arch/x86/mm/init_64.c

   1 /*
   2  *  linux/arch/x86_64/mm/init.c
   3  *
   4  *  Copyright (C) 1995  Linus Torvalds
   5  *  Copyright (C) 2000  Pavel Machek <pavel@suse.cz>
   6  *  Copyright (C) 2002,2003 Andi Kleen <ak@suse.de>
   7  */
   8
   9 #include <linux/signal.h>
  10 #include <linux/sched.h>
  11 #include <linux/kernel.h>
  12 #include <linux/errno.h>
  13 #include <linux/string.h>
  14 #include <linux/types.h>
  15 #include <linux/ptrace.h>
  16 #include <linux/mman.h>
  17 #include <linux/mm.h>
  18 #include <linux/swap.h>
  19 #include <linux/smp.h>
  20 #include <linux/init.h>
  21 #include <linux/initrd.h>
  22 #include <linux/pagemap.h>
  23 #include <linux/bootmem.h>
  24 #include <linux/proc_fs.h>
  25 #include <linux/pci.h>
  26 #include <linux/pfn.h>
  27 #include <linux/poison.h>
  28 #include <linux/dma-mapping.h>
  29 #include <linux/module.h>
  30 #include <linux/memory_hotplug.h>
  31 #include <linux/nmi.h>
  32
  33 #include <asm/processor.h>
  34 #include <asm/system.h>
  35 #include <asm/uaccess.h>
  36 #include <asm/pgtable.h>
  37 #include <asm/pgalloc.h>
  38 #include <asm/dma.h>
  39 #include <asm/fixmap.h>
  40 #include <asm/e820.h>
  41 #include <asm/apic.h>
  42 #include <asm/tlb.h>
  43 #include <asm/mmu_context.h>
  44 #include <asm/proto.h>
  45 #include <asm/smp.h>
  46 #include <asm/sections.h>
  47 #include <asm/kdebug.h>
  48 #include <asm/numa.h>
  49 #include <asm/cacheflush.h>
  50
  51 /*
  52  * end_pfn only includes RAM, while max_pfn_mapped includes all e820 entries.
  53  * The direct mapping extends to max_pfn_mapped, so that we can directly access
  54  * apertures, ACPI and other tables without having to play with fixmaps.
  55  */
  56 unsigned long max_low_pfn_mapped;
  57 unsigned long max_pfn_mapped;
  58
  59 static unsigned long dma_reserve __initdata;
  60
  61 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
  62
  63 int direct_gbpages
  64 #ifdef CONFIG_DIRECT_GBPAGES
  65                                 = 1
  66 #endif
  67 ;
  68
  69 static int __init parse_direct_gbpages_off(char *arg)
  70 {
  71         direct_gbpages = 0;
  72         return 0;
  73 }
  74 early_param("nogbpages", parse_direct_gbpages_off);
  75
  76 static int __init parse_direct_gbpages_on(char *arg)
  77 {
  78         direct_gbpages = 1;
  79         return 0;
  80 }
  81 early_param("gbpages", parse_direct_gbpages_on);
  82
  83 /*
  84  * NOTE: pagetable_init alloc all the fixmap pagetables contiguous on the
  85  * physical space so we can cache the place of the first one and move
  86  * around without checking the pgd every time.
  87  */
  88
  89 int after_bootmem;
  90
  91 /*
  92  * NOTE: This function is marked __ref because it calls __init function
  93  * (alloc_bootmem_pages). It's safe to do it ONLY when after_bootmem == 0.
  94  */
  95 static __ref void *spp_getpage(void)
  96 {
  97         void *ptr;
  98
  99         if (after_bootmem)
 100                 ptr = (void *) get_zeroed_page(GFP_ATOMIC);
 101         else
 102                 ptr = alloc_bootmem_pages(PAGE_SIZE);
 103
 104         if (!ptr || ((unsigned long)ptr & ~PAGE_MASK)) {
 105                 panic("set_pte_phys: cannot allocate page data %s\n",
 106                         after_bootmem ? "after bootmem" : "");
 107         }
 108
 109         pr_debug("spp_getpage %p\n", ptr);
 110
 111         return ptr;
 112 }
 113
 114 void
 115 set_pte_vaddr_pud(pud_t *pud_page, unsigned long vaddr, pte_t new_pte)
 116 {
 117         pud_t *pud;
 118         pmd_t *pmd;
 119         pte_t *pte;
 120
 121         pud = pud_page + pud_index(vaddr);
 122         if (pud_none(*pud)) {
 123                 pmd = (pmd_t *) spp_getpage();
 124                 pud_populate(&init_mm, pud, pmd);
 125                 if (pmd != pmd_offset(pud, 0)) {
 126                         printk(KERN_ERR "PAGETABLE BUG #01! %p <-> %p\n",
 127                                 pmd, pmd_offset(pud, 0));
 128                         return;
 129                 }
 130         }
 131         pmd = pmd_offset(pud, vaddr);
 132         if (pmd_none(*pmd)) {
 133                 pte = (pte_t *) spp_getpage();
 134                 pmd_populate_kernel(&init_mm, pmd, pte);
 135                 if (pte != pte_offset_kernel(pmd, 0)) {
 136                         printk(KERN_ERR "PAGETABLE BUG #02!\n");
 137                         return;
 138                 }
 139         }
 140
 141         pte = pte_offset_kernel(pmd, vaddr);
 142         if (!pte_none(*pte) && pte_val(new_pte) &&
 143             pte_val(*pte) != (pte_val(new_pte) & __supported_pte_mask))
 144                 pte_ERROR(*pte);
 145         set_pte(pte, new_pte);
 146
 147         /*
 148          * It's enough to flush this one mapping.
