arch/xen/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/pgalloc.h>
  16 #include <asm/io.h>
  17
  18 void pte_free(struct page *pte)
  19 {
  20         pte_t *ptep;
  21
  22         ptep = pfn_to_kaddr(page_to_pfn(pte));
  23
  24         xen_pte_unpin(__pa(ptep));
  25         make_page_writable(ptep);
  26         __free_page(pte);
  27 }
  28
  29 static inline pte_t *lookup_address(unsigned long address)
  30 {
  31         pgd_t *pgd = pgd_offset_k(address);
  32         pud_t *pud;
  33         pmd_t *pmd;
  34         pte_t *pte;
  35         if (pgd_none(*pgd))
  36                 return NULL;
  37         pud = pud_offset(pgd, address);
  38         if (!pud_present(*pud))
  39                 return NULL;
  40         pmd = pmd_offset(pud, address);
  41         if (!pmd_present(*pmd))
  42                 return NULL;
  43         if (pmd_large(*pmd))
  44                 return (pte_t *)pmd;
  45         pte = pte_offset_kernel(pmd, address);
  46         if (pte && !pte_present(*pte))
  47                 pte = NULL;
  48         return pte;
  49 }
  50
  51 static struct page *split_large_page(unsigned long address, pgprot_t prot,
  52                                      pgprot_t ref_prot)
  53 {
  54         int i;
  55         unsigned long addr;
  56         struct page *base = alloc_pages(GFP_KERNEL, 0);
  57         pte_t *pbase;
  58         if (!base)
  59                 return NULL;
  60         address = __pa(address);
  61         addr = address & LARGE_PAGE_MASK;
  62         pbase = (pte_t *)page_address(base);
  63         for (i = 0; i < PTRS_PER_PTE; i++, addr += PAGE_SIZE) {
  64                 pbase[i] = pfn_pte(addr >> PAGE_SHIFT,
  65                                    addr == address ? prot : ref_prot);
  66         }
  67         return base;
  68 }
  69
  70
  71 static void flush_kernel_map(void *address)
  72 {
  73         if (0 && address && cpu_has_clflush) {
  74                 /* is this worth it? */
  75                 int i;
  76                 for (i = 0; i < PAGE_SIZE; i += boot_cpu_data.x86_clflush_size)
  77                         asm volatile("clflush (%0)" :: "r" (address + i));
  78         } else
  79                 asm volatile("wbinvd":::"memory");
  80         if (address)
  81                 __flush_tlb_one((unsigned long) address);
  82         else
  83                 __flush_tlb_all();
  84 }
  85
  86
  87 static inline void flush_map(unsigned long address)
  88 {
  89         on_each_cpu(flush_kernel_map, (void *)address, 1, 1);
  90 }
  91
  92 struct deferred_page {
  93         struct deferred_page *next;
  94         struct page *fpage;
  95         unsigned long address;
  96 };
  97 static struct deferred_page *df_list; /* protected by init_mm.mmap_sem */
  98
  99 static inline void save_page(unsigned long address, struct page *fpage)
 100 {
 101         struct deferred_page *df;
 102         df = kmalloc(sizeof(struct deferred_page), GFP_KERNEL);
 103         if (!df) {
 104                 flush_map(address);
 105                 __free_page(fpage);
 106         } else {
 107                 df->next = df_list;
 108                 df->fpage = fpage;
 109                 df->address = address;
 110                 df_list = df;
 111         }
 112 }
 113
 114 /*
 115  * No more special protections in this 2/4MB area - revert to a
 116  * large page again.
 117  */
 118 static void revert_page(unsigned long address, pgprot_t ref_prot)
 119 {
 120         pgd_t *pgd;
 121         pud_t *pud;
 122         pmd_t *pmd;
 123         pte_t large_pte;
 124
 125         pgd = pgd_offset_k(address);
 126         BUG_ON(pgd_none(*pgd));
 127         pud = pud_offset(pgd,address);
 128         BUG_ON(pud_none(*pud));
 129         pmd = pmd_offset(pud, address);
 130         BUG_ON(pmd_val(*pmd) & _PAGE_PSE);
 131         pgprot_val(ref_prot) |= _PAGE_PSE;
 132         large_pte = mk_pte_phys(__pa(address) & LARGE_PAGE_MASK, ref_prot);
 133         set_pte((pte_t *)pmd, large_pte);
 134 }
 135
 136 static int
 137 __change_page_attr(unsigned long address, unsigned long pfn, pgprot_t prot,
 138                                    pgprot_t ref_prot)
 139 {
 140         pte_t *kpte;
 141         struct page *kpte_page;
 142         unsigned kpte_flags;
 143         kpte = lookup_address(address);
 144         if (!kpte) return 0;
 145         kpte_page = virt_to_page(((unsigned long)kpte) & PAGE_MASK);
 146         kpte_flags = pte_val(*kpte);
 147         if (pgprot_val(prot) != pgprot_val(ref_prot)) {
 148                 if ((kpte_flags & _PAGE_PSE) == 0) {
 149                         set_pte(kpte, pfn_pte(pfn, prot));
 150                 } else {
 151                         /*
 152                          * split_large_page will take the reference for this change_page_attr
 153                          * on the split page.
