1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (C) 2023 ARM Ltd.
4 */
5
6#include <linux/mm.h>
7#include <linux/efi.h>
8#include <linux/export.h>
9#include <asm/tlbflush.h>
10
11static inline bool mm_is_user(struct mm_struct *mm)
12{
13 /*
14 * Don't attempt to apply the contig bit to kernel mappings, because
15 * dynamically adding/removing the contig bit can cause page faults.
16 * These racing faults are ok for user space, since they get serialized
17 * on the PTL. But kernel mappings can't tolerate faults.
18 */
19 if (unlikely(mm_is_efi(mm)))
20 return false;
21 return mm != &init_mm;
22}
23
24static inline pte_t *contpte_align_down(pte_t *ptep)
25{
26 return PTR_ALIGN_DOWN(ptep, sizeof(*ptep) * CONT_PTES);
27}
28
29static void contpte_try_unfold_partial(struct mm_struct *mm, unsigned long addr,
30 pte_t *ptep, unsigned int nr)
31{
32 /*
33 * Unfold any partially covered contpte block at the beginning and end
34 * of the range.
35 */
36
37 if (ptep != contpte_align_down(ptep) || nr < CONT_PTES)
38 contpte_try_unfold(mm, addr, ptep, __ptep_get(ptep));
39
40 if (ptep + nr != contpte_align_down(ptep: ptep + nr)) {
41 unsigned long last_addr = addr + PAGE_SIZE * (nr - 1);
42 pte_t *last_ptep = ptep + nr - 1;
43
44 contpte_try_unfold(mm, last_addr, last_ptep,
45 __ptep_get(last_ptep));
46 }
47}
48
49static void contpte_convert(struct mm_struct *mm, unsigned long addr,
50 pte_t *ptep, pte_t pte)
51{
52 struct vm_area_struct vma = TLB_FLUSH_VMA(mm, 0);
53 unsigned long start_addr;
54 pte_t *start_ptep;
55 int i;
56
57 start_ptep = ptep = contpte_align_down(ptep);
58 start_addr = addr = ALIGN_DOWN(addr, CONT_PTE_SIZE);
59 pte = pfn_pte(ALIGN_DOWN(pte_pfn(pte), CONT_PTES), pte_pgprot(pte));
60
61 for (i = 0; i < CONT_PTES; i++, ptep++, addr += PAGE_SIZE) {
62 pte_t ptent = __ptep_get_and_clear(mm, addr, ptep);
63
64 if (pte_dirty(pte: ptent))
65 pte = pte_mkdirty(pte);
66
67 if (pte_young(pte: ptent))
68 pte = pte_mkyoung(pte);
69 }
70
71 __flush_tlb_range(&vma, start_addr, addr, PAGE_SIZE, true, 3);
72
73 __set_ptes(mm, start_addr, start_ptep, pte, CONT_PTES);
74}
75
76void __contpte_try_fold(struct mm_struct *mm, unsigned long addr,
77 pte_t *ptep, pte_t pte)
78{
79 /*
80 * We have already checked that the virtual and pysical addresses are
81 * correctly aligned for a contpte mapping in contpte_try_fold() so the
82 * remaining checks are to ensure that the contpte range is fully
83 * covered by a single folio, and ensure that all the ptes are valid
84 * with contiguous PFNs and matching prots. We ignore the state of the
85 * access and dirty bits for the purpose of deciding if its a contiguous
86 * range; the folding process will generate a single contpte entry which
87 * has a single access and dirty bit. Those 2 bits are the logical OR of
88 * their respective bits in the constituent pte entries. In order to
89 * ensure the contpte range is covered by a single folio, we must
90 * recover the folio from the pfn, but special mappings don't have a
91 * folio backing them. Fortunately contpte_try_fold() already checked
92 * that the pte is not special - we never try to fold special mappings.
93 * Note we can't use vm_normal_page() for this since we don't have the
94 * vma.
