]>
Commit | Line | Data |
---|---|---|
b1c10bea HD |
1 | /* |
2 | * Lockless get_user_pages_fast for MIPS | |
3 | * | |
4 | * Copyright (C) 2008 Nick Piggin | |
5 | * Copyright (C) 2008 Novell Inc. | |
6 | * Copyright (C) 2011 Ralf Baechle | |
7 | */ | |
8 | #include <linux/sched.h> | |
9 | #include <linux/mm.h> | |
10 | #include <linux/vmstat.h> | |
11 | #include <linux/highmem.h> | |
12 | #include <linux/swap.h> | |
13 | #include <linux/hugetlb.h> | |
14 | ||
15 | #include <asm/pgtable.h> | |
16 | ||
17 | static inline pte_t gup_get_pte(pte_t *ptep) | |
18 | { | |
19 | #if defined(CONFIG_64BIT_PHYS_ADDR) && defined(CONFIG_CPU_MIPS32) | |
20 | pte_t pte; | |
21 | ||
22 | retry: | |
23 | pte.pte_low = ptep->pte_low; | |
24 | smp_rmb(); | |
25 | pte.pte_high = ptep->pte_high; | |
26 | smp_rmb(); | |
27 | if (unlikely(pte.pte_low != ptep->pte_low)) | |
28 | goto retry; | |
29 | ||
30 | return pte; | |
31 | #else | |
32 | return ACCESS_ONCE(*ptep); | |
33 | #endif | |
34 | } | |
35 | ||
36 | static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end, | |
37 | int write, struct page **pages, int *nr) | |
38 | { | |
39 | pte_t *ptep = pte_offset_map(&pmd, addr); | |
40 | do { | |
41 | pte_t pte = gup_get_pte(ptep); | |
42 | struct page *page; | |
43 | ||
44 | if (!pte_present(pte) || | |
45 | pte_special(pte) || (write && !pte_write(pte))) { | |
46 | pte_unmap(ptep); | |
47 | return 0; | |
48 | } | |
49 | VM_BUG_ON(!pfn_valid(pte_pfn(pte))); | |
50 | page = pte_page(pte); | |
51 | get_page(page); | |
52 | SetPageReferenced(page); | |
53 | pages[*nr] = page; | |
54 | (*nr)++; | |
55 | ||
56 | } while (ptep++, addr += PAGE_SIZE, addr != end); | |
57 | ||
58 | pte_unmap(ptep - 1); | |
59 | return 1; | |
60 | } | |
61 | ||
62 | static inline void get_head_page_multiple(struct page *page, int nr) | |
63 | { | |
64 | VM_BUG_ON(page != compound_head(page)); | |
65 | VM_BUG_ON(page_count(page) == 0); | |
66 | atomic_add(nr, &page->_count); | |
67 | SetPageReferenced(page); | |
68 | } | |
69 | ||
70 | static int gup_huge_pmd(pmd_t pmd, unsigned long addr, unsigned long end, | |
71 | int write, struct page **pages, int *nr) | |
72 | { | |
73 | pte_t pte = *(pte_t *)&pmd; | |
74 | struct page *head, *page; | |
75 | int refs; | |
76 | ||
77 | if (write && !pte_write(pte)) | |
78 | return 0; | |
79 | /* hugepages are never "special" */ | |
80 | VM_BUG_ON(pte_special(pte)); | |
81 | VM_BUG_ON(!pfn_valid(pte_pfn(pte))); | |
82 | ||
83 | refs = 0; | |
84 | head = pte_page(pte); | |
85 | page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT); | |
86 | do { | |
87 | VM_BUG_ON(compound_head(page) != head); | |
88 | pages[*nr] = page; | |
89 | if (PageTail(page)) | |
90 | get_huge_page_tail(page); | |
91 | (*nr)++; | |
92 | page++; | |
93 | refs++; | |
94 | } while (addr += PAGE_SIZE, addr != end); | |
95 | ||
96 | get_head_page_multiple(head, refs); | |
97 | return 1; | |
98 | } | |
99 | ||
100 | static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end, | |
101 | int write, struct page **pages, int *nr) | |
102 | { | |
103 | unsigned long next; | |
104 | pmd_t *pmdp; | |
105 | ||
106 | pmdp = pmd_offset(&pud, addr); | |
107 | do { | |
108 | pmd_t pmd = *pmdp; | |
109 | ||
110 | next = pmd_addr_end(addr, end); | |
111 | /* | |
112 | * The pmd_trans_splitting() check below explains why | |
113 | * pmdp_splitting_flush has to flush the tlb, to stop | |
