Commit | Line | Data |
---|---|---|
80217147 MS |
1 | /* |
2 | * Lockless get_user_pages_fast for s390 | |
3 | * | |
4 | * Copyright IBM Corp. 2010 | |
5 | * Author(s): Martin Schwidefsky <schwidefsky@de.ibm.com> | |
6 | */ | |
7 | #include <linux/sched.h> | |
8 | #include <linux/mm.h> | |
9 | #include <linux/hugetlb.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 inline int gup_pte_range(pmd_t *pmdp, pmd_t pmd, unsigned long addr, | |
21 | unsigned long end, int write, struct page **pages, int *nr) | |
22 | { | |
25591b07 | 23 | unsigned long mask; |
80217147 MS |
24 | pte_t *ptep, pte; |
25 | struct page *page; | |
26 | ||
25591b07 | 27 | mask = (write ? _PAGE_RO : 0) | _PAGE_INVALID | _PAGE_SPECIAL; |
80217147 MS |
28 | |
29 | ptep = ((pte_t *) pmd_deref(pmd)) + pte_index(addr); | |
30 | do { | |
31 | pte = *ptep; | |
32 | barrier(); | |
25591b07 | 33 | if ((pte_val(pte) & mask) != 0) |
80217147 MS |
34 | return 0; |
35 | VM_BUG_ON(!pfn_valid(pte_pfn(pte))); | |
36 | page = pte_page(pte); | |
37 | if (!page_cache_get_speculative(page)) | |
38 | return 0; | |
39 | if (unlikely(pte_val(pte) != pte_val(*ptep))) { | |
40 | put_page(page); | |
41 | return 0; | |
42 | } | |
43 | pages[*nr] = page; | |
44 | (*nr)++; | |
45 | ||
46 | } while (ptep++, addr += PAGE_SIZE, addr != end); | |
47 | ||
48 | return 1; | |
49 | } | |
50 | ||
51 | static inline int gup_huge_pmd(pmd_t *pmdp, pmd_t pmd, unsigned long addr, | |
52 | unsigned long end, int write, struct page **pages, int *nr) | |
53 | { | |
54 | unsigned long mask, result; | |
220a2eb2 | 55 | struct page *head, *page, *tail; |
80217147 MS |
56 | int refs; |
57 | ||
58 | result = write ? 0 : _SEGMENT_ENTRY_RO; | |
59 | mask = result | _SEGMENT_ENTRY_INV; | |
60 | if ((pmd_val(pmd) & mask) != result) | |
61 | return 0; | |
62 | VM_BUG_ON(!pfn_valid(pmd_val(pmd) >> PAGE_SHIFT)); | |
63 | ||
64 | refs = 0; | |
65 | head = pmd_page(pmd); | |
66 | page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT); | |
220a2eb2 | 67 | tail = page; |
80217147 MS |
68 | do { |
69 | VM_BUG_ON(compound_head(page) != head); | |
70 | pages[*nr] = page; | |
71 | (*nr)++; | |
72 | page++; | |
73 | refs++; | |
74 | } while (addr += PAGE_SIZE, addr != end); | |
75 | ||
76 | if (!page_cache_add_speculative(head, refs)) { | |
77 | *nr -= refs; | |
78 | return 0; | |
79 | } | |
80 | ||
81 | if (unlikely(pmd_val(pmd) != pmd_val(*pmdp))) { | |
82 | *nr -= refs; | |
83 | while (refs--) | |
84 | put_page(head); | |
0693bc9c AA |
85 | return 0; |
86 | } | |
87 | ||
88 | /* | |
89 | * Any tail page need their mapcount reference taken before we | |
90 | * return. | |
91 | */ | |
92 | while (refs--) { | |
93 | if (PageTail(tail)) | |
94 | get_huge_page_tail(tail); | |
95 | tail++; | |
80217147 MS |
96 | } |
97 | ||
98 | return 1; | |
99 | } | |
100 | ||
101 | ||
102 | static inline int gup_pmd_range(pud_t *pudp, pud_t pud, unsigned long addr, | |
103 | unsigned long end, int write, struct page **pages, int *nr) | |
104 | { | |
105 | unsigned long next; | |
