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 #include <asm/tlb.h>
  13
  14 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
  15
  16 unsigned long __initdata e820_table_start;
  17 unsigned long __meminitdata e820_table_end;
  18 unsigned long __meminitdata e820_table_top;
  19
  20 int after_bootmem;
  21
  22 int direct_gbpages
  23 #ifdef CONFIG_DIRECT_GBPAGES
  24                                 = 1
  25 #endif
  26 ;
  27
  28 int nx_enabled;
  29
  30 #if defined(CONFIG_X86_64) || defined(CONFIG_X86_PAE)
  31 static int disable_nx __cpuinitdata;
  32
  33 /*
  34  * noexec = on|off
  35  *
  36  * Control non-executable mappings for processes.
  37  *
  38  * on      Enable
  39  * off     Disable
  40  */
  41 static int __init noexec_setup(char *str)
  42 {
  43         if (!str)
  44                 return -EINVAL;
  45         if (!strncmp(str, "on", 2)) {
  46                 __supported_pte_mask |= _PAGE_NX;
  47                 disable_nx = 0;
  48         } else if (!strncmp(str, "off", 3)) {
  49                 disable_nx = 1;
  50                 __supported_pte_mask &= ~_PAGE_NX;
  51         }
  52         return 0;
  53 }
  54 early_param("noexec", noexec_setup);
  55 #endif
  56
  57 #ifdef CONFIG_X86_PAE
  58 static void __init set_nx(void)
  59 {
  60         unsigned int v[4], l, h;
  61
  62         if (cpu_has_pae && (cpuid_eax(0x80000000) > 0x80000001)) {
  63                 cpuid(0x80000001, &v[0], &v[1], &v[2], &v[3]);
  64
  65                 if ((v[3] & (1 << 20)) && !disable_nx) {
  66                         rdmsr(MSR_EFER, l, h);
  67                         l |= EFER_NX;
  68                         wrmsr(MSR_EFER, l, h);
  69                         nx_enabled = 1;
  70                         __supported_pte_mask |= _PAGE_NX;
  71                 }
  72         }
  73 }
  74 #else
  75 static inline void set_nx(void)
  76 {
  77 }
  78 #endif
  79
  80 #ifdef CONFIG_X86_64
  81 void __cpuinit check_efer(void)
  82 {
  83         unsigned long efer;
  84
  85         rdmsrl(MSR_EFER, efer);
  86         if (!(efer & EFER_NX) || disable_nx)
  87                 __supported_pte_mask &= ~_PAGE_NX;
  88 }
  89 #endif
  90
  91 static void __init find_early_table_space(unsigned long end, int use_pse,
  92                                           int use_gbpages)
  93 {
  94         unsigned long puds, pmds, ptes, tables, start;
  95
  96         puds = (end + PUD_SIZE - 1) >> PUD_SHIFT;
  97         tables = roundup(puds * sizeof(pud_t), PAGE_SIZE);
  98
  99         if (use_gbpages) {
 100                 unsigned long extra;
 101
 102                 extra = end - ((end>>PUD_SHIFT) << PUD_SHIFT);
 103                 pmds = (extra + PMD_SIZE - 1) >> PMD_SHIFT;
 104         } else
 105                 pmds = (end + PMD_SIZE - 1) >> PMD_SHIFT;
 106
 107         tables += roundup(pmds * sizeof(pmd_t), PAGE_SIZE);
 108
 109         if (use_pse) {
 110                 unsigned long extra;
 111
 112                 extra = end - ((end>>PMD_SHIFT) << PMD_SHIFT);
 113 #ifdef CONFIG_X86_32
 114                 extra += PMD_SIZE;
 115 #endif
 116                 ptes = (extra + PAGE_SIZE - 1) >> PAGE_SHIFT;
 117         } else
 118                 ptes = (end + PAGE_SIZE - 1) >> PAGE_SHIFT;
 119
 120         tables += roundup(ptes * sizeof(pte_t), PAGE_SIZE);
 121
 122 #ifdef CONFIG_X86_32
 123         /* for fixmap */
 124         tables += roundup(__end_of_fixed_addresses * sizeof(pte_t), PAGE_SIZE);
 125 #endif
 126
 127         /*
 128          * RED-PEN putting page tables only on node 0 could
 129          * cause a hotspot and fill up ZONE_DMA. The page tables
 130          * need roughly 0.5KB per GB.
