]>
Commit | Line | Data |
---|---|---|
b2441318 | 1 | // SPDX-License-Identifier: GPL-2.0 |
df077ac4 DM |
2 | /* |
3 | * Lockless get_user_pages_fast for sparc, cribbed from powerpc | |
4 | * | |
5 | * Copyright (C) 2008 Nick Piggin | |
6 | * Copyright (C) 2008 Novell Inc. | |
7 | */ | |
8 | ||
9 | #include <linux/sched.h> | |
10 | #include <linux/mm.h> | |
11 | #include <linux/vmstat.h> | |
12 | #include <linux/pagemap.h> | |
13 | #include <linux/rwsem.h> | |
14 | #include <asm/pgtable.h> | |
15 | ||
16 | /* | |
17 | * The performance critical leaf functions are made noinline otherwise gcc | |
18 | * inlines everything into a single function which results in too much | |
19 | * register pressure. | |
20 | */ | |
21 | static noinline int gup_pte_range(pmd_t pmd, unsigned long addr, | |
22 | unsigned long end, int write, struct page **pages, int *nr) | |
23 | { | |
24 | unsigned long mask, result; | |
25 | pte_t *ptep; | |
26 | ||
27 | if (tlb_type == hypervisor) { | |
28 | result = _PAGE_PRESENT_4V|_PAGE_P_4V; | |
29 | if (write) | |
30 | result |= _PAGE_WRITE_4V; | |
31 | } else { | |
32 | result = _PAGE_PRESENT_4U|_PAGE_P_4U; | |
33 | if (write) | |
34 | result |= _PAGE_WRITE_4U; | |
35 | } | |
36 | mask = result | _PAGE_SPECIAL; | |
37 | ||
38 | ptep = pte_offset_kernel(&pmd, addr); | |
39 | do { | |
40 | struct page *page, *head; | |
41 | pte_t pte = *ptep; | |
42 | ||
43 | if ((pte_val(pte) & mask) != result) | |
44 | return 0; | |
45 | VM_BUG_ON(!pfn_valid(pte_pfn(pte))); | |
46 | ||
47 | /* The hugepage case is simplified on sparc64 because | |
48 | * we encode the sub-page pfn offsets into the | |
49 | * hugepage PTEs. We could optimize this in the future | |
50 | * use page_cache_add_speculative() for the hugepage case. | |
51 | */ | |
52 | page = pte_page(pte); | |
53 | head = compound_head(page); | |
54 | if (!page_cache_get_speculative(head)) | |
55 | return 0; | |
56 | if (unlikely(pte_val(pte) != pte_val(*ptep))) { | |
57 | put_page(head); | |
58 | return 0; | |
59 | } | |
60 | ||
61 | pages[*nr] = page; | |
62 | (*nr)++; | |
63 | } while (ptep++, addr += PAGE_SIZE, addr != end); | |
64 | ||
65 | return 1; | |
66 | } | |
67 | ||
89a77915 DM |
68 | static int gup_huge_pmd(pmd_t *pmdp, pmd_t pmd, unsigned long addr, |
69 | unsigned long end, int write, struct page **pages, | |
70 | int *nr) | |
71 | { | |
ddc58f27 | 72 | struct page *head, *page; |
89a77915 DM |
73 | int refs; |
74 | ||
04df419d | 75 | if (!(pmd_val(pmd) & _PAGE_VALID)) |
a7b9403f DM |
76 | return 0; |
77 | ||
f6f37321 | 78 | if (write && !pmd_write(pmd)) |
89a77915 DM |
79 | return 0; |
80 | ||
81 | refs = 0; | |
dbd2667a NG |
82 | page = pmd_page(pmd) + ((addr & ~PMD_MASK) >> PAGE_SHIFT); |
83 | head = compound_head(page); | |
89a77915 DM |
84 | do { |
85 | VM_BUG_ON(compound_head(page) != head); | |
86 | pages[*nr] = page; | |
87 | (*nr)++; | |
88 | page++; | |
89 | refs++; | |
90 | } while (addr += PAGE_SIZE, addr != end); | |
91 | ||
92 | if (!page_cache_add_speculative(head, refs)) { | |
93 | *nr -= refs; | |
94 | return 0; | |
95 | } | |
