SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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_pfn_mapped;
  57
  58 static unsigned long dma_reserve __initdata;
  59
  60 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
  61
  62 int direct_gbpages __meminitdata
  63 #ifdef CONFIG_DIRECT_GBPAGES
  64                                 = 1
  65 #endif
  66 ;
  67
  68 static int __init parse_direct_gbpages_off(char *arg)
  69 {
  70         direct_gbpages = 0;
  71         return 0;
  72 }
  73 early_param("nogbpages", parse_direct_gbpages_off);
  74
  75 static int __init parse_direct_gbpages_on(char *arg)
  76 {
  77         direct_gbpages = 1;
  78         return 0;
  79 }
  80 early_param("gbpages", parse_direct_gbpages_on);
  81
  82 /*
  83  * NOTE: pagetable_init alloc all the fixmap pagetables contiguous on the
  84  * physical space so we can cache the place of the first one and move
  85  * around without checking the pgd every time.
  86  */
  87
  88 void show_mem(void)
  89 {
  90         long i, total = 0, reserved = 0;
  91         long shared = 0, cached = 0;
  92         struct page *page;
  93         pg_data_t *pgdat;
  94
  95         printk(KERN_INFO "Mem-info:\n");
  96         show_free_areas();
  97         for_each_online_pgdat(pgdat) {
  98                 for (i = 0; i < pgdat->node_spanned_pages; ++i) {
  99                         /*
 100                          * This loop can take a while with 256 GB and
 101                          * 4k pages so defer the NMI watchdog:
 102                          */
 103                         if (unlikely(i % MAX_ORDER_NR_PAGES == 0))
 104                                 touch_nmi_watchdog();
 105
 106                         if (!pfn_valid(pgdat->node_start_pfn + i))
 107                                 continue;
 108
 109                         page = pfn_to_page(pgdat->node_start_pfn + i);
 110                         total++;
 111                         if (PageReserved(page))
 112                                 reserved++;
 113                         else if (PageSwapCache(page))
 114                                 cached++;
 115                         else if (page_count(page))
 116                                 shared += page_count(page) - 1;
 117                 }
 118         }
 119         printk(KERN_INFO "%lu pages of RAM\n",          total);
 120         printk(KERN_INFO "%lu reserved pages\n",        reserved);
 121         printk(KERN_INFO "%lu pages shared\n",          shared);
 122         printk(KERN_INFO "%lu pages swap cached\n",     cached);
 123 }
 124
 125 int after_bootmem;
 126
 127 static __init void *spp_getpage(void)
 128 {
 129         void *ptr;
 130
 131         if (after_bootmem)
 132                 ptr = (void *) get_zeroed_page(GFP_ATOMIC);
 133         else
 134                 ptr = alloc_bootmem_pages(PAGE_SIZE);
 135
 136         if (!ptr || ((unsigned long)ptr & ~PAGE_MASK)) {
 137                 panic("set_pte_phys: cannot allocate page data %s\n",
 138                         after_bootmem ? "after bootmem" : "");
 139         }
 140
 141         pr_debug("spp_getpage %p\n", ptr);
 142
 143         return ptr;
 144 }
 145
 146 void
 147 set_pte_vaddr_pud(pud_t *pud_page, unsigned long vaddr, pte_t new_pte)
 148 {
 149         pud_t *pud;
 150         pmd_t *pmd;
 151         pte_t *pte;
 152
 153         pud = pud_page + pud_index(vaddr);
 154         if (pud_none(*pud)) {
 155                 pmd = (pmd_t *) spp_getpage();
 156                 pud_populate(&init_mm, pud, pmd);
 157                 if (pmd != pmd_offset(pud, 0)) {
 158                         printk(KERN_ERR "PAGETABLE BUG #01! %p <-> %p\n",
 159                                 pmd, pmd_offset(pud, 0));
 160                         return;
 161                 }
 162         }
 163         pmd = pmd_offset(pud, vaddr);
 164         if (pmd_none(*pmd)) {
 165                 pte = (pte_t *) spp_getpage();
 166                 pmd_populate_kernel(&init_mm, pmd, pte);
 167                 if (pte != pte_offset_kernel(pmd, 0)) {
 168                         printk(KERN_ERR "PAGETABLE BUG #02!\n");
 169                         return;
 170                 }
 171         }
 172
 173         pte = pte_offset_kernel(pmd, vaddr);
 174         if (!pte_none(*pte) && pte_val(new_pte) &&
 175             pte_val(*pte) != (pte_val(new_pte) & __supported_pte_mask))
 176                 pte_ERROR(*pte);
 177         set_pte(pte, new_pte);
 178
 179         /*
 180          * It's enough to flush this one mapping.
