Jeremy Fitzhardinge | 4f76cd3 | 2008-03-17 16:36:55 -0700 | [diff] [blame^] | 1 | #include <linux/mm.h> |
| 2 | #include <asm/pgalloc.h> |
| 3 | #include <asm/tlb.h> |
| 4 | |
| 5 | pte_t *pte_alloc_one_kernel(struct mm_struct *mm, unsigned long address) |
| 6 | { |
| 7 | return (pte_t *)__get_free_page(GFP_KERNEL|__GFP_REPEAT|__GFP_ZERO); |
| 8 | } |
| 9 | |
| 10 | pgtable_t pte_alloc_one(struct mm_struct *mm, unsigned long address) |
| 11 | { |
| 12 | struct page *pte; |
| 13 | |
| 14 | #ifdef CONFIG_HIGHPTE |
| 15 | pte = alloc_pages(GFP_KERNEL|__GFP_HIGHMEM|__GFP_REPEAT|__GFP_ZERO, 0); |
| 16 | #else |
| 17 | pte = alloc_pages(GFP_KERNEL|__GFP_REPEAT|__GFP_ZERO, 0); |
| 18 | #endif |
| 19 | if (pte) |
| 20 | pgtable_page_ctor(pte); |
| 21 | return pte; |
| 22 | } |
| 23 | |
| 24 | #ifdef CONFIG_X86_64 |
| 25 | static inline void pgd_list_add(pgd_t *pgd) |
| 26 | { |
| 27 | struct page *page = virt_to_page(pgd); |
| 28 | unsigned long flags; |
| 29 | |
| 30 | spin_lock_irqsave(&pgd_lock, flags); |
| 31 | list_add(&page->lru, &pgd_list); |
| 32 | spin_unlock_irqrestore(&pgd_lock, flags); |
| 33 | } |
| 34 | |
| 35 | static inline void pgd_list_del(pgd_t *pgd) |
| 36 | { |
| 37 | struct page *page = virt_to_page(pgd); |
| 38 | unsigned long flags; |
| 39 | |
| 40 | spin_lock_irqsave(&pgd_lock, flags); |
| 41 | list_del(&page->lru); |
| 42 | spin_unlock_irqrestore(&pgd_lock, flags); |
| 43 | } |
| 44 | |
| 45 | pgd_t *pgd_alloc(struct mm_struct *mm) |
| 46 | { |
| 47 | unsigned boundary; |
| 48 | pgd_t *pgd = (pgd_t *)__get_free_page(GFP_KERNEL|__GFP_REPEAT); |
| 49 | if (!pgd) |
| 50 | return NULL; |
| 51 | pgd_list_add(pgd); |
| 52 | /* |
| 53 | * Copy kernel pointers in from init. |
| 54 | * Could keep a freelist or slab cache of those because the kernel |
| 55 | * part never changes. |
| 56 | */ |
| 57 | boundary = pgd_index(__PAGE_OFFSET); |
| 58 | memset(pgd, 0, boundary * sizeof(pgd_t)); |
| 59 | memcpy(pgd + boundary, |
| 60 | init_level4_pgt + boundary, |
| 61 | (PTRS_PER_PGD - boundary) * sizeof(pgd_t)); |
| 62 | return pgd; |
| 63 | } |
| 64 | |
| 65 | void pgd_free(struct mm_struct *mm, pgd_t *pgd) |
| 66 | { |
| 67 | BUG_ON((unsigned long)pgd & (PAGE_SIZE-1)); |
| 68 | pgd_list_del(pgd); |
| 69 | free_page((unsigned long)pgd); |
| 70 | } |
| 71 | #else |
| 72 | /* |
| 73 | * List of all pgd's needed for non-PAE so it can invalidate entries |
| 74 | * in both cached and uncached pgd's; not needed for PAE since the |
| 75 | * kernel pmd is shared. If PAE were not to share the pmd a similar |
| 76 | * tactic would be needed. This is essentially codepath-based locking |
