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 split_large_page(pte_t *kpte, unsigned long address)
  65 {
  66         pgprot_t ref_prot = pte_pgprot(pte_clrhuge(*kpte));
  67         unsigned long addr;
  68         pte_t *pbase, *tmp;
  69         struct page *base;
  70         int i, level;
  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         struct page *kpte_page;
 113         unsigned long address;
 114         int level, err = 0;
 115         pte_t *kpte;
 116
 117         BUG_ON(PageHighMem(page));
 118         address = (unsigned long)page_address(page);
 119
 120 repeat:
 121         kpte = lookup_address(address, &level);
 122         if (!kpte)
 123                 return -EINVAL;
 124
 125         kpte_page = virt_to_page(kpte);
 126         BUG_ON(PageLRU(kpte_page));
 127         BUG_ON(PageCompound(kpte_page));
 128
 129         /*
 130          * Better fail early if someone sets the kernel text to NX.
 131          * Does not cover __inittext
 132          */
 133         BUG_ON(address >= (unsigned long)&_text &&
 134                 address < (unsigned long)&_etext &&
 135                (pgprot_val(prot) & _PAGE_NX));
 136
 137         if (level == 3) {
 138                 set_pte_atomic(kpte, mk_pte(page, canon_pgprot(prot)));
 139         } else {
 140                 err = split_large_page(kpte, address);
 141                 if (!err)
 142                         goto repeat;
 143         }
 144         return err;
 145 }
 146
 147 /*
 148  * Change the page attributes of an page in the linear mapping.
 149  *
 150  * This should be used when a page is mapped with a different caching policy
 151  * than write-back somewhere - some CPUs do not like it when mappings with
 152  * different caching policies exist. This changes the page attributes of the
 153  * in kernel linear mapping too.
 154  *
 155  * The caller needs to ensure that there are no conflicting mappings elsewhere.
 156  * This function only deals with the kernel linear map.
 157  *
 158  * Caller must call global_flush_tlb() after this.
 159  */
 160 int change_page_attr(struct page *page, int numpages, pgprot_t prot)
 161 {
 162         int err = 0, i;
 163
 164         for (i = 0; i < numpages; i++, page++) {
 165                 err = __change_page_attr(page, prot);
 166                 if (err)
 167                         break;
 168         }
 169
 170         return err;
 171 }
 172 EXPORT_SYMBOL(change_page_attr);
 173
 174 int change_page_attr_addr(unsigned long addr, int numpages, pgprot_t prot)
 175 {
 176         int i;
 177         unsigned long pfn = (__pa(addr) >> PAGE_SHIFT);
 178
 179         for (i = 0; i < numpages; i++) {
 180                 if (!pfn_valid(pfn + i)) {
 181                         WARN_ON_ONCE(1);
 182                         break;
 183                 } else {
 184                         int level;
 185                         pte_t *pte = lookup_address(addr + i*PAGE_SIZE, &level);
 186                         BUG_ON(pte && pte_none(*pte));
 187                 }
 188         }
 189
 190         return change_page_attr(virt_to_page(addr), i, prot);
 191 }
 192
 193 static void flush_kernel_map(void *arg)
 194 {
 195         /*
 196          * Flush all to work around Errata in early athlons regarding
 197          * large page flushing.
 198          */
 199         __flush_tlb_all();
 200
 201         if (boot_cpu_data.x86_model >= 4)
 202                 wbinvd();
 203 }
 204
 205 void global_flush_tlb(void)
 206 {
 207         BUG_ON(irqs_disabled());
 208
 209         on_each_cpu(flush_kernel_map, NULL, 1, 1);
 210 }
 211 EXPORT_SYMBOL(global_flush_tlb);
 212
 213 #ifdef CONFIG_DEBUG_PAGEALLOC
 214 void kernel_map_pages(struct page *page, int numpages, int enable)
 215 {
 216         if (PageHighMem(page))
 217                 return;
 218         if (!enable) {
 219                 debug_check_no_locks_freed(page_address(page),
 220                                            numpages * PAGE_SIZE);
 221         }
 222
 223         /*
 224          * the return value is ignored - the calls cannot fail,
 225          * large pages are disabled at boot time.
 226          */
 227         change_page_attr(page, numpages, enable ? PAGE_KERNEL : __pgprot(0));
 228
 229         /*
 230          * we should perform an IPI and flush all tlbs,
 231          * but that can deadlock->flush only current cpu.
 232          */
 233         __flush_tlb_all();
 234 }
 235 #endif