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         pmd_t *pmd;
  21         pte_t *pte;
  22         if (!pgd || !pgd_present(*pgd))
  23                 return NULL;
  24         pmd = pmd_offset(pgd, address);
  25         if (!pmd_present(*pmd))
  26                 return NULL;
  27         if (pmd_large(*pmd))
  28                 return (pte_t *)pmd;
  29         pte = pte_offset_kernel(pmd, address);
  30         if (pte && !pte_present(*pte))
  31                 pte = NULL;
  32         return pte;
  33 }
  34
  35 static struct page *split_large_page(unsigned long address, pgprot_t prot)
  36 {
  37         int i;
  38         unsigned long addr;
  39         struct page *base = alloc_pages(GFP_KERNEL, 0);
  40         pte_t *pbase;
  41         if (!base)
  42                 return NULL;
  43         address = __pa(address);
  44         addr = address & LARGE_PAGE_MASK;
  45         pbase = (pte_t *)page_address(base);
  46         for (i = 0; i < PTRS_PER_PTE; i++, addr += PAGE_SIZE) {
  47                 pbase[i] = pfn_pte(addr >> PAGE_SHIFT,
  48                                    addr == address ? prot : PAGE_KERNEL);
  49         }
  50         return base;
  51 }
  52
  53
  54 static void flush_kernel_map(void *address)
  55 {
  56         if (0 && address && cpu_has_clflush) {
  57                 /* is this worth it? */
  58                 int i;
  59                 for (i = 0; i < PAGE_SIZE; i += boot_cpu_data.x86_clflush_size)
  60                         asm volatile("clflush (%0)" :: "r" (address + i));
  61         } else
  62                 asm volatile("wbinvd":::"memory");
  63         __flush_tlb_one(address);
  64 }
  65
  66
  67 static inline void flush_map(unsigned long address)
  68 {
  69         on_each_cpu(flush_kernel_map, (void *)address, 1, 1);
  70 }
  71
  72 struct deferred_page {
  73         struct deferred_page *next;
  74         struct page *fpage;
  75         unsigned long address;
  76 };
  77 static struct deferred_page *df_list; /* protected by init_mm.mmap_sem */
  78
  79 static inline void save_page(unsigned long address, struct page *fpage)
  80 {
  81         struct deferred_page *df;
  82         df = kmalloc(sizeof(struct deferred_page), GFP_KERNEL);
  83         if (!df) {
  84                 flush_map(address);
  85                 __free_page(fpage);
  86         } else {
  87                 df->next = df_list;
  88                 df->fpage = fpage;
  89                 df->address = address;
  90                 df_list = df;
  91         }
  92 }
  93
  94 /*
  95  * No more special protections in this 2/4MB area - revert to a
  96  * large page again.
  97  */
  98 static void revert_page(struct page *kpte_page, unsigned long address)
  99 {
 100        pgd_t *pgd;
 101        pmd_t *pmd;
 102        pte_t large_pte;
 103
 104        pgd = pgd_offset_k(address);
 105        pmd = pmd_offset(pgd, address);
 106        BUG_ON(pmd_val(*pmd) & _PAGE_PSE);
 107        large_pte = mk_pte_phys(__pa(address) & LARGE_PAGE_MASK, PAGE_KERNEL_LARGE);
 108        set_pte((pte_t *)pmd, large_pte);
 109 }
 110
 111 static int
 112 __change_page_attr(unsigned long address, struct page *page, pgprot_t prot)
 113 {
 114         pte_t *kpte;
 115         struct page *kpte_page;
 116         unsigned kpte_flags;
 117
 118         kpte = lookup_address(address);
 119         if (!kpte) return 0;
 120         kpte_page = virt_to_page(((unsigned long)kpte) & PAGE_MASK);
 121         kpte_flags = pte_val(*kpte);
 122         if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL)) {
 123                 if ((kpte_flags & _PAGE_PSE) == 0) {
 124                         pte_t old = *kpte;
 125                         pte_t standard = mk_pte(page, PAGE_KERNEL);
 126
 127                         set_pte(kpte, mk_pte(page, prot));
 128                         if (pte_same(old,standard))
 129                                 atomic_inc(&kpte_page->count);
 130                 } else {
 131                         struct page *split = split_large_page(address, prot);
 132                         if (!split)
 133                                 return -ENOMEM;
 134                         atomic_inc(&kpte_page->count);
 135                         set_pte(kpte,mk_pte(split, PAGE_KERNEL));
 136                 }
 137         } else if ((kpte_flags & _PAGE_PSE) == 0) {
 138                 set_pte(kpte, mk_pte(page, PAGE_KERNEL));
 139                 atomic_dec(&kpte_page->count);
 140         }
 141
 142         if (atomic_read(&kpte_page->count) == 1) {
 143                 save_page(address, kpte_page);
 144                 revert_page(kpte_page, address);
 145         }
 146         return 0;
 147 }
 148
 149 /*
 150  * Change the page attributes of an page in the linear mapping.
 151  *
 152  * This should be used when a page is mapped with a different caching policy
 153  * than write-back somewhere - some CPUs do not like it when mappings with
 154  * different caching policies exist. This changes the page attributes of the
 155  * in kernel linear mapping too.
 156  *
 157  * The caller needs to ensure that there are no conflicting mappings elsewhere.
 158  * This function only deals with the kernel linear map.
 159  *
 160  * Caller must call global_flush_tlb() after this.
 161  */
 162 int change_page_attr(struct page *page, int numpages, pgprot_t prot)
 163 {
 164         int err = 0;
 165         int i;
 166
 167         down_write(&init_mm.mmap_sem);
 168         for (i = 0; i < numpages; !err && i++, page++) {
 169                 unsigned long address = (unsigned long)page_address(page);
 170                 err = __change_page_attr(address, page, prot);
 171                 if (err)
 172                         break;
 173                 /* Handle kernel mapping too which aliases part of the lowmem */
 174                 if (page_to_phys(page) < KERNEL_TEXT_SIZE) {
 175                         unsigned long addr2 = __START_KERNEL_map + page_to_phys(page);
 176                         err = __change_page_attr(addr2, page, prot);
 177                 }
 178         }
 179         up_write(&init_mm.mmap_sem);
 180         return err;
 181 }
 182
 183 void global_flush_tlb(void)
 184 {
 185         struct deferred_page *df, *next_df;
 186
 187         down_read(&init_mm.mmap_sem);
 188         df = xchg(&df_list, NULL);
 189         up_read(&init_mm.mmap_sem);
 190         flush_map((df && !df->next) ? df->address : 0);
 191         for (; df; df = next_df) {
 192                 next_df = df->next;
 193                 if (df->fpage)
 194                         __free_page(df->fpage);
 195                 kfree(df);
 196         }
 197 }
 198
 199 EXPORT_SYMBOL(change_page_attr);
 200 EXPORT_SYMBOL(global_flush_tlb);