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 = roundup((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 unsigned long __meminit
 274 phys_pte_init(pte_t *pte_page, unsigned long addr, unsigned long end)
 275 {
 276         unsigned pages = 0;
 277         unsigned long last_map_addr = end;
 278         int i;
 279
 280         pte_t *pte = pte_page + pte_index(addr);
 281
 282         for(i = pte_index(addr); i < PTRS_PER_PTE; i++, addr += PAGE_SIZE, pte++) {
 283
 284                 if (addr >= end) {
 285                         if (!after_bootmem) {
 286                                 for(; i < PTRS_PER_PTE; i++, pte++)
 287                                         set_pte(pte, __pte(0));
 288                         }
 289                         break;
 290                 }
 291
 292                 if (pte_val(*pte))
 293                         continue;
 294
 295                 if (0)
 296                         printk("   pte=%p addr=%lx pte=%016lx\n",
 297                                pte, addr, pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL).pte);
 298                 set_pte(pte, pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL));
 299                 last_map_addr = (addr & PAGE_MASK) + PAGE_SIZE;
 300                 pages++;
 301         }
 302         update_page_count(PG_LEVEL_4K, pages);
 303
 304         return last_map_addr;
 305 }
 306
 307 static unsigned long __meminit
 308 phys_pte_update(pmd_t *pmd, unsigned long address, unsigned long end)
 309 {
 310         pte_t *pte = (pte_t *)pmd_page_vaddr(*pmd);
 311
 312         return phys_pte_init(pte, address, end);
 313 }
 314
 315 static unsigned long __meminit
 316 phys_pmd_init(pmd_t *pmd_page, unsigned long address, unsigned long end,
 317                          unsigned long page_size_mask)
 318 {
 319         unsigned long pages = 0;
 320         unsigned long last_map_addr = end;
 321         unsigned long start = address;
 322
 323         int i = pmd_index(address);
 324
 325         for (; i < PTRS_PER_PMD; i++, address += PMD_SIZE) {
 326                 unsigned long pte_phys;
 327                 pmd_t *pmd = pmd_page + pmd_index(address);
 328                 pte_t *pte;
 329
 330                 if (address >= end) {
 331                         if (!after_bootmem) {
 332                                 for (; i < PTRS_PER_PMD; i++, pmd++)
 333                                         set_pmd(pmd, __pmd(0));
 334                         }
 335                         break;
 336                 }
 337
 338                 if (pmd_val(*pmd)) {
 339                         if (!pmd_large(*pmd)) {
 340                                 spin_lock(&init_mm.page_table_lock);
 341                                 last_map_addr = phys_pte_update(pmd, address,
 342                                                                 end);
 343                                 spin_unlock(&init_mm.page_table_lock);
 344                         }
 345                         /* Count entries we're using from level2_ident_pgt */
 346                         if (start == 0)
 347                                 pages++;
 348                         continue;
 349                 }
 350
 351                 if (page_size_mask & (1<<PG_LEVEL_2M)) {
 352                         pages++;
 353                         spin_lock(&init_mm.page_table_lock);
 354                         set_pte((pte_t *)pmd,
 355                                 pfn_pte(address >> PAGE_SHIFT, PAGE_KERNEL_LARGE));
 356                         spin_unlock(&init_mm.page_table_lock);
 357                         last_map_addr = (address & PMD_MASK) + PMD_SIZE;
 358                         continue;
 359                 }
 360
 361                 pte = alloc_low_page(&pte_phys);
 362                 last_map_addr = phys_pte_init(pte, address, end);
 363                 unmap_low_page(pte);
 364
 365                 spin_lock(&init_mm.page_table_lock);
 366                 pmd_populate_kernel(&init_mm, pmd, __va(pte_phys));
 367                 spin_unlock(&init_mm.page_table_lock);
 368         }
 369         update_page_count(PG_LEVEL_2M, pages);
