1 // SPDX-License-Identifier: GPL-2.0 1 // SPDX-License-Identifier: GPL-2.0
2 /* 2 /*
3 * mm/pgtable-generic.c 3 * mm/pgtable-generic.c
4 * 4 *
5 * Generic pgtable methods declared in linux/ 5 * Generic pgtable methods declared in linux/pgtable.h
6 * 6 *
7 * Copyright (C) 2010 Linus Torvalds 7 * Copyright (C) 2010 Linus Torvalds
8 */ 8 */
9 9
10 #include <linux/pagemap.h> 10 #include <linux/pagemap.h>
11 #include <linux/hugetlb.h> 11 #include <linux/hugetlb.h>
12 #include <linux/pgtable.h> 12 #include <linux/pgtable.h>
13 #include <linux/swap.h> <<
14 #include <linux/swapops.h> <<
15 #include <linux/mm_inline.h> <<
16 #include <asm/pgalloc.h> <<
17 #include <asm/tlb.h> 13 #include <asm/tlb.h>
18 14
19 /* 15 /*
20 * If a p?d_bad entry is found while walking p 16 * If a p?d_bad entry is found while walking page tables, report
21 * the error, before resetting entry to p?d_no 17 * the error, before resetting entry to p?d_none. Usually (but
22 * very seldom) called out from the p?d_none_o 18 * very seldom) called out from the p?d_none_or_clear_bad macros.
23 */ 19 */
24 20
25 void pgd_clear_bad(pgd_t *pgd) 21 void pgd_clear_bad(pgd_t *pgd)
26 { 22 {
27 pgd_ERROR(*pgd); 23 pgd_ERROR(*pgd);
28 pgd_clear(pgd); 24 pgd_clear(pgd);
29 } 25 }
30 26
31 #ifndef __PAGETABLE_P4D_FOLDED 27 #ifndef __PAGETABLE_P4D_FOLDED
32 void p4d_clear_bad(p4d_t *p4d) 28 void p4d_clear_bad(p4d_t *p4d)
33 { 29 {
34 p4d_ERROR(*p4d); 30 p4d_ERROR(*p4d);
35 p4d_clear(p4d); 31 p4d_clear(p4d);
36 } 32 }
37 #endif 33 #endif
38 34
39 #ifndef __PAGETABLE_PUD_FOLDED 35 #ifndef __PAGETABLE_PUD_FOLDED
40 void pud_clear_bad(pud_t *pud) 36 void pud_clear_bad(pud_t *pud)
41 { 37 {
42 pud_ERROR(*pud); 38 pud_ERROR(*pud);
43 pud_clear(pud); 39 pud_clear(pud);
44 } 40 }
45 #endif 41 #endif
46 42
47 /* 43 /*
48 * Note that the pmd variant below can't be st 44 * Note that the pmd variant below can't be stub'ed out just as for p4d/pud
49 * above. pmd folding is special and typically 45 * above. pmd folding is special and typically pmd_* macros refer to upper
50 * level even when folded 46 * level even when folded
51 */ 47 */
52 void pmd_clear_bad(pmd_t *pmd) 48 void pmd_clear_bad(pmd_t *pmd)
53 { 49 {
54 pmd_ERROR(*pmd); 50 pmd_ERROR(*pmd);
55 pmd_clear(pmd); 51 pmd_clear(pmd);
56 } 52 }
57 53
58 #ifndef __HAVE_ARCH_PTEP_SET_ACCESS_FLAGS 54 #ifndef __HAVE_ARCH_PTEP_SET_ACCESS_FLAGS
59 /* 55 /*
60 * Only sets the access flags (dirty, accessed 56 * Only sets the access flags (dirty, accessed), as well as write
61 * permission. Furthermore, we know it always 57 * permission. Furthermore, we know it always gets set to a "more
62 * permissive" setting, which allows most arch 58 * permissive" setting, which allows most architectures to optimize
63 * this. We return whether the PTE actually ch 59 * this. We return whether the PTE actually changed, which in turn
64 * instructs the caller to do things like upda 60 * instructs the caller to do things like update__mmu_cache. This
65 * used to be done in the caller, but sparc ne 61 * used to be done in the caller, but sparc needs minor faults to
66 * force that call on sun4c so we changed this 62 * force that call on sun4c so we changed this macro slightly
67 */ 63 */
68 int ptep_set_access_flags(struct vm_area_struc 64 int ptep_set_access_flags(struct vm_area_struct *vma,
69 unsigned long addres 65 unsigned long address, pte_t *ptep,
