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 unsigned long __init find_early_fixmap_space(void)
  32 {
  33         unsigned long size = 0;
  34 #ifdef CONFIG_X86_32
  35         int kmap_begin_pmd_idx, kmap_end_pmd_idx;
  36         int fixmap_begin_pmd_idx, fixmap_end_pmd_idx;
  37         int btmap_begin_pmd_idx;
  38
  39         fixmap_begin_pmd_idx =
  40                 __fix_to_virt(__end_of_fixed_addresses - 1) >> PMD_SHIFT;
  41         /*
  42          * fixmap_end_pmd_idx is the end of the fixmap minus the PMD that
  43          * has been defined in the data section by head_32.S (see
  44          * initial_pg_fixmap).
  45          * Note: This is similar to what early_ioremap_page_table_range_init
  46          * does except that the "end" has PMD_SIZE expunged as per previous
  47          * comment.
  48          */
  49         fixmap_end_pmd_idx = (FIXADDR_TOP - 1) >> PMD_SHIFT;
  50         btmap_begin_pmd_idx = __fix_to_virt(FIX_BTMAP_BEGIN) >> PMD_SHIFT;
  51         kmap_begin_pmd_idx = __fix_to_virt(FIX_KMAP_END) >> PMD_SHIFT;
  52         kmap_end_pmd_idx = __fix_to_virt(FIX_KMAP_BEGIN) >> PMD_SHIFT;
  53
  54         size = fixmap_end_pmd_idx - fixmap_begin_pmd_idx;
  55         /*
  56          * early_ioremap_init has already allocated a PMD at
  57          * btmap_begin_pmd_idx
  58          */
  59         if (btmap_begin_pmd_idx < fixmap_end_pmd_idx)
  60                 size--;
  61
  62 #ifdef CONFIG_HIGHMEM
  63         /*
  64          * see page_table_kmap_check: if the kmap spans multiple PMDs, make
  65          * sure the pte pages are allocated contiguously. It might need up
  66          * to two additional pte pages to replace the page declared by
  67          * head_32.S and the one allocated by early_ioremap_init, if they
  68          * are even partially used for the kmap.
  69          */
  70         if (kmap_begin_pmd_idx != kmap_end_pmd_idx) {
  71                 if (kmap_end_pmd_idx == fixmap_end_pmd_idx)
  72                         size++;
  73                 if (btmap_begin_pmd_idx >= kmap_begin_pmd_idx &&
  74                                 btmap_begin_pmd_idx <= kmap_end_pmd_idx)
  75                         size++;
  76         }
  77 #endif
  78 #endif
  79         return (size * PMD_SIZE + PAGE_SIZE - 1) >> PAGE_SHIFT;
  80 }
  81
  82 static void __init find_early_table_space(unsigned long start,
  83                 unsigned long end, int use_pse, int use_gbpages)
  84 {
  85         unsigned long pmds = 0, ptes = 0, tables = 0, good_end = end,
  86                                   pud_mapped = 0, pmd_mapped = 0, size = end - start;
  87         phys_addr_t base;
  88
  89         pud_mapped = DIV_ROUND_UP(PFN_PHYS(max_pfn_mapped),
  90                         (PUD_SIZE * PTRS_PER_PUD));
  91         pud_mapped *= (PUD_SIZE * PTRS_PER_PUD);
  92         pmd_mapped = DIV_ROUND_UP(PFN_PHYS(max_pfn_mapped),
  93                         (PMD_SIZE * PTRS_PER_PMD));
  94         pmd_mapped *= (PMD_SIZE * PTRS_PER_PMD);
  95
  96         /*
  97          * On x86_64 do not limit the size we need to cover with 4KB pages
  98          * depending on the initial allocation because head_64.S always uses
  99          * 2MB pages.
