Martin Schwidefsky | 8021714 | 2010-10-25 16:10:11 +0200 | [diff] [blame] | 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 | { |
Martin Schwidefsky | 25591b0 | 2010-11-10 10:05:51 +0100 | [diff] [blame] | 23 | unsigned long mask; |
Martin Schwidefsky | 8021714 | 2010-10-25 16:10:11 +0200 | [diff] [blame] | 24 | pte_t *ptep, pte; |
| 25 | struct page *page; |
| 26 | |
Martin Schwidefsky | 25591b0 | 2010-11-10 10:05:51 +0100 | [diff] [blame] | 27 | mask = (write ? _PAGE_RO : 0) | _PAGE_INVALID | _PAGE_SPECIAL; |
Martin Schwidefsky | 8021714 | 2010-10-25 16:10:11 +0200 | [diff] [blame] | 28 | |
| 29 | ptep = ((pte_t *) pmd_deref(pmd)) + pte_index(addr); |
| 30 | do { |
| 31 | pte = *ptep; |
| 32 | barrier(); |
Martin Schwidefsky | 25591b0 | 2010-11-10 10:05:51 +0100 | [diff] [blame] | 33 | if ((pte_val(pte) & mask) != 0) |
Martin Schwidefsky | 8021714 | 2010-10-25 16:10:11 +0200 | [diff] [blame] | 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; |
| 55 | struct page *head, *page; |
| 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); |
| 67 | do { |
| 68 | VM_BUG_ON(compound_head(page) != head); |
| 69 | pages[*nr] = page; |
| 70 | (*nr)++; |
| 71 | page++; |
| 72 | refs++; |
| 73 | } while (addr += PAGE_SIZE, addr != end); |
| 74 | |
| 75 | if (!page_cache_add_speculative(head, refs)) { |
| 76 | *nr -= refs; |
| 77 | return 0; |
| 78 | } |
| 79 | |
| 80 | if (unlikely(pmd_val(pmd) != pmd_val(*pmdp))) { |
| 81 | *nr -= refs; |
| 82 | while (refs--) |
| 83 | put_page(head); |
| 84 | } |
| 85 | |
| 86 | return 1; |
| 87 | } |
| 88 | |
| 89 | |
| 90 | static inline int gup_pmd_range(pud_t *pudp, pud_t pud, unsigned long addr, |
| 91 | unsigned long end, int write, struct page **pages, int *nr) |
| 92 | { |
| 93 | unsigned long next; |
| 94 | pmd_t *pmdp, pmd; |
| 95 | |
| 96 | pmdp = (pmd_t *) pudp; |
| 97 | #ifdef CONFIG_64BIT |
| 98 | if ((pud_val(pud) & _REGION_ENTRY_TYPE_MASK) == _REGION_ENTRY_TYPE_R3) |
| 99 | pmdp = (pmd_t *) pud_deref(pud); |
| 100 | pmdp += pmd_index(addr); |
| 101 | #endif |
| 102 | do { |
| 103 | pmd = *pmdp; |
| 104 | barrier(); |
| 105 | next = pmd_addr_end(addr, end); |
| 106 | if (pmd_none(pmd)) |
| 107 | return 0; |
| 108 | if (unlikely(pmd_huge(pmd))) { |
| 109 | if (!gup_huge_pmd(pmdp, pmd, addr, next, |
| 110 | write, pages, nr)) |
| 111 | return 0; |
| 112 | } else if (!gup_pte_range(pmdp, pmd, addr, next, |
| 113 | write, pages, nr)) |
| 114 | return 0; |
| 115 | } while (pmdp++, addr = next, addr != end); |
| 116 | |
| 117 | return 1; |
| 118 | } |
| 119 | |
| 120 | static inline int gup_pud_range(pgd_t *pgdp, pgd_t pgd, unsigned long addr, |
| 121 | unsigned long end, int write, struct page **pages, int *nr) |
| 122 | { |
| 123 | unsigned long next; |
| 124 | pud_t *pudp, pud; |
| 125 | |
| 126 | pudp = (pud_t *) pgdp; |
| 127 | #ifdef CONFIG_64BIT |
| 128 | if ((pgd_val(pgd) & _REGION_ENTRY_TYPE_MASK) == _REGION_ENTRY_TYPE_R2) |
| 129 | pudp = (pud_t *) pgd_deref(pgd); |
