long end, const struct mm_walk_ops *ops,
579 		void *private)
580 {
581 	if (!check_ops_valid(ops))
582 		return -EINVAL;
583 
584 	return walk_page_range_mm(mm, start, end, ops, private);
585 }
586 
587 /**
588  * walk_page_range_novma - walk a range of pagetables not backed by a vma
589  * @mm:		mm_struct representing the target process of page table walk
590  * @start:	start address of the virtual address range
591  * @end:	end address of the virtual address range
592  * @ops:	operation to call during the walk
593  * @pgd:	pgd to walk if different from mm->pgd
594  * @private:	private data for callbacks' usage
595  *
596  * Similar to walk_page_range() but can walk any page tables even if they are
597  * not backed by VMAs. Because 'unusual' entries may be walked this function
598  * will also not lock the PTEs for the pte_entry() callback. This is useful for
599  * walking the kernel pages tables or page tables for firmware.
600  *
601  * Note: Be careful to walk the kernel pages tables, the caller may be need to
602  * take other effective approaches (mmap lock may be insufficient) to prevent
603  * the intermediate kernel page tables belonging to the specified address range
604  * from being freed (e.g. memory hot-remove).
605  */
606 int walk_page_range_novma(struct mm_struct *mm, unsigned long start,
607 			  unsigned long end, const struct mm_walk_ops *ops,
608 			  pgd_t *pgd,
609 			  void *private)
610 {
611 	struct mm_walk walk = {
612 		.ops		= ops,
613 		.mm		= mm,
614 		.pgd		= pgd,
615 		.private	= private,
616 		.no_vma		= true
617 	};
618 
619 	if (start >= end || !walk.mm)
620 		return -EINVAL;
621 	if (!check_ops_valid(ops))
622 		return -EINVAL;
623 
624 	/*
625 	 * 1) For walking the user virtual address space:
626 	 *
627 	 * The mmap lock protects the page walker from changes to the page
628 	 * tables during the walk.  However a read lock is insufficient to
629 	 * protect those areas which don't have a VMA as munmap() detaches
630 	 * the VMAs before downgrading to a read lock and actually tearing
631 	 * down PTEs/page tables. In which case, the mmap write lock should
632 	 * be hold.
633 	 *
634 	 * 2) For walking the kernel virtual address space:
635 	 *
636 	 * The kernel intermediate page tables usually do not be freed, so
637 	 * the mmap map read lock is sufficient. But there are some exceptions.
638 	 * E.g. memory hot-remove. In which case, the mmap lock is insufficient
639 	 * to prevent the intermediate kernel pages tables belonging to the
640 	 * specified address range from being freed. The caller should take
641 	 * other actions to prevent this race.
642 	 */
643 	if (mm == &init_mm)
644 		mmap_assert_locked(walk.mm);
645 	else
646 		mmap_assert_write_locked(walk.mm);
647 
648 	return walk_pgd_range(start, end, &walk);
649 }
650 
651 int walk_page_range_vma(struct vm_area_struct *vma, unsigned long start,
652 			unsigned long end, const struct mm_walk_ops *ops,
653 			void *private)
654 {
655 	struct mm_walk walk = {
656 		.ops		= ops,
657 		.mm		= vma->vm_mm,
658 		.vma		= vma,
659 		.private	= private,
660 	};
661 
662 	if (start >= end || !walk.mm)
663 		return -EINVAL;
664 	if (start < vma->vm_start || end > vma->vm_end)
665 		return -EINVAL;
666 	if (!check_ops_valid(ops))
667 		return -EINVAL;
668 
669 	process_mm_walk_lock(walk.mm, ops->walk_lock);
670 	process_vma_walk_lock(vma, ops->walk_lock);
671 	return __walk_page_range(start, end, &walk);
672 }
673 
674 int walk_page_vma(struct vm_area_struct *vma, const struct mm_walk_ops *ops,
675 		void *private)
676 {
677 	struct mm_walk walk = {
678 		.ops		= ops,
679 		.mm		= vma->vm_mm,
680 		.vma		= vma,
681 		.private	= private,
682 	};
683 
684 	if (!walk.mm)
685 		return -EINVAL;
686 	if (!check_ops_valid(ops))
687 		return -EINVAL;
688 
689 	process_mm_walk_lock(walk.mm, ops->walk_lock);
690 	process_vma_walk_lock(vma, ops->walk_lock);
691 	return __walk_page_range(vma->vm_start, vma->vm_end, &walk);
692 }
693 
694 /**
695  * walk_page_mapping - walk all memory areas mapped into a struct address_space.