 370         return last_map_addr;
 371 }
 372
 373 static unsigned long __meminit
 374 phys_pmd_update(pud_t *pud, unsigned long address, unsigned long end,
 375                          unsigned long page_size_mask)
 376 {
 377         pmd_t *pmd = pmd_offset(pud, 0);
 378         unsigned long last_map_addr;
 379
 380         last_map_addr = phys_pmd_init(pmd, address, end, page_size_mask);
 381         __flush_tlb_all();
 382         return last_map_addr;
 383 }
 384
 385 static unsigned long __meminit
 386 phys_pud_init(pud_t *pud_page, unsigned long addr, unsigned long end,
 387                          unsigned long page_size_mask)
 388 {
 389         unsigned long pages = 0;
 390         unsigned long last_map_addr = end;
 391         int i = pud_index(addr);
 392
 393         for (; i < PTRS_PER_PUD; i++, addr = (addr & PUD_MASK) + PUD_SIZE) {
 394                 unsigned long pmd_phys;
 395                 pud_t *pud = pud_page + pud_index(addr);
 396                 pmd_t *pmd;
 397
 398                 if (addr >= end)
 399                         break;
 400
 401                 if (!after_bootmem &&
 402                                 !e820_any_mapped(addr, addr+PUD_SIZE, 0)) {
 403                         set_pud(pud, __pud(0));
 404                         continue;
 405                 }
 406
 407                 if (pud_val(*pud)) {
 408                         if (!pud_large(*pud))
 409                                 last_map_addr = phys_pmd_update(pud, addr, end,
 410                                                          page_size_mask);
 411                         continue;
 412                 }
 413
 414                 if (page_size_mask & (1<<PG_LEVEL_1G)) {
 415                         pages++;
 416                         spin_lock(&init_mm.page_table_lock);
 417                         set_pte((pte_t *)pud,
 418                                 pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL_LARGE));
 419                         spin_unlock(&init_mm.page_table_lock);
 420                         last_map_addr = (addr & PUD_MASK) + PUD_SIZE;
 421                         continue;
 422                 }
 423
 424                 pmd = alloc_low_page(&pmd_phys);
 425                 last_map_addr = phys_pmd_init(pmd, addr, end, page_size_mask);
 426                 unmap_low_page(pmd);
 427
 428                 spin_lock(&init_mm.page_table_lock);
 429                 pud_populate(&init_mm, pud, __va(pmd_phys));
 430                 spin_unlock(&init_mm.page_table_lock);
 431         }
 432         __flush_tlb_all();
 433         update_page_count(PG_LEVEL_1G, pages);
 434
 435         return last_map_addr;
 436 }
 437
 438 static unsigned long __meminit
 439 phys_pud_update(pgd_t *pgd, unsigned long addr, unsigned long end,
 440                  unsigned long page_size_mask)
 441 {
 442         pud_t *pud;
 443
 444         pud = (pud_t *)pgd_page_vaddr(*pgd);
 445
 446         return phys_pud_init(pud, addr, end, page_size_mask);
 447 }
 448
 449 static void __init find_early_table_space(unsigned long end)
 450 {
 451         unsigned long puds, pmds, ptes, tables, start;
 452
 453         puds = (end + PUD_SIZE - 1) >> PUD_SHIFT;
 454         tables = roundup(puds * sizeof(pud_t), PAGE_SIZE);
 455         if (direct_gbpages) {
 456                 unsigned long extra;
 457                 extra = end - ((end>>PUD_SHIFT) << PUD_SHIFT);
 458                 pmds = (extra + PMD_SIZE - 1) >> PMD_SHIFT;
 459         } else
 460                 pmds = (end + PMD_SIZE - 1) >> PMD_SHIFT;
 461         tables += roundup(pmds * sizeof(pmd_t), PAGE_SIZE);
 462
 463         if (cpu_has_pse) {
 464                 unsigned long extra;
 465                 extra = end - ((end>>PMD_SHIFT) << PMD_SHIFT);
 466                 ptes = (extra + PAGE_SIZE - 1) >> PAGE_SHIFT;
 467         } else
 468                 ptes = (end + PAGE_SIZE - 1) >> PAGE_SHIFT;
 469         tables += roundup(ptes * sizeof(pte_t), PAGE_SIZE);
 470
 471         /*
 472          * RED-PEN putting page tables only on node 0 could
 473          * cause a hotspot and fill up ZONE_DMA. The page tables
 474          * need roughly 0.5KB per GB.
