arch/x86/mm/init.c

   1 #include <linux/ioport.h>
   2 #include <linux/swap.h>
   3
   4 #include <asm/cacheflush.h>
   5 #include <asm/e820.h>
   6 #include <asm/init.h>
   7 #include <asm/page.h>
   8 #include <asm/page_types.h>
   9 #include <asm/sections.h>
  10 #include <asm/system.h>
  11 #include <asm/tlbflush.h>
  12
  13 unsigned long __initdata e820_table_start;
  14 unsigned long __meminitdata e820_table_end;
  15 unsigned long __meminitdata e820_table_top;
  16
  17 enum bootmem_state bootmem_state = BEFORE_BOOTMEM;
  18
  19 int direct_gbpages
  20 #ifdef CONFIG_DIRECT_GBPAGES
  21                                 = 1
  22 #endif
  23 ;
  24
  25 static void __init find_early_table_space(unsigned long end, int use_pse,
  26                                           int use_gbpages)
  27 {
  28         unsigned long puds, pmds, ptes, tables, start;
  29
  30         puds = (end + PUD_SIZE - 1) >> PUD_SHIFT;
  31         tables = roundup(puds * sizeof(pud_t), PAGE_SIZE);
  32
  33         if (use_gbpages) {
  34                 unsigned long extra;
  35
  36                 extra = end - ((end>>PUD_SHIFT) << PUD_SHIFT);
  37                 pmds = (extra + PMD_SIZE - 1) >> PMD_SHIFT;
  38         } else
  39                 pmds = (end + PMD_SIZE - 1) >> PMD_SHIFT;
  40
  41         tables += roundup(pmds * sizeof(pmd_t), PAGE_SIZE);
  42
  43         if (use_pse) {
  44                 unsigned long extra;
  45
  46                 extra = end - ((end>>PMD_SHIFT) << PMD_SHIFT);
  47 #ifdef CONFIG_X86_32
  48                 extra += PMD_SIZE;
  49 #endif
  50                 ptes = (extra + PAGE_SIZE - 1) >> PAGE_SHIFT;
  51         } else
  52                 ptes = (end + PAGE_SIZE - 1) >> PAGE_SHIFT;
  53
  54         tables += roundup(ptes * sizeof(pte_t), PAGE_SIZE);
  55
  56 #ifdef CONFIG_X86_32
  57         /* for fixmap */
  58         tables += roundup(__end_of_fixed_addresses * sizeof(pte_t), PAGE_SIZE);
  59 #endif
  60
  61         /*
  62          * RED-PEN putting page tables only on node 0 could
  63          * cause a hotspot and fill up ZONE_DMA. The page tables
  64          * need roughly 0.5KB per GB.
  65          */
  66 #ifdef CONFIG_X86_32
  67         start = 0x7000;
  68         e820_table_start = find_e820_area(start, max_pfn_mapped<<PAGE_SHIFT,
  69                                         tables, PAGE_SIZE);
  70 #else /* CONFIG_X86_64 */
  71         start = 0x8000;
  72         e820_table_start = find_e820_area(start, end, tables, PAGE_SIZE);
  73 #endif
  74         if (e820_table_start == -1UL)
  75                 panic("Cannot find space for the kernel page tables");
  76
  77         e820_table_start >>= PAGE_SHIFT;
  78         e820_table_end = e820_table_start;
  79         e820_table_top = e820_table_start + (tables >> PAGE_SHIFT);
  80
  81         printk(KERN_DEBUG "kernel direct mapping tables up to %lx @ %lx-%lx\n",
  82                 end, e820_table_start << PAGE_SHIFT, e820_table_top << PAGE_SHIFT);
  83 }
  84
  85 struct map_range {
  86         unsigned long start;
  87         unsigned long end;
  88         unsigned page_size_mask;
  89 };
  90
  91 #ifdef CONFIG_X86_32
  92 #define NR_RANGE_MR 3
  93 #else /* CONFIG_X86_64 */
  94 #define NR_RANGE_MR 5
  95 #endif
  96
  97 static int __meminit save_mr(struct map_range *mr, int nr_range,
  98                              unsigned long start_pfn, unsigned long end_pfn,
  99                              unsigned long page_size_mask)
 100 {
 101         if (start_pfn < end_pfn) {
 102                 if (nr_range >= NR_RANGE_MR)
 103                         panic("run out of range for init_memory_mapping\n");
 104                 mr[nr_range].start = start_pfn<<PAGE_SHIFT;
 105                 mr[nr_range].end   = end_pfn<<PAGE_SHIFT;
 106                 mr[nr_range].page_size_mask = page_size_mask;
 107                 nr_range++;
 108         }
 109
 110         return nr_range;
 111 }
 112
 113 #ifdef CONFIG_X86_64
 114 static void __init init_gbpages(void)
 115 {
 116         if (direct_gbpages && cpu_has_gbpages)
 117                 printk(KERN_INFO "Using GB pages for direct mapping\n");
 118         else
 119                 direct_gbpages = 0;
 120 }
 121 #else
 122 static inline void init_gbpages(void)
 123 {
 124 }
 125 #endif
 126
 127 /*
 128  * Setup the direct mapping of the physical memory at PAGE_OFFSET.