 181          * (PGE mappings get flushed as well)
 182          */
 183         __flush_tlb_one(vaddr);
 184 }
 185
 186 void
 187 set_pte_vaddr(unsigned long vaddr, pte_t pteval)
 188 {
 189         pgd_t *pgd;
 190         pud_t *pud_page;
 191
 192         pr_debug("set_pte_vaddr %lx to %lx\n", vaddr, native_pte_val(pteval));
 193
 194         pgd = pgd_offset_k(vaddr);
 195         if (pgd_none(*pgd)) {
 196                 printk(KERN_ERR
 197                         "PGD FIXMAP MISSING, it should be setup in head.S!\n");
 198                 return;
 199         }
 200         pud_page = (pud_t*)pgd_page_vaddr(*pgd);
 201         set_pte_vaddr_pud(pud_page, vaddr, pteval);
 202 }
 203
 204 /*
 205  * The head.S code sets up the kernel high mapping:
 206  *
 207  *   from __START_KERNEL_map to __START_KERNEL_map + size (== _end-_text)
 208  *
 209  * phys_addr holds the negative offset to the kernel, which is added
 210  * to the compile time generated pmds. This results in invalid pmds up
 211  * to the point where we hit the physaddr 0 mapping.
 212  *
 213  * We limit the mappings to the region from _text to _end.  _end is
 214  * rounded up to the 2MB boundary. This catches the invalid pmds as
 215  * well, as they are located before _text:
 216  */
 217 void __init cleanup_highmap(void)
 218 {
 219         unsigned long vaddr = __START_KERNEL_map;
 220         unsigned long end = round_up((unsigned long)_end, PMD_SIZE) - 1;
 221         pmd_t *pmd = level2_kernel_pgt;
 222         pmd_t *last_pmd = pmd + PTRS_PER_PMD;
 223
 224         for (; pmd < last_pmd; pmd++, vaddr += PMD_SIZE) {
 225                 if (pmd_none(*pmd))
 226                         continue;
 227                 if (vaddr < (unsigned long) _text || vaddr > end)
 228                         set_pmd(pmd, __pmd(0));
 229         }
 230 }
 231
 232 static unsigned long __initdata table_start;
 233 static unsigned long __meminitdata table_end;
 234 static unsigned long __meminitdata table_top;
 235
 236 static __meminit void *alloc_low_page(unsigned long *phys)
 237 {
 238         unsigned long pfn = table_end++;
 239         void *adr;
 240
 241         if (after_bootmem) {
 242                 adr = (void *)get_zeroed_page(GFP_ATOMIC);
 243                 *phys = __pa(adr);
 244
 245                 return adr;
 246         }
 247
 248         if (pfn >= table_top)
 249                 panic("alloc_low_page: ran out of memory");
 250
 251         adr = early_ioremap(pfn * PAGE_SIZE, PAGE_SIZE);
 252         memset(adr, 0, PAGE_SIZE);
 253         *phys  = pfn * PAGE_SIZE;
 254         return adr;
 255 }
 256
 257 static __meminit void unmap_low_page(void *adr)
 258 {
 259         if (after_bootmem)
 260                 return;
 261
 262         early_iounmap(adr, PAGE_SIZE);
 263 }
 264
 265 static void __meminit
 266 phys_pte_init(pte_t *pte_page, unsigned long addr, unsigned long end)
 267 {
 268         unsigned pages = 0;
 269         int i;
 270         pte_t *pte = pte_page + pte_index(addr);
 271
 272         for(i = pte_index(addr); i < PTRS_PER_PTE; i++, addr += PAGE_SIZE, pte++) {
 273
 274                 if (addr >= end) {
 275                         if (!after_bootmem) {
 276                                 for(; i < PTRS_PER_PTE; i++, pte++)
 277                                         set_pte(pte, __pte(0));
 278                         }
 279                         break;
 280                 }
 281
 282                 if (pte_val(*pte))
 283                         continue;
 284
 285                 if (0)
 286                         printk("   pte=%p addr=%lx pte=%016lx\n",
 287                                pte, addr, pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL).pte);
 288                 set_pte(pte, pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL));
 289                 pages++;
 290         }
 291         update_page_count(PG_LEVEL_4K, pages);
 292 }
 293
