Chris Metcalf | 867e359 | 2010-05-28 23:09:12 -0400 | [diff] [blame] | 1 | /* |
| 2 | * Copyright 2010 Tilera Corporation. All Rights Reserved. |
| 3 | * |
| 4 | * This program is free software; you can redistribute it and/or |
| 5 | * modify it under the terms of the GNU General Public License |
| 6 | * as published by the Free Software Foundation, version 2. |
| 7 | * |
| 8 | * This program is distributed in the hope that it will be useful, but |
| 9 | * WITHOUT ANY WARRANTY; without even the implied warranty of |
| 10 | * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or |
| 11 | * NON INFRINGEMENT. See the GNU General Public License for |
| 12 | * more details. |
| 13 | * |
| 14 | * TILE Huge TLB Page Support for Kernel. |
| 15 | * Taken from i386 hugetlb implementation: |
| 16 | * Copyright (C) 2002, Rohit Seth <rohit.seth@intel.com> |
| 17 | */ |
| 18 | |
| 19 | #include <linux/init.h> |
| 20 | #include <linux/fs.h> |
| 21 | #include <linux/mm.h> |
| 22 | #include <linux/hugetlb.h> |
| 23 | #include <linux/pagemap.h> |
| 24 | #include <linux/smp_lock.h> |
| 25 | #include <linux/slab.h> |
| 26 | #include <linux/err.h> |
| 27 | #include <linux/sysctl.h> |
| 28 | #include <linux/mman.h> |
| 29 | #include <asm/tlb.h> |
| 30 | #include <asm/tlbflush.h> |
| 31 | |
| 32 | pte_t *huge_pte_alloc(struct mm_struct *mm, |
| 33 | unsigned long addr, unsigned long sz) |
| 34 | { |
| 35 | pgd_t *pgd; |
| 36 | pud_t *pud; |
| 37 | pte_t *pte = NULL; |
| 38 | |
| 39 | /* We do not yet support multiple huge page sizes. */ |
| 40 | BUG_ON(sz != PMD_SIZE); |
| 41 | |
| 42 | pgd = pgd_offset(mm, addr); |
| 43 | pud = pud_alloc(mm, pgd, addr); |
| 44 | if (pud) |
| 45 | pte = (pte_t *) pmd_alloc(mm, pud, addr); |
| 46 | BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte)); |
| 47 | |
| 48 | return pte; |
| 49 | } |
| 50 | |
| 51 | pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr) |
| 52 | { |
| 53 | pgd_t *pgd; |
| 54 | pud_t *pud; |
| 55 | pmd_t *pmd = NULL; |
| 56 | |
| 57 | pgd = pgd_offset(mm, addr); |
| 58 | if (pgd_present(*pgd)) { |
| 59 | pud = pud_offset(pgd, addr); |
| 60 | if (pud_present(*pud)) |
| 61 | pmd = pmd_offset(pud, addr); |
| 62 | } |
| 63 | return (pte_t *) pmd; |
| 64 | } |
| 65 | |
| 66 | #ifdef HUGETLB_TEST |
| 67 | struct page *follow_huge_addr(struct mm_struct *mm, unsigned long address, |
| 68 | int write) |
| 69 | { |
| 70 | unsigned long start = address; |
| 71 | int length = 1; |
| 72 | int nr; |
| 73 | struct page *page; |
| 74 | struct vm_area_struct *vma; |
| 75 | |
| 76 | vma = find_vma(mm, addr); |
| 77 | if (!vma || !is_vm_hugetlb_page(vma)) |
| 78 | return ERR_PTR(-EINVAL); |
| 79 | |
| 80 | pte = huge_pte_offset(mm, address); |
| 81 | |
| 82 | /* hugetlb should be locked, and hence, prefaulted */ |
| 83 | WARN_ON(!pte || pte_none(*pte)); |
| 84 | |
| 85 | page = &pte_page(*pte)[vpfn % (HPAGE_SIZE/PAGE_SIZE)]; |
| 86 | |
| 87 | WARN_ON(!PageHead(page)); |
| 88 | |
| 89 | return page; |
| 90 | } |
| 91 | |
| 92 | int pmd_huge(pmd_t pmd) |
| 93 | { |
| 94 | return 0; |
| 95 | } |
| 96 | |
| 97 | int pud_huge(pud_t pud) |
| 98 | { |
| 99 | return 0; |
| 100 | } |
| 101 | |
| 102 | struct page *follow_huge_pmd(struct mm_struct *mm, unsigned long address, |
| 103 | pmd_t *pmd, int write) |
| 104 | { |
| 105 | return NULL; |
| 106 | } |
| 107 | |
| 108 | #else |
| 109 | |
| 110 | struct page *follow_huge_addr(struct mm_struct *mm, unsigned long address, |
| 111 | int write) |
| 112 | { |
| 113 | return ERR_PTR(-EINVAL); |
| 114 | } |
| 115 | |
| 116 | int pmd_huge(pmd_t pmd) |
| 117 | { |
| 118 | return !!(pmd_val(pmd) & _PAGE_HUGE_PAGE); |
| 119 | } |
| 120 | |
| 121 | int pud_huge(pud_t pud) |
| 122 | { |
| 123 | return !!(pud_val(pud) & _PAGE_HUGE_PAGE); |
