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/tlb.h>
  19 #include <asm/tlbflush.h>
  20
  21 static pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr)
  22 {
  23         pgd_t *pgd;
  24         pmd_t *pmd = NULL;
  25
  26         pgd = pgd_offset(mm, addr);
  27         pmd = pmd_alloc(mm, pgd, addr);
  28         return (pte_t *) pmd;
  29 }
  30
  31 static pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
  32 {
  33         pgd_t *pgd;
  34         pmd_t *pmd = NULL;
  35
  36         pgd = pgd_offset(mm, addr);
  37         pmd = pmd_offset(pgd, addr);
  38         return (pte_t *) pmd;
  39 }
  40
  41 static void set_huge_pte(struct mm_struct *mm, struct vm_area_struct *vma, struct page *page, pte_t * page_table, int write_access)
  42 {
  43         pte_t entry;
  44
  45         mm->rss += (HPAGE_SIZE / PAGE_SIZE);
  46         if (write_access) {
  47                 entry =
  48                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
  49         } else
  50                 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
  51         entry = pte_mkyoung(entry);
  52         mk_pte_huge(entry);
  53         set_pte(page_table, entry);
  54 }
  55
  56 /*
  57  * This function checks for proper alignment of input addr and len parameters.
  58  */
  59 int is_aligned_hugepage_range(unsigned long addr, unsigned long len)
  60 {
  61         if (len & ~HPAGE_MASK)
  62                 return -EINVAL;
  63         if (addr & ~HPAGE_MASK)
  64                 return -EINVAL;
  65         return 0;
  66 }
  67
  68 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
  69                         struct vm_area_struct *vma)
  70 {
  71         pte_t *src_pte, *dst_pte, entry;
  72         struct page *ptepage;
  73         unsigned long addr = vma->vm_start;
  74         unsigned long end = vma->vm_end;
  75
  76         while (addr < end) {
  77                 dst_pte = huge_pte_alloc(dst, addr);
  78                 if (!dst_pte)
  79                         goto nomem;
  80                 src_pte = huge_pte_offset(src, addr);
  81                 entry = *src_pte;
  82                 ptepage = pte_page(entry);
  83                 get_page(ptepage);
  84                 set_pte(dst_pte, entry);
  85                 dst->rss += (HPAGE_SIZE / PAGE_SIZE);
  86                 addr += HPAGE_SIZE;
  87         }
  88         return 0;
  89
  90 nomem:
  91         return -ENOMEM;
  92 }
  93
  94 int
  95 follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
  96                     struct page **pages, struct vm_area_struct **vmas,
  97                     unsigned long *position, int *length, int i)
  98 {
  99         unsigned long vpfn, vaddr = *position;
 100         int remainder = *length;
 101
 102         WARN_ON(!is_vm_hugetlb_page(vma));
 103
 104         vpfn = vaddr/PAGE_SIZE;
 105         while (vaddr < vma->vm_end && remainder) {
 106
 107                 if (pages) {
 108                         pte_t *pte;
 109                         struct page *page;
 110
 111                         pte = huge_pte_offset(mm, vaddr);
 112
 113                         /* hugetlb should be locked, and hence, prefaulted */
 114                         WARN_ON(!pte || pte_none(*pte));
 115
 116                         page = &pte_page(*pte)[vpfn % (HPAGE_SIZE/PAGE_SIZE)];
 117
 118                         WARN_ON(!PageCompound(page));
 119
 120                         get_page(page);
 121                         pages[i] = page;
 122                 }
 123
 124                 if (vmas)
 125                         vmas[i] = vma;
 126
 127                 vaddr += PAGE_SIZE;
 128                 ++vpfn;
 129                 --remainder;
 130                 ++i;
 131         }
 132
 133         *length = remainder;
 134         *position = vaddr;
 135
 136         return i;
 137 }
 138
 139 #if 0   /* This is just for testing */
 140 struct page *
 141 follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
 142 {
 143         unsigned long start = address;
 144         int length = 1;
 145         int nr;
 146         struct page *page;
 147         struct vm_area_struct *vma;
 148
 149         vma = find_vma(mm, addr);
 150         if (!vma || !is_vm_hugetlb_page(vma))
 151                 return ERR_PTR(-EINVAL);
 152
 153         pte = huge_pte_offset(mm, address);
 154
