arch/x86/mm/pageattr_32.c

   1 /*
   2  * Copyright 2002 Andi Kleen, SuSE Labs.
   3  * Thanks to Ben LaHaise for precious feedback.
   4  */
   5
   6 #include <linux/highmem.h>
   7 #include <linux/module.h>
   8 #include <linux/sched.h>
   9 #include <linux/slab.h>
  10 #include <linux/mm.h>
  11
  12 #include <asm/processor.h>
  13 #include <asm/tlbflush.h>
  14 #include <asm/sections.h>
  15 #include <asm/uaccess.h>
  16 #include <asm/pgalloc.h>
  17
  18 pte_t *lookup_address(unsigned long address, int *level)
  19 {
  20         pgd_t *pgd = pgd_offset_k(address);
  21         pud_t *pud;
  22         pmd_t *pmd;
  23
  24         if (pgd_none(*pgd))
  25                 return NULL;
  26         pud = pud_offset(pgd, address);
  27         if (pud_none(*pud))
  28                 return NULL;
  29         pmd = pmd_offset(pud, address);
  30         if (pmd_none(*pmd))
  31                 return NULL;
  32         *level = 2;
  33         if (pmd_large(*pmd))
  34                 return (pte_t *)pmd;
  35         *level = 3;
  36
  37         return pte_offset_kernel(pmd, address);
  38 }
  39
  40 static void set_pmd_pte(pte_t *kpte, unsigned long address, pte_t pte)
  41 {
  42         unsigned long flags;
  43         struct page *page;
  44
  45         /* change init_mm */
  46         set_pte_atomic(kpte, pte);
  47         if (SHARED_KERNEL_PMD)
  48                 return;
  49
  50         spin_lock_irqsave(&pgd_lock, flags);
  51         for (page = pgd_list; page; page = (struct page *)page->index) {
  52                 pgd_t *pgd;
  53                 pud_t *pud;
  54                 pmd_t *pmd;
  55
  56                 pgd = (pgd_t *)page_address(page) + pgd_index(address);
  57                 pud = pud_offset(pgd, address);
  58                 pmd = pmd_offset(pud, address);
  59                 set_pte_atomic((pte_t *)pmd, pte);
  60         }
  61         spin_unlock_irqrestore(&pgd_lock, flags);
  62 }
  63
  64 static int
  65 split_large_page(pte_t *kpte, unsigned long address, pgprot_t ref_prot)
  66 {
  67         int i, level;
  68         unsigned long addr;
  69         pte_t *pbase, *tmp;
  70         struct page *base;
  71
  72         base = alloc_pages(GFP_KERNEL, 0);
  73         if (!base)
  74                 return -ENOMEM;
  75
  76         down_write(&init_mm.mmap_sem);
  77         /*
  78          * Check for races, another CPU might have split this page
  79          * up for us already:
  80          */
  81         tmp = lookup_address(address, &level);
  82         if (tmp != kpte) {
  83                 WARN_ON_ONCE(1);
  84                 goto out_unlock;
  85         }
  86
  87         address = __pa(address);
  88         addr = address & LARGE_PAGE_MASK;
  89         pbase = (pte_t *)page_address(base);
  90         paravirt_alloc_pt(&init_mm, page_to_pfn(base));
  91
  92         for (i = 0; i < PTRS_PER_PTE; i++, addr += PAGE_SIZE)
  93                 set_pte(&pbase[i], pfn_pte(addr >> PAGE_SHIFT, ref_prot));
  94
  95         /*
  96          * Install the new, split up pagetable:
  97          */
  98         set_pmd_pte(kpte, address, mk_pte(base, ref_prot));
  99         base = NULL;
 100
 101 out_unlock:
 102         up_write(&init_mm.mmap_sem);
 103
 104         if (base)
 105                 __free_pages(base, 0);
 106
 107         return 0;
 108 }
 109
 110 static int __change_page_attr(struct page *page, pgprot_t prot)
 111 {
 112         pgprot_t ref_prot = PAGE_KERNEL;
 113         struct page *kpte_page;
 114         unsigned long address;
 115         int level, err = 0;
 116         pgprot_t oldprot;
 117         pte_t *kpte;
 118
 119         BUG_ON(PageHighMem(page));
 120         address = (unsigned long)page_address(page);
 121
 122 repeat:
 123         kpte = lookup_address(address, &level);
 124         if (!kpte)
 125                 return -EINVAL;
 126
 127         oldprot = pte_pgprot(*kpte);
 128         kpte_page = virt_to_page(kpte);
 129         BUG_ON(PageLRU(kpte_page));
 130         BUG_ON(PageCompound(kpte_page));
 131
 132         /*
 133          * Better fail early if someone sets the kernel text to NX.
