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 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
  69         good_end = max_pfn_mapped << PAGE_SHIFT;
  70 #endif
  71
  72         base = memblock_find_in_range(start, good_end, tables, PAGE_SIZE);
  73         if (base == MEMBLOCK_ERROR)
  74                 panic("Cannot find space for the kernel page tables");
  75
  76         pgt_buf_start = base >> PAGE_SHIFT;
  77         pgt_buf_end = pgt_buf_start;
  78         pgt_buf_top = pgt_buf_start + (tables >> PAGE_SHIFT);
  79
  80         printk(KERN_DEBUG "kernel direct mapping tables up to %lx @ %lx-%lx\n",
  81                 end, pgt_buf_start << PAGE_SHIFT, pgt_buf_top << PAGE_SHIFT);
  82 }
  83
  84 void __init native_pagetable_reserve(u64 start, u64 end)
  85 {
  86         memblock_x86_reserve_range(start, end, "PGTABLE");
  87 }
  88
  89 struct map_range {
  90         unsigned long start;
  91         unsigned long end;
  92         unsigned page_size_mask;
  93 };
  94
  95 #ifdef CONFIG_X86_32
  96 #define NR_RANGE_MR 3
  97 #else /* CONFIG_X86_64 */
  98 #define NR_RANGE_MR 5
  99 #endif
 100
 101 static int __meminit save_mr(struct map_range *mr, int nr_range,
 102                              unsigned long start_pfn, unsigned long end_pfn,
 103                              unsigned long page_size_mask)
 104 {
 105         if (start_pfn < end_pfn) {
 106                 if (nr_range >= NR_RANGE_MR)
 107                         panic("run out of range for init_memory_mapping\n");
 108                 mr[nr_range].start = start_pfn<<PAGE_SHIFT;
 109                 mr[nr_range].end   = end_pfn<<PAGE_SHIFT;
 110                 mr[nr_range].page_size_mask = page_size_mask;
 111                 nr_range++;
 112         }
 113
 114         return nr_range;
 115 }
 116
 117 /*
 118  * Setup the direct mapping of the physical memory at PAGE_OFFSET.
 119  * This runs before bootmem is initialized and gets pages directly from
 120  * the physical memory. To access them they are temporarily mapped.
 121  */
 122 unsigned long __init_refok init_memory_mapping(unsigned long start,
 123                                                unsigned long end)
 124 {
 125         unsigned long page_size_mask = 0;
 126         unsigned long start_pfn, end_pfn;
 127         unsigned long ret = 0;
 128         unsigned long pos;
 129
 130         struct map_range mr[NR_RANGE_MR];
 131         int nr_range, i;
 132         int use_pse, use_gbpages;
 133
 134         printk(KERN_INFO "init_memory_mapping: %016lx-%016lx\n", start, end);
 135
 136 #if defined(CONFIG_DEBUG_PAGEALLOC) || defined(CONFIG_KMEMCHECK)
 137         /*
 138          * For CONFIG_DEBUG_PAGEALLOC, identity mapping will use small pages.
 139          * This will simplify cpa(), which otherwise needs to support splitting
 140          * large pages into small in interrupt context, etc.
 141          */
 142         use_pse = use_gbpages = 0;
 143 #else
 144         use_pse = cpu_has_pse;
 145         use_gbpages = direct_gbpages;
 146 #endif
 147
 148         /* Enable PSE if available */
 149         if (cpu_has_pse)
 150                 set_in_cr4(X86_CR4_PSE);
 151
 152         /* Enable PGE if available */
 153         if (cpu_has_pge) {
 154                 set_in_cr4(X86_CR4_PGE);
 155                 __supported_pte_mask |= _PAGE_GLOBAL;
 156         }
 157
 158         if (use_gbpages)
 159                 page_size_mask |= 1 << PG_LEVEL_1G;
 160         if (use_pse)
 161                 page_size_mask |= 1 << PG_LEVEL_2M;
 162
 163         memset(mr, 0, sizeof(mr));
 164         nr_range = 0;
 165
 166         /* head if not big page alignment ? */
 167         start_pfn = start >> PAGE_SHIFT;
 168         pos = start_pfn << PAGE_SHIFT;
 169 #ifdef CONFIG_X86_32
 170         /*
 171          * Don't use a large page for the first 2/4MB of memory
 172          * because there are often fixed size MTRRs in there
 173          * and overlapping MTRRs into large pages can cause
 174          * slowdowns.
