arch/x86/mm/pageattr_32.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 static DEFINE_SPINLOCK(cpa_lock);
  19 static struct list_head df_list = LIST_HEAD_INIT(df_list);
  20
  21 pte_t *lookup_address(unsigned long address, int *level)
  22 {
  23         pgd_t *pgd = pgd_offset_k(address);
  24         pud_t *pud;
  25         pmd_t *pmd;
  26
  27         if (pgd_none(*pgd))
  28                 return NULL;
  29         pud = pud_offset(pgd, address);
  30         if (pud_none(*pud))
  31                 return NULL;
  32         pmd = pmd_offset(pud, address);
  33         if (pmd_none(*pmd))
  34                 return NULL;
  35         *level = 2;
  36         if (pmd_large(*pmd))
  37                 return (pte_t *)pmd;
  38         *level = 3;
  39
  40         return pte_offset_kernel(pmd, address);
  41 }
  42
  43 static struct page *
  44 split_large_page(unsigned long address, pgprot_t prot, pgprot_t ref_prot)
  45 {
  46         unsigned long addr;
  47         struct page *base;
  48         pte_t *pbase;
  49         int i;
  50
  51         spin_unlock_irq(&cpa_lock);
  52         base = alloc_pages(GFP_KERNEL, 0);
  53         spin_lock_irq(&cpa_lock);
  54         if (!base)
  55                 return NULL;
  56
  57         /*
  58          * page_private is used to track the number of entries in
  59          * the page table page that have non standard attributes.
  60          */
  61         SetPagePrivate(base);
  62         page_private(base) = 0;
  63
  64         address = __pa(address);
  65         addr = address & LARGE_PAGE_MASK;
  66         pbase = (pte_t *)page_address(base);
  67         paravirt_alloc_pt(&init_mm, page_to_pfn(base));
  68
  69         for (i = 0; i < PTRS_PER_PTE; i++, addr += PAGE_SIZE) {
  70                 set_pte(&pbase[i], pfn_pte(addr >> PAGE_SHIFT,
  71                                            addr == address ? prot : ref_prot));
  72         }
  73         return base;
  74 }
  75
  76 static void cache_flush_page(struct page *p)
  77 {
  78         void *addr = page_address(p);
  79         int i;
  80
  81         for (i = 0; i < PAGE_SIZE; i += boot_cpu_data.x86_clflush_size)
  82                 clflush(addr + i);
  83 }
  84
  85 static void flush_kernel_map(void *arg)
  86 {
  87         struct list_head *lh = (struct list_head *)arg;
  88         struct page *p;
  89
  90         /*
  91          * Flush all to work around Errata in early athlons regarding
  92          * large page flushing.
  93          */
  94         __flush_tlb_all();
  95
  96         /* High level code is not ready for clflush yet */
  97         if (0 && cpu_has_clflush) {
  98                 list_for_each_entry(p, lh, lru)
  99                         cache_flush_page(p);
 100         } else {
 101                 if (boot_cpu_data.x86_model >= 4)
 102                         wbinvd();
 103         }
 104 }
 105
 106 static void set_pmd_pte(pte_t *kpte, unsigned long address, pte_t pte)
 107 {
 108         unsigned long flags;
 109         struct page *page;
 110
 111         /* change init_mm */
 112         set_pte_atomic(kpte, pte);
 113         if (SHARED_KERNEL_PMD)
 114                 return;
 115
 116         spin_lock_irqsave(&pgd_lock, flags);
 117         for (page = pgd_list; page; page = (struct page *)page->index) {
 118                 pgd_t *pgd;
 119                 pud_t *pud;
 120                 pmd_t *pmd;
 121
 122                 pgd = (pgd_t *)page_address(page) + pgd_index(address);
 123                 pud = pud_offset(pgd, address);
 124                 pmd = pmd_offset(pud, address);
 125                 set_pte_atomic((pte_t *)pmd, pte);
 126         }
 127         spin_unlock_irqrestore(&pgd_lock, flags);
 128 }
 129
 130 /*
 131  * No more special protections in this 2/4MB area - revert to a large
 132  * page again.
