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 unsigned long __initdata pgt_buf_start;
  20 unsigned long __meminitdata pgt_buf_end;
  21 unsigned long __meminitdata pgt_buf_top;
  22
  23 int after_bootmem;
  24
  25 int direct_gbpages
  26 #ifdef CONFIG_DIRECT_GBPAGES
  27                                 = 1
  28 #endif
  29 ;
  30
  31 static void __init find_early_table_space(unsigned long end, int use_pse,
  32                                           int use_gbpages)
  33 {
  34         unsigned long puds, pmds, ptes, tables, start = 0, good_end = end;
  35         phys_addr_t base;
  36
  37         puds = (end + PUD_SIZE - 1) >> PUD_SHIFT;
  38         tables = roundup(puds * sizeof(pud_t), PAGE_SIZE);
  39
  40         if (use_gbpages) {
  41                 unsigned long extra;
  42
  43                 extra = end - ((end>>PUD_SHIFT) << PUD_SHIFT);
  44                 pmds = (extra + PMD_SIZE - 1) >> PMD_SHIFT;
  45         } else
  46                 pmds = (end + PMD_SIZE - 1) >> PMD_SHIFT;
  47
  48         tables += roundup(pmds * sizeof(pmd_t), PAGE_SIZE);
  49
  50         if (use_pse) {
  51                 unsigned long extra;
  52
  53                 extra = end - ((end>>PMD_SHIFT) << PMD_SHIFT);
  54 #ifdef CONFIG_X86_32
  55                 extra += PMD_SIZE;
  56 #endif
  57                 ptes = (extra + PAGE_SIZE - 1) >> PAGE_SHIFT;
  58         } else
  59                 ptes = (end + PAGE_SIZE - 1) >> PAGE_SHIFT;
  60
  61         tables += roundup(ptes * sizeof(pte_t), PAGE_SIZE);
  62
  63 #ifdef CONFIG_X86_32
  64         /* for fixmap */
  65         tables += roundup(__end_of_fixed_addresses * sizeof(pte_t), PAGE_SIZE);
  66 #endif
  67         good_end = max_pfn_mapped << PAGE_SHIFT;
  68
  69         base = memblock_find_in_range(start, good_end, tables, PAGE_SIZE);
  70         if (base == MEMBLOCK_ERROR)
  71                 panic("Cannot find space for the kernel page tables");
  72
  73         pgt_buf_start = base >> PAGE_SHIFT;
  74         pgt_buf_end = pgt_buf_start;
  75         pgt_buf_top = pgt_buf_start + (tables >> PAGE_SHIFT);
  76
  77         printk(KERN_DEBUG "kernel direct mapping tables up to %lx @ %lx-%lx\n",
  78                 end, pgt_buf_start << PAGE_SHIFT, pgt_buf_top << PAGE_SHIFT);
  79 }
  80
  81 void __init native_pagetable_reserve(u64 start, u64 end)
  82 {
  83         memblock_x86_reserve_range(start, end, "PGTABLE");
  84 }
  85
  86 struct map_range {
  87         unsigned long start;
  88         unsigned long end;
  89         unsigned page_size_mask;
  90 };
  91
  92 #ifdef CONFIG_X86_32
  93 #define NR_RANGE_MR 3
  94 #else /* CONFIG_X86_64 */
  95 #define NR_RANGE_MR 5
  96 #endif
  97
  98 static int __meminit save_mr(struct map_range *mr, int nr_range,
  99                              unsigned long start_pfn, unsigned long end_pfn,
 100                              unsigned long page_size_mask)
 101 {
 102         if (start_pfn < end_pfn) {
 103                 if (nr_range >= NR_RANGE_MR)
 104                         panic("run out of range for init_memory_mapping\n");
 105                 mr[nr_range].start = start_pfn<<PAGE_SHIFT;
 106                 mr[nr_range].end   = end_pfn<<PAGE_SHIFT;
 107                 mr[nr_range].page_size_mask = page_size_mask;
 108                 nr_range++;
 109         }
 110
 111         return nr_range;
 112 }
 113
 114 /*
 115  * Setup the direct mapping of the physical memory at PAGE_OFFSET.
