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         addr = address & LARGE_PAGE_MASK;
  55         pbase = (pte_t *)page_address(base);
  56         for (i = 0; i < PTRS_PER_PTE; i++, addr += PAGE_SIZE) {
  57                 pbase[i] = pfn_pte(addr >> PAGE_SHIFT,
  58                                    addr == address ? prot : ref_prot);
  59         }
  60         return base;
  61 }
  62
  63 static void cache_flush_page(void *adr)
  64 {
  65         int i;
  66         for (i = 0; i < PAGE_SIZE; i += boot_cpu_data.x86_clflush_size)
  67                 asm volatile("clflush (%0)" :: "r" (adr + i));
  68 }
  69
  70 static void flush_kernel_map(void *arg)
  71 {
  72         struct list_head *l = (struct list_head *)arg;
  73         struct page *pg;
  74
  75         /* When clflush is available always use it because it is
  76            much cheaper than WBINVD */
  77         if (!cpu_has_clflush)
  78                 asm volatile("wbinvd" ::: "memory");
  79         list_for_each_entry(pg, l, lru) {
  80                 void *adr = page_address(pg);
  81                 if (cpu_has_clflush)
  82                         cache_flush_page(adr);
  83         }
  84         __flush_tlb_all();
  85 }
  86
  87 static inline void flush_map(struct list_head *l)
  88 {
  89         on_each_cpu(flush_kernel_map, l, 1, 1);
  90 }
  91
  92 static LIST_HEAD(deferred_pages); /* protected by init_mm.mmap_sem */
  93
  94 static inline void save_page(struct page *fpage)
  95 {
  96         list_add(&fpage->lru, &deferred_pages);
  97 }
  98
  99 /*
 100  * No more special protections in this 2/4MB area - revert to a
 101  * large page again.
 102  */
 103 static void revert_page(unsigned long address, unsigned long pfn, pgprot_t ref_prot)
 104 {
 105         pgd_t *pgd;
 106         pud_t *pud;
 107         pmd_t *pmd;
 108         pte_t large_pte;
 109
 110         pgd = pgd_offset_k(address);
 111         BUG_ON(pgd_none(*pgd));
 112         pud = pud_offset(pgd,address);
 113         BUG_ON(pud_none(*pud));
 114         pmd = pmd_offset(pud, address);
 115         BUG_ON(pmd_val(*pmd) & _PAGE_PSE);
 116         large_pte = pfn_pte(pfn, ref_prot);
 117         large_pte = pte_mkhuge(large_pte);
 118         set_pte((pte_t *)pmd, large_pte);
 119 }
 120
 121 static int
 122 __change_page_attr(unsigned long address, unsigned long pfn, pgprot_t prot,
 123                                    pgprot_t ref_prot)
 124 {
 125         pte_t *kpte;
 126         struct page *kpte_page;
 127         pgprot_t ref_prot2;
 128         kpte = lookup_address(address);
 129         if (!kpte) return 0;
 130         kpte_page = virt_to_page(((unsigned long)kpte) & PAGE_MASK);
 131         if (pgprot_val(prot) != pgprot_val(ref_prot)) {
 132                 if (!pte_huge(*kpte)) {
 133                         set_pte(kpte, pfn_pte(pfn, prot));
 134                 } else {
 135                         /*
 136                          * split_large_page will take the reference for this
 137                          * change_page_attr on the split page.
 138                          */
 139                         struct page *split;
 140                         ref_prot2 = pte_pgprot(pte_clrhuge(*kpte));
 141                         split = split_large_page(pfn << PAGE_SHIFT, prot,
 142                                                         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, pfn, 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         unsigned long phys_base_pfn = __pa_symbol(__START_KERNEL_map) >> PAGE_SHIFT;
 182         int err = 0, kernel_map = 0;
 183         int i;
 184
 185         if (address >= __START_KERNEL_map
 186             && address < __START_KERNEL_map + KERNEL_TEXT_SIZE) {
 187                 address = (unsigned long)__va(__pa(address));
 188                 kernel_map = 1;
 189         }
 190
 191         down_write(&init_mm.mmap_sem);
 192         for (i = 0; i < numpages; i++, address += PAGE_SIZE) {
 193                 unsigned long pfn = __pa(address) >> PAGE_SHIFT;
 194
 195                 if (!kernel_map || pte_present(pfn_pte(0, prot))) {
 196                         err = __change_page_attr(address, pfn, prot, PAGE_KERNEL);
 197                         if (err)
 198                                 break;
 199                 }
 200                 /* Handle kernel mapping too which aliases part of the
 201                  * lowmem */
 202                 if ((pfn >= phys_base_pfn) &&
 203                         ((pfn - phys_base_pfn) < (KERNEL_TEXT_SIZE >> PAGE_SHIFT))) {
 204                         unsigned long addr2;
 205                         pgprot_t prot2;
 206                         addr2 = __START_KERNEL_map + ((pfn - phys_base_pfn) << PAGE_SHIFT);
 207                         /* Make sure the kernel mappings stay executable */
 208                         prot2 = pte_pgprot(pte_mkexec(pfn_pte(0, prot)));
 209                         err = __change_page_attr(addr2, pfn, prot2,
 210                                                  PAGE_KERNEL_EXEC);
 211                 }
 212         }
 213         up_write(&init_mm.mmap_sem);
 214         return err;
 215 }
 216
 217 /* Don't call this for MMIO areas that may not have a mem_map entry */
 218 int change_page_attr(struct page *page, int numpages, pgprot_t prot)
 219 {
 220         unsigned long addr = (unsigned long)page_address(page);
 221         return change_page_attr_addr(addr, numpages, prot);
 222 }
 223
 224 void global_flush_tlb(void)
 225 {
 226         struct page *pg, *next;
 227         struct list_head l;
 228
 229         down_read(&init_mm.mmap_sem);
 230         list_replace_init(&deferred_pages, &l);
 231         up_read(&init_mm.mmap_sem);
 232
 233         flush_map(&l);
 234
 235         list_for_each_entry_safe(pg, next, &l, lru) {
 236                 ClearPagePrivate(pg);
 237                 __free_page(pg);
 238         }
 239 }
 240
 241 EXPORT_SYMBOL(change_page_attr);
 242 EXPORT_SYMBOL(global_flush_tlb);