 175          */
 176         if (pos == 0)
 177                 end_pfn = 1<<(PMD_SHIFT - PAGE_SHIFT);
 178         else
 179                 end_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
 180                                  << (PMD_SHIFT - PAGE_SHIFT);
 181 #else /* CONFIG_X86_64 */
 182         end_pfn = ((pos + (PMD_SIZE - 1)) >> PMD_SHIFT)
 183                         << (PMD_SHIFT - PAGE_SHIFT);
 184 #endif
 185         if (end_pfn > (end >> PAGE_SHIFT))
 186                 end_pfn = end >> PAGE_SHIFT;
 187         if (start_pfn < end_pfn) {
 188                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
 189                 pos = end_pfn << PAGE_SHIFT;
 190         }
 191
 192         /* big page (2M) range */
 193         start_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
 194                          << (PMD_SHIFT - PAGE_SHIFT);
 195 #ifdef CONFIG_X86_32
 196         end_pfn = (end>>PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
 197 #else /* CONFIG_X86_64 */
 198         end_pfn = ((pos + (PUD_SIZE - 1))>>PUD_SHIFT)
 199                          << (PUD_SHIFT - PAGE_SHIFT);
 200         if (end_pfn > ((end>>PMD_SHIFT)<<(PMD_SHIFT - PAGE_SHIFT)))
 201                 end_pfn = ((end>>PMD_SHIFT)<<(PMD_SHIFT - PAGE_SHIFT));
 202 #endif
 203
 204         if (start_pfn < end_pfn) {
 205                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 206                                 page_size_mask & (1<<PG_LEVEL_2M));
 207                 pos = end_pfn << PAGE_SHIFT;
 208         }
 209
 210 #ifdef CONFIG_X86_64
 211         /* big page (1G) range */
 212         start_pfn = ((pos + (PUD_SIZE - 1))>>PUD_SHIFT)
 213                          << (PUD_SHIFT - PAGE_SHIFT);
 214         end_pfn = (end >> PUD_SHIFT) << (PUD_SHIFT - PAGE_SHIFT);
 215         if (start_pfn < end_pfn) {
 216                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 217                                 page_size_mask &
 218                                  ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G)));
 219                 pos = end_pfn << PAGE_SHIFT;
 220         }
 221
 222         /* tail is not big page (1G) alignment */
 223         start_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
 224                          << (PMD_SHIFT - PAGE_SHIFT);
 225         end_pfn = (end >> PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
 226         if (start_pfn < end_pfn) {
 227                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 228                                 page_size_mask & (1<<PG_LEVEL_2M));
 229                 pos = end_pfn << PAGE_SHIFT;
 230         }
 231 #endif
 232
 233         /* tail is not big page (2M) alignment */
 234         start_pfn = pos>>PAGE_SHIFT;
 235         end_pfn = end>>PAGE_SHIFT;
 236         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
 237
 238         /* try to merge same page size and continuous */
 239         for (i = 0; nr_range > 1 && i < nr_range - 1; i++) {
 240                 unsigned long old_start;
 241                 if (mr[i].end != mr[i+1].start ||
 242                     mr[i].page_size_mask != mr[i+1].page_size_mask)
 243                         continue;
 244                 /* move it */
 245                 old_start = mr[i].start;
 246                 memmove(&mr[i], &mr[i+1],
 247                         (nr_range - 1 - i) * sizeof(struct map_range));
 248                 mr[i--].start = old_start;
 249                 nr_range--;
 250         }
 251
 252         for (i = 0; i < nr_range; i++)
 253                 printk(KERN_DEBUG " %010lx - %010lx page %s\n",
 254                                 mr[i].start, mr[i].end,
 255                         (mr[i].page_size_mask & (1<<PG_LEVEL_1G))?"1G":(
 256                          (mr[i].page_size_mask & (1<<PG_LEVEL_2M))?"2M":"4k"));
 257
 258         /*
 259          * Find space for the kernel direct mapping tables.
 260          *
 261          * Later we should allocate these tables in the local node of the
 262          * memory mapped. Unfortunately this is done currently before the
 263          * nodes are discovered.
 264          */
 265         if (!after_bootmem)
 266                 find_early_table_space(end, use_pse, use_gbpages);
 267
 268         for (i = 0; i < nr_range; i++)
 269                 ret = kernel_physical_mapping_init(mr[i].start, mr[i].end,
 270                                                    mr[i].page_size_mask);
 271
 272 #ifdef CONFIG_X86_32
 273         early_ioremap_page_table_range_init();
 274
 275         load_cr3(swapper_pg_dir);
 276 #endif
 277
 278         __flush_tlb_all();
 279
 280         /*
 281          * Reserve the kernel pagetable pages we used (pgt_buf_start -
 282          * pgt_buf_end) and free the other ones (pgt_buf_end - pgt_buf_top)
 283          * so that they can be reused for other purposes.