 116  * This runs before bootmem is initialized and gets pages directly from
 117  * the physical memory. To access them they are temporarily mapped.
 118  */
 119 unsigned long __init_refok init_memory_mapping(unsigned long start,
 120                                                unsigned long end)
 121 {
 122         unsigned long page_size_mask = 0;
 123         unsigned long start_pfn, end_pfn;
 124         unsigned long ret = 0;
 125         unsigned long pos;
 126
 127         struct map_range mr[NR_RANGE_MR];
 128         int nr_range, i;
 129         int use_pse, use_gbpages;
 130
 131         printk(KERN_INFO "init_memory_mapping: %016lx-%016lx\n", start, end);
 132
 133 #if defined(CONFIG_DEBUG_PAGEALLOC) || defined(CONFIG_KMEMCHECK)
 134         /*
 135          * For CONFIG_DEBUG_PAGEALLOC, identity mapping will use small pages.
 136          * This will simplify cpa(), which otherwise needs to support splitting
 137          * large pages into small in interrupt context, etc.
 138          */
 139         use_pse = use_gbpages = 0;
 140 #else
 141         use_pse = cpu_has_pse;
 142         use_gbpages = direct_gbpages;
 143 #endif
 144
 145         /* Enable PSE if available */
 146         if (cpu_has_pse)
 147                 set_in_cr4(X86_CR4_PSE);
 148
 149         /* Enable PGE if available */
 150         if (cpu_has_pge) {
 151                 set_in_cr4(X86_CR4_PGE);
 152                 __supported_pte_mask |= _PAGE_GLOBAL;
 153         }
 154
 155         if (use_gbpages)
 156                 page_size_mask |= 1 << PG_LEVEL_1G;
 157         if (use_pse)
 158                 page_size_mask |= 1 << PG_LEVEL_2M;
 159
 160         memset(mr, 0, sizeof(mr));
 161         nr_range = 0;
 162
 163         /* head if not big page alignment ? */
 164         start_pfn = start >> PAGE_SHIFT;
 165         pos = start_pfn << PAGE_SHIFT;
 166 #ifdef CONFIG_X86_32
 167         /*
 168          * Don't use a large page for the first 2/4MB of memory
 169          * because there are often fixed size MTRRs in there
 170          * and overlapping MTRRs into large pages can cause
 171          * slowdowns.
 172          */
 173         if (pos == 0)
 174                 end_pfn = 1<<(PMD_SHIFT - PAGE_SHIFT);
 175         else
 176                 end_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
 177                                  << (PMD_SHIFT - PAGE_SHIFT);
 178 #else /* CONFIG_X86_64 */
 179         end_pfn = ((pos + (PMD_SIZE - 1)) >> PMD_SHIFT)
 180                         << (PMD_SHIFT - PAGE_SHIFT);
 181 #endif
 182         if (end_pfn > (end >> PAGE_SHIFT))
 183                 end_pfn = end >> PAGE_SHIFT;
 184         if (start_pfn < end_pfn) {
 185                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
 186                 pos = end_pfn << PAGE_SHIFT;
 187         }
 188
 189         /* big page (2M) range */
 190         start_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
 191                          << (PMD_SHIFT - PAGE_SHIFT);
 192 #ifdef CONFIG_X86_32
 193         end_pfn = (end>>PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
 194 #else /* CONFIG_X86_64 */
 195         end_pfn = ((pos + (PUD_SIZE - 1))>>PUD_SHIFT)
 196                          << (PUD_SHIFT - PAGE_SHIFT);
 197         if (end_pfn > ((end>>PMD_SHIFT)<<(PMD_SHIFT - PAGE_SHIFT)))
 198                 end_pfn = ((end>>PMD_SHIFT)<<(PMD_SHIFT - PAGE_SHIFT));
 199 #endif
 200
 201         if (start_pfn < end_pfn) {
 202                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 203                                 page_size_mask & (1<<PG_LEVEL_2M));
 204                 pos = end_pfn << PAGE_SHIFT;
 205         }
 206
 207 #ifdef CONFIG_X86_64
 208         /* big page (1G) range */
 209         start_pfn = ((pos + (PUD_SIZE - 1))>>PUD_SHIFT)
