xref: /openbmc/linux/mm/pagewalk.c (revision d3964221)
1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/mm.h>
3 #include <linux/highmem.h>
4 #include <linux/sched.h>
5 #include <linux/hugetlb.h>
6 
7 static int walk_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
8 			  struct mm_walk *walk)
9 {
10 	pte_t *pte;
11 	int err = 0;
12 
13 	pte = pte_offset_map(pmd, addr);
14 	for (;;) {
15 		err = walk->pte_entry(pte, addr, addr + PAGE_SIZE, walk);
16 		if (err)
17 		       break;
18 		addr += PAGE_SIZE;
19 		if (addr == end)
20 			break;
21 		pte++;
22 	}
23 
24 	pte_unmap(pte);
25 	return err;
26 }
27 
28 static int walk_pmd_range(pud_t *pud, unsigned long addr, unsigned long end,
29 			  struct mm_walk *walk)
30 {
31 	pmd_t *pmd;
32 	unsigned long next;
33 	int err = 0;
34 
35 	pmd = pmd_offset(pud, addr);
36 	do {
37 again:
38 		next = pmd_addr_end(addr, end);
39 		if (pmd_none(*pmd) || !walk->vma) {
40 			if (walk->pte_hole)
41 				err = walk->pte_hole(addr, next, walk);
42 			if (err)
43 				break;
44 			continue;
45 		}
46 		/*
47 		 * This implies that each ->pmd_entry() handler
48 		 * needs to know about pmd_trans_huge() pmds
49 		 */
50 		if (walk->pmd_entry)
51 			err = walk->pmd_entry(pmd, addr, next, walk);
52 		if (err)
53 			break;
54 
55 		/*
56 		 * Check this here so we only break down trans_huge
57 		 * pages when we _need_ to
58 		 */
59 		if (!walk->pte_entry)
60 			continue;
61 
62 		split_huge_pmd(walk->vma, pmd, addr);
63 		if (pmd_trans_unstable(pmd))
64 			goto again;
65 		err = walk_pte_range(pmd, addr, next, walk);
66 		if (err)
67 			break;
68 	} while (pmd++, addr = next, addr != end);
69 
70 	return err;
71 }
72 
73 static int walk_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end,
74 			  struct mm_walk *walk)
75 {
76 	pud_t *pud;
77 	unsigned long next;
78 	int err = 0;
79 
80 	pud = pud_offset(p4d, addr);
81 	do {
82  again:
83 		next = pud_addr_end(addr, end);
84 		if (pud_none(*pud) || !walk->vma) {
85 			if (walk->pte_hole)
86 				err = walk->pte_hole(addr, next, walk);
87 			if (err)
88 				break;
89 			continue;
90 		}
91 
92 		if (walk->pud_entry) {
93 			spinlock_t *ptl = pud_trans_huge_lock(pud, walk->vma);
94 
95 			if (ptl) {
96 				err = walk->pud_entry(pud, addr, next, walk);
97 				spin_unlock(ptl);
98 				if (err)
99 					break;
100 				continue;
101 			}
102 		}
103 
104 		split_huge_pud(walk->vma, pud, addr);
105 		if (pud_none(*pud))
106 			goto again;
107 
108 		if (walk->pmd_entry || walk->pte_entry)
109 			err = walk_pmd_range(pud, addr, next, walk);
110 		if (err)
111 			break;
112 	} while (pud++, addr = next, addr != end);
113 
114 	return err;
115 }
116 
117 static int walk_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
118 			  struct mm_walk *walk)
119 {
120 	p4d_t *p4d;
121 	unsigned long next;
122 	int err = 0;
123 
124 	p4d = p4d_offset(pgd, addr);
125 	do {
126 		next = p4d_addr_end(addr, end);
127 		if (p4d_none_or_clear_bad(p4d)) {
128 			if (walk->pte_hole)
129 				err = walk->pte_hole(addr, next, walk);