 294 static void __meminit
 295 phys_pte_update(pmd_t *pmd, unsigned long address, unsigned long end)
 296 {
 297         pte_t *pte = (pte_t *)pmd_page_vaddr(*pmd);
 298
 299         phys_pte_init(pte, address, end);
 300 }
 301
 302 static unsigned long __meminit
 303 phys_pmd_init(pmd_t *pmd_page, unsigned long address, unsigned long end)
 304 {
 305         unsigned long pages = 0;
 306
 307         int i = pmd_index(address);
 308
 309         for (; i < PTRS_PER_PMD; i++, address += PMD_SIZE) {
 310                 unsigned long pte_phys;
 311                 pmd_t *pmd = pmd_page + pmd_index(address);
 312                 pte_t *pte;
 313
 314                 if (address >= end) {
 315                         if (!after_bootmem) {
 316                                 for (; i < PTRS_PER_PMD; i++, pmd++)
 317                                         set_pmd(pmd, __pmd(0));
 318                         }
 319                         break;
 320                 }
 321
 322                 if (pmd_val(*pmd)) {
 323                         phys_pte_update(pmd, address, end);
 324                         continue;
 325                 }
 326
 327                 if (cpu_has_pse) {
 328                         pages++;
 329                         set_pte((pte_t *)pmd,
 330                                 pfn_pte(address >> PAGE_SHIFT, PAGE_KERNEL_LARGE));
 331                         continue;
 332                 }
 333
 334                 pte = alloc_low_page(&pte_phys);
 335                 phys_pte_init(pte, address, end);
 336                 unmap_low_page(pte);
 337
 338                 pmd_populate_kernel(&init_mm, pmd, __va(pte_phys));
 339         }
 340         update_page_count(PG_LEVEL_2M, pages);
 341         return address;
 342 }
 343
 344 static unsigned long __meminit
 345 phys_pmd_update(pud_t *pud, unsigned long address, unsigned long end)
 346 {
 347         pmd_t *pmd = pmd_offset(pud, 0);
 348         unsigned long last_map_addr;
 349
 350         spin_lock(&init_mm.page_table_lock);
 351         last_map_addr = phys_pmd_init(pmd, address, end);
 352         spin_unlock(&init_mm.page_table_lock);
 353         __flush_tlb_all();
 354         return last_map_addr;
 355 }
 356
 357 static unsigned long __meminit
 358 phys_pud_init(pud_t *pud_page, unsigned long addr, unsigned long end)
 359 {
 360         unsigned long pages = 0;
 361         unsigned long last_map_addr = end;
 362         int i = pud_index(addr);
 363
 364         for (; i < PTRS_PER_PUD; i++, addr = (addr & PUD_MASK) + PUD_SIZE) {
 365                 unsigned long pmd_phys;
 366                 pud_t *pud = pud_page + pud_index(addr);
 367                 pmd_t *pmd;
 368
 369                 if (addr >= end)
 370                         break;
 371
 372                 if (!after_bootmem &&
 373                                 !e820_any_mapped(addr, addr+PUD_SIZE, 0)) {
 374                         set_pud(pud, __pud(0));
 375                         continue;
 376                 }
 377
 378                 if (pud_val(*pud)) {
 379                         if (!pud_large(*pud))
 380                                 last_map_addr = phys_pmd_update(pud, addr, end);
 381                         continue;
 382                 }
 383
 384                 if (direct_gbpages) {
 385                         pages++;
 386                         set_pte((pte_t *)pud,
 387                                 pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL_LARGE));
 388                         last_map_addr = (addr & PUD_MASK) + PUD_SIZE;
 389                         continue;
 390                 }
 391
 392                 pmd = alloc_low_page(&pmd_phys);
 393
 394                 spin_lock(&init_mm.page_table_lock);
 395                 last_map_addr = phys_pmd_init(pmd, addr, end);
 396                 unmap_low_page(pmd);
 397                 pud_populate(&init_mm, pud, __va(pmd_phys));
