arch/x86/mm/pageattr_64.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 void clflush_cache_range(void *addr, int size)
  13 {
  14         int i;
  15
  16         for (i = 0; i < size; i += boot_cpu_data.x86_clflush_size)
  17                 clflush(addr+i);
  18 }
  19
  20 #include <asm/processor.h>
  21 #include <asm/tlbflush.h>
  22 #include <asm/sections.h>
  23 #include <asm/uaccess.h>
  24 #include <asm/pgalloc.h>
  25
  26 pte_t *lookup_address(unsigned long address, int *level)
  27 {
  28         pgd_t *pgd = pgd_offset_k(address);
  29         pud_t *pud;
  30         pmd_t *pmd;
  31
  32         if (pgd_none(*pgd))
  33                 return NULL;
  34         pud = pud_offset(pgd, address);
  35         if (pud_none(*pud))
  36                 return NULL;
  37         pmd = pmd_offset(pud, address);
  38         if (pmd_none(*pmd))
  39                 return NULL;
  40         *level = 3;
  41         if (pmd_large(*pmd))
  42                 return (pte_t *)pmd;
  43         *level = 4;
  44
  45         return pte_offset_kernel(pmd, address);
  46 }
  47
  48 static void __set_pmd_pte(pte_t *kpte, unsigned long address, pte_t pte)
  49 {
  50         /* change init_mm */
  51         set_pte_atomic(kpte, pte);
  52 #ifdef CONFIG_X86_32
  53         if (SHARED_KERNEL_PMD)
  54                 return;
  55         {
  56                 struct page *page;
  57
  58                 for (page = pgd_list; page; page = (struct page *)page->index) {
  59                         pgd_t *pgd;
  60                         pud_t *pud;
  61                         pmd_t *pmd;
  62
  63                         pgd = (pgd_t *)page_address(page) + pgd_index(address);
  64                         pud = pud_offset(pgd, address);
  65                         pmd = pmd_offset(pud, address);
  66                         set_pte_atomic((pte_t *)pmd, pte);
  67                 }
  68         }
  69 #endif
  70 }
  71
  72 static int split_large_page(pte_t *kpte, unsigned long address)
  73 {
  74         pgprot_t ref_prot = pte_pgprot(pte_clrhuge(*kpte));
  75         gfp_t gfp_flags = GFP_KERNEL;
  76         unsigned long flags;
  77         unsigned long addr;
  78         pte_t *pbase, *tmp;
  79         struct page *base;
  80         int i, level;
  81
  82
  83 #ifdef CONFIG_DEBUG_PAGEALLOC
  84         gfp_flags = GFP_ATOMIC;
  85 #endif
  86         base = alloc_pages(gfp_flags, 0);
  87         if (!base)
  88                 return -ENOMEM;
  89
  90         spin_lock_irqsave(&pgd_lock, flags);
  91         /*
  92          * Check for races, another CPU might have split this page
  93          * up for us already:
  94          */
  95         tmp = lookup_address(address, &level);
  96         if (tmp != kpte) {
  97                 WARN_ON_ONCE(1);
  98                 goto out_unlock;
  99         }
 100
 101         address = __pa(address);
 102         addr = address & LARGE_PAGE_MASK;
 103         pbase = (pte_t *)page_address(base);
 104 #ifdef CONFIG_X86_32
 105         paravirt_alloc_pt(&init_mm, page_to_pfn(base));
 106 #endif
 107
 108         for (i = 0; i < PTRS_PER_PTE; i++, addr += PAGE_SIZE)
 109                 set_pte(&pbase[i], pfn_pte(addr >> PAGE_SHIFT, ref_prot));
 110
 111         /*
 112          * Install the new, split up pagetable:
 113          */
 114         __set_pmd_pte(kpte, address, mk_pte(base, ref_prot));
 115         base = NULL;
 116
 117 out_unlock:
 118         spin_unlock_irqrestore(&pgd_lock, flags);
 119
 120         if (base)
 121                 __free_pages(base, 0);
 122
 123         return 0;
 124 }
 125
 126 static int
 127 __change_page_attr(unsigned long address, struct page *page, pgprot_t prot)
 128 {
 129         struct page *kpte_page;
 130         int level, err = 0;
 131         pte_t *kpte;
 132
 133         BUG_ON(PageHighMem(page));
 134
 135 repeat:
 136         kpte = lookup_address(address, &level);
 137         if (!kpte)
 138                 return -EINVAL;
 139
 140         kpte_page = virt_to_page(kpte);
 141         BUG_ON(PageLRU(kpte_page));
 142         BUG_ON(PageCompound(kpte_page));
 143
 144         /*
 145          * Better fail early if someone sets the kernel text to NX.
 146          * Does not cover __inittext
 147          */
 148         BUG_ON(address >= (unsigned long)&_text &&
 149                 address < (unsigned long)&_etext &&
 150                (pgprot_val(prot) & _PAGE_NX));
 151
 152         if (level == 4) {
 153                 set_pte_atomic(kpte, mk_pte(page, canon_pgprot(prot)));
 154         } else {
 155                 err = split_large_page(kpte, address);
 156                 if (!err)
 157                         goto repeat;
 158         }
 159         return err;
 160 }
 161
 162 /**
 163  * change_page_attr_addr - Change page table attributes in linear mapping
 164  * @address: Virtual address in linear mapping.
