1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/mm.h>
3 #include <linux/rmap.h>
4 #include <linux/hugetlb.h>
5 #include <linux/swap.h>
6 #include <linux/swapops.h>
7 
8 #include "internal.h"
9 
not_found(struct page_vma_mapped_walk * pvmw)10 static inline bool not_found(struct page_vma_mapped_walk *pvmw)
11 {
12 	page_vma_mapped_walk_done(pvmw);
13 	return false;
14 }
15 
map_pte(struct page_vma_mapped_walk * pvmw,pmd_t * pmdvalp,spinlock_t ** ptlp)16 static bool map_pte(struct page_vma_mapped_walk *pvmw, pmd_t *pmdvalp,
17 		    spinlock_t **ptlp)
18 {
19 	pte_t ptent;
20 
21 	if (pvmw->flags & PVMW_SYNC) {
22 		/* Use the stricter lookup */
23 		pvmw->pte = pte_offset_map_lock(pvmw->vma->vm_mm, pvmw->pmd,
24 						pvmw->address, &pvmw->ptl);
25 		*ptlp = pvmw->ptl;
26 		return !!pvmw->pte;
27 	}
28 
29 again:
30 	/*
31 	 * It is important to return the ptl corresponding to pte,
32 	 * in case *pvmw->pmd changes underneath us; so we need to
33 	 * return it even when choosing not to lock, in case caller
34 	 * proceeds to loop over next ptes, and finds a match later.
35 	 * Though, in most cases, page lock already protects this.
36 	 */
37 	pvmw->pte = pte_offset_map_rw_nolock(pvmw->vma->vm_mm, pvmw->pmd,
38 					     pvmw->address, pmdvalp, ptlp);
39 	if (!pvmw->pte)
40 		return false;
41 
42 	ptent = ptep_get(pvmw->pte);
43 
44 	if (pvmw->flags & PVMW_MIGRATION) {
45 		if (!is_swap_pte(ptent))
46 			return false;
47 	} else if (is_swap_pte(ptent)) {
48 		swp_entry_t entry;
49 		/*
50 		 * Handle un-addressable ZONE_DEVICE memory.
51 		 *
52 		 * We get here when we are trying to unmap a private
53 		 * device page from the process address space. Such
54 		 * page is not CPU accessible and thus is mapped as
55 		 * a special swap entry, nonetheless it still does
56 		 * count as a valid regular mapping for the page
57 		 * (and is accounted as such in page maps count).
58 		 *
59 		 * So handle this special case as if it was a normal
60 		 * page mapping ie lock CPU page table and return true.
61 		 *
62 		 * For more details on device private memory see HMM
63 		 * (include/linux/hmm.h or mm/hmm.c).
64 		 */
65 		entry = pte_to_swp_entry(ptent);
66 		if (!is_device_private_entry(entry) &&
67 		    !is_device_exclusive_entry(entry))
68 			return false;
69 	} else if (!pte_present(ptent)) {
70 		return false;
71 	}
72 	spin_lock(*ptlp);
73 	if (unlikely(!pmd_same(*pmdvalp, pmdp_get_lockless(pvmw->pmd)))) {
74 		pte_unmap_unlock(pvmw->pte, *ptlp);
75 		goto again;
76 	}
77 	pvmw->ptl = *ptlp;
78 
79 	return true;
80 }
81 
82 /**
83  * check_pte - check if [pvmw->pfn, @pvmw->pfn + @pvmw->nr_pages) is
84  * mapped at the @pvmw->pte
85  * @pvmw: page_vma_mapped_walk struct, includes a pair pte and pfn range
86  * for checking
87  * @pte_nr: the number of small pages described by @pvmw->pte.
88  *
89  * page_vma_mapped_walk() found a place where pfn range is *potentially*
90  * mapped. check_pte() has to validate this.
91  *
92  * pvmw->pte may point to empty PTE, swap PTE or PTE pointing to
93  * arbitrary page.