 149          * (PGE mappings get flushed as well)
 150          */
 151         __flush_tlb_one(vaddr);
 152 }
 153
 154 void
 155 set_pte_vaddr(unsigned long vaddr, pte_t pteval)
 156 {
 157         pgd_t *pgd;
 158         pud_t *pud_page;
 159
 160         pr_debug("set_pte_vaddr %lx to %lx\n", vaddr, native_pte_val(pteval));
 161
 162         pgd = pgd_offset_k(vaddr);
 163         if (pgd_none(*pgd)) {
 164                 printk(KERN_ERR
 165                         "PGD FIXMAP MISSING, it should be setup in head.S!\n");
 166                 return;
 167         }
 168         pud_page = (pud_t*)pgd_page_vaddr(*pgd);
 169         set_pte_vaddr_pud(pud_page, vaddr, pteval);
 170 }
 171
 172 /*
 173  * Create large page table mappings for a range of physical addresses.
 174  */
 175 static void __init __init_extra_mapping(unsigned long phys, unsigned long size,
 176                                                 pgprot_t prot)
 177 {
 178         pgd_t *pgd;
 179         pud_t *pud;
 180         pmd_t *pmd;
 181
 182         BUG_ON((phys & ~PMD_MASK) || (size & ~PMD_MASK));
 183         for (; size; phys += PMD_SIZE, size -= PMD_SIZE) {
 184                 pgd = pgd_offset_k((unsigned long)__va(phys));
 185                 if (pgd_none(*pgd)) {
 186                         pud = (pud_t *) spp_getpage();
 187                         set_pgd(pgd, __pgd(__pa(pud) | _KERNPG_TABLE |
 188                                                 _PAGE_USER));
 189                 }
 190                 pud = pud_offset(pgd, (unsigned long)__va(phys));
 191                 if (pud_none(*pud)) {
 192                         pmd = (pmd_t *) spp_getpage();
 193                         set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE |
 194                                                 _PAGE_USER));
 195                 }
 196                 pmd = pmd_offset(pud, phys);
 197                 BUG_ON(!pmd_none(*pmd));
 198                 set_pmd(pmd, __pmd(phys | pgprot_val(prot)));
 199         }
 200 }
 201
 202 void __init init_extra_mapping_wb(unsigned long phys, unsigned long size)
 203 {
 204         __init_extra_mapping(phys, size, PAGE_KERNEL_LARGE);
 205 }
 206
 207 void __init init_extra_mapping_uc(unsigned long phys, unsigned long size)
 208 {
 209         __init_extra_mapping(phys, size, PAGE_KERNEL_LARGE_NOCACHE);
 210 }
 211
 212 /*
 213  * The head.S code sets up the kernel high mapping:
 214  *
 215  *   from __START_KERNEL_map to __START_KERNEL_map + size (== _end-_text)
 216  *
 217  * phys_addr holds the negative offset to the kernel, which is added
 218  * to the compile time generated pmds. This results in invalid pmds up
 219  * to the point where we hit the physaddr 0 mapping.
 220  *
 221  * We limit the mappings to the region from _text to _end.  _end is
 222  * rounded up to the 2MB boundary. This catches the invalid pmds as
 223  * well, as they are located before _text:
 224  */
 225 void __init cleanup_highmap(void)
 226 {
 227         unsigned long vaddr = __START_KERNEL_map;
 228         unsigned long end = round_up((unsigned long)_end, PMD_SIZE) - 1;
 229         pmd_t *pmd = level2_kernel_pgt;
 230         pmd_t *last_pmd = pmd + PTRS_PER_PMD;
 231
 232         for (; pmd < last_pmd; pmd++, vaddr += PMD_SIZE) {
 233                 if (pmd_none(*pmd))
 234                         continue;
 235                 if (vaddr < (unsigned long) _text || vaddr > end)
 236                         set_pmd(pmd, __pmd(0));
 237         }
 238 }
 239
 240 static unsigned long __initdata table_start;
 241 static unsigned long __meminitdata table_end;
 242 static unsigned long __meminitdata table_top;
 243
 244 static __ref void *alloc_low_page(unsigned long *phys)
 245 {
 246         unsigned long pfn = table_end++;
 247         void *adr;
 248
 249         if (after_bootmem) {
 250                 adr = (void *)get_zeroed_page(GFP_ATOMIC);
 251                 *phys = __pa(adr);
 252
 253                 return adr;
 254         }
 255
 256         if (pfn >= table_top)
 257                 panic("alloc_low_page: ran out of memory");
 258
 259         adr = early_ioremap(pfn * PAGE_SIZE, PAGE_SIZE);
 260         memset(adr, 0, PAGE_SIZE);
 261         *phys  = pfn * PAGE_SIZE;
 262         return adr;
 263 }
 264
 265 static __ref void unmap_low_page(void *adr)
 266 {
 267         if (after_bootmem)
 268                 return;
 269
 270         early_iounmap(adr, PAGE_SIZE);
 271 }
 272
 273 static int physical_mapping_iter;
 274
 275 static unsigned long __meminit
 276 phys_pte_init(pte_t *pte_page, unsigned long addr, unsigned long end)
 277 {
 278         unsigned pages = 0;
 279         unsigned long last_map_addr = end;
 280         int i;
 281
 282         pte_t *pte = pte_page + pte_index(addr);
 283
 284         for(i = pte_index(addr); i < PTRS_PER_PTE; i++, addr += PAGE_SIZE, pte++) {
 285
 286                 if (addr >= end) {
 287                         if (!after_bootmem) {
 288                                 for(; i < PTRS_PER_PTE; i++, pte++)
 289                                         set_pte(pte, __pte(0));
 290                         }
 291                         break;
 292                 }
 293
 294                 if (pte_val(*pte))
 295                         goto repeat_set_pte;
 296
 297                 if (0)
 298                         printk("   pte=%p addr=%lx pte=%016lx\n",
 299                                pte, addr, pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL).pte);
