arch/ia64/mm/hugetlbpage.c

   1 /*
   2  * IA-64 Huge TLB Page Support for Kernel.
   3  *
   4  * Copyright (C) 2002-2004 Rohit Seth <rohit.seth@intel.com>
   5  * Copyright (C) 2003-2004 Ken Chen <kenneth.w.chen@intel.com>
   6  *
   7  * Sep, 2003: add numa support
   8  * Feb, 2004: dynamic hugetlb page size via boot parameter
   9  */
  10
  11 #include <linux/init.h>
  12 #include <linux/fs.h>
  13 #include <linux/mm.h>
  14 #include <linux/hugetlb.h>
  15 #include <linux/pagemap.h>
  16 #include <linux/module.h>
  17 #include <linux/slab.h>
  18 #include <linux/sysctl.h>
  19 #include <linux/log2.h>
  20 #include <asm/mman.h>
  21 #include <asm/pgalloc.h>
  22 #include <asm/tlb.h>
  23 #include <asm/tlbflush.h>
  24
  25 unsigned int hpage_shift = HPAGE_SHIFT_DEFAULT;
  26 EXPORT_SYMBOL(hpage_shift);
  27
  28 pte_t *
  29 huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz)
  30 {
  31         unsigned long taddr = htlbpage_to_page(addr);
  32         pgd_t *pgd;
  33         pud_t *pud;
  34         pmd_t *pmd;
  35         pte_t *pte = NULL;
  36
  37         pgd = pgd_offset(mm, taddr);
  38         pud = pud_alloc(mm, pgd, taddr);
  39         if (pud) {
  40                 pmd = pmd_alloc(mm, pud, taddr);
  41                 if (pmd)
  42                         pte = pte_alloc_map(mm, pmd, taddr);
  43         }
  44         return pte;
  45 }
  46
  47 pte_t *
  48 huge_pte_offset (struct mm_struct *mm, unsigned long addr)
  49 {
  50         unsigned long taddr = htlbpage_to_page(addr);
  51         pgd_t *pgd;
  52         pud_t *pud;
  53         pmd_t *pmd;
  54         pte_t *pte = NULL;
  55
  56         pgd = pgd_offset(mm, taddr);
  57         if (pgd_present(*pgd)) {
  58                 pud = pud_offset(pgd, taddr);
  59                 if (pud_present(*pud)) {
  60                         pmd = pmd_offset(pud, taddr);
  61                         if (pmd_present(*pmd))
  62                                 pte = pte_offset_map(pmd, taddr);
  63                 }
  64         }
  65
  66         return pte;
  67 }
  68
  69 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
  70 {
  71         return 0;
  72 }
  73
  74 #define mk_pte_huge(entry) { pte_val(entry) |= _PAGE_P; }
  75
  76 /*
  77  * Don't actually need to do any preparation, but need to make sure
  78  * the address is in the right region.
  79  */
  80 int prepare_hugepage_range(struct file *file,
  81                         unsigned long addr, unsigned long len)
  82 {
  83         if (len & ~HPAGE_MASK)
  84                 return -EINVAL;
  85         if (addr & ~HPAGE_MASK)
  86                 return -EINVAL;
  87         if (REGION_NUMBER(addr) != RGN_HPAGE)
  88                 return -EINVAL;
  89
  90         return 0;
  91 }
  92
  93 struct page *follow_huge_addr(struct mm_struct *mm, unsigned long addr, int write)
  94 {
  95         struct page *page;
  96         pte_t *ptep;
  97
  98         if (REGION_NUMBER(addr) != RGN_HPAGE)
  99                 return ERR_PTR(-EINVAL);
 100
 101         ptep = huge_pte_offset(mm, addr);
 102         if (!ptep || pte_none(*ptep))
 103                 return NULL;
 104         page = pte_page(*ptep);
 105         page += ((addr & ~HPAGE_MASK) >> PAGE_SHIFT);
 106         return page;
 107 }
 108 int pmd_huge(pmd_t pmd)
 109 {
 110         return 0;
 111 }
 112
 113 int pud_huge(pud_t pud)
 114 {
 115         return 0;
 116 }
 117
 118 struct page *
 119 follow_huge_pmd(struct mm_struct *mm, unsigned long address, pmd_t *pmd, int write)
 120 {
 121         return NULL;
 122 }
 123
 124 void hugetlb_free_pgd_range(struct mmu_gather *tlb,
 125                         unsigned long addr, unsigned long end,
 126                         unsigned long floor, unsigned long ceiling)
 127 {
 128         /*
 129          * This is called to free hugetlb page tables.
