arch/x86/mm/hugetlbpage.c

   1 /*
   2  * IA-32 Huge TLB Page Support for Kernel.
   3  *
   4  * Copyright (C) 2002, Rohit Seth <rohit.seth@intel.com>
   5  */
   6
   7 #include <linux/init.h>
   8 #include <linux/fs.h>
   9 #include <linux/mm.h>
  10 #include <linux/sched/mm.h>
  11 #include <linux/hugetlb.h>
  12 #include <linux/pagemap.h>
  13 #include <linux/err.h>
  14 #include <linux/sysctl.h>
  15 #include <linux/compat.h>
  16 #include <asm/mman.h>
  17 #include <asm/tlb.h>
  18 #include <asm/tlbflush.h>
  19 #include <asm/pgalloc.h>
  20 #include <asm/elf.h>
  21 #include <asm/mpx.h>
  22
  23 #if 0   /* This is just for testing */
  24 struct page *
  25 follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
  26 {
  27         unsigned long start = address;
  28         int length = 1;
  29         int nr;
  30         struct page *page;
  31         struct vm_area_struct *vma;
  32
  33         vma = find_vma(mm, addr);
  34         if (!vma || !is_vm_hugetlb_page(vma))
  35                 return ERR_PTR(-EINVAL);
  36
  37         pte = huge_pte_offset(mm, address, vma_mmu_pagesize(vma));
  38
  39         /* hugetlb should be locked, and hence, prefaulted */
  40         WARN_ON(!pte || pte_none(*pte));
  41
  42         page = &pte_page(*pte)[vpfn % (HPAGE_SIZE/PAGE_SIZE)];
  43
  44         WARN_ON(!PageHead(page));
  45
  46         return page;
  47 }
  48
  49 int pmd_huge(pmd_t pmd)
  50 {
  51         return 0;
  52 }
  53
  54 int pud_huge(pud_t pud)
  55 {
  56         return 0;
  57 }
  58
  59 #else
  60
  61 /*
  62  * pmd_huge() returns 1 if @pmd is hugetlb related entry, that is normal
  63  * hugetlb entry or non-present (migration or hwpoisoned) hugetlb entry.
  64  * Otherwise, returns 0.
  65  */
  66 int pmd_huge(pmd_t pmd)
  67 {
  68         return !pmd_none(pmd) &&
  69                 (pmd_val(pmd) & (_PAGE_PRESENT|_PAGE_PSE)) != _PAGE_PRESENT;
  70 }
  71
  72 int pud_huge(pud_t pud)
  73 {
  74         return !!(pud_val(pud) & _PAGE_PSE);
  75 }
  76 #endif
  77
  78 #ifdef CONFIG_HUGETLB_PAGE
  79 static unsigned long hugetlb_get_unmapped_area_bottomup(struct file *file,
  80                 unsigned long addr, unsigned long len,
  81                 unsigned long pgoff, unsigned long flags)
  82 {
  83         struct hstate *h = hstate_file(file);
  84         struct vm_unmapped_area_info info;
  85
  86         info.flags = 0;
  87         info.length = len;
  88         info.low_limit = get_mmap_base(1);
  89
  90         /*
  91          * If hint address is above DEFAULT_MAP_WINDOW, look for unmapped area
  92          * in the full address space.
  93          */
  94         info.high_limit = in_compat_syscall() ?
  95                 task_size_32bit() : task_size_64bit(addr > DEFAULT_MAP_WINDOW);
  96
  97         info.align_mask = PAGE_MASK & ~huge_page_mask(h);
  98         info.align_offset = 0;
  99         return vm_unmapped_area(&info);
 100 }
 101
 102 static unsigned long hugetlb_get_unmapped_area_topdown(struct file *file,
 103                 unsigned long addr, unsigned long len,
 104                 unsigned long pgoff, unsigned long flags)
 105 {
 106         struct hstate *h = hstate_file(file);
 107         struct vm_unmapped_area_info info;
 108
 109         info.flags = VM_UNMAPPED_AREA_TOPDOWN;
 110         info.length = len;
 111         info.low_limit = PAGE_SIZE;
 112         info.high_limit = get_mmap_base(0);
 113
 114         /*
 115          * If hint address is above DEFAULT_MAP_WINDOW, look for unmapped area
 116          * in the full address space.