| 77 | * against pageattr.c; it is the unique case in which a valid change |
| 78 | * of kernel pagetables can't be lazily synchronized by vmalloc faults. |
| 79 | * vmalloc faults work because attached pagetables are never freed. |
| 80 | * -- wli |
| 81 | */ |
| 82 | static inline void pgd_list_add(pgd_t *pgd) |
| 83 | { |
| 84 | struct page *page = virt_to_page(pgd); |
| 85 | |
| 86 | list_add(&page->lru, &pgd_list); |
| 87 | } |
| 88 | |
| 89 | static inline void pgd_list_del(pgd_t *pgd) |
| 90 | { |
| 91 | struct page *page = virt_to_page(pgd); |
| 92 | |
| 93 | list_del(&page->lru); |
| 94 | } |
| 95 | |
| 96 | #define UNSHARED_PTRS_PER_PGD \ |
| 97 | (SHARED_KERNEL_PMD ? USER_PTRS_PER_PGD : PTRS_PER_PGD) |
| 98 | |
| 99 | static void pgd_ctor(void *p) |
| 100 | { |
| 101 | pgd_t *pgd = p; |
| 102 | unsigned long flags; |
| 103 | |
| 104 | /* Clear usermode parts of PGD */ |
| 105 | memset(pgd, 0, USER_PTRS_PER_PGD*sizeof(pgd_t)); |
| 106 | |
| 107 | spin_lock_irqsave(&pgd_lock, flags); |
| 108 | |
| 109 | /* If the pgd points to a shared pagetable level (either the |
| 110 | ptes in non-PAE, or shared PMD in PAE), then just copy the |
| 111 | references from swapper_pg_dir. */ |
| 112 | if (PAGETABLE_LEVELS == 2 || |
| 113 | (PAGETABLE_LEVELS == 3 && SHARED_KERNEL_PMD)) { |
| 114 | clone_pgd_range(pgd + USER_PTRS_PER_PGD, |
| 115 | swapper_pg_dir + USER_PTRS_PER_PGD, |
| 116 | KERNEL_PGD_PTRS); |
| 117 | paravirt_alloc_pd_clone(__pa(pgd) >> PAGE_SHIFT, |
| 118 | __pa(swapper_pg_dir) >> PAGE_SHIFT, |
| 119 | USER_PTRS_PER_PGD, |
| 120 | KERNEL_PGD_PTRS); |
| 121 | } |
| 122 | |
| 123 | /* list required to sync kernel mapping updates */ |
| 124 | if (!SHARED_KERNEL_PMD) |
| 125 | pgd_list_add(pgd); |
| 126 | |
| 127 | spin_unlock_irqrestore(&pgd_lock, flags); |
| 128 | } |
| 129 | |
| 130 | static void pgd_dtor(void *pgd) |
| 131 | { |
| 132 | unsigned long flags; /* can be called from interrupt context */ |
| 133 | |
| 134 | if (SHARED_KERNEL_PMD) |
| 135 | return; |
| 136 | |
| 137 | spin_lock_irqsave(&pgd_lock, flags); |
| 138 | pgd_list_del(pgd); |
| 139 | spin_unlock_irqrestore(&pgd_lock, flags); |
| 140 | } |
| 141 | |
| 142 | #ifdef CONFIG_X86_PAE |
| 143 | /* |
| 144 | * Mop up any pmd pages which may still be attached to the pgd. |
| 145 | * Normally they will be freed by munmap/exit_mmap, but any pmd we |
| 146 | * preallocate which never got a corresponding vma will need to be |
| 147 | * freed manually. |
| 148 | */ |
| 149 | static void pgd_mop_up_pmds(struct mm_struct *mm, pgd_t *pgdp) |
| 150 | { |
| 151 | int i; |
| 152 | |
| 153 | for(i = 0; i < UNSHARED_PTRS_PER_PGD; i++) { |
| 154 | pgd_t pgd = pgdp[i]; |
