arch/i386/mm/pageattr.c

   1 /*
   2  * Copyright 2002 Andi Kleen, SuSE Labs.
   3  * Thanks to Ben LaHaise for precious feedback.
   4  */
   5
   6 #include <linux/mm.h>
   7 #include <linux/sched.h>
   8 #include <linux/highmem.h>
   9 #include <linux/module.h>
  10 #include <linux/slab.h>
  11 #include <asm/uaccess.h>
  12 #include <asm/processor.h>
  13 #include <asm/tlbflush.h>
  14 #include <asm/pgalloc.h>
  15 #include <asm/sections.h>
  16
  17 static DEFINE_SPINLOCK(cpa_lock);
  18 static struct list_head df_list = LIST_HEAD_INIT(df_list);
  19
  20
  21 pte_t *lookup_address(unsigned long address)
  22 {
  23         pgd_t *pgd = pgd_offset_k(address);
  24         pud_t *pud;
  25         pmd_t *pmd;
  26         if (pgd_none(*pgd))
  27                 return NULL;
  28         pud = pud_offset(pgd, address);
  29         if (pud_none(*pud))
  30                 return NULL;
  31         pmd = pmd_offset(pud, address);
  32         if (pmd_none(*pmd))
  33                 return NULL;
  34         if (pmd_large(*pmd))
  35                 return (pte_t *)pmd;
  36         return pte_offset_kernel(pmd, address);
  37 }
  38
  39 static struct page *split_large_page(unsigned long address, pgprot_t prot,
  40                                         pgprot_t ref_prot)
  41 {
  42         int i;
  43         unsigned long addr;
  44         struct page *base;
  45         pte_t *pbase;
  46
  47         spin_unlock_irq(&cpa_lock);
  48         base = alloc_pages(GFP_KERNEL, 0);
  49         spin_lock_irq(&cpa_lock);
  50         if (!base)
  51                 return NULL;
  52
  53         /*
  54          * page_private is used to track the number of entries in
  55          * the page table page that have non standard attributes.
  56          */
  57         SetPagePrivate(base);
  58         page_private(base) = 0;
  59
  60         address = __pa(address);
  61         addr = address & LARGE_PAGE_MASK;
  62         pbase = (pte_t *)page_address(base);
  63         paravirt_alloc_pt(&init_mm, page_to_pfn(base));
  64         for (i = 0; i < PTRS_PER_PTE; i++, addr += PAGE_SIZE) {
  65                set_pte(&pbase[i], pfn_pte(addr >> PAGE_SHIFT,
  66                                           addr == address ? prot : ref_prot));
  67         }
  68         return base;
  69 }
  70
  71 static void cache_flush_page(struct page *p)
  72 {
  73         unsigned long adr = (unsigned long)page_address(p);
  74         int i;
  75         for (i = 0; i < PAGE_SIZE; i += boot_cpu_data.x86_clflush_size)
  76                 asm volatile("clflush (%0)" :: "r" (adr + i));
  77 }
  78
  79 static void flush_kernel_map(void *arg)
  80 {
  81         struct list_head *lh = (struct list_head *)arg;
  82         struct page *p;
  83
  84         /* High level code is not ready for clflush yet */
  85         if (0 && cpu_has_clflush) {
  86                 list_for_each_entry (p, lh, lru)
  87                         cache_flush_page(p);
  88         } else if (boot_cpu_data.x86_model >= 4)
  89                 wbinvd();
  90
  91         /* Flush all to work around Errata in early athlons regarding
  92          * large page flushing.
  93          */
  94         __flush_tlb_all();
  95 }
  96
  97 static void set_pmd_pte(pte_t *kpte, unsigned long address, pte_t pte)
  98 {
  99         struct page *page;
 100         unsigned long flags;
 101
 102         set_pte_atomic(kpte, pte);      /* change init_mm */
 103         if (SHARED_KERNEL_PMD)
 104                 return;
 105
 106         spin_lock_irqsave(&pgd_lock, flags);
 107         for (page = pgd_list; page; page = (struct page *)page->index) {
 108                 pgd_t *pgd;
 109                 pud_t *pud;
 110                 pmd_t *pmd;
 111                 pgd = (pgd_t *)page_address(page) + pgd_index(address);
 112                 pud = pud_offset(pgd, address);
 113                 pmd = pmd_offset(pud, address);
 114                 set_pte_atomic((pte_t *)pmd, pte);
 115         }
 116         spin_unlock_irqrestore(&pgd_lock, flags);
 117 }
 118
 119 /*
 120  * No more special protections in this 2/4MB area - revert to a
 121  * large page again.