70 pte_t entry, int dir 66 pte_t entry, int dirty)
71 { 67 {
72 int changed = !pte_same(ptep_get(ptep) !! 68 int changed = !pte_same(*ptep, entry);
73 if (changed) { 69 if (changed) {
74 set_pte_at(vma->vm_mm, address 70 set_pte_at(vma->vm_mm, address, ptep, entry);
75 flush_tlb_fix_spurious_fault(v !! 71 flush_tlb_fix_spurious_fault(vma, address);
76 } 72 }
77 return changed; 73 return changed;
78 } 74 }
79 #endif 75 #endif
80 76
81 #ifndef __HAVE_ARCH_PTEP_CLEAR_YOUNG_FLUSH 77 #ifndef __HAVE_ARCH_PTEP_CLEAR_YOUNG_FLUSH
82 int ptep_clear_flush_young(struct vm_area_stru 78 int ptep_clear_flush_young(struct vm_area_struct *vma,
83 unsigned long addre 79 unsigned long address, pte_t *ptep)
84 { 80 {
85 int young; 81 int young;
86 young = ptep_test_and_clear_young(vma, 82 young = ptep_test_and_clear_young(vma, address, ptep);
87 if (young) 83 if (young)
88 flush_tlb_page(vma, address); 84 flush_tlb_page(vma, address);
89 return young; 85 return young;
90 } 86 }
91 #endif 87 #endif
92 88
93 #ifndef __HAVE_ARCH_PTEP_CLEAR_FLUSH 89 #ifndef __HAVE_ARCH_PTEP_CLEAR_FLUSH
94 pte_t ptep_clear_flush(struct vm_area_struct * 90 pte_t ptep_clear_flush(struct vm_area_struct *vma, unsigned long address,
95 pte_t *ptep) 91 pte_t *ptep)
96 { 92 {
97 struct mm_struct *mm = (vma)->vm_mm; 93 struct mm_struct *mm = (vma)->vm_mm;
98 pte_t pte; 94 pte_t pte;
99 pte = ptep_get_and_clear(mm, address, 95 pte = ptep_get_and_clear(mm, address, ptep);
100 if (pte_accessible(mm, pte)) 96 if (pte_accessible(mm, pte))
101 flush_tlb_page(vma, address); 97 flush_tlb_page(vma, address);
102 return pte; 98 return pte;
103 } 99 }
104 #endif 100 #endif
105 101
106 #ifdef CONFIG_TRANSPARENT_HUGEPAGE 102 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
107 103
108 #ifndef __HAVE_ARCH_PMDP_SET_ACCESS_FLAGS 104 #ifndef __HAVE_ARCH_PMDP_SET_ACCESS_FLAGS
109 int pmdp_set_access_flags(struct vm_area_struc 105 int pmdp_set_access_flags(struct vm_area_struct *vma,
110 unsigned long addres 106 unsigned long address, pmd_t *pmdp,
111 pmd_t entry, int dir 107 pmd_t entry, int dirty)
112 { 108 {
113 int changed = !pmd_same(*pmdp, entry); 109 int changed = !pmd_same(*pmdp, entry);
114 VM_BUG_ON(address & ~HPAGE_PMD_MASK); 110 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
115 if (changed) { 111 if (changed) {
116 set_pmd_at(vma->vm_mm, address 112 set_pmd_at(vma->vm_mm, address, pmdp, entry);
117 flush_pmd_tlb_range(vma, addre 113 flush_pmd_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
118 } 114 }
119 return changed; 115 return changed;
120 } 116 }
121 #endif 117 #endif
122 118
123 #ifndef __HAVE_ARCH_PMDP_CLEAR_YOUNG_FLUSH 119 #ifndef __HAVE_ARCH_PMDP_CLEAR_YOUNG_FLUSH
124 int pmdp_clear_flush_young(struct vm_area_stru 120 int pmdp_clear_flush_young(struct vm_area_struct *vma,
125 unsigned long addre 121 unsigned long address, pmd_t *pmdp)
126 { 122 {
127 int young; 123 int young;
128 VM_BUG_ON(address & ~HPAGE_PMD_MASK); 124 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
129 young = pmdp_test_and_clear_young(vma, 125 young = pmdp_test_and_clear_young(vma, address, pmdp);
130 if (young) 126 if (young)
131 flush_pmd_tlb_range(vma, addre 127 flush_pmd_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
132 return young; 128 return young;
133 } 129 }
134 #endif 130 #endif
135 131
