arch/x86/mm/init.c

   1 #include <linux/gfp.h>
   2 #include <linux/initrd.h>
   3 #include <linux/ioport.h>
   4 #include <linux/swap.h>
   5 #include <linux/memblock.h>
   6
   7 #include <asm/cacheflush.h>
   8 #include <asm/e820.h>
   9 #include <asm/init.h>
  10 #include <asm/page.h>
  11 #include <asm/page_types.h>
  12 #include <asm/sections.h>
  13 #include <asm/setup.h>
  14 #include <asm/system.h>
  15 #include <asm/tlbflush.h>
  16 #include <asm/tlb.h>
  17 #include <asm/proto.h>
  18
  19 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
  20
  21 unsigned long __initdata e820_table_start;
  22 unsigned long __meminitdata e820_table_end;
  23 unsigned long __meminitdata e820_table_top;
  24
  25 int after_bootmem;
  26
  27 int direct_gbpages
  28 #ifdef CONFIG_DIRECT_GBPAGES
  29                                 = 1
  30 #endif
  31 ;
  32
  33 static void __init find_early_table_space(unsigned long end, int use_pse,
  34                                           int use_gbpages)
  35 {
  36         unsigned long puds, pmds, ptes, tables, start;
  37         phys_addr_t base;
  38
  39         puds = (end + PUD_SIZE - 1) >> PUD_SHIFT;
  40         tables = roundup(puds * sizeof(pud_t), PAGE_SIZE);
  41
  42         if (use_gbpages) {
  43                 unsigned long extra;
  44
  45                 extra = end - ((end>>PUD_SHIFT) << PUD_SHIFT);
  46                 pmds = (extra + PMD_SIZE - 1) >> PMD_SHIFT;
  47         } else
  48                 pmds = (end + PMD_SIZE - 1) >> PMD_SHIFT;
  49
  50         tables += roundup(pmds * sizeof(pmd_t), PAGE_SIZE);
  51
  52         if (use_pse) {
  53                 unsigned long extra;
  54
  55                 extra = end - ((end>>PMD_SHIFT) << PMD_SHIFT);
  56 #ifdef CONFIG_X86_32
  57                 extra += PMD_SIZE;
  58 #endif
  59                 ptes = (extra + PAGE_SIZE - 1) >> PAGE_SHIFT;
  60         } else
  61                 ptes = (end + PAGE_SIZE - 1) >> PAGE_SHIFT;
  62
  63         tables += roundup(ptes * sizeof(pte_t), PAGE_SIZE);
  64
  65 #ifdef CONFIG_X86_32
  66         /* for fixmap */
  67         tables += roundup(__end_of_fixed_addresses * sizeof(pte_t), PAGE_SIZE);
  68 #endif
  69
  70         /*
  71          * RED-PEN putting page tables only on node 0 could
  72          * cause a hotspot and fill up ZONE_DMA. The page tables
  73          * need roughly 0.5KB per GB.
  74          */
  75 #ifdef CONFIG_X86_32
  76         start = 0x7000;
  77 #else
  78         start = 0x8000;
  79 #endif
  80         base = memblock_find_in_range(start, max_pfn_mapped<<PAGE_SHIFT,
  81                                         tables, PAGE_SIZE);
  82         if (base == MEMBLOCK_ERROR)
  83                 panic("Cannot find space for the kernel page tables");
  84
  85         e820_table_start = base >> PAGE_SHIFT;
  86         e820_table_end = e820_table_start;
  87         e820_table_top = e820_table_start + (tables >> PAGE_SHIFT);
  88
  89         printk(KERN_DEBUG "kernel direct mapping tables up to %lx @ %lx-%lx\n",
  90                 end, e820_table_start << PAGE_SHIFT, e820_table_top << PAGE_SHIFT);
  91 }
  92
  93 struct map_range {
  94         unsigned long start;
  95         unsigned long end;
  96         unsigned page_size_mask;
  97 };
  98
  99 #ifdef CONFIG_X86_32
 100 #define NR_RANGE_MR 3
 101 #else /* CONFIG_X86_64 */
 102 #define NR_RANGE_MR 5
 103 #endif
 104
 105 static int __meminit save_mr(struct map_range *mr, int nr_range,
 106                              unsigned long start_pfn, unsigned long end_pfn,
 107                              unsigned long page_size_mask)
 108 {
 109         if (start_pfn < end_pfn) {
 110                 if (nr_range >= NR_RANGE_MR)
 111                         panic("run out of range for init_memory_mapping\n");
 112                 mr[nr_range].start = start_pfn<<PAGE_SHIFT;
 113                 mr[nr_range].end   = end_pfn<<PAGE_SHIFT;
 114                 mr[nr_range].page_size_mask = page_size_mask;
 115                 nr_range++;
 116         }
 117
 118         return nr_range;
 119 }
 120
 121 /*
 122  * Setup the direct mapping of the physical memory at PAGE_OFFSET.