696  * @mapping: Pointer to the struct address_space
697  * @first_index: First page offset in the address_space
698  * @nr: Number of incremental page offsets to cover
699  * @ops:	operation to call during the walk
700  * @private:	private data for callbacks' usage
701  *
702  * This function walks all memory areas mapped into a struct address_space.
703  * The walk is limited to only the given page-size index range, but if
704  * the index boundaries cross a huge page-table entry, that entry will be
705  * included.
706  *
707  * Also see walk_page_range() for additional information.
708  *
709  * Locking:
710  *   This function can't require that the struct mm_struct::mmap_lock is held,
711  *   since @mapping may be mapped by multiple processes. Instead
712  *   @mapping->i_mmap_rwsem must be held. This might have implications in the
713  *   callbacks, and it's up tho the caller to ensure that the
714  *   struct mm_struct::mmap_lock is not needed.
715  *
716  *   Also this means that a caller can't rely on the struct
717  *   vm_area_struct::vm_flags to be constant across a call,
718  *   except for immutable flags. Callers requiring this shouldn't use
719  *   this function.
720  *
721  * Return: 0 on success, negative error code on failure, positive number on
722  * caller defined premature termination.
723  */
724 int walk_page_mapping(struct address_space *mapping, pgoff_t first_index,
725 		      pgoff_t nr, const struct mm_walk_ops *ops,
726 		      void *private)
727 {
728 	struct mm_walk walk = {
729 		.ops		= ops,
730 		.private	= private,
731 	};
732 	struct vm_area_struct *vma;
733 	pgoff_t vba, vea, cba, cea;
734 	unsigned long start_addr, end_addr;
735 	int err = 0;
736 
737 	if (!check_ops_valid(ops))
738 		return -EINVAL;
739 
740 	lockdep_assert_held(&mapping->i_mmap_rwsem);
741 	vma_interval_tree_foreach(vma, &mapping->i_mmap, first_index,
742 				  first_index + nr - 1) {
743 		/* Clip to the vma */
744 		vba = vma->vm_pgoff;
745 		vea = vba + vma_pages(vma);
746 		cba = first_index;
747 		cba = max(cba, vba);
748 		cea = first_index + nr;
749 		cea = min(cea, vea);
750 
751 		start_addr = ((cba - vba) << PAGE_SHIFT) + vma->vm_start;
752 		end_addr = ((cea - vba) << PAGE_SHIFT) + vma->vm_start;
753 		if (start_addr >= end_addr)
754 			continue;
755 
756 		walk.vma = vma;
757 		walk.mm = vma->vm_mm;
758 
759 		err = walk_page_test(vma->vm_start, vma->vm_end, &walk);
760 		if (err > 0) {
761 			err = 0;
762 			break;
763 		} else if (err < 0)
764 			break;
765 
766 		err = __walk_page_range(start_addr, end_addr, &walk);
767 		if (err)
768 			break;
769 	}
770 
771 	return err;
772 }
773 
774 /**
775  * folio_walk_start - walk the page tables to a folio
776  * @fw: filled with information on success.
777  * @vma: the VMA.
778  * @addr: the virtual address to use for the page table walk.
779  * @flags: flags modifying which folios to walk to.
780  *
781  * Walk the page tables using @addr in a given @vma to a mapped folio and
782  * return the folio, making sure that the page table entry referenced by
783  * @addr cannot change until folio_walk_end() was called.
