]>
Commit | Line | Data |
---|---|---|
df077ac4 DM |
1 | /* |
2 | * Lockless get_user_pages_fast for sparc, cribbed from powerpc | |
3 | * | |
4 | * Copyright (C) 2008 Nick Piggin | |
5 | * Copyright (C) 2008 Novell Inc. | |
6 | */ | |
7 | ||
8 | #include <linux/sched.h> | |
9 | #include <linux/mm.h> | |
10 | #include <linux/vmstat.h> | |
11 | #include <linux/pagemap.h> | |
12 | #include <linux/rwsem.h> | |
13 | #include <asm/pgtable.h> | |
14 | ||
15 | /* | |
16 | * The performance critical leaf functions are made noinline otherwise gcc | |
17 | * inlines everything into a single function which results in too much | |
18 | * register pressure. | |
19 | */ | |
20 | static noinline int gup_pte_range(pmd_t pmd, unsigned long addr, | |
21 | unsigned long end, int write, struct page **pages, int *nr) | |
22 | { | |
23 | unsigned long mask, result; | |
24 | pte_t *ptep; | |
25 | ||
26 | if (tlb_type == hypervisor) { | |
27 | result = _PAGE_PRESENT_4V|_PAGE_P_4V; | |
28 | if (write) | |
29 | result |= _PAGE_WRITE_4V; | |
30 | } else { | |
31 | result = _PAGE_PRESENT_4U|_PAGE_P_4U; | |
32 | if (write) | |
33 | result |= _PAGE_WRITE_4U; | |
34 | } | |
35 | mask = result | _PAGE_SPECIAL; | |
36 | ||
37 | ptep = pte_offset_kernel(&pmd, addr); | |
38 | do { | |
39 | struct page *page, *head; | |
40 | pte_t pte = *ptep; | |
41 | ||
42 | if ((pte_val(pte) & mask) != result) | |
43 | return 0; | |
44 | VM_BUG_ON(!pfn_valid(pte_pfn(pte))); | |
45 | ||
46 | /* The hugepage case is simplified on sparc64 because | |
47 | * we encode the sub-page pfn offsets into the | |
48 | * hugepage PTEs. We could optimize this in the future | |
49 | * use page_cache_add_speculative() for the hugepage case. | |
50 | */ | |
51 | page = pte_page(pte); | |
52 | head = compound_head(page); | |
53 | if (!page_cache_get_speculative(head)) | |
54 | return 0; | |
55 | if (unlikely(pte_val(pte) != pte_val(*ptep))) { | |
56 | put_page(head); | |
57 | return 0; | |
58 | } | |
59 | ||
60 | pages[*nr] = page; | |
61 | (*nr)++; | |
62 | } while (ptep++, addr += PAGE_SIZE, addr != end); | |
63 | ||
64 | return 1; | |
65 | } | |
66 | ||
89a77915 DM |
67 | static int gup_huge_pmd(pmd_t *pmdp, pmd_t pmd, unsigned long addr, |
68 | unsigned long end, int write, struct page **pages, | |
69 | int *nr) | |
70 | { | |
ddc58f27 | 71 | struct page *head, *page; |
89a77915 DM |
72 | int refs; |
73 | ||
04df419d | 74 | if (!(pmd_val(pmd) & _PAGE_VALID)) |
a7b9403f DM |
75 | return 0; |
76 | ||
77 | if (write && !pmd_write(pmd)) | |
89a77915 DM |
78 | return 0; |
79 | ||
80 | refs = 0; | |
81 | head = pmd_page(pmd); | |
82 | page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT); | |
89a77915 DM |
83 | do { |
84 | VM_BUG_ON(compound_head(page) != head); | |
85 | pages[*nr] = page; | |
86 | (*nr)++; | |
87 | page++; | |
88 | refs++; | |
89 | } while (addr += PAGE_SIZE, addr != end); | |
90 | ||
91 | if (!page_cache_add_speculative(head, refs)) { | |
92 | *nr -= refs; | |
93 | return 0; | |
94 | } | |
95 | ||
96 | if (unlikely(pmd_val(pmd) != pmd_val(*pmdp))) { | |
97 | *nr -= refs; | |
98 | while (refs--) | |
99 | put_page(head); | |
100 | return 0; | |
101 | } | |
102 | ||
89a77915 DM |
103 | return 1; |
104 | } | |
105 | ||
df077ac4 DM |
106 | static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end, |
107 | int write, struct page **pages, int *nr) | |
108 | { | |
109 | unsigned long next; | |
110 | pmd_t *pmdp; | |
111 | ||
112 | pmdp = pmd_offset(&pud, addr); | |
113 | do { | |
114 | pmd_t pmd = *pmdp; | |
115 | ||
116 | next = pmd_addr_end(addr, end); | |
99f1bc01 | 117 | if (pmd_none(pmd)) |
df077ac4 | 118 | return 0; |
89a77915 DM |
119 | if (unlikely(pmd_large(pmd))) { |
120 | if (!gup_huge_pmd(pmdp, pmd, addr, next, | |
121 | write, pages, nr)) | |
122 | return 0; | |
123 | } else if (!gup_pte_range(pmd, addr, next, write, | |
124 | pages, nr)) | |
df077ac4 DM |
125 | return 0; |
