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/setup.h>
  11 #include <asm/system.h>
  12 #include <asm/tlbflush.h>
  13
  14 unsigned long __initdata e820_table_start;
  15 unsigned long __meminitdata e820_table_end;
  16 unsigned long __meminitdata e820_table_top;
  17
  18 int after_bootmem;
  19
  20 int direct_gbpages
  21 #ifdef CONFIG_DIRECT_GBPAGES
  22                                 = 1
  23 #endif
  24 ;
  25
  26 static void __init find_early_table_space(unsigned long end, int use_pse,
  27                                           int use_gbpages)
  28 {
  29         unsigned long puds, pmds, ptes, tables, start;
  30
  31         puds = (end + PUD_SIZE - 1) >> PUD_SHIFT;
  32         tables = roundup(puds * sizeof(pud_t), PAGE_SIZE);
  33
  34         if (use_gbpages) {
  35                 unsigned long extra;
  36
  37                 extra = end - ((end>>PUD_SHIFT) << PUD_SHIFT);
  38                 pmds = (extra + PMD_SIZE - 1) >> PMD_SHIFT;
  39         } else
  40                 pmds = (end + PMD_SIZE - 1) >> PMD_SHIFT;
  41
  42         tables += roundup(pmds * sizeof(pmd_t), PAGE_SIZE);
  43
  44         if (use_pse) {
  45                 unsigned long extra;
  46
  47                 extra = end - ((end>>PMD_SHIFT) << PMD_SHIFT);
  48 #ifdef CONFIG_X86_32
  49                 extra += PMD_SIZE;
  50 #endif
  51                 ptes = (extra + PAGE_SIZE - 1) >> PAGE_SHIFT;
  52         } else
  53                 ptes = (end + PAGE_SIZE - 1) >> PAGE_SHIFT;
  54
  55         tables += roundup(ptes * sizeof(pte_t), PAGE_SIZE);
  56
  57 #ifdef CONFIG_X86_32
  58         /* for fixmap */
  59         tables += roundup(__end_of_fixed_addresses * sizeof(pte_t), PAGE_SIZE);
  60 #endif
  61
  62         /*
  63          * RED-PEN putting page tables only on node 0 could
  64          * cause a hotspot and fill up ZONE_DMA. The page tables
  65          * need roughly 0.5KB per GB.
  66          */
  67 #ifdef CONFIG_X86_32
  68         start = 0x7000;
  69         e820_table_start = find_e820_area(start, max_pfn_mapped<<PAGE_SHIFT,
  70                                         tables, PAGE_SIZE);
  71 #else /* CONFIG_X86_64 */
  72         start = 0x8000;
  73         e820_table_start = find_e820_area(start, end, tables, PAGE_SIZE);
  74 #endif
  75         if (e820_table_start == -1UL)
  76                 panic("Cannot find space for the kernel page tables");
  77
  78         e820_table_start >>= PAGE_SHIFT;
  79         e820_table_end = e820_table_start;
  80         e820_table_top = e820_table_start + (tables >> PAGE_SHIFT);
  81
  82         printk(KERN_DEBUG "kernel direct mapping tables up to %lx @ %lx-%lx\n",
  83                 end, e820_table_start << PAGE_SHIFT, e820_table_top << PAGE_SHIFT);
  84 }
  85
  86 struct map_range {
  87         unsigned long start;
  88         unsigned long end;
  89         unsigned page_size_mask;
  90 };
  91
  92 #ifdef CONFIG_X86_32
  93 #define NR_RANGE_MR 3
  94 #else /* CONFIG_X86_64 */
  95 #define NR_RANGE_MR 5
  96 #endif
  97
  98 static int __meminit save_mr(struct map_range *mr, int nr_range,
  99                              unsigned long start_pfn, unsigned long end_pfn,
 100                              unsigned long page_size_mask)
 101 {
 102         if (start_pfn < end_pfn) {
 103                 if (nr_range >= NR_RANGE_MR)
 104                         panic("run out of range for init_memory_mapping\n");
 105                 mr[nr_range].start = start_pfn<<PAGE_SHIFT;
 106                 mr[nr_range].end   = end_pfn<<PAGE_SHIFT;
 107                 mr[nr_range].page_size_mask = page_size_mask;
 108                 nr_range++;
 109         }
 110
 111         return nr_range;
 112 }
 113
 114 #ifdef CONFIG_X86_64
 115 static void __init init_gbpages(void)
 116 {
 117         if (direct_gbpages && cpu_has_gbpages)
 118                 printk(KERN_INFO "Using GB pages for direct mapping\n");
 119         else
 120                 direct_gbpages = 0;
 121 }
 122 #else
 123 static inline void init_gbpages(void)
 124 {
 125 }
 126 #endif
 127
 128 /*
 129  * Setup the direct mapping of the physical memory at PAGE_OFFSET.