95 */
96
97 unsigned long folio_start, folio_end;
98 unsigned long cont_start, cont_end;
99 pte_t expected_pte, subpte;
100 struct folio *folio;
101 struct page *page;
102 unsigned long pfn;
103 pte_t *orig_ptep;
104 pgprot_t prot;
105
106 int i;
107
108 if (!mm_is_user(mm))
109 return;
110
111 page = pte_page(pte);
112 folio = page_folio(page);
113 folio_start = addr - (page - &folio->page) * PAGE_SIZE;
114 folio_end = folio_start + folio_nr_pages(folio) * PAGE_SIZE;
115 cont_start = ALIGN_DOWN(addr, CONT_PTE_SIZE);
116 cont_end = cont_start + CONT_PTE_SIZE;
117
118 if (folio_start > cont_start || folio_end < cont_end)
119 return;
120
121 pfn = ALIGN_DOWN(pte_pfn(pte), CONT_PTES);
122 prot = pte_pgprot(pte_mkold(pte_mkclean(pte)));
123 expected_pte = pfn_pte(page_nr: pfn, pgprot: prot);
124 orig_ptep = ptep;
125 ptep = contpte_align_down(ptep);
126
127 for (i = 0; i < CONT_PTES; i++) {
128 subpte = pte_mkold(pte: pte_mkclean(__ptep_get(ptep)));
129 if (!pte_same(a: subpte, b: expected_pte))
130 return;
131 expected_pte = pte_advance_pfn(pte: expected_pte, nr: 1);
132 ptep++;
133 }
134
135 pte = pte_mkcont(pte);
136 contpte_convert(mm, addr, ptep: orig_ptep, pte);
137}
138EXPORT_SYMBOL_GPL(__contpte_try_fold);
139
140void __contpte_try_unfold(struct mm_struct *mm, unsigned long addr,
141 pte_t *ptep, pte_t pte)
142{
143 /*
144 * We have already checked that the ptes are contiguous in
145 * contpte_try_unfold(), so just check that the mm is user space.
146 */
147 if (!mm_is_user(mm))
148 return;
149
150 pte = pte_mknoncont(pte);
151 contpte_convert(mm, addr, ptep, pte);
152}
153EXPORT_SYMBOL_GPL(__contpte_try_unfold);
154
155pte_t contpte_ptep_get(pte_t *ptep, pte_t orig_pte)
156{
157 /*
158 * Gather access/dirty bits, which may be populated in any of the ptes
159 * of the contig range. We are guaranteed to be holding the PTL, so any
160 * contiguous range cannot be unfolded or otherwise modified under our
161 * feet.
162 */
163
164 pte_t pte;
165 int i;
166
167 ptep = contpte_align_down(ptep);
168
169 for (i = 0; i < CONT_PTES; i++, ptep++) {
170 pte = __ptep_get(ptep);
171
172 if (pte_dirty(pte))
173 orig_pte = pte_mkdirty(pte: orig_pte);
174
175 if (pte_young(pte))
176 orig_pte = pte_mkyoung(pte: orig_pte);
177 }
178
179 return orig_pte;
180}
181EXPORT_SYMBOL_GPL(contpte_ptep_get);
182
183pte_t contpte_ptep_get_lockless(pte_t *orig_ptep)
184{
185 /*
186 * The ptep_get_lockless() API requires us to read and return *orig_ptep
187 * so that it is self-consistent, without the PTL held, so we may be
188 * racing with other threads modifying the pte. Usually a READ_ONCE()
189 * would suffice, but for the contpte case, we also need to gather the
190 * access and dirty bits from across all ptes in the contiguous block,
191 * and we can't read all of those neighbouring ptes atomically, so any
192 * contiguous range may be unfolded/modified/refolded under our feet.
193 * Therefore we ensure we read a _consistent_ contpte range by checking
194 * that all ptes in the range are valid and have CONT_PTE set, that all
195 * pfns are contiguous and that all pgprots are the same (ignoring
196 * access/dirty). If we find a pte that is not consistent, then we must
197 * be racing with an update so start again. If the target pte does not
198 * have CONT_PTE set then that is considered consistent on its own
199 * because it is not part of a contpte range.