126 | } while (pmdp++, addr = next, addr != end); | |
127 | ||
128 | return 1; | |
129 | } | |
130 | ||
131 | static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end, | |
132 | int write, struct page **pages, int *nr) | |
133 | { | |
134 | unsigned long next; | |
135 | pud_t *pudp; | |
136 | ||
137 | pudp = pud_offset(&pgd, addr); | |
138 | do { | |
139 | pud_t pud = *pudp; | |
140 | ||
141 | next = pud_addr_end(addr, end); | |
142 | if (pud_none(pud)) | |
143 | return 0; | |
144 | if (!gup_pmd_range(pud, addr, next, write, pages, nr)) | |
145 | return 0; | |
146 | } while (pudp++, addr = next, addr != end); | |
147 | ||
148 | return 1; | |
149 | } | |
150 | ||
06090e8e DM |
151 | int __get_user_pages_fast(unsigned long start, int nr_pages, int write, |
152 | struct page **pages) | |
153 | { | |
154 | struct mm_struct *mm = current->mm; | |
155 | unsigned long addr, len, end; | |
156 | unsigned long next, flags; | |
157 | pgd_t *pgdp; | |
158 | int nr = 0; | |
159 | ||
160 | start &= PAGE_MASK; | |
161 | addr = start; | |
162 | len = (unsigned long) nr_pages << PAGE_SHIFT; | |
163 | end = start + len; | |
164 | ||
165 | local_irq_save(flags); | |
166 | pgdp = pgd_offset(mm, addr); | |
167 | do { | |
168 | pgd_t pgd = *pgdp; | |
169 | ||
170 | next = pgd_addr_end(addr, end); | |
171 | if (pgd_none(pgd)) | |
172 | break; | |
173 | if (!gup_pud_range(pgd, addr, next, write, pages, &nr)) | |
174 | break; | |
175 | } while (pgdp++, addr = next, addr != end); | |
176 | local_irq_restore(flags); | |
177 | ||
178 | return nr; | |
179 | } | |
180 | ||
df077ac4 DM |
181 | int get_user_pages_fast(unsigned long start, int nr_pages, int write, |
182 | struct page **pages) | |
183 | { | |
184 | struct mm_struct *mm = current->mm; | |
185 | unsigned long addr, len, end; | |
186 | unsigned long next; | |
187 | pgd_t *pgdp; | |
188 | int nr = 0; | |
189 | ||
190 | start &= PAGE_MASK; | |
191 | addr = start; | |
192 | len = (unsigned long) nr_pages << PAGE_SHIFT; | |
193 | end = start + len; | |
194 | ||
195 | /* | |
196 | * XXX: batch / limit 'nr', to avoid large irq off latency | |
197 | * needs some instrumenting to determine the common sizes used by | |
198 | * important workloads (eg. DB2), and whether limiting the batch size | |
199 | * will decrease performance. | |
200 | * | |
201 | * It seems like we're in the clear for the moment. Direct-IO is | |
202 | * the main guy that batches up lots of get_user_pages, and even | |
203 | * they are limited to 64-at-a-time which is not so many. | |
204 | */ | |
205 | /* | |
206 | * This doesn't prevent pagetable teardown, but does prevent | |
207 | * the pagetables from being freed on sparc. | |
208 | * | |
209 | * So long as we atomically load page table pointers versus teardown, | |
210 | * we can follow the address down to the the page and take a ref on it. | |
211 | */ | |
212 | local_irq_disable(); | |
213 | ||
214 | pgdp = pgd_offset(mm, addr); | |
215 | do { | |
216 | pgd_t pgd = *pgdp; | |
217 | ||
218 | next = pgd_addr_end(addr, end); | |
219 | if (pgd_none(pgd)) | |
220 | goto slow; | |
221 | if (!gup_pud_range(pgd, addr, next, write, pages, &nr)) | |
222 | goto slow; | |
223 | } while (pgdp++, addr = next, addr != end); | |
224 | ||
225 | local_irq_enable(); | |
226 | ||
227 | VM_BUG_ON(nr != (end - start) >> PAGE_SHIFT); | |
228 | return nr; | |
229 | ||
230 | { | |
231 | int ret; | |
232 | ||
233 | slow: | |
234 | local_irq_enable(); | |
235 | ||
236 | /* Try to get the remaining pages with get_user_pages */ | |
237 | start += nr << PAGE_SHIFT; | |
238 | pages += nr; | |
239 | ||
d4edcf0d | 240 | ret = get_user_pages_unlocked(start, |
a7b78075 | 241 | (end - start) >> PAGE_SHIFT, write, 0, pages); |
df077ac4 DM |
242 | |
243 | /* Have to be a bit careful with return values */ | |
244 | if (nr > 0) { | |
245 | if (ret < 0) | |
246 | ret = nr; | |
247 | else | |
248 | ret += nr; | |
249 | } | |
250 | ||
251 | return ret; | |
252 | } | |
253 | } |