arch/sparc64/mm/hugetlbpage.c

   1 /*
   2  * SPARC64 Huge TLB page support.
   3  *
   4  * Copyright (C) 2002, 2003 David S. Miller (davem@redhat.com)
   5  */
   6
   7 #include <linux/config.h>
   8 #include <linux/init.h>
   9 #include <linux/module.h>
  10 #include <linux/fs.h>
  11 #include <linux/mm.h>
  12 #include <linux/hugetlb.h>
  13 #include <linux/pagemap.h>
  14 #include <linux/smp_lock.h>
  15 #include <linux/slab.h>
  16 #include <linux/sysctl.h>
  17
  18 #include <asm/mman.h>
  19 #include <asm/pgalloc.h>
  20 #include <asm/tlb.h>
  21 #include <asm/tlbflush.h>
  22 #include <asm/cacheflush.h>
  23
  24 static pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr)
  25 {
  26         pgd_t *pgd;
  27         pud_t *pud;
  28         pmd_t *pmd;
  29         pte_t *pte = NULL;
  30
  31         pgd = pgd_offset(mm, addr);
  32         if (pgd) {
  33                 pud = pud_offset(pgd, addr);
  34                 if (pud) {
  35                         pmd = pmd_alloc(mm, pud, addr);
  36                         if (pmd)
  37                                 pte = pte_alloc_map(mm, pmd, addr);
  38                 }
  39         }
  40         return pte;
  41 }
  42
  43 static pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
  44 {
  45         pgd_t *pgd;
  46         pud_t *pud;
  47         pmd_t *pmd;
  48         pte_t *pte = NULL;
  49
  50         pgd = pgd_offset(mm, addr);
  51         if (pgd) {
  52                 pud = pud_offset(pgd, addr);
  53                 if (pud) {
  54                         pmd = pmd_offset(pud, addr);
  55                         if (pmd)
  56                                 pte = pte_offset_map(pmd, addr);
  57                 }
  58         }
  59         return pte;
  60 }
  61
  62 #define mk_pte_huge(entry) do { pte_val(entry) |= _PAGE_SZHUGE; } while (0)
  63
  64 static void set_huge_pte(struct mm_struct *mm, struct vm_area_struct *vma,
  65                          struct page *page, pte_t * page_table, int write_access)
  66 {
  67         unsigned long i;
  68         pte_t entry;
  69
  70         vx_rsspages_add(mm, HPAGE_SIZE / PAGE_SIZE);
  71
  72         if (write_access)
  73                 entry = pte_mkwrite(pte_mkdirty(mk_pte(page,
  74                                                        vma->vm_page_prot)));
  75         else
  76                 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
  77         entry = pte_mkyoung(entry);
  78         mk_pte_huge(entry);
  79
  80         for (i = 0; i < (1 << HUGETLB_PAGE_ORDER); i++) {
  81                 set_pte(page_table, entry);
  82                 page_table++;
  83
  84                 pte_val(entry) += PAGE_SIZE;
  85         }
  86 }
  87
  88 /*
  89  * This function checks for proper alignment of input addr and len parameters.
  90  */
  91 int is_aligned_hugepage_range(unsigned long addr, unsigned long len)
  92 {
  93         if (len & ~HPAGE_MASK)
  94                 return -EINVAL;
  95         if (addr & ~HPAGE_MASK)
  96                 return -EINVAL;
  97         return 0;
  98 }
  99
 100 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
 101                             struct vm_area_struct *vma)
 102 {
 103         pte_t *src_pte, *dst_pte, entry;
 104         struct page *ptepage;
 105         unsigned long addr = vma->vm_start;
 106         unsigned long end = vma->vm_end;
 107         int i;
 108
 109         while (addr < end) {
 110                 dst_pte = huge_pte_alloc(dst, addr);
 111                 if (!dst_pte)
 112                         goto nomem;
 113                 src_pte = huge_pte_offset(src, addr);
 114                 BUG_ON(!src_pte || pte_none(*src_pte));
 115                 entry = *src_pte;
 116                 ptepage = pte_page(entry);
 117                 get_page(ptepage);
 118                 for (i = 0; i < (1 << HUGETLB_PAGE_ORDER); i++) {
 119                         set_pte(dst_pte, entry);
 120                         pte_val(entry) += PAGE_SIZE;
 121                         dst_pte++;
 122                 }
 123                 vx_rsspages_add(dst, HPAGE_SIZE / PAGE_SIZE);
 124                 addr += HPAGE_SIZE;
 125         }
 126         return 0;
 127
 128 nomem:
 129         return -ENOMEM;
 130 }
 131
 132 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
 133                         struct page **pages, struct vm_area_struct **vmas,
 134                         unsigned long *position, int *length, int i)
 135 {
 136         unsigned long vaddr = *position;
 137         int remainder = *length;
 138
 139         WARN_ON(!is_vm_hugetlb_page(vma));
