xref: /openbmc/linux/mm/page_vma_mapped.c (revision 4ed91d48259d9ddd378424d008f2e6559f7e78f8)
1 #include <linux/mm.h>
2 #include <linux/rmap.h>
3 #include <linux/hugetlb.h>
4 #include <linux/swap.h>
5 #include <linux/swapops.h>
6 
7 #include "internal.h"
8 
9 static inline bool check_pmd(struct page_vma_mapped_walk *pvmw)
10 {
11 	pmd_t pmde;
12 	/*
13 	 * Make sure we don't re-load pmd between present and !trans_huge check.
14 	 * We need a consistent view.
15 	 */
16 	pmde = READ_ONCE(*pvmw->pmd);
17 	return pmd_present(pmde) && !pmd_trans_huge(pmde);
18 }
19 
20 static inline bool not_found(struct page_vma_mapped_walk *pvmw)
21 {
22 	page_vma_mapped_walk_done(pvmw);
23 	return false;
24 }
25 
26 static bool map_pte(struct page_vma_mapped_walk *pvmw)
27 {
28 	pvmw->pte = pte_offset_map(pvmw->pmd, pvmw->address);
29 	if (!(pvmw->flags & PVMW_SYNC)) {
30 		if (pvmw->flags & PVMW_MIGRATION) {
31 			if (!is_swap_pte(*pvmw->pte))
32 				return false;
33 		} else {
34 			if (!pte_present(*pvmw->pte))
35 				return false;
36 		}
37 	}
38 	pvmw->ptl = pte_lockptr(pvmw->vma->vm_mm, pvmw->pmd);
39 	spin_lock(pvmw->ptl);
40 	return true;
41 }
42 
43 static bool check_pte(struct page_vma_mapped_walk *pvmw)
44 {
45 	if (pvmw->flags & PVMW_MIGRATION) {
46 #ifdef CONFIG_MIGRATION
47 		swp_entry_t entry;
48 		if (!is_swap_pte(*pvmw->pte))
49 			return false;
50 		entry = pte_to_swp_entry(*pvmw->pte);
51 		if (!is_migration_entry(entry))
52 			return false;
53 		if (migration_entry_to_page(entry) - pvmw->page >=
54 				hpage_nr_pages(pvmw->page)) {
55 			return false;
56 		}
57 		if (migration_entry_to_page(entry) < pvmw->page)
58 			return false;
59 #else
60 		WARN_ON_ONCE(1);
61 #endif
62 	} else {
63 		if (!pte_present(*pvmw->pte))
64 			return false;
65 
66 		/* THP can be referenced by any subpage */
67 		if (pte_page(*pvmw->pte) - pvmw->page >=
68 				hpage_nr_pages(pvmw->page)) {
69 			return false;
70 		}
71 		if (pte_page(*pvmw->pte) < pvmw->page)
72 			return false;
73 	}
74 
75 	return true;
76 }
77 
78 /**
79  * page_vma_mapped_walk - check if @pvmw->page is mapped in @pvmw->vma at
80  * @pvmw->address
81  * @pvmw: pointer to struct page_vma_mapped_walk. page, vma, address and flags
82  * must be set. pmd, pte and ptl must be NULL.
83  *
84  * Returns true if the page is mapped in the vma. @pvmw->pmd and @pvmw->pte point
85  * to relevant page table entries. @pvmw->ptl is locked. @pvmw->address is
86  * adjusted if needed (for PTE-mapped THPs).
87  *
88  * If @pvmw->pmd is set but @pvmw->pte is not, you have found PMD-mapped page
89  * (usually THP). For PTE-mapped THP, you should run page_vma_mapped_walk() in
90  * a loop to find all PTEs that map the THP.
91  *
92  * For HugeTLB pages, @pvmw->pte is set to the relevant page table entry
93  * regardless of which page table level the page is mapped at. @pvmw->pmd is
94  * NULL.
95  *
96  * Retruns false if there are no more page table entries for the page in
97  * the vma. @pvmw->ptl is unlocked and @pvmw->pte is unmapped.
98  *
99  * If you need to stop the walk before page_vma_mapped_walk() returned false,
100  * use page_vma_mapped_walk_done(). It will do the housekeeping.