 130  * This runs before bootmem is initialized and gets pages directly from
 131  * the physical memory. To access them they are temporarily mapped.
 132  */
 133 unsigned long __init_refok init_memory_mapping(unsigned long start,
 134                                                unsigned long end)
 135 {
 136         unsigned long page_size_mask = 0;
 137         unsigned long start_pfn, end_pfn;
 138         unsigned long ret = 0;
 139         unsigned long pos;
 140
 141         struct map_range mr[NR_RANGE_MR];
 142         int nr_range, i;
 143         int use_pse, use_gbpages;
 144
 145         printk(KERN_INFO "init_memory_mapping: %016lx-%016lx\n", start, end);
 146
 147         if (!after_bootmem)
 148                 init_gbpages();
 149
 150 #ifdef CONFIG_DEBUG_PAGEALLOC
 151         /*
 152          * For CONFIG_DEBUG_PAGEALLOC, identity mapping will use small pages.
 153          * This will simplify cpa(), which otherwise needs to support splitting
 154          * large pages into small in interrupt context, etc.
 155          */
 156         use_pse = use_gbpages = 0;
 157 #else
 158         use_pse = cpu_has_pse;
 159         use_gbpages = direct_gbpages;
 160 #endif
 161
 162 #ifdef CONFIG_X86_32
 163 #ifdef CONFIG_X86_PAE
 164         set_nx();
 165         if (nx_enabled)
 166                 printk(KERN_INFO "NX (Execute Disable) protection: active\n");
 167 #endif
 168
 169         /* Enable PSE if available */
 170         if (cpu_has_pse)
 171                 set_in_cr4(X86_CR4_PSE);
 172
 173         /* Enable PGE if available */
 174         if (cpu_has_pge) {
 175                 set_in_cr4(X86_CR4_PGE);
 176                 __supported_pte_mask |= _PAGE_GLOBAL;
 177         }
 178 #endif
 179
 180         if (use_gbpages)
 181                 page_size_mask |= 1 << PG_LEVEL_1G;
 182         if (use_pse)
 183                 page_size_mask |= 1 << PG_LEVEL_2M;
 184
 185         memset(mr, 0, sizeof(mr));
 186         nr_range = 0;
 187
 188         /* head if not big page alignment ? */
 189         start_pfn = start >> PAGE_SHIFT;
 190         pos = start_pfn << PAGE_SHIFT;
 191 #ifdef CONFIG_X86_32
 192         /*
 193          * Don't use a large page for the first 2/4MB of memory
 194          * because there are often fixed size MTRRs in there
 195          * and overlapping MTRRs into large pages can cause
 196          * slowdowns.