| 155 | |
| 156 | if (pgd_val(pgd) != 0) { |
| 157 | pmd_t *pmd = (pmd_t *)pgd_page_vaddr(pgd); |
| 158 | |
| 159 | pgdp[i] = native_make_pgd(0); |
| 160 | |
| 161 | paravirt_release_pd(pgd_val(pgd) >> PAGE_SHIFT); |
| 162 | pmd_free(mm, pmd); |
| 163 | } |
| 164 | } |
| 165 | } |
| 166 | |
| 167 | /* |
| 168 | * In PAE mode, we need to do a cr3 reload (=tlb flush) when |
| 169 | * updating the top-level pagetable entries to guarantee the |
| 170 | * processor notices the update. Since this is expensive, and |
| 171 | * all 4 top-level entries are used almost immediately in a |
| 172 | * new process's life, we just pre-populate them here. |
| 173 | * |
| 174 | * Also, if we're in a paravirt environment where the kernel pmd is |
| 175 | * not shared between pagetables (!SHARED_KERNEL_PMDS), we allocate |
| 176 | * and initialize the kernel pmds here. |
| 177 | */ |
| 178 | static int pgd_prepopulate_pmd(struct mm_struct *mm, pgd_t *pgd) |
| 179 | { |
| 180 | pud_t *pud; |
| 181 | unsigned long addr; |
| 182 | int i; |
| 183 | |
| 184 | pud = pud_offset(pgd, 0); |
| 185 | for (addr = i = 0; i < UNSHARED_PTRS_PER_PGD; |
| 186 | i++, pud++, addr += PUD_SIZE) { |
| 187 | pmd_t *pmd = pmd_alloc_one(mm, addr); |
| 188 | |
| 189 | if (!pmd) { |
| 190 | pgd_mop_up_pmds(mm, pgd); |
| 191 | return 0; |
| 192 | } |
| 193 | |
| 194 | if (i >= USER_PTRS_PER_PGD) |
| 195 | memcpy(pmd, (pmd_t *)pgd_page_vaddr(swapper_pg_dir[i]), |
| 196 | sizeof(pmd_t) * PTRS_PER_PMD); |
| 197 | |
| 198 | pud_populate(mm, pud, pmd); |
| 199 | } |
| 200 | |
| 201 | return 1; |
| 202 | } |
| 203 | #else /* !CONFIG_X86_PAE */ |
| 204 | /* No need to prepopulate any pagetable entries in non-PAE modes. */ |
| 205 | static int pgd_prepopulate_pmd(struct mm_struct *mm, pgd_t *pgd) |
| 206 | { |
| 207 | return 1; |
| 208 | } |
| 209 | |
| 210 | static void pgd_mop_up_pmds(struct mm_struct *mm, pgd_t *pgd) |
| 211 | { |
| 212 | } |
| 213 | #endif /* CONFIG_X86_PAE */ |
| 214 | |
| 215 | pgd_t *pgd_alloc(struct mm_struct *mm) |
| 216 | { |
| 217 | pgd_t *pgd = (pgd_t *)__get_free_page(GFP_KERNEL | __GFP_ZERO); |
| 218 | |
| 219 | /* so that alloc_pd can use it */ |
| 220 | mm->pgd = pgd; |
| 221 | if (pgd) |
| 222 | pgd_ctor(pgd); |
| 223 | |
| 224 | if (pgd && !pgd_prepopulate_pmd(mm, pgd)) { |
| 225 | pgd_dtor(pgd); |
| 226 | free_page((unsigned long)pgd); |
| 227 | pgd = NULL; |
| 228 | } |
| 229 | |
| 230 | return pgd; |
| 231 | } |
| 232 | |
| 233 | void pgd_free(struct mm_struct *mm, pgd_t *pgd) |
| 234 | { |
| 235 | pgd_mop_up_pmds(mm, pgd); |
| 236 | pgd_dtor(pgd); |
| 237 | free_page((unsigned long)pgd); |
| 238 | } |
| 239 | #endif |