 123  * This runs before bootmem is initialized and gets pages directly from
 124  * the physical memory. To access them they are temporarily mapped.
 125  */
 126 unsigned long __init_refok init_memory_mapping(unsigned long start,
 127                                                unsigned long end)
 128 {
 129         unsigned long page_size_mask = 0;
 130         unsigned long start_pfn, end_pfn;
 131         unsigned long ret = 0;
 132         unsigned long pos;
 133
 134         struct map_range mr[NR_RANGE_MR];
 135         int nr_range, i;
 136         int use_pse, use_gbpages;
 137
 138         printk(KERN_INFO "init_memory_mapping: %016lx-%016lx\n", start, end);
 139
 140 #if defined(CONFIG_DEBUG_PAGEALLOC) || defined(CONFIG_KMEMCHECK)
 141         /*
 142          * For CONFIG_DEBUG_PAGEALLOC, identity mapping will use small pages.
 143          * This will simplify cpa(), which otherwise needs to support splitting
 144          * large pages into small in interrupt context, etc.
 145          */
 146         use_pse = use_gbpages = 0;
 147 #else
 148         use_pse = cpu_has_pse;
 149         use_gbpages = direct_gbpages;
 150 #endif
 151
 152         /* Enable PSE if available */
 153         if (cpu_has_pse)
 154                 set_in_cr4(X86_CR4_PSE);
 155
 156         /* Enable PGE if available */
 157         if (cpu_has_pge) {
 158                 set_in_cr4(X86_CR4_PGE);
 159                 __supported_pte_mask |= _PAGE_GLOBAL;
 160         }
 161
 162         if (use_gbpages)
 163                 page_size_mask |= 1 << PG_LEVEL_1G;
 164         if (use_pse)
 165                 page_size_mask |= 1 << PG_LEVEL_2M;
 166
 167         memset(mr, 0, sizeof(mr));
 168         nr_range = 0;
 169
 170         /* head if not big page alignment ? */
 171         start_pfn = start >> PAGE_SHIFT;
 172         pos = start_pfn << PAGE_SHIFT;
 173 #ifdef CONFIG_X86_32
 174         /*
 175          * Don't use a large page for the first 2/4MB of memory
 176          * because there are often fixed size MTRRs in there
 177          * and overlapping MTRRs into large pages can cause
 178          * slowdowns.
