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 #include <linux/bootmem.h>      /* for max_low_pfn */
   7
   8 #include <asm/cacheflush.h>
   9 #include <asm/e820.h>
  10 #include <asm/init.h>
  11 #include <asm/page.h>
  12 #include <asm/page_types.h>
  13 #include <asm/sections.h>
  14 #include <asm/setup.h>
  15 #include <asm/tlbflush.h>
  16 #include <asm/tlb.h>
  17 #include <asm/proto.h>
  18 #include <asm/dma.h>            /* for MAX_DMA_PFN */
  19
  20 unsigned long __initdata pgt_buf_start;
  21 unsigned long __meminitdata pgt_buf_end;
  22 unsigned long __meminitdata pgt_buf_top;
  23
  24 int after_bootmem;
  25
  26 int direct_gbpages
  27 #ifdef CONFIG_DIRECT_GBPAGES
  28                                 = 1
  29 #endif
  30 ;
  31
  32 struct map_range {
  33         unsigned long start;
  34         unsigned long end;
  35         unsigned page_size_mask;
  36 };
  37
  38 static void __init find_early_table_space(struct map_range *mr, unsigned long end,
  39                                           int use_pse, int use_gbpages)
  40 {
  41         unsigned long puds, pmds, ptes, tables, start = 0, good_end = end;
  42         phys_addr_t base;
  43
  44         puds = (end + PUD_SIZE - 1) >> PUD_SHIFT;
  45         tables = roundup(puds * sizeof(pud_t), PAGE_SIZE);
  46
  47         if (use_gbpages) {
  48                 unsigned long extra;
  49
  50                 extra = end - ((end>>PUD_SHIFT) << PUD_SHIFT);
  51                 pmds = (extra + PMD_SIZE - 1) >> PMD_SHIFT;
  52         } else
  53                 pmds = (end + PMD_SIZE - 1) >> PMD_SHIFT;
  54
  55         tables += roundup(pmds * sizeof(pmd_t), PAGE_SIZE);
  56
  57         if (use_pse) {
  58                 unsigned long extra;
  59
  60                 extra = end - ((end>>PMD_SHIFT) << PMD_SHIFT);
  61 #ifdef CONFIG_X86_32
  62                 extra += PMD_SIZE;
  63 #endif
  64                 /* The first 2/4M doesn't use large pages. */
  65                 extra += mr->end - mr->start;
  66
  67                 ptes = (extra + PAGE_SIZE - 1) >> PAGE_SHIFT;
  68         } else
  69                 ptes = (end + PAGE_SIZE - 1) >> PAGE_SHIFT;
  70
  71         tables += roundup(ptes * sizeof(pte_t), PAGE_SIZE);
  72
  73 #ifdef CONFIG_X86_32
  74         /* for fixmap */
  75         tables += roundup(__end_of_fixed_addresses * sizeof(pte_t), PAGE_SIZE);
  76 #endif
  77         good_end = max_pfn_mapped << PAGE_SHIFT;
  78
  79         base = memblock_find_in_range(start, good_end, tables, PAGE_SIZE);
  80         if (!base)
  81                 panic("Cannot find space for the kernel page tables");
  82
  83         pgt_buf_start = base >> PAGE_SHIFT;
  84         pgt_buf_end = pgt_buf_start;
  85         pgt_buf_top = pgt_buf_start + (tables >> PAGE_SHIFT);
  86
  87         printk(KERN_DEBUG "kernel direct mapping tables up to %lx @ %lx-%lx\n",
  88                 end, pgt_buf_start << PAGE_SHIFT, pgt_buf_top << PAGE_SHIFT);
  89 }
  90
  91 void __init native_pagetable_reserve(u64 start, u64 end)
  92 {
  93         memblock_reserve(start, end - start);
  94 }
  95
  96 #ifdef CONFIG_X86_32
  97 #define NR_RANGE_MR 3
  98 #else /* CONFIG_X86_64 */
  99 #define NR_RANGE_MR 5
 100 #endif
 101
 102 static int __meminit save_mr(struct map_range *mr, int nr_range,
 103                              unsigned long start_pfn, unsigned long end_pfn,
 104                              unsigned long page_size_mask)
 105 {
 106         if (start_pfn < end_pfn) {
 107                 if (nr_range >= NR_RANGE_MR)
 108                         panic("run out of range for init_memory_mapping\n");
 109                 mr[nr_range].start = start_pfn<<PAGE_SHIFT;
 110                 mr[nr_range].end   = end_pfn<<PAGE_SHIFT;
 111                 mr[nr_range].page_size_mask = page_size_mask;
 112                 nr_range++;
 113         }
 114
 115         return nr_range;
 116 }
 117
 118 /*
 119  * Setup the direct mapping of the physical memory at PAGE_OFFSET.
