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 | } | |
e0d85a36 AA |
59 | if (head != page) |
60 | get_huge_page_tail(page); | |
df077ac4 DM |
61 | |
62 | pages[*nr] = page; | |
63 | (*nr)++; | |
64 | } while (ptep++, addr += PAGE_SIZE, addr != end); | |
65 | ||
66 | return 1; | |
67 | } | |
68 | ||
89a77915 DM |
69 | static int gup_huge_pmd(pmd_t *pmdp, pmd_t pmd, unsigned long addr, |
70 | unsigned long end, int write, struct page **pages, | |
71 | int *nr) | |
72 | { | |
73 | struct page *head, *page, *tail; | |
89a77915 DM |
74 | int refs; |
75 | ||
04df419d | 76 | if (!(pmd_val(pmd) & _PAGE_VALID)) |
a7b9403f DM |
77 | return 0; |
78 | ||
79 | if (write && !pmd_write(pmd)) | |
89a77915 DM |
80 | return 0; |
81 | ||
82 | refs = 0; | |
83 | head = pmd_page(pmd); | |
84 | page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT); | |
85 | tail = page; | |
86 | do { | |
87 | VM_BUG_ON(compound_head(page) != head); | |
88 | pages[*nr] = page; | |
89 | (*nr)++; | |
90 | page++; | |
91 | refs++; | |
92 | } while (addr += PAGE_SIZE, addr != end); | |
93 | ||
94 | if (!page_cache_add_speculative(head, refs)) { | |
95 | *nr -= refs; | |
96 | return 0; | |
97 | } | |
98 | ||
99 | if (unlikely(pmd_val(pmd) != pmd_val(*pmdp))) { | |
100 | *nr -= refs; | |
101 | while (refs--) | |
102 | put_page(head); | |
103 | return 0; | |
104 | } | |
105 | ||
106 | /* Any tail page need their mapcount reference taken before we | |
107 | * return. | |
108 | */ | |
109 | while (refs--) { | |
110 | if (PageTail(tail)) | |
111 | get_huge_page_tail(tail); | |
112 | tail++; | |
113 | } | |
114 | ||
115 | return 1; | |
116 | } | |
117 | ||
df077ac4 DM |
118 | static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end, |
119 | int write, struct page **pages, int *nr) | |
120 | { | |
121 | unsigned long next; | |
122 | pmd_t *pmdp; | |
123 | ||
124 | pmdp = pmd_offset(&pud, addr); | |
125 | do { | |
126 | pmd_t pmd = *pmdp; | |
127 | ||
128 | next = pmd_addr_end(addr, end); | |
89a77915 | 129 | if (pmd_none(pmd) || pmd_trans_splitting(pmd)) |
df077ac4 | 130 | return 0; |
89a77915 DM |
131 | if (unlikely(pmd_large(pmd))) { |
132 | if (!gup_huge_pmd(pmdp, pmd, addr, next, | |
133 | write, pages, nr)) | |
134 | return 0; | |
135 | } else if (!gup_pte_range(pmd, addr, next, write, | |
136 | pages, nr)) | |
df077ac4 DM |
137 | return 0; |
138 | } while (pmdp++, addr = next, addr != end); | |
139 | ||
140 | return 1; | |
141 | } | |
142 | ||
143 | static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end, | |
144 | int write, struct page **pages, int *nr) | |
145 | { | |
146 | unsigned long next; | |
147 | pud_t *pudp; | |
148 | ||
149 | pudp = pud_offset(&pgd, addr); | |
150 | do { | |
151 | pud_t pud = *pudp; | |
152 | ||
153 | next = pud_addr_end(addr, end); | |
154 | if (pud_none(pud)) | |
155 | return 0; | |
156 | if (!gup_pmd_range(pud, addr, next, write, pages, nr)) | |
157 | return 0; | |
158 | } while (pudp++, addr = next, addr != end); | |
159 | ||
160 | return 1; | |
161 | } | |
162 | ||
163 | int get_user_pages_fast(unsigned long start, int nr_pages, int write, | |
164 | struct page **pages) | |
165 | { | |
166 | struct mm_struct *mm = current->mm; | |
167 | unsigned long addr, len, end; | |
168 | unsigned long next; | |
169 | pgd_t *pgdp; | |
170 | int nr = 0; | |
171 | ||
172 | start &= PAGE_MASK; | |
173 | addr = start; | |
174 | len = (unsigned long) nr_pages << PAGE_SHIFT; | |
175 | end = start + len; | |
176 | ||
177 | /* | |
178 | * XXX: batch / limit 'nr', to avoid large irq off latency | |
179 | * needs some instrumenting to determine the common sizes used by | |
180 | * important workloads (eg. DB2), and whether limiting the batch size | |
181 | * will decrease performance. | |
182 | * | |
183 | * It seems like we're in the clear for the moment. Direct-IO is | |
184 | * the main guy that batches up lots of get_user_pages, and even | |
185 | * they are limited to 64-at-a-time which is not so many. | |
186 | */ | |
187 | /* | |
188 | * This doesn't prevent pagetable teardown, but does prevent | |
189 | * the pagetables from being freed on sparc. | |
190 | * | |
191 | * So long as we atomically load page table pointers versus teardown, | |
192 | * we can follow the address down to the the page and take a ref on it. | |
193 | */ | |
194 | local_irq_disable(); | |
195 | ||
196 | pgdp = pgd_offset(mm, addr); | |
197 | do { | |
198 | pgd_t pgd = *pgdp; | |
199 | ||
200 | next = pgd_addr_end(addr, end); | |
201 | if (pgd_none(pgd)) | |
202 | goto slow; | |
203 | if (!gup_pud_range(pgd, addr, next, write, pages, &nr)) | |
204 | goto slow; | |
205 | } while (pgdp++, addr = next, addr != end); | |
206 | ||
207 | local_irq_enable(); | |
208 | ||
209 | VM_BUG_ON(nr != (end - start) >> PAGE_SHIFT); | |
210 | return nr; | |
211 | ||
212 | { | |
213 | int ret; | |
214 | ||
215 | slow: | |
216 | local_irq_enable(); | |
217 | ||
218 | /* Try to get the remaining pages with get_user_pages */ | |
219 | start += nr << PAGE_SHIFT; | |
220 | pages += nr; | |
221 | ||
222 | down_read(&mm->mmap_sem); | |
223 | ret = get_user_pages(current, mm, start, | |
224 | (end - start) >> PAGE_SHIFT, write, 0, pages, NULL); | |
225 | up_read(&mm->mmap_sem); | |
226 | ||
227 | /* Have to be a bit careful with return values */ | |
228 | if (nr > 0) { | |
229 | if (ret < 0) | |
230 | ret = nr; | |
231 | else | |
232 | ret += nr; | |
233 | } | |
234 | ||
235 | return ret; | |
236 | } | |
237 | } |