Commit | Line | Data |
---|---|---|
667b279b PM |
1 | /* |
2 | * Lockless get_user_pages_fast for SuperH | |
3 | * | |
4 | * Copyright (C) 2009 - 2010 Paul Mundt | |
5 | * | |
6 | * Cloned from the x86 and PowerPC versions, by: | |
7 | * | |
8 | * Copyright (C) 2008 Nick Piggin | |
9 | * Copyright (C) 2008 Novell Inc. | |
10 | */ | |
11 | #include <linux/sched.h> | |
12 | #include <linux/mm.h> | |
13 | #include <linux/vmstat.h> | |
14 | #include <linux/highmem.h> | |
15 | #include <asm/pgtable.h> | |
16 | ||
17 | static inline pte_t gup_get_pte(pte_t *ptep) | |
18 | { | |
19 | #ifndef CONFIG_X2TLB | |
20 | return ACCESS_ONCE(*ptep); | |
21 | #else | |
22 | /* | |
23 | * With get_user_pages_fast, we walk down the pagetables without | |
24 | * taking any locks. For this we would like to load the pointers | |
25 | * atomically, but that is not possible with 64-bit PTEs. What | |
26 | * we do have is the guarantee that a pte will only either go | |
27 | * from not present to present, or present to not present or both | |
28 | * -- it will not switch to a completely different present page | |
29 | * without a TLB flush in between; something that we are blocking | |
30 | * by holding interrupts off. | |
31 | * | |
32 | * Setting ptes from not present to present goes: | |
33 | * ptep->pte_high = h; | |
34 | * smp_wmb(); | |
35 | * ptep->pte_low = l; | |
36 | * | |
37 | * And present to not present goes: | |
38 | * ptep->pte_low = 0; | |
39 | * smp_wmb(); | |
40 | * ptep->pte_high = 0; | |
41 | * | |
42 | * We must ensure here that the load of pte_low sees l iff pte_high | |
43 | * sees h. We load pte_high *after* loading pte_low, which ensures we | |
44 | * don't see an older value of pte_high. *Then* we recheck pte_low, | |
45 | * which ensures that we haven't picked up a changed pte high. We might | |
46 | * have got rubbish values from pte_low and pte_high, but we are | |
47 | * guaranteed that pte_low will not have the present bit set *unless* | |
48 | * it is 'l'. And get_user_pages_fast only operates on present ptes, so | |
49 | * we're safe. | |
50 | * | |
51 | * gup_get_pte should not be used or copied outside gup.c without being | |
52 | * very careful -- it does not atomically load the pte or anything that | |
53 | * is likely to be useful for you. | |
54 | */ | |
55 | pte_t pte; | |
56 | ||
57 | retry: | |
58 | pte.pte_low = ptep->pte_low; | |
59 | smp_rmb(); | |
60 | pte.pte_high = ptep->pte_high; | |
61 | smp_rmb(); | |
62 | if (unlikely(pte.pte_low != ptep->pte_low)) | |
63 | goto retry; | |
64 | ||
65 | return pte; | |
66 | #endif | |
67 | } | |
68 | ||
69 | /* | |
70 | * The performance critical leaf functions are made noinline otherwise gcc | |
71 | * inlines everything into a single function which results in too much | |
72 | * register pressure. | |
73 | */ | |
74 | static noinline int gup_pte_range(pmd_t pmd, unsigned long addr, | |
75 | unsigned long end, int write, struct page **pages, int *nr) | |
76 | { | |
77 | u64 mask, result; | |
78 | pte_t *ptep; | |
79 | ||
80 | #ifdef CONFIG_X2TLB | |
81 | result = _PAGE_PRESENT | _PAGE_EXT(_PAGE_EXT_KERN_READ | _PAGE_EXT_USER_READ); | |
82 | if (write) | |
83 | result |= _PAGE_EXT(_PAGE_EXT_KERN_WRITE | _PAGE_EXT_USER_WRITE); | |
84 | #elif defined(CONFIG_SUPERH64) | |
85 | result = _PAGE_PRESENT | _PAGE_USER | _PAGE_READ; | |
86 | if (write) | |
87 | result |= _PAGE_WRITE; | |
88 | #else | |
89 | result = _PAGE_PRESENT | _PAGE_USER; | |
90 | if (write) | |
91 | result |= _PAGE_RW; | |
92 | #endif | |
93 | ||
94 | mask = result | _PAGE_SPECIAL; | |
95 | ||
96 | ptep = pte_offset_map(&pmd, addr); | |
97 | do { | |
98 | pte_t pte = gup_get_pte(ptep); | |
99 | struct page *page; | |
100 | ||
101 | if ((pte_val(pte) & mask) != result) { | |
102 | pte_unmap(ptep); | |
103 | return 0; | |
104 | } | |
105 | VM_BUG_ON(!pfn_valid(pte_pfn(pte))); | |
106 | page = pte_page(pte); | |
107 | get_page(page); | |
108 | pages[*nr] = page; | |
109 | (*nr)++; | |
110 | ||
111 | } while (ptep++, addr += PAGE_SIZE, addr != end); | |
112 | pte_unmap(ptep - 1); | |
113 | ||
114 | return 1; | |
115 | } | |
116 | ||
117 | static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end, | |
118 | int write, struct page **pages, int *nr) | |
119 | { | |
120 | unsigned long next; | |
121 | pmd_t *pmdp; | |
122 | ||
123 | pmdp = pmd_offset(&pud, addr); | |
124 | do { | |
125 | pmd_t pmd = *pmdp; | |
126 | ||
127 | next = pmd_addr_end(addr, end); | |
128 | if (pmd_none(pmd)) | |
129 | return 0; | |
