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