136 #ifndef __HAVE_ARCH_PMDP_HUGE_CLEAR_FLUSH 132 #ifndef __HAVE_ARCH_PMDP_HUGE_CLEAR_FLUSH
137 pmd_t pmdp_huge_clear_flush(struct vm_area_str 133 pmd_t pmdp_huge_clear_flush(struct vm_area_struct *vma, unsigned long address,
138 pmd_t *pmdp) 134 pmd_t *pmdp)
139 { 135 {
140 pmd_t pmd; 136 pmd_t pmd;
141 VM_BUG_ON(address & ~HPAGE_PMD_MASK); 137 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
142 VM_BUG_ON(pmd_present(*pmdp) && !pmd_t 138 VM_BUG_ON(pmd_present(*pmdp) && !pmd_trans_huge(*pmdp) &&
143 !pmd_devmap(*pmdp)) 139 !pmd_devmap(*pmdp));
144 pmd = pmdp_huge_get_and_clear(vma->vm_ 140 pmd = pmdp_huge_get_and_clear(vma->vm_mm, address, pmdp);
145 flush_pmd_tlb_range(vma, address, addr 141 flush_pmd_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
146 return pmd; 142 return pmd;
147 } 143 }
148 144
149 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_P 145 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
150 pud_t pudp_huge_clear_flush(struct vm_area_str 146 pud_t pudp_huge_clear_flush(struct vm_area_struct *vma, unsigned long address,
151 pud_t *pudp) 147 pud_t *pudp)
152 { 148 {
153 pud_t pud; 149 pud_t pud;
154 150
155 VM_BUG_ON(address & ~HPAGE_PUD_MASK); 151 VM_BUG_ON(address & ~HPAGE_PUD_MASK);
156 VM_BUG_ON(!pud_trans_huge(*pudp) && !p 152 VM_BUG_ON(!pud_trans_huge(*pudp) && !pud_devmap(*pudp));
157 pud = pudp_huge_get_and_clear(vma->vm_ 153 pud = pudp_huge_get_and_clear(vma->vm_mm, address, pudp);
158 flush_pud_tlb_range(vma, address, addr 154 flush_pud_tlb_range(vma, address, address + HPAGE_PUD_SIZE);
159 return pud; 155 return pud;
160 } 156 }
161 #endif 157 #endif
162 #endif 158 #endif
163 159
164 #ifndef __HAVE_ARCH_PGTABLE_DEPOSIT 160 #ifndef __HAVE_ARCH_PGTABLE_DEPOSIT
165 void pgtable_trans_huge_deposit(struct mm_stru 161 void pgtable_trans_huge_deposit(struct mm_struct *mm, pmd_t *pmdp,
166 pgtable_t pgta 162 pgtable_t pgtable)
167 { 163 {
168 assert_spin_locked(pmd_lockptr(mm, pmd 164 assert_spin_locked(pmd_lockptr(mm, pmdp));
169 165
170 /* FIFO */ 166 /* FIFO */
171 if (!pmd_huge_pte(mm, pmdp)) 167 if (!pmd_huge_pte(mm, pmdp))
172 INIT_LIST_HEAD(&pgtable->lru); 168 INIT_LIST_HEAD(&pgtable->lru);
173 else 169 else
174 list_add(&pgtable->lru, &pmd_h 170 list_add(&pgtable->lru, &pmd_huge_pte(mm, pmdp)->lru);
175 pmd_huge_pte(mm, pmdp) = pgtable; 171 pmd_huge_pte(mm, pmdp) = pgtable;
176 } 172 }
177 #endif 173 #endif
178 174
179 #ifndef __HAVE_ARCH_PGTABLE_WITHDRAW 175 #ifndef __HAVE_ARCH_PGTABLE_WITHDRAW
180 /* no "address" argument so destroys page colo 176 /* no "address" argument so destroys page coloring of some arch */
181 pgtable_t pgtable_trans_huge_withdraw(struct m 177 pgtable_t pgtable_trans_huge_withdraw(struct mm_struct *mm, pmd_t *pmdp)
182 { 178 {
183 pgtable_t pgtable; 179 pgtable_t pgtable;
184 180
185 assert_spin_locked(pmd_lockptr(mm, pmd 181 assert_spin_locked(pmd_lockptr(mm, pmdp));
186 182
187 /* FIFO */ 183 /* FIFO */
188 pgtable = pmd_huge_pte(mm, pmdp); 184 pgtable = pmd_huge_pte(mm, pmdp);
189 pmd_huge_pte(mm, pmdp) = list_first_en 185 pmd_huge_pte(mm, pmdp) = list_first_entry_or_null(&pgtable->lru,
190 186 struct page, lru);
191 if (pmd_huge_pte(mm, pmdp)) 187 if (pmd_huge_pte(mm, pmdp))
192 list_del(&pgtable->lru); 188 list_del(&pgtable->lru);
193 return pgtable; 189 return pgtable;
194 } 190 }
195 #endif 191 #endif
196 192