 398                 spin_unlock(&init_mm.page_table_lock);
 399
 400         }
 401         __flush_tlb_all();
 402         update_page_count(PG_LEVEL_1G, pages);
 403
 404         return last_map_addr;
 405 }
 406
 407 static unsigned long __meminit
 408 phys_pud_update(pgd_t *pgd, unsigned long addr, unsigned long end)
 409 {
 410         pud_t *pud;
 411
 412         pud = (pud_t *)pgd_page_vaddr(*pgd);
 413
 414         return phys_pud_init(pud, addr, end);
 415 }
 416
 417 static void __init find_early_table_space(unsigned long end)
 418 {
 419         unsigned long puds, tables, start;
 420
 421         puds = (end + PUD_SIZE - 1) >> PUD_SHIFT;
 422         tables = round_up(puds * sizeof(pud_t), PAGE_SIZE);
 423         if (!direct_gbpages) {
 424                 unsigned long pmds = (end + PMD_SIZE - 1) >> PMD_SHIFT;
 425                 tables += round_up(pmds * sizeof(pmd_t), PAGE_SIZE);
 426         }
 427         if (!cpu_has_pse) {
 428                 unsigned long ptes = (end + PAGE_SIZE - 1) >> PAGE_SHIFT;
 429                 tables += round_up(ptes * sizeof(pte_t), PAGE_SIZE);
 430         }
 431
 432         /*
 433          * RED-PEN putting page tables only on node 0 could
 434          * cause a hotspot and fill up ZONE_DMA. The page tables
 435          * need roughly 0.5KB per GB.
 436          */
 437         start = 0x8000;
 438         table_start = find_e820_area(start, end, tables, PAGE_SIZE);
 439         if (table_start == -1UL)
 440                 panic("Cannot find space for the kernel page tables");
 441
 442         table_start >>= PAGE_SHIFT;
 443         table_end = table_start;
 444         table_top = table_start + (tables >> PAGE_SHIFT);
 445
 446         printk(KERN_DEBUG "kernel direct mapping tables up to %lx @ %lx-%lx\n",
 447                 end, table_start << PAGE_SHIFT, table_top << PAGE_SHIFT);
 448 }
 449
 450 static void __init init_gbpages(void)
 451 {
 452         if (direct_gbpages && cpu_has_gbpages)
 453                 printk(KERN_INFO "Using GB pages for direct mapping\n");
 454         else
 455                 direct_gbpages = 0;
 456 }
 457
 458 #ifdef CONFIG_MEMTEST
 459
 460 static void __init memtest(unsigned long start_phys, unsigned long size,
 461                                  unsigned pattern)
 462 {
 463         unsigned long i;
 464         unsigned long *start;
 465         unsigned long start_bad;
 466         unsigned long last_bad;
 467         unsigned long val;
 468         unsigned long start_phys_aligned;
 469         unsigned long count;
 470         unsigned long incr;
 471
 472         switch (pattern) {
 473         case 0:
 474                 val = 0UL;
 475                 break;
 476         case 1:
 477                 val = -1UL;
 478                 break;
 479         case 2:
 480                 val = 0x5555555555555555UL;
 481                 break;
 482         case 3:
 483                 val = 0xaaaaaaaaaaaaaaaaUL;
 484                 break;
 485         default:
 486                 return;
 487         }
 488
 489         incr = sizeof(unsigned long);
 490         start_phys_aligned = ALIGN(start_phys, incr);
 491         count = (size - (start_phys_aligned - start_phys))/incr;
 492         start = __va(start_phys_aligned);
 493         start_bad = 0;
 494         last_bad = 0;
 495
 496         for (i = 0; i < count; i++)
 497                 start[i] = val;
 498         for (i = 0; i < count; i++, start++, start_phys_aligned += incr) {
 499                 if (*start != val) {
 500                         if (start_phys_aligned == last_bad + incr) {
 501                                 last_bad += incr;
 502                         } else {
 503                                 if (start_bad) {
 504                                         printk(KERN_CONT "\n  %016lx bad mem addr %016lx - %016lx reserved",
 505                                                 val, start_bad, last_bad + incr);
