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