 300                 pages++;
 301 repeat_set_pte:
 302                 set_pte(pte, pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL));
 303                 last_map_addr = (addr & PAGE_MASK) + PAGE_SIZE;
 304         }
 305
 306         if (physical_mapping_iter == 1)
 307                 update_page_count(PG_LEVEL_4K, pages);
 308
 309         return last_map_addr;
 310 }
 311
 312 static unsigned long __meminit
 313 phys_pte_update(pmd_t *pmd, unsigned long address, unsigned long end)
 314 {
 315         pte_t *pte = (pte_t *)pmd_page_vaddr(*pmd);
 316
 317         return phys_pte_init(pte, address, end);
 318 }
 319
 320 static unsigned long __meminit
 321 phys_pmd_init(pmd_t *pmd_page, unsigned long address, unsigned long end,
 322                          unsigned long page_size_mask)
 323 {
 324         unsigned long pages = 0;
 325         unsigned long last_map_addr = end;
 326
 327         int i = pmd_index(address);
 328
 329         for (; i < PTRS_PER_PMD; i++, address += PMD_SIZE) {
 330                 unsigned long pte_phys;
 331                 pmd_t *pmd = pmd_page + pmd_index(address);
 332                 pte_t *pte;
 333
 334                 if (address >= end) {
 335                         if (!after_bootmem) {
 336                                 for (; i < PTRS_PER_PMD; i++, pmd++)
 337                                         set_pmd(pmd, __pmd(0));
 338                         }
 339                         break;
 340                 }
 341
 342                 if (pmd_val(*pmd)) {
 343                         if (!pmd_large(*pmd)) {
 344                                 spin_lock(&init_mm.page_table_lock);
 345                                 last_map_addr = phys_pte_update(pmd, address,
 346                                                                 end);
 347                                 spin_unlock(&init_mm.page_table_lock);
 348                                 continue;
 349                         }
 350                         goto repeat_set_pte;
 351                 }
 352
 353                 if (page_size_mask & (1<<PG_LEVEL_2M)) {
 354                         pages++;
 355 repeat_set_pte:
 356                         spin_lock(&init_mm.page_table_lock);
 357                         set_pte((pte_t *)pmd,
 358                                 pfn_pte(address >> PAGE_SHIFT, PAGE_KERNEL_LARGE));
 359                         spin_unlock(&init_mm.page_table_lock);
 360                         last_map_addr = (address & PMD_MASK) + PMD_SIZE;
 361                         continue;
 362                 }
 363
 364                 pte = alloc_low_page(&pte_phys);
 365                 last_map_addr = phys_pte_init(pte, address, end);
 366                 unmap_low_page(pte);
 367
 368                 spin_lock(&init_mm.page_table_lock);
 369                 pmd_populate_kernel(&init_mm, pmd, __va(pte_phys));
 370                 spin_unlock(&init_mm.page_table_lock);
 371         }
 372         if (physical_mapping_iter == 1)
 373                 update_page_count(PG_LEVEL_2M, pages);
 374         return last_map_addr;
 375 }
 376
 377 static unsigned long __meminit
 378 phys_pmd_update(pud_t *pud, unsigned long address, unsigned long end,
 379                          unsigned long page_size_mask)
 380 {
 381         pmd_t *pmd = pmd_offset(pud, 0);
 382         unsigned long last_map_addr;
 383
 384         last_map_addr = phys_pmd_init(pmd, address, end, page_size_mask);
 385         __flush_tlb_all();
 386         return last_map_addr;
 387 }
 388
 389 static unsigned long __meminit
 390 phys_pud_init(pud_t *pud_page, unsigned long addr, unsigned long end,
 391                          unsigned long page_size_mask)
 392 {
 393         unsigned long pages = 0;
 394         unsigned long last_map_addr = end;
 395         int i = pud_index(addr);
 396
 397         for (; i < PTRS_PER_PUD; i++, addr = (addr & PUD_MASK) + PUD_SIZE) {
 398                 unsigned long pmd_phys;
 399                 pud_t *pud = pud_page + pud_index(addr);
 400                 pmd_t *pmd;
 401
 402                 if (addr >= end)
 403                         break;
 404
 405                 if (!after_bootmem &&
 406                                 !e820_any_mapped(addr, addr+PUD_SIZE, 0)) {
 407                         set_pud(pud, __pud(0));
 408                         continue;
 409                 }
 410
 411                 if (pud_val(*pud)) {
 412                         if (!pud_large(*pud)) {
 413                                 last_map_addr = phys_pmd_update(pud, addr, end,
 414                                                          page_size_mask);
 415                                 continue;
 416                         }
 417
 418                         goto repeat_set_pte;
 419                 }
 420
 421                 if (page_size_mask & (1<<PG_LEVEL_1G)) {
 422                         pages++;
 423 repeat_set_pte:
 424                         spin_lock(&init_mm.page_table_lock);
 425                         set_pte((pte_t *)pud,
 426                                 pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL_LARGE));
 427                         spin_unlock(&init_mm.page_table_lock);
 428                         last_map_addr = (addr & PUD_MASK) + PUD_SIZE;
 429                         continue;
 430                 }
 431
 432                 pmd = alloc_low_page(&pmd_phys);
 433                 last_map_addr = phys_pmd_init(pmd, addr, end, page_size_mask);