 475          */
 476         start = 0x8000;
 477         table_start = find_e820_area(start, end, tables, PAGE_SIZE);
 478         if (table_start == -1UL)
 479                 panic("Cannot find space for the kernel page tables");
 480
 481         table_start >>= PAGE_SHIFT;
 482         table_end = table_start;
 483         table_top = table_start + (tables >> PAGE_SHIFT);
 484
 485         printk(KERN_DEBUG "kernel direct mapping tables up to %lx @ %lx-%lx\n",
 486                 end, table_start << PAGE_SHIFT, table_top << PAGE_SHIFT);
 487 }
 488
 489 static void __init init_gbpages(void)
 490 {
 491         if (direct_gbpages && cpu_has_gbpages)
 492                 printk(KERN_INFO "Using GB pages for direct mapping\n");
 493         else
 494                 direct_gbpages = 0;
 495 }
 496
 497 static unsigned long __init kernel_physical_mapping_init(unsigned long start,
 498                                                 unsigned long end,
 499                                                 unsigned long page_size_mask)
 500 {
 501
 502         unsigned long next, last_map_addr = end;
 503
 504         start = (unsigned long)__va(start);
 505         end = (unsigned long)__va(end);
 506
 507         for (; start < end; start = next) {
 508                 pgd_t *pgd = pgd_offset_k(start);
 509                 unsigned long pud_phys;
 510                 pud_t *pud;
 511
 512                 next = (start + PGDIR_SIZE) & PGDIR_MASK;
 513                 if (next > end)
 514                         next = end;
 515
 516                 if (pgd_val(*pgd)) {
 517                         last_map_addr = phys_pud_update(pgd, __pa(start),
 518                                                  __pa(end), page_size_mask);
 519                         continue;
 520                 }
 521
 522                 pud = alloc_low_page(&pud_phys);
 523                 last_map_addr = phys_pud_init(pud, __pa(start), __pa(next),
 524                                                  page_size_mask);
 525                 unmap_low_page(pud);
 526
 527                 spin_lock(&init_mm.page_table_lock);
 528                 pgd_populate(&init_mm, pgd, __va(pud_phys));
 529                 spin_unlock(&init_mm.page_table_lock);
 530         }
 531
 532         return last_map_addr;
 533 }
 534
 535 struct map_range {
 536         unsigned long start;
 537         unsigned long end;
 538         unsigned page_size_mask;
 539 };
 540
 541 #define NR_RANGE_MR 5
 542
 543 static int save_mr(struct map_range *mr, int nr_range,
 544                    unsigned long start_pfn, unsigned long end_pfn,
 545                    unsigned long page_size_mask)
 546 {
 547
 548         if (start_pfn < end_pfn) {
 549                 if (nr_range >= NR_RANGE_MR)
 550                         panic("run out of range for init_memory_mapping\n");
 551                 mr[nr_range].start = start_pfn<<PAGE_SHIFT;
 552                 mr[nr_range].end   = end_pfn<<PAGE_SHIFT;
 553                 mr[nr_range].page_size_mask = page_size_mask;
 554                 nr_range++;
 555         }
 556
 557         return nr_range;
 558 }
 559
 560 /*
 561  * Setup the direct mapping of the physical memory at PAGE_OFFSET.
 562  * This runs before bootmem is initialized and gets pages directly from
 563  * the physical memory. To access them they are temporarily mapped.
 564  */
 565 unsigned long __init_refok init_memory_mapping(unsigned long start,
 566                                                unsigned long end)
 567 {
 568         unsigned long last_map_addr = 0;
 569         unsigned long page_size_mask = 0;
 570         unsigned long start_pfn, end_pfn;
 571
 572         struct map_range mr[NR_RANGE_MR];
 573         int nr_range, i;
 574
 575         printk(KERN_INFO "init_memory_mapping\n");
 576
 577         /*
 578          * Find space for the kernel direct mapping tables.
 579          *
 580          * Later we should allocate these tables in the local node of the
 581          * memory mapped. Unfortunately this is done currently before the
 582          * nodes are discovered.