784  *
785  * As default, this function returns only folios that are not special (e.g., not
786  * the zeropage) and never returns folios that are supposed to be ignored by the
787  * VM as documented by vm_normal_page(). If requested, zeropages will be
788  * returned as well.
789  *
790  * As default, this function only considers present page table entries.
791  * If requested, it will also consider migration entries.
792  *
793  * If this function returns NULL it might either indicate "there is nothing" or
794  * "there is nothing suitable".
795  *
796  * On success, @fw is filled and the function returns the folio while the PTL
797  * is still held and folio_walk_end() must be called to clean up,
798  * releasing any held locks. The returned folio must *not* be used after the
799  * call to folio_walk_end(), unless a short-term folio reference is taken before
800  * that call.
801  *
802  * @fw->page will correspond to the page that is effectively referenced by
803  * @addr. However, for migration entries and shared zeropages @fw->page is
804  * set to NULL. Note that large folios might be mapped by multiple page table
805  * entries, and this function will always only lookup a single entry as
806  * specified by @addr, which might or might not cover more than a single page of
807  * the returned folio.
808  *
809  * This function must *not* be used as a naive replacement for
810  * get_user_pages() / pin_user_pages(), especially not to perform DMA or
811  * to carelessly modify page content. This function may *only* be used to grab
812  * short-term folio references, never to grab long-term folio references.
813  *
814  * Using the page table entry pointers in @fw for reading or modifying the
815  * entry should be avoided where possible: however, there might be valid
816  * use cases.
817  *
818  * WARNING: Modifying page table entries in hugetlb VMAs requires a lot of care.
819  * For example, PMD page table sharing might require prior unsharing. Also,
820  * logical hugetlb entries might span multiple physical page table entries,
821  * which *must* be modified in a single operation (set_huge_pte_at(),
822  * huge_ptep_set_*, ...). Note that the page table entry stored in @fw might
823  * not correspond to the first physical entry of a logical hugetlb entry.
824  *
825  * The mmap lock must be held in read mode.
826  *
827  * Return: folio pointer on success, otherwise NULL.
828  */
829 struct folio *folio_walk_start(struct folio_walk *fw,
830 		struct vm_area_struct *vma, unsigned long addr,
831 		folio_walk_flags_t flags)
832 {
833 	unsigned long entry_size;
834 	bool expose_page = true;
835 	struct page *page;
836 	pud_t *pudp, pud;
837 	pmd_t *pmdp, pmd;
838 	pte_t *ptep, pte;
839 	spinlock_t *ptl;
840 	pgd_t *pgdp;
841 	p4d_t *p4dp;
842 
843 	mmap_assert_locked(vma->vm_mm);
844 	vma_pgtable_walk_begin(vma);
845 
846 	if (WARN_ON_ONCE(addr < vma->vm_start || addr >= vma->vm_end))
847 		goto not_found;
848 
849 	pgdp = pgd_offset(vma->vm_mm, addr);
850 	if (pgd_none_or_clear_bad(pgdp))
851 		goto not_found;
852 
853 	p4dp = p4d_offset(pgdp, addr);
854 	if (p4d_none_or_clear_bad(p4dp))
855 		goto not_found;
856 
857 	pudp = pud_offset(p4dp, addr);
858 	pud = pudp_get(pudp);
859 	if (pud_none(pud))
860 		goto not_found;
861 	if (IS_ENABLED(CONFIG_PGTABLE_HAS_HUGE_LEAVES) &&
862 	    (!pud_present(pud) || pud_leaf(pud))) {
863 		ptl = pud_lock(vma->vm_mm, pudp);
864 		pud = pudp_get(pudp);
865 
866 		entry_size = PUD_SIZE;
867 		fw->level = FW_LEVEL_PUD;
868 		fw->pudp = pudp;
869 		fw->pud = pud;
870 
871 		/*
872 		 * TODO: FW_MIGRATION support for PUD migration entries
873 		 * once there are relevant users.
