arch/x86_64/mm/pageattr.c

   1 /*
   2  * Copyright 2002 Andi Kleen, SuSE Labs.
   3  * Thanks to Ben LaHaise for precious feedback.
   4  */
   5
   6 #include <linux/config.h>
   7 #include <linux/mm.h>
   8 #include <linux/sched.h>
   9 #include <linux/highmem.h>
  10 #include <linux/module.h>
  11 #include <linux/slab.h>
  12 #include <asm/uaccess.h>
  13 #include <asm/processor.h>
  14 #include <asm/tlbflush.h>
  15 #include <asm/io.h>
  16
  17 static inline pte_t *lookup_address(unsigned long address)
  18 {
  19         pgd_t *pgd = pgd_offset_k(address);
  20         pud_t *pud;
  21         pmd_t *pmd;
  22         pte_t *pte;
  23         if (pgd_none(*pgd))
  24                 return NULL;
  25         pud = pud_offset(pgd, address);
  26         if (!pud_present(*pud))
  27                 return NULL;
  28         pmd = pmd_offset(pud, address);
  29         if (!pmd_present(*pmd))
  30                 return NULL;
  31         if (pmd_large(*pmd))
  32                 return (pte_t *)pmd;
  33         pte = pte_offset_kernel(pmd, address);
  34         if (pte && !pte_present(*pte))
  35                 pte = NULL;
  36         return pte;
  37 }
  38
  39 static struct page *split_large_page(unsigned long address, pgprot_t prot,
  40                                      pgprot_t ref_prot)
  41 {
  42         int i;
  43         unsigned long addr;
  44         struct page *base = alloc_pages(GFP_KERNEL, 0);
  45         pte_t *pbase;
  46         if (!base)
  47                 return NULL;
  48         /*
  49          * page_private is used to track the number of entries in
  50          * the page table page have non standard attributes.
  51          */
  52         SetPagePrivate(base);
  53         page_private(base) = 0;
  54
  55         address = __pa(address);
  56         addr = address & LARGE_PAGE_MASK;
  57         pbase = (pte_t *)page_address(base);
  58         for (i = 0; i < PTRS_PER_PTE; i++, addr += PAGE_SIZE) {
  59                 pbase[i] = pfn_pte(addr >> PAGE_SHIFT,
  60                                    addr == address ? prot : ref_prot);
  61         }
  62         return base;
  63 }
  64
  65
  66 static void flush_kernel_map(void *address)
  67 {
  68         if (0 && address && cpu_has_clflush) {
  69                 /* is this worth it? */
  70                 int i;
  71                 for (i = 0; i < PAGE_SIZE; i += boot_cpu_data.x86_clflush_size)
  72                         asm volatile("clflush (%0)" :: "r" (address + i));
  73         } else
  74                 asm volatile("wbinvd":::"memory");
  75         if (address)
  76                 __flush_tlb_one(address);
  77         else
  78                 __flush_tlb_all();
  79 }
  80
  81
  82 static inline void flush_map(unsigned long address)
  83 {
  84         on_each_cpu(flush_kernel_map, (void *)address, 1, 1);
  85 }
  86
  87 static struct page *deferred_pages; /* protected by init_mm.mmap_sem */
  88
  89 static inline void save_page(struct page *fpage)
  90 {
  91         fpage->lru.next = (struct list_head *)deferred_pages;
  92         deferred_pages = fpage;
  93 }
  94
  95 /*
  96  * No more special protections in this 2/4MB area - revert to a
  97  * large page again.
  98  */
  99 static void revert_page(unsigned long address, pgprot_t ref_prot)
 100 {
 101         pgd_t *pgd;
 102         pud_t *pud;
 103         pmd_t *pmd;
 104         pte_t large_pte;
 105
 106         pgd = pgd_offset_k(address);
 107         BUG_ON(pgd_none(*pgd));
 108         pud = pud_offset(pgd,address);
 109         BUG_ON(pud_none(*pud));
 110         pmd = pmd_offset(pud, address);
 111         BUG_ON(pmd_val(*pmd) & _PAGE_PSE);
 112         pgprot_val(ref_prot) |= _PAGE_PSE;
 113         large_pte = mk_pte_phys(__pa(address) & LARGE_PAGE_MASK, ref_prot);
 114         set_pte((pte_t *)pmd, large_pte);
 115 }
 116
 117 static int
 118 __change_page_attr(unsigned long address, unsigned long pfn, pgprot_t prot,
 119                                    pgprot_t ref_prot)
 120 {
 121         pte_t *kpte;
 122         struct page *kpte_page;
 123         unsigned kpte_flags;
 124         pgprot_t ref_prot2;
 125         kpte = lookup_address(address);
 126         if (!kpte) return 0;
 127         kpte_page = virt_to_page(((unsigned long)kpte) & PAGE_MASK);
 128         kpte_flags = pte_val(*kpte);
 129         if (pgprot_val(prot) != pgprot_val(ref_prot)) {
 130                 if ((kpte_flags & _PAGE_PSE) == 0) {
 131                         set_pte(kpte, pfn_pte(pfn, prot));
 132                 } else {
 133                         /*
 134                          * split_large_page will take the reference for this
 135                          * change_page_attr on the split page.