 120  * This runs before bootmem is initialized and gets pages directly from
 121  * the physical memory. To access them they are temporarily mapped.
 122  */
 123 unsigned long __init_refok init_memory_mapping(unsigned long start,
 124                                                unsigned long end)
 125 {
 126         unsigned long page_size_mask = 0;
 127         unsigned long start_pfn, end_pfn;
 128         unsigned long ret = 0;
 129         unsigned long pos;
 130
 131         struct map_range mr[NR_RANGE_MR];
 132         int nr_range, i;
 133         int use_pse, use_gbpages;
 134
 135         printk(KERN_INFO "init_memory_mapping: %016lx-%016lx\n", start, end);
 136
 137 #if defined(CONFIG_DEBUG_PAGEALLOC) || defined(CONFIG_KMEMCHECK)
 138         /*
 139          * For CONFIG_DEBUG_PAGEALLOC, identity mapping will use small pages.
 140          * This will simplify cpa(), which otherwise needs to support splitting
 141          * large pages into small in interrupt context, etc.
 142          */
 143         use_pse = use_gbpages = 0;
 144 #else
 145         use_pse = cpu_has_pse;
 146         use_gbpages = direct_gbpages;
 147 #endif
 148
 149         /* Enable PSE if available */
 150         if (cpu_has_pse)
 151                 set_in_cr4(X86_CR4_PSE);
 152
 153         /* Enable PGE if available */
 154         if (cpu_has_pge) {
 155                 set_in_cr4(X86_CR4_PGE);
 156                 __supported_pte_mask |= _PAGE_GLOBAL;
 157         }
 158
 159         if (use_gbpages)
 160                 page_size_mask |= 1 << PG_LEVEL_1G;
 161         if (use_pse)
 162                 page_size_mask |= 1 << PG_LEVEL_2M;
 163
 164         memset(mr, 0, sizeof(mr));
 165         nr_range = 0;
 166
 167         /* head if not big page alignment ? */
 168         start_pfn = start >> PAGE_SHIFT;
 169         pos = start_pfn << PAGE_SHIFT;
 170 #ifdef CONFIG_X86_32
 171         /*
 172          * Don't use a large page for the first 2/4MB of memory
 173          * because there are often fixed size MTRRs in there
 174          * and overlapping MTRRs into large pages can cause
 175          * slowdowns.
 176          */
 177         if (pos == 0)
 178                 end_pfn = 1<<(PMD_SHIFT - PAGE_SHIFT);
 179         else
 180                 end_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
 181                                  << (PMD_SHIFT - PAGE_SHIFT);
 182 #else /* CONFIG_X86_64 */
 183         end_pfn = ((pos + (PMD_SIZE - 1)) >> PMD_SHIFT)
 184                         << (PMD_SHIFT - PAGE_SHIFT);
 185 #endif
 186         if (end_pfn > (end >> PAGE_SHIFT))
 187                 end_pfn = end >> PAGE_SHIFT;
 188         if (start_pfn < end_pfn) {
 189                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
 190                 pos = end_pfn << PAGE_SHIFT;
 191         }
 192
 193         /* big page (2M) range */
 194         start_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
 195                          << (PMD_SHIFT - PAGE_SHIFT);
 196 #ifdef CONFIG_X86_32
 197         end_pfn = (end>>PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
 198 #else /* CONFIG_X86_64 */
 199         end_pfn = ((pos + (PUD_SIZE - 1))>>PUD_SHIFT)
 200                          << (PUD_SHIFT - PAGE_SHIFT);
 201         if (end_pfn > ((end>>PMD_SHIFT)<<(PMD_SHIFT - PAGE_SHIFT)))
 202                 end_pfn = ((end>>PMD_SHIFT)<<(PMD_SHIFT - PAGE_SHIFT));
 203 #endif
 204
 205         if (start_pfn < end_pfn) {
 206                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 207                                 page_size_mask & (1<<PG_LEVEL_2M));
 208                 pos = end_pfn << PAGE_SHIFT;
 209         }
 210
 211 #ifdef CONFIG_X86_64
 212         /* big page (1G) range */
 213         start_pfn = ((pos + (PUD_SIZE - 1))>>PUD_SHIFT)
 214                          << (PUD_SHIFT - PAGE_SHIFT);
 215         end_pfn = (end >> PUD_SHIFT) << (PUD_SHIFT - PAGE_SHIFT);
 216         if (start_pfn < end_pfn) {
 217                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 218                                 page_size_mask &