 506                                         reserve_early(start_bad, last_bad - start_bad, "BAD RAM");
 507                                 }
 508                                 start_bad = last_bad = start_phys_aligned;
 509                         }
 510                 }
 511         }
 512         if (start_bad) {
 513                 printk(KERN_CONT "\n  %016lx bad mem addr %016lx - %016lx reserved",
 514                         val, start_bad, last_bad + incr);
 515                 reserve_early(start_bad, last_bad - start_bad, "BAD RAM");
 516         }
 517
 518 }
 519
 520 /* default is disabled */
 521 static int memtest_pattern __initdata;
 522
 523 static int __init parse_memtest(char *arg)
 524 {
 525         if (arg)
 526                 memtest_pattern = simple_strtoul(arg, NULL, 0);
 527         return 0;
 528 }
 529
 530 early_param("memtest", parse_memtest);
 531
 532 static void __init early_memtest(unsigned long start, unsigned long end)
 533 {
 534         u64 t_start, t_size;
 535         unsigned pattern;
 536
 537         if (!memtest_pattern)
 538                 return;
 539
 540         printk(KERN_INFO "early_memtest: pattern num %d", memtest_pattern);
 541         for (pattern = 0; pattern < memtest_pattern; pattern++) {
 542                 t_start = start;
 543                 t_size = 0;
 544                 while (t_start < end) {
 545                         t_start = find_e820_area_size(t_start, &t_size, 1);
 546
 547                         /* done ? */
 548                         if (t_start >= end)
 549                                 break;
 550                         if (t_start + t_size > end)
 551                                 t_size = end - t_start;
 552
 553                         printk(KERN_CONT "\n  %016llx - %016llx pattern %d",
 554                                 (unsigned long long)t_start,
 555                                 (unsigned long long)t_start + t_size, pattern);
 556
 557                         memtest(t_start, t_size, pattern);
 558
 559                         t_start += t_size;
 560                 }
 561         }
 562         printk(KERN_CONT "\n");
 563 }
 564 #else
 565 static void __init early_memtest(unsigned long start, unsigned long end)
 566 {
 567 }
 568 #endif
 569
 570 /*
 571  * Setup the direct mapping of the physical memory at PAGE_OFFSET.
 572  * This runs before bootmem is initialized and gets pages directly from
 573  * the physical memory. To access them they are temporarily mapped.
 574  */
 575 unsigned long __init_refok init_memory_mapping(unsigned long start, unsigned long end)
 576 {
 577         unsigned long next, last_map_addr = end;
 578         unsigned long start_phys = start, end_phys = end;
 579
 580         printk(KERN_INFO "init_memory_mapping\n");
 581
 582         /*
 583          * Find space for the kernel direct mapping tables.
 584          *
 585          * Later we should allocate these tables in the local node of the
 586          * memory mapped. Unfortunately this is done currently before the
 587          * nodes are discovered.
 588          */
 589         if (!after_bootmem) {
 590                 init_gbpages();
 591                 find_early_table_space(end);
 592         }
 593
 594         start = (unsigned long)__va(start);
 595         end = (unsigned long)__va(end);
 596
 597         for (; start < end; start = next) {
 598                 pgd_t *pgd = pgd_offset_k(start);
 599                 unsigned long pud_phys;
 600                 pud_t *pud;
 601
 602                 next = start + PGDIR_SIZE;
 603                 if (next > end)
 604                         next = end;
 605
 606                 if (pgd_val(*pgd)) {
 607                         last_map_addr = phys_pud_update(pgd, __pa(start), __pa(end));
 608                         continue;
 609                 }
 610
 611                 if (after_bootmem)
 612                         pud = pud_offset(pgd, start & PGDIR_MASK);