 154                          */
 155                         struct page *split = split_large_page(address, prot, ref_prot);
 156                         if (!split)
 157                                 return -ENOMEM;
 158                         set_pte(kpte,mk_pte(split, ref_prot));
 159                         kpte_page = split;
 160                 }
 161                 get_page(kpte_page);
 162         } else if ((kpte_flags & _PAGE_PSE) == 0) {
 163                 set_pte(kpte, pfn_pte(pfn, ref_prot));
 164                 __put_page(kpte_page);
 165         } else
 166                 BUG();
 167
 168         /* on x86-64 the direct mapping set at boot is not using 4k pages */
 169         BUG_ON(PageReserved(kpte_page));
 170
 171         switch (page_count(kpte_page)) {
 172         case 1:
 173                 save_page(address, kpte_page);
 174                 revert_page(address, ref_prot);
 175                 break;
 176         case 0:
 177                 BUG(); /* memleak and failed 2M page regeneration */
 178         }
 179         return 0;
 180 }
 181
 182 /*
 183  * Change the page attributes of an page in the linear mapping.
 184  *
 185  * This should be used when a page is mapped with a different caching policy
 186  * than write-back somewhere - some CPUs do not like it when mappings with
 187  * different caching policies exist. This changes the page attributes of the
 188  * in kernel linear mapping too.
 189  *
 190  * The caller needs to ensure that there are no conflicting mappings elsewhere.
 191  * This function only deals with the kernel linear map.
 192  *
 193  * Caller must call global_flush_tlb() after this.
 194  */
 195 int change_page_attr_addr(unsigned long address, int numpages, pgprot_t prot)
 196 {
 197         int err = 0;
 198         int i;
 199
 200         down_write(&init_mm.mmap_sem);
 201         for (i = 0; i < numpages; i++, address += PAGE_SIZE) {
 202                 unsigned long pfn = __pa(address) >> PAGE_SHIFT;
 203
 204                 err = __change_page_attr(address, pfn, prot, PAGE_KERNEL);
 205                 if (err)
 206                         break;
 207                 /* Handle kernel mapping too which aliases part of the
 208                  * lowmem */
 209                 if (__pa(address) < KERNEL_TEXT_SIZE) {
 210                         unsigned long addr2;
 211                         pgprot_t prot2 = prot;
 212                         addr2 = __START_KERNEL_map + __pa(address);
 213                         pgprot_val(prot2) &= ~_PAGE_NX;
 214                         err = __change_page_attr(addr2, pfn, prot2, PAGE_KERNEL_EXEC);
 215                 }
 216         }
 217         up_write(&init_mm.mmap_sem);
 218         return err;
 219 }
 220
 221 /* Don't call this for MMIO areas that may not have a mem_map entry */
 222 int change_page_attr(struct page *page, int numpages, pgprot_t prot)
 223 {
 224         unsigned long addr = (unsigned long)page_address(page);
 225         return change_page_attr_addr(addr, numpages, prot);
 226 }
 227
 228 void global_flush_tlb(void)
 229 {
 230         struct deferred_page *df, *next_df;
 231
 232         down_read(&init_mm.mmap_sem);
 233         df = xchg(&df_list, NULL);
 234         up_read(&init_mm.mmap_sem);
 235         if (!df)
 236                 return;
 237         flush_map((df && !df->next) ? df->address : 0);
 238         for (; df; df = next_df) {
 239                 next_df = df->next;
 240                 if (df->fpage)
 241                         __free_page(df->fpage);
 242                 kfree(df);
 243         }
 244 }
 245
 246 EXPORT_SYMBOL(change_page_attr);
 247 EXPORT_SYMBOL(global_flush_tlb);