xref: /linux/mm/page_vma_mapped.c (revision 7203ca412fc8e8a0588e9adc0f777d3163f8dff3)
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/leafops.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 	bool is_migration;
20 	pte_t ptent;
21 
22 	if (pvmw->flags & PVMW_SYNC) {
23 		/* Use the stricter lookup */
24 		pvmw->pte = pte_offset_map_lock(pvmw->vma->vm_mm, pvmw->pmd,
25 						pvmw->address, &pvmw->ptl);
26 		*ptlp = pvmw->ptl;
27 		return !!pvmw->pte;
28 	}
29 
30 	is_migration = pvmw->flags & PVMW_MIGRATION;
31 again:
32 	/*
33 	 * It is important to return the ptl corresponding to pte,
34 	 * in case *pvmw->pmd changes underneath us; so we need to
35 	 * return it even when choosing not to lock, in case caller
36 	 * proceeds to loop over next ptes, and finds a match later.
37 	 * Though, in most cases, page lock already protects this.
38 	 */
39 	pvmw->pte = pte_offset_map_rw_nolock(pvmw->vma->vm_mm, pvmw->pmd,
40 					     pvmw->address, pmdvalp, ptlp);
41 	if (!pvmw->pte)
42 		return false;
43 
44 	ptent = ptep_get(pvmw->pte);
45 
46 	if (pte_none(ptent)) {
47 		return false;
48 	} else if (pte_present(ptent)) {
49 		if (is_migration)
50 			return false;
51 	} else if (!is_migration) {
52 		softleaf_t entry;
53 
54 		/*
55 		 * Handle un-addressable ZONE_DEVICE memory.
56 		 *
57 		 * We get here when we are trying to unmap a private
58 		 * device page from the process address space. Such
59 		 * page is not CPU accessible and thus is mapped as
60 		 * a special swap entry, nonetheless it still does
61 		 * count as a valid regular mapping for the page
62 		 * (and is accounted as such in page maps count).
63 		 *
64 		 * So handle this special case as if it was a normal
65 		 * page mapping ie lock CPU page table and return true.
66 		 *
67 		 * For more details on device private memory see HMM
68 		 * (include/linux/hmm.h or mm/hmm.c).
69 		 */
70 		entry = softleaf_from_pte(ptent);
71 		if (!softleaf_is_device_private(entry) &&
72 		    !softleaf_is_device_exclusive(entry))
73 			return false;
74 	}
75 	spin_lock(*ptlp);
76 	if (unlikely(!pmd_same(*pmdvalp, pmdp_get_lockless(pvmw->pmd)))) {
77 		pte_unmap_unlock(pvmw->pte, *ptlp);
78 		goto again;
79 	}
80 	pvmw->ptl = *ptlp;
81 
82 	return true;
83 }
84 
85 /**
86  * check_pte - check if [pvmw->pfn, @pvmw->pfn + @pvmw->nr_pages) is
87  * mapped at the @pvmw->pte
88  * @pvmw: page_vma_mapped_walk struct, includes a pair pte and pfn range
89  * for checking
90  * @pte_nr: the number of small pages described by @pvmw->pte.
91  *
92  * page_vma_mapped_walk() found a place where pfn range is *potentially*
93  * mapped. check_pte() has to validate this.
94  *
95  * pvmw->pte may point to empty PTE, swap PTE or PTE pointing to
96  * arbitrary page.
97  *
98  * If PVMW_MIGRATION flag is set, returns true if @pvmw->pte contains migration
99  * entry that points to [pvmw->pfn, @pvmw->pfn + @pvmw->nr_pages)
100  *
101  * If PVMW_MIGRATION flag is not set, returns true if pvmw->pte points to
102  * [pvmw->pfn, @pvmw->pfn + @pvmw->nr_pages)
103  *
104  * Otherwise, return false.
