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 *dummy)
  73 {
  74         /* Could use CLFLUSH here if the CPU supports it (Hammer,P4) */
  75         if (boot_cpu_data.x86_model >= 4)
  76                 wbinvd();
  77         /* Flush all to work around Errata in early athlons regarding
  78          * large page flushing.
  79          */
  80         __flush_tlb_all();
  81 }
  82
  83 static void set_pmd_pte(pte_t *kpte, unsigned long address, pte_t pte)
  84 {
  85         struct page *page;
  86         unsigned long flags;
  87
  88         set_pte_atomic(kpte, pte);      /* change init_mm */
  89         if (HAVE_SHARED_KERNEL_PMD)
  90                 return;
  91
  92         spin_lock_irqsave(&pgd_lock, flags);
  93         for (page = pgd_list; page; page = (struct page *)page->index) {
  94                 pgd_t *pgd;
  95                 pud_t *pud;
  96                 pmd_t *pmd;
  97                 pgd = (pgd_t *)page_address(page) + pgd_index(address);
  98                 pud = pud_offset(pgd, address);
  99                 pmd = pmd_offset(pud, address);
 100                 set_pte_atomic((pte_t *)pmd, pte);
 101         }
 102         spin_unlock_irqrestore(&pgd_lock, flags);
 103 }
 104
 105 /*
 106  * No more special protections in this 2/4MB area - revert to a
 107  * large page again.
 108  */
 109 static inline void revert_page(struct page *kpte_page, unsigned long address)
 110 {
 111         pgprot_t ref_prot;
 112         pte_t *linear;
 113
 114         ref_prot =
 115         ((address & LARGE_PAGE_MASK) < (unsigned long)&_etext)
 116                 ? PAGE_KERNEL_LARGE_EXEC : PAGE_KERNEL_LARGE;
 117
 118         linear = (pte_t *)
 119                 pmd_offset(pud_offset(pgd_offset_k(address), address), address);
 120         set_pmd_pte(linear,  address,
 121                     pfn_pte((__pa(address) & LARGE_PAGE_MASK) >> PAGE_SHIFT,
 122                             ref_prot));
 123 }
 124
 125 static int
 126 __change_page_attr(struct page *page, pgprot_t prot)
 127 {
 128         pte_t *kpte;
 129         unsigned long address;
 130         struct page *kpte_page;
 131
 132         BUG_ON(PageHighMem(page));
 133         address = (unsigned long)page_address(page);
 134
 135         if (address >= (unsigned long)__start_rodata && address <= (unsigned long)__end_rodata &&
 136                 (pgprot_val(prot) & _PAGE_RW)) {
 137                 pgprot_val(prot) &= ~(_PAGE_RW);
 138                 add_taint(TAINT_MACHINE_CHECK);
 139         }
 140
 141         kpte = lookup_address(address);
 142         if (!kpte)
 143                 return -EINVAL;
 144         kpte_page = virt_to_page(kpte);
 145         if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL)) {
 146                 if ((pte_val(*kpte) & _PAGE_PSE) == 0) {
 147                         set_pte_atomic(kpte, mk_pte(page, prot));
 148                 } else {
 149                         pgprot_t ref_prot;
 150                         struct page *split;
 151
 152                         ref_prot =
 153                         ((address & LARGE_PAGE_MASK) < (unsigned long)&_etext)
 154                                 ? PAGE_KERNEL_EXEC : PAGE_KERNEL;
 155                         split = split_large_page(address, prot, ref_prot);
 156                         if (!split)
 157                                 return -ENOMEM;
 158                         set_pmd_pte(kpte,address,mk_pte(split, ref_prot));
 159                         kpte_page = split;
 160                 }
 161                 page_private(kpte_page)++;
 162         } else if ((pte_val(*kpte) & _PAGE_PSE) == 0) {
 163                 set_pte_atomic(kpte, mk_pte(page, PAGE_KERNEL));
 164                 BUG_ON(page_private(kpte_page) == 0);
 165                 page_private(kpte_page)--;
 166         } else
 167                 BUG();
 168
 169         /*
 170          * If the pte was reserved, it means it was created at boot
 171          * time (not via split_large_page) and in turn we must not
 172          * replace it with a largepage.
 173          */
 174         if (!PageReserved(kpte_page)) {
 175                 if (cpu_has_pse && (page_private(kpte_page) == 0)) {
 176                         ClearPagePrivate(kpte_page);
 177                         list_add(&kpte_page->lru, &df_list);
 178                         revert_page(kpte_page, address);
 179                 }
 180         }
 181         return 0;
 182 }
 183
 184 static inline void flush_map(void)
 185 {
 186         on_each_cpu(flush_kernel_map, NULL, 1, 1);
 187 }
 188
 189 /*
 190  * Change the page attributes of an page in the linear mapping.
 191  *
 192  * This should be used when a page is mapped with a different caching policy
 193  * than write-back somewhere - some CPUs do not like it when mappings with
 194  * different caching policies exist. This changes the page attributes of the
 195  * in kernel linear mapping too.
 196  *
 197  * The caller needs to ensure that there are no conflicting mappings elsewhere.
 198  * This function only deals with the kernel linear map.
 199  *
 200  * Caller must call global_flush_tlb() after this.
 201  */
 202 int change_page_attr(struct page *page, int numpages, pgprot_t prot)
 203 {
 204         int err = 0;
 205         int i;
 206         unsigned long flags;
 207
 208         spin_lock_irqsave(&cpa_lock, flags);
 209         for (i = 0; i < numpages; i++, page++) {
 210                 err = __change_page_attr(page, prot);
 211                 if (err)
 212                         break;
 213         }
 214         spin_unlock_irqrestore(&cpa_lock, flags);
 215         return err;
 216 }
 217
 218 void global_flush_tlb(void)
 219 {
 220         struct list_head l;
 221         struct page *pg, *next;
 222
 223         BUG_ON(irqs_disabled());
 224
 225         spin_lock_irq(&cpa_lock);
 226         list_replace_init(&df_list, &l);
 227         spin_unlock_irq(&cpa_lock);
 228         flush_map();
 229         list_for_each_entry_safe(pg, next, &l, lru)
 230                 __free_page(pg);
 231 }
 232
 233 #ifdef CONFIG_DEBUG_PAGEALLOC
 234 void kernel_map_pages(struct page *page, int numpages, int enable)
 235 {
 236         if (PageHighMem(page))
 237                 return;
 238         if (!enable)
 239                 debug_check_no_locks_freed(page_address(page),
 240                                            numpages * PAGE_SIZE);
 241
 242         /* the return value is ignored - the calls cannot fail,
 243          * large pages are disabled at boot time.
 244          */
 245         change_page_attr(page, numpages, enable ? PAGE_KERNEL : __pgprot(0));
 246         /* we should perform an IPI and flush all tlbs,
 247          * but that can deadlock->flush only current cpu.
 248          */
 249         __flush_tlb_all();
 250 }
 251 #endif
 252
 253 EXPORT_SYMBOL(change_page_attr);
 254 EXPORT_SYMBOL(global_flush_tlb);