 179          */
 180         if (pos == 0)
 181                 end_pfn = 1<<(PMD_SHIFT - PAGE_SHIFT);
 182         else
 183                 end_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
 184                                  << (PMD_SHIFT - PAGE_SHIFT);
 185 #else /* CONFIG_X86_64 */
 186         end_pfn = ((pos + (PMD_SIZE - 1)) >> PMD_SHIFT)
 187                         << (PMD_SHIFT - PAGE_SHIFT);
 188 #endif
 189         if (end_pfn > (end >> PAGE_SHIFT))
 190                 end_pfn = end >> PAGE_SHIFT;
 191         if (start_pfn < end_pfn) {
 192                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
 193                 pos = end_pfn << PAGE_SHIFT;
 194         }
 195
 196         /* big page (2M) range */
 197         start_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
 198                          << (PMD_SHIFT - PAGE_SHIFT);
 199 #ifdef CONFIG_X86_32
 200         end_pfn = (end>>PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
 201 #else /* CONFIG_X86_64 */
 202         end_pfn = ((pos + (PUD_SIZE - 1))>>PUD_SHIFT)
 203                          << (PUD_SHIFT - PAGE_SHIFT);
 204         if (end_pfn > ((end>>PMD_SHIFT)<<(PMD_SHIFT - PAGE_SHIFT)))
 205                 end_pfn = ((end>>PMD_SHIFT)<<(PMD_SHIFT - PAGE_SHIFT));
 206 #endif
 207
 208         if (start_pfn < end_pfn) {
 209                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 210                                 page_size_mask & (1<<PG_LEVEL_2M));
 211                 pos = end_pfn << PAGE_SHIFT;
 212         }
 213
 214 #ifdef CONFIG_X86_64
 215         /* big page (1G) range */
 216         start_pfn = ((pos + (PUD_SIZE - 1))>>PUD_SHIFT)
 217                          << (PUD_SHIFT - PAGE_SHIFT);
 218         end_pfn = (end >> PUD_SHIFT) << (PUD_SHIFT - PAGE_SHIFT);
 219         if (start_pfn < end_pfn) {
 220                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 221                                 page_size_mask &
 222                                  ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G)));
 223                 pos = end_pfn << PAGE_SHIFT;
 224         }
 225
 226         /* tail is not big page (1G) alignment */
 227         start_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
 228                          << (PMD_SHIFT - PAGE_SHIFT);
 229         end_pfn = (end >> PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
 230         if (start_pfn < end_pfn) {
 231                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 232                                 page_size_mask & (1<<PG_LEVEL_2M));
 233                 pos = end_pfn << PAGE_SHIFT;
 234         }
 235 #endif
 236
 237         /* tail is not big page (2M) alignment */
 238         start_pfn = pos>>PAGE_SHIFT;
 239         end_pfn = end>>PAGE_SHIFT;
 240         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
 241
 242         /* try to merge same page size and continuous */
 243         for (i = 0; nr_range > 1 && i < nr_range - 1; i++) {
 244                 unsigned long old_start;
 245                 if (mr[i].end != mr[i+1].start ||
 246                     mr[i].page_size_mask != mr[i+1].page_size_mask)
 247                         continue;
 248                 /* move it */
 249                 old_start = mr[i].start;
 250                 memmove(&mr[i], &mr[i+1],
 251                         (nr_range - 1 - i) * sizeof(struct map_range));
 252                 mr[i--].start = old_start;
 253                 nr_range--;
 254         }
 255
 256         for (i = 0; i < nr_range; i++)
 257                 printk(KERN_DEBUG " %010lx - %010lx page %s\n",
 258                                 mr[i].start, mr[i].end,
 259                         (mr[i].page_size_mask & (1<<PG_LEVEL_1G))?"1G":(
 260                          (mr[i].page_size_mask & (1<<PG_LEVEL_2M))?"2M":"4k"));
 261
 262         /*
 263          * Find space for the kernel direct mapping tables.
 264          *
 265          * Later we should allocate these tables in the local node of the
 266          * memory mapped. Unfortunately this is done currently before the
 267          * nodes are discovered.
 268          */
 269         if (!after_bootmem)
 270                 find_early_table_space(end, use_pse, use_gbpages);
 271
 272         for (i = 0; i < nr_range; i++)
 273                 ret = kernel_physical_mapping_init(mr[i].start, mr[i].end,
 274                                                    mr[i].page_size_mask);
 275
 276 #ifdef CONFIG_X86_32
 277         early_ioremap_page_table_range_init();
 278
 279         load_cr3(swapper_pg_dir);
 280 #endif
 281
 282 #ifdef CONFIG_X86_64
 283         if (!after_bootmem && !start) {
 284                 pud_t *pud;
 285                 pmd_t *pmd;
 286
 287                 mmu_cr4_features = read_cr4();
 288
 289                 /*
 290                  * _brk_end cannot change anymore, but it and _end may be
 291                  * located on different 2M pages. cleanup_highmap(), however,
 292                  * can only consider _end when it runs, so destroy any
 293                  * mappings beyond _brk_end here.