874 		 */
875 		if (!pud_present(pud) || pud_devmap(pud) || pud_special(pud)) {
876 			spin_unlock(ptl);
877 			goto not_found;
878 		} else if (!pud_leaf(pud)) {
879 			spin_unlock(ptl);
880 			goto pmd_table;
881 		}
882 		/*
883 		 * TODO: vm_normal_page_pud() will be handy once we want to
884 		 * support PUD mappings in VM_PFNMAP|VM_MIXEDMAP VMAs.
885 		 */
886 		page = pud_page(pud);
887 		goto found;
888 	}
889 
890 pmd_table:
891 	VM_WARN_ON_ONCE(!pud_present(pud) || pud_leaf(pud));
892 	pmdp = pmd_offset(pudp, addr);
893 	pmd = pmdp_get_lockless(pmdp);
894 	if (pmd_none(pmd))
895 		goto not_found;
896 	if (IS_ENABLED(CONFIG_PGTABLE_HAS_HUGE_LEAVES) &&
897 	    (!pmd_present(pmd) || pmd_leaf(pmd))) {
898 		ptl = pmd_lock(vma->vm_mm, pmdp);
899 		pmd = pmdp_get(pmdp);
900 
901 		entry_size = PMD_SIZE;
902 		fw->level = FW_LEVEL_PMD;
903 		fw->pmdp = pmdp;
904 		fw->pmd = pmd;
905 
906 		if (pmd_none(pmd)) {
907 			spin_unlock(ptl);
908 			goto not_found;
909 		} else if (pmd_present(pmd) && !pmd_leaf(pmd)) {
910 			spin_unlock(ptl);
911 			goto pte_table;
912 		} else if (pmd_present(pmd)) {
913 			page = vm_normal_page_pmd(vma, addr, pmd);
914 			if (page) {
915 				goto found;
916 			} else if ((flags & FW_ZEROPAGE) &&
917 				    is_huge_zero_pmd(pmd)) {
918 				page = pfn_to_page(pmd_pfn(pmd));
919 				expose_page = false;
920 				goto found;
921 			}
922 		} else if ((flags & FW_MIGRATION) &&
923 			   is_pmd_migration_entry(pmd)) {
924 			swp_entry_t entry = pmd_to_swp_entry(pmd);
925 
926 			page = pfn_swap_entry_to_page(entry);
927 			expose_page = false;
928 			goto found;
929 		}
930 		spin_unlock(ptl);
931 		goto not_found;
932 	}
933 
934 pte_table:
935 	VM_WARN_ON_ONCE(!pmd_present(pmd) || pmd_leaf(pmd));
936 	ptep = pte_offset_map_lock(vma->vm_mm, pmdp, addr, &ptl);
937 	if (!ptep)
938 		goto not_found;
939 	pte = ptep_get(ptep);
940 
941 	entry_size = PAGE_SIZE;
942 	fw->level = FW_LEVEL_PTE;
943 	fw->ptep = ptep;
944 	fw->pte = pte;
945 
946 	if (pte_present(pte)) {
947 		page = vm_normal_page(vma, addr, pte);
948 		if (page)
949 			goto found;
950 		if ((flags & FW_ZEROPAGE) &&
951 		    is_zero_pfn(pte_pfn(pte))) {
952 			page = pfn_to_page(pte_pfn(pte));
953 			expose_page = false;
954 			goto found;
955 		}
956 	} else if (!pte_none(pte)) {
957 		swp_entry_t entry = pte_to_swp_entry(pte);
958 
959 		if ((flags & FW_MIGRATION) &&
960 		    is_migration_entry(entry)) {
961 			page = pfn_swap_entry_to_page(entry);
962 			expose_page = false;
963 			goto found;
964 		}
965 	}
966 	pte_unmap_unlock(ptep, ptl);
967 not_found:
968 	vma_pgtable_walk_end(vma);
969 	return NULL;
970 found:
971 	if (expose_page)
972 		/* Note: Offset from the mapped page, not the folio start. */
973 		fw->page = nth_page(page, (addr & (entry_size - 1)) >> PAGE_SHIFT);
974 	else
975 		fw->page = NULL;
976 	fw->ptl = ptl;
977 	return page_folio(page);
978 }
979