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         if (! mm->used_hugetlb)
 150                 return ERR_PTR(-EINVAL);
 151
 152         vma = find_vma(mm, addr);
 153         if (!vma || !is_vm_hugetlb_page(vma))
 154                 return ERR_PTR(-EINVAL);
 155
 156         pte = huge_pte_offset(mm, address);
 157
 158         /* hugetlb should be locked, and hence, prefaulted */
 159         WARN_ON(!pte || pte_none(*pte));
 160
 161         page = &pte_page(*pte)[vpfn % (HPAGE_SIZE/PAGE_SIZE)];
 162
 163         WARN_ON(!PageCompound(page));
 164
 165         return page;
 166 }
 167
 168 int pmd_huge(pmd_t pmd)
 169 {
 170         return 0;
 171 }
 172
 173 struct page *
 174 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
 175                 pmd_t *pmd, int write)
 176 {
 177         return NULL;
 178 }
 179
 180 #else
 181
 182 struct page *
 183 follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
 184 {
 185         return ERR_PTR(-EINVAL);
 186 }
 187
 188 int pmd_huge(pmd_t pmd)
 189 {
 190         return !!(pmd_val(pmd) & _PAGE_PSE);
 191 }
 192
 193 struct page *
 194 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
 195                 pmd_t *pmd, int write)
 196 {
 197         struct page *page;
 198
 199         page = pte_page(*(pte_t *)pmd);
 200         if (page)
 201                 page += ((address & ~HPAGE_MASK) >> PAGE_SHIFT);
 202         return page;
 203 }
 204 #endif
 205
 206 void unmap_hugepage_range(struct vm_area_struct *vma,
 207                 unsigned long start, unsigned long end)
 208 {
 209         struct mm_struct *mm = vma->vm_mm;
 210         unsigned long address;
 211         pte_t pte;
 212         struct page *page;
 213
 214         BUG_ON(start & (HPAGE_SIZE - 1));
 215         BUG_ON(end & (HPAGE_SIZE - 1));
 216
 217         for (address = start; address < end; address += HPAGE_SIZE) {
 218                 pte = ptep_get_and_clear(huge_pte_offset(mm, address));
 219                 if (pte_none(pte))
 220                         continue;
 221                 page = pte_page(pte);
 222                 put_page(page);
 223         }
 224         mm->rss -= (end - start) >> PAGE_SHIFT;
 225         flush_tlb_range(vma, start, end);
 226 }
 227
 228 int hugetlb_prefault(struct address_space *mapping, struct vm_area_struct *vma)
 229 {
 230         struct mm_struct *mm = current->mm;
 231         unsigned long addr;
 232         int ret = 0;
 233
 234         BUG_ON(vma->vm_start & ~HPAGE_MASK);
 235         BUG_ON(vma->vm_end & ~HPAGE_MASK);
 236
 237         spin_lock(&mm->page_table_lock);
 238         for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
 239                 unsigned long idx;
 240                 pte_t *pte = huge_pte_alloc(mm, addr);
 241                 struct page *page;
 242
 243                 if (!pte) {
 244                         ret = -ENOMEM;
 245                         goto out;
 246                 }
 247
 248                 if (!pte_none(*pte)) {
 249                         pmd_t *pmd = (pmd_t *) pte;
 250
 251                         page = pmd_page(*pmd);
 252                         pmd_clear(pmd);
 253                         dec_page_state(nr_page_table_pages);
 254                         page_cache_release(page);
 255                 }
 256
 257                 idx = ((addr - vma->vm_start) >> HPAGE_SHIFT)
 258                         + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
 259                 page = find_get_page(mapping, idx);
 260                 if (!page) {
 261                         /* charge the fs quota first */
 262                         if (hugetlb_get_quota(mapping)) {
 263                                 ret = -ENOMEM;
 264                                 goto out;
 265                         }
 266                         page = alloc_huge_page();
 267                         if (!page) {
 268                                 hugetlb_put_quota(mapping);
 269                                 ret = -ENOMEM;
 270                                 goto out;
 271                         }
 272                         ret = add_to_page_cache(page, mapping, idx, GFP_ATOMIC);
 273                         if (! ret) {
 274                                 unlock_page(page);
 275                         } else {
 276                                 hugetlb_put_quota(mapping);
 277                                 free_huge_page(page);
 278                                 goto out;
 279                         }
 280                 }
 281                 set_huge_pte(mm, vma, page, pte, vma->vm_flags & VM_WRITE);
 282         }
 283 out:
 284         spin_unlock(&mm->page_table_lock);
 285         return ret;
 286 }