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                                      pgprot_t ref_prot)
  37 {
  38         int i;
  39         unsigned long addr;
  40         struct page *base = alloc_pages(GFP_KERNEL, 0);
  41         pte_t *pbase;
  42         if (!base)
  43                 return NULL;
  44         address = __pa(address);
  45         addr = address & LARGE_PAGE_MASK;
  46         pbase = (pte_t *)page_address(base);
  47         for (i = 0; i < PTRS_PER_PTE; i++, addr += PAGE_SIZE) {
  48                 pbase[i] = pfn_pte(addr >> PAGE_SHIFT,
  49                                    addr == address ? prot : ref_prot);
  50         }
  51         return base;
  52 }
  53
  54
  55 static void flush_kernel_map(void *address)
  56 {
  57         if (0 && address && cpu_has_clflush) {
  58                 /* is this worth it? */
  59                 int i;
  60                 for (i = 0; i < PAGE_SIZE; i += boot_cpu_data.x86_clflush_size)
  61                         asm volatile("clflush (%0)" :: "r" (address + i));
  62         } else
  63                 asm volatile("wbinvd":::"memory");
  64         __flush_tlb_one(address);
  65 }
  66
  67
  68 static inline void flush_map(unsigned long address)
  69 {
  70         on_each_cpu(flush_kernel_map, (void *)address, 1, 1);
  71 }
  72
  73 struct deferred_page {
  74         struct deferred_page *next;
  75         struct page *fpage;
  76         unsigned long address;
  77 };
  78 static struct deferred_page *df_list; /* protected by init_mm.mmap_sem */
  79
  80 static inline void save_page(unsigned long address, struct page *fpage)
  81 {
  82         struct deferred_page *df;
  83         df = kmalloc(sizeof(struct deferred_page), GFP_KERNEL);
  84         if (!df) {
  85                 flush_map(address);
  86                 __free_page(fpage);
  87         } else {
  88                 df->next = df_list;
  89                 df->fpage = fpage;
  90                 df->address = address;
  91                 df_list = df;
  92         }
  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        pmd_t *pmd;
 103        pte_t large_pte;
 104
 105        pgd = pgd_offset_k(address);
 106        pmd = pmd_offset(pgd, address);
 107        BUG_ON(pmd_val(*pmd) & _PAGE_PSE);
 108        pgprot_val(ref_prot) |= _PAGE_PSE;
 109        large_pte = mk_pte_phys(__pa(address) & LARGE_PAGE_MASK, ref_prot);
 110        set_pte((pte_t *)pmd, large_pte);
 111 }
 112
 113 static int
 114 __change_page_attr(unsigned long address, struct page *page, pgprot_t prot,
 115                    pgprot_t ref_prot)
 116 {
 117         pte_t *kpte;
 118         struct page *kpte_page;
 119         unsigned kpte_flags;
 120
 121         kpte = lookup_address(address);
 122         if (!kpte) return 0;
 123         kpte_page = virt_to_page(((unsigned long)kpte) & PAGE_MASK);
 124         kpte_flags = pte_val(*kpte);
 125         if (pgprot_val(prot) != pgprot_val(ref_prot)) {
 126                 if ((kpte_flags & _PAGE_PSE) == 0) {
 127                         pte_t old = *kpte;
 128                         pte_t standard = mk_pte(page, ref_prot);
 129
 130                         set_pte(kpte, mk_pte(page, prot));
 131                         if (pte_same(old,standard))
 132                                 get_page(kpte_page);
 133                 } else {
 134                         struct page *split = split_large_page(address, prot, ref_prot);
 135                         if (!split)
 136                                 return -ENOMEM;
 137                         get_page(kpte_page);
 138                         set_pte(kpte,mk_pte(split, ref_prot));
 139                 }
 140         } else if ((kpte_flags & _PAGE_PSE) == 0) {
 141                 set_pte(kpte, mk_pte(page, ref_prot));
 142                 __put_page(kpte_page);
 143         }
 144
 145         if (page_count(kpte_page) == 1) {
 146                 save_page(address, kpte_page);
 147                 revert_page(address, ref_prot);
 148         }
 149         return 0;
 150 }
 151
 152 /*
 153  * Change the page attributes of an page in the linear mapping.
 154  *
 155  * This should be used when a page is mapped with a different caching policy
 156  * than write-back somewhere - some CPUs do not like it when mappings with
 157  * different caching policies exist. This changes the page attributes of the
 158  * in kernel linear mapping too.
 159  *
 160  * The caller needs to ensure that there are no conflicting mappings elsewhere.
 161  * This function only deals with the kernel linear map.
 162  *
 163  * Caller must call global_flush_tlb() after this.
 164  */
 165 int change_page_attr(struct page *page, int numpages, pgprot_t prot)
 166 {
 167         int err = 0;
 168         int i;
 169
 170         down_write(&init_mm.mmap_sem);
 171         for (i = 0; i < numpages; !err && i++, page++) {
 172                 unsigned long address = (unsigned long)page_address(page);
 173                 err = __change_page_attr(address, page, prot, PAGE_KERNEL);
 174                 if (err)
 175                         break;
 176                 /* Handle kernel mapping too which aliases part of the
 177                  * lowmem */
 178                 /* Disabled right now. Fixme */
 179                 if (0 && page_to_phys(page) < KERNEL_TEXT_SIZE) {
 180                         unsigned long addr2;
 181                         addr2 = __START_KERNEL_map + page_to_phys(page);
 182                         err = __change_page_attr(addr2, page, prot,
 183                                                  PAGE_KERNEL_EXECUTABLE);
 184                 }
 185         }
 186         up_write(&init_mm.mmap_sem);
 187         return err;
 188 }
 189
 190 void global_flush_tlb(void)
 191 {
 192         struct deferred_page *df, *next_df;
 193
 194         down_read(&init_mm.mmap_sem);
 195         df = xchg(&df_list, NULL);
 196         up_read(&init_mm.mmap_sem);
 197         flush_map((df && !df->next) ? df->address : 0);
 198         for (; df; df = next_df) {
 199                 next_df = df->next;
 200                 if (df->fpage)
 201                         __free_page(df->fpage);
 202                 kfree(df);
 203         }
 204 }
 205
 206 EXPORT_SYMBOL(change_page_attr);
 207 EXPORT_SYMBOL(global_flush_tlb);