 155         /* hugetlb should be locked, and hence, prefaulted */
 156         WARN_ON(!pte || pte_none(*pte));
 157
 158         page = &pte_page(*pte)[vpfn % (HPAGE_SIZE/PAGE_SIZE)];
 159
 160         WARN_ON(!PageCompound(page));
 161
 162         return page;
 163 }
 164
 165 int pmd_huge(pmd_t pmd)
 166 {
 167         return 0;
 168 }
 169
 170 struct page *
 171 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
 172                 pmd_t *pmd, int write)
 173 {
 174         return NULL;
 175 }
 176
 177 #else
 178
 179 struct page *
 180 follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
 181 {
 182         return ERR_PTR(-EINVAL);
 183 }
 184
 185 int pmd_huge(pmd_t pmd)
 186 {
 187         return !!(pmd_val(pmd) & _PAGE_PSE);
 188 }
 189
 190 struct page *
 191 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
 192                 pmd_t *pmd, int write)
 193 {
 194         struct page *page;
 195
 196         page = pte_page(*(pte_t *)pmd);
 197         if (page)
 198                 page += ((address & ~HPAGE_MASK) >> PAGE_SHIFT);
 199         return page;
 200 }
 201 #endif
 202
 203 void unmap_hugepage_range(struct vm_area_struct *vma,
 204                 unsigned long start, unsigned long end)
 205 {
 206         struct mm_struct *mm = vma->vm_mm;
 207         unsigned long address;
 208         pte_t pte;
 209         struct page *page;
 210
 211         BUG_ON(start & (HPAGE_SIZE - 1));
 212         BUG_ON(end & (HPAGE_SIZE - 1));
 213
 214         for (address = start; address < end; address += HPAGE_SIZE) {
 215                 pte = ptep_get_and_clear(huge_pte_offset(mm, address));
 216                 if (pte_none(pte))
 217                         continue;
 218                 page = pte_page(pte);
 219                 put_page(page);
 220         }
 221         mm->rss -= (end - start) >> PAGE_SHIFT;
 222         flush_tlb_range(vma, start, end);
 223 }
 224
 225 int hugetlb_prefault(struct address_space *mapping, struct vm_area_struct *vma)
 226 {
 227         struct mm_struct *mm = current->mm;
 228         unsigned long addr;
 229         int ret = 0;
 230
 231         BUG_ON(vma->vm_start & ~HPAGE_MASK);
 232         BUG_ON(vma->vm_end & ~HPAGE_MASK);
 233
 234         spin_lock(&mm->page_table_lock);
 235         for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
 236                 unsigned long idx;
 237                 pte_t *pte = huge_pte_alloc(mm, addr);
 238                 struct page *page;
 239
 240                 if (!pte) {
 241                         ret = -ENOMEM;
 242                         goto out;
 243                 }
 244
 245                 if (!pte_none(*pte)) {
 246                         pmd_t *pmd = (pmd_t *) pte;
 247
 248                         page = pmd_page(*pmd);
 249                         pmd_clear(pmd);
 250                         dec_page_state(nr_page_table_pages);
 251                         page_cache_release(page);
 252                 }
 253
 254                 idx = ((addr - vma->vm_start) >> HPAGE_SHIFT)
 255                         + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
 256                 page = find_get_page(mapping, idx);
 257                 if (!page) {
 258                         /* charge the fs quota first */
 259                         if (hugetlb_get_quota(mapping)) {
 260                                 ret = -ENOMEM;
 261                                 goto out;
 262                         }
 263                         page = alloc_huge_page();
 264                         if (!page) {
 265                                 hugetlb_put_quota(mapping);
 266                                 ret = -ENOMEM;
 267                                 goto out;
 268                         }
 269                         ret = add_to_page_cache(page, mapping, idx, GFP_ATOMIC);
 270                         if (! ret) {
 271                                 unlock_page(page);
 272                         } else {
 273                                 hugetlb_put_quota(mapping);
 274                                 free_huge_page(page);
 275                                 goto out;
 276                         }
 277                 }
 278                 set_huge_pte(mm, vma, page, pte, vma->vm_flags & VM_WRITE);
 279         }
 280 out:
 281         spin_unlock(&mm->page_table_lock);
 282         return ret;
 283 }