 134          * Does not cover __inittext
 135          */
 136         BUG_ON(address >= (unsigned long)&_text &&
 137                 address < (unsigned long)&_etext &&
 138                (pgprot_val(prot) & _PAGE_NX));
 139
 140         if ((address & LARGE_PAGE_MASK) < (unsigned long)&_etext)
 141                 ref_prot = PAGE_KERNEL_EXEC;
 142
 143         ref_prot = canon_pgprot(ref_prot);
 144         prot = canon_pgprot(prot);
 145
 146         if (level == 3) {
 147                 set_pte_atomic(kpte, mk_pte(page, prot));
 148         } else {
 149                 err = split_large_page(kpte, address, ref_prot);
 150                 if (!err)
 151                         goto repeat;
 152         }
 153         return err;
 154 }
 155
 156 /*
 157  * Change the page attributes of an page in the linear mapping.
 158  *
 159  * This should be used when a page is mapped with a different caching policy
 160  * than write-back somewhere - some CPUs do not like it when mappings with
 161  * different caching policies exist. This changes the page attributes of the
 162  * in kernel linear mapping too.
 163  *
 164  * The caller needs to ensure that there are no conflicting mappings elsewhere.
 165  * This function only deals with the kernel linear map.
 166  *
 167  * Caller must call global_flush_tlb() after this.
 168  */
 169 int change_page_attr(struct page *page, int numpages, pgprot_t prot)
 170 {
 171         int err = 0, i;
 172
 173         for (i = 0; i < numpages; i++, page++) {
 174                 err = __change_page_attr(page, prot);
 175                 if (err)
 176                         break;
 177         }
 178
 179         return err;
 180 }
 181 EXPORT_SYMBOL(change_page_attr);
 182
 183 int change_page_attr_addr(unsigned long addr, int numpages, pgprot_t prot)
 184 {
 185         int i;
 186         unsigned long pfn = (__pa(addr) >> PAGE_SHIFT);
 187
 188         for (i = 0; i < numpages; i++) {
 189                 if (!pfn_valid(pfn + i)) {
 190                         WARN_ON_ONCE(1);
 191                         break;
 192                 } else {
 193                         int level;
 194                         pte_t *pte = lookup_address(addr + i*PAGE_SIZE, &level);
 195                         BUG_ON(pte && pte_none(*pte));
 196                 }
 197         }
 198
 199         return change_page_attr(virt_to_page(addr), i, prot);
 200 }
 201
 202 static void flush_kernel_map(void *arg)
 203 {
 204         /*
 205          * Flush all to work around Errata in early athlons regarding
 206          * large page flushing.
 207          */
 208         __flush_tlb_all();
 209
 210         if (boot_cpu_data.x86_model >= 4)
 211                 wbinvd();
 212 }
 213
 214 void global_flush_tlb(void)
 215 {
 216         BUG_ON(irqs_disabled());
 217
 218         on_each_cpu(flush_kernel_map, NULL, 1, 1);
 219 }
 220 EXPORT_SYMBOL(global_flush_tlb);
 221
 222 #ifdef CONFIG_DEBUG_PAGEALLOC
 223 void kernel_map_pages(struct page *page, int numpages, int enable)
 224 {
 225         if (PageHighMem(page))
 226                 return;
 227         if (!enable) {
 228                 debug_check_no_locks_freed(page_address(page),
 229                                            numpages * PAGE_SIZE);
 230         }
 231
 232         /*
 233          * the return value is ignored - the calls cannot fail,
 234          * large pages are disabled at boot time.
 235          */
 236         change_page_attr(page, numpages, enable ? PAGE_KERNEL : __pgprot(0));
 237
 238         /*
 239          * we should perform an IPI and flush all tlbs,
 240          * but that can deadlock->flush only current cpu.
 241          */
 242         __flush_tlb_all();
 243 }
 244 #endif