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