arch/x86/mm/init.c

   1 #include <linux/initrd.h>
   2 #include <linux/ioport.h>
   3 #include <linux/swap.h>
   4
   5 #include <asm/cacheflush.h>
   6 #include <asm/e820.h>
   7 #include <asm/init.h>
   8 #include <asm/page.h>
   9 #include <asm/page_types.h>
  10 #include <asm/sections.h>
  11 #include <asm/setup.h>
  12 #include <asm/system.h>
  13 #include <asm/tlbflush.h>
  14 #include <asm/tlb.h>
  15 #include <asm/proto.h>
  16
  17 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
  18
  19 unsigned long __initdata e820_table_start;
  20 unsigned long __meminitdata e820_table_end;
  21 unsigned long __meminitdata e820_table_top;
  22
  23 int after_bootmem;
  24
  25 int direct_gbpages
  26 #ifdef CONFIG_DIRECT_GBPAGES
  27                                 = 1
  28 #endif
  29 ;
  30
  31 int nx_enabled;
  32
  33 #if defined(CONFIG_X86_64) || defined(CONFIG_X86_PAE)
  34 static int disable_nx __cpuinitdata;
  35
  36 /*
  37  * noexec = on|off
  38  *
  39  * Control non-executable mappings for processes.
  40  *
  41  * on      Enable
  42  * off     Disable
  43  */
  44 static int __init noexec_setup(char *str)
  45 {
  46         if (!str)
  47                 return -EINVAL;
  48         if (!strncmp(str, "on", 2)) {
  49                 __supported_pte_mask |= _PAGE_NX;
  50                 disable_nx = 0;
  51         } else if (!strncmp(str, "off", 3)) {
  52                 disable_nx = 1;
  53                 __supported_pte_mask &= ~_PAGE_NX;
  54         }
  55         return 0;
  56 }
  57 early_param("noexec", noexec_setup);
  58 #endif
  59
  60 #ifdef CONFIG_X86_PAE
  61 static void __init set_nx(void)
  62 {
  63         unsigned int v[4], l, h;
  64
  65         if (cpu_has_pae && (cpuid_eax(0x80000000) > 0x80000001)) {
  66                 cpuid(0x80000001, &v[0], &v[1], &v[2], &v[3]);
  67
  68                 if ((v[3] & (1 << 20)) && !disable_nx) {
  69                         rdmsr(MSR_EFER, l, h);
  70                         l |= EFER_NX;
  71                         wrmsr(MSR_EFER, l, h);
  72                         nx_enabled = 1;
  73                         __supported_pte_mask |= _PAGE_NX;
  74                 }
  75         }
  76 }
  77 #else
  78 static inline void set_nx(void)
  79 {
  80 }
  81 #endif
  82
  83 #ifdef CONFIG_X86_64
  84 void __cpuinit check_efer(void)
  85 {
  86         unsigned long efer;
  87
  88         rdmsrl(MSR_EFER, efer);
  89         if (!(efer & EFER_NX) || disable_nx)
  90                 __supported_pte_mask &= ~_PAGE_NX;
  91 }
  92 #endif
  93
  94 static void __init find_early_table_space(unsigned long end, int use_pse,
  95                                           int use_gbpages)
  96 {
  97         unsigned long puds, pmds, ptes, tables, start;
  98
  99         puds = (end + PUD_SIZE - 1) >> PUD_SHIFT;
 100         tables = roundup(puds * sizeof(pud_t), PAGE_SIZE);
 101
 102         if (use_gbpages) {
 103                 unsigned long extra;
 104
 105                 extra = end - ((end>>PUD_SHIFT) << PUD_SHIFT);
 106                 pmds = (extra + PMD_SIZE - 1) >> PMD_SHIFT;
 107         } else
 108                 pmds = (end + PMD_SIZE - 1) >> PMD_SHIFT;
 109
 110         tables += roundup(pmds * sizeof(pmd_t), PAGE_SIZE);
 111
 112         if (use_pse) {
 113                 unsigned long extra;
 114
 115                 extra = end - ((end>>PMD_SHIFT) << PMD_SHIFT);
 116 #ifdef CONFIG_X86_32
 117                 extra += PMD_SIZE;
 118 #endif
 119                 ptes = (extra + PAGE_SIZE - 1) >> PAGE_SHIFT;
 120         } else
 121                 ptes = (end + PAGE_SIZE - 1) >> PAGE_SHIFT;
 122
 123         tables += roundup(ptes * sizeof(pte_t), PAGE_SIZE);
 124
 125 #ifdef CONFIG_X86_32
 126         /* for fixmap */
 127         tables += roundup(__end_of_fixed_addresses * sizeof(pte_t), PAGE_SIZE);
 128 #endif
 129
 130         /*
 131          * RED-PEN putting page tables only on node 0 could
 132          * cause a hotspot and fill up ZONE_DMA. The page tables
 133          * need roughly 0.5KB per GB.