 583          */
 584         if (!after_bootmem)
 585                 init_gbpages();
 586
 587         if (direct_gbpages)
 588                 page_size_mask |= 1 << PG_LEVEL_1G;
 589         if (cpu_has_pse)
 590                 page_size_mask |= 1 << PG_LEVEL_2M;
 591
 592         memset(mr, 0, sizeof(mr));
 593         nr_range = 0;
 594
 595         /* head if not big page alignment ?*/
 596         start_pfn = start >> PAGE_SHIFT;
 597         end_pfn = ((start + (PMD_SIZE - 1)) >> PMD_SHIFT)
 598                         << (PMD_SHIFT - PAGE_SHIFT);
 599         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
 600
 601         /* big page (2M) range*/
 602         start_pfn = ((start + (PMD_SIZE - 1))>>PMD_SHIFT)
 603                          << (PMD_SHIFT - PAGE_SHIFT);
 604         end_pfn = ((start + (PUD_SIZE - 1))>>PUD_SHIFT)
 605                          << (PUD_SHIFT - PAGE_SHIFT);
 606         if (end_pfn > ((end>>PUD_SHIFT)<<(PUD_SHIFT - PAGE_SHIFT)))
 607                 end_pfn = ((end>>PUD_SHIFT)<<(PUD_SHIFT - PAGE_SHIFT));
 608         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 609                         page_size_mask & (1<<PG_LEVEL_2M));
 610
 611         /* big page (1G) range */
 612         start_pfn = end_pfn;
 613         end_pfn = (end>>PUD_SHIFT) << (PUD_SHIFT - PAGE_SHIFT);
 614         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 615                                 page_size_mask &
 616                                  ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G)));
 617
 618         /* tail is not big page (1G) alignment */
 619         start_pfn = end_pfn;
 620         end_pfn = (end>>PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
 621         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 622                         page_size_mask & (1<<PG_LEVEL_2M));
 623
 624         /* tail is not big page (2M) alignment */
 625         start_pfn = end_pfn;
 626         end_pfn = end>>PAGE_SHIFT;
 627         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
 628
 629         /* try to merge same page size and continuous */
 630         for (i = 0; nr_range > 1 && i < nr_range - 1; i++) {
 631                 unsigned long old_start;
 632                 if (mr[i].end != mr[i+1].start ||
 633                     mr[i].page_size_mask != mr[i+1].page_size_mask)
 634                         continue;
 635                 /* move it */
 636                 old_start = mr[i].start;
 637                 memmove(&mr[i], &mr[i+1],
 638                          (nr_range - 1 - i) * sizeof (struct map_range));
 639                 mr[i].start = old_start;
 640                 nr_range--;
 641         }
 642
 643         for (i = 0; i < nr_range; i++)
 644                 printk(KERN_DEBUG " %010lx - %010lx page %s\n",
 645                                 mr[i].start, mr[i].end,
 646                         (mr[i].page_size_mask & (1<<PG_LEVEL_1G))?"1G":(
 647                          (mr[i].page_size_mask & (1<<PG_LEVEL_2M))?"2M":"4k"));
 648
 649         if (!after_bootmem)
 650                 find_early_table_space(end);
 651
 652         for (i = 0; i < nr_range; i++)
 653                 last_map_addr = kernel_physical_mapping_init(
 654                                         mr[i].start, mr[i].end,
 655                                         mr[i].page_size_mask);
 656
 657         if (!after_bootmem)
 658                 mmu_cr4_features = read_cr4();
 659         __flush_tlb_all();
 660
