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
  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 {
  41         int i;
  42         unsigned long addr;
  43         struct page *base;
  44         pte_t *pbase;
  45
  46         spin_unlock_irq(&cpa_lock);
  47         base = alloc_pages(GFP_KERNEL, 0);
  48         spin_lock_irq(&cpa_lock);
  49         if (!base)
  50                 return NULL;
  51
  52         address = __pa(address);
  53         addr = address & LARGE_PAGE_MASK;
  54         pbase = (pte_t *)page_address(base);
  55         for (i = 0; i < PTRS_PER_PTE; i++, addr += PAGE_SIZE) {
  56                set_pte(&pbase[i], pfn_pte(addr >> PAGE_SHIFT,
  57                                           addr == address ? prot : PAGE_KERNEL));
  58         }
  59         return base;
  60 }
  61
  62 static void flush_kernel_map(void *dummy)
  63 {
  64         /* Could use CLFLUSH here if the CPU supports it (Hammer,P4) */
  65         if (boot_cpu_data.x86_model >= 4)
  66                 wbinvd();
  67         /* Flush all to work around Errata in early athlons regarding
  68          * large page flushing.
  69          */
  70         __flush_tlb_all();
  71 }
  72
  73 static void set_pmd_pte(pte_t *kpte, unsigned long address, pte_t pte)
  74 {
  75         struct page *page;
  76         unsigned long flags;
  77
  78         set_pte_atomic(kpte, pte);      /* change init_mm */
  79         if (PTRS_PER_PMD > 1)
  80                 return;
  81
  82         spin_lock_irqsave(&pgd_lock, flags);
  83         for (page = pgd_list; page; page = (struct page *)page->index) {
  84                 pgd_t *pgd;
  85                 pud_t *pud;
  86                 pmd_t *pmd;
  87                 pgd = (pgd_t *)page_address(page) + pgd_index(address);
  88                 pud = pud_offset(pgd, address);
  89                 pmd = pmd_offset(pud, address);
  90                 set_pte_atomic((pte_t *)pmd, pte);
  91         }
  92         spin_unlock_irqrestore(&pgd_lock, flags);
  93 }
  94
  95 /*
  96  * No more special protections in this 2/4MB area - revert to a
  97  * large page again.
  98  */
  99 static inline void revert_page(struct page *kpte_page, unsigned long address)
 100 {
 101         pte_t *linear = (pte_t *)
 102                 pmd_offset(pud_offset(pgd_offset_k(address), address), address);
 103         set_pmd_pte(linear,  address,
 104                     pfn_pte((__pa(address) & LARGE_PAGE_MASK) >> PAGE_SHIFT,
 105                             PAGE_KERNEL_LARGE));
 106 }
 107
 108 static int
 109 __change_page_attr(struct page *page, pgprot_t prot)
 110 {
 111         pte_t *kpte;
 112         unsigned long address;
 113         struct page *kpte_page;
 114
 115         BUG_ON(PageHighMem(page));
 116         address = (unsigned long)page_address(page);
 117
 118         kpte = lookup_address(address);
 119         if (!kpte)
 120                 return -EINVAL;
 121         kpte_page = virt_to_page(kpte);
 122         if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL)) {
 123                 if ((pte_val(*kpte) & _PAGE_PSE) == 0) {
 124                         set_pte_atomic(kpte, mk_pte(page, prot));
 125                 } else {
 126                         struct page *split = split_large_page(address, prot);
 127                         if (!split)
 128                                 return -ENOMEM;
 129                         set_pmd_pte(kpte,address,mk_pte(split, PAGE_KERNEL));
 130                         kpte_page = split;
 131                 }
 132                 get_page(kpte_page);
 133         } else if ((pte_val(*kpte) & _PAGE_PSE) == 0) {
 134                 set_pte_atomic(kpte, mk_pte(page, PAGE_KERNEL));
 135                 __put_page(kpte_page);
 136         } else
 137                 BUG();
 138
 139         /*
 140          * If the pte was reserved, it means it was created at boot
 141          * time (not via split_large_page) and in turn we must not
 142          * replace it with a largepage.
 143          */
 144         if (!PageReserved(kpte_page)) {
 145                 /* memleak and potential failed 2M page regeneration */
 146                 BUG_ON(!page_count(kpte_page));
 147
 148                 if (cpu_has_pse && (page_count(kpte_page) == 1)) {
 149                         list_add(&kpte_page->lru, &df_list);
 150                         revert_page(kpte_page, address);
 151                 }
 152         }
 153         return 0;
 154 }
 155
 156 static inline void flush_map(void)
 157 {
 158         on_each_cpu(flush_kernel_map, NULL, 1, 1);
 159 }
 160
 161 /*
 162  * Change the page attributes of an page in the linear mapping.
 163  *
 164  * This should be used when a page is mapped with a different caching policy
 165  * than write-back somewhere - some CPUs do not like it when mappings with
 166  * different caching policies exist. This changes the page attributes of the
 167  * in kernel linear mapping too.
 168  *
 169  * The caller needs to ensure that there are no conflicting mappings elsewhere.
 170  * This function only deals with the kernel linear map.
 171  *
 172  * Caller must call global_flush_tlb() after this.
 173  */
 174 int change_page_attr(struct page *page, int numpages, pgprot_t prot)
 175 {
 176         int err = 0;
 177         int i;
 178         unsigned long flags;
 179
 180         spin_lock_irqsave(&cpa_lock, flags);
 181         for (i = 0; i < numpages; i++, page++) {
 182                 err = __change_page_attr(page, prot);
 183                 if (err)
 184                         break;
 185         }
 186         spin_unlock_irqrestore(&cpa_lock, flags);
 187         return err;
 188 }
 189
 190 void global_flush_tlb(void)
 191 {
 192         LIST_HEAD(l);
 193         struct page *pg, *next;
 194
 195         BUG_ON(irqs_disabled());
 196
 197         spin_lock_irq(&cpa_lock);
 198         list_splice_init(&df_list, &l);
 199         spin_unlock_irq(&cpa_lock);
 200         flush_map();
 201         list_for_each_entry_safe(pg, next, &l, lru)
 202                 __free_page(pg);
 203 }
 204
 205 #ifdef CONFIG_DEBUG_PAGEALLOC
 206 void kernel_map_pages(struct page *page, int numpages, int enable)
 207 {
 208         if (PageHighMem(page))
 209                 return;
 210         /* the return value is ignored - the calls cannot fail,
 211          * large pages are disabled at boot time.
 212          */
 213         change_page_attr(page, numpages, enable ? PAGE_KERNEL : __pgprot(0));
 214         /* we should perform an IPI and flush all tlbs,
 215          * but that can deadlock->flush only current cpu.
 216          */
 217         __flush_tlb_all();
 218 }
 219 #endif
 220
 221 EXPORT_SYMBOL(change_page_attr);
 222 EXPORT_SYMBOL(global_flush_tlb);