| 124 | } |
| 125 | |
| 126 | struct page *follow_huge_pmd(struct mm_struct *mm, unsigned long address, |
| 127 | pmd_t *pmd, int write) |
| 128 | { |
| 129 | struct page *page; |
| 130 | |
| 131 | page = pte_page(*(pte_t *)pmd); |
| 132 | if (page) |
| 133 | page += ((address & ~PMD_MASK) >> PAGE_SHIFT); |
| 134 | return page; |
| 135 | } |
| 136 | |
| 137 | struct page *follow_huge_pud(struct mm_struct *mm, unsigned long address, |
| 138 | pud_t *pud, int write) |
| 139 | { |
| 140 | struct page *page; |
| 141 | |
| 142 | page = pte_page(*(pte_t *)pud); |
| 143 | if (page) |
| 144 | page += ((address & ~PUD_MASK) >> PAGE_SHIFT); |
| 145 | return page; |
| 146 | } |
| 147 | |
| 148 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) |
| 149 | { |
| 150 | return 0; |
| 151 | } |
| 152 | |
| 153 | #endif |
| 154 | |
| 155 | #ifdef HAVE_ARCH_HUGETLB_UNMAPPED_AREA |
| 156 | static unsigned long hugetlb_get_unmapped_area_bottomup(struct file *file, |
| 157 | unsigned long addr, unsigned long len, |
| 158 | unsigned long pgoff, unsigned long flags) |
| 159 | { |
| 160 | struct hstate *h = hstate_file(file); |
| 161 | struct mm_struct *mm = current->mm; |
| 162 | struct vm_area_struct *vma; |
| 163 | unsigned long start_addr; |
| 164 | |
| 165 | if (len > mm->cached_hole_size) { |
| 166 | start_addr = mm->free_area_cache; |
| 167 | } else { |
| 168 | start_addr = TASK_UNMAPPED_BASE; |
| 169 | mm->cached_hole_size = 0; |
| 170 | } |
| 171 | |
| 172 | full_search: |
| 173 | addr = ALIGN(start_addr, huge_page_size(h)); |
| 174 | |
| 175 | for (vma = find_vma(mm, addr); ; vma = vma->vm_next) { |
| 176 | /* At this point: (!vma || addr < vma->vm_end). */ |
| 177 | if (TASK_SIZE - len < addr) { |
| 178 | /* |
| 179 | * Start a new search - just in case we missed |
| 180 | * some holes. |
| 181 | */ |
| 182 | if (start_addr != TASK_UNMAPPED_BASE) { |
| 183 | start_addr = TASK_UNMAPPED_BASE; |
| 184 | mm->cached_hole_size = 0; |
| 185 | goto full_search; |
| 186 | } |
| 187 | return -ENOMEM; |
| 188 | } |
| 189 | if (!vma || addr + len <= vma->vm_start) { |
| 190 | mm->free_area_cache = addr + len; |
| 191 | return addr; |
| 192 | } |
| 193 | if (addr + mm->cached_hole_size < vma->vm_start) |
| 194 | mm->cached_hole_size = vma->vm_start - addr; |
| 195 | addr = ALIGN(vma->vm_end, huge_page_size(h)); |
| 196 | } |
| 197 | } |
| 198 | |
| 199 | static unsigned long hugetlb_get_unmapped_area_topdown(struct file *file, |
| 200 | unsigned long addr0, unsigned long len, |
| 201 | unsigned long pgoff, unsigned long flags) |
| 202 | { |
| 203 | struct hstate *h = hstate_file(file); |
| 204 | struct mm_struct *mm = current->mm; |
| 205 | struct vm_area_struct *vma, *prev_vma; |
| 206 | unsigned long base = mm->mmap_base, addr = addr0; |
| 207 | unsigned long largest_hole = mm->cached_hole_size; |
| 208 | int first_time = 1; |
| 209 | |
| 210 | /* don't allow allocations above current base */ |
| 211 | if (mm->free_area_cache > base) |
| 212 | mm->free_area_cache = base; |
| 213 | |
| 214 | if (len <= largest_hole) { |
| 215 | largest_hole = 0; |
| 216 | mm->free_area_cache = base; |
| 217 | } |
| 218 | try_again: |
| 219 | /* make sure it can fit in the remaining address space */ |
| 220 | if (mm->free_area_cache < len) |
| 221 | goto fail; |
| 222 | |
| 223 | /* either no address requested or cant fit in requested address hole */ |
| 224 | addr = (mm->free_area_cache - len) & huge_page_mask(h); |
| 225 | do { |
| 226 | /* |
| 227 | * Lookup failure means no vma is above this address, |
| 228 | * i.e. return with success: |
| 229 | */ |
| 230 | vma = find_vma_prev(mm, addr, &prev_vma); |
| 231 | if (!vma) { |
| 232 | return addr; |
| 233 | break; |
| 234 | } |
| 235 | |
| 236 | /* |
