arch/i386/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 <linux/kernel.h>
  12 #include <asm/sections.h>
  13 #include <asm/uaccess.h>
  14 #include <asm/processor.h>
  15 #include <asm/tlbflush.h>
  16 #include <asm/pgalloc.h>
  17 #include <asm/sections.h>
  18
  19 static DEFINE_SPINLOCK(cpa_lock);
  20 static struct list_head df_list = LIST_HEAD_INIT(df_list);
  21
  22
  23 pte_t *lookup_address(unsigned long address)
  24 {
  25         pgd_t *pgd = pgd_offset_k(address);
  26         pud_t *pud;
  27         pmd_t *pmd;
  28         if (pgd_none(*pgd))
  29                 return NULL;
  30         pud = pud_offset(pgd, address);
  31         if (pud_none(*pud))
  32                 return NULL;
  33         pmd = pmd_offset(pud, address);
  34         if (pmd_none(*pmd))
  35                 return NULL;
  36         if (pmd_large(*pmd))
  37                 return (pte_t *)pmd;
  38         return pte_offset_kernel(pmd, address);
  39 }
  40
  41 static struct page *split_large_page(unsigned long address, pgprot_t prot,
  42                                         pgprot_t ref_prot)
  43 {
  44         int i;
  45         unsigned long addr;
  46         struct page *base;
  47         pte_t *pbase;
  48
  49         spin_unlock_irq(&cpa_lock);
  50         base = alloc_pages(GFP_KERNEL, 0);
  51         spin_lock_irq(&cpa_lock);
  52         if (!base)
  53                 return NULL;
  54
  55         /*
  56          * page_private is used to track the number of entries in
  57          * the page table page that 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                set_pte(&pbase[i], pfn_pte(addr >> PAGE_SHIFT,
  67                                           addr == address ? prot : ref_prot));
  68         }
  69         return base;
  70 }
  71
  72 static void flush_kernel_map(void *arg)
  73 {
  74         unsigned long adr = (unsigned long)arg;
  75
  76         if (adr && cpu_has_clflush) {
  77                 int i;
  78                 for (i = 0; i < PAGE_SIZE; i += boot_cpu_data.x86_clflush_size)
  79                         asm volatile("clflush (%0)" :: "r" (adr + i));
  80         } else if (boot_cpu_data.x86_model >= 4)
  81                 wbinvd();
  82
  83         /* Flush all to work around Errata in early athlons regarding
  84          * large page flushing.
  85          */
  86         __flush_tlb_all();
  87 }
  88
  89 static void set_pmd_pte(pte_t *kpte, unsigned long address, pte_t pte)
  90 {
  91         struct page *page;
  92         unsigned long flags;
  93
  94         set_pte_atomic(kpte, pte);      /* change init_mm */
  95         if (HAVE_SHARED_KERNEL_PMD)
  96                 return;
  97
  98         spin_lock_irqsave(&pgd_lock, flags);
  99         for (page = pgd_list; page; page = (struct page *)page->index) {
 100                 pgd_t *pgd;
 101                 pud_t *pud;
 102                 pmd_t *pmd;
 103                 pgd = (pgd_t *)page_address(page) + pgd_index(address);
 104                 pud = pud_offset(pgd, address);
 105                 pmd = pmd_offset(pud, address);
 106                 set_pte_atomic((pte_t *)pmd, pte);
 107         }
 108         spin_unlock_irqrestore(&pgd_lock, flags);
 109 }
 110
 111 /*
 112  * No more special protections in this 2/4MB area - revert to a
 113  * large page again.
