git.oblomov.eu Git - linux-2.6/blob - 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/mm.h>
   7 #include <linux/sched.h>
   8 #include <linux/highmem.h>
   9 #include <linux/module.h>
  10 #include <linux/slab.h>
  11 #include <asm/uaccess.h>
  12 #include <asm/processor.h>
  13 #include <asm/tlbflush.h>
  14 #include <asm/io.h>
  15
  16 static inline pte_t *lookup_address(unsigned long address)
  17 {
  18         pgd_t *pgd = pgd_offset_k(address);
  19         pud_t *pud;
  20         pmd_t *pmd;
  21         pte_t *pte;
  22         if (pgd_none(*pgd))
  23                 return NULL;
  24         pud = pud_offset(pgd, address);
  25         if (!pud_present(*pud))
  26                 return NULL;
  27         pmd = pmd_offset(pud, address);
  28         if (!pmd_present(*pmd))
  29                 return NULL;
  30         if (pmd_large(*pmd))
  31                 return (pte_t *)pmd;
  32         pte = pte_offset_kernel(pmd, address);
  33         if (pte && !pte_present(*pte))
  34                 pte = NULL;
  35         return pte;
  36 }
  37
  38 static struct page *split_large_page(unsigned long address, pgprot_t prot,
  39                                      pgprot_t ref_prot)
  40 {
  41         int i;
  42         unsigned long addr;
  43         struct page *base = alloc_pages(GFP_KERNEL, 0);
  44         pte_t *pbase;
  45         if (!base)
  46                 return NULL;
  47         /*
  48          * page_private is used to track the number of entries in
  49          * the page table page have non standard attributes.
  50          */
  51         SetPagePrivate(base);
  52         page_private(base) = 0;
  53
  54         address = __pa(address);
  55         addr = address & LARGE_PAGE_MASK;
  56         pbase = (pte_t *)page_address(base);
  57         for (i = 0; i < PTRS_PER_PTE; i++, addr += PAGE_SIZE) {
  58                 pbase[i] = pfn_pte(addr >> PAGE_SHIFT,
  59                                    addr == address ? prot : ref_prot);
  60         }
  61         return base;
  62 }
  63
  64 static void cache_flush_page(void *adr)
  65 {
  66         int i;
  67         for (i = 0; i < PAGE_SIZE; i += boot_cpu_data.x86_clflush_size)
  68                 asm volatile("clflush (%0)" :: "r" (adr + i));
  69 }
  70
  71 static void flush_kernel_map(void *arg)
  72 {
  73         struct list_head *l = (struct list_head *)arg;
  74         struct page *pg;
  75
  76         /* When clflush is available always use it because it is
  77            much cheaper than WBINVD */
  78         if (!cpu_has_clflush)
  79                 asm volatile("wbinvd" ::: "memory");
  80         list_for_each_entry(pg, l, lru) {
  81                 void *adr = page_address(pg);
  82                 if (cpu_has_clflush)
  83                         cache_flush_page(adr);
  84                 __flush_tlb_one(adr);
  85         }
  86 }
  87
  88 static inline void flush_map(struct list_head *l)
  89 {
  90         on_each_cpu(flush_kernel_map, l, 1, 1);
  91 }
  92
  93 static LIST_HEAD(deferred_pages); /* protected by init_mm.mmap_sem */
  94
  95 static inline void save_page(struct page *fpage)
  96 {
  97         list_add(&fpage->lru, &deferred_pages);
  98 }
  99
 100 /*
 101  * No more special protections in this 2/4MB area - revert to a
 102  * large page again.
 103  */
 104 static void revert_page(unsigned long address, pgprot_t ref_prot)
 105 {
 106         pgd_t *pgd;
 107         pud_t *pud;
 108         pmd_t *pmd;
 109         pte_t large_pte;
 110
 111         pgd = pgd_offset_k(address);
 112         BUG_ON(pgd_none(*pgd));
 113         pud = pud_offset(pgd,address);
 114         BUG_ON(pud_none(*pud));
 115         pmd = pmd_offset(pud, address);
 116         BUG_ON(pmd_val(*pmd) & _PAGE_PSE);
 117         large_pte = mk_pte_phys(__pa(address) & LARGE_PAGE_MASK, ref_prot);
 118         large_pte = pte_mkhuge(large_pte);
 119         set_pte((pte_t *)pmd, large_pte);
 120 }
 121
 122 static int
 123 __change_page_attr(unsigned long address, unsigned long pfn, pgprot_t prot,
 124                                    pgprot_t ref_prot)
 125 {
 126         pte_t *kpte;
 127         struct page *kpte_page;
 128         pgprot_t ref_prot2;
 129         kpte = lookup_address(address);
 130         if (!kpte) return 0;
 131         kpte_page = virt_to_page(((unsigned long)kpte) & PAGE_MASK);
 132         if (pgprot_val(prot) != pgprot_val(ref_prot)) {
 133                 if (!pte_huge(*kpte)) {
 134                         set_pte(kpte, pfn_pte(pfn, prot));
 135                 } else {
 136                         /*
 137                          * split_large_page will take the reference for this
 138                          * change_page_attr on the split page.