 197          */
 198         if (pos == 0)
 199                 end_pfn = 1<<(PMD_SHIFT - PAGE_SHIFT);
 200         else
 201                 end_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
 202                                  << (PMD_SHIFT - PAGE_SHIFT);
 203 #else /* CONFIG_X86_64 */
 204         end_pfn = ((pos + (PMD_SIZE - 1)) >> PMD_SHIFT)
 205                         << (PMD_SHIFT - PAGE_SHIFT);
 206 #endif
 207         if (end_pfn > (end >> PAGE_SHIFT))
 208                 end_pfn = end >> PAGE_SHIFT;
 209         if (start_pfn < end_pfn) {
 210                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
 211                 pos = end_pfn << PAGE_SHIFT;
 212         }
 213
 214         /* big page (2M) range */
 215         start_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
 216                          << (PMD_SHIFT - PAGE_SHIFT);
 217 #ifdef CONFIG_X86_32
 218         end_pfn = (end>>PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
 219 #else /* CONFIG_X86_64 */
 220         end_pfn = ((pos + (PUD_SIZE - 1))>>PUD_SHIFT)
 221                          << (PUD_SHIFT - PAGE_SHIFT);
 222         if (end_pfn > ((end>>PMD_SHIFT)<<(PMD_SHIFT - PAGE_SHIFT)))
 223                 end_pfn = ((end>>PMD_SHIFT)<<(PMD_SHIFT - PAGE_SHIFT));
 224 #endif
 225
 226         if (start_pfn < end_pfn) {
 227                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 228                                 page_size_mask & (1<<PG_LEVEL_2M));
 229                 pos = end_pfn << PAGE_SHIFT;
 230         }
 231
 232 #ifdef CONFIG_X86_64
 233         /* big page (1G) range */
 234         start_pfn = ((pos + (PUD_SIZE - 1))>>PUD_SHIFT)
 235                          << (PUD_SHIFT - PAGE_SHIFT);
 236         end_pfn = (end >> PUD_SHIFT) << (PUD_SHIFT - PAGE_SHIFT);
 237         if (start_pfn < end_pfn) {
 238                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 239                                 page_size_mask &
 240                                  ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G)));
 241                 pos = end_pfn << PAGE_SHIFT;
 242         }
 243
 244         /* tail is not big page (1G) alignment */
 245         start_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
 246                          << (PMD_SHIFT - PAGE_SHIFT);
 247         end_pfn = (end >> PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
 248         if (start_pfn < end_pfn) {
 249                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 250                                 page_size_mask & (1<<PG_LEVEL_2M));
 251                 pos = end_pfn << PAGE_SHIFT;
 252         }
 253 #endif
 254
 255         /* tail is not big page (2M) alignment */
 256         start_pfn = pos>>PAGE_SHIFT;
 257         end_pfn = end>>PAGE_SHIFT;
 258         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
 259
 260         /* try to merge same page size and continuous */
 261         for (i = 0; nr_range > 1 && i < nr_range - 1; i++) {
 262                 unsigned long old_start;
 263                 if (mr[i].end != mr[i+1].start ||
 264                     mr[i].page_size_mask != mr[i+1].page_size_mask)
 265                         continue;
 266                 /* move it */
 267                 old_start = mr[i].start;
 268                 memmove(&mr[i], &mr[i+1],
 269                         (nr_range - 1 - i) * sizeof(struct map_range));
 270                 mr[i--].start = old_start;
 271                 nr_range--;
 272         }
 273
 274         for (i = 0; i < nr_range; i++)
 275                 printk(KERN_DEBUG " %010lx - %010lx page %s\n",
 276                                 mr[i].start, mr[i].end,
 277                         (mr[i].page_size_mask & (1<<PG_LEVEL_1G))?"1G":(
 278                          (mr[i].page_size_mask & (1<<PG_LEVEL_2M))?"2M":"4k"));
 279
 280         /*
 281          * Find space for the kernel direct mapping tables.
 282          *
 283          * Later we should allocate these tables in the local node of the
 284          * memory mapped. Unfortunately this is done currently before the
 285          * nodes are discovered.
 286          */
 287         if (!after_bootmem)
 288                 find_early_table_space(end, use_pse, use_gbpages);
 289
 290 #ifdef CONFIG_X86_32
 291         for (i = 0; i < nr_range; i++)
 292                 kernel_physical_mapping_init(mr[i].start, mr[i].end,
 293                                              mr[i].page_size_mask);
 294         ret = end;
 295 #else /* CONFIG_X86_64 */
 296         for (i = 0; i < nr_range; i++)
 297                 ret = kernel_physical_mapping_init(mr[i].start, mr[i].end,
 298                                                    mr[i].page_size_mask);
 299 #endif
 300
 301 #ifdef CONFIG_X86_32
 302         early_ioremap_page_table_range_init();
 303
 304         load_cr3(swapper_pg_dir);
 305 #endif
 306
 307 #ifdef CONFIG_X86_64
 308         if (!after_bootmem && !start) {
 309                 pud_t *pud;
 310                 pmd_t *pmd;
 311
 312                 mmu_cr4_features = read_cr4();
 313
 314                 /*
 315                  * _brk_end cannot change anymore, but it and _end may be
 316                  * located on different 2M pages. cleanup_highmap(), however,
 317                  * can only consider _end when it runs, so destroy any
 318                  * mappings beyond _brk_end here.