197 #ifndef __HAVE_ARCH_PMDP_INVALIDATE 193 #ifndef __HAVE_ARCH_PMDP_INVALIDATE
198 pmd_t pmdp_invalidate(struct vm_area_struct *v 194 pmd_t pmdp_invalidate(struct vm_area_struct *vma, unsigned long address,
199 pmd_t *pmdp) 195 pmd_t *pmdp)
200 { 196 {
201 VM_WARN_ON_ONCE(!pmd_present(*pmdp)); <<
202 pmd_t old = pmdp_establish(vma, addres 197 pmd_t old = pmdp_establish(vma, address, pmdp, pmd_mkinvalid(*pmdp));
203 flush_pmd_tlb_range(vma, address, addr 198 flush_pmd_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
204 return old; 199 return old;
205 } 200 }
206 #endif 201 #endif
207 202
208 #ifndef __HAVE_ARCH_PMDP_INVALIDATE_AD <<
209 pmd_t pmdp_invalidate_ad(struct vm_area_struct <<
210 pmd_t *pmdp) <<
211 { <<
212 VM_WARN_ON_ONCE(!pmd_present(*pmdp)); <<
213 return pmdp_invalidate(vma, address, p <<
214 } <<
215 #endif <<
216 <<
217 #ifndef pmdp_collapse_flush 203 #ifndef pmdp_collapse_flush
218 pmd_t pmdp_collapse_flush(struct vm_area_struc 204 pmd_t pmdp_collapse_flush(struct vm_area_struct *vma, unsigned long address,
219 pmd_t *pmdp) 205 pmd_t *pmdp)
220 { 206 {
221 /* 207 /*
222 * pmd and hugepage pte format are sam 208 * pmd and hugepage pte format are same. So we could
223 * use the same function. 209 * use the same function.
224 */ 210 */
225 pmd_t pmd; 211 pmd_t pmd;
226 212
227 VM_BUG_ON(address & ~HPAGE_PMD_MASK); 213 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
228 VM_BUG_ON(pmd_trans_huge(*pmdp)); 214 VM_BUG_ON(pmd_trans_huge(*pmdp));
229 pmd = pmdp_huge_get_and_clear(vma->vm_ 215 pmd = pmdp_huge_get_and_clear(vma->vm_mm, address, pmdp);
230 216
231 /* collapse entails shooting down ptes 217 /* collapse entails shooting down ptes not pmd */
232 flush_tlb_range(vma, address, address 218 flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
233 return pmd; 219 return pmd;
234 } 220 }
235 #endif 221 #endif
236 <<
237 /* arch define pte_free_defer in asm/pgalloc.h <<
238 #ifndef pte_free_defer <<
239 static void pte_free_now(struct rcu_head *head <<
240 { <<
241 struct page *page; <<
242 <<
243 page = container_of(head, struct page, <<
244 pte_free(NULL /* mm not passed and not <<
245 } <<
246 <<
247 void pte_free_defer(struct mm_struct *mm, pgta <<
248 { <<
249 struct page *page; <<
250 <<
251 page = pgtable; <<
252 call_rcu(&page->rcu_head, pte_free_now <<
253 } <<
254 #endif /* pte_free_defer */ <<
255 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */ 222 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
256 <<
257 #if defined(CONFIG_GUP_GET_PXX_LOW_HIGH) && \ <<
258 (defined(CONFIG_SMP) || defined(CONFIG <<
259 /* <<
260 * See the comment above ptep_get_lockless() i <<
261 * the barriers in pmdp_get_lockless() cannot <<
262 * pmd_high actually belongs with the value in <<
263 * off blocks the TLB flush between present up <<
264 * successful __pte_offset_map() points to a p <<
265 */ <<
266 static unsigned long pmdp_get_lockless_start(v <<
267 { <<
268 unsigned long irqflags; <<
269 <<
270 local_irq_save(irqflags); <<
271 return irqflags; <<
272 } <<
273 static void pmdp_get_lockless_end(unsigned lon <<
274 { <<
275 local_irq_restore(irqflags); <<
276 } <<
277 #else <<
278 static unsigned long pmdp_get_lockless_start(v <<
279 static void pmdp_get_lockless_end(unsigned lon <<
280 #endif <<
281 <<
282 pte_t *__pte_offset_map(pmd_t *pmd, unsigned l <<
283 { <<
284 unsigned long irqflags; <<