106 | pmd_t *pmdp, pmd; | |
107 | ||
108 | pmdp = (pmd_t *) pudp; | |
109 | #ifdef CONFIG_64BIT | |
110 | if ((pud_val(pud) & _REGION_ENTRY_TYPE_MASK) == _REGION_ENTRY_TYPE_R3) | |
111 | pmdp = (pmd_t *) pud_deref(pud); | |
112 | pmdp += pmd_index(addr); | |
113 | #endif | |
114 | do { | |
115 | pmd = *pmdp; | |
116 | barrier(); | |
117 | next = pmd_addr_end(addr, end); | |
75077afb GS |
118 | /* |
119 | * The pmd_trans_splitting() check below explains why | |
120 | * pmdp_splitting_flush() has to serialize with | |
121 | * smp_call_function() against our disabled IRQs, to stop | |
122 | * this gup-fast code from running while we set the | |
123 | * splitting bit in the pmd. Returning zero will take | |
124 | * the slow path that will call wait_split_huge_page() | |
125 | * if the pmd is still in splitting state. | |
126 | */ | |
127 | if (pmd_none(pmd) || pmd_trans_splitting(pmd)) | |
80217147 | 128 | return 0; |
156152f8 | 129 | if (unlikely(pmd_large(pmd))) { |
80217147 MS |
130 | if (!gup_huge_pmd(pmdp, pmd, addr, next, |
131 | write, pages, nr)) | |
132 | return 0; | |
133 | } else if (!gup_pte_range(pmdp, pmd, addr, next, | |
134 | write, pages, nr)) | |
135 | return 0; | |
136 | } while (pmdp++, addr = next, addr != end); | |
137 | ||
138 | return 1; | |
139 | } | |
140 | ||
141 | static inline int gup_pud_range(pgd_t *pgdp, pgd_t pgd, unsigned long addr, | |
142 | unsigned long end, int write, struct page **pages, int *nr) | |
143 | { | |
144 | unsigned long next; | |
145 | pud_t *pudp, pud; | |
146 | ||
147 | pudp = (pud_t *) pgdp; | |
148 | #ifdef CONFIG_64BIT | |
149 | if ((pgd_val(pgd) & _REGION_ENTRY_TYPE_MASK) == _REGION_ENTRY_TYPE_R2) | |
150 | pudp = (pud_t *) pgd_deref(pgd); | |
151 | pudp += pud_index(addr); | |
152 | #endif | |
153 | do { | |
154 | pud = *pudp; | |
155 | barrier(); | |
156 | next = pud_addr_end(addr, end); | |
157 | if (pud_none(pud)) | |
158 | return 0; | |
159 | if (!gup_pmd_range(pudp, pud, addr, next, write, pages, nr)) | |
160 | return 0; | |
161 | } while (pudp++, addr = next, addr != end); | |
162 | ||
163 | return 1; | |
164 | } | |
165 | ||
34cda992 GS |
166 | /* |
167 | * Like get_user_pages_fast() except its IRQ-safe in that it won't fall | |
168 | * back to the regular GUP. | |
169 | */ | |
170 | int __get_user_pages_fast(unsigned long start, int nr_pages, int write, | |
171 | struct page **pages) | |
172 | { | |
173 | struct mm_struct *mm = current->mm; | |
174 | unsigned long addr, len, end; | |
175 | unsigned long next, flags; | |
176 | pgd_t *pgdp, pgd; | |
177 | int nr = 0; | |
178 | ||
179 | start &= PAGE_MASK; | |
180 | addr = start; | |
181 | len = (unsigned long) nr_pages << PAGE_SHIFT; | |
182 | end = start + len; | |
516bad44 | 183 | if ((end < start) || (end > TASK_SIZE)) |
34cda992 GS |
184 | return 0; |
185 | ||
186 | local_irq_save(flags); | |
187 | pgdp = pgd_offset(mm, addr); | |
188 | do { | |
189 | pgd = *pgdp; | |
190 | barrier(); | |
191 | next = pgd_addr_end(addr, end); | |
192 | if (pgd_none(pgd)) | |
193 | break; | |
