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/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/sections.h>
  17
  18 static DEFINE_SPINLOCK(cpa_lock);
  19 static struct list_head df_list = LIST_HEAD_INIT(df_list);
  20
  21
  22 pte_t *lookup_address(unsigned long address)
  23 {
  24         pgd_t *pgd = pgd_offset_k(address);
  25         pud_t *pud;
  26         pmd_t *pmd;
  27         if (pgd_none(*pgd))
  28                 return NULL;
  29         pud = pud_offset(pgd, address);
  30         if (pud_none(*pud))
  31                 return NULL;
  32         pmd = pmd_offset(pud, address);
  33         if (pmd_none(*pmd))
  34                 return NULL;
  35         if (pmd_large(*pmd))
  36                 return (pte_t *)pmd;
  37         return pte_offset_kernel(pmd, address);
  38 }
  39
  40 static struct page *split_large_page(unsigned long address, pgprot_t prot,
  41                                         pgprot_t ref_prot)
  42 {
  43         int i;
  44         unsigned long addr;
  45         struct page *base;
  46         pte_t *pbase;
  47
  48         spin_unlock_irq(&cpa_lock);
  49         base = alloc_pages(GFP_KERNEL, 0);
  50         spin_lock_irq(&cpa_lock);
  51         if (!base)
  52                 return NULL;
  53
  54         /*
  55          * page_private is used to track the number of entries in
  56          * the page table page that have non standard attributes.
  57          */
  58         SetPagePrivate(base);
  59         page_private(base) = 0;
  60
  61         address = __pa(address);
  62         addr = address & LARGE_PAGE_MASK;
  63         pbase = (pte_t *)page_address(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 *dummy)
  72 {
  73         /* Could use CLFLUSH here if the CPU supports it (Hammer,P4) */
  74         if (boot_cpu_data.x86_model >= 4)
  75                 wbinvd();
  76         /* Flush all to work around Errata in early athlons regarding
  77          * large page flushing.
  78          */
  79         __flush_tlb_all();
  80 }
  81
  82 static void set_pmd_pte(pte_t *kpte, unsigned long address, pte_t pte)
  83 {
  84         struct page *page;
  85         unsigned long flags;
  86
  87         set_pte_atomic(kpte, pte);      /* change init_mm */
  88 #ifndef HAVE_SHARED_KERNEL_PMD
  89         if (PTRS_PER_PMD > 1)
  90 #else
  91         if (HAVE_SHARED_KERNEL_PMD)
  92 #endif
  93                 return;
  94
  95         spin_lock_irqsave(&pgd_lock, flags);
  96         for (page = pgd_list; page; page = (struct page *)page->index) {
  97                 pgd_t *pgd;
  98                 pud_t *pud;
  99                 pmd_t *pmd;
 100                 pgd = (pgd_t *)page_address(page) + pgd_index(address);
 101                 pud = pud_offset(pgd, address);
 102                 pmd = pmd_offset(pud, address);
 103                 set_pte_atomic((pte_t *)pmd, pte);
 104         }
 105         spin_unlock_irqrestore(&pgd_lock, flags);
 106 }
 107
 108 /*
 109  * No more special protections in this 2/4MB area - revert to a
 110  * large page again.