 114  */
 115 static inline void revert_page(struct page *kpte_page, unsigned long address)
 116 {
 117         pgprot_t ref_prot;
 118         pte_t *linear;
 119
 120         ref_prot =
 121         ((address & LARGE_PAGE_MASK) < (unsigned long)&_etext)
 122                 ? PAGE_KERNEL_LARGE_EXEC : PAGE_KERNEL_LARGE;
 123
 124         linear = (pte_t *)
 125                 pmd_offset(pud_offset(pgd_offset_k(address), address), address);
 126         set_pmd_pte(linear,  address,
 127                     pfn_pte((__pa(address) & LARGE_PAGE_MASK) >> PAGE_SHIFT,
 128                             ref_prot));
 129 }
 130
 131 static int
 132 __change_page_attr(struct page *page, pgprot_t prot)
 133 {
 134         pte_t *kpte;
 135         unsigned long address;
 136         struct page *kpte_page;
 137
 138         BUG_ON(PageHighMem(page));
 139         address = (unsigned long)page_address(page);
 140
 141         if (address >= (unsigned long)__start_rodata && address <= (unsigned long)__end_rodata &&
 142                 (pgprot_val(prot) & _PAGE_RW)) {
 143                 pgprot_val(prot) &= ~(_PAGE_RW);
 144                 add_taint(TAINT_MACHINE_CHECK);
 145         }
 146
 147         kpte = lookup_address(address);
 148         if (!kpte)
 149                 return -EINVAL;
 150         kpte_page = virt_to_page(kpte);
 151         if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL)) {
 152                 if ((pte_val(*kpte) & _PAGE_PSE) == 0) {
 153                         set_pte_atomic(kpte, mk_pte(page, prot));
 154                 } else {
 155                         pgprot_t ref_prot;
 156                         struct page *split;
 157
 158                         ref_prot =
 159                         ((address & LARGE_PAGE_MASK) < (unsigned long)&_etext)
 160                                 ? PAGE_KERNEL_EXEC : PAGE_KERNEL;
 161                         split = split_large_page(address, prot, ref_prot);
 162                         if (!split)
 163                                 return -ENOMEM;
 164                         set_pmd_pte(kpte,address,mk_pte(split, ref_prot));
 165                         kpte_page = split;
 166                 }
 167                 page_private(kpte_page)++;
 168         } else if ((pte_val(*kpte) & _PAGE_PSE) == 0) {
 169                 set_pte_atomic(kpte, mk_pte(page, PAGE_KERNEL));
 170                 BUG_ON(page_private(kpte_page) == 0);
 171                 page_private(kpte_page)--;
 172         } else
 173                 BUG();
 174
 175         /*
 176          * If the pte was reserved, it means it was created at boot
 177          * time (not via split_large_page) and in turn we must not
 178          * replace it with a largepage.
 179          */
 180         if (!PageReserved(kpte_page)) {
 181                 if (cpu_has_pse && (page_private(kpte_page) == 0)) {
 182                         ClearPagePrivate(kpte_page);
 183                         list_add(&kpte_page->lru, &df_list);
 184                         revert_page(kpte_page, address);
 185                 }
 186         }
 187         return 0;
 188 }
 189
 190 static inline void flush_map(void *adr)
 191 {
 192         on_each_cpu(flush_kernel_map, adr, 1, 1);
 193 }
 194
 195 /*
 196  * Change the page attributes of an page in the linear mapping.
 197  *
 198  * This should be used when a page is mapped with a different caching policy
 199  * than write-back somewhere - some CPUs do not like it when mappings with
 200  * different caching policies exist. This changes the page attributes of the
 201  * in kernel linear mapping too.
 202  *
 203  * The caller needs to ensure that there are no conflicting mappings elsewhere.
 204  * This function only deals with the kernel linear map.
 205  *
 206  * Caller must call global_flush_tlb() after this.
 207  */
 208 int change_page_attr(struct page *page, int numpages, pgprot_t prot)
 209 {
 210         int err = 0;
 211         int i;
 212         unsigned long flags;
 213
 214         spin_lock_irqsave(&cpa_lock, flags);
 215         for (i = 0; i < numpages; i++, page++) {
 216                 err = __change_page_attr(page, prot);
 217                 if (err)
 218                         break;
 219         }
 220         spin_unlock_irqrestore(&cpa_lock, flags);
 221         return err;
 222 }
 223
 224 void global_flush_tlb(void)
 225 {
 226         struct list_head l;
 227         struct page *pg, *next;
 228
 229         BUG_ON(irqs_disabled());
 230
 231         spin_lock_irq(&cpa_lock);
 232         list_replace_init(&df_list, &l);
 233         spin_unlock_irq(&cpa_lock);
 234         if (!cpu_has_clflush)
 235                 flush_map(0);
 236         list_for_each_entry_safe(pg, next, &l, lru) {
 237                 if (cpu_has_clflush)
 238                         flush_map(page_address(pg));
 239                 __free_page(pg);
 240         }
 241 }
 242
 243 #ifdef CONFIG_DEBUG_PAGEALLOC
 244 void kernel_map_pages(struct page *page, int numpages, int enable)
 245 {
 246         if (PageHighMem(page))
 247                 return;
 248         if (!enable)
 249                 debug_check_no_locks_freed(page_address(page),
 250                                            numpages * PAGE_SIZE);
 251
 252         /* the return value is ignored - the calls cannot fail,
 253          * large pages are disabled at boot time.
 254          */
 255         change_page_attr(page, numpages, enable ? PAGE_KERNEL : __pgprot(0));
 256         /* we should perform an IPI and flush all tlbs,
 257          * but that can deadlock->flush only current cpu.
 258          */
 259         __flush_tlb_all();
 260 }
 261 #endif
 262
 263 EXPORT_SYMBOL(change_page_attr);
 264 EXPORT_SYMBOL(global_flush_tlb);