 319                  */
 320                 pud = pud_offset(pgd_offset_k(_brk_end), _brk_end);
 321                 pmd = pmd_offset(pud, _brk_end - 1);
 322                 while (++pmd <= pmd_offset(pud, (unsigned long)_end - 1))
 323                         pmd_clear(pmd);
 324         }
 325 #endif
 326         __flush_tlb_all();
 327
 328         if (!after_bootmem && e820_table_end > e820_table_start)
 329                 reserve_early(e820_table_start << PAGE_SHIFT,
 330                                  e820_table_end << PAGE_SHIFT, "PGTABLE");
 331
 332         if (!after_bootmem)
 333                 early_memtest(start, end);
 334
 335         return ret >> PAGE_SHIFT;
 336 }
 337
 338
 339 /*
 340  * devmem_is_allowed() checks to see if /dev/mem access to a certain address
 341  * is valid. The argument is a physical page number.
 342  *
 343  *
 344  * On x86, access has to be given to the first megabyte of ram because that area
 345  * contains bios code and data regions used by X and dosemu and similar apps.
 346  * Access has to be given to non-kernel-ram areas as well, these contain the PCI
 347  * mmio resources as well as potential bios/acpi data regions.
 348  */
 349 int devmem_is_allowed(unsigned long pagenr)
 350 {
 351         if (pagenr <= 256)
 352                 return 1;
 353         if (iomem_is_exclusive(pagenr << PAGE_SHIFT))
 354                 return 0;
 355         if (!page_is_ram(pagenr))
 356                 return 1;
 357         return 0;
 358 }
 359
 360 void free_init_pages(char *what, unsigned long begin, unsigned long end)
 361 {
 362         unsigned long addr = begin;
 363
 364         if (addr >= end)
 365                 return;
 366
 367         /*
 368          * If debugging page accesses then do not free this memory but
 369          * mark them not present - any buggy init-section access will
 370          * create a kernel page fault:
 371          */
 372 #ifdef CONFIG_DEBUG_PAGEALLOC
 373         printk(KERN_INFO "debug: unmapping init memory %08lx..%08lx\n",
 374                 begin, PAGE_ALIGN(end));
 375         set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
 376 #else
 377         /*
 378          * We just marked the kernel text read only above, now that
 379          * we are going to free part of that, we need to make that
 380          * writeable first.
 381          */
 382         set_memory_rw(begin, (end - begin) >> PAGE_SHIFT);
 383
 384         printk(KERN_INFO "Freeing %s: %luk freed\n", what, (end - begin) >> 10);
 385
 386         for (; addr < end; addr += PAGE_SIZE) {
 387                 ClearPageReserved(virt_to_page(addr));
 388                 init_page_count(virt_to_page(addr));
 389                 memset((void *)(addr & ~(PAGE_SIZE-1)),
 390                         POISON_FREE_INITMEM, PAGE_SIZE);
 391                 free_page(addr);
 392                 totalram_pages++;
 393         }
 394 #endif
 395 }
 396
 397 void free_initmem(void)
 398 {
 399         free_init_pages("unused kernel memory",
 400                         (unsigned long)(&__init_begin),
 401                         (unsigned long)(&__init_end));
 402 }
 403
 404 #ifdef CONFIG_BLK_DEV_INITRD
 405 void free_initrd_mem(unsigned long start, unsigned long end)
 406 {
 407         free_init_pages("initrd memory", start, end);
 408 }
 409 #endif