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