114 | * this gup-fast code from running while we set the | |
115 | * splitting bit in the pmd. Returning zero will take | |
116 | * the slow path that will call wait_split_huge_page() | |
117 | * if the pmd is still in splitting state. gup-fast | |
118 | * can't because it has irq disabled and | |
119 | * wait_split_huge_page() would never return as the | |
120 | * tlb flush IPI wouldn't run. | |
121 | */ | |
122 | if (pmd_none(pmd) || pmd_trans_splitting(pmd)) | |
123 | return 0; | |
124 | if (unlikely(pmd_huge(pmd))) { | |
125 | if (!gup_huge_pmd(pmd, addr, next, write, pages,nr)) | |
126 | return 0; | |
127 | } else { | |
128 | if (!gup_pte_range(pmd, addr, next, write, pages,nr)) | |
129 | return 0; | |
130 | } | |
131 | } while (pmdp++, addr = next, addr != end); | |
132 | ||
133 | return 1; | |
134 | } | |
135 | ||
136 | static int gup_huge_pud(pud_t pud, unsigned long addr, unsigned long end, | |
137 | int write, struct page **pages, int *nr) | |
138 | { | |
139 | pte_t pte = *(pte_t *)&pud; | |
140 | struct page *head, *page; | |
141 | int refs; | |
142 | ||
143 | if (write && !pte_write(pte)) | |
144 | return 0; | |
145 | /* hugepages are never "special" */ | |
146 | VM_BUG_ON(pte_special(pte)); | |
147 | VM_BUG_ON(!pfn_valid(pte_pfn(pte))); | |
148 | ||
149 | refs = 0; | |
150 | head = pte_page(pte); | |
151 | page = head + ((addr & ~PUD_MASK) >> PAGE_SHIFT); | |
152 | do { | |
153 | VM_BUG_ON(compound_head(page) != head); | |
154 | pages[*nr] = page; | |
155 | (*nr)++; | |
156 | page++; | |
157 | refs++; | |
158 | } while (addr += PAGE_SIZE, addr != end); | |
159 | ||
160 | get_head_page_multiple(head, refs); | |
161 | return 1; | |
162 | } | |
163 | ||
164 | static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end, | |
165 | int write, struct page **pages, int *nr) | |
166 | { | |
167 | unsigned long next; | |
168 | pud_t *pudp; | |
169 | ||
170 | pudp = pud_offset(&pgd, addr); | |
171 | do { | |
172 | pud_t pud = *pudp; | |
173 | ||
174 | next = pud_addr_end(addr, end); | |
175 | if (pud_none(pud)) | |
176 | return 0; | |
177 | if (unlikely(pud_huge(pud))) { | |
178 | if (!gup_huge_pud(pud, addr, next, write, pages,nr)) | |
179 | return 0; | |
180 | } else { | |
181 | if (!gup_pmd_range(pud, addr, next, write, pages,nr)) | |
182 | return 0; | |
183 | } | |
184 | } while (pudp++, addr = next, addr != end); | |
185 | ||
186 | return 1; | |
187 | } | |
188 | ||
189 | /* | |
190 | * Like get_user_pages_fast() except its IRQ-safe in that it won't fall | |
191 | * back to the regular GUP. | |
192 | */ | |
193 | int __get_user_pages_fast(unsigned long start, int nr_pages, int write, | |
194 | struct page **pages) | |
195 | { | |
196 | struct mm_struct *mm = current->mm; | |
197 | unsigned long addr, len, end; | |
198 | unsigned long next; | |
199 | unsigned long flags; | |
200 | pgd_t *pgdp; | |
201 | int nr = 0; | |
202 | ||
203 | start &= PAGE_MASK; | |
204 | addr = start; | |
205 | len = (unsigned long) nr_pages << PAGE_SHIFT; | |
206 | end = start + len; | |
207 | if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ, | |
208 | (void __user *)start, len))) | |
209 | return 0; | |
210 | ||
211 | /* | |
212 | * XXX: batch / limit 'nr', to avoid large irq off latency | |
213 | * needs some instrumenting to determine the common sizes used by | |
214 | * important workloads (eg. DB2), and whether limiting the batch | |
215 | * size will decrease performance. | |