285 pmd_t pmdval; <<
286 <<
287 rcu_read_lock(); <<
288 irqflags = pmdp_get_lockless_start(); <<
289 pmdval = pmdp_get_lockless(pmd); <<
290 pmdp_get_lockless_end(irqflags); <<
291 <<
292 if (pmdvalp) <<
293 *pmdvalp = pmdval; <<
294 if (unlikely(pmd_none(pmdval) || is_pm <<
295 goto nomap; <<
296 if (unlikely(pmd_trans_huge(pmdval) || <<
297 goto nomap; <<
298 if (unlikely(pmd_bad(pmdval))) { <<
299 pmd_clear_bad(pmd); <<
300 goto nomap; <<
301 } <<
302 return __pte_map(&pmdval, addr); <<
303 nomap: <<
304 rcu_read_unlock(); <<
305 return NULL; <<
306 } <<
307 <<
308 pte_t *pte_offset_map_nolock(struct mm_struct <<
309 unsigned long add <<
310 { <<
311 pmd_t pmdval; <<
312 pte_t *pte; <<
313 <<
314 pte = __pte_offset_map(pmd, addr, &pmd <<
315 if (likely(pte)) <<
316 *ptlp = pte_lockptr(mm, &pmdva <<
317 return pte; <<
318 } <<
319 <<
320 /* <<
321 * pte_offset_map_lock(mm, pmd, addr, ptlp), a <<
322 * __pte_offset_map_lock() below, is usually c <<
323 * addr, reached by walking down the mm's pgd, <<
324 * holding mmap_lock or vma lock for read or f <<
325 * context, while holding file's i_mmap_lock o <<
326 * write). In a few cases, it may be used with <<
327 * copied to or constructed on the stack. <<
328 * <<
329 * When successful, it returns the pte pointer <<
330 * kmapped if necessary (when CONFIG_HIGHPTE), <<
331 * modification by software, with a pointer to <<
332 * configs mm->page_table_lock, in SPLIT_PTLOC <<
333 * struct page). pte_unmap_unlock(pte, ptl) t <<
334 * <<
335 * But it is unsuccessful, returning NULL with <<
336 * page table at *pmd: if, for example, the pa <<
337 * or replaced by the huge pmd of a THP. (Whe <<
338 * after acquiring the ptlock, and retried int <<
339 * page table can be safely removed or replace <<
340 * <<
341 * pte_offset_map(pmd, addr), and its internal <<
342 * just returns the pte pointer for addr, its <<
343 * or NULL if there is no page table at *pmd. <<
344 * page table, so cannot normally be used when <<
345 * or when entries read must be stable. But i <<
346 * that even when page table is racily removed <<
347 * and disconnected table. Until pte_unmap(pt <<
348 * afterwards. <<
349 * <<
350 * pte_offset_map_nolock(mm, pmd, addr, ptlp), <<
351 * but when successful, it also outputs a poin <<
352 * pte_offset_map_lock() does, but in this cas <<
353 * the caller to avoid a later pte_lockptr(mm, <<
354 * act on a changed *pmd: pte_offset_map_noloc <<
355 * pointer for the page table that it returns. <<
356 * recheck *pmd once the lock is taken; in pra <<
357 * either the mmap_lock for write, or pte_same <<
358 * <<
359 * Note that free_pgtables(), used after unmap <<
360 * exiting the whole mm, does not take page ta <<
361 * table, and may not use RCU at all: "outside <<
362 * pte_offset_map() and co once the vma is det <<
363 */ <<
364 pte_t *__pte_offset_map_lock(struct mm_struct <<
365 unsigned long add <<
366 { <<
367 spinlock_t *ptl; <<
368 pmd_t pmdval; <<
369 pte_t *pte; <<
370 again: <<
371 pte = __pte_offset_map(pmd, addr, &pmd <<
372 if (unlikely(!pte)) <<
373 return pte; <<
374 ptl = pte_lockptr(mm, &pmdval); <<
375 spin_lock(ptl); <<
376 if (likely(pmd_same(pmdval, pmdp_get_l <<
377 *ptlp = ptl; <<
378 return pte; <<
379 } <<
380 pte_unmap_unlock(pte, ptl); <<
381 goto again; <<
382 } <<
383 223
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.