Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * SPARC64 Huge TLB page support. | |
3 | * | |
f6b83f07 | 4 | * Copyright (C) 2002, 2003, 2006 David S. Miller (davem@davemloft.net) |
1da177e4 LT |
5 | */ |
6 | ||
7 | #include <linux/config.h> | |
8 | #include <linux/init.h> | |
9 | #include <linux/module.h> | |
10 | #include <linux/fs.h> | |
11 | #include <linux/mm.h> | |
12 | #include <linux/hugetlb.h> | |
13 | #include <linux/pagemap.h> | |
14 | #include <linux/smp_lock.h> | |
15 | #include <linux/slab.h> | |
16 | #include <linux/sysctl.h> | |
17 | ||
18 | #include <asm/mman.h> | |
19 | #include <asm/pgalloc.h> | |
20 | #include <asm/tlb.h> | |
21 | #include <asm/tlbflush.h> | |
22 | #include <asm/cacheflush.h> | |
23 | #include <asm/mmu_context.h> | |
24 | ||
f6b83f07 DM |
25 | /* Slightly simplified from the non-hugepage variant because by |
26 | * definition we don't have to worry about any page coloring stuff | |
27 | */ | |
28 | #define VA_EXCLUDE_START (0x0000080000000000UL - (1UL << 32UL)) | |
29 | #define VA_EXCLUDE_END (0xfffff80000000000UL + (1UL << 32UL)) | |
30 | ||
31 | static unsigned long hugetlb_get_unmapped_area_bottomup(struct file *filp, | |
32 | unsigned long addr, | |
33 | unsigned long len, | |
34 | unsigned long pgoff, | |
35 | unsigned long flags) | |
36 | { | |
37 | struct mm_struct *mm = current->mm; | |
38 | struct vm_area_struct * vma; | |
39 | unsigned long task_size = TASK_SIZE; | |
40 | unsigned long start_addr; | |
41 | ||
42 | if (test_thread_flag(TIF_32BIT)) | |
43 | task_size = STACK_TOP32; | |
44 | if (unlikely(len >= VA_EXCLUDE_START)) | |
45 | return -ENOMEM; | |
46 | ||
47 | if (len > mm->cached_hole_size) { | |
48 | start_addr = addr = mm->free_area_cache; | |
49 | } else { | |
50 | start_addr = addr = TASK_UNMAPPED_BASE; | |
51 | mm->cached_hole_size = 0; | |
52 | } | |
53 | ||
54 | task_size -= len; | |
55 | ||
56 | full_search: | |
57 | addr = ALIGN(addr, HPAGE_SIZE); | |
58 | ||
59 | for (vma = find_vma(mm, addr); ; vma = vma->vm_next) { | |
60 | /* At this point: (!vma || addr < vma->vm_end). */ | |
61 | if (addr < VA_EXCLUDE_START && | |
62 | (addr + len) >= VA_EXCLUDE_START) { | |
63 | addr = VA_EXCLUDE_END; | |
64 | vma = find_vma(mm, VA_EXCLUDE_END); | |
65 | } | |
66 | if (unlikely(task_size < addr)) { | |
67 | if (start_addr != TASK_UNMAPPED_BASE) { | |
68 | start_addr = addr = TASK_UNMAPPED_BASE; | |
69 | mm->cached_hole_size = 0; | |
70 | goto full_search; | |
71 | } | |
72 | return -ENOMEM; | |
73 | } | |
74 | if (likely(!vma || addr + len <= vma->vm_start)) { | |
75 | /* | |
76 | * Remember the place where we stopped the search: | |
77 | */ | |
78 | mm->free_area_cache = addr + len; | |
79 | return addr; | |
80 | } | |
81 | if (addr + mm->cached_hole_size < vma->vm_start) | |
82 | mm->cached_hole_size = vma->vm_start - addr; | |
83 | ||
84 | addr = ALIGN(vma->vm_end, HPAGE_SIZE); | |
85 | } | |
86 | } | |
87 | ||
88 | static unsigned long | |
89 | hugetlb_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0, | |
90 | const unsigned long len, | |
91 | const unsigned long pgoff, | |
92 | const unsigned long flags) | |
93 | { | |
94 | struct vm_area_struct *vma; | |
95 | struct mm_struct *mm = current->mm; | |
96 | unsigned long addr = addr0; | |
97 | ||
98 | /* This should only ever run for 32-bit processes. */ | |
99 | BUG_ON(!test_thread_flag(TIF_32BIT)); | |
100 | ||
101 | /* check if free_area_cache is useful for us */ | |
102 | if (len <= mm->cached_hole_size) { | |
103 | mm->cached_hole_size = 0; | |
