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 int page_size_mask;
  39
  40 static void __init probe_page_size_mask(void)
  41 {
  42 #if !defined(CONFIG_DEBUG_PAGEALLOC) && !defined(CONFIG_KMEMCHECK)
  43         /*
  44          * For CONFIG_DEBUG_PAGEALLOC, identity mapping will use small pages.
  45          * This will simplify cpa(), which otherwise needs to support splitting
  46          * large pages into small in interrupt context, etc.
  47          */
  48         if (direct_gbpages)
  49                 page_size_mask |= 1 << PG_LEVEL_1G;
  50         if (cpu_has_pse)
  51                 page_size_mask |= 1 << PG_LEVEL_2M;
  52 #endif
  53
  54         /* Enable PSE if available */
  55         if (cpu_has_pse)
  56                 set_in_cr4(X86_CR4_PSE);
  57
  58         /* Enable PGE if available */
  59         if (cpu_has_pge) {
  60                 set_in_cr4(X86_CR4_PGE);
  61                 __supported_pte_mask |= _PAGE_GLOBAL;
  62         }
  63 }
  64 void __init native_pagetable_reserve(u64 start, u64 end)
  65 {
  66         memblock_reserve(start, end - start);
  67 }
  68
  69 #ifdef CONFIG_X86_32
  70 #define NR_RANGE_MR 3
  71 #else /* CONFIG_X86_64 */
  72 #define NR_RANGE_MR 5
  73 #endif
  74
  75 static int __meminit save_mr(struct map_range *mr, int nr_range,
  76                              unsigned long start_pfn, unsigned long end_pfn,
  77                              unsigned long page_size_mask)
  78 {
  79         if (start_pfn < end_pfn) {
  80                 if (nr_range >= NR_RANGE_MR)
  81                         panic("run out of range for init_memory_mapping\n");
  82                 mr[nr_range].start = start_pfn<<PAGE_SHIFT;
  83                 mr[nr_range].end   = end_pfn<<PAGE_SHIFT;
  84                 mr[nr_range].page_size_mask = page_size_mask;
  85                 nr_range++;
  86         }
  87
  88         return nr_range;
  89 }
  90
  91 static int __meminit split_mem_range(struct map_range *mr, int nr_range,
  92                                      unsigned long start,
  93                                      unsigned long end)
  94 {
  95         unsigned long start_pfn, end_pfn;
  96         unsigned long pos;
  97         int i;
  98
  99         /* head if not big page alignment ? */
 100         start_pfn = start >> PAGE_SHIFT;
 101         pos = start_pfn << PAGE_SHIFT;
 102 #ifdef CONFIG_X86_32
 103         /*
 104          * Don't use a large page for the first 2/4MB of memory
 105          * because there are often fixed size MTRRs in there
 106          * and overlapping MTRRs into large pages can cause
 107          * slowdowns.
 108          */
 109         if (pos == 0)
 110                 end_pfn = 1<<(PMD_SHIFT - PAGE_SHIFT);
 111         else
 112                 end_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
 113                                  << (PMD_SHIFT - PAGE_SHIFT);
 114 #else /* CONFIG_X86_64 */
 115         end_pfn = ((pos + (PMD_SIZE - 1)) >> PMD_SHIFT)
 116                         << (PMD_SHIFT - PAGE_SHIFT);
 117 #endif
 118         if (end_pfn > (end >> PAGE_SHIFT))
 119                 end_pfn = end >> PAGE_SHIFT;
 120         if (start_pfn < end_pfn) {
 121                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
 122                 pos = end_pfn << PAGE_SHIFT;
 123         }
 124
 125         /* big page (2M) range */
 126         start_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
 127                          << (PMD_SHIFT - PAGE_SHIFT);
 128 #ifdef CONFIG_X86_32
 129         end_pfn = (end>>PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
 130 #else /* CONFIG_X86_64 */
 131         end_pfn = ((pos + (PUD_SIZE - 1))>>PUD_SHIFT)
 132                          << (PUD_SHIFT - PAGE_SHIFT);
 133         if (end_pfn > ((end>>PMD_SHIFT)<<(PMD_SHIFT - PAGE_SHIFT)))
 134                 end_pfn = ((end>>PMD_SHIFT)<<(PMD_SHIFT - PAGE_SHIFT));
 135 #endif
 136
 137         if (start_pfn < end_pfn) {
 138                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 139                                 page_size_mask & (1<<PG_LEVEL_2M));
 140                 pos = end_pfn << PAGE_SHIFT;
 141         }
 142
 143 #ifdef CONFIG_X86_64
 144         /* big page (1G) range */
 145         start_pfn = ((pos + (PUD_SIZE - 1))>>PUD_SHIFT)
 146                          << (PUD_SHIFT - PAGE_SHIFT);
 147         end_pfn = (end >> PUD_SHIFT) << (PUD_SHIFT - PAGE_SHIFT);
 148         if (start_pfn < end_pfn) {
 149                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 150                                 page_size_mask &