 100          */
 101 #ifdef CONFIG_X86_32
 102         if (start < PFN_PHYS(max_pfn_mapped)) {
 103                 if (PFN_PHYS(max_pfn_mapped) < end)
 104                         size -= PFN_PHYS(max_pfn_mapped) - start;
 105                 else
 106                         size = 0;
 107         }
 108 #endif
 109
 110 #ifndef __PAGETABLE_PUD_FOLDED
 111         if (end > pud_mapped) {
 112                 unsigned long puds;
 113                 if (start < pud_mapped)
 114                         puds = (end - pud_mapped + PUD_SIZE - 1) >> PUD_SHIFT;
 115                 else
 116                         puds = (end - start + PUD_SIZE - 1) >> PUD_SHIFT;
 117                 tables += roundup(puds * sizeof(pud_t), PAGE_SIZE);
 118         }
 119 #endif
 120
 121         if (use_gbpages) {
 122                 unsigned long extra;
 123
 124                 extra = end - ((end>>PUD_SHIFT) << PUD_SHIFT);
 125                 pmds = (extra + PMD_SIZE - 1) >> PMD_SHIFT;
 126         }
 127 #ifndef __PAGETABLE_PMD_FOLDED
 128         else if (end > pmd_mapped) {
 129                 if (start < pmd_mapped)
 130                         pmds = (end - pmd_mapped + PMD_SIZE - 1) >> PMD_SHIFT;
 131                 else
 132                         pmds = (end - start + PMD_SIZE - 1) >> PMD_SHIFT;
 133         }
 134 #endif
 135
 136         tables += roundup(pmds * sizeof(pmd_t), PAGE_SIZE);
 137
 138         if (use_pse) {
 139                 unsigned long extra;
 140
 141                 extra = end - ((end>>PMD_SHIFT) << PMD_SHIFT);
 142                 ptes = (extra + PAGE_SIZE - 1) >> PAGE_SHIFT;
 143         } else
 144                 ptes = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
 145
 146         ptes += find_early_fixmap_space();
 147
 148         tables += roundup(ptes * sizeof(pte_t), PAGE_SIZE);
 149
 150         if (!tables)
 151                 return;
 152
 153 #ifdef CONFIG_X86_32
 154         good_end = max_pfn_mapped << PAGE_SHIFT;
 155 #endif
 156
 157         base = memblock_find_in_range(0x00, good_end, tables, PAGE_SIZE);
 158         if (base == MEMBLOCK_ERROR)
 159                 panic("Cannot find space for the kernel page tables");
 160
 161         pgt_buf_start = base >> PAGE_SHIFT;
 162         pgt_buf_end = pgt_buf_start;
 163         pgt_buf_top = pgt_buf_start + (tables >> PAGE_SHIFT);
 164
 165         printk(KERN_DEBUG "kernel direct mapping tables up to %lx @ %lx-%lx\n",
 166                 end, pgt_buf_start << PAGE_SHIFT, pgt_buf_top << PAGE_SHIFT);
 167
 168         if (pgt_buf_top > pgt_buf_start)
 169                 memblock_x86_reserve_range(pgt_buf_start << PAGE_SHIFT,
 170                                  pgt_buf_top << PAGE_SHIFT, "PGTABLE");
 171 }
 172
 173 struct map_range {
 174         unsigned long start;
 175         unsigned long end;
 176         unsigned page_size_mask;
 177 };
 178
 179 #ifdef CONFIG_X86_32
 180 #define NR_RANGE_MR 3
 181 #else /* CONFIG_X86_64 */
 182 #define NR_RANGE_MR 5
 183 #endif
 184
 185 static int __meminit save_mr(struct map_range *mr, int nr_range,
 186                              unsigned long start_pfn, unsigned long end_pfn,
 187                              unsigned long page_size_mask)
 188 {
 189         if (start_pfn < end_pfn) {
 190                 if (nr_range >= NR_RANGE_MR)
 191                         panic("run out of range for init_memory_mapping\n");
 192                 mr[nr_range].start = start_pfn<<PAGE_SHIFT;
 193                 mr[nr_range].end   = end_pfn<<PAGE_SHIFT;
 194                 mr[nr_range].page_size_mask = page_size_mask;
 195                 nr_range++;
 196         }
 197
 198         return nr_range;
 199 }
 200
 201 /*
 202  * Setup the direct mapping of the physical memory at PAGE_OFFSET.
 203  * This runs before bootmem is initialized and gets pages directly from
 204  * the physical memory. To access them they are temporarily mapped.
 205  */
 206 unsigned long __init_refok init_memory_mapping(unsigned long start,
 207                                                unsigned long end)
 208 {
 209         unsigned long page_size_mask = 0;
 210         unsigned long start_pfn, end_pfn;
 211         unsigned long ret = 0;
 212         unsigned long pos;
 213
 214         struct map_range mr[NR_RANGE_MR];
 215         int nr_range, i;
 216         int use_pse, use_gbpages;
 217
 218         printk(KERN_INFO "init_memory_mapping: %016lx-%016lx\n", start, end);
 219
 220 #if defined(CONFIG_DEBUG_PAGEALLOC) || defined(CONFIG_KMEMCHECK)
 221         /*
 222          * For CONFIG_DEBUG_PAGEALLOC, identity mapping will use small pages.
 223          * This will simplify cpa(), which otherwise needs to support splitting
 224          * large pages into small in interrupt context, etc.
