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 #include <asm/mmu_context.h>
  24
  25 static pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr)
  26 {
  27         pgd_t *pgd;
  28         pud_t *pud;
  29         pmd_t *pmd;
  30         pte_t *pte = NULL;
  31
  32         pgd = pgd_offset(mm, addr);
  33         if (pgd) {
  34                 pud = pud_offset(pgd, addr);
  35                 if (pud) {
  36                         pmd = pmd_alloc(mm, pud, addr);
  37                         if (pmd)
  38                                 pte = pte_alloc_map(mm, pmd, addr);
  39                 }
  40         }
  41         return pte;
  42 }
  43
  44 static pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
  45 {
  46         pgd_t *pgd;
  47         pud_t *pud;
  48         pmd_t *pmd;
  49         pte_t *pte = NULL;
  50
  51         pgd = pgd_offset(mm, addr);
  52         if (pgd) {
  53                 pud = pud_offset(pgd, addr);
  54                 if (pud) {
  55                         pmd = pmd_offset(pud, addr);
  56                         if (pmd)
  57                                 pte = pte_offset_map(pmd, addr);
  58                 }
  59         }
  60         return pte;
  61 }
  62
  63 #define mk_pte_huge(entry) do { pte_val(entry) |= _PAGE_SZHUGE; } while (0)
  64
  65 static void set_huge_pte(struct mm_struct *mm, struct vm_area_struct *vma,
  66                          unsigned long addr,
  67                          struct page *page, pte_t * page_table, int write_access)
  68 {
  69         unsigned long i;
  70         pte_t entry;
  71
  72         add_mm_counter(mm, rss, HPAGE_SIZE / PAGE_SIZE);
  73
  74         if (write_access)
  75                 entry = pte_mkwrite(pte_mkdirty(mk_pte(page,
  76                                                        vma->vm_page_prot)));
  77         else
  78                 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
  79         entry = pte_mkyoung(entry);
  80         mk_pte_huge(entry);
  81
  82         for (i = 0; i < (1 << HUGETLB_PAGE_ORDER); i++) {
  83                 set_pte_at(mm, addr, page_table, entry);
  84                 page_table++;
  85                 addr += PAGE_SIZE;
  86
  87                 pte_val(entry) += PAGE_SIZE;
  88         }
  89 }
  90
  91 /*
  92  * This function checks for proper alignment of input addr and len parameters.
  93  */
  94 int is_aligned_hugepage_range(unsigned long addr, unsigned long len)
  95 {
  96         if (len & ~HPAGE_MASK)
  97                 return -EINVAL;
  98         if (addr & ~HPAGE_MASK)
  99                 return -EINVAL;
 100         return 0;
 101 }
 102
 103 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
 104                             struct vm_area_struct *vma)
 105 {
 106         pte_t *src_pte, *dst_pte, entry;
 107         struct page *ptepage;
 108         unsigned long addr = vma->vm_start;
 109         unsigned long end = vma->vm_end;
 110         int i;
 111
 112         while (addr < end) {
 113                 dst_pte = huge_pte_alloc(dst, addr);
 114                 if (!dst_pte)
 115                         goto nomem;
 116                 src_pte = huge_pte_offset(src, addr);
 117                 BUG_ON(!src_pte || pte_none(*src_pte));
 118                 entry = *src_pte;
 119                 ptepage = pte_page(entry);
 120                 get_page(ptepage);
 121                 for (i = 0; i < (1 << HUGETLB_PAGE_ORDER); i++) {
 122                         set_pte_at(dst, addr, dst_pte, entry);
 123                         pte_val(entry) += PAGE_SIZE;
 124                         dst_pte++;
 125                         addr += PAGE_SIZE;
 126                 }
 127                 add_mm_counter(dst, rss, HPAGE_SIZE / PAGE_SIZE);
 128         }
 129         return 0;
 130
 131 nomem:
 132         return -ENOMEM;
 133 }
 134
 135 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
 136                         struct page **pages, struct vm_area_struct **vmas,
 137                         unsigned long *position, int *length, int i)
 138 {
 139         unsigned long vaddr = *position;
 140         int remainder = *length;
 141
 142         WARN_ON(!is_vm_hugetlb_page(vma));
 143
 144         while (vaddr < vma->vm_end && remainder) {
 145                 if (pages) {
 146                         pte_t *pte;
 147                         struct page *page;
 148
 149                         pte = huge_pte_offset(mm, vaddr);
 150
 151                         /* hugetlb should be locked, and hence, prefaulted */
 152                         BUG_ON(!pte || pte_none(*pte));
 153
 154                         page = pte_page(*pte);
 155
 156                         WARN_ON(!PageCompound(page));
 157
 158                         get_page(page);
 159                         pages[i] = page;
 160                 }
 161
 162                 if (vmas)
 163                         vmas[i] = vma;
 164
 165                 vaddr += PAGE_SIZE;
 166                 --remainder;
 167                 ++i;
 168         }
 169
 170         *length = remainder;
 171         *position = vaddr;
 172
 173         return i;
 174 }
 175
 176 struct page *follow_huge_addr(struct mm_struct *mm,