 129  * This runs before bootmem is initialized and gets pages directly from
 130  * the physical memory. To access them they are temporarily mapped.
 131  */
 132 unsigned long __init_refok init_memory_mapping(unsigned long start,
 133                                                unsigned long end)
 134 {
 135         unsigned long page_size_mask = 0;
 136         unsigned long start_pfn, end_pfn;
 137         unsigned long ret = 0;
 138         unsigned long pos;
 139
 140         struct map_range mr[NR_RANGE_MR];
 141         int nr_range, i;
 142         int use_pse, use_gbpages;
 143
 144         printk(KERN_INFO "init_memory_mapping: %016lx-%016lx\n", start, end);
 145
 146         if (bootmem_state == BEFORE_BOOTMEM)
 147                 init_gbpages();
 148
 149 #if defined(CONFIG_DEBUG_PAGEALLOC) || defined(CONFIG_KMEMCHECK)
 150         /*
 151          * For CONFIG_DEBUG_PAGEALLOC, identity mapping will use small pages.
 152          * This will simplify cpa(), which otherwise needs to support splitting
 153          * large pages into small in interrupt context, etc.
 154          */
 155         use_pse = use_gbpages = 0;
 156 #else
 157         use_pse = cpu_has_pse;
 158         use_gbpages = direct_gbpages;
 159 #endif
 160
 161 #ifdef CONFIG_X86_32
 162 #ifdef CONFIG_X86_PAE
 163         set_nx();
 164         if (nx_enabled)
 165                 printk(KERN_INFO "NX (Execute Disable) protection: active\n");
 166 #endif
 167
 168         /* Enable PSE if available */
 169         if (cpu_has_pse)
 170                 set_in_cr4(X86_CR4_PSE);
 171
 172         /* Enable PGE if available */
 173         if (cpu_has_pge) {
 174                 set_in_cr4(X86_CR4_PGE);
 175                 __supported_pte_mask |= _PAGE_GLOBAL;
 176         }
 177 #endif
 178
 179         if (use_gbpages)
 180                 page_size_mask |= 1 << PG_LEVEL_1G;
 181         if (use_pse)
 182                 page_size_mask |= 1 << PG_LEVEL_2M;
 183
 184         memset(mr, 0, sizeof(mr));
 185         nr_range = 0;
 186
 187         /* head if not big page alignment ? */
 188         start_pfn = start >> PAGE_SHIFT;
 189         pos = start_pfn << PAGE_SHIFT;
 190 #ifdef CONFIG_X86_32
 191         /*
 192          * Don't use a large page for the first 2/4MB of memory
 193          * because there are often fixed size MTRRs in there
 194          * and overlapping MTRRs into large pages can cause
 195          * slowdowns.
