arch/i386/mm/hugetlbpage.c

   1 /*
   2  * IA-32 Huge TLB Page Support for Kernel.
   3  *
   4  * Copyright (C) 2002, Rohit Seth <rohit.seth@intel.com>
   5  */
   6
   7 #include <linux/config.h>
   8 #include <linux/init.h>
   9 #include <linux/fs.h>
  10 #include <linux/mm.h>
  11 #include <linux/hugetlb.h>
  12 #include <linux/pagemap.h>
  13 #include <linux/smp_lock.h>
  14 #include <linux/slab.h>
  15 #include <linux/err.h>
  16 #include <linux/sysctl.h>
  17 #include <asm/mman.h>
  18 #include <asm/pgalloc.h>
  19 #include <asm/tlb.h>
  20 #include <asm/tlbflush.h>
  21
  22 static pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr)
  23 {
  24         pgd_t *pgd;
  25         pmd_t *pmd = NULL;
  26
  27         pgd = pgd_offset(mm, addr);
  28         pmd = pmd_alloc(mm, pgd, addr);
  29         return (pte_t *) pmd;
  30 }
  31
  32 static pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
  33 {
  34         pgd_t *pgd;
  35         pmd_t *pmd = NULL;
  36
  37         pgd = pgd_offset(mm, addr);
  38         pmd = pmd_offset(pgd, addr);
  39         return (pte_t *) pmd;
  40 }
  41
  42 static void set_huge_pte(struct mm_struct *mm, struct vm_area_struct *vma, struct page *page, pte_t * page_table, int write_access)
  43 {
  44         pte_t entry;
  45
  46         mm->rss += (HPAGE_SIZE / PAGE_SIZE);
  47         if (write_access) {
  48                 entry =
  49                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
  50         } else
  51                 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
  52         entry = pte_mkyoung(entry);
  53         mk_pte_huge(entry);
  54         set_pte(page_table, entry);
  55 }
  56
  57 /*
  58  * This function checks for proper alignment of input addr and len parameters.
  59  */
  60 int is_aligned_hugepage_range(unsigned long addr, unsigned long len)
  61 {
  62         if (len & ~HPAGE_MASK)
  63                 return -EINVAL;
  64         if (addr & ~HPAGE_MASK)
  65                 return -EINVAL;
  66         return 0;
  67 }
  68
  69 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
  70                         struct vm_area_struct *vma)
  71 {
  72         pte_t *src_pte, *dst_pte, entry;
  73         struct page *ptepage;
  74         unsigned long addr = vma->vm_start;
  75         unsigned long end = vma->vm_end;
  76
  77         while (addr < end) {
  78                 dst_pte = huge_pte_alloc(dst, addr);
  79                 if (!dst_pte)
  80                         goto nomem;
  81                 src_pte = huge_pte_offset(src, addr);
  82                 entry = *src_pte;
  83                 ptepage = pte_page(entry);
  84                 get_page(ptepage);
  85                 set_pte(dst_pte, entry);
  86                 dst->rss += (HPAGE_SIZE / PAGE_SIZE);
  87                 addr += HPAGE_SIZE;
  88         }
  89         return 0;
  90
  91 nomem:
  92         return -ENOMEM;
  93 }
  94
  95 int
  96 follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
  97                     struct page **pages, struct vm_area_struct **vmas,
  98                     unsigned long *position, int *length, int i)
  99 {
 100         unsigned long vpfn, vaddr = *position;
 101         int remainder = *length;
 102
 103         WARN_ON(!is_vm_hugetlb_page(vma));
 104
 105         vpfn = vaddr/PAGE_SIZE;
 106         while (vaddr < vma->vm_end && remainder) {
 107
 108                 if (pages) {
 109                         pte_t *pte;
 110                         struct page *page;
 111
 112                         pte = huge_pte_offset(mm, vaddr);
 113
 114                         /* hugetlb should be locked, and hence, prefaulted */
 115                         WARN_ON(!pte || pte_none(*pte));
 116
 117                         page = &pte_page(*pte)[vpfn % (HPAGE_SIZE/PAGE_SIZE)];
 118
 119                         WARN_ON(!PageCompound(page));
 120
 121                         get_page(page);
 122                         pages[i] = page;
 123                 }
 124
 125                 if (vmas)
 126                         vmas[i] = vma;
 127
 128                 vaddr += PAGE_SIZE;
 129                 ++vpfn;
 130                 --remainder;
 131                 ++i;
 132         }
 133
 134         *length = remainder;
 135         *position = vaddr;
 136
 137         return i;
 138 }
 139
 140 #if 0   /* This is just for testing */
 141 struct page *
 142 follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
 143 {