 613                 else
 614                         pud = alloc_low_page(&pud_phys);
 615
 616                 last_map_addr = phys_pud_init(pud, __pa(start), __pa(next));
 617                 unmap_low_page(pud);
 618                 if (!after_bootmem)
 619                         pgd_populate(&init_mm, pgd_offset_k(start),
 620                                      __va(pud_phys));
 621         }
 622
 623         if (!after_bootmem)
 624                 mmu_cr4_features = read_cr4();
 625         __flush_tlb_all();
 626
 627         if (!after_bootmem)
 628                 reserve_early(table_start << PAGE_SHIFT,
 629                                  table_end << PAGE_SHIFT, "PGTABLE");
 630
 631         if (!after_bootmem)
 632                 early_memtest(start_phys, end_phys);
 633
 634         return last_map_addr >> PAGE_SHIFT;
 635 }
 636
 637 #ifndef CONFIG_NUMA
 638 void __init initmem_init(unsigned long start_pfn, unsigned long end_pfn)
 639 {
 640         unsigned long bootmap_size, bootmap;
 641
 642         bootmap_size = bootmem_bootmap_pages(end_pfn)<<PAGE_SHIFT;
 643         bootmap = find_e820_area(0, end_pfn<<PAGE_SHIFT, bootmap_size,
 644                                  PAGE_SIZE);
 645         if (bootmap == -1L)
 646                 panic("Cannot find bootmem map of size %ld\n", bootmap_size);
 647         /* don't touch min_low_pfn */
 648         bootmap_size = init_bootmem_node(NODE_DATA(0), bootmap >> PAGE_SHIFT,
 649                                          0, end_pfn);
 650         e820_register_active_regions(0, start_pfn, end_pfn);
 651         free_bootmem_with_active_regions(0, end_pfn);
 652         early_res_to_bootmem(0, end_pfn<<PAGE_SHIFT);
 653         reserve_bootmem(bootmap, bootmap_size, BOOTMEM_DEFAULT);
 654 }
 655
 656 void __init paging_init(void)
 657 {
 658         unsigned long max_zone_pfns[MAX_NR_ZONES];
 659
 660         memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
 661         max_zone_pfns[ZONE_DMA] = MAX_DMA_PFN;
 662         max_zone_pfns[ZONE_DMA32] = MAX_DMA32_PFN;
 663         max_zone_pfns[ZONE_NORMAL] = max_pfn;
 664
 665         memory_present(0, 0, max_pfn);
 666         sparse_init();
 667         free_area_init_nodes(max_zone_pfns);
 668 }
 669 #endif
 670
 671 /*
 672  * Memory hotplug specific functions
 673  */
 674 #ifdef CONFIG_MEMORY_HOTPLUG
 675 /*
 676  * Memory is added always to NORMAL zone. This means you will never get
 677  * additional DMA/DMA32 memory.
 678  */
 679 int arch_add_memory(int nid, u64 start, u64 size)
 680 {
 681         struct pglist_data *pgdat = NODE_DATA(nid);
 682         struct zone *zone = pgdat->node_zones + ZONE_NORMAL;
 683         unsigned long last_mapped_pfn, start_pfn = start >> PAGE_SHIFT;
 684         unsigned long nr_pages = size >> PAGE_SHIFT;
 685         int ret;
 686
 687         last_mapped_pfn = init_memory_mapping(start, start + size-1);
 688         if (last_mapped_pfn > max_pfn_mapped)
 689                 max_pfn_mapped = last_mapped_pfn;
 690
 691         ret = __add_pages(zone, start_pfn, nr_pages);
 692         WARN_ON(1);
 693
 694         return ret;
 695 }
 696 EXPORT_SYMBOL_GPL(arch_add_memory);
 697
 698 #if !defined(CONFIG_ACPI_NUMA) && defined(CONFIG_NUMA)
 699 int memory_add_physaddr_to_nid(u64 start)
 700 {
 701         return 0;
 702 }
 703 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
 704 #endif
 705
 706 #endif /* CONFIG_MEMORY_HOTPLUG */
 707
 708 /*
 709  * devmem_is_allowed() checks to see if /dev/mem access to a certain address
 710  * is valid. The argument is a physical page number.
 711  *
 712  *
 713  * On x86, access has to be given to the first megabyte of ram because that area
 714  * contains bios code and data regions used by X and dosemu and similar apps.
 715  * Access has to be given to non-kernel-ram areas as well, these contain the PCI
 716  * mmio resources as well as potential bios/acpi data regions.