 661         if (!after_bootmem && table_end > table_start)
 662                 reserve_early(table_start << PAGE_SHIFT,
 663                                  table_end << PAGE_SHIFT, "PGTABLE");
 664
 665         printk(KERN_INFO "last_map_addr: %lx end: %lx\n",
 666                          last_map_addr, end);
 667
 668         if (!after_bootmem)
 669                 early_memtest(start, end);
 670
 671         return last_map_addr >> PAGE_SHIFT;
 672 }
 673
 674 #ifndef CONFIG_NUMA
 675 void __init initmem_init(unsigned long start_pfn, unsigned long end_pfn)
 676 {
 677         unsigned long bootmap_size, bootmap;
 678
 679         bootmap_size = bootmem_bootmap_pages(end_pfn)<<PAGE_SHIFT;
 680         bootmap = find_e820_area(0, end_pfn<<PAGE_SHIFT, bootmap_size,
 681                                  PAGE_SIZE);
 682         if (bootmap == -1L)
 683                 panic("Cannot find bootmem map of size %ld\n", bootmap_size);
 684         /* don't touch min_low_pfn */
 685         bootmap_size = init_bootmem_node(NODE_DATA(0), bootmap >> PAGE_SHIFT,
 686                                          0, end_pfn);
 687         e820_register_active_regions(0, start_pfn, end_pfn);
 688         free_bootmem_with_active_regions(0, end_pfn);
 689         early_res_to_bootmem(0, end_pfn<<PAGE_SHIFT);
 690         reserve_bootmem(bootmap, bootmap_size, BOOTMEM_DEFAULT);
 691 }
 692
 693 void __init paging_init(void)
 694 {
 695         unsigned long max_zone_pfns[MAX_NR_ZONES];
 696
 697         memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
 698         max_zone_pfns[ZONE_DMA] = MAX_DMA_PFN;
 699         max_zone_pfns[ZONE_DMA32] = MAX_DMA32_PFN;
 700         max_zone_pfns[ZONE_NORMAL] = max_pfn;
 701
 702         memory_present(0, 0, max_pfn);
 703         sparse_init();
 704         free_area_init_nodes(max_zone_pfns);
 705 }
 706 #endif
 707
 708 /*
 709  * Memory hotplug specific functions
 710  */
 711 #ifdef CONFIG_MEMORY_HOTPLUG
 712 /*
 713  * Memory is added always to NORMAL zone. This means you will never get
 714  * additional DMA/DMA32 memory.
 715  */
 716 int arch_add_memory(int nid, u64 start, u64 size)
 717 {
 718         struct pglist_data *pgdat = NODE_DATA(nid);
 719         struct zone *zone = pgdat->node_zones + ZONE_NORMAL;
 720         unsigned long last_mapped_pfn, start_pfn = start >> PAGE_SHIFT;
 721         unsigned long nr_pages = size >> PAGE_SHIFT;
 722         int ret;
 723
 724         last_mapped_pfn = init_memory_mapping(start, start + size-1);
 725         if (last_mapped_pfn > max_pfn_mapped)
 726                 max_pfn_mapped = last_mapped_pfn;
 727
 728         ret = __add_pages(zone, start_pfn, nr_pages);
 729         WARN_ON(1);
 730
 731         return ret;
 732 }
 733 EXPORT_SYMBOL_GPL(arch_add_memory);
 734
 735 #if !defined(CONFIG_ACPI_NUMA) && defined(CONFIG_NUMA)
 736 int memory_add_physaddr_to_nid(u64 start)
 737 {
 738         return 0;
 739 }
 740 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
 741 #endif
 742
 743 #endif /* CONFIG_MEMORY_HOTPLUG */
 744
 745 /*
 746  * devmem_is_allowed() checks to see if /dev/mem access to a certain address
 747  * is valid. The argument is a physical page number.
 748  *
 749  *
 750  * On x86, access has to be given to the first megabyte of ram because that area
 751  * contains bios code and data regions used by X and dosemu and similar apps.
 752  * Access has to be given to non-kernel-ram areas as well, these contain the PCI
 753  * mmio resources as well as potential bios/acpi data regions.