 225          */
 226         use_pse = use_gbpages = 0;
 227 #else
 228         use_pse = cpu_has_pse;
 229         use_gbpages = direct_gbpages;
 230 #endif
 231
 232         /* Enable PSE if available */
 233         if (cpu_has_pse)
 234                 set_in_cr4(X86_CR4_PSE);
 235
 236         /* Enable PGE if available */
 237         if (cpu_has_pge) {
 238                 set_in_cr4(X86_CR4_PGE);
 239                 __supported_pte_mask |= _PAGE_GLOBAL;
 240         }
 241
 242         if (use_gbpages)
 243                 page_size_mask |= 1 << PG_LEVEL_1G;
 244         if (use_pse)
 245                 page_size_mask |= 1 << PG_LEVEL_2M;
 246
 247         memset(mr, 0, sizeof(mr));
 248         nr_range = 0;
 249
 250         /* head if not big page alignment ? */
 251         start_pfn = start >> PAGE_SHIFT;
 252         pos = start_pfn << PAGE_SHIFT;
 253 #ifdef CONFIG_X86_32
 254         /*
 255          * Don't use a large page for the first 2/4MB of memory
 256          * because there are often fixed size MTRRs in there
 257          * and overlapping MTRRs into large pages can cause
 258          * slowdowns.
 259          */
 260         if (pos == 0)
 261                 end_pfn = 1<<(PMD_SHIFT - PAGE_SHIFT);
 262         else
 263                 end_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
 264                                  << (PMD_SHIFT - PAGE_SHIFT);
 265 #else /* CONFIG_X86_64 */
 266         end_pfn = ((pos + (PMD_SIZE - 1)) >> PMD_SHIFT)
 267                         << (PMD_SHIFT - PAGE_SHIFT);
 268 #endif
 269         if (end_pfn > (end >> PAGE_SHIFT))
 270                 end_pfn = end >> PAGE_SHIFT;
 271         if (start_pfn < end_pfn) {
 272                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
 273                 pos = end_pfn << PAGE_SHIFT;
 274         }
 275
 276         /* big page (2M) range */
 277         start_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
 278                          << (PMD_SHIFT - PAGE_SHIFT);
 279 #ifdef CONFIG_X86_32
 280         end_pfn = (end>>PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
 281 #else /* CONFIG_X86_64 */
 282         end_pfn = ((pos + (PUD_SIZE - 1))>>PUD_SHIFT)
 283                          << (PUD_SHIFT - PAGE_SHIFT);
 284         if (end_pfn > ((end>>PMD_SHIFT)<<(PMD_SHIFT - PAGE_SHIFT)))
 285                 end_pfn = ((end>>PMD_SHIFT)<<(PMD_SHIFT - PAGE_SHIFT));
 286 #endif
 287
 288         if (start_pfn < end_pfn) {
 289                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 290                                 page_size_mask & (1<<PG_LEVEL_2M));
 291                 pos = end_pfn << PAGE_SHIFT;
 292         }
 293
 294 #ifdef CONFIG_X86_64
 295         /* big page (1G) range */
 296         start_pfn = ((pos + (PUD_SIZE - 1))>>PUD_SHIFT)
 297                          << (PUD_SHIFT - PAGE_SHIFT);
 298         end_pfn = (end >> PUD_SHIFT) << (PUD_SHIFT - PAGE_SHIFT);
 299         if (start_pfn < end_pfn) {
 300                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 301                                 page_size_mask &
 302                                  ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G)));
 303                 pos = end_pfn << PAGE_SHIFT;
 304         }
 305
 306         /* tail is not big page (1G) alignment */
 307         start_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
 308                          << (PMD_SHIFT - PAGE_SHIFT);
 309         end_pfn = (end >> PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
 310         if (start_pfn < end_pfn) {
 311                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 312                                 page_size_mask & (1<<PG_LEVEL_2M));
 313                 pos = end_pfn << PAGE_SHIFT;
 314         }
 315 #endif
 316
 317         /* tail is not big page (2M) alignment */
 318         start_pfn = pos>>PAGE_SHIFT;
 319         end_pfn = end>>PAGE_SHIFT;
 320         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
 321
 322         /* try to merge same page size and continuous */
 323         for (i = 0; nr_range > 1 && i < nr_range - 1; i++) {
 324                 unsigned long old_start;
 325                 if (mr[i].end != mr[i+1].start ||
 326                     mr[i].page_size_mask != mr[i+1].page_size_mask)
 327                         continue;
 328                 /* move it */
 329                 old_start = mr[i].start;
 330                 memmove(&mr[i], &mr[i+1],
 331                         (nr_range - 1 - i) * sizeof(struct map_range));
 332                 mr[i--].start = old_start;
 333                 nr_range--;
 334         }
 335
 336         for (i = 0; i < nr_range; i++)
 337                 printk(KERN_DEBUG " %010lx - %010lx page %s\n",
 338                                 mr[i].start, mr[i].end,
 339                         (mr[i].page_size_mask & (1<<PG_LEVEL_1G))?"1G":(
 340                          (mr[i].page_size_mask & (1<<PG_LEVEL_2M))?"2M":"4k"));
 341
 342         /*
 343          * Find space for the kernel direct mapping tables.