| 130 | pudp += pud_index(addr); |
| 131 | #endif |
| 132 | do { |
| 133 | pud = *pudp; |
| 134 | barrier(); |
| 135 | next = pud_addr_end(addr, end); |
| 136 | if (pud_none(pud)) |
| 137 | return 0; |
| 138 | if (!gup_pmd_range(pudp, pud, addr, next, write, pages, nr)) |
| 139 | return 0; |
| 140 | } while (pudp++, addr = next, addr != end); |
| 141 | |
| 142 | return 1; |
| 143 | } |
| 144 | |
| 145 | /** |
| 146 | * get_user_pages_fast() - pin user pages in memory |
| 147 | * @start: starting user address |
| 148 | * @nr_pages: number of pages from start to pin |
| 149 | * @write: whether pages will be written to |
| 150 | * @pages: array that receives pointers to the pages pinned. |
| 151 | * Should be at least nr_pages long. |
| 152 | * |
| 153 | * Attempt to pin user pages in memory without taking mm->mmap_sem. |
| 154 | * If not successful, it will fall back to taking the lock and |
| 155 | * calling get_user_pages(). |
| 156 | * |
| 157 | * Returns number of pages pinned. This may be fewer than the number |
| 158 | * requested. If nr_pages is 0 or negative, returns 0. If no pages |
| 159 | * were pinned, returns -errno. |
| 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 | pgd_t *pgdp, pgd; |
| 168 | int nr = 0; |
| 169 | |
| 170 | start &= PAGE_MASK; |
| 171 | addr = start; |
| 172 | len = (unsigned long) nr_pages << PAGE_SHIFT; |
| 173 | end = start + len; |
| 174 | if (end < start) |
| 175 | goto slow_irqon; |
| 176 | |
| 177 | /* |
| 178 | * local_irq_disable() doesn't prevent pagetable teardown, but does |
| 179 | * prevent the pagetables from being freed on s390. |
| 180 | * |
| 181 | * So long as we atomically load page table pointers versus teardown, |
| 182 | * we can follow the address down to the the page and take a ref on it. |
| 183 | */ |
| 184 | local_irq_disable(); |
| 185 | pgdp = pgd_offset(mm, addr); |
| 186 | do { |
| 187 | pgd = *pgdp; |
| 188 | barrier(); |
| 189 | next = pgd_addr_end(addr, end); |
| 190 | if (pgd_none(pgd)) |
| 191 | goto slow; |
| 192 | if (!gup_pud_range(pgdp, pgd, addr, next, write, pages, &nr)) |
| 193 | goto slow; |
| 194 | } while (pgdp++, addr = next, addr != end); |
| 195 | local_irq_enable(); |
| 196 | |
| 197 | VM_BUG_ON(nr != (end - start) >> PAGE_SHIFT); |
| 198 | return nr; |
| 199 | |
| 200 | { |
| 201 | int ret; |
| 202 | slow: |
| 203 | local_irq_enable(); |
| 204 | slow_irqon: |
| 205 | /* Try to get the remaining pages with get_user_pages */ |
| 206 | start += nr << PAGE_SHIFT; |
| 207 | pages += nr; |
| 208 | |
| 209 | down_read(&mm->mmap_sem); |
| 210 | ret = get_user_pages(current, mm, start, |
| 211 | (end - start) >> PAGE_SHIFT, write, 0, pages, NULL); |
| 212 | up_read(&mm->mmap_sem); |
| 213 | |
| 214 | /* Have to be a bit careful with return values */ |
| 215 | if (nr > 0) { |
| 216 | if (ret < 0) |
| 217 | ret = nr; |
| 218 | else |
| 219 | ret += nr; |
| 220 | } |
| 221 | |
| 222 | return ret; |
| 223 | } |
| 224 | } |