 177                               unsigned long address, int write)
 178 {
 179         return ERR_PTR(-EINVAL);
 180 }
 181
 182 int pmd_huge(pmd_t pmd)
 183 {
 184         return 0;
 185 }
 186
 187 struct page *follow_huge_pmd(struct mm_struct *mm, unsigned long address,
 188                              pmd_t *pmd, int write)
 189 {
 190         return NULL;
 191 }
 192
 193 void unmap_hugepage_range(struct vm_area_struct *vma,
 194                           unsigned long start, unsigned long end)
 195 {
 196         struct mm_struct *mm = vma->vm_mm;
 197         unsigned long address;
 198         pte_t *pte;
 199         struct page *page;
 200         int i;
 201
 202         BUG_ON(start & (HPAGE_SIZE - 1));
 203         BUG_ON(end & (HPAGE_SIZE - 1));
 204
 205         for (address = start; address < end; address += HPAGE_SIZE) {
 206                 pte = huge_pte_offset(mm, address);
 207                 BUG_ON(!pte);
 208                 if (pte_none(*pte))
 209                         continue;
 210                 page = pte_page(*pte);
 211                 put_page(page);
 212                 for (i = 0; i < (1 << HUGETLB_PAGE_ORDER); i++) {
 213                         pte_clear(mm, address+(i*PAGE_SIZE), pte);
 214                         pte++;
 215                 }
 216         }
 217         add_mm_counter(mm, rss, -((end - start) >> PAGE_SHIFT));
 218         flush_tlb_range(vma, start, end);
 219 }
 220
 221 static void context_reload(void *__data)
 222 {
 223         struct mm_struct *mm = __data;
 224
 225         if (mm == current->mm)
 226                 load_secondary_context(mm);
 227 }
 228
 229 int hugetlb_prefault(struct address_space *mapping, struct vm_area_struct *vma)
 230 {
 231         struct mm_struct *mm = current->mm;
 232         unsigned long addr;
 233         int ret = 0;
 234
 235         /* On UltraSPARC-III+ and later, configure the second half of
 236          * the Data-TLB for huge pages.
 237          */
 238         if (tlb_type == cheetah_plus) {
 239                 unsigned long ctx;
 240
 241                 spin_lock(&ctx_alloc_lock);
 242                 ctx = mm->context.sparc64_ctx_val;
 243                 ctx &= ~CTX_PGSZ_MASK;
 244                 ctx |= CTX_PGSZ_BASE << CTX_PGSZ0_SHIFT;
 245                 ctx |= CTX_PGSZ_HUGE << CTX_PGSZ1_SHIFT;
 246
 247                 if (ctx != mm->context.sparc64_ctx_val) {
 248                         /* When changing the page size fields, we
 249                          * must perform a context flush so that no
 250                          * stale entries match.  This flush must
 251                          * occur with the original context register
 252                          * settings.
 253                          */
 254                         do_flush_tlb_mm(mm);
 255
 256                         /* Reload the context register of all processors
 257                          * also executing in this address space.
 258                          */
 259                         mm->context.sparc64_ctx_val = ctx;
 260                         on_each_cpu(context_reload, mm, 0, 0);
 261                 }
 262                 spin_unlock(&ctx_alloc_lock);
 263         }
 264
 265         BUG_ON(vma->vm_start & ~HPAGE_MASK);
 266         BUG_ON(vma->vm_end & ~HPAGE_MASK);
 267
 268         spin_lock(&mm->page_table_lock);
 269         for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
 270                 unsigned long idx;
 271                 pte_t *pte = huge_pte_alloc(mm, addr);
 272                 struct page *page;
 273
 274                 if (!pte) {
 275                         ret = -ENOMEM;
 276                         goto out;
 277                 }
 278                 if (!pte_none(*pte))
 279                         continue;
 280
 281                 idx = ((addr - vma->vm_start) >> HPAGE_SHIFT)
 282                         + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
 283                 page = find_get_page(mapping, idx);
 284                 if (!page) {
 285                         /* charge the fs quota first */
 286                         if (hugetlb_get_quota(mapping)) {
 287                                 ret = -ENOMEM;
 288                                 goto out;
 289                         }
 290                         page = alloc_huge_page();
 291                         if (!page) {
 292                                 hugetlb_put_quota(mapping);
 293                                 ret = -ENOMEM;
 294                                 goto out;
 295                         }
 296                         ret = add_to_page_cache(page, mapping, idx, GFP_ATOMIC);
 297                         if (! ret) {
 298                                 unlock_page(page);
 299                         } else {
 300                                 hugetlb_put_quota(mapping);
 301                                 free_huge_page(page);
 302                                 goto out;
 303                         }
 304                 }
 305                 set_huge_pte(mm, vma, addr, page, pte, vma->vm_flags & VM_WRITE);
 306         }
 307 out:
 308         spin_unlock(&mm->page_table_lock);
 309         return ret;
 310 }