 196          */
 197         if (pos == 0)
 198                 end_pfn = 1<<(PMD_SHIFT - PAGE_SHIFT);
 199         else
 200                 end_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
 201                                  << (PMD_SHIFT - PAGE_SHIFT);
 202 #else /* CONFIG_X86_64 */
 203         end_pfn = ((pos + (PMD_SIZE - 1)) >> PMD_SHIFT)
 204                         << (PMD_SHIFT - PAGE_SHIFT);
 205 #endif
 206         if (end_pfn > (end >> PAGE_SHIFT))
 207                 end_pfn = end >> PAGE_SHIFT;
 208         if (start_pfn < end_pfn) {
 209                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
 210                 pos = end_pfn << PAGE_SHIFT;
 211         }
 212
 213         /* big page (2M) range */
 214         start_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
 215                          << (PMD_SHIFT - PAGE_SHIFT);
 216 #ifdef CONFIG_X86_32
 217         end_pfn = (end>>PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
 218 #else /* CONFIG_X86_64 */
 219         end_pfn = ((pos + (PUD_SIZE - 1))>>PUD_SHIFT)
 220                          << (PUD_SHIFT - PAGE_SHIFT);
 221         if (end_pfn > ((end>>PMD_SHIFT)<<(PMD_SHIFT - PAGE_SHIFT)))
 222                 end_pfn = ((end>>PMD_SHIFT)<<(PMD_SHIFT - PAGE_SHIFT));
 223 #endif
 224
 225         if (start_pfn < end_pfn) {
 226                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 227                                 page_size_mask & (1<<PG_LEVEL_2M));
 228                 pos = end_pfn << PAGE_SHIFT;
 229         }
 230
 231 #ifdef CONFIG_X86_64
 232         /* big page (1G) range */
 233         start_pfn = ((pos + (PUD_SIZE - 1))>>PUD_SHIFT)
 234                          << (PUD_SHIFT - PAGE_SHIFT);
 235         end_pfn = (end >> PUD_SHIFT) << (PUD_SHIFT - PAGE_SHIFT);
 236         if (start_pfn < end_pfn) {
 237                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 238                                 page_size_mask &
 239                                  ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G)));
 240                 pos = end_pfn << PAGE_SHIFT;
 241         }
 242
 243         /* tail is not big page (1G) alignment */
 244         start_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
 245                          << (PMD_SHIFT - PAGE_SHIFT);
 246         end_pfn = (end >> PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
 247         if (start_pfn < end_pfn) {
 248                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 249                                 page_size_mask & (1<<PG_LEVEL_2M));
 250                 pos = end_pfn << PAGE_SHIFT;
 251         }
 252 #endif
 253
 254         /* tail is not big page (2M) alignment */
 255         start_pfn = pos>>PAGE_SHIFT;
 256         end_pfn = end>>PAGE_SHIFT;
 257         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
 258
 259         /* try to merge same page size and continuous */
 260         for (i = 0; nr_range > 1 && i < nr_range - 1; i++) {
 261                 unsigned long old_start;
 262                 if (mr[i].end != mr[i+1].start ||
 263                     mr[i].page_size_mask != mr[i+1].page_size_mask)
 264                         continue;
 265                 /* move it */
 266                 old_start = mr[i].start;
 267                 memmove(&mr[i], &mr[i+1],
 268                         (nr_range - 1 - i) * sizeof(struct map_range));
 269                 mr[i--].start = old_start;
 270                 nr_range--;
 271         }
 272
 273         for (i = 0; i < nr_range; i++)
 274                 printk(KERN_DEBUG " %010lx - %010lx page %s\n",
 275                                 mr[i].start, mr[i].end,
 276                         (mr[i].page_size_mask & (1<<PG_LEVEL_1G))?"1G":(
 277                          (mr[i].page_size_mask & (1<<PG_LEVEL_2M))?"2M":"4k"));
 278
 279         /*
 280          * Find space for the kernel direct mapping tables.
 281          *
 282          * Later we should allocate these tables in the local node of the
 283          * memory mapped. Unfortunately this is done currently before the
 284          * nodes are discovered.