 434                 unmap_low_page(pmd);
 435
 436                 spin_lock(&init_mm.page_table_lock);
 437                 pud_populate(&init_mm, pud, __va(pmd_phys));
 438                 spin_unlock(&init_mm.page_table_lock);
 439         }
 440         __flush_tlb_all();
 441
 442         if (physical_mapping_iter == 1)
 443                 update_page_count(PG_LEVEL_1G, pages);
 444
 445         return last_map_addr;
 446 }
 447
 448 static unsigned long __meminit
 449 phys_pud_update(pgd_t *pgd, unsigned long addr, unsigned long end,
 450                  unsigned long page_size_mask)
 451 {
 452         pud_t *pud;
 453
 454         pud = (pud_t *)pgd_page_vaddr(*pgd);
 455
 456         return phys_pud_init(pud, addr, end, page_size_mask);
 457 }
 458
 459 static void __init find_early_table_space(unsigned long end)
 460 {
 461         unsigned long puds, pmds, ptes, tables, start;
 462
 463         puds = (end + PUD_SIZE - 1) >> PUD_SHIFT;
 464         tables = round_up(puds * sizeof(pud_t), PAGE_SIZE);
 465         if (direct_gbpages) {
 466                 unsigned long extra;
 467                 extra = end - ((end>>PUD_SHIFT) << PUD_SHIFT);
 468                 pmds = (extra + PMD_SIZE - 1) >> PMD_SHIFT;
 469         } else
 470                 pmds = (end + PMD_SIZE - 1) >> PMD_SHIFT;
 471         tables += round_up(pmds * sizeof(pmd_t), PAGE_SIZE);
 472
 473         if (cpu_has_pse) {
 474                 unsigned long extra;
 475                 extra = end - ((end>>PMD_SHIFT) << PMD_SHIFT);
 476                 ptes = (extra + PAGE_SIZE - 1) >> PAGE_SHIFT;
 477         } else
 478                 ptes = (end + PAGE_SIZE - 1) >> PAGE_SHIFT;
 479         tables += round_up(ptes * sizeof(pte_t), PAGE_SIZE);
 480
 481         /*
 482          * RED-PEN putting page tables only on node 0 could
 483          * cause a hotspot and fill up ZONE_DMA. The page tables
 484          * need roughly 0.5KB per GB.
 485          */
 486         start = 0x8000;
 487         table_start = find_e820_area(start, end, tables, PAGE_SIZE);
 488         if (table_start == -1UL)
 489                 panic("Cannot find space for the kernel page tables");
 490
 491         table_start >>= PAGE_SHIFT;
 492         table_end = table_start;
 493         table_top = table_start + (tables >> PAGE_SHIFT);
 494
 495         printk(KERN_DEBUG "kernel direct mapping tables up to %lx @ %lx-%lx\n",
 496                 end, table_start << PAGE_SHIFT, table_top << PAGE_SHIFT);
 497 }
 498
 499 static void __init init_gbpages(void)
 500 {
 501         if (direct_gbpages && cpu_has_gbpages)
 502                 printk(KERN_INFO "Using GB pages for direct mapping\n");
 503         else
 504                 direct_gbpages = 0;
 505 }
 506
 507 static int is_kernel(unsigned long pfn)
 508 {
 509         unsigned long pg_addresss = pfn << PAGE_SHIFT;
 510
 511         if (pg_addresss >= (unsigned long) __pa(_text) &&
 512             pg_addresss <= (unsigned long) __pa(_end))
 513                 return 1;
 514
 515         return 0;
 516 }
 517
 518 static unsigned long __init kernel_physical_mapping_init(unsigned long start,
 519                                                 unsigned long end,
 520                                                 unsigned long page_size_mask)
 521 {
 522
 523         unsigned long next, last_map_addr;
 524         u64 cached_supported_pte_mask = __supported_pte_mask;
 525         unsigned long cache_start = start;
 526         unsigned long cache_end = end;
 527
 528         /*
 529          * First iteration will setup identity mapping using large/small pages
 530          * based on page_size_mask, with other attributes same as set by
 531          * the early code in head_64.S
 532          *
 533          * Second iteration will setup the appropriate attributes
 534          * as desired for the kernel identity mapping.
 535          *
 536          * This two pass mechanism conforms to the TLB app note which says:
 537          *
 538          *     "Software should not write to a paging-structure entry in a way
 539          *      that would change, for any linear address, both the page size
 540          *      and either the page frame or attributes."
 541          *
 542          * For now, only difference between very early PTE attributes used in
 543          * head_64.S and here is _PAGE_NX.
 544          */
 545         BUILD_BUG_ON((__PAGE_KERNEL_LARGE & ~__PAGE_KERNEL_IDENT_LARGE_EXEC)
 546                      != _PAGE_NX);
 547         __supported_pte_mask &= ~(_PAGE_NX);
 548         physical_mapping_iter = 1;
 549
 550 repeat:
 551         last_map_addr = cache_end;
 552
 553         start = (unsigned long)__va(cache_start);
 554         end = (unsigned long)__va(cache_end);
 555
 556         for (; start < end; start = next) {
 557                 pgd_t *pgd = pgd_offset_k(start);
 558                 unsigned long pud_phys;
 559                 pud_t *pud;
 560
 561                 next = (start + PGDIR_SIZE) & PGDIR_MASK;
 562                 if (next > end)
 563                         next = end;
 564
 565                 if (pgd_val(*pgd)) {
 566                         /*
 567                          * Static identity mappings will be overwritten
 568                          * with run-time mappings. For example, this allows
 569                          * the static 0-1GB identity mapping to be mapped
 570                          * non-executable with this.