 130          *
 131          * The offset of these addresses from the base of the hugetlb
 132          * region must be scaled down by HPAGE_SIZE/PAGE_SIZE so that
 133          * the standard free_pgd_range will free the right page tables.
 134          *
 135          * If floor and ceiling are also in the hugetlb region, they
 136          * must likewise be scaled down; but if outside, left unchanged.
 137          */
 138
 139         addr = htlbpage_to_page(addr);
 140         end  = htlbpage_to_page(end);
 141         if (REGION_NUMBER(floor) == RGN_HPAGE)
 142                 floor = htlbpage_to_page(floor);
 143         if (REGION_NUMBER(ceiling) == RGN_HPAGE)
 144                 ceiling = htlbpage_to_page(ceiling);
 145
 146         free_pgd_range(tlb, addr, end, floor, ceiling);
 147 }
 148
 149 unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
 150                 unsigned long pgoff, unsigned long flags)
 151 {
 152         struct vm_area_struct *vmm;
 153
 154         if (len > RGN_MAP_LIMIT)
 155                 return -ENOMEM;
 156         if (len & ~HPAGE_MASK)
 157                 return -EINVAL;
 158
 159         /* Handle MAP_FIXED */
 160         if (flags & MAP_FIXED) {
 161                 if (prepare_hugepage_range(file, addr, len))
 162                         return -EINVAL;
 163                 return addr;
 164         }
 165
 166         /* This code assumes that RGN_HPAGE != 0. */
 167         if ((REGION_NUMBER(addr) != RGN_HPAGE) || (addr & (HPAGE_SIZE - 1)))
 168                 addr = HPAGE_REGION_BASE;
 169         else
 170                 addr = ALIGN(addr, HPAGE_SIZE);
 171         for (vmm = find_vma(current->mm, addr); ; vmm = vmm->vm_next) {
 172                 /* At this point:  (!vmm || addr < vmm->vm_end). */
 173                 if (REGION_OFFSET(addr) + len > RGN_MAP_LIMIT)
 174                         return -ENOMEM;
 175                 if (!vmm || (addr + len) <= vmm->vm_start)
 176                         return addr;
 177                 addr = ALIGN(vmm->vm_end, HPAGE_SIZE);
 178         }
 179 }
 180
 181 static int __init hugetlb_setup_sz(char *str)
 182 {
 183         u64 tr_pages;
 184         unsigned long long size;
 185
 186         if (ia64_pal_vm_page_size(&tr_pages, NULL) != 0)
 187                 /*
 188                  * shouldn't happen, but just in case.
 189                  */
 190                 tr_pages = 0x15557000UL;
 191
 192         size = memparse(str, &str);
 193         if (*str || !is_power_of_2(size) || !(tr_pages & size) ||
 194                 size <= PAGE_SIZE ||
 195                 size >= (1UL << PAGE_SHIFT << MAX_ORDER)) {
 196                 printk(KERN_WARNING "Invalid huge page size specified\n");
 197                 return 1;
 198         }
 199
 200         hpage_shift = __ffs(size);
 201         /*
 202          * boot cpu already executed ia64_mmu_init, and has HPAGE_SHIFT_DEFAULT
 203          * override here with new page shift.
 204          */
 205         ia64_set_rr(HPAGE_REGION_BASE, hpage_shift << 2);
 206         return 0;
 207 }
 208 early_param("hugepagesz", hugetlb_setup_sz);