 134          */
 135 #ifdef CONFIG_X86_32
 136         start = 0x7000;
 137 #else
 138         start = 0x8000;
 139 #endif
 140         e820_table_start = find_e820_area(start, max_pfn_mapped<<PAGE_SHIFT,
 141                                         tables, PAGE_SIZE);
 142         if (e820_table_start == -1UL)
 143                 panic("Cannot find space for the kernel page tables");
 144
 145         e820_table_start >>= PAGE_SHIFT;
 146         e820_table_end = e820_table_start;
 147         e820_table_top = e820_table_start + (tables >> PAGE_SHIFT);
 148
 149         printk(KERN_DEBUG "kernel direct mapping tables up to %lx @ %lx-%lx\n",
 150                 end, e820_table_start << PAGE_SHIFT, e820_table_top << PAGE_SHIFT);
 151 }
 152
 153 struct map_range {
 154         unsigned long start;
 155         unsigned long end;
 156         unsigned page_size_mask;
 157 };
 158
 159 #ifdef CONFIG_X86_32
 160 #define NR_RANGE_MR 3
 161 #else /* CONFIG_X86_64 */
 162 #define NR_RANGE_MR 5
 163 #endif
 164
 165 static int __meminit save_mr(struct map_range *mr, int nr_range,
 166                              unsigned long start_pfn, unsigned long end_pfn,
 167                              unsigned long page_size_mask)
 168 {
 169         if (start_pfn < end_pfn) {
 170                 if (nr_range >= NR_RANGE_MR)
 171                         panic("run out of range for init_memory_mapping\n");
 172                 mr[nr_range].start = start_pfn<<PAGE_SHIFT;
 173                 mr[nr_range].end   = end_pfn<<PAGE_SHIFT;
 174                 mr[nr_range].page_size_mask = page_size_mask;
 175                 nr_range++;
 176         }
 177
 178         return nr_range;
 179 }
 180
 181 /*
 182  * Setup the direct mapping of the physical memory at PAGE_OFFSET.
 183  * This runs before bootmem is initialized and gets pages directly from
 184  * the physical memory. To access them they are temporarily mapped.
 185  */
 186 unsigned long __init_refok init_memory_mapping(unsigned long start,
 187                                                unsigned long end)
 188 {
 189         unsigned long page_size_mask = 0;
 190         unsigned long start_pfn, end_pfn;
 191         unsigned long ret = 0;
 192         unsigned long pos;
 193
 194         struct map_range mr[NR_RANGE_MR];
 195         int nr_range, i;
 196         int use_pse, use_gbpages;
 197
 198         printk(KERN_INFO "init_memory_mapping: %016lx-%016lx\n", start, end);
 199
 200 #if defined(CONFIG_DEBUG_PAGEALLOC) || defined(CONFIG_KMEMCHECK)
 201         /*
 202          * For CONFIG_DEBUG_PAGEALLOC, identity mapping will use small pages.
 203          * This will simplify cpa(), which otherwise needs to support splitting
 204          * large pages into small in interrupt context, etc.
 205          */
 206         use_pse = use_gbpages = 0;
 207 #else
 208         use_pse = cpu_has_pse;
 209         use_gbpages = direct_gbpages;
 210 #endif
 211
 212         set_nx();
 213         if (nx_enabled)
 214                 printk(KERN_INFO "NX (Execute Disable) protection: active\n");
 215
 216         /* Enable PSE if available */
 217         if (cpu_has_pse)
 218                 set_in_cr4(X86_CR4_PSE);
 219
 220         /* Enable PGE if available */
 221         if (cpu_has_pge) {
 222                 set_in_cr4(X86_CR4_PGE);
 223                 __supported_pte_mask |= _PAGE_GLOBAL;
 224         }
 225
 226         if (use_gbpages)
 227                 page_size_mask |= 1 << PG_LEVEL_1G;
 228         if (use_pse)
 229                 page_size_mask |= 1 << PG_LEVEL_2M;
 230
 231         memset(mr, 0, sizeof(mr));
 232         nr_range = 0;
 233
 234         /* head if not big page alignment ? */
 235         start_pfn = start >> PAGE_SHIFT;
 236         pos = start_pfn << PAGE_SHIFT;
 237 #ifdef CONFIG_X86_32
 238         /*
 239          * Don't use a large page for the first 2/4MB of memory
 240          * because there are often fixed size MTRRs in there
 241          * and overlapping MTRRs into large pages can cause
 242          * slowdowns.