 571                          */
 572                         if (is_kernel(pte_pfn(*((pte_t *) pgd))))
 573                                 goto realloc;
 574
 575                         last_map_addr = phys_pud_update(pgd, __pa(start),
 576                                                  __pa(end), page_size_mask);
 577                         continue;
 578                 }
 579
 580 realloc:
 581                 pud = alloc_low_page(&pud_phys);
 582                 last_map_addr = phys_pud_init(pud, __pa(start), __pa(next),
 583                                                  page_size_mask);
 584                 unmap_low_page(pud);
 585
 586                 spin_lock(&init_mm.page_table_lock);
 587                 pgd_populate(&init_mm, pgd, __va(pud_phys));
 588                 spin_unlock(&init_mm.page_table_lock);
 589         }
 590         __flush_tlb_all();
 591
 592         if (physical_mapping_iter == 1) {
 593                 physical_mapping_iter = 2;
 594                 /*
 595                  * Second iteration will set the actual desired PTE attributes.
 596                  */
 597                 __supported_pte_mask = cached_supported_pte_mask;
 598                 goto repeat;
 599         }
 600
 601         return last_map_addr;
 602 }
 603
 604 struct map_range {
 605         unsigned long start;
 606         unsigned long end;
 607         unsigned page_size_mask;
 608 };
 609
 610 #define NR_RANGE_MR 5
 611
 612 static int save_mr(struct map_range *mr, int nr_range,
 613                    unsigned long start_pfn, unsigned long end_pfn,
 614                    unsigned long page_size_mask)
 615 {
 616
 617         if (start_pfn < end_pfn) {
 618                 if (nr_range >= NR_RANGE_MR)
 619                         panic("run out of range for init_memory_mapping\n");
 620                 mr[nr_range].start = start_pfn<<PAGE_SHIFT;
 621                 mr[nr_range].end   = end_pfn<<PAGE_SHIFT;
 622                 mr[nr_range].page_size_mask = page_size_mask;
 623                 nr_range++;
 624         }
 625
 626         return nr_range;
 627 }
 628
 629 /*
 630  * Setup the direct mapping of the physical memory at PAGE_OFFSET.
 631  * This runs before bootmem is initialized and gets pages directly from
 632  * the physical memory. To access them they are temporarily mapped.
 633  */
 634 unsigned long __init_refok init_memory_mapping(unsigned long start,
 635                                                unsigned long end)
 636 {
 637         unsigned long last_map_addr = 0;
 638         unsigned long page_size_mask = 0;
 639         unsigned long start_pfn, end_pfn;
 640
 641         struct map_range mr[NR_RANGE_MR];
 642         int nr_range, i;
 643
 644         printk(KERN_INFO "init_memory_mapping\n");
 645
 646         /*
 647          * Find space for the kernel direct mapping tables.
 648          *
 649          * Later we should allocate these tables in the local node of the
 650          * memory mapped. Unfortunately this is done currently before the
 651          * nodes are discovered.
 652          */
 653         if (!after_bootmem)
 654                 init_gbpages();
 655
 656         if (direct_gbpages)
 657                 page_size_mask |= 1 << PG_LEVEL_1G;
 658         if (cpu_has_pse)
 659                 page_size_mask |= 1 << PG_LEVEL_2M;
 660
 661         memset(mr, 0, sizeof(mr));
 662         nr_range = 0;
 663
 664         /* head if not big page alignment ?*/
 665         start_pfn = start >> PAGE_SHIFT;
 666         end_pfn = ((start + (PMD_SIZE - 1)) >> PMD_SHIFT)
 667                         << (PMD_SHIFT - PAGE_SHIFT);
 668         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
 669
 670         /* big page (2M) range*/
 671         start_pfn = ((start + (PMD_SIZE - 1))>>PMD_SHIFT)
 672                          << (PMD_SHIFT - PAGE_SHIFT);
 673         end_pfn = ((start + (PUD_SIZE - 1))>>PUD_SHIFT)
 674                          << (PUD_SHIFT - PAGE_SHIFT);
 675         if (end_pfn > ((end>>PUD_SHIFT)<<(PUD_SHIFT - PAGE_SHIFT)))
 676                 end_pfn = ((end>>PUD_SHIFT)<<(PUD_SHIFT - PAGE_SHIFT));
 677         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 678                         page_size_mask & (1<<PG_LEVEL_2M));
 679
 680         /* big page (1G) range */
 681         start_pfn = end_pfn;
 682         end_pfn = (end>>PUD_SHIFT) << (PUD_SHIFT - PAGE_SHIFT);
 683         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 684                                 page_size_mask &
 685                                  ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G)));
 686
 687         /* tail is not big page (1G) alignment */
 688         start_pfn = end_pfn;
 689         end_pfn = (end>>PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
 690         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 691                         page_size_mask & (1<<PG_LEVEL_2M));
 692
 693         /* tail is not big page (2M) alignment */
 694         start_pfn = end_pfn;
 695         end_pfn = end>>PAGE_SHIFT;