 285          */
 286         if (bootmem_state == BEFORE_BOOTMEM)
 287                 find_early_table_space(end, use_pse, use_gbpages);
 288
 289 #ifdef CONFIG_X86_32
 290         for (i = 0; i < nr_range; i++)
 291                 kernel_physical_mapping_init(mr[i].start, mr[i].end,
 292                                              mr[i].page_size_mask);
 293         ret = end;
 294 #else /* CONFIG_X86_64 */
 295         for (i = 0; i < nr_range; i++)
 296                 ret = kernel_physical_mapping_init(mr[i].start, mr[i].end,
 297                                                    mr[i].page_size_mask);
 298 #endif
 299
 300 #ifdef CONFIG_X86_32
 301         early_ioremap_page_table_range_init();
 302
 303         load_cr3(swapper_pg_dir);
 304 #endif
 305
 306 #ifdef CONFIG_X86_64
 307         if (bootmem_state == BEFORE_BOOTMEM)
 308                 mmu_cr4_features = read_cr4();
 309 #endif
 310         __flush_tlb_all();
 311
 312         if (bootmem_state == BEFORE_BOOTMEM &&
 313             e820_table_end > e820_table_start)
 314                 reserve_early(e820_table_start << PAGE_SHIFT,
 315                                  e820_table_end << PAGE_SHIFT, "PGTABLE");
 316
 317         if (bootmem_state == BEFORE_BOOTMEM)
 318                 early_memtest(start, end);
 319
 320         return ret >> PAGE_SHIFT;
 321 }
 322
 323
 324 /*
 325  * devmem_is_allowed() checks to see if /dev/mem access to a certain address
 326  * is valid. The argument is a physical page number.
 327  *
 328  *
 329  * On x86, access has to be given to the first megabyte of ram because that area
 330  * contains bios code and data regions used by X and dosemu and similar apps.
 331  * Access has to be given to non-kernel-ram areas as well, these contain the PCI
 332  * mmio resources as well as potential bios/acpi data regions.
 333  */
 334 int devmem_is_allowed(unsigned long pagenr)
 335 {
 336         if (pagenr <= 256)
 337                 return 1;
 338         if (iomem_is_exclusive(pagenr << PAGE_SHIFT))
 339                 return 0;
 340         if (!page_is_ram(pagenr))
 341                 return 1;
 342         return 0;
 343 }
 344
 345 void free_init_pages(char *what, unsigned long begin, unsigned long end)
 346 {
 347         unsigned long addr = begin;
 348
 349         if (addr >= end)
 350                 return;
 351
 352         /*
 353          * If debugging page accesses then do not free this memory but
 354          * mark them not present - any buggy init-section access will
 355          * create a kernel page fault:
 356          */
 357 #ifdef CONFIG_DEBUG_PAGEALLOC
 358         printk(KERN_INFO "debug: unmapping init memory %08lx..%08lx\n",
 359                 begin, PAGE_ALIGN(end));
 360         set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
 361 #else
 362         /*
 363          * We just marked the kernel text read only above, now that
 364          * we are going to free part of that, we need to make that
 365          * writeable first.
 366          */
 367         set_memory_rw(begin, (end - begin) >> PAGE_SHIFT);
 368
 369         printk(KERN_INFO "Freeing %s: %luk freed\n", what, (end - begin) >> 10);
 370
 371         for (; addr < end; addr += PAGE_SIZE) {
 372                 ClearPageReserved(virt_to_page(addr));
 373                 init_page_count(virt_to_page(addr));
 374                 memset((void *)(addr & ~(PAGE_SIZE-1)),
 375                         POISON_FREE_INITMEM, PAGE_SIZE);
 376                 free_page(addr);
 377                 totalram_pages++;
 378         }
 379 #endif
 380 }
 381
 382 void free_initmem(void)
 383 {
 384         free_init_pages("unused kernel memory",
 385                         (unsigned long)(&__init_begin),
 386                         (unsigned long)(&__init_end));
 387 }
 388
 389 #ifdef CONFIG_BLK_DEV_INITRD
 390 void free_initrd_mem(unsigned long start, unsigned long end)
 391 {
 392         free_init_pages("initrd memory", start, end);
 393 }
 394 #endif