| 237 | * new region fits between prev_vma->vm_end and |
| 238 | * vma->vm_start, use it: |
| 239 | */ |
| 240 | if (addr + len <= vma->vm_start && |
| 241 | (!prev_vma || (addr >= prev_vma->vm_end))) { |
| 242 | /* remember the address as a hint for next time */ |
| 243 | mm->cached_hole_size = largest_hole; |
| 244 | mm->free_area_cache = addr; |
| 245 | return addr; |
| 246 | } else { |
| 247 | /* pull free_area_cache down to the first hole */ |
| 248 | if (mm->free_area_cache == vma->vm_end) { |
| 249 | mm->free_area_cache = vma->vm_start; |
| 250 | mm->cached_hole_size = largest_hole; |
| 251 | } |
| 252 | } |
| 253 | |
| 254 | /* remember the largest hole we saw so far */ |
| 255 | if (addr + largest_hole < vma->vm_start) |
| 256 | largest_hole = vma->vm_start - addr; |
| 257 | |
| 258 | /* try just below the current vma->vm_start */ |
| 259 | addr = (vma->vm_start - len) & huge_page_mask(h); |
| 260 | |
| 261 | } while (len <= vma->vm_start); |
| 262 | |
| 263 | fail: |
| 264 | /* |
| 265 | * if hint left us with no space for the requested |
| 266 | * mapping then try again: |
| 267 | */ |
| 268 | if (first_time) { |
| 269 | mm->free_area_cache = base; |
| 270 | largest_hole = 0; |
| 271 | first_time = 0; |
| 272 | goto try_again; |
| 273 | } |
| 274 | /* |
| 275 | * A failed mmap() very likely causes application failure, |
| 276 | * so fall back to the bottom-up function here. This scenario |
| 277 | * can happen with large stack limits and large mmap() |
| 278 | * allocations. |
| 279 | */ |
| 280 | mm->free_area_cache = TASK_UNMAPPED_BASE; |
| 281 | mm->cached_hole_size = ~0UL; |
| 282 | addr = hugetlb_get_unmapped_area_bottomup(file, addr0, |
| 283 | len, pgoff, flags); |
| 284 | |
| 285 | /* |
| 286 | * Restore the topdown base: |
| 287 | */ |
| 288 | mm->free_area_cache = base; |
| 289 | mm->cached_hole_size = ~0UL; |
| 290 | |
| 291 | return addr; |
| 292 | } |
| 293 | |
| 294 | unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr, |
| 295 | unsigned long len, unsigned long pgoff, unsigned long flags) |
| 296 | { |
| 297 | struct hstate *h = hstate_file(file); |
| 298 | struct mm_struct *mm = current->mm; |
| 299 | struct vm_area_struct *vma; |
| 300 | |
| 301 | if (len & ~huge_page_mask(h)) |
| 302 | return -EINVAL; |
| 303 | if (len > TASK_SIZE) |
| 304 | return -ENOMEM; |
| 305 | |
| 306 | if (flags & MAP_FIXED) { |
| 307 | if (prepare_hugepage_range(file, addr, len)) |
| 308 | return -EINVAL; |
| 309 | return addr; |
| 310 | } |
| 311 | |
| 312 | if (addr) { |
| 313 | addr = ALIGN(addr, huge_page_size(h)); |
| 314 | vma = find_vma(mm, addr); |
| 315 | if (TASK_SIZE - len >= addr && |
| 316 | (!vma || addr + len <= vma->vm_start)) |
| 317 | return addr; |
| 318 | } |
| 319 | if (current->mm->get_unmapped_area == arch_get_unmapped_area) |
| 320 | return hugetlb_get_unmapped_area_bottomup(file, addr, len, |
| 321 | pgoff, flags); |
| 322 | else |
| 323 | return hugetlb_get_unmapped_area_topdown(file, addr, len, |
| 324 | pgoff, flags); |
| 325 | } |
| 326 | |
| 327 | static __init int setup_hugepagesz(char *opt) |
| 328 | { |
| 329 | unsigned long ps = memparse(opt, &opt); |
| 330 | if (ps == PMD_SIZE) { |
| 331 | hugetlb_add_hstate(PMD_SHIFT - PAGE_SHIFT); |
| 332 | } else if (ps == PUD_SIZE) { |
| 333 | hugetlb_add_hstate(PUD_SHIFT - PAGE_SHIFT); |
| 334 | } else { |
Chris Metcalf | 0707ad3 | 2010-06-25 17:04:17 -0400 | [diff] [blame] | 335 | pr_err("hugepagesz: Unsupported page size %lu M\n", |
Chris Metcalf | 867e359 | 2010-05-28 23:09:12 -0400 | [diff] [blame] | 336 | ps >> 20); |
| 337 | return 0; |
| 338 | } |
| 339 | return 1; |
| 340 | } |
| 341 | __setup("hugepagesz=", setup_hugepagesz); |
| 342 | |
| 343 | #endif /*HAVE_ARCH_HUGETLB_UNMAPPED_AREA*/ |