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 | { | |
23 | unsigned long mask, result; | |
24 | pte_t *ptep, pte; | |
25 | struct page *page; | |
26 | ||
27 | result = write ? 0 : _PAGE_RO; | |
28 | mask = result | _PAGE_INVALID | _PAGE_SPECIAL; | |
29 | ||
30 | ptep = ((pte_t *) pmd_deref(pmd)) + pte_index(addr); | |
31 | do { | |
32 | pte = *ptep; | |
33 | barrier(); | |
34 | if ((pte_val(pte) & mask) != result) | |
35 | return 0; | |
36 | VM_BUG_ON(!pfn_valid(pte_pfn(pte))); | |
37 | page = pte_page(pte); | |
38 | if (!page_cache_get_speculative(page)) | |
39 | return 0; | |
40 | if (unlikely(pte_val(pte) != pte_val(*ptep))) { | |
41 | put_page(page); | |
42 | return 0; | |
43 | } | |
44 | pages[*nr] = page; | |
45 | (*nr)++; | |
46 | ||
47 | } while (ptep++, addr += PAGE_SIZE, addr != end); | |
48 | ||
49 | return 1; | |
50 | } | |
51 | ||
52 | static inline int gup_huge_pmd(pmd_t *pmdp, pmd_t pmd, unsigned long addr, | |
53 | unsigned long end, int write, struct page **pages, int *nr) | |
54 | { | |
55 | unsigned long mask, result; | |
56 | struct page *head, *page; | |
57 | int refs; | |
58 | ||
59 | result = write ? 0 : _SEGMENT_ENTRY_RO; | |
60 | mask = result | _SEGMENT_ENTRY_INV; | |
61 | if ((pmd_val(pmd) & mask) != result) | |
62 | return 0; | |
63 | VM_BUG_ON(!pfn_valid(pmd_val(pmd) >> PAGE_SHIFT)); | |
64 | ||
65 | refs = 0; | |
66 | head = pmd_page(pmd); | |
67 | page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT); | |
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); | |
85 | } | |
86 | ||
87 | return 1; | |
88 | } | |
89 | ||
90 | ||
91 | static inline int gup_pmd_range(pud_t *pudp, pud_t pud, unsigned long addr, | |
92 | unsigned long end, int write, struct page **pages, int *nr) | |
93 | { | |
94 | unsigned long next; | |
95 | pmd_t *pmdp, pmd; | |
96 | ||
97 | pmdp = (pmd_t *) pudp; | |
98 | #ifdef CONFIG_64BIT | |
99 | if ((pud_val(pud) & _REGION_ENTRY_TYPE_MASK) == _REGION_ENTRY_TYPE_R3) | |
100 | pmdp = (pmd_t *) pud_deref(pud); | |
101 | pmdp += pmd_index(addr); | |
102 | #endif | |
103 | do { | |
104 | pmd = *pmdp; | |
105 | barrier(); | |
106 | next = pmd_addr_end(addr, end); | |
107 | if (pmd_none(pmd)) | |
108 | return 0; | |
109 | if (unlikely(pmd_huge(pmd))) { | |
110 | if (!gup_huge_pmd(pmdp, pmd, addr, next, | |
111 | write, pages, nr)) | |
112 | return 0; | |
113 | } else if (!gup_pte_range(pmdp, pmd, addr, next, | |
114 | write, pages, nr)) | |
115 | return 0; | |
116 | } while (pmdp++, addr = next, addr != end); | |
117 | ||
118 | return 1; | |
119 | } | |
120 | ||
121 | static inline int gup_pud_range(pgd_t *pgdp, pgd_t pgd, unsigned long addr, | |
122 | unsigned long end, int write, struct page **pages, int *nr) | |
123 | { | |
124 | unsigned long next; | |
125 | pud_t *pudp, pud; | |
126 | ||
127 | pudp = (pud_t *) pgdp; | |
128 | #ifdef CONFIG_64BIT | |
129 | if ((pgd_val(pgd) & _REGION_ENTRY_TYPE_MASK) == _REGION_ENTRY_TYPE_R2) | |
130 | pudp = (pud_t *) pgd_deref(pgd); | |