 754  */
 755 int devmem_is_allowed(unsigned long pagenr)
 756 {
 757         if (pagenr <= 256)
 758                 return 1;
 759         if (!page_is_ram(pagenr))
 760                 return 1;
 761         return 0;
 762 }
 763
 764
 765 static struct kcore_list kcore_mem, kcore_vmalloc, kcore_kernel,
 766                          kcore_modules, kcore_vsyscall;
 767
 768 void __init mem_init(void)
 769 {
 770         long codesize, reservedpages, datasize, initsize;
 771
 772         pci_iommu_alloc();
 773
 774         /* clear_bss() already clear the empty_zero_page */
 775
 776         reservedpages = 0;
 777
 778         /* this will put all low memory onto the freelists */
 779 #ifdef CONFIG_NUMA
 780         totalram_pages = numa_free_all_bootmem();
 781 #else
 782         totalram_pages = free_all_bootmem();
 783 #endif
 784         reservedpages = max_pfn - totalram_pages -
 785                                         absent_pages_in_range(0, max_pfn);
 786         after_bootmem = 1;
 787
 788         codesize =  (unsigned long) &_etext - (unsigned long) &_text;
 789         datasize =  (unsigned long) &_edata - (unsigned long) &_etext;
 790         initsize =  (unsigned long) &__init_end - (unsigned long) &__init_begin;
 791
 792         /* Register memory areas for /proc/kcore */
 793         kclist_add(&kcore_mem, __va(0), max_low_pfn << PAGE_SHIFT);
 794         kclist_add(&kcore_vmalloc, (void *)VMALLOC_START,
 795                    VMALLOC_END-VMALLOC_START);
 796         kclist_add(&kcore_kernel, &_stext, _end - _stext);
 797         kclist_add(&kcore_modules, (void *)MODULES_VADDR, MODULES_LEN);
 798         kclist_add(&kcore_vsyscall, (void *)VSYSCALL_START,
 799                                  VSYSCALL_END - VSYSCALL_START);
 800
 801         printk(KERN_INFO "Memory: %luk/%luk available (%ldk kernel code, "
 802                                 "%ldk reserved, %ldk data, %ldk init)\n",
 803                 (unsigned long) nr_free_pages() << (PAGE_SHIFT-10),
 804                 max_pfn << (PAGE_SHIFT-10),
 805                 codesize >> 10,
 806                 reservedpages << (PAGE_SHIFT-10),
 807                 datasize >> 10,
 808                 initsize >> 10);
 809
 810         cpa_init();
 811 }
 812
 813 void free_init_pages(char *what, unsigned long begin, unsigned long end)
 814 {
 815         unsigned long addr = begin;
 816
 817         if (addr >= end)
 818                 return;
 819
 820         /*
 821          * If debugging page accesses then do not free this memory but
 822          * mark them not present - any buggy init-section access will
 823          * create a kernel page fault:
 824          */
 825 #ifdef CONFIG_DEBUG_PAGEALLOC
 826         printk(KERN_INFO "debug: unmapping init memory %08lx..%08lx\n",
 827                 begin, PAGE_ALIGN(end));
 828         set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
 829 #else
 830         printk(KERN_INFO "Freeing %s: %luk freed\n", what, (end - begin) >> 10);
 831
 832         for (; addr < end; addr += PAGE_SIZE) {
 833                 ClearPageReserved(virt_to_page(addr));
 834                 init_page_count(virt_to_page(addr));
 835                 memset((void *)(addr & ~(PAGE_SIZE-1)),
 836                         POISON_FREE_INITMEM, PAGE_SIZE);
 837                 free_page(addr);
 838                 totalram_pages++;
 839         }
 840 #endif
 841 }
 842
 843 void free_initmem(void)
 844 {
 845         free_init_pages("unused kernel memory",
 846                         (unsigned long)(&__init_begin),
 847                         (unsigned long)(&__init_end));
 848 }
 849
 850 #ifdef CONFIG_DEBUG_RODATA
 851 const int rodata_test_data = 0xC3;
 852 EXPORT_SYMBOL_GPL(rodata_test_data);
 853
 854 void mark_rodata_ro(void)
 855 {
 856         unsigned long start = PFN_ALIGN(_stext), end = PFN_ALIGN(__end_rodata);
 857         unsigned long rodata_start =
 858                 ((unsigned long)__start_rodata + PAGE_SIZE - 1) & PAGE_MASK;
 859
 860 #ifdef CONFIG_DYNAMIC_FTRACE
 861         /* Dynamic tracing modifies the kernel text section */
 862         start = rodata_start;
 863 #endif
 864
 865         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
 866                (end - start) >> 10);
 867         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
 868
 869         /*
 870          * The rodata section (but not the kernel text!) should also be
 871          * not-executable.