 243          */
 244         if (pos == 0)
 245                 end_pfn = 1<<(PMD_SHIFT - PAGE_SHIFT);
 246         else
 247                 end_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
 248                                  << (PMD_SHIFT - PAGE_SHIFT);
 249 #else /* CONFIG_X86_64 */
 250         end_pfn = ((pos + (PMD_SIZE - 1)) >> PMD_SHIFT)
 251                         << (PMD_SHIFT - PAGE_SHIFT);
 252 #endif
 253         if (end_pfn > (end >> PAGE_SHIFT))
 254                 end_pfn = end >> PAGE_SHIFT;
 255         if (start_pfn < end_pfn) {
 256                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
 257                 pos = end_pfn << PAGE_SHIFT;
 258         }
 259
 260         /* big page (2M) range */
 261         start_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
 262                          << (PMD_SHIFT - PAGE_SHIFT);
 263 #ifdef CONFIG_X86_32
 264         end_pfn = (end>>PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
 265 #else /* CONFIG_X86_64 */
 266         end_pfn = ((pos + (PUD_SIZE - 1))>>PUD_SHIFT)
 267                          << (PUD_SHIFT - PAGE_SHIFT);
 268         if (end_pfn > ((end>>PMD_SHIFT)<<(PMD_SHIFT - PAGE_SHIFT)))
 269                 end_pfn = ((end>>PMD_SHIFT)<<(PMD_SHIFT - PAGE_SHIFT));
 270 #endif
 271
 272         if (start_pfn < end_pfn) {
 273                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 274                                 page_size_mask & (1<<PG_LEVEL_2M));
 275                 pos = end_pfn << PAGE_SHIFT;
 276         }
 277
 278 #ifdef CONFIG_X86_64
 279         /* big page (1G) range */
 280         start_pfn = ((pos + (PUD_SIZE - 1))>>PUD_SHIFT)
 281                          << (PUD_SHIFT - PAGE_SHIFT);
 282         end_pfn = (end >> PUD_SHIFT) << (PUD_SHIFT - PAGE_SHIFT);
 283         if (start_pfn < end_pfn) {
 284                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 285                                 page_size_mask &
 286                                  ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G)));
 287                 pos = end_pfn << PAGE_SHIFT;
 288         }
 289
 290         /* tail is not big page (1G) alignment */
 291         start_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
 292                          << (PMD_SHIFT - PAGE_SHIFT);
 293         end_pfn = (end >> PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
 294         if (start_pfn < end_pfn) {
 295                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 296                                 page_size_mask & (1<<PG_LEVEL_2M));
 297                 pos = end_pfn << PAGE_SHIFT;
 298         }
 299 #endif
 300
 301         /* tail is not big page (2M) alignment */
 302         start_pfn = pos>>PAGE_SHIFT;
 303         end_pfn = end>>PAGE_SHIFT;
 304         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
 305
 306         /* try to merge same page size and continuous */
 307         for (i = 0; nr_range > 1 && i < nr_range - 1; i++) {
 308                 unsigned long old_start;
 309                 if (mr[i].end != mr[i+1].start ||
 310                     mr[i].page_size_mask != mr[i+1].page_size_mask)
 311                         continue;
 312                 /* move it */
 313                 old_start = mr[i].start;
 314                 memmove(&mr[i], &mr[i+1],
 315                         (nr_range - 1 - i) * sizeof(struct map_range));
 316                 mr[i--].start = old_start;
 317                 nr_range--;
 318         }
 319
 320         for (i = 0; i < nr_range; i++)
 321                 printk(KERN_DEBUG " %010lx - %010lx page %s\n",
 322                                 mr[i].start, mr[i].end,
 323                         (mr[i].page_size_mask & (1<<PG_LEVEL_1G))?"1G":(
 324                          (mr[i].page_size_mask & (1<<PG_LEVEL_2M))?"2M":"4k"));
 325
 326         /*
 327          * Find space for the kernel direct mapping tables.
 328          *
 329          * Later we should allocate these tables in the local node of the
 330          * memory mapped. Unfortunately this is done currently before the
 331          * nodes are discovered.
 332          */
 333         if (!after_bootmem)
 334                 find_early_table_space(end, use_pse, use_gbpages);
 335
 336 #ifdef CONFIG_X86_32
 337         for (i = 0; i < nr_range; i++)
 338                 kernel_physical_mapping_init(mr[i].start, mr[i].end,
 339                                              mr[i].page_size_mask);
 340         ret = end;
 341 #else /* CONFIG_X86_64 */
 342         for (i = 0; i < nr_range; i++)
 343                 ret = kernel_physical_mapping_init(mr[i].start, mr[i].end,
 344                                                    mr[i].page_size_mask);
 345 #endif
 346
 347 #ifdef CONFIG_X86_32
 348         early_ioremap_page_table_range_init();
 349
 350         load_cr3(swapper_pg_dir);
 351 #endif
 352
 353 #ifdef CONFIG_X86_64
 354         if (!after_bootmem && !start) {
 355                 pud_t *pud;
 356                 pmd_t *pmd;
 357
 358                 mmu_cr4_features = read_cr4();
 359
 360                 /*
 361                  * _brk_end cannot change anymore, but it and _end may be
 362                  * located on different 2M pages. cleanup_highmap(), however,
 363                  * can only consider _end when it runs, so destroy any
 364                  * mappings beyond _brk_end here.
 365                  */
 366                 pud = pud_offset(pgd_offset_k(_brk_end), _brk_end);
 367                 pmd = pmd_offset(pud, _brk_end - 1);
 368                 while (++pmd <= pmd_offset(pud, (unsigned long)_end - 1))
 369                         pmd_clear(pmd);
 370         }
 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