 696         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
 697
 698         /* try to merge same page size and continuous */
 699         for (i = 0; nr_range > 1 && i < nr_range - 1; i++) {
 700                 unsigned long old_start;
 701                 if (mr[i].end != mr[i+1].start ||
 702                     mr[i].page_size_mask != mr[i+1].page_size_mask)
 703                         continue;
 704                 /* move it */
 705                 old_start = mr[i].start;
 706                 memmove(&mr[i], &mr[i+1],
 707                          (nr_range - 1 - i) * sizeof (struct map_range));
 708                 mr[i].start = old_start;
 709                 nr_range--;
 710         }
 711
 712         for (i = 0; i < nr_range; i++)
 713                 printk(KERN_DEBUG " %010lx - %010lx page %s\n",
 714                                 mr[i].start, mr[i].end,
 715                         (mr[i].page_size_mask & (1<<PG_LEVEL_1G))?"1G":(
 716                          (mr[i].page_size_mask & (1<<PG_LEVEL_2M))?"2M":"4k"));
 717
 718         if (!after_bootmem)
 719                 find_early_table_space(end);
 720
 721         for (i = 0; i < nr_range; i++)
 722                 last_map_addr = kernel_physical_mapping_init(
 723                                         mr[i].start, mr[i].end,
 724                                         mr[i].page_size_mask);
 725
 726         if (!after_bootmem)
 727                 mmu_cr4_features = read_cr4();
 728         __flush_tlb_all();
 729
 730         if (!after_bootmem && table_end > table_start)
 731                 reserve_early(table_start << PAGE_SHIFT,
 732                                  table_end << PAGE_SHIFT, "PGTABLE");
 733
 734         printk(KERN_INFO "last_map_addr: %lx end: %lx\n",
 735                          last_map_addr, end);
 736
 737         if (!after_bootmem)
 738                 early_memtest(start, end);
 739
 740         return last_map_addr >> PAGE_SHIFT;
 741 }
 742
 743 #ifndef CONFIG_NUMA
 744 void __init initmem_init(unsigned long start_pfn, unsigned long end_pfn)
 745 {
 746         unsigned long bootmap_size, bootmap;
 747
 748         bootmap_size = bootmem_bootmap_pages(end_pfn)<<PAGE_SHIFT;
 749         bootmap = find_e820_area(0, end_pfn<<PAGE_SHIFT, bootmap_size,
 750                                  PAGE_SIZE);
 751         if (bootmap == -1L)
 752                 panic("Cannot find bootmem map of size %ld\n", bootmap_size);
 753         /* don't touch min_low_pfn */
 754         bootmap_size = init_bootmem_node(NODE_DATA(0), bootmap >> PAGE_SHIFT,
 755                                          0, end_pfn);
 756         e820_register_active_regions(0, start_pfn, end_pfn);
 757         free_bootmem_with_active_regions(0, end_pfn);
 758         early_res_to_bootmem(0, end_pfn<<PAGE_SHIFT);
 759         reserve_bootmem(bootmap, bootmap_size, BOOTMEM_DEFAULT);
 760 }
 761
 762 void __init paging_init(void)
 763 {
 764         unsigned long max_zone_pfns[MAX_NR_ZONES];
 765
 766         memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
 767         max_zone_pfns[ZONE_DMA] = MAX_DMA_PFN;
 768         max_zone_pfns[ZONE_DMA32] = MAX_DMA32_PFN;
 769         max_zone_pfns[ZONE_NORMAL] = max_pfn;
 770
 771         memory_present(0, 0, max_pfn);
 772         sparse_init();
 773         free_area_init_nodes(max_zone_pfns);
 774 }
 775 #endif
 776
 777 /*
 778  * Memory hotplug specific functions
 779  */
 780 #ifdef CONFIG_MEMORY_HOTPLUG
 781 /*
 782  * Memory is added always to NORMAL zone. This means you will never get
 783  * additional DMA/DMA32 memory.
 784  */
 785 int arch_add_memory(int nid, u64 start, u64 size)
 786 {
 787         struct pglist_data *pgdat = NODE_DATA(nid);
 788         struct zone *zone = pgdat->node_zones + ZONE_NORMAL;
 789         unsigned long last_mapped_pfn, start_pfn = start >> PAGE_SHIFT;
 790         unsigned long nr_pages = size >> PAGE_SHIFT;
 791         int ret;
 792
 793         last_mapped_pfn = init_memory_mapping(start, start + size-1);
 794         if (last_mapped_pfn > max_pfn_mapped)
 795                 max_pfn_mapped = last_mapped_pfn;
 796
 797         ret = __add_pages(zone, start_pfn, nr_pages);
 798         WARN_ON(1);
 799
 800         return ret;
 801 }
 802 EXPORT_SYMBOL_GPL(arch_add_memory);
 803
 804 #if !defined(CONFIG_ACPI_NUMA) && defined(CONFIG_NUMA)
 805 int memory_add_physaddr_to_nid(u64 start)
 806 {
 807         return 0;
 808 }
 809 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
 810 #endif
 811
 812 #endif /* CONFIG_MEMORY_HOTPLUG */
 813
 814 /*
 815  * devmem_is_allowed() checks to see if /dev/mem access to a certain address
 816  * is valid. The argument is a physical page number.
 817  *
 818  *
 819  * On x86, access has to be given to the first megabyte of ram because that area
 820  * contains bios code and data regions used by X and dosemu and similar apps.
 821  * Access has to be given to non-kernel-ram areas as well, these contain the PCI
 822  * mmio resources as well as potential bios/acpi data regions.