200 */
201
202 pgprot_t orig_prot;
203 unsigned long pfn;
204 pte_t orig_pte;
205 pgprot_t prot;
206 pte_t *ptep;
207 pte_t pte;
208 int i;
209
210retry:
211 orig_pte = __ptep_get(orig_ptep);
212
213 if (!pte_valid_cont(orig_pte))
214 return orig_pte;
215
216 orig_prot = pte_pgprot(pte_mkold(pte_mkclean(orig_pte)));
217 ptep = contpte_align_down(ptep: orig_ptep);
218 pfn = pte_pfn(pte: orig_pte) - (orig_ptep - ptep);
219
220 for (i = 0; i < CONT_PTES; i++, ptep++, pfn++) {
221 pte = __ptep_get(ptep);
222 prot = pte_pgprot(pte_mkold(pte_mkclean(pte)));
223
224 if (!pte_valid_cont(pte) ||
225 pte_pfn(pte) != pfn ||
226 pgprot_val(prot) != pgprot_val(orig_prot))
227 goto retry;
228
229 if (pte_dirty(pte))
230 orig_pte = pte_mkdirty(pte: orig_pte);
231
232 if (pte_young(pte))
233 orig_pte = pte_mkyoung(pte: orig_pte);
234 }
235
236 return orig_pte;
237}
238EXPORT_SYMBOL_GPL(contpte_ptep_get_lockless);
239
240void contpte_set_ptes(struct mm_struct *mm, unsigned long addr,
241 pte_t *ptep, pte_t pte, unsigned int nr)
242{
243 unsigned long next;
244 unsigned long end;
245 unsigned long pfn;
246 pgprot_t prot;
247
248 /*
249 * The set_ptes() spec guarantees that when nr > 1, the initial state of
250 * all ptes is not-present. Therefore we never need to unfold or
251 * otherwise invalidate a range before we set the new ptes.
252 * contpte_set_ptes() should never be called for nr < 2.
253 */
254 VM_WARN_ON(nr == 1);
255
256 if (!mm_is_user(mm))
257 return __set_ptes(mm, addr, ptep, pte, nr);
258
259 end = addr + (nr << PAGE_SHIFT);
260 pfn = pte_pfn(pte);
261 prot = pte_pgprot(pte);
262
263 do {
264 next = pte_cont_addr_end(addr, end);
265 nr = (next - addr) >> PAGE_SHIFT;
266 pte = pfn_pte(page_nr: pfn, pgprot: prot);
267
268 if (((addr | next | (pfn << PAGE_SHIFT)) & ~CONT_PTE_MASK) == 0)
269 pte = pte_mkcont(pte);
270 else
271 pte = pte_mknoncont(pte);
272
273 __set_ptes(mm, addr, ptep, pte, nr);
274
275 addr = next;
276 ptep += nr;
277 pfn += nr;
278
279 } while (addr != end);
280}
281EXPORT_SYMBOL_GPL(contpte_set_ptes);
282
283void contpte_clear_full_ptes(struct mm_struct *mm, unsigned long addr,
284 pte_t *ptep, unsigned int nr, int full)
285{
286 contpte_try_unfold_partial(mm, addr, ptep, nr);
287 __clear_full_ptes(mm, addr, ptep, nr, full);
288}
289EXPORT_SYMBOL_GPL(contpte_clear_full_ptes);
290
291pte_t contpte_get_and_clear_full_ptes(struct mm_struct *mm,
292 unsigned long addr, pte_t *ptep,
293 unsigned int nr, int full)
294{
295 contpte_try_unfold_partial(mm, addr, ptep, nr);
296 return __get_and_clear_full_ptes(mm, addr, ptep, nr, full);
297}
298EXPORT_SYMBOL_GPL(contpte_get_and_clear_full_ptes);
299
300int contpte_ptep_test_and_clear_young(struct vm_area_struct *vma,
301 unsigned long addr, pte_t *ptep)
302{
303 /*
304 * ptep_clear_flush_young() technically requires us to clear the access
305 * flag for a _single_ pte. However, the core-mm code actually tracks
306 * access/dirty per folio, not per page. And since we only create a
307 * contig range when the range is covered by a single folio, we can get
308 * away with clearing young for the whole contig range here, so we avoid
309 * having to unfold.