 111  */
 112 static inline void revert_page(struct page *kpte_page, unsigned long address)
 113 {
 114         pgprot_t ref_prot;
 115         pte_t *linear;
 116
 117         ref_prot =
 118         ((address & LARGE_PAGE_MASK) < (unsigned long)&_etext)
 119                 ? PAGE_KERNEL_LARGE_EXEC : PAGE_KERNEL_LARGE;
 120
 121         linear = (pte_t *)
 122                 pmd_offset(pud_offset(pgd_offset_k(address), address), address);
 123         set_pmd_pte(linear,  address,
 124                     pfn_pte((__pa(address) & LARGE_PAGE_MASK) >> PAGE_SHIFT,
 125                             ref_prot));
 126 }
 127
 128 static int
 129 __change_page_attr(struct page *page, pgprot_t prot)
 130 {
 131         pte_t *kpte;
 132         unsigned long address;
 133         struct page *kpte_page;
 134
 135         BUG_ON(PageHighMem(page));
 136         address = (unsigned long)page_address(page);
 137
 138         kpte = lookup_address(address);
 139         if (!kpte)
 140                 return -EINVAL;
 141         kpte_page = virt_to_page(kpte);
 142         if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL)) {
 143                 if ((pte_val(*kpte) & _PAGE_PSE) == 0) {
 144                         set_pte_atomic(kpte, mk_pte(page, prot));
 145                 } else {
 146                         pgprot_t ref_prot;
 147                         struct page *split;
 148
 149                         ref_prot =
 150                         ((address & LARGE_PAGE_MASK) < (unsigned long)&_etext)
 151                                 ? PAGE_KERNEL_EXEC : PAGE_KERNEL;
 152                         split = split_large_page(address, prot, ref_prot);
 153                         if (!split)
 154                                 return -ENOMEM;
 155                         set_pmd_pte(kpte,address,mk_pte(split, ref_prot));
 156                         kpte_page = split;
 157                 }
 158                 page_private(kpte_page)++;
 159         } else if ((pte_val(*kpte) & _PAGE_PSE) == 0) {
 160                 set_pte_atomic(kpte, mk_pte(page, PAGE_KERNEL));
 161                 BUG_ON(page_private(kpte_page) == 0);
 162                 page_private(kpte_page)--;
 163         } else
 164                 BUG();
 165
 166         /*
 167          * If the pte was reserved, it means it was created at boot
 168          * time (not via split_large_page) and in turn we must not
 169          * replace it with a largepage.
 170          */
 171         if (!PageReserved(kpte_page)) {
 172                 if (cpu_has_pse && (page_private(kpte_page) == 0)) {
 173                         ClearPagePrivate(kpte_page);
 174                         list_add(&kpte_page->lru, &df_list);
 175                         revert_page(kpte_page, address);
 176                 }
 177         }
 178         return 0;
 179 }
 180
 181 static inline void flush_map(void)
 182 {
 183         on_each_cpu(flush_kernel_map, NULL, 1, 1);
 184 }
 185
 186 /*
 187  * Change the page attributes of an page in the linear mapping.
 188  *
 189  * This should be used when a page is mapped with a different caching policy
 190  * than write-back somewhere - some CPUs do not like it when mappings with
 191  * different caching policies exist. This changes the page attributes of the
 192  * in kernel linear mapping too.
 193  *
 194  * The caller needs to ensure that there are no conflicting mappings elsewhere.
 195  * This function only deals with the kernel linear map.
 196  *
 197  * Caller must call global_flush_tlb() after this.
 198  */
 199 int change_page_attr(struct page *page, int numpages, pgprot_t prot)
 200 {
 201         int err = 0;
 202         int i;
 203         unsigned long flags;
 204
 205         spin_lock_irqsave(&cpa_lock, flags);
 206         for (i = 0; i < numpages; i++, page++) {
 207                 err = __change_page_attr(page, prot);
 208                 if (err)
 209                         break;
 210         }
 211         spin_unlock_irqrestore(&cpa_lock, flags);
 212         return err;
 213 }
 214
 215 void global_flush_tlb(void)
 216 {
 217         LIST_HEAD(l);
 218         struct page *pg, *next;
 219
 220         BUG_ON(irqs_disabled());
 221
 222         spin_lock_irq(&cpa_lock);
 223         list_splice_init(&df_list, &l);
 224         spin_unlock_irq(&cpa_lock);
 225         flush_map();
 226         list_for_each_entry_safe(pg, next, &l, lru)
 227                 __free_page(pg);
 228 }
 229
 230 #ifdef CONFIG_DEBUG_PAGEALLOC
 231 void kernel_map_pages(struct page *page, int numpages, int enable)
 232 {
 233         if (PageHighMem(page))
 234                 return;
 235         if (!enable)
 236                 mutex_debug_check_no_locks_freed(page_address(page),
 237                                                  numpages * PAGE_SIZE);
 238
 239         /* the return value is ignored - the calls cannot fail,
 240          * large pages are disabled at boot time.
 241          */
 242         change_page_attr(page, numpages, enable ? PAGE_KERNEL : __pgprot(0));
 243         /* we should perform an IPI and flush all tlbs,
 244          * but that can deadlock->flush only current cpu.
 245          */
 246         __flush_tlb_all();
 247 }
 248 #endif
 249
 250 EXPORT_SYMBOL(change_page_attr);
 251 EXPORT_SYMBOL(global_flush_tlb);