arch/x86/mm/pageattr_64.c

   1 /*
   2  * Copyright 2002 Andi Kleen, SuSE Labs.
   3  * Thanks to Ben LaHaise for precious feedback.
   4  */
   5
   6 #include <linux/highmem.h>
   7 #include <linux/module.h>
   8 #include <linux/sched.h>
   9 #include <linux/slab.h>
  10 #include <linux/mm.h>
  11
  12 #include <asm/processor.h>
  13 #include <asm/tlbflush.h>
  14 #include <asm/uaccess.h>
  15 #include <asm/io.h>
  16
  17 pte_t *lookup_address(unsigned long address, int *level)
  18 {
  19         pgd_t *pgd = pgd_offset_k(address);
  20         pud_t *pud;
  21         pmd_t *pmd;
  22         pte_t *pte;
  23
  24         if (pgd_none(*pgd))
  25                 return NULL;
  26         pud = pud_offset(pgd, address);
  27         if (!pud_present(*pud))
  28                 return NULL;
  29         pmd = pmd_offset(pud, address);
  30         if (!pmd_present(*pmd))
  31                 return NULL;
  32         *level = 3;
  33         if (pmd_large(*pmd))
  34                 return (pte_t *)pmd;
  35         *level = 4;
  36
  37         pte = pte_offset_kernel(pmd, address);
  38         if (pte && !pte_present(*pte))
  39                 pte = NULL;
  40
  41         return pte;
  42 }
  43
  44 static struct page *
  45 split_large_page(unsigned long address, pgprot_t prot, pgprot_t ref_prot)
  46 {
  47         unsigned long addr;
  48         struct page *base;
  49         pte_t *pbase;
  50         int i;
  51
  52         base = alloc_pages(GFP_KERNEL, 0);
  53         if (!base)
  54                 return NULL;
  55         /*
  56          * page_private is used to track the number of entries in
  57          * the page table page have non standard attributes.
  58          */
  59         SetPagePrivate(base);
  60         page_private(base) = 0;
  61
  62         address = __pa(address);
  63         addr = address & LARGE_PAGE_MASK;
  64         pbase = (pte_t *)page_address(base);
  65         for (i = 0; i < PTRS_PER_PTE; i++, addr += PAGE_SIZE) {
  66                 pbase[i] = pfn_pte(addr >> PAGE_SHIFT,
  67                                    addr == address ? prot : ref_prot);
  68         }
  69         return base;
  70 }
  71
  72 void clflush_cache_range(void *addr, int size)
  73 {
  74         int i;
  75
  76         for (i = 0; i < size; i += boot_cpu_data.x86_clflush_size)
  77                 clflush(addr+i);
  78 }
  79
  80 static void flush_kernel_map(void *arg)
  81 {
  82         struct list_head *l = (struct list_head *)arg;
  83         struct page *pg;
  84
  85         __flush_tlb_all();
  86
  87         /* When clflush is available always use it because it is
  88            much cheaper than WBINVD. */
  89         /* clflush is still broken. Disable for now. */
  90         if (1 || !cpu_has_clflush) {
  91                 wbinvd();
  92         } else {
  93                 list_for_each_entry(pg, l, lru) {
  94                         void *addr = page_address(pg);
  95
  96                         clflush_cache_range(addr, PAGE_SIZE);
  97                 }
  98         }
  99 }
 100
 101 static inline void flush_map(struct list_head *l)
 102 {
 103         on_each_cpu(flush_kernel_map, l, 1, 1);
 104 }
 105
 106 static LIST_HEAD(deferred_pages); /* protected by init_mm.mmap_sem */
 107
 108 static inline void save_page(struct page *fpage)
 109 {
 110         if (!test_and_set_bit(PG_arch_1, &fpage->flags))
 111                 list_add(&fpage->lru, &deferred_pages);
 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         unsigned long pfn;
 121         pgd_t *pgd;
 122         pud_t *pud;
 123         pmd_t *pmd;
 124         pte_t large_pte;
 125
 126         pgd = pgd_offset_k(address);
 127         BUG_ON(pgd_none(*pgd));
 128         pud = pud_offset(pgd, address);
 129         BUG_ON(pud_none(*pud));
 130         pmd = pmd_offset(pud, address);
 131         BUG_ON(pmd_val(*pmd) & _PAGE_PSE);
 132         pfn = (__pa(address) & LARGE_PAGE_MASK) >> PAGE_SHIFT;
 133         large_pte = pfn_pte(pfn, ref_prot);
 134         large_pte = pte_mkhuge(large_pte);
 135
 136         set_pte((pte_t *)pmd, large_pte);
 137 }
 138
 139 static int
 140 __change_page_attr(unsigned long address, unsigned long pfn, pgprot_t prot,
 141                    pgprot_t ref_prot)
 142 {
 143         struct page *kpte_page;
 144         pgprot_t ref_prot2;
 145         pte_t *kpte;
 146         int level;
 147
 148         kpte = lookup_address(address, &level);
 149         if (!kpte)
 150                 return 0;
 151
 152         kpte_page = virt_to_page(kpte);
 153         BUG_ON(PageLRU(kpte_page));
 154         BUG_ON(PageCompound(kpte_page));
 155         if (pgprot_val(prot) != pgprot_val(ref_prot)) {
 156                 if (!pte_huge(*kpte)) {
 157                         set_pte(kpte, pfn_pte(pfn, prot));
 158                 } else {
 159                         /*
 160                          * split_large_page will take the reference for this
 161                          * change_page_attr on the split page.