 133  */
 134 static inline void revert_page(struct page *kpte_page, unsigned long address)
 135 {
 136         pgprot_t ref_prot;
 137         pte_t *linear;
 138
 139         ref_prot =
 140         ((address & LARGE_PAGE_MASK) < (unsigned long)&_etext)
 141                 ? PAGE_KERNEL_LARGE_EXEC : PAGE_KERNEL_LARGE;
 142
 143         linear = (pte_t *)
 144                 pmd_offset(pud_offset(pgd_offset_k(address), address), address);
 145         set_pmd_pte(linear,  address,
 146                     pfn_pte((__pa(address) & LARGE_PAGE_MASK) >> PAGE_SHIFT,
 147                             ref_prot));
 148 }
 149
 150 static inline void save_page(struct page *kpte_page)
 151 {
 152         if (!test_and_set_bit(PG_arch_1, &kpte_page->flags))
 153                 list_add(&kpte_page->lru, &df_list);
 154 }
 155
 156 static int __change_page_attr(struct page *page, pgprot_t prot)
 157 {
 158         struct page *kpte_page;
 159         unsigned long address;
 160         pte_t *kpte;
 161         int level;
 162
 163         BUG_ON(PageHighMem(page));
 164         address = (unsigned long)page_address(page);
 165
 166         kpte = lookup_address(address, &level);
 167         if (!kpte)
 168                 return -EINVAL;
 169
 170         kpte_page = virt_to_page(kpte);
 171         BUG_ON(PageLRU(kpte_page));
 172         BUG_ON(PageCompound(kpte_page));
 173
 174         if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL)) {
 175                 if (level == 3) {
 176                         set_pte_atomic(kpte, mk_pte(page, prot));
 177                 } else {
 178                         struct page *split;
 179                         pgprot_t ref_prot;
 180
 181                         ref_prot =
 182                         ((address & LARGE_PAGE_MASK) < (unsigned long)&_etext)
 183                                 ? PAGE_KERNEL_EXEC : PAGE_KERNEL;
 184                         split = split_large_page(address, prot, ref_prot);
 185                         if (!split)
 186                                 return -ENOMEM;
 187
 188                         set_pmd_pte(kpte, address, mk_pte(split, ref_prot));
 189                         kpte_page = split;
 190                 }
 191                 page_private(kpte_page)++;
 192         } else {
 193                 if (level == 3) {
 194                         set_pte_atomic(kpte, mk_pte(page, PAGE_KERNEL));
 195                         BUG_ON(page_private(kpte_page) == 0);
 196                         page_private(kpte_page)--;
 197                 } else
 198                         BUG();
 199         }
 200
 201         /*
 202          * If the pte was reserved, it means it was created at boot
 203          * time (not via split_large_page) and in turn we must not
 204          * replace it with a largepage.
 205          */
 206
 207         save_page(kpte_page);
 208         if (!PageReserved(kpte_page)) {
 209                 if (cpu_has_pse && (page_private(kpte_page) == 0)) {
 210                         paravirt_release_pt(page_to_pfn(kpte_page));
 211                         revert_page(kpte_page, address);
 212                 }
 213         }
 214         return 0;
 215 }
 216
 217 static inline void flush_map(struct list_head *l)
 218 {
 219         on_each_cpu(flush_kernel_map, l, 1, 1);
 220 }
 221
 222 /*
 223  * Change the page attributes of an page in the linear mapping.
 224  *
 225  * This should be used when a page is mapped with a different caching policy
 226  * than write-back somewhere - some CPUs do not like it when mappings with
 227  * different caching policies exist. This changes the page attributes of the
 228  * in kernel linear mapping too.
 229  *
 230  * The caller needs to ensure that there are no conflicting mappings elsewhere.
 231  * This function only deals with the kernel linear map.
 232  *
 233  * Caller must call global_flush_tlb() after this.
 234  */
 235 int change_page_attr(struct page *page, int numpages, pgprot_t prot)
 236 {
 237         unsigned long flags;
 238         int err = 0, i;
 239
 240         spin_lock_irqsave(&cpa_lock, flags);
 241         for (i = 0; i < numpages; i++, page++) {
 242                 err = __change_page_attr(page, prot);
 243                 if (err)
 244                         break;
 245         }
 246         spin_unlock_irqrestore(&cpa_lock, flags);
 247
 248         return err;
 249 }
 250 EXPORT_SYMBOL(change_page_attr);
 251
 252 void global_flush_tlb(void)
 253 {
 254         struct page *pg, *next;
 255         struct list_head l;
 256
 257         BUG_ON(irqs_disabled());
 258
 259         spin_lock_irq(&cpa_lock);
 260         list_replace_init(&df_list, &l);
 261         spin_unlock_irq(&cpa_lock);
 262         flush_map(&l);
 263         list_for_each_entry_safe(pg, next, &l, lru) {
 264                 list_del(&pg->lru);
 265                 clear_bit(PG_arch_1, &pg->flags);
 266                 if (PageReserved(pg) || !cpu_has_pse || page_private(pg) != 0)
 267                         continue;
 268                 ClearPagePrivate(pg);
 269                 __free_page(pg);
 270         }
 271 }
 272 EXPORT_SYMBOL(global_flush_tlb);
 273
 274 #ifdef CONFIG_DEBUG_PAGEALLOC
 275 void kernel_map_pages(struct page *page, int numpages, int enable)
 276 {
 277         if (PageHighMem(page))
 278                 return;
 279         if (!enable) {
 280                 debug_check_no_locks_freed(page_address(page),
 281                                            numpages * PAGE_SIZE);
 282         }
 283
 284         /*
 285          * the return value is ignored - the calls cannot fail,
 286          * large pages are disabled at boot time.
 287          */
 288         change_page_attr(page, numpages, enable ? PAGE_KERNEL : __pgprot(0));
 289
 290         /*
 291          * we should perform an IPI and flush all tlbs,
 292          * but that can deadlock->flush only current cpu.
 293          */
 294         __flush_tlb_all();
 295 }
 296 #endif