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 <asm/uaccess.h>
  12 #include <asm/processor.h>
  13 #include <asm/tlbflush.h>
  14 #include <asm/pgalloc.h>
  15 #include <asm/sections.h>
  16
  17 static DEFINE_SPINLOCK(cpa_lock);
  18 static struct list_head df_list = LIST_HEAD_INIT(df_list);
  19
  20
  21 pte_t *lookup_address(unsigned long address)
  22 {
  23         pgd_t *pgd = pgd_offset_k(address);
  24         pud_t *pud;
  25         pmd_t *pmd;
  26         if (pgd_none(*pgd))
  27                 return NULL;
  28         pud = pud_offset(pgd, address);
  29         if (pud_none(*pud))
  30                 return NULL;
  31         pmd = pmd_offset(pud, address);
  32         if (pmd_none(*pmd))
  33                 return NULL;
  34         if (pmd_large(*pmd))
  35                 return (pte_t *)pmd;
  36         return pte_offset_kernel(pmd, address);
  37 }
  38
  39 static struct page *split_large_page(unsigned long address, pgprot_t prot,
  40                                         pgprot_t ref_prot)
  41 {
  42         int i;
  43         unsigned long addr;
  44         struct page *base;
  45         pte_t *pbase;
  46
  47         spin_unlock_irq(&cpa_lock);
  48         base = alloc_pages(GFP_KERNEL, 0);
  49         spin_lock_irq(&cpa_lock);
  50         if (!base)
  51                 return NULL;
  52
  53         /*
  54          * page_private is used to track the number of entries in
  55          * the page table page that have non standard attributes.
  56          */
  57         SetPagePrivate(base);
  58         page_private(base) = 0;
  59
  60         address = __pa(address);
  61         addr = address & LARGE_PAGE_MASK;
  62         pbase = (pte_t *)page_address(base);
  63         paravirt_alloc_pt(page_to_pfn(base));
  64         for (i = 0; i < PTRS_PER_PTE; i++, addr += PAGE_SIZE) {
  65                set_pte(&pbase[i], pfn_pte(addr >> PAGE_SHIFT,
  66                                           addr == address ? prot : ref_prot));
  67         }
  68         return base;
  69 }
  70
  71 static void flush_kernel_map(void *arg)
  72 {
  73         unsigned long adr = (unsigned long)arg;
  74
  75         if (adr && cpu_has_clflush) {
  76                 int i;
  77                 for (i = 0; i < PAGE_SIZE; i += boot_cpu_data.x86_clflush_size)
  78                         asm volatile("clflush (%0)" :: "r" (adr + i));
  79         } else if (boot_cpu_data.x86_model >= 4)
  80                 wbinvd();
  81
  82         /* Flush all to work around Errata in early athlons regarding
  83          * large page flushing.
  84          */
  85         __flush_tlb_all();
  86 }
  87
  88 static void set_pmd_pte(pte_t *kpte, unsigned long address, pte_t pte)
  89 {
  90         struct page *page;
  91         unsigned long flags;
  92
  93         set_pte_atomic(kpte, pte);      /* change init_mm */
  94         if (SHARED_KERNEL_PMD)
  95                 return;
  96
  97         spin_lock_irqsave(&pgd_lock, flags);
  98         for (page = pgd_list; page; page = (struct page *)page->index) {
  99                 pgd_t *pgd;
 100                 pud_t *pud;
 101                 pmd_t *pmd;
 102                 pgd = (pgd_t *)page_address(page) + pgd_index(address);
 103                 pud = pud_offset(pgd, address);
 104                 pmd = pmd_offset(pud, address);
 105                 set_pte_atomic((pte_t *)pmd, pte);
 106         }
 107         spin_unlock_irqrestore(&pgd_lock, flags);
 108 }
 109
 110 /*
 111  * No more special protections in this 2/4MB area - revert to a
 112  * large page again.
