arch/x86/mm/pageattr.c

   1 /*
   2  * Copyright 2002 Andi Kleen, SuSE Labs.
   3  * Thanks to Ben LaHaise for precious feedback.
   4  */
   5 #include <linux/highmem.h>
   6 #include <linux/module.h>
   7 #include <linux/sched.h>
   8 #include <linux/slab.h>
   9 #include <linux/mm.h>
  10
  11 void clflush_cache_range(void *addr, int size)
  12 {
  13         int i;
  14
  15         for (i = 0; i < size; i += boot_cpu_data.x86_clflush_size)
  16                 clflush(addr+i);
  17 }
  18
  19 #include <asm/processor.h>
  20 #include <asm/tlbflush.h>
  21 #include <asm/sections.h>
  22 #include <asm/uaccess.h>
  23 #include <asm/pgalloc.h>
  24
  25 /*
  26  * We allow the BIOS range to be executable:
  27  */
  28 #define BIOS_BEGIN              0x000a0000
  29 #define BIOS_END                0x00100000
  30
  31 static inline pgprot_t check_exec(pgprot_t prot, unsigned long address)
  32 {
  33         if (__pa(address) >= BIOS_BEGIN && __pa(address) < BIOS_END)
  34                 pgprot_val(prot) &= ~_PAGE_NX;
  35         /*
  36          * Better fail early if someone sets the kernel text to NX.
  37          * Does not cover __inittext
  38          */
  39         BUG_ON(address >= (unsigned long)&_text &&
  40                 address < (unsigned long)&_etext &&
  41                (pgprot_val(prot) & _PAGE_NX));
  42
  43         return prot;
  44 }
  45
  46 pte_t *lookup_address(unsigned long address, int *level)
  47 {
  48         pgd_t *pgd = pgd_offset_k(address);
  49         pud_t *pud;
  50         pmd_t *pmd;
  51
  52         *level = PG_LEVEL_NONE;
  53
  54         if (pgd_none(*pgd))
  55                 return NULL;
  56         pud = pud_offset(pgd, address);
  57         if (pud_none(*pud))
  58                 return NULL;
  59         pmd = pmd_offset(pud, address);
  60         if (pmd_none(*pmd))
  61                 return NULL;
  62
  63         *level = PG_LEVEL_2M;
  64         if (pmd_large(*pmd))
  65                 return (pte_t *)pmd;
  66
  67         *level = PG_LEVEL_4K;
  68         return pte_offset_kernel(pmd, address);
  69 }
  70
  71 static void __set_pmd_pte(pte_t *kpte, unsigned long address, pte_t pte)
  72 {
  73         /* change init_mm */
  74         set_pte_atomic(kpte, pte);
  75 #ifdef CONFIG_X86_32
  76         if (!SHARED_KERNEL_PMD) {
  77                 struct page *page;
  78
  79                 for (page = pgd_list; page; page = (struct page *)page->index) {
  80                         pgd_t *pgd;
  81                         pud_t *pud;
  82                         pmd_t *pmd;
  83
  84                         pgd = (pgd_t *)page_address(page) + pgd_index(address);
  85                         pud = pud_offset(pgd, address);
  86                         pmd = pmd_offset(pud, address);
  87                         set_pte_atomic((pte_t *)pmd, pte);
  88                 }
  89         }
  90 #endif
  91 }
  92
  93 static int split_large_page(pte_t *kpte, unsigned long address)
  94 {
  95         pgprot_t ref_prot = pte_pgprot(pte_clrhuge(*kpte));
  96         gfp_t gfp_flags = GFP_KERNEL;
  97         unsigned long flags;
  98         unsigned long addr;
  99         pte_t *pbase, *tmp;
 100         struct page *base;
 101         int i, level;
 102
 103 #ifdef CONFIG_DEBUG_PAGEALLOC
 104         gfp_flags = GFP_ATOMIC;
 105 #endif
 106         base = alloc_pages(gfp_flags, 0);
 107         if (!base)
 108                 return -ENOMEM;
 109
 110         spin_lock_irqsave(&pgd_lock, flags);
 111         /*
 112          * Check for races, another CPU might have split this page
 113          * up for us already:
 114          */
 115         tmp = lookup_address(address, &level);
 116         if (tmp != kpte) {
 117                 WARN_ON_ONCE(1);
 118                 goto out_unlock;
 119         }
 120
 121         address = __pa(address);
 122         addr = address & LARGE_PAGE_MASK;
 123         pbase = (pte_t *)page_address(base);
 124 #ifdef CONFIG_X86_32
 125         paravirt_alloc_pt(&init_mm, page_to_pfn(base));
 126 #endif
 127
 128         for (i = 0; i < PTRS_PER_PTE; i++, addr += PAGE_SIZE)
 129                 set_pte(&pbase[i], pfn_pte(addr >> PAGE_SHIFT, ref_prot));
 130
 131         /*
 132          * Install the new, split up pagetable:
 133          */
 134         __set_pmd_pte(kpte, address, mk_pte(base, ref_prot));
 135         base = NULL;
 136
 137 out_unlock:
 138         spin_unlock_irqrestore(&pgd_lock, flags);
 139
 140         if (base)
 141                 __free_pages(base, 0);
 142
 143         return 0;
 144 }
 145
 146 static int
 147 __change_page_attr(unsigned long address, struct page *page, pgprot_t prot)
 148 {
 149         struct page *kpte_page;
 150         int level, err = 0;
 151         pte_t *kpte;
 152
 153         BUG_ON(PageHighMem(page));
 154
 155 repeat:
 156         kpte = lookup_address(address, &level);
 157         if (!kpte)
 158                 return -EINVAL;
 159
 160         kpte_page = virt_to_page(kpte);
 161         BUG_ON(PageLRU(kpte_page));
 162         BUG_ON(PageCompound(kpte_page));
 163
 164         prot = check_exec(prot, address);
 165
 166         if (level == PG_LEVEL_4K) {
 167                 set_pte_atomic(kpte, mk_pte(page, canon_pgprot(prot)));
 168         } else {
 169                 err = split_large_page(kpte, address);
 170                 if (!err)
 171                         goto repeat;
 172         }
 173         return err;
 174 }
 175
 176 /**
 177  * change_page_attr_addr - Change page table attributes in linear mapping
 178  * @address: Virtual address in linear mapping.