 823  */
 824 int devmem_is_allowed(unsigned long pagenr)
 825 {
 826         if (pagenr <= 256)
 827                 return 1;
 828         if (!page_is_ram(pagenr))
 829                 return 1;
 830         return 0;
 831 }
 832
 833
 834 static struct kcore_list kcore_mem, kcore_vmalloc, kcore_kernel,
 835                          kcore_modules, kcore_vsyscall;
 836
 837 void __init mem_init(void)
 838 {
 839         long codesize, reservedpages, datasize, initsize;
 840
 841         pci_iommu_alloc();
 842
 843         /* clear_bss() already clear the empty_zero_page */
 844
 845         reservedpages = 0;
 846
 847         /* this will put all low memory onto the freelists */
 848 #ifdef CONFIG_NUMA
 849         totalram_pages = numa_free_all_bootmem();
 850 #else
 851         totalram_pages = free_all_bootmem();
 852 #endif
 853         reservedpages = max_pfn - totalram_pages -
 854                                         absent_pages_in_range(0, max_pfn);
 855         after_bootmem = 1;
 856
 857         codesize =  (unsigned long) &_etext - (unsigned long) &_text;
 858         datasize =  (unsigned long) &_edata - (unsigned long) &_etext;
 859         initsize =  (unsigned long) &__init_end - (unsigned long) &__init_begin;
 860
 861         /* Register memory areas for /proc/kcore */
 862         kclist_add(&kcore_mem, __va(0), max_low_pfn << PAGE_SHIFT);
 863         kclist_add(&kcore_vmalloc, (void *)VMALLOC_START,
 864                    VMALLOC_END-VMALLOC_START);
 865         kclist_add(&kcore_kernel, &_stext, _end - _stext);
 866         kclist_add(&kcore_modules, (void *)MODULES_VADDR, MODULES_LEN);
 867         kclist_add(&kcore_vsyscall, (void *)VSYSCALL_START,
 868                                  VSYSCALL_END - VSYSCALL_START);
 869
 870         printk(KERN_INFO "Memory: %luk/%luk available (%ldk kernel code, "
 871                                 "%ldk reserved, %ldk data, %ldk init)\n",
 872                 (unsigned long) nr_free_pages() << (PAGE_SHIFT-10),
 873                 max_pfn << (PAGE_SHIFT-10),
 874                 codesize >> 10,
 875                 reservedpages << (PAGE_SHIFT-10),
 876                 datasize >> 10,
 877                 initsize >> 10);
 878
 879         cpa_init();
 880 }
 881
 882 void free_init_pages(char *what, unsigned long begin, unsigned long end)
 883 {
 884         unsigned long addr = begin;
 885
 886         if (addr >= end)
 887                 return;
 888
 889         /*
 890          * If debugging page accesses then do not free this memory but
 891          * mark them not present - any buggy init-section access will
 892          * create a kernel page fault:
 893          */
 894 #ifdef CONFIG_DEBUG_PAGEALLOC
 895         printk(KERN_INFO "debug: unmapping init memory %08lx..%08lx\n",
 896                 begin, PAGE_ALIGN(end));
 897         set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
 898 #else
 899         printk(KERN_INFO "Freeing %s: %luk freed\n", what, (end - begin) >> 10);
 900
 901         for (; addr < end; addr += PAGE_SIZE) {
 902                 ClearPageReserved(virt_to_page(addr));
 903                 init_page_count(virt_to_page(addr));
 904                 memset((void *)(addr & ~(PAGE_SIZE-1)),
 905                         POISON_FREE_INITMEM, PAGE_SIZE);
 906                 free_page(addr);
 907                 totalram_pages++;
 908         }
 909 #endif
 910 }
 911
 912 void free_initmem(void)
 913 {
 914         free_init_pages("unused kernel memory",
 915                         (unsigned long)(&__init_begin),
 916                         (unsigned long)(&__init_end));
 917 }
 918
 919 #ifdef CONFIG_DEBUG_RODATA
 920 const int rodata_test_data = 0xC3;
 921 EXPORT_SYMBOL_GPL(rodata_test_data);
 922
 923 void mark_rodata_ro(void)
 924 {
 925         unsigned long start = PFN_ALIGN(_stext), end = PFN_ALIGN(__end_rodata);
 926         unsigned long rodata_start =
 927                 ((unsigned long)__start_rodata + PAGE_SIZE - 1) & PAGE_MASK;
 928
 929 #ifdef CONFIG_DYNAMIC_FTRACE
 930         /* Dynamic tracing modifies the kernel text section */
 931         start = rodata_start;
 932 #endif
 933
 934         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
 935                (end - start) >> 10);
 936         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
 937
 938         /*
 939          * The rodata section (but not the kernel text!) should also be
 940          * not-executable.
 941          */
 942         set_memory_nx(rodata_start, (end - rodata_start) >> PAGE_SHIFT);
 943
 944         rodata_test();
 945
 946 #ifdef CONFIG_CPA_DEBUG
 947         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
 948         set_memory_rw(start, (end-start) >> PAGE_SHIFT);
 949
 950         printk(KERN_INFO "Testing CPA: again\n");
 951         set_memory_ro(start, (end-start) >> PAGE_SHIFT);
 952 #endif
 953 }
 954
 955 #endif
 956
 957 #ifdef CONFIG_BLK_DEV_INITRD
 958 void free_initrd_mem(unsigned long start, unsigned long end)
 959 {
 960         free_init_pages("initrd memory", start, end);
 961 }
 962 #endif
 963
 964 int __init reserve_bootmem_generic(unsigned long phys, unsigned long len,
 965                                    int flags)
 966 {
 967 #ifdef CONFIG_NUMA
 968         int nid, next_nid;
 969         int ret;
 970 #endif
 971         unsigned long pfn = phys >> PAGE_SHIFT;
 972
 973         if (pfn >= max_pfn) {
 974                 /*
 975                  * This can happen with kdump kernels when accessing
 976                  * firmware tables:
 977                  */
 978                 if (pfn < max_pfn_mapped)
 979                         return -EFAULT;
 980
 981                 printk(KERN_ERR "reserve_bootmem: illegal reserve %lx %lu\n",
 982                                 phys, len);
 983                 return -EFAULT;
 984         }
 985
 986         /* Should check here against the e820 map to avoid double free */
 987 #ifdef CONFIG_NUMA
 988         nid = phys_to_nid(phys);
 989         next_nid = phys_to_nid(phys + len - 1);
 990         if (nid == next_nid)
 991                 ret = reserve_bootmem_node(NODE_DATA(nid), phys, len, flags);
 992         else
 993                 ret = reserve_bootmem(phys, len, flags);
 994
 995         if (ret != 0)
 996                 return ret;
 997
 998 #else
 999         reserve_bootmem(phys, len, BOOTMEM_DEFAULT);