130 			if (err)
131 				break;
132 			continue;
133 		}
134 		if (walk->pmd_entry || walk->pte_entry)
135 			err = walk_pud_range(p4d, addr, next, walk);
136 		if (err)
137 			break;
138 	} while (p4d++, addr = next, addr != end);
139 
140 	return err;
141 }
142 
143 static int walk_pgd_range(unsigned long addr, unsigned long end,
144 			  struct mm_walk *walk)
145 {
146 	pgd_t *pgd;
147 	unsigned long next;
148 	int err = 0;
149 
150 	pgd = pgd_offset(walk->mm, addr);
151 	do {
152 		next = pgd_addr_end(addr, end);
153 		if (pgd_none_or_clear_bad(pgd)) {
154 			if (walk->pte_hole)
155 				err = walk->pte_hole(addr, next, walk);
156 			if (err)
157 				break;
158 			continue;
159 		}
160 		if (walk->pmd_entry || walk->pte_entry)
161 			err = walk_p4d_range(pgd, addr, next, walk);
162 		if (err)
163 			break;
164 	} while (pgd++, addr = next, addr != end);
165 
166 	return err;
167 }
168 
169 #ifdef CONFIG_HUGETLB_PAGE
170 static unsigned long hugetlb_entry_end(struct hstate *h, unsigned long addr,
171 				       unsigned long end)
172 {
173 	unsigned long boundary = (addr & huge_page_mask(h)) + huge_page_size(h);
174 	return boundary < end ? boundary : end;
175 }
176 
177 static int walk_hugetlb_range(unsigned long addr, unsigned long end,
178 			      struct mm_walk *walk)
179 {
180 	struct vm_area_struct *vma = walk->vma;
181 	struct hstate *h = hstate_vma(vma);
182 	unsigned long next;
183 	unsigned long hmask = huge_page_mask(h);
184 	unsigned long sz = huge_page_size(h);
185 	pte_t *pte;
186 	int err = 0;
187 
188 	do {
189 		next = hugetlb_entry_end(h, addr, end);
190 		pte = huge_pte_offset(walk->mm, addr & hmask, sz);
191 		if (pte && walk->hugetlb_entry)
192 			err = walk->hugetlb_entry(pte, hmask, addr, next, walk);
193 		if (err)
194 			break;
195 	} while (addr = next, addr != end);
196 
197 	return err;
198 }
199 
200 #else /* CONFIG_HUGETLB_PAGE */
201 static int walk_hugetlb_range(unsigned long addr, unsigned long end,
202 			      struct mm_walk *walk)
203 {
204 	return 0;
205 }
206 
207 #endif /* CONFIG_HUGETLB_PAGE */
208 
209 /*
210  * Decide whether we really walk over the current vma on [@start, @end)
211  * or skip it via the returned value. Return 0 if we do walk over the
212  * current vma, and return 1 if we skip the vma. Negative values means
213  * error, where we abort the current walk.
214  */
215 static int walk_page_test(unsigned long start, unsigned long end,
216 			struct mm_walk *walk)
217 {
218 	struct vm_area_struct *vma = walk->vma;
219 
220 	if (walk->test_walk)
221 		return walk->test_walk(start, end, walk);
222 
223 	/*
224 	 * vma(VM_PFNMAP) doesn't have any valid struct pages behind VM_PFNMAP
225 	 * range, so we don't walk over it as we do for normal vmas. However,
226 	 * Some callers are interested in handling hole range and they don't
227 	 * want to just ignore any single address range. Such users certainly
228 	 * define their ->pte_hole() callbacks, so let's delegate them to handle
229 	 * vma(VM_PFNMAP).