 344          *
 345          * Later we should allocate these tables in the local node of the
 346          * memory mapped. Unfortunately this is done currently before the
 347          * nodes are discovered.
 348          */
 349         if (!after_bootmem)
 350                 find_early_table_space(start, end, use_pse, use_gbpages);
 351
 352         for (i = 0; i < nr_range; i++)
 353                 ret = kernel_physical_mapping_init(mr[i].start, mr[i].end,
 354                                                    mr[i].page_size_mask);
 355
 356 #ifdef CONFIG_X86_32
 357         early_ioremap_page_table_range_init();
 358
 359         load_cr3(swapper_pg_dir);
 360 #endif
 361
 362         __flush_tlb_all();
 363
 364         if (pgt_buf_end != pgt_buf_top)
 365                 printk(KERN_DEBUG "initial kernel pagetable allocation wasted %lx"
 366                                 " pages\n", pgt_buf_top - pgt_buf_end);
 367
 368         if (!after_bootmem)
 369                 early_memtest(start, end);
 370
 371         return ret >> PAGE_SHIFT;
 372 }
 373
 374
 375 /*
 376  * devmem_is_allowed() checks to see if /dev/mem access to a certain address
 377  * is valid. The argument is a physical page number.
 378  *
 379  *
 380  * On x86, access has to be given to the first megabyte of ram because that area
 381  * contains bios code and data regions used by X and dosemu and similar apps.
 382  * Access has to be given to non-kernel-ram areas as well, these contain the PCI
 383  * mmio resources as well as potential bios/acpi data regions.
 384  */
 385 int devmem_is_allowed(unsigned long pagenr)
 386 {
 387         if (pagenr <= 256)
 388                 return 1;
 389         if (iomem_is_exclusive(pagenr << PAGE_SHIFT))
 390                 return 0;
 391         if (!page_is_ram(pagenr))
 392                 return 1;
 393         return 0;
 394 }
 395
 396 void free_init_pages(char *what, unsigned long begin, unsigned long end)
 397 {
 398         unsigned long addr;
 399         unsigned long begin_aligned, end_aligned;
 400
 401         /* Make sure boundaries are page aligned */
 402         begin_aligned = PAGE_ALIGN(begin);
 403         end_aligned   = end & PAGE_MASK;
 404
 405         if (WARN_ON(begin_aligned != begin || end_aligned != end)) {
 406                 begin = begin_aligned;
 407                 end   = end_aligned;
 408         }
 409
 410         if (begin >= end)
 411                 return;
 412
 413         addr = begin;
 414
 415         /*
 416          * If debugging page accesses then do not free this memory but
 417          * mark them not present - any buggy init-section access will
 418          * create a kernel page fault:
 419          */
 420 #ifdef CONFIG_DEBUG_PAGEALLOC
 421         printk(KERN_INFO "debug: unmapping init memory %08lx..%08lx\n",
 422                 begin, end);
 423         set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
 424 #else
 425         /*
 426          * We just marked the kernel text read only above, now that
 427          * we are going to free part of that, we need to make that
 428          * writeable and non-executable first.
 429          */
 430         set_memory_nx(begin, (end - begin) >> PAGE_SHIFT);
 431         set_memory_rw(begin, (end - begin) >> PAGE_SHIFT);
 432
 433         printk(KERN_INFO "Freeing %s: %luk freed\n", what, (end - begin) >> 10);
 434
 435         for (; addr < end; addr += PAGE_SIZE) {
 436                 ClearPageReserved(virt_to_page(addr));
 437                 init_page_count(virt_to_page(addr));
 438                 memset((void *)addr, POISON_FREE_INITMEM, PAGE_SIZE);
 439                 free_page(addr);
 440                 totalram_pages++;
 441         }
 442 #endif
 443 }
 444
 445 void free_initmem(void)
 446 {
 447         free_init_pages("unused kernel memory",
 448                         (unsigned long)(&__init_begin),
 449                         (unsigned long)(&__init_end));
 450 }
 451
 452 #ifdef CONFIG_BLK_DEV_INITRD
 453 void free_initrd_mem(unsigned long start, unsigned long end)
 454 {
 455         /*
 456          * end could be not aligned, and We can not align that,
 457          * decompresser could be confused by aligned initrd_end
 458          * We already reserve the end partial page before in
 459          *   - i386_start_kernel()
 460          *   - x86_64_start_kernel()
 461          *   - relocate_initrd()
 462          * So here We can do PAGE_ALIGN() safely to get partial page to be freed
 463          */
 464         free_init_pages("initrd memory", start, PAGE_ALIGN(end));
 465 }
 466 #endif