 131          */
 132 #ifdef CONFIG_X86_32
 133         start = 0x7000;
 134         e820_table_start = find_e820_area(start, max_pfn_mapped<<PAGE_SHIFT,
 135                                         tables, PAGE_SIZE);
 136 #else /* CONFIG_X86_64 */
 137         start = 0x8000;
 138         e820_table_start = find_e820_area(start, end, tables, PAGE_SIZE);
 139 #endif
 140         if (e820_table_start == -1UL)
 141                 panic("Cannot find space for the kernel page tables");
 142
 143         e820_table_start >>= PAGE_SHIFT;
 144         e820_table_end = e820_table_start;
 145         e820_table_top = e820_table_start + (tables >> PAGE_SHIFT);
 146
 147         printk(KERN_DEBUG "kernel direct mapping tables up to %lx @ %lx-%lx\n",
 148                 end, e820_table_start << PAGE_SHIFT, e820_table_top << PAGE_SHIFT);
 149 }
 150
 151 struct map_range {
 152         unsigned long start;
 153         unsigned long end;
 154         unsigned page_size_mask;
 155 };
 156
 157 #ifdef CONFIG_X86_32
 158 #define NR_RANGE_MR 3
 159 #else /* CONFIG_X86_64 */
 160 #define NR_RANGE_MR 5
 161 #endif
 162
 163 static int __meminit save_mr(struct map_range *mr, int nr_range,
 164                              unsigned long start_pfn, unsigned long end_pfn,
 165                              unsigned long page_size_mask)
 166 {
 167         if (start_pfn < end_pfn) {
 168                 if (nr_range >= NR_RANGE_MR)
 169                         panic("run out of range for init_memory_mapping\n");
 170                 mr[nr_range].start = start_pfn<<PAGE_SHIFT;
 171                 mr[nr_range].end   = end_pfn<<PAGE_SHIFT;
 172                 mr[nr_range].page_size_mask = page_size_mask;
 173                 nr_range++;
 174         }
 175
 176         return nr_range;
 177 }
 178
 179 #ifdef CONFIG_X86_64
 180 static void __init init_gbpages(void)
 181 {
 182         if (direct_gbpages && cpu_has_gbpages)
 183                 printk(KERN_INFO "Using GB pages for direct mapping\n");
 184         else
 185                 direct_gbpages = 0;
 186 }
 187 #else
 188 static inline void init_gbpages(void)
 189 {
 190 }
 191 #endif
 192
 193 /*
 194  * Setup the direct mapping of the physical memory at PAGE_OFFSET.
 195  * This runs before bootmem is initialized and gets pages directly from
 196  * the physical memory. To access them they are temporarily mapped.
 197  */
 198 unsigned long __init_refok init_memory_mapping(unsigned long start,
 199                                                unsigned long end)
 200 {
 201         unsigned long page_size_mask = 0;
 202         unsigned long start_pfn, end_pfn;
 203         unsigned long ret = 0;
 204         unsigned long pos;
 205
 206         struct map_range mr[NR_RANGE_MR];
 207         int nr_range, i;
 208         int use_pse, use_gbpages;
 209
 210         printk(KERN_INFO "init_memory_mapping: %016lx-%016lx\n", start, end);
 211
 212         if (!after_bootmem)
 213                 init_gbpages();
 214
 215 #ifdef CONFIG_DEBUG_PAGEALLOC
 216         /*
 217          * For CONFIG_DEBUG_PAGEALLOC, identity mapping will use small pages.
 218          * This will simplify cpa(), which otherwise needs to support splitting
 219          * large pages into small in interrupt context, etc.
 220          */
 221         use_pse = use_gbpages = 0;
 222 #else
 223         use_pse = cpu_has_pse;
 224         use_gbpages = direct_gbpages;
 225 #endif
 226
 227         set_nx();
 228         if (nx_enabled)
 229                 printk(KERN_INFO "NX (Execute Disable) protection: active\n");
 230
 231         /* Enable PSE if available */
 232         if (cpu_has_pse)
 233                 set_in_cr4(X86_CR4_PSE);
 234
 235         /* Enable PGE if available */
 236         if (cpu_has_pge) {
 237                 set_in_cr4(X86_CR4_PGE);
 238                 __supported_pte_mask |= _PAGE_GLOBAL;
 239         }
 240
 241         if (use_gbpages)
 242                 page_size_mask |= 1 << PG_LEVEL_1G;
 243         if (use_pse)
 244                 page_size_mask |= 1 << PG_LEVEL_2M;
 245
 246         memset(mr, 0, sizeof(mr));
 247         nr_range = 0;
 248
 249         /* head if not big page alignment ? */
 250         start_pfn = start >> PAGE_SHIFT;
 251         pos = start_pfn << PAGE_SHIFT;
 252 #ifdef CONFIG_X86_32
 253         /*
 254          * Don't use a large page for the first 2/4MB of memory
 255          * because there are often fixed size MTRRs in there
 256          * and overlapping MTRRs into large pages can cause
 257          * slowdowns.