 140
 141         while (vaddr < vma->vm_end && remainder) {
 142                 if (pages) {
 143                         pte_t *pte;
 144                         struct page *page;
 145
 146                         pte = huge_pte_offset(mm, vaddr);
 147
 148                         /* hugetlb should be locked, and hence, prefaulted */
 149                         BUG_ON(!pte || pte_none(*pte));
 150
 151                         page = pte_page(*pte);
 152
 153                         WARN_ON(!PageCompound(page));
 154
 155                         get_page(page);
 156                         pages[i] = page;
 157                 }
 158
 159                 if (vmas)
 160                         vmas[i] = vma;
 161
 162                 vaddr += PAGE_SIZE;
 163                 --remainder;
 164                 ++i;
 165         }
 166
 167         *length = remainder;
 168         *position = vaddr;
 169
 170         return i;
 171 }
 172
 173 struct page *follow_huge_addr(struct mm_struct *mm,
 174                               unsigned long address, int write)
 175 {
 176         return ERR_PTR(-EINVAL);
 177 }
 178
 179 int pmd_huge(pmd_t pmd)
 180 {
 181         return 0;
 182 }
 183
 184 struct page *follow_huge_pmd(struct mm_struct *mm, unsigned long address,
 185                              pmd_t *pmd, int write)
 186 {
 187         return NULL;
 188 }
 189
 190 void unmap_hugepage_range(struct vm_area_struct *vma,
 191                           unsigned long start, unsigned long end)
 192 {
 193         struct mm_struct *mm = vma->vm_mm;
 194         unsigned long address;
 195         pte_t *pte;
 196         struct page *page;
 197         int i;
 198
 199         BUG_ON(start & (HPAGE_SIZE - 1));
 200         BUG_ON(end & (HPAGE_SIZE - 1));
 201
 202         for (address = start; address < end; address += HPAGE_SIZE) {
 203                 pte = huge_pte_offset(mm, address);
 204                 BUG_ON(!pte);
 205                 if (pte_none(*pte))
 206                         continue;
 207                 page = pte_page(*pte);
 208                 put_page(page);
 209                 for (i = 0; i < (1 << HUGETLB_PAGE_ORDER); i++) {
 210                         pte_clear(pte);
 211                         pte++;
 212                 }
 213         }
 214         vx_rsspages_sub(mm, (end - start) >> PAGE_SHIFT);
 215         flush_tlb_range(vma, start, end);
 216 }
 217
 218 int hugetlb_prefault(struct address_space *mapping, struct vm_area_struct *vma)
 219 {
 220         struct mm_struct *mm = current->mm;
 221         unsigned long addr;
 222         int ret = 0;
 223
 224         BUG_ON(vma->vm_start & ~HPAGE_MASK);
 225         BUG_ON(vma->vm_end & ~HPAGE_MASK);
 226
 227         spin_lock(&mm->page_table_lock);
 228         for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
 229                 unsigned long idx;
 230                 pte_t *pte = huge_pte_alloc(mm, addr);
 231                 struct page *page;
 232
 233                 if (!pte) {
 234                         ret = -ENOMEM;
 235                         goto out;
 236                 }
 237                 if (!pte_none(*pte))
 238                         continue;
 239
 240                 idx = ((addr - vma->vm_start) >> HPAGE_SHIFT)
 241                         + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
 242                 page = find_get_page(mapping, idx);
 243                 if (!page) {
 244                         /* charge the fs quota first */
 245                         if (hugetlb_get_quota(mapping)) {
 246                                 ret = -ENOMEM;
 247                                 goto out;
 248                         }
 249                         page = alloc_huge_page();
 250                         if (!page) {
 251                                 hugetlb_put_quota(mapping);
 252                                 ret = -ENOMEM;
 253                                 goto out;
 254                         }
 255                         ret = add_to_page_cache(page, mapping, idx, GFP_ATOMIC);
 256                         if (! ret) {
 257                                 unlock_page(page);
 258                         } else {
 259                                 hugetlb_put_quota(mapping);
 260                                 free_huge_page(page);
 261                                 goto out;
 262                         }
 263                 }
 264                 set_huge_pte(mm, vma, page, pte, vma->vm_flags & VM_WRITE);
 265         }
 266 out:
 267         spin_unlock(&mm->page_table_lock);
 268         return ret;
 269 }