 219                                  ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G)));
 220                 pos = end_pfn << PAGE_SHIFT;
 221         }
 222
 223         /* tail is not big page (1G) alignment */
 224         start_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
 225                          << (PMD_SHIFT - PAGE_SHIFT);
 226         end_pfn = (end >> PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
 227         if (start_pfn < end_pfn) {
 228                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 229                                 page_size_mask & (1<<PG_LEVEL_2M));
 230                 pos = end_pfn << PAGE_SHIFT;
 231         }
 232 #endif
 233
 234         /* tail is not big page (2M) alignment */
 235         start_pfn = pos>>PAGE_SHIFT;
 236         end_pfn = end>>PAGE_SHIFT;
 237         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
 238
 239         /* try to merge same page size and continuous */
 240         for (i = 0; nr_range > 1 && i < nr_range - 1; i++) {
 241                 unsigned long old_start;
 242                 if (mr[i].end != mr[i+1].start ||
 243                     mr[i].page_size_mask != mr[i+1].page_size_mask)
 244                         continue;
 245                 /* move it */
 246                 old_start = mr[i].start;
 247                 memmove(&mr[i], &mr[i+1],
 248                         (nr_range - 1 - i) * sizeof(struct map_range));
 249                 mr[i--].start = old_start;
 250                 nr_range--;
 251         }
 252
 253         for (i = 0; i < nr_range; i++)
 254                 printk(KERN_DEBUG " %010lx - %010lx page %s\n",
 255                                 mr[i].start, mr[i].end,
 256                         (mr[i].page_size_mask & (1<<PG_LEVEL_1G))?"1G":(
 257                          (mr[i].page_size_mask & (1<<PG_LEVEL_2M))?"2M":"4k"));
 258
 259         /*
 260          * Find space for the kernel direct mapping tables.
 261          *
 262          * Later we should allocate these tables in the local node of the
 263          * memory mapped. Unfortunately this is done currently before the
 264          * nodes are discovered.
 265          */
 266         if (!after_bootmem)
 267                 find_early_table_space(&mr[0], end, use_pse, use_gbpages);
 268
 269         for (i = 0; i < nr_range; i++)
 270                 ret = kernel_physical_mapping_init(mr[i].start, mr[i].end,
 271                                                    mr[i].page_size_mask);
 272
 273 #ifdef CONFIG_X86_32
 274         early_ioremap_page_table_range_init();
 275
 276         load_cr3(swapper_pg_dir);
 277 #endif
 278
 279         __flush_tlb_all();
 280
 281         /*
 282          * Reserve the kernel pagetable pages we used (pgt_buf_start -
 283          * pgt_buf_end) and free the other ones (pgt_buf_end - pgt_buf_top)
 284          * so that they can be reused for other purposes.
 285          *
 286          * On native it just means calling memblock_reserve, on Xen it also
 287          * means marking RW the pagetable pages that we allocated before
 288          * but that haven't been used.
 289          *
 290          * In fact on xen we mark RO the whole range pgt_buf_start -
 291          * pgt_buf_top, because we have to make sure that when
 292          * init_memory_mapping reaches the pagetable pages area, it maps
 293          * RO all the pagetable pages, including the ones that are beyond
 294          * pgt_buf_end at that time.
 295          */
 296         if (!after_bootmem && pgt_buf_end > pgt_buf_start)
 297                 x86_init.mapping.pagetable_reserve(PFN_PHYS(pgt_buf_start),
 298                                 PFN_PHYS(pgt_buf_end));
 299
 300         if (!after_bootmem)
 301                 early_memtest(start, end);
 302
 303         return ret >> PAGE_SHIFT;
 304 }
 305
 306
 307 /*
 308  * devmem_is_allowed() checks to see if /dev/mem access to a certain address
 309  * is valid. The argument is a physical page number.
