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