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         // mm->rss += (HPAGE_SIZE / PAGE_SIZE);
  71         vx_rsspages_add(mm, HPAGE_SIZE / PAGE_SIZE);
  72
  73         if (write_access)
  74                 entry = pte_mkwrite(pte_mkdirty(mk_pte(page,
  75                                                        vma->vm_page_prot)));
  76         else
  77                 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
  78         entry = pte_mkyoung(entry);
  79         mk_pte_huge(entry);
  80
  81         for (i = 0; i < (1 << HUGETLB_PAGE_ORDER); i++) {
  82                 set_pte(page_table, entry);
  83                 page_table++;
  84
  85                 pte_val(entry) += PAGE_SIZE;
  86         }
  87 }
  88
  89 /*
  90  * This function checks for proper alignment of input addr and len parameters.
  91  */
  92 int is_aligned_hugepage_range(unsigned long addr, unsigned long len)
  93 {
  94         if (len & ~HPAGE_MASK)
  95                 return -EINVAL;
  96         if (addr & ~HPAGE_MASK)
  97                 return -EINVAL;
  98         return 0;
  99 }
 100
 101 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
 102                             struct vm_area_struct *vma)
 103 {
 104         pte_t *src_pte, *dst_pte, entry;
 105         struct page *ptepage;
 106         unsigned long addr = vma->vm_start;
 107         unsigned long end = vma->vm_end;
 108         int i;
 109
 110         while (addr < end) {
 111                 dst_pte = huge_pte_alloc(dst, addr);
 112                 if (!dst_pte)
 113                         goto nomem;
 114                 src_pte = huge_pte_offset(src, addr);
 115                 BUG_ON(!src_pte || pte_none(*src_pte));
 116                 entry = *src_pte;
 117                 ptepage = pte_page(entry);
 118                 get_page(ptepage);
 119                 for (i = 0; i < (1 << HUGETLB_PAGE_ORDER); i++) {
 120                         set_pte(dst_pte, entry);
 121                         pte_val(entry) += PAGE_SIZE;
 122                         dst_pte++;
 123                 }
 124                 // dst->rss += (HPAGE_SIZE / PAGE_SIZE);
 125                 vx_rsspages_add(dst, HPAGE_SIZE / PAGE_SIZE);
 126                 addr += HPAGE_SIZE;
 127         }
 128         return 0;
 129
 130 nomem:
 131         return -ENOMEM;
 132 }
 133
 134 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
 135                         struct page **pages, struct vm_area_struct **vmas,
 136                         unsigned long *position, int *length, int i)
 137 {
 138         unsigned long vaddr = *position;
 139         int remainder = *length;
 140
 141         WARN_ON(!is_vm_hugetlb_page(vma));
 142
 143         while (vaddr < vma->vm_end && remainder) {
 144                 if (pages) {
 145                         pte_t *pte;
 146                         struct page *page;
 147
 148                         pte = huge_pte_offset(mm, vaddr);
 149
 150                         /* hugetlb should be locked, and hence, prefaulted */
 151                         BUG_ON(!pte || pte_none(*pte));
 152
 153                         page = pte_page(*pte);
 154
 155                         WARN_ON(!PageCompound(page));
 156
 157                         get_page(page);
 158                         pages[i] = page;
 159                 }
 160
 161                 if (vmas)
 162                         vmas[i] = vma;
 163
 164                 vaddr += PAGE_SIZE;
 165                 --remainder;
 166                 ++i;
 167         }
 168
 169         *length = remainder;
 170         *position = vaddr;
 171
 172         return i;
 173 }
 174
 175 struct page *follow_huge_addr(struct mm_struct *mm,
 176                               unsigned long address, int write)
 177 {
 178         return ERR_PTR(-EINVAL);
 179 }
 180
 181 int pmd_huge(pmd_t pmd)
 182 {
 183         return 0;
 184 }
 185
 186 struct page *follow_huge_pmd(struct mm_struct *mm, unsigned long address,
 187                              pmd_t *pmd, int write)
 188 {
 189         return NULL;
 190 }
 191
 192 void unmap_hugepage_range(struct vm_area_struct *vma,
 193                           unsigned long start, unsigned long end)
 194 {
 195         struct mm_struct *mm = vma->vm_mm;
 196         unsigned long address;
 197         pte_t *pte;
 198         struct page *page;
 199         int i;
 200
 201         BUG_ON(start & (HPAGE_SIZE - 1));
 202         BUG_ON(end & (HPAGE_SIZE - 1));
 203
 204         for (address = start; address < end; address += HPAGE_SIZE) {
 205                 pte = huge_pte_offset(mm, address);
 206                 BUG_ON(!pte);
 207                 if (pte_none(*pte))
 208                         continue;
 209                 page = pte_page(*pte);
 210                 put_page(page);
 211                 for (i = 0; i < (1 << HUGETLB_PAGE_ORDER); i++) {
 212                         pte_clear(pte);
 213                         pte++;
 214                 }
 215         }
 216         // mm->rss -= (end - start) >> PAGE_SHIFT;
 217         vx_rsspages_sub(mm, (end - start) >> PAGE_SHIFT);
 218         flush_tlb_range(vma, start, end);
 219 }
 220
 221 int hugetlb_prefault(struct address_space *mapping, struct vm_area_struct *vma)
 222 {
 223         struct mm_struct *mm = current->mm;
 224         unsigned long addr;
 225         int ret = 0;
 226
 227         BUG_ON(vma->vm_start & ~HPAGE_MASK);
 228         BUG_ON(vma->vm_end & ~HPAGE_MASK);
 229
 230         spin_lock(&mm->page_table_lock);
 231         for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
 232                 unsigned long idx;
 233                 pte_t *pte = huge_pte_alloc(mm, addr);
 234                 struct page *page;
 235
 236                 if (!pte) {
 237                         ret = -ENOMEM;
 238                         goto out;
 239                 }
 240                 if (!pte_none(*pte))
 241                         continue;
 242
 243                 idx = ((addr - vma->vm_start) >> HPAGE_SHIFT)
 244                         + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
 245                 page = find_get_page(mapping, idx);
 246                 if (!page) {
 247                         /* charge the fs quota first */
 248                         if (hugetlb_get_quota(mapping)) {
 249                                 ret = -ENOMEM;
 250                                 goto out;
 251                         }
 252                         page = alloc_huge_page();
 253                         if (!page) {
 254                                 hugetlb_put_quota(mapping);
 255                                 ret = -ENOMEM;
 256                                 goto out;
 257                         }
 258                         ret = add_to_page_cache(page, mapping, idx, GFP_ATOMIC);
 259                         if (! ret) {
 260                                 unlock_page(page);
 261                         } else {
 262                                 hugetlb_put_quota(mapping);
 263                                 free_huge_page(page);
 264                                 goto out;
 265                         }
 266                 }
 267                 set_huge_pte(mm, vma, page, pte, vma->vm_flags & VM_WRITE);
 268         }
 269 out:
 270         spin_unlock(&mm->page_table_lock);
 271         return ret;
 272 }