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