 179  * @numpages: Number of pages to change
 180  * @prot:    New page table attribute (PAGE_*)
 181  *
 182  * Change page attributes of a page in the direct mapping. This is a variant
 183  * of change_page_attr() that also works on memory holes that do not have
 184  * mem_map entry (pfn_valid() is false).
 185  *
 186  * See change_page_attr() documentation for more details.
 187  */
 188
 189 int change_page_attr_addr(unsigned long address, int numpages, pgprot_t prot)
 190 {
 191         int err = 0, kernel_map = 0, i;
 192
 193 #ifdef CONFIG_X86_64
 194         if (address >= __START_KERNEL_map &&
 195                         address < __START_KERNEL_map + KERNEL_TEXT_SIZE) {
 196
 197                 address = (unsigned long)__va(__pa(address));
 198                 kernel_map = 1;
 199         }
 200 #endif
 201
 202         for (i = 0; i < numpages; i++, address += PAGE_SIZE) {
 203                 unsigned long pfn = __pa(address) >> PAGE_SHIFT;
 204
 205                 if (!kernel_map || pte_present(pfn_pte(0, prot))) {
 206                         err = __change_page_attr(address, pfn_to_page(pfn), prot);
 207                         if (err)
 208                                 break;
 209                 }
 210 #ifdef CONFIG_X86_64
 211                 /*
 212                  * Handle kernel mapping too which aliases part of
 213                  * lowmem:
 214                  */
 215                 if (__pa(address) < KERNEL_TEXT_SIZE) {
 216                         unsigned long addr2;
 217                         pgprot_t prot2;
 218
 219                         addr2 = __START_KERNEL_map + __pa(address);
 220                         /* Make sure the kernel mappings stay executable */
 221                         prot2 = pte_pgprot(pte_mkexec(pfn_pte(0, prot)));
 222                         err = __change_page_attr(addr2, pfn_to_page(pfn), prot2);
 223                 }
 224 #endif
 225         }
 226
 227         return err;
 228 }
 229
 230 /**
 231  * change_page_attr - Change page table attributes in the linear mapping.
 232  * @page: First page to change
 233  * @numpages: Number of pages to change
 234  * @prot: New protection/caching type (PAGE_*)
 235  *
 236  * Returns 0 on success, otherwise a negated errno.
 237  *
 238  * This should be used when a page is mapped with a different caching policy
 239  * than write-back somewhere - some CPUs do not like it when mappings with
 240  * different caching policies exist. This changes the page attributes of the
 241  * in kernel linear mapping too.
 242  *
 243  * Caller must call global_flush_tlb() later to make the changes active.
 244  *
 245  * The caller needs to ensure that there are no conflicting mappings elsewhere
 246  * (e.g. in user space) * This function only deals with the kernel linear map.
 247  *
 248  * For MMIO areas without mem_map use change_page_attr_addr() instead.
 249  */
 250 int change_page_attr(struct page *page, int numpages, pgprot_t prot)
 251 {
 252         unsigned long addr = (unsigned long)page_address(page);
 253
 254         return change_page_attr_addr(addr, numpages, prot);
 255 }
 256 EXPORT_SYMBOL(change_page_attr);
 257
 258 static void flush_kernel_map(void *arg)
 259 {
 260         /*
 261          * Flush all to work around Errata in early athlons regarding
 262          * large page flushing.
 263          */
 264         __flush_tlb_all();
 265
 266         if (boot_cpu_data.x86_model >= 4)
 267                 wbinvd();
 268 }
 269
 270 void global_flush_tlb(void)
 271 {
 272         BUG_ON(irqs_disabled());
 273
 274         on_each_cpu(flush_kernel_map, NULL, 1, 1);
 275 }
 276 EXPORT_SYMBOL(global_flush_tlb);
 277
 278 #ifdef CONFIG_DEBUG_PAGEALLOC
 279 void kernel_map_pages(struct page *page, int numpages, int enable)
 280 {
 281         if (PageHighMem(page))
 282                 return;
 283         if (!enable) {
 284                 debug_check_no_locks_freed(page_address(page),
 285                                            numpages * PAGE_SIZE);
 286         }
 287
 288         /*
 289          * If page allocator is not up yet then do not call c_p_a():
 290          */
 291         if (!debug_pagealloc_enabled)
 292                 return;
 293
 294         /*
 295          * The return value is ignored - the calls cannot fail,
 296          * large pages are disabled at boot time:
 297          */
 298         change_page_attr(page, numpages, enable ? PAGE_KERNEL : __pgprot(0));
 299
 300         /*
 301          * We should perform an IPI and flush all tlbs,
 302          * but that can deadlock->flush only current cpu:
 303          */
 304         __flush_tlb_all();
 305 }
 306 #endif