 258          */
 259         if (pos == 0)
 260                 end_pfn = 1<<(PMD_SHIFT - PAGE_SHIFT);
 261         else
 262                 end_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
 263                                  << (PMD_SHIFT - PAGE_SHIFT);
 264 #else /* CONFIG_X86_64 */
 265         end_pfn = ((pos + (PMD_SIZE - 1)) >> PMD_SHIFT)
 266                         << (PMD_SHIFT - PAGE_SHIFT);
 267 #endif
 268         if (end_pfn > (end >> PAGE_SHIFT))
 269                 end_pfn = end >> PAGE_SHIFT;
 270         if (start_pfn < end_pfn) {
 271                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
 272                 pos = end_pfn << PAGE_SHIFT;
 273         }
 274
 275         /* big page (2M) range */
 276         start_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
 277                          << (PMD_SHIFT - PAGE_SHIFT);
 278 #ifdef CONFIG_X86_32
 279         end_pfn = (end>>PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
 280 #else /* CONFIG_X86_64 */
 281         end_pfn = ((pos + (PUD_SIZE - 1))>>PUD_SHIFT)
 282                          << (PUD_SHIFT - PAGE_SHIFT);
 283         if (end_pfn > ((end>>PMD_SHIFT)<<(PMD_SHIFT - PAGE_SHIFT)))
 284                 end_pfn = ((end>>PMD_SHIFT)<<(PMD_SHIFT - PAGE_SHIFT));
 285 #endif
 286
 287         if (start_pfn < end_pfn) {
 288                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 289                                 page_size_mask & (1<<PG_LEVEL_2M));
 290                 pos = end_pfn << PAGE_SHIFT;
 291         }
 292
 293 #ifdef CONFIG_X86_64
 294         /* big page (1G) range */
 295         start_pfn = ((pos + (PUD_SIZE - 1))>>PUD_SHIFT)
 296                          << (PUD_SHIFT - PAGE_SHIFT);
 297         end_pfn = (end >> PUD_SHIFT) << (PUD_SHIFT - PAGE_SHIFT);
 298         if (start_pfn < end_pfn) {
 299                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 300                                 page_size_mask &
 301                                  ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G)));
 302                 pos = end_pfn << PAGE_SHIFT;
 303         }
 304
 305         /* tail is not big page (1G) alignment */
 306         start_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
 307                          << (PMD_SHIFT - PAGE_SHIFT);
 308         end_pfn = (end >> PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
 309         if (start_pfn < end_pfn) {
 310                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 311                                 page_size_mask & (1<<PG_LEVEL_2M));
 312                 pos = end_pfn << PAGE_SHIFT;
 313         }
 314 #endif
 315
 316         /* tail is not big page (2M) alignment */
 317         start_pfn = pos>>PAGE_SHIFT;
 318         end_pfn = end>>PAGE_SHIFT;
 319         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
 320
 321         /* try to merge same page size and continuous */
 322         for (i = 0; nr_range > 1 && i < nr_range - 1; i++) {
 323                 unsigned long old_start;
 324                 if (mr[i].end != mr[i+1].start ||
 325                     mr[i].page_size_mask != mr[i+1].page_size_mask)
 326                         continue;
 327                 /* move it */
 328                 old_start = mr[i].start;
 329                 memmove(&mr[i], &mr[i+1],
 330                         (nr_range - 1 - i) * sizeof(struct map_range));
 331                 mr[i--].start = old_start;
 332                 nr_range--;
 333         }
 334
 335         for (i = 0; i < nr_range; i++)
 336                 printk(KERN_DEBUG " %010lx - %010lx page %s\n",
 337                                 mr[i].start, mr[i].end,
 338                         (mr[i].page_size_mask & (1<<PG_LEVEL_1G))?"1G":(
 339                          (mr[i].page_size_mask & (1<<PG_LEVEL_2M))?"2M":"4k"));
 340
 341         /*
 342          * Find space for the kernel direct mapping tables.