130 | if (!gup_pte_range(pmd, addr, next, write, pages, nr)) | |
131 | return 0; | |
132 | } while (pmdp++, addr = next, addr != end); | |
133 | ||
134 | return 1; | |
135 | } | |
136 | ||
137 | static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end, | |
138 | int write, struct page **pages, int *nr) | |
139 | { | |
140 | unsigned long next; | |
141 | pud_t *pudp; | |
142 | ||
143 | pudp = pud_offset(&pgd, addr); | |
144 | do { | |
145 | pud_t pud = *pudp; | |
146 | ||
147 | next = pud_addr_end(addr, end); | |
148 | if (pud_none(pud)) | |
149 | return 0; | |
150 | if (!gup_pmd_range(pud, addr, next, write, pages, nr)) | |
151 | return 0; | |
152 | } while (pudp++, addr = next, addr != end); | |
153 | ||
154 | return 1; | |
155 | } | |
156 | ||
157 | /* | |
158 | * Like get_user_pages_fast() except its IRQ-safe in that it won't fall | |
159 | * back to the regular GUP. | |
160 | */ | |
161 | int __get_user_pages_fast(unsigned long start, int nr_pages, int write, | |
162 | struct page **pages) | |
163 | { | |
164 | struct mm_struct *mm = current->mm; | |
165 | unsigned long addr, len, end; | |
166 | unsigned long next; | |
167 | unsigned long flags; | |
168 | pgd_t *pgdp; | |
169 | int nr = 0; | |
170 | ||
171 | start &= PAGE_MASK; | |
172 | addr = start; | |
173 | len = (unsigned long) nr_pages << PAGE_SHIFT; | |
174 | end = start + len; | |
175 | if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ, | |
176 | (void __user *)start, len))) | |
177 | return 0; | |
178 | ||
179 | /* | |
180 | * This doesn't prevent pagetable teardown, but does prevent | |
181 | * the pagetables and pages from being freed. | |
182 | */ | |
183 | local_irq_save(flags); | |
184 | pgdp = pgd_offset(mm, addr); | |
185 | do { | |
186 | pgd_t pgd = *pgdp; | |
187 | ||
188 | next = pgd_addr_end(addr, end); | |
189 | if (pgd_none(pgd)) | |
190 | break; | |
191 | if (!gup_pud_range(pgd, addr, next, write, pages, &nr)) | |
192 | break; | |
193 | } while (pgdp++, addr = next, addr != end); | |
194 | local_irq_restore(flags); | |
195 | ||
196 | return nr; | |
197 | } | |
198 | ||
199 | /** | |
200 | * get_user_pages_fast() - pin user pages in memory | |
201 | * @start: starting user address | |
202 | * @nr_pages: number of pages from start to pin | |
203 | * @write: whether pages will be written to | |
204 | * @pages: array that receives pointers to the pages pinned. | |
205 | * Should be at least nr_pages long. | |
206 | * | |
207 | * Attempt to pin user pages in memory without taking mm->mmap_sem. | |
208 | * If not successful, it will fall back to taking the lock and | |
209 | * calling get_user_pages(). | |
210 | * | |
211 | * Returns number of pages pinned. This may be fewer than the number | |
212 | * requested. If nr_pages is 0 or negative, returns 0. If no pages | |
213 | * were pinned, returns -errno. | |
214 | */ | |
215 | int get_user_pages_fast(unsigned long start, int nr_pages, int write, | |
216 | struct page **pages) | |
217 | { | |
218 | struct mm_struct *mm = current->mm; | |
219 | unsigned long addr, len, end; | |
220 | unsigned long next; | |
221 | pgd_t *pgdp; | |
222 | int nr = 0; | |
223 | ||
224 | start &= PAGE_MASK; | |
225 | addr = start; | |
226 | len = (unsigned long) nr_pages << PAGE_SHIFT; | |
227 | ||
228 | end = start + len; | |
229 | if (end < start) | |
230 | goto slow_irqon; | |
231 | ||
232 | local_irq_disable(); | |
233 | pgdp = pgd_offset(mm, addr); | |
234 | do { | |
235 | pgd_t pgd = *pgdp; | |
236 | ||
237 | next = pgd_addr_end(addr, end); | |
238 | if (pgd_none(pgd)) | |
239 | goto slow; | |
240 | if (!gup_pud_range(pgd, addr, next, write, pages, &nr)) | |
241 | goto slow; | |
242 | } while (pgdp++, addr = next, addr != end); | |
243 | local_irq_enable(); | |
244 | ||
245 | VM_BUG_ON(nr != (end - start) >> PAGE_SHIFT); | |
246 | return nr; | |
247 | ||
248 | { | |
249 | int ret; | |
250 | ||
251 | slow: | |
252 | local_irq_enable(); | |
253 | slow_irqon: | |
254 | /* Try to get the remaining pages with get_user_pages */ | |
255 | start += nr << PAGE_SHIFT; | |
256 | pages += nr; | |
257 | ||
258 | down_read(&mm->mmap_sem); | |
259 | ret = get_user_pages(current, mm, start, | |
260 | (end - start) >> PAGE_SHIFT, write, 0, pages, NULL); | |
261 | up_read(&mm->mmap_sem); | |
262 | ||
263 | /* Have to be a bit careful with return values */ | |
264 | if (nr > 0) { | |
265 | if (ret < 0) | |
266 | ret = nr; | |
267 | else | |
268 | ret += nr; | |
269 | } | |
270 | ||
271 | return ret; | |
272 | } | |
273 | } |