216 | * | |
217 | * It seems like we're in the clear for the moment. Direct-IO is | |
218 | * the main guy that batches up lots of get_user_pages, and even | |
219 | * they are limited to 64-at-a-time which is not so many. | |
220 | */ | |
221 | /* | |
222 | * This doesn't prevent pagetable teardown, but does prevent | |
223 | * the pagetables and pages from being freed. | |
224 | * | |
225 | * So long as we atomically load page table pointers versus teardown, | |
226 | * we can follow the address down to the page and take a ref on it. | |
227 | */ | |
228 | local_irq_save(flags); | |
229 | pgdp = pgd_offset(mm, addr); | |
230 | do { | |
231 | pgd_t pgd = *pgdp; | |
232 | ||
233 | next = pgd_addr_end(addr, end); | |
234 | if (pgd_none(pgd)) | |
235 | break; | |
236 | if (!gup_pud_range(pgd, addr, next, write, pages, &nr)) | |
237 | break; | |
238 | } while (pgdp++, addr = next, addr != end); | |
239 | local_irq_restore(flags); | |
240 | ||
241 | return nr; | |
242 | } | |
243 | ||
244 | /** | |
245 | * get_user_pages_fast() - pin user pages in memory | |
246 | * @start: starting user address | |
247 | * @nr_pages: number of pages from start to pin | |
248 | * @write: whether pages will be written to | |
249 | * @pages: array that receives pointers to the pages pinned. | |
250 | * Should be at least nr_pages long. | |
251 | * | |
252 | * Attempt to pin user pages in memory without taking mm->mmap_sem. | |
253 | * If not successful, it will fall back to taking the lock and | |
254 | * calling get_user_pages(). | |
255 | * | |
256 | * Returns number of pages pinned. This may be fewer than the number | |
257 | * requested. If nr_pages is 0 or negative, returns 0. If no pages | |
258 | * were pinned, returns -errno. | |
259 | */ | |
260 | int get_user_pages_fast(unsigned long start, int nr_pages, int write, | |
261 | struct page **pages) | |
262 | { | |
263 | struct mm_struct *mm = current->mm; | |
264 | unsigned long addr, len, end; | |
265 | unsigned long next; | |
266 | pgd_t *pgdp; | |
267 | int ret, nr = 0; | |
268 | ||
269 | start &= PAGE_MASK; | |
270 | addr = start; | |
271 | len = (unsigned long) nr_pages << PAGE_SHIFT; | |
272 | ||
273 | end = start + len; | |
274 | if (end < start) | |
275 | goto slow_irqon; | |
276 | ||
277 | /* XXX: batch / limit 'nr' */ | |
278 | local_irq_disable(); | |
279 | pgdp = pgd_offset(mm, addr); | |
280 | do { | |
281 | pgd_t pgd = *pgdp; | |
282 | ||
283 | next = pgd_addr_end(addr, end); | |
284 | if (pgd_none(pgd)) | |
285 | goto slow; | |
286 | if (!gup_pud_range(pgd, addr, next, write, pages, &nr)) | |
287 | goto slow; | |
288 | } while (pgdp++, addr = next, addr != end); | |
289 | local_irq_enable(); | |
290 | ||
291 | VM_BUG_ON(nr != (end - start) >> PAGE_SHIFT); | |
292 | return nr; | |
293 | slow: | |
294 | local_irq_enable(); | |
295 | ||
296 | slow_irqon: | |
297 | /* Try to get the remaining pages with get_user_pages */ | |
298 | start += nr << PAGE_SHIFT; | |
299 | pages += nr; | |
300 | ||
301 | down_read(&mm->mmap_sem); | |
302 | ret = get_user_pages(current, mm, start, | |
303 | (end - start) >> PAGE_SHIFT, | |
304 | write, 0, pages, NULL); | |
305 | up_read(&mm->mmap_sem); | |
306 | ||
307 | /* Have to be a bit careful with return values */ | |
308 | if (nr > 0) { | |
309 | if (ret < 0) | |
310 | ret = nr; | |
311 | else | |
312 | ret += nr; | |
313 | } | |
314 | return ret; | |
315 | } |