 117          */
 118         if (addr > DEFAULT_MAP_WINDOW && !in_compat_syscall())
 119                 info.high_limit += TASK_SIZE_MAX - DEFAULT_MAP_WINDOW;
 120
 121         info.align_mask = PAGE_MASK & ~huge_page_mask(h);
 122         info.align_offset = 0;
 123         addr = vm_unmapped_area(&info);
 124
 125         /*
 126          * A failed mmap() very likely causes application failure,
 127          * so fall back to the bottom-up function here. This scenario
 128          * can happen with large stack limits and large mmap()
 129          * allocations.
 130          */
 131         if (addr & ~PAGE_MASK) {
 132                 VM_BUG_ON(addr != -ENOMEM);
 133                 info.flags = 0;
 134                 info.low_limit = TASK_UNMAPPED_BASE;
 135                 info.high_limit = TASK_SIZE_LOW;
 136                 addr = vm_unmapped_area(&info);
 137         }
 138
 139         return addr;
 140 }
 141
 142 unsigned long
 143 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 144                 unsigned long len, unsigned long pgoff, unsigned long flags)
 145 {
 146         struct hstate *h = hstate_file(file);
 147         struct mm_struct *mm = current->mm;
 148         struct vm_area_struct *vma;
 149
 150         if (len & ~huge_page_mask(h))
 151                 return -EINVAL;
 152
 153         addr = mpx_unmapped_area_check(addr, len, flags);
 154         if (IS_ERR_VALUE(addr))
 155                 return addr;
 156
 157         if (len > TASK_SIZE)
 158                 return -ENOMEM;
 159
 160         if (flags & MAP_FIXED) {
 161                 if (prepare_hugepage_range(file, addr, len))
 162                         return -EINVAL;
 163                 return addr;
 164         }
 165
 166         if (addr) {
 167                 addr = ALIGN(addr, huge_page_size(h));
 168                 vma = find_vma(mm, addr);
 169                 if (TASK_SIZE - len >= addr &&
 170                     (!vma || addr + len <= vm_start_gap(vma)))
 171                         return addr;
 172         }
 173         if (mm->get_unmapped_area == arch_get_unmapped_area)
 174                 return hugetlb_get_unmapped_area_bottomup(file, addr, len,
 175                                 pgoff, flags);
 176         else
 177                 return hugetlb_get_unmapped_area_topdown(file, addr, len,
 178                                 pgoff, flags);
 179 }
 180 #endif /* CONFIG_HUGETLB_PAGE */
 181
 182 #ifdef CONFIG_X86_64
 183 static __init int setup_hugepagesz(char *opt)
 184 {
 185         unsigned long ps = memparse(opt, &opt);
 186         if (ps == PMD_SIZE) {
 187                 hugetlb_add_hstate(PMD_SHIFT - PAGE_SHIFT);
 188         } else if (ps == PUD_SIZE && boot_cpu_has(X86_FEATURE_GBPAGES)) {
 189                 hugetlb_add_hstate(PUD_SHIFT - PAGE_SHIFT);
 190         } else {
 191                 hugetlb_bad_size();
 192                 printk(KERN_ERR "hugepagesz: Unsupported page size %lu M\n",
 193                         ps >> 20);
 194                 return 0;
 195         }
 196         return 1;
 197 }
 198 __setup("hugepagesz=", setup_hugepagesz);
 199
 200 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
 201 static __init int gigantic_pages_init(void)
 202 {
 203         /* With compaction or CMA we can allocate gigantic pages at runtime */
 204         if (boot_cpu_has(X86_FEATURE_GBPAGES) && !size_to_hstate(1UL << PUD_SHIFT))
 205                 hugetlb_add_hstate(PUD_SHIFT - PAGE_SHIFT);
 206         return 0;
 207 }
 208 arch_initcall(gigantic_pages_init);
 209 #endif
 210 #endif