 210                          << (PUD_SHIFT - PAGE_SHIFT);
 211         end_pfn = (end >> PUD_SHIFT) << (PUD_SHIFT - PAGE_SHIFT);
 212         if (start_pfn < end_pfn) {
 213                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 214                                 page_size_mask &
 215                                  ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G)));
 216                 pos = end_pfn << PAGE_SHIFT;
 217         }
 218
 219         /* tail is not big page (1G) alignment */
 220         start_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
 221                          << (PMD_SHIFT - PAGE_SHIFT);
 222         end_pfn = (end >> PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
 223         if (start_pfn < end_pfn) {
 224                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 225                                 page_size_mask & (1<<PG_LEVEL_2M));
 226                 pos = end_pfn << PAGE_SHIFT;
 227         }
 228 #endif
 229
 230         /* tail is not big page (2M) alignment */
 231         start_pfn = pos>>PAGE_SHIFT;
 232         end_pfn = end>>PAGE_SHIFT;
 233         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
 234
 235         /* try to merge same page size and continuous */
 236         for (i = 0; nr_range > 1 && i < nr_range - 1; i++) {
 237                 unsigned long old_start;
 238                 if (mr[i].end != mr[i+1].start ||
 239                     mr[i].page_size_mask != mr[i+1].page_size_mask)
 240                         continue;
 241                 /* move it */
 242                 old_start = mr[i].start;
 243                 memmove(&mr[i], &mr[i+1],
 244                         (nr_range - 1 - i) * sizeof(struct map_range));
 245                 mr[i--].start = old_start;
 246                 nr_range--;
 247         }
 248
 249         for (i = 0; i < nr_range; i++)
 250                 printk(KERN_DEBUG " %010lx - %010lx page %s\n",
 251                                 mr[i].start, mr[i].end,
 252                         (mr[i].page_size_mask & (1<<PG_LEVEL_1G))?"1G":(
 253                          (mr[i].page_size_mask & (1<<PG_LEVEL_2M))?"2M":"4k"));
 254
 255         /*
 256          * Find space for the kernel direct mapping tables.
 257          *
 258          * Later we should allocate these tables in the local node of the
 259          * memory mapped. Unfortunately this is done currently before the
 260          * nodes are discovered.
 261          */
 262         if (!after_bootmem)
 263                 find_early_table_space(end, use_pse, use_gbpages);
 264
 265         for (i = 0; i < nr_range; i++)
 266                 ret = kernel_physical_mapping_init(mr[i].start, mr[i].end,
 267                                                    mr[i].page_size_mask);
 268
 269 #ifdef CONFIG_X86_32
 270         early_ioremap_page_table_range_init();
 271
 272         load_cr3(swapper_pg_dir);
 273 #endif
 274
 275         __flush_tlb_all();
 276
 277         /*
 278          * Reserve the kernel pagetable pages we used (pgt_buf_start -
 279          * pgt_buf_end) and free the other ones (pgt_buf_end - pgt_buf_top)
 280          * so that they can be reused for other purposes.
 281          *
 282          * On native it just means calling memblock_x86_reserve_range, on Xen it
 283          * also means marking RW the pagetable pages that we allocated before
 284          * but that haven't been used.
 285          *
 286          * In fact on xen we mark RO the whole range pgt_buf_start -
 287          * pgt_buf_top, because we have to make sure that when
 288          * init_memory_mapping reaches the pagetable pages area, it maps
 289          * RO all the pagetable pages, including the ones that are beyond
 290          * pgt_buf_end at that time.