 144         unsigned long start = address;
 145         int length = 1;
 146         int nr;
 147         struct page *page;
 148         struct vm_area_struct *vma;
 149
 150         if (! mm->used_hugetlb)
 151                 return ERR_PTR(-EINVAL);
 152
 153         vma = find_vma(mm, addr);
 154         if (!vma || !is_vm_hugetlb_page(vma))
 155                 return ERR_PTR(-EINVAL);
 156
 157         pte = huge_pte_offset(mm, address);
 158
 159         /* hugetlb should be locked, and hence, prefaulted */
 160         WARN_ON(!pte || pte_none(*pte));
 161
 162         page = &pte_page(*pte)[vpfn % (HPAGE_SIZE/PAGE_SIZE)];
 163
 164         WARN_ON(!PageCompound(page));
 165
 166         return page;
 167 }
 168
 169 int pmd_huge(pmd_t pmd)
 170 {
 171         return 0;
 172 }
 173
 174 struct page *
 175 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
 176                 pmd_t *pmd, int write)
 177 {
 178         return NULL;
 179 }
 180
 181 #else
 182
 183 struct page *
 184 follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
 185 {
 186         return ERR_PTR(-EINVAL);
 187 }
 188
 189 int pmd_huge(pmd_t pmd)
 190 {
 191         return !!(pmd_val(pmd) & _PAGE_PSE);
 192 }
 193
 194 struct page *
 195 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
 196                 pmd_t *pmd, int write)
 197 {
 198         struct page *page;
 199
 200         page = pte_page(*(pte_t *)pmd);
 201         if (page)
 202                 page += ((address & ~HPAGE_MASK) >> PAGE_SHIFT);
 203         return page;
 204 }
 205 #endif
 206
 207 void unmap_hugepage_range(struct vm_area_struct *vma,
 208                 unsigned long start, unsigned long end)
 209 {
 210         struct mm_struct *mm = vma->vm_mm;
 211         unsigned long address;
 212         pte_t pte;
 213         struct page *page;
 214
 215         BUG_ON(start & (HPAGE_SIZE - 1));
 216         BUG_ON(end & (HPAGE_SIZE - 1));
 217
 218         for (address = start; address < end; address += HPAGE_SIZE) {
 219                 pte = ptep_get_and_clear(huge_pte_offset(mm, address));
 220                 if (pte_none(pte))
 221                         continue;
 222                 page = pte_page(pte);
 223                 put_page(page);
 224         }
 225         mm->rss -= (end - start) >> PAGE_SHIFT;
 226         flush_tlb_range(vma, start, end);
 227 }
 228
 229 int hugetlb_prefault(struct address_space *mapping, struct vm_area_struct *vma)
 230 {
 231         struct mm_struct *mm = current->mm;
 232         unsigned long addr;
 233         int ret = 0;
 234
 235         BUG_ON(vma->vm_start & ~HPAGE_MASK);
 236         BUG_ON(vma->vm_end & ~HPAGE_MASK);
 237
 238         spin_lock(&mm->page_table_lock);
 239         for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
 240                 unsigned long idx;
 241                 pte_t *pte = huge_pte_alloc(mm, addr);
 242                 struct page *page;
 243
 244                 if (!pte) {
 245                         ret = -ENOMEM;
 246                         goto out;
 247                 }
 248                 if (!pte_none(*pte))
 249                         continue;
 250
 251                 idx = ((addr - vma->vm_start) >> HPAGE_SHIFT)
 252                         + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
 253                 page = find_get_page(mapping, idx);
 254                 if (!page) {
 255                         /* charge the fs quota first */
 256                         if (hugetlb_get_quota(mapping)) {
 257                                 ret = -ENOMEM;
 258                                 goto out;
 259                         }
 260                         page = alloc_huge_page();
 261                         if (!page) {
 262                                 hugetlb_put_quota(mapping);
 263                                 ret = -ENOMEM;
 264                                 goto out;
 265                         }
 266                         ret = add_to_page_cache(page, mapping, idx, GFP_ATOMIC);
 267                         unlock_page(page);
 268                         if (ret) {
 269                                 hugetlb_put_quota(mapping);
 270                                 free_huge_page(page);
 271                                 goto out;
 272                         }
 273                 }
 274                 set_huge_pte(mm, vma, page, pte, vma->vm_flags & VM_WRITE);
 275         }
 276 out:
 277         spin_unlock(&mm->page_table_lock);
 278         return ret;
 279 }