 151                                  ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G)));
 152                 pos = end_pfn << PAGE_SHIFT;
 153         }
 154
 155         /* tail is not big page (1G) alignment */
 156         start_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
 157                          << (PMD_SHIFT - PAGE_SHIFT);
 158         end_pfn = (end >> PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
 159         if (start_pfn < end_pfn) {
 160                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 161                                 page_size_mask & (1<<PG_LEVEL_2M));
 162                 pos = end_pfn << PAGE_SHIFT;
 163         }
 164 #endif
 165
 166         /* tail is not big page (2M) alignment */
 167         start_pfn = pos>>PAGE_SHIFT;
 168         end_pfn = end>>PAGE_SHIFT;
 169         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
 170
 171         /* try to merge same page size and continuous */
 172         for (i = 0; nr_range > 1 && i < nr_range - 1; i++) {
 173                 unsigned long old_start;
 174                 if (mr[i].end != mr[i+1].start ||
 175                     mr[i].page_size_mask != mr[i+1].page_size_mask)
 176                         continue;
 177                 /* move it */
 178                 old_start = mr[i].start;
 179                 memmove(&mr[i], &mr[i+1],
 180                         (nr_range - 1 - i) * sizeof(struct map_range));
 181                 mr[i--].start = old_start;
 182                 nr_range--;
 183         }
 184
 185         for (i = 0; i < nr_range; i++)
 186                 printk(KERN_DEBUG " [mem %#010lx-%#010lx] page %s\n",
 187                                 mr[i].start, mr[i].end - 1,
 188                         (mr[i].page_size_mask & (1<<PG_LEVEL_1G))?"1G":(
 189                          (mr[i].page_size_mask & (1<<PG_LEVEL_2M))?"2M":"4k"));
 190
 191         return nr_range;
 192 }
 193
 194 /*
 195  * First calculate space needed for kernel direct mapping page tables to cover
 196  * mr[0].start to mr[nr_range - 1].end, while accounting for possible 2M and 1GB
 197  * pages. Then find enough contiguous space for those page tables.
 198  */
 199 static void __init find_early_table_space(struct map_range *mr, int nr_range)
 200 {
 201         int i;
 202         unsigned long puds = 0, pmds = 0, ptes = 0, tables;
 203         unsigned long start = 0, good_end;
 204         phys_addr_t base;
 205
 206         for (i = 0; i < nr_range; i++) {
 207                 unsigned long range, extra;
 208
 209                 range = mr[i].end - mr[i].start;
 210                 puds += (range + PUD_SIZE - 1) >> PUD_SHIFT;
 211
 212                 if (mr[i].page_size_mask & (1 << PG_LEVEL_1G)) {
 213                         extra = range - ((range >> PUD_SHIFT) << PUD_SHIFT);
 214                         pmds += (extra + PMD_SIZE - 1) >> PMD_SHIFT;
 215                 } else {
 216                         pmds += (range + PMD_SIZE - 1) >> PMD_SHIFT;
 217                 }
 218
 219                 if (mr[i].page_size_mask & (1 << PG_LEVEL_2M)) {
 220                         extra = range - ((range >> PMD_SHIFT) << PMD_SHIFT);
 221 #ifdef CONFIG_X86_32
 222                         extra += PMD_SIZE;
 223 #endif
 224                         ptes += (extra + PAGE_SIZE - 1) >> PAGE_SHIFT;
 225                 } else {
 226                         ptes += (range + PAGE_SIZE - 1) >> PAGE_SHIFT;
 227                 }
 228         }
 229
 230         tables = roundup(puds * sizeof(pud_t), PAGE_SIZE);
 231         tables += roundup(pmds * sizeof(pmd_t), PAGE_SIZE);
 232         tables += roundup(ptes * sizeof(pte_t), PAGE_SIZE);
 233
 234 #ifdef CONFIG_X86_32
 235         /* for fixmap */
 236         tables += roundup(__end_of_fixed_addresses * sizeof(pte_t), PAGE_SIZE);
 237 #endif
 238         good_end = max_pfn_mapped << PAGE_SHIFT;
 239
 240         base = memblock_find_in_range(start, good_end, tables, PAGE_SIZE);
 241         if (!base)
 242                 panic("Cannot find space for the kernel page tables");
 243
 244         pgt_buf_start = base >> PAGE_SHIFT;
 245         pgt_buf_end = pgt_buf_start;
 246         pgt_buf_top = pgt_buf_start + (tables >> PAGE_SHIFT);
 247
 248         printk(KERN_DEBUG "kernel direct mapping tables up to %#lx @ [mem %#010lx-%#010lx]\n",
 249                 mr[nr_range - 1].end - 1, pgt_buf_start << PAGE_SHIFT,
 250                 (pgt_buf_top << PAGE_SHIFT) - 1);
 251 }
 252
 253 /*
 254  * Setup the direct mapping of the physical memory at PAGE_OFFSET.