131 | pudp += pud_index(addr); | |
132 | #endif | |
133 | do { | |
134 | pud = *pudp; | |
135 | barrier(); | |
136 | next = pud_addr_end(addr, end); | |
137 | if (pud_none(pud)) | |
138 | return 0; | |
139 | if (!gup_pmd_range(pudp, pud, addr, next, write, pages, nr)) | |
140 | return 0; | |
141 | } while (pudp++, addr = next, addr != end); | |
142 | ||
143 | return 1; | |
144 | } | |
145 | ||
146 | /** | |
147 | * get_user_pages_fast() - pin user pages in memory | |
148 | * @start: starting user address | |
149 | * @nr_pages: number of pages from start to pin | |
150 | * @write: whether pages will be written to | |
151 | * @pages: array that receives pointers to the pages pinned. | |
152 | * Should be at least nr_pages long. | |
153 | * | |
154 | * Attempt to pin user pages in memory without taking mm->mmap_sem. | |
155 | * If not successful, it will fall back to taking the lock and | |
156 | * calling get_user_pages(). | |
157 | * | |
158 | * Returns number of pages pinned. This may be fewer than the number | |
159 | * requested. If nr_pages is 0 or negative, returns 0. If no pages | |
160 | * were pinned, returns -errno. | |
161 | */ | |
162 | int get_user_pages_fast(unsigned long start, int nr_pages, int write, | |
163 | struct page **pages) | |
164 | { | |
165 | struct mm_struct *mm = current->mm; | |
166 | unsigned long addr, len, end; | |
167 | unsigned long next; | |
168 | pgd_t *pgdp, pgd; | |
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 (end < start) | |
176 | goto slow_irqon; | |
177 | ||
178 | /* | |
179 | * local_irq_disable() doesn't prevent pagetable teardown, but does | |
180 | * prevent the pagetables from being freed on s390. | |
181 | * | |
182 | * So long as we atomically load page table pointers versus teardown, | |
183 | * we can follow the address down to the the page and take a ref on it. | |
184 | */ | |
185 | local_irq_disable(); | |
186 | pgdp = pgd_offset(mm, addr); | |
187 | do { | |
188 | pgd = *pgdp; | |
189 | barrier(); | |
190 | next = pgd_addr_end(addr, end); | |
191 | if (pgd_none(pgd)) | |
192 | goto slow; | |
193 | if (!gup_pud_range(pgdp, pgd, addr, next, write, pages, &nr)) | |
194 | goto slow; | |
195 | } while (pgdp++, addr = next, addr != end); | |
196 | local_irq_enable(); | |
197 | ||
198 | VM_BUG_ON(nr != (end - start) >> PAGE_SHIFT); | |
199 | return nr; | |
200 | ||
201 | { | |
202 | int ret; | |
203 | slow: | |
204 | local_irq_enable(); | |
205 | slow_irqon: | |
206 | /* Try to get the remaining pages with get_user_pages */ | |
207 | start += nr << PAGE_SHIFT; | |
208 | pages += nr; | |
209 | ||
210 | down_read(&mm->mmap_sem); | |
211 | ret = get_user_pages(current, mm, start, | |
212 | (end - start) >> PAGE_SHIFT, write, 0, pages, NULL); | |
213 | up_read(&mm->mmap_sem); | |
214 | ||
215 | /* Have to be a bit careful with return values */ | |
216 | if (nr > 0) { | |
217 | if (ret < 0) | |
218 | ret = nr; | |
219 | else | |
220 | ret += nr; | |
221 | } | |
222 | ||
223 | return ret; | |
224 | } | |
225 | } |