94  *
95  * If PVMW_MIGRATION flag is set, returns true if @pvmw->pte contains migration
96  * entry that points to [pvmw->pfn, @pvmw->pfn + @pvmw->nr_pages)
97  *
98  * If PVMW_MIGRATION flag is not set, returns true if pvmw->pte points to
99  * [pvmw->pfn, @pvmw->pfn + @pvmw->nr_pages)
100  *
101  * Otherwise, return false.
102  *
103  */
check_pte(struct page_vma_mapped_walk * pvmw,unsigned long pte_nr)104 static bool check_pte(struct page_vma_mapped_walk *pvmw, unsigned long pte_nr)
105 {
106 	unsigned long pfn;
107 	pte_t ptent = ptep_get(pvmw->pte);
108 
109 	if (pvmw->flags & PVMW_MIGRATION) {
110 		swp_entry_t entry;
111 		if (!is_swap_pte(ptent))
112 			return false;
113 		entry = pte_to_swp_entry(ptent);
114 
115 		if (!is_migration_entry(entry) &&
116 		    !is_device_exclusive_entry(entry))
117 			return false;
118 
119 		pfn = swp_offset_pfn(entry);
120 	} else if (is_swap_pte(ptent)) {
121 		swp_entry_t entry;
122 
123 		/* Handle un-addressable ZONE_DEVICE memory */
124 		entry = pte_to_swp_entry(ptent);
125 		if (!is_device_private_entry(entry) &&
126 		    !is_device_exclusive_entry(entry))
127 			return false;
128 
129 		pfn = swp_offset_pfn(entry);
130 	} else {
131 		if (!pte_present(ptent))
132 			return false;
133 
134 		pfn = pte_pfn(ptent);
135 	}
136 
137 	if ((pfn + pte_nr - 1) < pvmw->pfn)
138 		return false;
139 	if (pfn > (pvmw->pfn + pvmw->nr_pages - 1))
140 		return false;
141 	return true;
142 }
143 
144 /* Returns true if the two ranges overlap.  Careful to not overflow. */
check_pmd(unsigned long pfn,struct page_vma_mapped_walk * pvmw)145 static bool check_pmd(unsigned long pfn, struct page_vma_mapped_walk *pvmw)
146 {
147 	if ((pfn + HPAGE_PMD_NR - 1) < pvmw->pfn)
148 		return false;
149 	if (pfn > pvmw->pfn + pvmw->nr_pages - 1)
150 		return false;
151 	return true;
152 }
153 
step_forward(struct page_vma_mapped_walk * pvmw,unsigned long size)154 static void step_forward(struct page_vma_mapped_walk *pvmw, unsigned long size)
155 {
156 	pvmw->address = (pvmw->address + size) & ~(size - 1);
157 	if (!pvmw->address)
158 		pvmw->address = ULONG_MAX;
159 }
160 
161 /**
162  * page_vma_mapped_walk - check if @pvmw->pfn is mapped in @pvmw->vma at
163  * @pvmw->address
164  * @pvmw: pointer to struct page_vma_mapped_walk. page, vma, address and flags
165  * must be set. pmd, pte and ptl must be NULL.
166  *
167  * Returns true if the page is mapped in the vma. @pvmw->pmd and @pvmw->pte point
168  * to relevant page table entries. @pvmw->ptl is locked. @pvmw->address is
169  * adjusted if needed (for PTE-mapped THPs).
170  *
171  * If @pvmw->pmd is set but @pvmw->pte is not, you have found PMD-mapped page
172  * (usually THP). For PTE-mapped THP, you should run page_vma_mapped_walk() in
173  * a loop to find all PTEs that map the THP.
174  *
175  * For HugeTLB pages, @pvmw->pte is set to the relevant page table entry
176  * regardless of which page table level the page is mapped at. @pvmw->pmd is
177  * NULL.