 872          */
 873         set_memory_nx(rodata_start, (end - rodata_start) >> PAGE_SHIFT);
 874
 875         rodata_test();
 876
 877 #ifdef CONFIG_CPA_DEBUG
 878         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
 879         set_memory_rw(start, (end-start) >> PAGE_SHIFT);
 880
 881         printk(KERN_INFO "Testing CPA: again\n");
 882         set_memory_ro(start, (end-start) >> PAGE_SHIFT);
 883 #endif
 884 }
 885
 886 #endif
 887
 888 #ifdef CONFIG_BLK_DEV_INITRD
 889 void free_initrd_mem(unsigned long start, unsigned long end)
 890 {
 891         free_init_pages("initrd memory", start, end);
 892 }
 893 #endif
 894
 895 int __init reserve_bootmem_generic(unsigned long phys, unsigned long len,
 896                                    int flags)
 897 {
 898 #ifdef CONFIG_NUMA
 899         int nid, next_nid;
 900         int ret;
 901 #endif
 902         unsigned long pfn = phys >> PAGE_SHIFT;
 903
 904         if (pfn >= max_pfn) {
 905                 /*
 906                  * This can happen with kdump kernels when accessing
 907                  * firmware tables:
 908                  */
 909                 if (pfn < max_pfn_mapped)
 910                         return -EFAULT;
 911
 912                 printk(KERN_ERR "reserve_bootmem: illegal reserve %lx %lu\n",
 913                                 phys, len);
 914                 return -EFAULT;
 915         }
 916
 917         /* Should check here against the e820 map to avoid double free */
 918 #ifdef CONFIG_NUMA
 919         nid = phys_to_nid(phys);
 920         next_nid = phys_to_nid(phys + len - 1);
 921         if (nid == next_nid)
 922                 ret = reserve_bootmem_node(NODE_DATA(nid), phys, len, flags);
 923         else
 924                 ret = reserve_bootmem(phys, len, flags);
 925
 926         if (ret != 0)
 927                 return ret;
 928
 929 #else
 930         reserve_bootmem(phys, len, BOOTMEM_DEFAULT);
 931 #endif
 932
 933         if (phys+len <= MAX_DMA_PFN*PAGE_SIZE) {
 934                 dma_reserve += len / PAGE_SIZE;
 935                 set_dma_reserve(dma_reserve);
 936         }
 937
 938         return 0;
 939 }
 940
 941 int kern_addr_valid(unsigned long addr)
 942 {
 943         unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
 944         pgd_t *pgd;
 945         pud_t *pud;
 946         pmd_t *pmd;
 947         pte_t *pte;
 948
 949         if (above != 0 && above != -1UL)
 950                 return 0;
 951
 952         pgd = pgd_offset_k(addr);
 953         if (pgd_none(*pgd))
 954                 return 0;
 955
 956         pud = pud_offset(pgd, addr);
 957         if (pud_none(*pud))
 958                 return 0;
 959
 960         pmd = pmd_offset(pud, addr);
 961         if (pmd_none(*pmd))
 962                 return 0;
 963
 964         if (pmd_large(*pmd))
 965                 return pfn_valid(pmd_pfn(*pmd));
 966
 967         pte = pte_offset_kernel(pmd, addr);
 968         if (pte_none(*pte))
 969                 return 0;
 970
 971         return pfn_valid(pte_pfn(*pte));
 972 }
 973
 974 /*
 975  * A pseudo VMA to allow ptrace access for the vsyscall page.  This only
 976  * covers the 64bit vsyscall page now. 32bit has a real VMA now and does
 977  * not need special handling anymore:
 978  */
 979 static struct vm_area_struct gate_vma = {
 980         .vm_start       = VSYSCALL_START,
 981         .vm_end         = VSYSCALL_START + (VSYSCALL_MAPPED_PAGES * PAGE_SIZE),
 982         .vm_page_prot   = PAGE_READONLY_EXEC,
 983         .vm_flags       = VM_READ | VM_EXEC
 984 };
 985
 986 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
 987 {
 988 #ifdef CONFIG_IA32_EMULATION
 989         if (test_tsk_thread_flag(tsk, TIF_IA32))
 990                 return NULL;
 991 #endif
 992         return &gate_vma;
 993 }
 994
 995 int in_gate_area(struct task_struct *task, unsigned long addr)
 996 {
 997         struct vm_area_struct *vma = get_gate_vma(task);
 998
 999         if (!vma)
1000                 return 0;
1001
1002         return (addr >= vma->vm_start) && (addr < vma->vm_end);
1003 }
1004
1005 /*
1006  * Use this when you have no reliable task/vma, typically from interrupt