310 */
311
312 int young = 0;
313 int i;
314
315 ptep = contpte_align_down(ptep);
316 addr = ALIGN_DOWN(addr, CONT_PTE_SIZE);
317
318 for (i = 0; i < CONT_PTES; i++, ptep++, addr += PAGE_SIZE)
319 young |= __ptep_test_and_clear_young(vma, addr, ptep);
320
321 return young;
322}
323EXPORT_SYMBOL_GPL(contpte_ptep_test_and_clear_young);
324
325int contpte_ptep_clear_flush_young(struct vm_area_struct *vma,
326 unsigned long addr, pte_t *ptep)
327{
328 int young;
329
330 young = contpte_ptep_test_and_clear_young(vma, addr, ptep);
331
332 if (young) {
333 /*
334 * See comment in __ptep_clear_flush_young(); same rationale for
335 * eliding the trailing DSB applies here.
336 */
337 addr = ALIGN_DOWN(addr, CONT_PTE_SIZE);
338 __flush_tlb_range_nosync(vma, addr, addr + CONT_PTE_SIZE,
339 PAGE_SIZE, true, 3);
340 }
341
342 return young;
343}
344EXPORT_SYMBOL_GPL(contpte_ptep_clear_flush_young);
345
346void contpte_wrprotect_ptes(struct mm_struct *mm, unsigned long addr,
347 pte_t *ptep, unsigned int nr)
348{
349 /*
350 * If wrprotecting an entire contig range, we can avoid unfolding. Just
351 * set wrprotect and wait for the later mmu_gather flush to invalidate
352 * the tlb. Until the flush, the page may or may not be wrprotected.
353 * After the flush, it is guaranteed wrprotected. If it's a partial
354 * range though, we must unfold, because we can't have a case where
355 * CONT_PTE is set but wrprotect applies to a subset of the PTEs; this
356 * would cause it to continue to be unpredictable after the flush.
357 */
358
359 contpte_try_unfold_partial(mm, addr, ptep, nr);
360 __wrprotect_ptes(mm, addr, ptep, nr);
361}
362EXPORT_SYMBOL_GPL(contpte_wrprotect_ptes);
363
364int contpte_ptep_set_access_flags(struct vm_area_struct *vma,
365 unsigned long addr, pte_t *ptep,
366 pte_t entry, int dirty)
367{
368 unsigned long start_addr;
369 pte_t orig_pte;
370 int i;
371
372 /*
373 * Gather the access/dirty bits for the contiguous range. If nothing has
374 * changed, its a noop.
375 */
376 orig_pte = pte_mknoncont(ptep_get(ptep));
377 if (pte_val(pte: orig_pte) == pte_val(pte: entry))
378 return 0;
379
380 /*
381 * We can fix up access/dirty bits without having to unfold the contig
382 * range. But if the write bit is changing, we must unfold.
383 */
384 if (pte_write(pte: orig_pte) == pte_write(pte: entry)) {
385 /*
386 * For HW access management, we technically only need to update
387 * the flag on a single pte in the range. But for SW access
388 * management, we need to update all the ptes to prevent extra
389 * faults. Avoid per-page tlb flush in __ptep_set_access_flags()
390 * and instead flush the whole range at the end.
391 */
392 ptep = contpte_align_down(ptep);
393 start_addr = addr = ALIGN_DOWN(addr, CONT_PTE_SIZE);
394
395 for (i = 0; i < CONT_PTES; i++, ptep++, addr += PAGE_SIZE)
396 __ptep_set_access_flags(vma, addr, ptep, entry, 0);
397
398 if (dirty)
399 __flush_tlb_range(vma, start_addr, addr,
400 PAGE_SIZE, true, 3);
401 } else {
402 __contpte_try_unfold(vma->vm_mm, addr, ptep, orig_pte);
403 __ptep_set_access_flags(vma, addr, ptep, entry, dirty);
404 }
405
406 return 1;
407}
408EXPORT_SYMBOL_GPL(contpte_ptep_set_access_flags);
409

source code of linux/arch/arm64/mm/contpte.c