105  *
106  */
check_pte(struct page_vma_mapped_walk * pvmw,unsigned long pte_nr)107 static bool check_pte(struct page_vma_mapped_walk *pvmw, unsigned long pte_nr)
108 {
109 	unsigned long pfn;
110 	pte_t ptent = ptep_get(pvmw->pte);
111 
112 	if (pvmw->flags & PVMW_MIGRATION) {
113 		const softleaf_t entry = softleaf_from_pte(ptent);
114 
115 		if (!softleaf_is_migration(entry))
116 			return false;
117 
118 		pfn = softleaf_to_pfn(entry);
119 	} else if (pte_present(ptent)) {
120 		pfn = pte_pfn(ptent);
121 	} else {
122 		const softleaf_t entry = softleaf_from_pte(ptent);
123 
124 		/* Handle un-addressable ZONE_DEVICE memory */
125 		if (!softleaf_is_device_private(entry) &&
126 		    !softleaf_is_device_exclusive(entry))
127 			return false;
128 
129 		pfn = softleaf_to_pfn(entry);
130 	}
131 
132 	if ((pfn + pte_nr - 1) < pvmw->pfn)
133 		return false;
134 	if (pfn > (pvmw->pfn + pvmw->nr_pages - 1))
135 		return false;
136 	return true;
137 }
138 
139 /* Returns true if the two ranges overlap.  Careful to not overflow. */
check_pmd(unsigned long pfn,struct page_vma_mapped_walk * pvmw)140 static bool check_pmd(unsigned long pfn, struct page_vma_mapped_walk *pvmw)
141 {
142 	if ((pfn + HPAGE_PMD_NR - 1) < pvmw->pfn)
143 		return false;
144 	if (pfn > pvmw->pfn + pvmw->nr_pages - 1)
145 		return false;
146 	return true;
147 }
148 
step_forward(struct page_vma_mapped_walk * pvmw,unsigned long size)149 static void step_forward(struct page_vma_mapped_walk *pvmw, unsigned long size)
150 {
151 	pvmw->address = (pvmw->address + size) & ~(size - 1);
152 	if (!pvmw->address)
153 		pvmw->address = ULONG_MAX;
154 }
155 
156 /**
157  * page_vma_mapped_walk - check if @pvmw->pfn is mapped in @pvmw->vma at
158  * @pvmw->address
159  * @pvmw: pointer to struct page_vma_mapped_walk. page, vma, address and flags
160  * must be set. pmd, pte and ptl must be NULL.
161  *
162  * Returns true if the page is mapped in the vma. @pvmw->pmd and @pvmw->pte point
163  * to relevant page table entries. @pvmw->ptl is locked. @pvmw->address is
164  * adjusted if needed (for PTE-mapped THPs).
165  *
166  * If @pvmw->pmd is set but @pvmw->pte is not, you have found PMD-mapped page
167  * (usually THP). For PTE-mapped THP, you should run page_vma_mapped_walk() in
168  * a loop to find all PTEs that map the THP.
169  *
170  * For HugeTLB pages, @pvmw->pte is set to the relevant page table entry
171  * regardless of which page table level the page is mapped at. @pvmw->pmd is
172  * NULL.
173  *
174  * Returns false if there are no more page table entries for the page in
175  * the vma. @pvmw->ptl is unlocked and @pvmw->pte is unmapped.
176  *
177  * If you need to stop the walk before page_vma_mapped_walk() returned false,
178  * use page_vma_mapped_walk_done(). It will do the housekeeping.
179  */
page_vma_mapped_walk(struct page_vma_mapped_walk * pvmw)180 bool page_vma_mapped_walk(struct page_vma_mapped_walk *pvmw)
181 {
182 	struct vm_area_struct *vma = pvmw->vma;
183 	struct mm_struct *mm = vma->vm_mm;
184 	unsigned long end;
185 	spinlock_t *ptl;
186 	pgd_t *pgd;
187 	p4d_t *p4d;
188 	pud_t *pud;
189 	pmd_t pmde;
190 
191 	/* The only possible pmd mapping has been handled on last iteration */
192 	if (pvmw->pmd && !pvmw->pte)
193 		return not_found(pvmw);
194 
195 	if (unlikely(is_vm_hugetlb_page(vma))) {
196 		struct hstate *hstate = hstate_vma(vma);
197 		unsigned long size = huge_page_size(hstate);
198 		/* The only possible mapping was handled on last iteration */
199 		if (pvmw->pte)
200 			return not_found(pvmw);
201 		/*
202 		 * All callers that get here will already hold the
203 		 * i_mmap_rwsem.  Therefore, no additional locks need to be
204 		 * taken before calling hugetlb_walk().
205 		 */
206 		pvmw->pte = hugetlb_walk(vma, pvmw->address, size);
207 		if (!pvmw->pte)
208 			return false;
209 
210 		pvmw->ptl = huge_pte_lock(hstate, mm, pvmw->pte);
211 		if (!check_pte(pvmw, pages_per_huge_page(hstate)))
212 			return not_found(pvmw);
213 		return true;
214 	}
215 
216 	end = vma_address_end(pvmw);
217 	if (pvmw->pte)
218 		goto next_pte;
219 restart:
220 	do {
221 		pgd = pgd_offset(mm, pvmw->address);
222 		if (!pgd_present(*pgd)) {
223 			step_forward(pvmw, PGDIR_SIZE);
224 			continue;
225 		}
226 		p4d = p4d_offset(pgd, pvmw->address);
227 		if (!p4d_present(*p4d)) {
228 			step_forward(pvmw, P4D_SIZE);
229 			continue;
230 		}
231 		pud = pud_offset(p4d, pvmw->address);
232 		if (!pud_present(*pud)) {
233 			step_forward(pvmw, PUD_SIZE);
234 			continue;
235 		}
236 
237 		pvmw->pmd = pmd_offset(pud, pvmw->address);
238 		/*
239 		 * Make sure the pmd value isn't cached in a register by the
240 		 * compiler and used as a stale value after we've observed a
241 		 * subsequent update.