104 | mm->free_area_cache = mm->mmap_base; | |
105 | } | |
106 | ||
107 | /* either no address requested or can't fit in requested address hole */ | |
108 | addr = mm->free_area_cache & HPAGE_MASK; | |
109 | ||
110 | /* make sure it can fit in the remaining address space */ | |
111 | if (likely(addr > len)) { | |
112 | vma = find_vma(mm, addr-len); | |
113 | if (!vma || addr <= vma->vm_start) { | |
114 | /* remember the address as a hint for next time */ | |
115 | return (mm->free_area_cache = addr-len); | |
116 | } | |
117 | } | |
118 | ||
119 | if (unlikely(mm->mmap_base < len)) | |
120 | goto bottomup; | |
121 | ||
122 | addr = (mm->mmap_base-len) & HPAGE_MASK; | |
123 | ||
124 | do { | |
125 | /* | |
126 | * Lookup failure means no vma is above this address, | |
127 | * else if new region fits below vma->vm_start, | |
128 | * return with success: | |
129 | */ | |
130 | vma = find_vma(mm, addr); | |
131 | if (likely(!vma || addr+len <= vma->vm_start)) { | |
132 | /* remember the address as a hint for next time */ | |
133 | return (mm->free_area_cache = addr); | |
134 | } | |
135 | ||
136 | /* remember the largest hole we saw so far */ | |
137 | if (addr + mm->cached_hole_size < vma->vm_start) | |
138 | mm->cached_hole_size = vma->vm_start - addr; | |
139 | ||
140 | /* try just below the current vma->vm_start */ | |
141 | addr = (vma->vm_start-len) & HPAGE_MASK; | |
142 | } while (likely(len < vma->vm_start)); | |
143 | ||
144 | bottomup: | |
145 | /* | |
146 | * A failed mmap() very likely causes application failure, | |
147 | * so fall back to the bottom-up function here. This scenario | |
148 | * can happen with large stack limits and large mmap() | |
149 | * allocations. | |
150 | */ | |
151 | mm->cached_hole_size = ~0UL; | |
152 | mm->free_area_cache = TASK_UNMAPPED_BASE; | |
153 | addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags); | |
154 | /* | |
155 | * Restore the topdown base: | |
156 | */ | |
157 | mm->free_area_cache = mm->mmap_base; | |
158 | mm->cached_hole_size = ~0UL; | |
159 | ||
160 | return addr; | |
161 | } | |
162 | ||
163 | unsigned long | |
164 | hugetlb_get_unmapped_area(struct file *file, unsigned long addr, | |
165 | unsigned long len, unsigned long pgoff, unsigned long flags) | |
166 | { | |
167 | struct mm_struct *mm = current->mm; | |
168 | struct vm_area_struct *vma; | |
169 | unsigned long task_size = TASK_SIZE; | |
170 | ||
171 | if (test_thread_flag(TIF_32BIT)) | |
172 | task_size = STACK_TOP32; | |
173 | ||
174 | if (len & ~HPAGE_MASK) | |
175 | return -EINVAL; | |
176 | if (len > task_size) | |
177 | return -ENOMEM; | |
178 | ||
179 | if (addr) { | |
180 | addr = ALIGN(addr, HPAGE_SIZE); | |
181 | vma = find_vma(mm, addr); | |
182 | if (task_size - len >= addr && | |
183 | (!vma || addr + len <= vma->vm_start)) | |
184 | return addr; | |
185 | } | |
186 | if (mm->get_unmapped_area == arch_get_unmapped_area) | |
187 | return hugetlb_get_unmapped_area_bottomup(file, addr, len, | |
188 | pgoff, flags); | |
189 | else | |
190 | return hugetlb_get_unmapped_area_topdown(file, addr, len, | |
191 | pgoff, flags); | |
192 | } | |
193 | ||
63551ae0 | 194 | pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr) |
1da177e4 LT |
195 | { |
196 | pgd_t *pgd; | |
197 | pud_t *pud; | |
198 | pmd_t *pmd; | |
199 | pte_t *pte = NULL; | |
200 | ||
201 | pgd = pgd_offset(mm, addr); | |
202 | if (pgd) { | |
203 | pud = pud_offset(pgd, addr); | |
204 | if (pud) { | |
205 | pmd = pmd_alloc(mm, pud, addr); | |
206 | if (pmd) | |
207 | pte = pte_alloc_map(mm, pmd, addr); | |
208 | } | |