 122  */
 123 static inline void revert_page(struct page *kpte_page, unsigned long address)
 124 {
 125         pgprot_t ref_prot;
 126         pte_t *linear;
 127
 128         ref_prot =
 129         ((address & LARGE_PAGE_MASK) < (unsigned long)&_etext)
 130                 ? PAGE_KERNEL_LARGE_EXEC : PAGE_KERNEL_LARGE;
 131
 132         linear = (pte_t *)
 133                 pmd_offset(pud_offset(pgd_offset_k(address), address), address);
 134         set_pmd_pte(linear,  address,
 135                     pfn_pte((__pa(address) & LARGE_PAGE_MASK) >> PAGE_SHIFT,
 136                             ref_prot));
 137 }
 138
 139 static inline void save_page(struct page *kpte_page)
 140 {
 141         if (!test_and_set_bit(PG_arch_1, &kpte_page->flags))
 142                 list_add(&kpte_page->lru, &df_list);
 143 }
 144
 145 static int
 146 __change_page_attr(struct page *page, pgprot_t prot)
 147 {
 148         pte_t *kpte;
 149         unsigned long address;
 150         struct page *kpte_page;
 151
 152         BUG_ON(PageHighMem(page));
 153         address = (unsigned long)page_address(page);
 154
 155         kpte = lookup_address(address);
 156         if (!kpte)
 157                 return -EINVAL;
 158         kpte_page = virt_to_page(kpte);
 159         BUG_ON(PageLRU(kpte_page));
 160         BUG_ON(PageCompound(kpte_page));
 161
 162         if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL)) {
 163                 if (!pte_huge(*kpte)) {
 164                         set_pte_atomic(kpte, mk_pte(page, prot));
 165                 } else {
 166                         pgprot_t ref_prot;
 167                         struct page *split;
 168
 169                         ref_prot =
 170                         ((address & LARGE_PAGE_MASK) < (unsigned long)&_etext)
 171                                 ? PAGE_KERNEL_EXEC : PAGE_KERNEL;
 172                         split = split_large_page(address, prot, ref_prot);
 173                         if (!split)
 174                                 return -ENOMEM;
 175                         set_pmd_pte(kpte,address,mk_pte(split, ref_prot));
 176                         kpte_page = split;
 177                 }
 178                 page_private(kpte_page)++;
 179         } else if (!pte_huge(*kpte)) {
 180                 set_pte_atomic(kpte, mk_pte(page, PAGE_KERNEL));
 181                 BUG_ON(page_private(kpte_page) == 0);
 182                 page_private(kpte_page)--;
 183         } else
 184                 BUG();
 185
 186         /*
 187          * If the pte was reserved, it means it was created at boot
 188          * time (not via split_large_page) and in turn we must not
 189          * replace it with a largepage.
 190          */
 191
 192         save_page(kpte_page);
 193         if (!PageReserved(kpte_page)) {
 194                 if (cpu_has_pse && (page_private(kpte_page) == 0)) {
 195                         paravirt_release_pt(page_to_pfn(kpte_page));
 196                         revert_page(kpte_page, address);
 197                 }
 198         }
 199         return 0;
 200 }
 201
 202 static inline void flush_map(struct list_head *l)
 203 {
 204         on_each_cpu(flush_kernel_map, l, 1, 1);
 205 }
 206
 207 /*
 208  * Change the page attributes of an page in the linear mapping.
 209  *
 210  * This should be used when a page is mapped with a different caching policy
 211  * than write-back somewhere - some CPUs do not like it when mappings with
 212  * different caching policies exist. This changes the page attributes of the
 213  * in kernel linear mapping too.
 214  *
 215  * The caller needs to ensure that there are no conflicting mappings elsewhere.
 216  * This function only deals with the kernel linear map.
 217  *
 218  * Caller must call global_flush_tlb() after this.
 219  */
 220 int change_page_attr(struct page *page, int numpages, pgprot_t prot)
 221 {
 222         int err = 0;
 223         int i;
 224         unsigned long flags;
 225
 226         spin_lock_irqsave(&cpa_lock, flags);
 227         for (i = 0; i < numpages; i++, page++) {
 228                 err = __change_page_attr(page, prot);
 229                 if (err)
 230                         break;
 231         }
 232         spin_unlock_irqrestore(&cpa_lock, flags);
 233         return err;
 234 }
 235
 236 void global_flush_tlb(void)
 237 {
 238         struct list_head l;
 239         struct page *pg, *next;
 240
 241         BUG_ON(irqs_disabled());
 242
 243         spin_lock_irq(&cpa_lock);
 244         list_replace_init(&df_list, &l);
 245         spin_unlock_irq(&cpa_lock);
 246         flush_map(&l);
 247         list_for_each_entry_safe(pg, next, &l, lru) {
 248                 list_del(&pg->lru);
 249                 clear_bit(PG_arch_1, &pg->flags);
 250                 if (PageReserved(pg) || !cpu_has_pse || page_private(pg) != 0)
 251                         continue;
 252                 ClearPagePrivate(pg);
 253                 __free_page(pg);
 254         }
 255 }
 256
 257 #ifdef CONFIG_DEBUG_PAGEALLOC
 258 void kernel_map_pages(struct page *page, int numpages, int enable)
 259 {
 260         if (PageHighMem(page))
 261                 return;
 262         if (!enable)
 263                 debug_check_no_locks_freed(page_address(page),
 264                                            numpages * PAGE_SIZE);
 265
 266         /* the return value is ignored - the calls cannot fail,
 267          * large pages are disabled at boot time.
 268          */
 269         change_page_attr(page, numpages, enable ? PAGE_KERNEL : __pgprot(0));
 270         /* we should perform an IPI and flush all tlbs,
 271          * but that can deadlock->flush only current cpu.
 272          */
 273         __flush_tlb_all();
 274 }
 275 #endif
 276
 277 EXPORT_SYMBOL(change_page_attr);
 278 EXPORT_SYMBOL(global_flush_tlb);