178  *
179  * Returns false if there are no more page table entries for the page in
180  * the vma. @pvmw->ptl is unlocked and @pvmw->pte is unmapped.
181  *
182  * If you need to stop the walk before page_vma_mapped_walk() returned false,
183  * use page_vma_mapped_walk_done(). It will do the housekeeping.
184  */
page_vma_mapped_walk(struct page_vma_mapped_walk * pvmw)185 bool page_vma_mapped_walk(struct page_vma_mapped_walk *pvmw)
186 {
187 	struct vm_area_struct *vma = pvmw->vma;
188 	struct mm_struct *mm = vma->vm_mm;
189 	unsigned long end;
190 	spinlock_t *ptl;
191 	pgd_t *pgd;
192 	p4d_t *p4d;
193 	pud_t *pud;
194 	pmd_t pmde;
195 
196 	/* The only possible pmd mapping has been handled on last iteration */
197 	if (pvmw->pmd && !pvmw->pte)
198 		return not_found(pvmw);
199 
200 	if (unlikely(is_vm_hugetlb_page(vma))) {
201 		struct hstate *hstate = hstate_vma(vma);
202 		unsigned long size = huge_page_size(hstate);
203 		/* The only possible mapping was handled on last iteration */
204 		if (pvmw->pte)
205 			return not_found(pvmw);
206 		/*
207 		 * All callers that get here will already hold the
208 		 * i_mmap_rwsem.  Therefore, no additional locks need to be
209 		 * taken before calling hugetlb_walk().
210 		 */
211 		pvmw->pte = hugetlb_walk(vma, pvmw->address, size);
212 		if (!pvmw->pte)
213 			return false;
214 
215 		pvmw->ptl = huge_pte_lock(hstate, mm, pvmw->pte);
216 		if (!check_pte(pvmw, pages_per_huge_page(hstate)))
217 			return not_found(pvmw);
218 		return true;
219 	}
220 
221 	end = vma_address_end(pvmw);
222 	if (pvmw->pte)
223 		goto next_pte;
224 restart:
225 	do {
226 		pgd = pgd_offset(mm, pvmw->address);
227 		if (!pgd_present(*pgd)) {
228 			step_forward(pvmw, PGDIR_SIZE);
229 			continue;
230 		}
231 		p4d = p4d_offset(pgd, pvmw->address);
232 		if (!p4d_present(*p4d)) {
233 			step_forward(pvmw, P4D_SIZE);
234 			continue;
235 		}
236 		pud = pud_offset(p4d, pvmw->address);
237 		if (!pud_present(*pud)) {
238 			step_forward(pvmw, PUD_SIZE);
239 			continue;
240 		}
241 
242 		pvmw->pmd = pmd_offset(pud, pvmw->address);
243 		/*
244 		 * Make sure the pmd value isn't cached in a register by the
245 		 * compiler and used as a stale value after we've observed a
246 		 * subsequent update.
247 		 */
248 		pmde = pmdp_get_lockless(pvmw->pmd);
249 
250 		if (pmd_trans_huge(pmde) || is_pmd_migration_entry(pmde) ||
251 		    (pmd_present(pmde) && pmd_devmap(pmde))) {
252 			pvmw->ptl = pmd_lock(mm, pvmw->pmd);
253 			pmde = *pvmw->pmd;
254 			if (!pmd_present(pmde)) {
255 				swp_entry_t entry;
256 
257 				if (!thp_migration_supported() ||
258 				    !(pvmw->flags & PVMW_MIGRATION))
259 					return not_found(pvmw);
260 				entry = pmd_to_swp_entry(pmde);
261 				if (!is_migration_entry(entry) ||
262 				    !check_pmd(swp_offset_pfn(entry), pvmw))
263 					return not_found(pvmw);
264 				return true;
265 			}
266 			if (likely(pmd_trans_huge(pmde) || pmd_devmap(pmde))) {
267 				if (pvmw->flags & PVMW_MIGRATION)
268 					return not_found(pvmw);
269 				if (!check_pmd(pmd_pfn(pmde), pvmw))
270 					return not_found(pvmw);
271 				return true;
272 			}
273 			/* THP pmd was split under us: handle on pte level */
274 			spin_unlock(pvmw->ptl);
275 			pvmw->ptl = NULL;
276 		} else if (!pmd_present(pmde)) {
277 			/*
278 			 * If PVMW_SYNC, take and drop THP pmd lock so that we
279 			 * cannot return prematurely, while zap_huge_pmd() has
280 			 * cleared *pmd but not decremented compound_mapcount().