209 | } | |
210 | return pte; | |
211 | } | |
212 | ||
63551ae0 | 213 | pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr) |
1da177e4 LT |
214 | { |
215 | pgd_t *pgd; | |
216 | pud_t *pud; | |
217 | pmd_t *pmd; | |
218 | pte_t *pte = NULL; | |
219 | ||
f6b83f07 DM |
220 | addr &= HPAGE_MASK; |
221 | ||
1da177e4 | 222 | pgd = pgd_offset(mm, addr); |
f6b83f07 | 223 | if (!pgd_none(*pgd)) { |
1da177e4 | 224 | pud = pud_offset(pgd, addr); |
f6b83f07 | 225 | if (!pud_none(*pud)) { |
1da177e4 | 226 | pmd = pmd_offset(pud, addr); |
f6b83f07 | 227 | if (!pmd_none(*pmd)) |
1da177e4 LT |
228 | pte = pte_offset_map(pmd, addr); |
229 | } | |
230 | } | |
231 | return pte; | |
232 | } | |
233 | ||
234 | #define mk_pte_huge(entry) do { pte_val(entry) |= _PAGE_SZHUGE; } while (0) | |
235 | ||
63551ae0 DG |
236 | void set_huge_pte_at(struct mm_struct *mm, unsigned long addr, |
237 | pte_t *ptep, pte_t entry) | |
1da177e4 | 238 | { |
63551ae0 DG |
239 | int i; |
240 | ||
241 | for (i = 0; i < (1 << HUGETLB_PAGE_ORDER); i++) { | |
242 | set_pte_at(mm, addr, ptep, entry); | |
243 | ptep++; | |
244 | addr += PAGE_SIZE; | |
245 | pte_val(entry) += PAGE_SIZE; | |
246 | } | |
247 | } | |
1da177e4 | 248 | |
63551ae0 DG |
249 | pte_t huge_ptep_get_and_clear(struct mm_struct *mm, unsigned long addr, |
250 | pte_t *ptep) | |
251 | { | |
252 | pte_t entry; | |
253 | int i; | |
1da177e4 | 254 | |
63551ae0 | 255 | entry = *ptep; |
1da177e4 LT |
256 | |
257 | for (i = 0; i < (1 << HUGETLB_PAGE_ORDER); i++) { | |
63551ae0 | 258 | pte_clear(mm, addr, ptep); |
1da177e4 | 259 | addr += PAGE_SIZE; |
63551ae0 | 260 | ptep++; |
1da177e4 | 261 | } |
63551ae0 DG |
262 | |
263 | return entry; | |
1da177e4 LT |
264 | } |
265 | ||
1da177e4 LT |
266 | struct page *follow_huge_addr(struct mm_struct *mm, |
267 | unsigned long address, int write) | |
268 | { | |
269 | return ERR_PTR(-EINVAL); | |
270 | } | |
271 | ||
272 | int pmd_huge(pmd_t pmd) | |
273 | { | |
274 | return 0; | |
275 | } | |
276 | ||
277 | struct page *follow_huge_pmd(struct mm_struct *mm, unsigned long address, | |
278 | pmd_t *pmd, int write) | |
279 | { | |
280 | return NULL; | |
281 | } | |
282 | ||
1da177e4 LT |
283 | static void context_reload(void *__data) |
284 | { | |
285 | struct mm_struct *mm = __data; | |
286 | ||
287 | if (mm == current->mm) | |
288 | load_secondary_context(mm); | |
289 | } | |
290 | ||
63551ae0 | 291 | void hugetlb_prefault_arch_hook(struct mm_struct *mm) |
1da177e4 | 292 | { |
1da177e4 LT |
293 | /* On UltraSPARC-III+ and later, configure the second half of |
294 | * the Data-TLB for huge pages. | |
295 | */ | |
296 | if (tlb_type == cheetah_plus) { | |
297 | unsigned long ctx; | |
298 | ||
299 | spin_lock(&ctx_alloc_lock); | |
300 | ctx = mm->context.sparc64_ctx_val; | |
301 | ctx &= ~CTX_PGSZ_MASK; | |
302 | ctx |= CTX_PGSZ_BASE << CTX_PGSZ0_SHIFT; | |
303 | ctx |= CTX_PGSZ_HUGE << CTX_PGSZ1_SHIFT; | |
304 | ||
305 | if (ctx != mm->context.sparc64_ctx_val) { | |
306 | /* When changing the page size fields, we | |
307 | * must perform a context flush so that no | |
308 | * stale entries match. This flush must | |
309 | * occur with the original context register | |
310 | * settings. | |
311 | */ | |
312 | do_flush_tlb_mm(mm); | |
313 | ||
314 | /* Reload the context register of all processors | |
315 | * also executing in this address space. | |
316 | */ | |
317 | mm->context.sparc64_ctx_val = ctx; | |
318 | on_each_cpu(context_reload, mm, 0, 0); | |
319 | } | |
320 | spin_unlock(&ctx_alloc_lock); | |
321 | } | |
1da177e4 | 322 | } |