 139                          */
 140                         struct page *split;
 141                         ref_prot2 = pte_pgprot(pte_clrhuge(*kpte));
 142                         split = split_large_page(address, prot, ref_prot2);
 143                         if (!split)
 144                                 return -ENOMEM;
 145                         set_pte(kpte, mk_pte(split, ref_prot2));
 146                         kpte_page = split;
 147                 }
 148                 page_private(kpte_page)++;
 149         } else if (!pte_huge(*kpte)) {
 150                 set_pte(kpte, pfn_pte(pfn, ref_prot));
 151                 BUG_ON(page_private(kpte_page) == 0);
 152                 page_private(kpte_page)--;
 153         } else
 154                 BUG();
 155
 156         /* on x86-64 the direct mapping set at boot is not using 4k pages */
 157         BUG_ON(PageReserved(kpte_page));
 158
 159         if (page_private(kpte_page) == 0) {
 160                 save_page(kpte_page);
 161                 revert_page(address, ref_prot);
 162         }
 163         return 0;
 164 }
 165
 166 /*
 167  * Change the page attributes of an page in the linear mapping.
 168  *
 169  * This should be used when a page is mapped with a different caching policy
 170  * than write-back somewhere - some CPUs do not like it when mappings with
 171  * different caching policies exist. This changes the page attributes of the
 172  * in kernel linear mapping too.
 173  *
 174  * The caller needs to ensure that there are no conflicting mappings elsewhere.
 175  * This function only deals with the kernel linear map.
 176  *
 177  * Caller must call global_flush_tlb() after this.
 178  */
 179 int change_page_attr_addr(unsigned long address, int numpages, pgprot_t prot)
 180 {
 181         int err = 0;
 182         int i;
 183
 184         down_write(&init_mm.mmap_sem);
 185         for (i = 0; i < numpages; i++, address += PAGE_SIZE) {
 186                 unsigned long pfn = __pa(address) >> PAGE_SHIFT;
 187
 188                 err = __change_page_attr(address, pfn, prot, PAGE_KERNEL);
 189                 if (err)
 190                         break;
 191                 /* Handle kernel mapping too which aliases part of the
 192                  * lowmem */
 193                 if (__pa(address) < KERNEL_TEXT_SIZE) {
 194                         unsigned long addr2;
 195                         pgprot_t prot2;
 196                         addr2 = __START_KERNEL_map + __pa(address);
 197                         /* Make sure the kernel mappings stay executable */
 198                         prot2 = pte_pgprot(pte_mkexec(pfn_pte(0, prot)));
 199                         err = __change_page_attr(addr2, pfn, prot2,
 200                                                  PAGE_KERNEL_EXEC);
 201                 }
 202         }
 203         up_write(&init_mm.mmap_sem);
 204         return err;
 205 }
 206
 207 /* Don't call this for MMIO areas that may not have a mem_map entry */
 208 int change_page_attr(struct page *page, int numpages, pgprot_t prot)
 209 {
 210         unsigned long addr = (unsigned long)page_address(page);
 211         return change_page_attr_addr(addr, numpages, prot);
 212 }
 213
 214 void global_flush_tlb(void)
 215 {
 216         struct page *pg, *next;
 217         struct list_head l;
 218
 219         down_read(&init_mm.mmap_sem);
 220         list_replace_init(&deferred_pages, &l);
 221         up_read(&init_mm.mmap_sem);
 222
 223         flush_map(&l);
 224
 225         list_for_each_entry_safe(pg, next, &l, lru) {
 226                 ClearPagePrivate(pg);
 227                 __free_page(pg);
 228         }
 229 }
 230
 231 EXPORT_SYMBOL(change_page_attr);
 232 EXPORT_SYMBOL(global_flush_tlb);