 255  * This runs before bootmem is initialized and gets pages directly from
 256  * the physical memory. To access them they are temporarily mapped.
 257  */
 258 unsigned long __init_refok init_memory_mapping(unsigned long start,
 259                                                unsigned long end)
 260 {
 261         struct map_range mr[NR_RANGE_MR];
 262         unsigned long ret = 0;
 263         int nr_range, i;
 264
 265         pr_info("init_memory_mapping: [mem %#010lx-%#010lx]\n",
 266                start, end - 1);
 267
 268         memset(mr, 0, sizeof(mr));
 269         nr_range = split_mem_range(mr, 0, start, end);
 270
 271         /*
 272          * Find space for the kernel direct mapping tables.
 273          *
 274          * Later we should allocate these tables in the local node of the
 275          * memory mapped. Unfortunately this is done currently before the
 276          * nodes are discovered.
 277          */
 278         if (!after_bootmem)
 279                 find_early_table_space(mr, nr_range);
 280
 281         for (i = 0; i < nr_range; i++)
 282                 ret = kernel_physical_mapping_init(mr[i].start, mr[i].end,
 283                                                    mr[i].page_size_mask);
 284
 285 #ifdef CONFIG_X86_32
 286         early_ioremap_page_table_range_init();
 287
 288         load_cr3(swapper_pg_dir);
 289 #endif
 290
 291         __flush_tlb_all();
 292
 293         /*
 294          * Reserve the kernel pagetable pages we used (pgt_buf_start -
 295          * pgt_buf_end) and free the other ones (pgt_buf_end - pgt_buf_top)
 296          * so that they can be reused for other purposes.
 297          *
 298          * On native it just means calling memblock_reserve, on Xen it also
 299          * means marking RW the pagetable pages that we allocated before
 300          * but that haven't been used.
 301          *
 302          * In fact on xen we mark RO the whole range pgt_buf_start -
 303          * pgt_buf_top, because we have to make sure that when
 304          * init_memory_mapping reaches the pagetable pages area, it maps
 305          * RO all the pagetable pages, including the ones that are beyond
 306          * pgt_buf_end at that time.
 307          */
 308         if (!after_bootmem && pgt_buf_end > pgt_buf_start)
 309                 x86_init.mapping.pagetable_reserve(PFN_PHYS(pgt_buf_start),
 310                                 PFN_PHYS(pgt_buf_end));
 311
 312         if (!after_bootmem)
 313                 early_memtest(start, end);
 314
 315         return ret >> PAGE_SHIFT;
 316 }
 317
 318 void __init init_mem_mapping(void)
 319 {
 320         probe_page_size_mask();
 321
 322         /* max_pfn_mapped is updated here */
 323         max_low_pfn_mapped = init_memory_mapping(0, max_low_pfn<<PAGE_SHIFT);
 324         max_pfn_mapped = max_low_pfn_mapped;
 325
 326 #ifdef CONFIG_X86_64
 327         if (max_pfn > max_low_pfn) {
 328                 max_pfn_mapped = init_memory_mapping(1UL<<32,
 329                                                      max_pfn<<PAGE_SHIFT);
 330                 /* can we preseve max_low_pfn ?*/
 331                 max_low_pfn = max_pfn;
 332         }
 333 #endif
 334 }
 335
 336 /*
 337  * devmem_is_allowed() checks to see if /dev/mem access to a certain address
 338  * is valid. The argument is a physical page number.