230 	 */
231 	if (vma->vm_flags & VM_PFNMAP) {
232 		int err = 1;
233 		if (walk->pte_hole)
234 			err = walk->pte_hole(start, end, walk);
235 		return err ? err : 1;
236 	}
237 	return 0;
238 }
239 
240 static int __walk_page_range(unsigned long start, unsigned long end,
241 			struct mm_walk *walk)
242 {
243 	int err = 0;
244 	struct vm_area_struct *vma = walk->vma;
245 
246 	if (vma && is_vm_hugetlb_page(vma)) {
247 		if (walk->hugetlb_entry)
248 			err = walk_hugetlb_range(start, end, walk);
249 	} else
250 		err = walk_pgd_range(start, end, walk);
251 
252 	return err;
253 }
254 
255 /**
256  * walk_page_range - walk page table with caller specific callbacks
257  *
258  * Recursively walk the page table tree of the process represented by @walk->mm
259  * within the virtual address range [@start, @end). During walking, we can do
260  * some caller-specific works for each entry, by setting up pmd_entry(),
261  * pte_entry(), and/or hugetlb_entry(). If you don't set up for some of these
262  * callbacks, the associated entries/pages are just ignored.
263  * The return values of these callbacks are commonly defined like below:
264  *  - 0  : succeeded to handle the current entry, and if you don't reach the
265  *         end address yet, continue to walk.
266  *  - >0 : succeeded to handle the current entry, and return to the caller
267  *         with caller specific value.
268  *  - <0 : failed to handle the current entry, and return to the caller
269  *         with error code.
270  *
271  * Before starting to walk page table, some callers want to check whether
272  * they really want to walk over the current vma, typically by checking
273  * its vm_flags. walk_page_test() and @walk->test_walk() are used for this
274  * purpose.
275  *
276  * struct mm_walk keeps current values of some common data like vma and pmd,
277  * which are useful for the access from callbacks. If you want to pass some
278  * caller-specific data to callbacks, @walk->private should be helpful.
279  *
280  * Locking:
281  *   Callers of walk_page_range() and walk_page_vma() should hold
282  *   @walk->mm->mmap_sem, because these function traverse vma list and/or
283  *   access to vma's data.
284  */
285 int walk_page_range(unsigned long start, unsigned long end,
286 		    struct mm_walk *walk)
287 {
288 	int err = 0;
289 	unsigned long next;
290 	struct vm_area_struct *vma;
291 
292 	if (start >= end)
293 		return -EINVAL;
294 
295 	if (!walk->mm)
296 		return -EINVAL;
297 
298 	VM_BUG_ON_MM(!rwsem_is_locked(&walk->mm->mmap_sem), walk->mm);
299 
300 	vma = find_vma(walk->mm, start);
301 	do {
302 		if (!vma) { /* after the last vma */
303 			walk->vma = NULL;
304 			next = end;
305 		} else if (start < vma->vm_start) { /* outside vma */
306 			walk->vma = NULL;
307 			next = min(end, vma->vm_start);
308 		} else { /* inside vma */
309 			walk->vma = vma;
310 			next = min(end, vma->vm_end);
311 			vma = vma->vm_next;
312 
313 			err = walk_page_test(start, next, walk);
314 			if (err > 0) {
315 				/*
316 				 * positive return values are purely for
317 				 * controlling the pagewalk, so should never
318 				 * be passed to the callers.
319 				 */
320 				err = 0;
321 				continue;
322 			}
323 			if (err < 0)
324 				break;
325 		}
326 		if (walk->vma || walk->pte_hole)
327 			err = __walk_page_range(start, next, walk);
328 		if (err)
329 			break;
330 	} while (start = next, start < end);
331 	return err;
332 }
333 
334 int walk_page_vma(struct vm_area_struct *vma, struct mm_walk *walk)
335 {
336 	int err;
337 
338 	if (!walk->mm)
339 		return -EINVAL;
340 
341 	VM_BUG_ON(!rwsem_is_locked(&walk->mm->mmap_sem));
342 	VM_BUG_ON(!vma);
343 	walk->vma = vma;
344 	err = walk_page_test(vma->vm_start, vma->vm_end, walk);
345 	if (err > 0)
346 		return 0;
347 	if (err < 0)
348 		return err;
349 	return __walk_page_range(vma->vm_start, vma->vm_end, walk);
350 }
351