arch/x86/mm/pageattr_64.c

   1 /*
   2  * Copyright 2002 Andi Kleen, SuSE Labs.
   3  * Thanks to Ben LaHaise for precious feedback.
   4  */
   5
   6 #include <linux/highmem.h>
   7 #include <linux/module.h>
   8 #include <linux/sched.h>
   9 #include <linux/slab.h>
  10 #include <linux/mm.h>
  11
  12 void clflush_cache_range(void *addr, int size)
  13 {
  14         int i;
  15
  16         for (i = 0; i < size; i += boot_cpu_data.x86_clflush_size)
  17                 clflush(addr+i);
  18 }
  19
  20 #include <asm/processor.h>
  21 #include <asm/tlbflush.h>
  22 #include <asm/sections.h>
  23 #include <asm/uaccess.h>
  24 #include <asm/pgalloc.h>
  25
  26 pte_t *lookup_address(unsigned long address, int *level)
  27 {
  28         pgd_t *pgd = pgd_offset_k(address);
  29         pud_t *pud;
  30         pmd_t *pmd;
  31
  32         if (pgd_none(*pgd))
  33                 return NULL;
  34         pud = pud_offset(pgd, address);
  35         if (pud_none(*pud))
  36                 return NULL;
  37         pmd = pmd_offset(pud, address);
  38         if (pmd_none(*pmd))
  39                 return NULL;
  40         *level = 3;
  41         if (pmd_large(*pmd))
  42                 return (pte_t *)pmd;
  43         *level = 4;
  44
  45         return pte_offset_kernel(pmd, address);
  46 }
  47
  48 static struct page *
  49 split_large_page(unsigned long address, pgprot_t ref_prot)
  50 {
  51         unsigned long addr;
  52         struct page *base;
  53         pte_t *pbase;
  54         int i;
  55
  56         base = alloc_pages(GFP_KERNEL, 0);
  57         if (!base)
  58                 return NULL;
  59
  60         address = __pa(address);
  61         addr = address & LARGE_PAGE_MASK;
  62         pbase = (pte_t *)page_address(base);
  63         for (i = 0; i < PTRS_PER_PTE; i++, addr += PAGE_SIZE)
  64                 pbase[i] = pfn_pte(addr >> PAGE_SHIFT, ref_prot);
  65
  66         return base;
  67 }
  68
  69 static int
  70 __change_page_attr(unsigned long address, struct page *page, pgprot_t prot)
  71 {
  72         struct page *kpte_page;
  73         pte_t *kpte;
  74         pgprot_t ref_prot2, oldprot;
  75         int level;
  76
  77 repeat:
  78         kpte = lookup_address(address, &level);
  79         if (!kpte)
  80                 return 0;
  81
  82         kpte_page = virt_to_page(kpte);
  83         oldprot = pte_pgprot(*kpte);
  84         BUG_ON(PageLRU(kpte_page));
  85         BUG_ON(PageCompound(kpte_page));
  86         prot = canon_pgprot(prot);
  87
  88         if (level == 4) {
  89                 set_pte_atomic(kpte, mk_pte(page, prot));
  90         } else {
  91                 /*
  92                  * split_large_page will take the reference for this
  93                  * change_page_attr on the split page.
  94                  */
  95                 struct page *split;
  96
  97                 ref_prot2 = pte_pgprot(pte_clrhuge(*kpte));
  98                 split = split_large_page(address, ref_prot2);
  99                 if (!split)
 100                         return -ENOMEM;
 101                 pgprot_val(ref_prot2) &= ~_PAGE_NX;
 102                 set_pte_atomic(kpte, mk_pte(split, ref_prot2));
 103                 goto repeat;
 104         }
 105
 106         return 0;
 107 }
 108
 109 /**
 110  * change_page_attr_addr - Change page table attributes in linear mapping
 111  * @address: Virtual address in linear mapping.
 112  * @numpages: Number of pages to change
 113  * @prot:    New page table attribute (PAGE_*)
 114  *
 115  * Change page attributes of a page in the direct mapping. This is a variant
 116  * of change_page_attr() that also works on memory holes that do not have
 117  * mem_map entry (pfn_valid() is false).
 118  *
 119  * See change_page_attr() documentation for more details.
 120  */
 121
 122 int change_page_attr_addr(unsigned long address, int numpages, pgprot_t prot)
 123 {
 124         int err = 0, kernel_map = 0, i;
 125
 126         if (address >= __START_KERNEL_map &&
 127                         address < __START_KERNEL_map + KERNEL_TEXT_SIZE) {
 128
 129                 address = (unsigned long)__va(__pa(address));
 130                 kernel_map = 1;
 131         }
 132
 133         down_write(&init_mm.mmap_sem);
 134         for (i = 0; i < numpages; i++, address += PAGE_SIZE) {
 135                 unsigned long pfn = __pa(address) >> PAGE_SHIFT;
 136
 137                 if (!kernel_map || pte_present(pfn_pte(0, prot))) {
 138                         err = __change_page_attr(address, pfn_to_page(pfn), prot);
 139                         if (err)
 140                                 break;
 141                 }
 142                 /* Handle kernel mapping too which aliases part of the
 143                  * lowmem */
 144                 if (__pa(address) < KERNEL_TEXT_SIZE) {
 145                         unsigned long addr2;
 146                         pgprot_t prot2;
 147
 148                         addr2 = __START_KERNEL_map + __pa(address);
 149                         /* Make sure the kernel mappings stay executable */
 150                         prot2 = pte_pgprot(pte_mkexec(pfn_pte(0, prot)));
 151                         err = __change_page_attr(addr2, pfn_to_page(pfn), prot2);
 152                 }
 153         }
 154         up_write(&init_mm.mmap_sem);
 155
 156         return err;
 157 }
 158
 159 /**
 160  * change_page_attr - Change page table attributes in the linear mapping.
 161  * @page: First page to change
 162  * @numpages: Number of pages to change
 163  * @prot: New protection/caching type (PAGE_*)
 164  *
 165  * Returns 0 on success, otherwise a negated errno.
 166  *
 167  * This should be used when a page is mapped with a different caching policy
 168  * than write-back somewhere - some CPUs do not like it when mappings with
 169  * different caching policies exist. This changes the page attributes of the
 170  * in kernel linear mapping too.
 171  *
 172  * Caller must call global_flush_tlb() later to make the changes active.
 173  *
 174  * The caller needs to ensure that there are no conflicting mappings elsewhere
 175  * (e.g. in user space) * This function only deals with the kernel linear map.
 176  *
 177  * For MMIO areas without mem_map use change_page_attr_addr() instead.
 178  */
 179 int change_page_attr(struct page *page, int numpages, pgprot_t prot)
 180 {
 181         unsigned long addr = (unsigned long)page_address(page);
 182
 183         return change_page_attr_addr(addr, numpages, prot);
 184 }
 185 EXPORT_SYMBOL(change_page_attr);
 186
 187 static void flush_kernel_map(void *arg)
 188 {
 189         /*
 190          * Flush all to work around Errata in early athlons regarding
 191          * large page flushing.
 192          */
 193         __flush_tlb_all();
 194
 195         if (boot_cpu_data.x86_model >= 4)
 196                 wbinvd();
 197 }
 198
 199 void global_flush_tlb(void)
 200 {
 201         BUG_ON(irqs_disabled());
 202
 203         on_each_cpu(flush_kernel_map, NULL, 1, 1);
 204 }
 205 EXPORT_SYMBOL(global_flush_tlb);