 343          *
 344          * Later we should allocate these tables in the local node of the
 345          * memory mapped. Unfortunately this is done currently before the
 346          * nodes are discovered.
 347          */
 348         if (!after_bootmem)
 349                 find_early_table_space(end, use_pse, use_gbpages);
 350
 351 #ifdef CONFIG_X86_32
 352         for (i = 0; i < nr_range; i++)
 353                 kernel_physical_mapping_init(mr[i].start, mr[i].end,
 354                                              mr[i].page_size_mask);
 355         ret = end;
 356 #else /* CONFIG_X86_64 */
 357         for (i = 0; i < nr_range; i++)
 358                 ret = kernel_physical_mapping_init(mr[i].start, mr[i].end,
 359                                                    mr[i].page_size_mask);
 360 #endif
 361
 362 #ifdef CONFIG_X86_32
 363         early_ioremap_page_table_range_init();
 364
 365         load_cr3(swapper_pg_dir);
 366 #endif
 367
 368 #ifdef CONFIG_X86_64
 369         if (!after_bootmem)
 370                 mmu_cr4_features = read_cr4();
 371 #endif
 372         __flush_tlb_all();
 373
 374         if (!after_bootmem && e820_table_end > e820_table_start)
 375                 reserve_early(e820_table_start << PAGE_SHIFT,
 376                                  e820_table_end << PAGE_SHIFT, "PGTABLE");
 377
 378         if (!after_bootmem)
 379                 early_memtest(start, end);
 380
 381         return ret >> PAGE_SHIFT;
 382 }
 383
 384
 385 /*
 386  * devmem_is_allowed() checks to see if /dev/mem access to a certain address
 387  * is valid. The argument is a physical page number.
 388  *
 389  *
 390  * On x86, access has to be given to the first megabyte of ram because that area
 391  * contains bios code and data regions used by X and dosemu and similar apps.
 392  * Access has to be given to non-kernel-ram areas as well, these contain the PCI
 393  * mmio resources as well as potential bios/acpi data regions.
 394  */
 395 int devmem_is_allowed(unsigned long pagenr)
 396 {
 397         if (pagenr <= 256)
 398                 return 1;
 399         if (iomem_is_exclusive(pagenr << PAGE_SHIFT))
 400                 return 0;
 401         if (!page_is_ram(pagenr))
 402                 return 1;
 403         return 0;
 404 }
 405
 406 void free_init_pages(char *what, unsigned long begin, unsigned long end)
 407 {
 408         unsigned long addr = begin;
 409
 410         if (addr >= end)
 411                 return;
 412
 413         /*
 414          * If debugging page accesses then do not free this memory but
 415          * mark them not present - any buggy init-section access will
 416          * create a kernel page fault:
 417          */
 418 #ifdef CONFIG_DEBUG_PAGEALLOC
 419         printk(KERN_INFO "debug: unmapping init memory %08lx..%08lx\n",
 420                 begin, PAGE_ALIGN(end));
 421         set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
 422 #else
 423         /*
 424          * We just marked the kernel text read only above, now that
 425          * we are going to free part of that, we need to make that
 426          * writeable first.
 427          */
 428         set_memory_rw(begin, (end - begin) >> PAGE_SHIFT);
 429
 430         printk(KERN_INFO "Freeing %s: %luk freed\n", what, (end - begin) >> 10);
 431
 432         for (; addr < end; addr += PAGE_SIZE) {
 433                 ClearPageReserved(virt_to_page(addr));
 434                 init_page_count(virt_to_page(addr));
 435                 memset((void *)(addr & ~(PAGE_SIZE-1)),
 436                         POISON_FREE_INITMEM, PAGE_SIZE);
 437                 free_page(addr);
 438                 totalram_pages++;
 439         }
 440 #endif
 441 }
 442
 443 void free_initmem(void)
 444 {
 445         free_init_pages("unused kernel memory",
 446                         (unsigned long)(&__init_begin),
 447                         (unsigned long)(&__init_end));
 448 }
 449
 450 #ifdef CONFIG_BLK_DEV_INITRD
 451 void free_initrd_mem(unsigned long start, unsigned long end)
 452 {
 453         free_init_pages("initrd memory", start, end);
 454 }
 455 #endif