96 | ||
97 | if (unlikely(pmd_val(pmd) != pmd_val(*pmdp))) { | |
98 | *nr -= refs; | |
99 | while (refs--) | |
100 | put_page(head); | |
101 | return 0; | |
102 | } | |
103 | ||
89a77915 DM |
104 | return 1; |
105 | } | |
106 | ||
44382b01 NG |
107 | static int gup_huge_pud(pud_t *pudp, pud_t pud, unsigned long addr, |
108 | unsigned long end, int write, struct page **pages, | |
109 | int *nr) | |
110 | { | |
111 | struct page *head, *page; | |
112 | int refs; | |
113 | ||
114 | if (!(pud_val(pud) & _PAGE_VALID)) | |
115 | return 0; | |
116 | ||
f6f37321 | 117 | if (write && !pud_write(pud)) |
44382b01 NG |
118 | return 0; |
119 | ||
120 | refs = 0; | |
121 | page = pud_page(pud) + ((addr & ~PUD_MASK) >> PAGE_SHIFT); | |
122 | head = compound_head(page); | |
123 | do { | |
124 | VM_BUG_ON(compound_head(page) != head); | |
125 | pages[*nr] = page; | |
126 | (*nr)++; | |
127 | page++; | |
128 | refs++; | |
129 | } while (addr += PAGE_SIZE, addr != end); | |
130 | ||
131 | if (!page_cache_add_speculative(head, refs)) { | |
132 | *nr -= refs; | |
133 | return 0; | |
134 | } | |
135 | ||
136 | if (unlikely(pud_val(pud) != pud_val(*pudp))) { | |
137 | *nr -= refs; | |
138 | while (refs--) | |
139 | put_page(head); | |
140 | return 0; | |
141 | } | |
142 | ||
143 | return 1; | |
144 | } | |
145 | ||
df077ac4 DM |
146 | static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end, |
147 | int write, struct page **pages, int *nr) | |
148 | { | |
149 | unsigned long next; | |
150 | pmd_t *pmdp; | |
151 | ||
152 | pmdp = pmd_offset(&pud, addr); | |
153 | do { | |
154 | pmd_t pmd = *pmdp; | |
155 | ||
156 | next = pmd_addr_end(addr, end); | |
99f1bc01 | 157 | if (pmd_none(pmd)) |
df077ac4 | 158 | return 0; |
89a77915 DM |
159 | if (unlikely(pmd_large(pmd))) { |
160 | if (!gup_huge_pmd(pmdp, pmd, addr, next, | |
161 | write, pages, nr)) | |
162 | return 0; | |
163 | } else if (!gup_pte_range(pmd, addr, next, write, | |
164 | pages, nr)) | |
df077ac4 DM |
165 | return 0; |
166 | } while (pmdp++, addr = next, addr != end); | |
167 | ||
168 | return 1; | |
169 | } | |
170 | ||
171 | static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end, | |
172 | int write, struct page **pages, int *nr) | |
173 | { | |
174 | unsigned long next; | |
175 | pud_t *pudp; | |
176 | ||
177 | pudp = pud_offset(&pgd, addr); | |
178 | do { | |
179 | pud_t pud = *pudp; | |
180 | ||
181 | next = pud_addr_end(addr, end); | |
182 | if (pud_none(pud)) | |
183 | return 0; | |
44382b01 NG |
184 | if (unlikely(pud_large(pud))) { |
185 | if (!gup_huge_pud(pudp, pud, addr, next, | |
186 | write, pages, nr)) | |
187 | return 0; | |
188 | } else if (!gup_pmd_range(pud, addr, next, write, pages, nr)) | |
df077ac4 DM |
189 | return 0; |
190 | } while (pudp++, addr = next, addr != end); | |
191 | ||
192 | return 1; | |
193 | } | |
194 | ||
06090e8e DM |
195 | int __get_user_pages_fast(unsigned long start, int nr_pages, int write, |
196 | struct page **pages) | |
197 | { | |
198 | struct mm_struct *mm = current->mm; | |
199 | unsigned long addr, len, end; | |