281 			 */
282 			if ((pvmw->flags & PVMW_SYNC) &&
283 			    thp_vma_suitable_order(vma, pvmw->address,
284 						   PMD_ORDER) &&
285 			    (pvmw->nr_pages >= HPAGE_PMD_NR)) {
286 				spinlock_t *ptl = pmd_lock(mm, pvmw->pmd);
287 
288 				spin_unlock(ptl);
289 			}
290 			step_forward(pvmw, PMD_SIZE);
291 			continue;
292 		}
293 		if (!map_pte(pvmw, &pmde, &ptl)) {
294 			if (!pvmw->pte)
295 				goto restart;
296 			goto next_pte;
297 		}
298 this_pte:
299 		if (check_pte(pvmw, 1))
300 			return true;
301 next_pte:
302 		do {
303 			pvmw->address += PAGE_SIZE;
304 			if (pvmw->address >= end)
305 				return not_found(pvmw);
306 			/* Did we cross page table boundary? */
307 			if ((pvmw->address & (PMD_SIZE - PAGE_SIZE)) == 0) {
308 				if (pvmw->ptl) {
309 					spin_unlock(pvmw->ptl);
310 					pvmw->ptl = NULL;
311 				}
312 				pte_unmap(pvmw->pte);
313 				pvmw->pte = NULL;
314 				goto restart;
315 			}
316 			pvmw->pte++;
317 		} while (pte_none(ptep_get(pvmw->pte)));
318 
319 		if (!pvmw->ptl) {
320 			spin_lock(ptl);
321 			if (unlikely(!pmd_same(pmde, pmdp_get_lockless(pvmw->pmd)))) {
322 				pte_unmap_unlock(pvmw->pte, ptl);
323 				pvmw->pte = NULL;
324 				goto restart;
325 			}
326 			pvmw->ptl = ptl;
327 		}
328 		goto this_pte;
329 	} while (pvmw->address < end);
330 
331 	return false;
332 }
333 
334 #ifdef CONFIG_MEMORY_FAILURE
335 /**
336  * page_mapped_in_vma - check whether a page is really mapped in a VMA
337  * @page: the page to test
338  * @vma: the VMA to test
339  *
340  * Return: The address the page is mapped at if the page is in the range
341  * covered by the VMA and present in the page table.  If the page is
342  * outside the VMA or not present, returns -EFAULT.
343  * Only valid for normal file or anonymous VMAs.
344  */
page_mapped_in_vma(const struct page * page,struct vm_area_struct * vma)345 unsigned long page_mapped_in_vma(const struct page *page,
346 		struct vm_area_struct *vma)
347 {
348 	const struct folio *folio = page_folio(page);
349 	struct page_vma_mapped_walk pvmw = {
350 		.pfn = page_to_pfn(page),
351 		.nr_pages = 1,
352 		.vma = vma,
353 		.flags = PVMW_SYNC,
354 	};
355 
356 	pvmw.address = vma_address(vma, page_pgoff(folio, page), 1);
357 	if (pvmw.address == -EFAULT)
358 		goto out;
359 	if (!page_vma_mapped_walk(&pvmw))
360 		return -EFAULT;
361 	page_vma_mapped_walk_done(&pvmw);
362 out:
363 	return pvmw.address;
364 }
365 #endif
366