1007  * context. It is less reliable than using the task's vma and may give
1008  * false positives:
1009  */
1010 int in_gate_area_no_task(unsigned long addr)
1011 {
1012         return (addr >= VSYSCALL_START) && (addr < VSYSCALL_END);
1013 }
1014
1015 const char *arch_vma_name(struct vm_area_struct *vma)
1016 {
1017         if (vma->vm_mm && vma->vm_start == (long)vma->vm_mm->context.vdso)
1018                 return "[vdso]";
1019         if (vma == &gate_vma)
1020                 return "[vsyscall]";
1021         return NULL;
1022 }
1023
1024 #ifdef CONFIG_SPARSEMEM_VMEMMAP
1025 /*
1026  * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
1027  */
1028 static long __meminitdata addr_start, addr_end;
1029 static void __meminitdata *p_start, *p_end;
1030 static int __meminitdata node_start;
1031
1032 int __meminit
1033 vmemmap_populate(struct page *start_page, unsigned long size, int node)
1034 {
1035         unsigned long addr = (unsigned long)start_page;
1036         unsigned long end = (unsigned long)(start_page + size);
1037         unsigned long next;
1038         pgd_t *pgd;
1039         pud_t *pud;
1040         pmd_t *pmd;
1041
1042         for (; addr < end; addr = next) {
1043                 void *p = NULL;
1044
1045                 pgd = vmemmap_pgd_populate(addr, node);
1046                 if (!pgd)
1047                         return -ENOMEM;
1048
1049                 pud = vmemmap_pud_populate(pgd, addr, node);
1050                 if (!pud)
1051                         return -ENOMEM;
1052
1053                 if (!cpu_has_pse) {
1054                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1055                         pmd = vmemmap_pmd_populate(pud, addr, node);
1056
1057                         if (!pmd)
1058                                 return -ENOMEM;
1059
1060                         p = vmemmap_pte_populate(pmd, addr, node);
1061
1062                         if (!p)
1063                                 return -ENOMEM;
1064
1065                         addr_end = addr + PAGE_SIZE;
1066                         p_end = p + PAGE_SIZE;
1067                 } else {
1068                         next = pmd_addr_end(addr, end);
1069
1070                         pmd = pmd_offset(pud, addr);
1071                         if (pmd_none(*pmd)) {
1072                                 pte_t entry;
1073
1074                                 p = vmemmap_alloc_block(PMD_SIZE, node);
1075                                 if (!p)
1076                                         return -ENOMEM;
1077
1078                                 entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
1079                                                 PAGE_KERNEL_LARGE);
1080                                 set_pmd(pmd, __pmd(pte_val(entry)));
1081
1082                                 /* check to see if we have contiguous blocks */
1083                                 if (p_end != p || node_start != node) {
1084                                         if (p_start)
1085                                                 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1086                                                        addr_start, addr_end-1, p_start, p_end-1, node_start);
1087                                         addr_start = addr;
1088                                         node_start = node;
1089                                         p_start = p;
1090                                 }
1091
1092                                 addr_end = addr + PMD_SIZE;
1093                                 p_end = p + PMD_SIZE;
1094                         } else
1095                                 vmemmap_verify((pte_t *)pmd, node, addr, next);
1096                 }
1097
1098         }
1099         return 0;
1100 }
1101
1102 void __meminit vmemmap_populate_print_last(void)
1103 {
1104         if (p_start) {
1105                 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1106                         addr_start, addr_end-1, p_start, p_end-1, node_start);
1107                 p_start = NULL;
1108                 p_end = NULL;
1109                 node_start = 0;
1110         }
1111 }
1112 #endif