200 | unsigned long next, flags; | |
201 | pgd_t *pgdp; | |
202 | int nr = 0; | |
203 | ||
204 | start &= PAGE_MASK; | |
205 | addr = start; | |
206 | len = (unsigned long) nr_pages << PAGE_SHIFT; | |
207 | end = start + len; | |
208 | ||
209 | local_irq_save(flags); | |
210 | pgdp = pgd_offset(mm, addr); | |
211 | do { | |
212 | pgd_t pgd = *pgdp; | |
213 | ||
214 | next = pgd_addr_end(addr, end); | |
215 | if (pgd_none(pgd)) | |
216 | break; | |
217 | if (!gup_pud_range(pgd, addr, next, write, pages, &nr)) | |
218 | break; | |
219 | } while (pgdp++, addr = next, addr != end); | |
220 | local_irq_restore(flags); | |
221 | ||
222 | return nr; | |
223 | } | |
224 | ||
df077ac4 DM |
225 | int get_user_pages_fast(unsigned long start, int nr_pages, int write, |
226 | struct page **pages) | |
227 | { | |
228 | struct mm_struct *mm = current->mm; | |
229 | unsigned long addr, len, end; | |
230 | unsigned long next; | |
231 | pgd_t *pgdp; | |
232 | int nr = 0; | |
233 | ||
234 | start &= PAGE_MASK; | |
235 | addr = start; | |
236 | len = (unsigned long) nr_pages << PAGE_SHIFT; | |
237 | end = start + len; | |
238 | ||
239 | /* | |
240 | * XXX: batch / limit 'nr', to avoid large irq off latency | |
241 | * needs some instrumenting to determine the common sizes used by | |
242 | * important workloads (eg. DB2), and whether limiting the batch size | |
243 | * will decrease performance. | |
244 | * | |
245 | * It seems like we're in the clear for the moment. Direct-IO is | |
246 | * the main guy that batches up lots of get_user_pages, and even | |
247 | * they are limited to 64-at-a-time which is not so many. | |
248 | */ | |
249 | /* | |
250 | * This doesn't prevent pagetable teardown, but does prevent | |
251 | * the pagetables from being freed on sparc. | |
252 | * | |
253 | * So long as we atomically load page table pointers versus teardown, | |
254 | * we can follow the address down to the the page and take a ref on it. | |
255 | */ | |
256 | local_irq_disable(); | |
257 | ||
258 | pgdp = pgd_offset(mm, addr); | |
259 | do { | |
260 | pgd_t pgd = *pgdp; | |
261 | ||
262 | next = pgd_addr_end(addr, end); | |
263 | if (pgd_none(pgd)) | |
264 | goto slow; | |
265 | if (!gup_pud_range(pgd, addr, next, write, pages, &nr)) | |
266 | goto slow; | |
267 | } while (pgdp++, addr = next, addr != end); | |
268 | ||
269 | local_irq_enable(); | |
270 | ||
271 | VM_BUG_ON(nr != (end - start) >> PAGE_SHIFT); | |
272 | return nr; | |
273 | ||
274 | { | |
275 | int ret; | |
276 | ||
277 | slow: | |
278 | local_irq_enable(); | |
279 | ||
280 | /* Try to get the remaining pages with get_user_pages */ | |
281 | start += nr << PAGE_SHIFT; | |
282 | pages += nr; | |
283 | ||
d4edcf0d | 284 | ret = get_user_pages_unlocked(start, |
c164154f LS |
285 | (end - start) >> PAGE_SHIFT, pages, |
286 | write ? FOLL_WRITE : 0); | |
df077ac4 DM |
287 | |
288 | /* Have to be a bit careful with return values */ | |
289 | if (nr > 0) { | |
290 | if (ret < 0) | |
291 | ret = nr; | |
292 | else | |
293 | ret += nr; | |
294 | } | |
295 | ||
296 | return ret; | |
297 | } | |
298 | } |