 113  */
 114 static inline void revert_page(struct page *kpte_page, unsigned long address)
 115 {
 116         pgprot_t ref_prot;
 117         pte_t *linear;
 118
 119         ref_prot =
 120         ((address & LARGE_PAGE_MASK) < (unsigned long)&_etext)
 121                 ? PAGE_KERNEL_LARGE_EXEC : PAGE_KERNEL_LARGE;
 122
 123         linear = (pte_t *)
 124                 pmd_offset(pud_offset(pgd_offset_k(address), address), address);
 125         set_pmd_pte(linear,  address,
 126                     pfn_pte((__pa(address) & LARGE_PAGE_MASK) >> PAGE_SHIFT,
 127                             ref_prot));
 128 }
 129
 130 static int
 131 __change_page_attr(struct page *page, pgprot_t prot)
 132 {
 133         pte_t *kpte;
 134         unsigned long address;
 135         struct page *kpte_page;
 136
 137         BUG_ON(PageHighMem(page));
 138         address = (unsigned long)page_address(page);
 139
 140         kpte = lookup_address(address);
 141         if (!kpte)
 142                 return -EINVAL;
 143         kpte_page = virt_to_page(kpte);
 144         if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL)) {
 145                 if (!pte_huge(*kpte)) {
 146                         set_pte_atomic(kpte, mk_pte(page, prot));
 147                 } else {
 148                         pgprot_t ref_prot;
 149                         struct page *split;
 150
 151                         ref_prot =
 152                         ((address & LARGE_PAGE_MASK) < (unsigned long)&_etext)
 153                                 ? PAGE_KERNEL_EXEC : PAGE_KERNEL;
 154                         split = split_large_page(address, prot, ref_prot);
 155                         if (!split)
 156                                 return -ENOMEM;
 157                         set_pmd_pte(kpte,address,mk_pte(split, ref_prot));
 158                         kpte_page = split;
 159                 }
 160                 page_private(kpte_page)++;
 161         } else if (!pte_huge(*kpte)) {
 162                 set_pte_atomic(kpte, mk_pte(page, PAGE_KERNEL));
 163                 BUG_ON(page_private(kpte_page) == 0);
 164                 page_private(kpte_page)--;
 165         } else
 166                 BUG();
 167
 168         /*
 169          * If the pte was reserved, it means it was created at boot
 170          * time (not via split_large_page) and in turn we must not
 171          * replace it with a largepage.
 172          */
 173         if (!PageReserved(kpte_page)) {
 174                 if (cpu_has_pse && (page_private(kpte_page) == 0)) {
 175                         ClearPagePrivate(kpte_page);
 176                         paravirt_release_pt(page_to_pfn(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 *adr)
 185 {
 186         on_each_cpu(flush_kernel_map, adr, 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         if (!cpu_has_clflush)
 229                 flush_map(NULL);
 230         list_for_each_entry_safe(pg, next, &l, lru) {
 231                 if (cpu_has_clflush)
 232                         flush_map(page_address(pg));
 233                 __free_page(pg);
 234         }
 235 }
 236
 237 #ifdef CONFIG_DEBUG_PAGEALLOC
 238 void kernel_map_pages(struct page *page, int numpages, int enable)
 239 {
 240         if (PageHighMem(page))
 241                 return;
 242         if (!enable)
 243                 debug_check_no_locks_freed(page_address(page),
 244                                            numpages * PAGE_SIZE);
 245
 246         /* the return value is ignored - the calls cannot fail,
 247          * large pages are disabled at boot time.
 248          */
 249         change_page_attr(page, numpages, enable ? PAGE_KERNEL : __pgprot(0));
 250         /* we should perform an IPI and flush all tlbs,
 251          * but that can deadlock->flush only current cpu.
 252          */
 253         __flush_tlb_all();
 254 }
 255 #endif
 256
 257 EXPORT_SYMBOL(change_page_attr);
 258 EXPORT_SYMBOL(global_flush_tlb);