 291          */
 292         if (!after_bootmem && pgt_buf_end > pgt_buf_start)
 293                 x86_init.mapping.pagetable_reserve(PFN_PHYS(pgt_buf_start),
 294                                 PFN_PHYS(pgt_buf_end));
 295
 296         if (!after_bootmem)
 297                 early_memtest(start, end);
 298
 299         return ret >> PAGE_SHIFT;
 300 }
 301
 302
 303 /*
 304  * devmem_is_allowed() checks to see if /dev/mem access to a certain address
 305  * is valid. The argument is a physical page number.
 306  *
 307  *
 308  * On x86, access has to be given to the first megabyte of ram because that area
 309  * contains bios code and data regions used by X and dosemu and similar apps.
 310  * Access has to be given to non-kernel-ram areas as well, these contain the PCI
 311  * mmio resources as well as potential bios/acpi data regions.
 312  */
 313 int devmem_is_allowed(unsigned long pagenr)
 314 {
 315         if (pagenr <= 256)
 316                 return 1;
 317         if (iomem_is_exclusive(pagenr << PAGE_SHIFT))
 318                 return 0;
 319         if (!page_is_ram(pagenr))
 320                 return 1;
 321         return 0;
 322 }
 323
 324 void free_init_pages(char *what, unsigned long begin, unsigned long end)
 325 {
 326         unsigned long addr;
 327         unsigned long begin_aligned, end_aligned;
 328
 329         /* Make sure boundaries are page aligned */
 330         begin_aligned = PAGE_ALIGN(begin);
 331         end_aligned   = end & PAGE_MASK;
 332
 333         if (WARN_ON(begin_aligned != begin || end_aligned != end)) {
 334                 begin = begin_aligned;
 335                 end   = end_aligned;
 336         }
 337
 338         if (begin >= end)
 339                 return;
 340
 341         addr = begin;
 342
 343         /*
 344          * If debugging page accesses then do not free this memory but
 345          * mark them not present - any buggy init-section access will
 346          * create a kernel page fault:
 347          */
 348 #ifdef CONFIG_DEBUG_PAGEALLOC
 349         printk(KERN_INFO "debug: unmapping init memory %08lx..%08lx\n",
 350                 begin, end);
 351         set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
 352 #else
 353         /*
 354          * We just marked the kernel text read only above, now that
 355          * we are going to free part of that, we need to make that
 356          * writeable and non-executable first.
 357          */
 358         set_memory_nx(begin, (end - begin) >> PAGE_SHIFT);
 359         set_memory_rw(begin, (end - begin) >> PAGE_SHIFT);
 360
 361         printk(KERN_INFO "Freeing %s: %luk freed\n", what, (end - begin) >> 10);
 362
 363         for (; addr < end; addr += PAGE_SIZE) {
 364                 ClearPageReserved(virt_to_page(addr));
 365                 init_page_count(virt_to_page(addr));
 366                 memset((void *)addr, POISON_FREE_INITMEM, PAGE_SIZE);
 367                 free_page(addr);
 368                 totalram_pages++;
 369         }
 370 #endif
 371 }
 372
 373 void free_initmem(void)
 374 {
 375         free_init_pages("unused kernel memory",
 376                         (unsigned long)(&__init_begin),
 377                         (unsigned long)(&__init_end));
 378 }
 379
 380 #ifdef CONFIG_BLK_DEV_INITRD
 381 void free_initrd_mem(unsigned long start, unsigned long end)
 382 {
 383         /*
 384          * end could be not aligned, and We can not align that,
 385          * decompresser could be confused by aligned initrd_end
 386          * We already reserve the end partial page before in
 387          *   - i386_start_kernel()
 388          *   - x86_64_start_kernel()
 389          *   - relocate_initrd()
 390          * So here We can do PAGE_ALIGN() safely to get partial page to be freed
 391          */
 392         free_init_pages("initrd memory", start, PAGE_ALIGN(end));
 393 }
 394 #endif