101  */
102 bool page_vma_mapped_walk(struct page_vma_mapped_walk *pvmw)
103 {
104 	struct mm_struct *mm = pvmw->vma->vm_mm;
105 	struct page *page = pvmw->page;
106 	pgd_t *pgd;
107 	p4d_t *p4d;
108 	pud_t *pud;
109 
110 	/* The only possible pmd mapping has been handled on last iteration */
111 	if (pvmw->pmd && !pvmw->pte)
112 		return not_found(pvmw);
113 
114 	/* Only for THP, seek to next pte entry makes sense */
115 	if (pvmw->pte) {
116 		if (!PageTransHuge(pvmw->page) || PageHuge(pvmw->page))
117 			return not_found(pvmw);
118 		goto next_pte;
119 	}
120 
121 	if (unlikely(PageHuge(pvmw->page))) {
122 		/* when pud is not present, pte will be NULL */
123 		pvmw->pte = huge_pte_offset(mm, pvmw->address);
124 		if (!pvmw->pte)
125 			return false;
126 
127 		pvmw->ptl = huge_pte_lockptr(page_hstate(page), mm, pvmw->pte);
128 		spin_lock(pvmw->ptl);
129 		if (!check_pte(pvmw))
130 			return not_found(pvmw);
131 		return true;
132 	}
133 restart:
134 	pgd = pgd_offset(mm, pvmw->address);
135 	if (!pgd_present(*pgd))
136 		return false;
137 	p4d = p4d_offset(pgd, pvmw->address);
138 	if (!p4d_present(*p4d))
139 		return false;
140 	pud = pud_offset(p4d, pvmw->address);
141 	if (!pud_present(*pud))
142 		return false;
143 	pvmw->pmd = pmd_offset(pud, pvmw->address);
144 	if (pmd_trans_huge(*pvmw->pmd)) {
145 		pvmw->ptl = pmd_lock(mm, pvmw->pmd);
146 		if (!pmd_present(*pvmw->pmd))
147 			return not_found(pvmw);
148 		if (likely(pmd_trans_huge(*pvmw->pmd))) {
149 			if (pvmw->flags & PVMW_MIGRATION)
150 				return not_found(pvmw);
151 			if (pmd_page(*pvmw->pmd) != page)
152 				return not_found(pvmw);
153 			return true;
154 		} else {
155 			/* THP pmd was split under us: handle on pte level */
156 			spin_unlock(pvmw->ptl);
157 			pvmw->ptl = NULL;
158 		}
159 	} else {
160 		if (!check_pmd(pvmw))
161 			return false;
162 	}
163 	if (!map_pte(pvmw))
164 		goto next_pte;
165 	while (1) {
166 		if (check_pte(pvmw))
167 			return true;
168 next_pte:	do {
169 			pvmw->address += PAGE_SIZE;
170 			if (pvmw->address >=
171 					__vma_address(pvmw->page, pvmw->vma) +
172 					hpage_nr_pages(pvmw->page) * PAGE_SIZE)
173 				return not_found(pvmw);
174 			/* Did we cross page table boundary? */
175 			if (pvmw->address % PMD_SIZE == 0) {
176 				pte_unmap(pvmw->pte);
177 				if (pvmw->ptl) {
178 					spin_unlock(pvmw->ptl);
179 					pvmw->ptl = NULL;
180 				}
181 				goto restart;
182 			} else {
183 				pvmw->pte++;
184 			}
185 		} while (pte_none(*pvmw->pte));
186 
187 		if (!pvmw->ptl) {
188 			pvmw->ptl = pte_lockptr(mm, pvmw->pmd);
189 			spin_lock(pvmw->ptl);
190 		}
191 	}
192 }
193 
194 /**
195  * page_mapped_in_vma - check whether a page is really mapped in a VMA
196  * @page: the page to test
197  * @vma: the VMA to test
198  *
199  * Returns 1 if the page is mapped into the page tables of the VMA, 0
200  * if the page is not mapped into the page tables of this VMA.  Only
201  * valid for normal file or anonymous VMAs.
202  */
203 int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
204 {
205 	struct page_vma_mapped_walk pvmw = {
206 		.page = page,
207 		.vma = vma,
208 		.flags = PVMW_SYNC,
209 	};
210 	unsigned long start, end;
211 
212 	start = __vma_address(page, vma);
213 	end = start + PAGE_SIZE * (hpage_nr_pages(page) - 1);
214 
215 	if (unlikely(end < vma->vm_start || start >= vma->vm_end))
216 		return 0;
217 	pvmw.address = max(start, vma->vm_start);
218 	if (!page_vma_mapped_walk(&pvmw))
219 		return 0;
220 	page_vma_mapped_walk_done(&pvmw);
221 	return 1;
222 }
223