 310  *
 311  *
 312  * On x86, access has to be given to the first megabyte of ram because that area
 313  * contains bios code and data regions used by X and dosemu and similar apps.
 314  * Access has to be given to non-kernel-ram areas as well, these contain the PCI
 315  * mmio resources as well as potential bios/acpi data regions.
 316  */
 317 int devmem_is_allowed(unsigned long pagenr)
 318 {
 319         if (pagenr <= 256)
 320                 return 1;
 321         if (iomem_is_exclusive(pagenr << PAGE_SHIFT))
 322                 return 0;
 323         if (!page_is_ram(pagenr))
 324                 return 1;
 325         return 0;
 326 }
 327
 328 void free_init_pages(char *what, unsigned long begin, unsigned long end)
 329 {
 330         unsigned long addr;
 331         unsigned long begin_aligned, end_aligned;
 332
 333         /* Make sure boundaries are page aligned */
 334         begin_aligned = PAGE_ALIGN(begin);
 335         end_aligned   = end & PAGE_MASK;
 336
 337         if (WARN_ON(begin_aligned != begin || end_aligned != end)) {
 338                 begin = begin_aligned;
 339                 end   = end_aligned;
 340         }
 341
 342         if (begin >= end)
 343                 return;
 344
 345         addr = begin;
 346
 347         /*
 348          * If debugging page accesses then do not free this memory but
 349          * mark them not present - any buggy init-section access will
 350          * create a kernel page fault:
 351          */
 352 #ifdef CONFIG_DEBUG_PAGEALLOC
 353         printk(KERN_INFO "debug: unmapping init memory %08lx..%08lx\n",
 354                 begin, end);
 355         set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
 356 #else
 357         /*
 358          * We just marked the kernel text read only above, now that
 359          * we are going to free part of that, we need to make that
 360          * writeable and non-executable first.
 361          */
 362         set_memory_nx(begin, (end - begin) >> PAGE_SHIFT);
 363         set_memory_rw(begin, (end - begin) >> PAGE_SHIFT);
 364
 365         printk(KERN_INFO "Freeing %s: %luk freed\n", what, (end - begin) >> 10);
 366
 367         for (; addr < end; addr += PAGE_SIZE) {
 368                 ClearPageReserved(virt_to_page(addr));
 369                 init_page_count(virt_to_page(addr));
 370                 memset((void *)addr, POISON_FREE_INITMEM, PAGE_SIZE);
 371                 free_page(addr);
 372                 totalram_pages++;
 373         }
 374 #endif
 375 }
 376
 377 void free_initmem(void)
 378 {
 379         free_init_pages("unused kernel memory",
 380                         (unsigned long)(&__init_begin),
 381                         (unsigned long)(&__init_end));
 382 }
 383
 384 #ifdef CONFIG_BLK_DEV_INITRD
 385 void free_initrd_mem(unsigned long start, unsigned long end)
 386 {
 387         /*
 388          * end could be not aligned, and We can not align that,
 389          * decompresser could be confused by aligned initrd_end
 390          * We already reserve the end partial page before in
 391          *   - i386_start_kernel()
 392          *   - x86_64_start_kernel()
 393          *   - relocate_initrd()
 394          * So here We can do PAGE_ALIGN() safely to get partial page to be freed
 395          */
 396         free_init_pages("initrd memory", start, PAGE_ALIGN(end));
 397 }
 398 #endif
 399
 400 void __init zone_sizes_init(void)
 401 {
 402         unsigned long max_zone_pfns[MAX_NR_ZONES];
 403
 404         memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
 405
 406 #ifdef CONFIG_ZONE_DMA
 407         max_zone_pfns[ZONE_DMA]         = MAX_DMA_PFN;
 408 #endif
 409 #ifdef CONFIG_ZONE_DMA32
 410         max_zone_pfns[ZONE_DMA32]       = MAX_DMA32_PFN;
 411 #endif
 412         max_zone_pfns[ZONE_NORMAL]      = max_low_pfn;
 413 #ifdef CONFIG_HIGHMEM
 414         max_zone_pfns[ZONE_HIGHMEM]     = max_pfn;
 415 #endif
 416
 417         free_area_init_nodes(max_zone_pfns);
 418 }
 419