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