 284          *
 285          * On native it just means calling memblock_x86_reserve_range, on Xen it
 286          * also means marking RW the pagetable pages that we allocated before
 287          * but that haven't been used.
 288          *
 289          * In fact on xen we mark RO the whole range pgt_buf_start -
 290          * pgt_buf_top, because we have to make sure that when
 291          * init_memory_mapping reaches the pagetable pages area, it maps
 292          * RO all the pagetable pages, including the ones that are beyond
 293          * pgt_buf_end at that time.
 294          */
 295         if (!after_bootmem && pgt_buf_end > pgt_buf_start)
 296                 x86_init.mapping.pagetable_reserve(PFN_PHYS(pgt_buf_start),
 297                                 PFN_PHYS(pgt_buf_end));
 298
 299         if (!after_bootmem)
 300                 early_memtest(start, end);
 301
 302         return ret >> PAGE_SHIFT;
 303 }
 304
 305
 306 /*
 307  * devmem_is_allowed() checks to see if /dev/mem access to a certain address
 308  * is valid. The argument is a physical page number.
 309  *
 310  *
 311  * On x86, access has to be given to the first megabyte of ram because that area
 312  * contains bios code and data regions used by X and dosemu and similar apps.
 313  * Access has to be given to non-kernel-ram areas as well, these contain the PCI
 314  * mmio resources as well as potential bios/acpi data regions.
 315  */
 316 int devmem_is_allowed(unsigned long pagenr)
 317 {
 318         if (pagenr <= 256)
 319                 return 1;
 320         if (iomem_is_exclusive(pagenr << PAGE_SHIFT))
 321                 return 0;
 322         if (!page_is_ram(pagenr))
 323                 return 1;
 324         return 0;
 325 }
 326
 327 void free_init_pages(char *what, unsigned long begin, unsigned long end)
 328 {
 329         unsigned long addr;
 330         unsigned long begin_aligned, end_aligned;
 331
 332         /* Make sure boundaries are page aligned */
 333         begin_aligned = PAGE_ALIGN(begin);
 334         end_aligned   = end & PAGE_MASK;
 335
 336         if (WARN_ON(begin_aligned != begin || end_aligned != end)) {
 337                 begin = begin_aligned;
 338                 end   = end_aligned;
 339         }
 340
 341         if (begin >= end)
 342                 return;
 343
 344         addr = begin;
 345
 346         /*
 347          * If debugging page accesses then do not free this memory but
 348          * mark them not present - any buggy init-section access will
 349          * create a kernel page fault:
 350          */
 351 #ifdef CONFIG_DEBUG_PAGEALLOC
 352         printk(KERN_INFO "debug: unmapping init memory %08lx..%08lx\n",
 353                 begin, end);
 354         set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
 355 #else
 356         /*
 357          * We just marked the kernel text read only above, now that
 358          * we are going to free part of that, we need to make that
 359          * writeable and non-executable first.
 360          */
 361         set_memory_nx(begin, (end - begin) >> PAGE_SHIFT);
 362         set_memory_rw(begin, (end - begin) >> PAGE_SHIFT);
 363
 364         printk(KERN_INFO "Freeing %s: %luk freed\n", what, (end - begin) >> 10);
 365
 366         for (; addr < end; addr += PAGE_SIZE) {
 367                 ClearPageReserved(virt_to_page(addr));
 368                 init_page_count(virt_to_page(addr));
 369                 memset((void *)addr, POISON_FREE_INITMEM, PAGE_SIZE);
 370                 free_page(addr);
 371                 totalram_pages++;
 372         }
 373 #endif
 374 }
 375
 376 void free_initmem(void)
 377 {
 378         free_init_pages("unused kernel memory",
 379                         (unsigned long)(&__init_begin),
 380                         (unsigned long)(&__init_end));
 381 }
 382
 383 #ifdef CONFIG_BLK_DEV_INITRD
 384 void free_initrd_mem(unsigned long start, unsigned long end)
 385 {
 386         /*
 387          * end could be not aligned, and We can not align that,
 388          * decompresser could be confused by aligned initrd_end
 389          * We already reserve the end partial page before in
 390          *   - i386_start_kernel()
 391          *   - x86_64_start_kernel()
 392          *   - relocate_initrd()
 393          * So here We can do PAGE_ALIGN() safely to get partial page to be freed
 394          */
 395         free_init_pages("initrd memory", start, PAGE_ALIGN(end));
 396 }
 397 #endif