 339  *
 340  *
 341  * On x86, access has to be given to the first megabyte of ram because that area
 342  * contains bios code and data regions used by X and dosemu and similar apps.
 343  * Access has to be given to non-kernel-ram areas as well, these contain the PCI
 344  * mmio resources as well as potential bios/acpi data regions.
 345  */
 346 int devmem_is_allowed(unsigned long pagenr)
 347 {
 348         if (pagenr < 256)
 349                 return 1;
 350         if (iomem_is_exclusive(pagenr << PAGE_SHIFT))
 351                 return 0;
 352         if (!page_is_ram(pagenr))
 353                 return 1;
 354         return 0;
 355 }
 356
 357 void free_init_pages(char *what, unsigned long begin, unsigned long end)
 358 {
 359         unsigned long addr;
 360         unsigned long begin_aligned, end_aligned;
 361
 362         /* Make sure boundaries are page aligned */
 363         begin_aligned = PAGE_ALIGN(begin);
 364         end_aligned   = end & PAGE_MASK;
 365
 366         if (WARN_ON(begin_aligned != begin || end_aligned != end)) {
 367                 begin = begin_aligned;
 368                 end   = end_aligned;
 369         }
 370
 371         if (begin >= end)
 372                 return;
 373
 374         addr = begin;
 375
 376         /*
 377          * If debugging page accesses then do not free this memory but
 378          * mark them not present - any buggy init-section access will
 379          * create a kernel page fault:
 380          */
 381 #ifdef CONFIG_DEBUG_PAGEALLOC
 382         printk(KERN_INFO "debug: unmapping init [mem %#010lx-%#010lx]\n",
 383                 begin, end - 1);
 384         set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
 385 #else
 386         /*
 387          * We just marked the kernel text read only above, now that
 388          * we are going to free part of that, we need to make that
 389          * writeable and non-executable first.
 390          */
 391         set_memory_nx(begin, (end - begin) >> PAGE_SHIFT);
 392         set_memory_rw(begin, (end - begin) >> PAGE_SHIFT);
 393
 394         printk(KERN_INFO "Freeing %s: %luk freed\n", what, (end - begin) >> 10);
 395
 396         for (; addr < end; addr += PAGE_SIZE) {
 397                 ClearPageReserved(virt_to_page(addr));
 398                 init_page_count(virt_to_page(addr));
 399                 memset((void *)addr, POISON_FREE_INITMEM, PAGE_SIZE);
 400                 free_page(addr);
 401                 totalram_pages++;
 402         }
 403 #endif
 404 }
 405
 406 void free_initmem(void)
 407 {
 408         free_init_pages("unused kernel memory",
 409                         (unsigned long)(&__init_begin),
 410                         (unsigned long)(&__init_end));
 411 }
 412
 413 #ifdef CONFIG_BLK_DEV_INITRD
 414 void __init free_initrd_mem(unsigned long start, unsigned long end)
 415 {
 416         /*
 417          * end could be not aligned, and We can not align that,
 418          * decompresser could be confused by aligned initrd_end
 419          * We already reserve the end partial page before in
 420          *   - i386_start_kernel()
 421          *   - x86_64_start_kernel()
 422          *   - relocate_initrd()
 423          * So here We can do PAGE_ALIGN() safely to get partial page to be freed
 424          */
 425         free_init_pages("initrd memory", start, PAGE_ALIGN(end));
 426 }
 427 #endif
 428
 429 void __init zone_sizes_init(void)
 430 {
 431         unsigned long max_zone_pfns[MAX_NR_ZONES];
 432
 433         memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
 434
 435 #ifdef CONFIG_ZONE_DMA
 436         max_zone_pfns[ZONE_DMA]         = MAX_DMA_PFN;
 437 #endif
 438 #ifdef CONFIG_ZONE_DMA32
 439         max_zone_pfns[ZONE_DMA32]       = MAX_DMA32_PFN;
 440 #endif
 441         max_zone_pfns[ZONE_NORMAL]      = max_low_pfn;
 442 #ifdef CONFIG_HIGHMEM
 443         max_zone_pfns[ZONE_HIGHMEM]     = max_pfn;
 444 #endif
 445
 446         free_area_init_nodes(max_zone_pfns);
 447 }
 448