242 		 */
243 		pmde = pmdp_get_lockless(pvmw->pmd);
244 
245 		if (pmd_trans_huge(pmde) || pmd_is_migration_entry(pmde)) {
246 			pvmw->ptl = pmd_lock(mm, pvmw->pmd);
247 			pmde = *pvmw->pmd;
248 			if (!pmd_present(pmde)) {
249 				softleaf_t entry;
250 
251 				if (!thp_migration_supported() ||
252 				    !(pvmw->flags & PVMW_MIGRATION))
253 					return not_found(pvmw);
254 				entry = softleaf_from_pmd(pmde);
255 
256 				if (!softleaf_is_migration(entry) ||
257 				    !check_pmd(softleaf_to_pfn(entry), pvmw))
258 					return not_found(pvmw);
259 				return true;
260 			}
261 			if (likely(pmd_trans_huge(pmde))) {
262 				if (pvmw->flags & PVMW_MIGRATION)
263 					return not_found(pvmw);
264 				if (!check_pmd(pmd_pfn(pmde), pvmw))
265 					return not_found(pvmw);
266 				return true;
267 			}
268 			/* THP pmd was split under us: handle on pte level */
269 			spin_unlock(pvmw->ptl);
270 			pvmw->ptl = NULL;
271 		} else if (!pmd_present(pmde)) {
272 			/*
273 			 * If PVMW_SYNC, take and drop THP pmd lock so that we
274 			 * cannot return prematurely, while zap_huge_pmd() has
275 			 * cleared *pmd but not decremented compound_mapcount().
276 			 */
277 			const softleaf_t entry = softleaf_from_pmd(pmde);
278 
279 			if (softleaf_is_device_private(entry)) {
280 				pvmw->ptl = pmd_lock(mm, pvmw->pmd);
281 				return true;
282 			}
283 
284 			if ((pvmw->flags & PVMW_SYNC) &&
285 			    thp_vma_suitable_order(vma, pvmw->address,
286 						   PMD_ORDER) &&
287 			    (pvmw->nr_pages >= HPAGE_PMD_NR)) {
288 				spinlock_t *ptl = pmd_lock(mm, pvmw->pmd);
289 
290 				spin_unlock(ptl);
291 			}
292 			step_forward(pvmw, PMD_SIZE);
293 			continue;
294 		}
295 		if (!map_pte(pvmw, &pmde, &ptl)) {
296 			if (!pvmw->pte)
297 				goto restart;
298 			goto next_pte;
299 		}
300 this_pte:
301 		if (check_pte(pvmw, 1))
302 			return true;
303 next_pte:
304 		do {
305 			pvmw->address += PAGE_SIZE;
306 			if (pvmw->address >= end)
307 				return not_found(pvmw);
308 			/* Did we cross page table boundary? */
309 			if ((pvmw->address & (PMD_SIZE - PAGE_SIZE)) == 0) {
310 				if (pvmw->ptl) {
311 					spin_unlock(pvmw->ptl);
312 					pvmw->ptl = NULL;
313 				}
314 				pte_unmap(pvmw->pte);
315 				pvmw->pte = NULL;
316 				pvmw->flags |= PVMW_PGTABLE_CROSSED;
317 				goto restart;
318 			}
319 			pvmw->pte++;
320 		} while (pte_none(ptep_get(pvmw->pte)));
321 
322 		if (!pvmw->ptl) {
323 			spin_lock(ptl);
324 			if (unlikely(!pmd_same(pmde, pmdp_get_lockless(pvmw->pmd)))) {
325 				pte_unmap_unlock(pvmw->pte, ptl);
326 				pvmw->pte = NULL;
327 				goto restart;
328 			}
329 			pvmw->ptl = ptl;
330 		}
331 		goto this_pte;
332 	} while (pvmw->address < end);
333 
334 	return false;
335 }
336 
337 #ifdef CONFIG_MEMORY_FAILURE
338 /**
339  * page_mapped_in_vma - check whether a page is really mapped in a VMA
340  * @page: the page to test
341  * @vma: the VMA to test
342  *
343  * Return: The address the page is mapped at if the page is in the range
344  * covered by the VMA and present in the page table.  If the page is
345  * outside the VMA or not present, returns -EFAULT.
346  * Only valid for normal file or anonymous VMAs.
347  */
page_mapped_in_vma(const struct page * page,struct vm_area_struct * vma)348 unsigned long page_mapped_in_vma(const struct page *page,
349 		struct vm_area_struct *vma)
350 {
351 	const struct folio *folio = page_folio(page);
352 	struct page_vma_mapped_walk pvmw = {
353 		.pfn = page_to_pfn(page),
354 		.nr_pages = 1,
355 		.vma = vma,
356 		.flags = PVMW_SYNC,
357 	};
358 
359 	pvmw.address = vma_address(vma, page_pgoff(folio, page), 1);
360 	if (pvmw.address == -EFAULT)
361 		goto out;
362 	if (!page_vma_mapped_walk(&pvmw))
363 		return -EFAULT;
364 	page_vma_mapped_walk_done(&pvmw);
365 out:
366 	return pvmw.address;
367 }
368 #endif
369