 162                          */
 163                         struct page *split;
 164
 165                         ref_prot2 = pte_pgprot(pte_clrhuge(*kpte));
 166                         split = split_large_page(address, prot, ref_prot2);
 167                         if (!split)
 168                                 return -ENOMEM;
 169                         pgprot_val(ref_prot2) &= ~_PAGE_NX;
 170                         set_pte(kpte, mk_pte(split, ref_prot2));
 171                         kpte_page = split;
 172                 }
 173                 page_private(kpte_page)++;
 174         } else {
 175                 if (!pte_huge(*kpte)) {
 176                         set_pte(kpte, pfn_pte(pfn, ref_prot));
 177                         BUG_ON(page_private(kpte_page) == 0);
 178                         page_private(kpte_page)--;
 179                 } else
 180                         BUG();
 181         }
 182
 183         /* on x86-64 the direct mapping set at boot is not using 4k pages */
 184         BUG_ON(PageReserved(kpte_page));
 185
 186         save_page(kpte_page);
 187         if (page_private(kpte_page) == 0)
 188                 revert_page(address, ref_prot);
 189         return 0;
 190 }
 191
 192 /*
 193  * Change the page attributes of an page in the linear mapping.
 194  *
 195  * This should be used when a page is mapped with a different caching policy
 196  * than write-back somewhere - some CPUs do not like it when mappings with
 197  * different caching policies exist. This changes the page attributes of the
 198  * in kernel linear mapping too.
 199  *
 200  * The caller needs to ensure that there are no conflicting mappings elsewhere.
 201  * This function only deals with the kernel linear map.
 202  *
 203  * Caller must call global_flush_tlb() after this.
 204  */
 205 int change_page_attr_addr(unsigned long address, int numpages, pgprot_t prot)
 206 {
 207         int err = 0, kernel_map = 0, i;
 208
 209         if (address >= __START_KERNEL_map &&
 210                         address < __START_KERNEL_map + KERNEL_TEXT_SIZE) {
 211
 212                 address = (unsigned long)__va(__pa(address));
 213                 kernel_map = 1;
 214         }
 215
 216         down_write(&init_mm.mmap_sem);
 217         for (i = 0; i < numpages; i++, address += PAGE_SIZE) {
 218                 unsigned long pfn = __pa(address) >> PAGE_SHIFT;
 219
 220                 if (!kernel_map || pte_present(pfn_pte(0, prot))) {
 221                         err = __change_page_attr(address, pfn, prot,
 222                                                 PAGE_KERNEL);
 223                         if (err)
 224                                 break;
 225                 }
 226                 /* Handle kernel mapping too which aliases part of the
 227                  * lowmem */
 228                 if (__pa(address) < KERNEL_TEXT_SIZE) {
 229                         unsigned long addr2;
 230                         pgprot_t prot2;
 231
 232                         addr2 = __START_KERNEL_map + __pa(address);
 233                         /* Make sure the kernel mappings stay executable */
 234                         prot2 = pte_pgprot(pte_mkexec(pfn_pte(0, prot)));
 235                         err = __change_page_attr(addr2, pfn, prot2,
 236                                                  PAGE_KERNEL_EXEC);
 237                 }
 238         }
 239         up_write(&init_mm.mmap_sem);
 240
 241         return err;
 242 }
 243
 244 /* Don't call this for MMIO areas that may not have a mem_map entry */
 245 int change_page_attr(struct page *page, int numpages, pgprot_t prot)
 246 {
 247         unsigned long addr = (unsigned long)page_address(page);
 248
 249         return change_page_attr_addr(addr, numpages, prot);
 250 }
 251 EXPORT_SYMBOL(change_page_attr);
 252
 253 void global_flush_tlb(void)
 254 {
 255         struct page *pg, *next;
 256         struct list_head l;
 257
 258         /*
 259          * Write-protect the semaphore, to exclude two contexts
 260          * doing a list_replace_init() call in parallel and to
 261          * exclude new additions to the deferred_pages list:
 262          */
 263         down_write(&init_mm.mmap_sem);
 264         list_replace_init(&deferred_pages, &l);
 265         up_write(&init_mm.mmap_sem);
 266
 267         flush_map(&l);
 268
 269         list_for_each_entry_safe(pg, next, &l, lru) {
 270                 list_del(&pg->lru);
 271                 clear_bit(PG_arch_1, &pg->flags);
 272                 if (page_private(pg) != 0)
 273                         continue;
 274                 ClearPagePrivate(pg);
 275                 __free_page(pg);
 276         }
 277 }
 278 EXPORT_SYMBOL(global_flush_tlb);