 294                  */
 295                 pud = pud_offset(pgd_offset_k(_brk_end), _brk_end);
 296                 pmd = pmd_offset(pud, _brk_end - 1);
 297                 while (++pmd <= pmd_offset(pud, (unsigned long)_end - 1))
 298                         pmd_clear(pmd);
 299         }
 300 #endif
 301         __flush_tlb_all();
 302
 303         if (!after_bootmem && e820_table_end > e820_table_start)
 304                 memblock_x86_reserve_range(e820_table_start << PAGE_SHIFT,
 305                                  e820_table_end << PAGE_SHIFT, "PGTABLE");
 306
 307         if (!after_bootmem)
 308                 early_memtest(start, end);
 309
 310         return ret >> PAGE_SHIFT;
 311 }
 312
 313
 314 /*
 315  * devmem_is_allowed() checks to see if /dev/mem access to a certain address
 316  * is valid. The argument is a physical page number.
 317  *
 318  *
 319  * On x86, access has to be given to the first megabyte of ram because that area
 320  * contains bios code and data regions used by X and dosemu and similar apps.
 321  * Access has to be given to non-kernel-ram areas as well, these contain the PCI
 322  * mmio resources as well as potential bios/acpi data regions.
 323  */
 324 int devmem_is_allowed(unsigned long pagenr)
 325 {
 326         if (pagenr <= 256)
 327                 return 1;
 328         if (iomem_is_exclusive(pagenr << PAGE_SHIFT))
 329                 return 0;
 330         if (!page_is_ram(pagenr))
 331                 return 1;
 332         return 0;
 333 }
 334
 335 void free_init_pages(char *what, unsigned long begin, unsigned long end)
 336 {
 337         unsigned long addr;
 338         unsigned long begin_aligned, end_aligned;
 339
 340         /* Make sure boundaries are page aligned */
 341         begin_aligned = PAGE_ALIGN(begin);
 342         end_aligned   = end & PAGE_MASK;
 343
 344         if (WARN_ON(begin_aligned != begin || end_aligned != end)) {
 345                 begin = begin_aligned;
 346                 end   = end_aligned;
 347         }
 348
 349         if (begin >= end)
 350                 return;
 351
 352         addr = begin;
 353
 354         /*
 355          * If debugging page accesses then do not free this memory but
 356          * mark them not present - any buggy init-section access will
 357          * create a kernel page fault:
 358          */
 359 #ifdef CONFIG_DEBUG_PAGEALLOC
 360         printk(KERN_INFO "debug: unmapping init memory %08lx..%08lx\n",
 361                 begin, end);
 362         set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
 363 #else
 364         /*
 365          * We just marked the kernel text read only above, now that
 366          * we are going to free part of that, we need to make that
 367          * writeable and non-executable first.
 368          */
 369         set_memory_nx(begin, (end - begin) >> PAGE_SHIFT);
 370         set_memory_rw(begin, (end - begin) >> PAGE_SHIFT);
 371
 372         printk(KERN_INFO "Freeing %s: %luk freed\n", what, (end - begin) >> 10);
 373
 374         for (; addr < end; addr += PAGE_SIZE) {
 375                 ClearPageReserved(virt_to_page(addr));
 376                 init_page_count(virt_to_page(addr));
 377                 memset((void *)addr, POISON_FREE_INITMEM, PAGE_SIZE);
 378                 free_page(addr);
 379                 totalram_pages++;
 380         }
 381 #endif
 382 }
 383
 384 void free_initmem(void)
 385 {
 386         free_init_pages("unused kernel memory",
 387                         (unsigned long)(&__init_begin),
 388                         (unsigned long)(&__init_end));
 389 }
 390
 391 #ifdef CONFIG_BLK_DEV_INITRD
 392 void free_initrd_mem(unsigned long start, unsigned long end)
 393 {
 394         /*
 395          * end could be not aligned, and We can not align that,
 396          * decompresser could be confused by aligned initrd_end
 397          * We already reserve the end partial page before in
 398          *   - i386_start_kernel()
 399          *   - x86_64_start_kernel()
 400          *   - relocate_initrd()
 401          * So here We can do PAGE_ALIGN() safely to get partial page to be freed
 402          */
 403         free_init_pages("initrd memory", start, PAGE_ALIGN(end));
 404 }
 405 #endif