194 | if (!gup_pud_range(pgdp, pgd, addr, next, write, pages, &nr)) | |
195 | break; | |
196 | } while (pgdp++, addr = next, addr != end); | |
197 | local_irq_restore(flags); | |
198 | ||
199 | return nr; | |
200 | } | |
201 | ||
80217147 MS |
202 | /** |
203 | * get_user_pages_fast() - pin user pages in memory | |
204 | * @start: starting user address | |
205 | * @nr_pages: number of pages from start to pin | |
206 | * @write: whether pages will be written to | |
207 | * @pages: array that receives pointers to the pages pinned. | |
208 | * Should be at least nr_pages long. | |
209 | * | |
210 | * Attempt to pin user pages in memory without taking mm->mmap_sem. | |
211 | * If not successful, it will fall back to taking the lock and | |
212 | * calling get_user_pages(). | |
213 | * | |
214 | * Returns number of pages pinned. This may be fewer than the number | |
215 | * requested. If nr_pages is 0 or negative, returns 0. If no pages | |
216 | * were pinned, returns -errno. | |
217 | */ | |
218 | int get_user_pages_fast(unsigned long start, int nr_pages, int write, | |
219 | struct page **pages) | |
220 | { | |
221 | struct mm_struct *mm = current->mm; | |
222 | unsigned long addr, len, end; | |
223 | unsigned long next; | |
224 | pgd_t *pgdp, pgd; | |
225 | int nr = 0; | |
226 | ||
227 | start &= PAGE_MASK; | |
228 | addr = start; | |
229 | len = (unsigned long) nr_pages << PAGE_SHIFT; | |
230 | end = start + len; | |
d55c4c61 | 231 | if ((end < start) || (end > TASK_SIZE)) |
80217147 MS |
232 | goto slow_irqon; |
233 | ||
234 | /* | |
235 | * local_irq_disable() doesn't prevent pagetable teardown, but does | |
236 | * prevent the pagetables from being freed on s390. | |
237 | * | |
238 | * So long as we atomically load page table pointers versus teardown, | |
239 | * we can follow the address down to the the page and take a ref on it. | |
240 | */ | |
241 | local_irq_disable(); | |
242 | pgdp = pgd_offset(mm, addr); | |
243 | do { | |
244 | pgd = *pgdp; | |
245 | barrier(); | |
246 | next = pgd_addr_end(addr, end); | |
247 | if (pgd_none(pgd)) | |
248 | goto slow; | |
249 | if (!gup_pud_range(pgdp, pgd, addr, next, write, pages, &nr)) | |
250 | goto slow; | |
251 | } while (pgdp++, addr = next, addr != end); | |
252 | local_irq_enable(); | |
253 | ||
254 | VM_BUG_ON(nr != (end - start) >> PAGE_SHIFT); | |
255 | return nr; | |
256 | ||
257 | { | |
258 | int ret; | |
259 | slow: | |
260 | local_irq_enable(); | |
261 | slow_irqon: | |
262 | /* Try to get the remaining pages with get_user_pages */ | |
263 | start += nr << PAGE_SHIFT; | |
264 | pages += nr; | |
265 | ||
266 | down_read(&mm->mmap_sem); | |
267 | ret = get_user_pages(current, mm, start, | |
268 | (end - start) >> PAGE_SHIFT, write, 0, pages, NULL); | |
269 | up_read(&mm->mmap_sem); | |
270 | ||
271 | /* Have to be a bit careful with return values */ | |
272 | if (nr > 0) { | |
273 | if (ret < 0) | |
274 | ret = nr; | |
275 | else | |
276 | ret += nr; | |
277 | } | |
278 | ||
279 | return ret; | |
280 | } | |
281 | } |