1000 #endif
1001
1002         if (phys+len <= MAX_DMA_PFN*PAGE_SIZE) {
1003                 dma_reserve += len / PAGE_SIZE;
1004                 set_dma_reserve(dma_reserve);
1005         }
1006
1007         return 0;
1008 }
1009
1010 int kern_addr_valid(unsigned long addr)
1011 {
1012         unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
1013         pgd_t *pgd;
1014         pud_t *pud;
1015         pmd_t *pmd;
1016         pte_t *pte;
1017
1018         if (above != 0 && above != -1UL)
1019                 return 0;
1020
1021         pgd = pgd_offset_k(addr);
1022         if (pgd_none(*pgd))
1023                 return 0;
1024
1025         pud = pud_offset(pgd, addr);
1026         if (pud_none(*pud))
1027                 return 0;
1028
1029         pmd = pmd_offset(pud, addr);
1030         if (pmd_none(*pmd))
1031                 return 0;
1032
1033         if (pmd_large(*pmd))
1034                 return pfn_valid(pmd_pfn(*pmd));
1035
1036         pte = pte_offset_kernel(pmd, addr);
1037         if (pte_none(*pte))
1038                 return 0;
1039
1040         return pfn_valid(pte_pfn(*pte));
1041 }
1042
1043 /*
1044  * A pseudo VMA to allow ptrace access for the vsyscall page.  This only
1045  * covers the 64bit vsyscall page now. 32bit has a real VMA now and does
1046  * not need special handling anymore:
1047  */
1048 static struct vm_area_struct gate_vma = {
1049         .vm_start       = VSYSCALL_START,
1050         .vm_end         = VSYSCALL_START + (VSYSCALL_MAPPED_PAGES * PAGE_SIZE),
1051         .vm_page_prot   = PAGE_READONLY_EXEC,
1052         .vm_flags       = VM_READ | VM_EXEC
1053 };
1054
1055 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
1056 {
1057 #ifdef CONFIG_IA32_EMULATION
1058         if (test_tsk_thread_flag(tsk, TIF_IA32))
1059                 return NULL;
1060 #endif
1061         return &gate_vma;
1062 }
1063
1064 int in_gate_area(struct task_struct *task, unsigned long addr)
1065 {
1066         struct vm_area_struct *vma = get_gate_vma(task);
1067
1068         if (!vma)
1069                 return 0;
1070
1071         return (addr >= vma->vm_start) && (addr < vma->vm_end);
1072 }
1073
1074 /*
1075  * Use this when you have no reliable task/vma, typically from interrupt
1076  * context. It is less reliable than using the task's vma and may give
1077  * false positives:
1078  */
1079 int in_gate_area_no_task(unsigned long addr)
1080 {
1081         return (addr >= VSYSCALL_START) && (addr < VSYSCALL_END);
1082 }
1083
1084 const char *arch_vma_name(struct vm_area_struct *vma)
1085 {
1086         if (vma->vm_mm && vma->vm_start == (long)vma->vm_mm->context.vdso)
1087                 return "[vdso]";
1088         if (vma == &gate_vma)
1089                 return "[vsyscall]";
1090         return NULL;
1091 }
1092
1093 #ifdef CONFIG_SPARSEMEM_VMEMMAP
1094 /*
1095  * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
1096  */
1097 static long __meminitdata addr_start, addr_end;
1098 static void __meminitdata *p_start, *p_end;
1099 static int __meminitdata node_start;
1100
1101 int __meminit
1102 vmemmap_populate(struct page *start_page, unsigned long size, int node)
1103 {
1104         unsigned long addr = (unsigned long)start_page;
1105         unsigned long end = (unsigned long)(start_page + size);
1106         unsigned long next;
1107         pgd_t *pgd;
1108         pud_t *pud;
1109         pmd_t *pmd;
1110
1111         for (; addr < end; addr = next) {
1112                 void *p = NULL;
1113
1114                 pgd = vmemmap_pgd_populate(addr, node);
1115                 if (!pgd)
1116                         return -ENOMEM;
1117
1118                 pud = vmemmap_pud_populate(pgd, addr, node);
1119                 if (!pud)
1120                         return -ENOMEM;
1121
1122                 if (!cpu_has_pse) {
1123                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1124                         pmd = vmemmap_pmd_populate(pud, addr, node);
1125
1126                         if (!pmd)
1127                                 return -ENOMEM;
1128
1129                         p = vmemmap_pte_populate(pmd, addr, node);
1130
1131                         if (!p)
1132                                 return -ENOMEM;
1133
1134                         addr_end = addr + PAGE_SIZE;
1135                         p_end = p + PAGE_SIZE;
1136                 } else {
1137                         next = pmd_addr_end(addr, end);
1138
1139                         pmd = pmd_offset(pud, addr);
1140                         if (pmd_none(*pmd)) {
1141                                 pte_t entry;
1142
1143                                 p = vmemmap_alloc_block(PMD_SIZE, node);
1144                                 if (!p)
1145                                         return -ENOMEM;
1146
1147                                 entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
1148                                                 PAGE_KERNEL_LARGE);
1149                                 set_pmd(pmd, __pmd(pte_val(entry)));
1150
1151                                 /* check to see if we have contiguous blocks */
1152                                 if (p_end != p || node_start != node) {
1153                                         if (p_start)
1154                                                 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1155                                                        addr_start, addr_end-1, p_start, p_end-1, node_start);
1156                                         addr_start = addr;
1157                                         node_start = node;
1158                                         p_start = p;
1159                                 }
1160
1161                                 addr_end = addr + PMD_SIZE;
1162                                 p_end = p + PMD_SIZE;
1163                         } else
1164                                 vmemmap_verify((pte_t *)pmd, node, addr, next);
1165                 }
1166
1167         }
1168         return 0;
1169 }
1170
1171 void __meminit vmemmap_populate_print_last(void)
1172 {
1173         if (p_start) {
1174                 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1175                         addr_start, addr_end-1, p_start, p_end-1, node_start);
1176                 p_start = NULL;
1177                 p_end = NULL;
1178                 node_start = 0;
1179         }
1180 }
1181 #endif