 136                          */
 137
 138                         struct page *split;
 139                         ref_prot2 = __pgprot(pgprot_val(pte_pgprot(*lookup_address(address))) & ~(1<<_PAGE_BIT_PSE));
 140
 141                         split = split_large_page(address, prot, ref_prot2);
 142                         if (!split)
 143                                 return -ENOMEM;
 144                         set_pte(kpte,mk_pte(split, ref_prot2));
 145                         kpte_page = split;
 146                 }
 147                 page_private(kpte_page)++;
 148         } else if ((kpte_flags & _PAGE_PSE) == 0) {
 149                 set_pte(kpte, pfn_pte(pfn, ref_prot));
 150                 BUG_ON(page_private(kpte_page) == 0);
 151                 page_private(kpte_page)--;
 152         } else
 153                 BUG();
 154
 155         /* on x86-64 the direct mapping set at boot is not using 4k pages */
 156         BUG_ON(PageReserved(kpte_page));
 157
 158         if (page_private(kpte_page) == 0) {
 159                 save_page(kpte_page);
 160                 revert_page(address, ref_prot);
 161         }
 162         return 0;
 163 }
 164
 165 /*
 166  * Change the page attributes of an page in the linear mapping.
 167  *
 168  * This should be used when a page is mapped with a different caching policy
 169  * than write-back somewhere - some CPUs do not like it when mappings with
 170  * different caching policies exist. This changes the page attributes of the
 171  * in kernel linear mapping too.
 172  *
 173  * The caller needs to ensure that there are no conflicting mappings elsewhere.
 174  * This function only deals with the kernel linear map.
 175  *
 176  * Caller must call global_flush_tlb() after this.
 177  */
 178 int change_page_attr_addr(unsigned long address, int numpages, pgprot_t prot)
 179 {
 180         int err = 0;
 181         int i;
 182
 183         down_write(&init_mm.mmap_sem);
 184         for (i = 0; i < numpages; i++, address += PAGE_SIZE) {
 185                 unsigned long pfn = __pa(address) >> PAGE_SHIFT;
 186
 187                 err = __change_page_attr(address, pfn, prot, PAGE_KERNEL);
 188                 if (err)
 189                         break;
 190                 /* Handle kernel mapping too which aliases part of the
 191                  * lowmem */
 192                 if (__pa(address) < KERNEL_TEXT_SIZE) {
 193                         unsigned long addr2;
 194                         pgprot_t prot2 = prot;
 195                         addr2 = __START_KERNEL_map + __pa(address);
 196                         pgprot_val(prot2) &= ~_PAGE_NX;
 197                         err = __change_page_attr(addr2, pfn, prot2, PAGE_KERNEL_EXEC);
 198                 }
 199         }
 200         up_write(&init_mm.mmap_sem);
 201         return err;
 202 }
 203
 204 /* Don't call this for MMIO areas that may not have a mem_map entry */
 205 int change_page_attr(struct page *page, int numpages, pgprot_t prot)
 206 {
 207         unsigned long addr = (unsigned long)page_address(page);
 208         return change_page_attr_addr(addr, numpages, prot);
 209 }
 210
 211 void global_flush_tlb(void)
 212 {
 213         struct page *dpage;
 214
 215         down_read(&init_mm.mmap_sem);
 216         dpage = xchg(&deferred_pages, NULL);
 217         up_read(&init_mm.mmap_sem);
 218
 219         flush_map((dpage && !dpage->lru.next) ? (unsigned long)page_address(dpage) : 0);
 220         while (dpage) {
 221                 struct page *tmp = dpage;
 222                 dpage = (struct page *)dpage->lru.next;
 223                 ClearPagePrivate(tmp);
 224                 __free_page(tmp);
 225         }
 226 }
 227
 228 EXPORT_SYMBOL(change_page_attr);
 229 EXPORT_SYMBOL(global_flush_tlb);