 165  * @numpages: Number of pages to change
 166  * @prot:    New page table attribute (PAGE_*)
 167  *
 168  * Change page attributes of a page in the direct mapping. This is a variant
 169  * of change_page_attr() that also works on memory holes that do not have
 170  * mem_map entry (pfn_valid() is false).
 171  *
 172  * See change_page_attr() documentation for more details.
 173  */
 174
 175 int change_page_attr_addr(unsigned long address, int numpages, pgprot_t prot)
 176 {
 177         int err = 0, kernel_map = 0, i;
 178
 179 #ifdef CONFIG_X86_64
 180         if (address >= __START_KERNEL_map &&
 181                         address < __START_KERNEL_map + KERNEL_TEXT_SIZE) {
 182
 183                 address = (unsigned long)__va(__pa(address));
 184                 kernel_map = 1;
 185         }
 186 #endif
 187
 188         for (i = 0; i < numpages; i++, address += PAGE_SIZE) {
 189                 unsigned long pfn = __pa(address) >> PAGE_SHIFT;
 190
 191                 if (!kernel_map || pte_present(pfn_pte(0, prot))) {
 192                         err = __change_page_attr(address, pfn_to_page(pfn), prot);
 193                         if (err)
 194                                 break;
 195                 }
 196 #ifdef CONFIG_X86_64
 197                 /*
 198                  * Handle kernel mapping too which aliases part of
 199                  * lowmem:
 200                  */
 201                 if (__pa(address) < KERNEL_TEXT_SIZE) {
 202                         unsigned long addr2;
 203                         pgprot_t prot2;
 204
 205                         addr2 = __START_KERNEL_map + __pa(address);
 206                         /* Make sure the kernel mappings stay executable */
 207                         prot2 = pte_pgprot(pte_mkexec(pfn_pte(0, prot)));
 208                         err = __change_page_attr(addr2, pfn_to_page(pfn), prot2);
 209                 }
 210 #endif
 211         }
 212
 213         return err;
 214 }
 215
 216 /**
 217  * change_page_attr - Change page table attributes in the linear mapping.
 218  * @page: First page to change
 219  * @numpages: Number of pages to change
 220  * @prot: New protection/caching type (PAGE_*)
 221  *
 222  * Returns 0 on success, otherwise a negated errno.
 223  *
 224  * This should be used when a page is mapped with a different caching policy
 225  * than write-back somewhere - some CPUs do not like it when mappings with
 226  * different caching policies exist. This changes the page attributes of the
 227  * in kernel linear mapping too.
 228  *
 229  * Caller must call global_flush_tlb() later to make the changes active.
 230  *
 231  * The caller needs to ensure that there are no conflicting mappings elsewhere
 232  * (e.g. in user space) * This function only deals with the kernel linear map.
 233  *
 234  * For MMIO areas without mem_map use change_page_attr_addr() instead.
 235  */
 236 int change_page_attr(struct page *page, int numpages, pgprot_t prot)
 237 {
 238         unsigned long addr = (unsigned long)page_address(page);
 239
 240         return change_page_attr_addr(addr, numpages, prot);
 241 }
 242 EXPORT_SYMBOL(change_page_attr);
 243
 244 static void flush_kernel_map(void *arg)
 245 {
 246         /*
 247          * Flush all to work around Errata in early athlons regarding
 248          * large page flushing.
 249          */
 250         __flush_tlb_all();
 251
 252         if (boot_cpu_data.x86_model >= 4)
 253                 wbinvd();
 254 }
 255
 256 void global_flush_tlb(void)
 257 {
 258         BUG_ON(irqs_disabled());
 259
 260         on_each_cpu(flush_kernel_map, NULL, 1, 1);
 261 }
 262 EXPORT_SYMBOL(global_flush_tlb);
 263
 264 #ifdef CONFIG_DEBUG_PAGEALLOC
 265 void kernel_map_pages(struct page *page, int numpages, int enable)
 266 {
 267         if (PageHighMem(page))
 268                 return;
 269         if (!enable) {
 270                 debug_check_no_locks_freed(page_address(page),
 271                                            numpages * PAGE_SIZE);
 272         }
 273
 274         /*
 275          * If page allocator is not up yet then do not call c_p_a():
 276          */
 277         if (!debug_pagealloc_enabled)
 278                 return;
 279
 280         /*
 281          * the return value is ignored - the calls cannot fail,
 282          * large pages are disabled at boot time.
 283          */
 284         change_page_attr(page, numpages, enable ? PAGE_KERNEL : __pgprot(0));
 285
 286         /*
 287          * we should perform an IPI and flush all tlbs,
 288          * but that can deadlock->flush only current cpu.
 289          */
 290         __flush_tlb_all();
 291 }
 292 #endif