 717  */
 718 int devmem_is_allowed(unsigned long pagenr)
 719 {
 720         if (pagenr <= 256)
 721                 return 1;
 722         if (!page_is_ram(pagenr))
 723                 return 1;
 724         return 0;
 725 }
 726
 727
 728 static struct kcore_list kcore_mem, kcore_vmalloc, kcore_kernel,
 729                          kcore_modules, kcore_vsyscall;
 730
 731 void __init mem_init(void)
 732 {
 733         long codesize, reservedpages, datasize, initsize;
 734
 735         pci_iommu_alloc();
 736
 737         /* clear_bss() already clear the empty_zero_page */
 738
 739         reservedpages = 0;
 740
 741         /* this will put all low memory onto the freelists */
 742 #ifdef CONFIG_NUMA
 743         totalram_pages = numa_free_all_bootmem();
 744 #else
 745         totalram_pages = free_all_bootmem();
 746 #endif
 747         reservedpages = max_pfn - totalram_pages -
 748                                         absent_pages_in_range(0, max_pfn);
 749         after_bootmem = 1;
 750
 751         codesize =  (unsigned long) &_etext - (unsigned long) &_text;
 752         datasize =  (unsigned long) &_edata - (unsigned long) &_etext;
 753         initsize =  (unsigned long) &__init_end - (unsigned long) &__init_begin;
 754
 755         /* Register memory areas for /proc/kcore */
 756         kclist_add(&kcore_mem, __va(0), max_low_pfn << PAGE_SHIFT);
 757         kclist_add(&kcore_vmalloc, (void *)VMALLOC_START,
 758                    VMALLOC_END-VMALLOC_START);
 759         kclist_add(&kcore_kernel, &_stext, _end - _stext);
 760         kclist_add(&kcore_modules, (void *)MODULES_VADDR, MODULES_LEN);
 761         kclist_add(&kcore_vsyscall, (void *)VSYSCALL_START,
 762                                  VSYSCALL_END - VSYSCALL_START);
 763
 764         printk(KERN_INFO "Memory: %luk/%luk available (%ldk kernel code, "
 765                                 "%ldk reserved, %ldk data, %ldk init)\n",
 766                 (unsigned long) nr_free_pages() << (PAGE_SHIFT-10),
 767                 max_pfn << (PAGE_SHIFT-10),
 768                 codesize >> 10,
 769                 reservedpages << (PAGE_SHIFT-10),
 770                 datasize >> 10,
 771                 initsize >> 10);
 772
 773         cpa_init();
 774 }
 775
 776 void free_init_pages(char *what, unsigned long begin, unsigned long end)
 777 {
 778         unsigned long addr = begin;
 779
 780         if (addr >= end)
 781                 return;
 782
 783         /*
 784          * If debugging page accesses then do not free this memory but
 785          * mark them not present - any buggy init-section access will
 786          * create a kernel page fault:
 787          */
 788 #ifdef CONFIG_DEBUG_PAGEALLOC
 789         printk(KERN_INFO "debug: unmapping init memory %08lx..%08lx\n",
 790                 begin, PAGE_ALIGN(end));
 791         set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
 792 #else
 793         printk(KERN_INFO "Freeing %s: %luk freed\n", what, (end - begin) >> 10);
 794
 795         for (; addr < end; addr += PAGE_SIZE) {
 796                 ClearPageReserved(virt_to_page(addr));
 797                 init_page_count(virt_to_page(addr));
 798                 memset((void *)(addr & ~(PAGE_SIZE-1)),
 799                         POISON_FREE_INITMEM, PAGE_SIZE);
 800                 free_page(addr);
 801                 totalram_pages++;
 802         }
 803 #endif
 804 }
 805
 806 void free_initmem(void)
 807 {
 808         free_init_pages("unused kernel memory",
 809                         (unsigned long)(&__init_begin),
 810                         (unsigned long)(&__init_end));
 811 }
 812
 813 #ifdef CONFIG_DEBUG_RODATA
 814 const int rodata_test_data = 0xC3;
 815 EXPORT_SYMBOL_GPL(rodata_test_data);
 816
 817 void mark_rodata_ro(void)
 818 {
 819         unsigned long start = PFN_ALIGN(_stext), end = PFN_ALIGN(__end_rodata);
 820
 821         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
 822                (end - start) >> 10);
 823         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
 824
 825         /*
 826          * The rodata section (but not the kernel text!) should also be
 827          * not-executable.
 828          */
 829         start = ((unsigned long)__start_rodata + PAGE_SIZE - 1) & PAGE_MASK;
 830         set_memory_nx(start, (end - start) >> PAGE_SHIFT);
 831
 832         rodata_test();
 833
 834 #ifdef CONFIG_CPA_DEBUG
 835         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
 836         set_memory_rw(start, (end-start) >> PAGE_SHIFT);
 837
 838         printk(KERN_INFO "Testing CPA: again\n");
 839         set_memory_ro(start, (end-start) >> PAGE_SHIFT);
 840 #endif
 841 }
 842
 843 #endif
 844
 845 #ifdef CONFIG_BLK_DEV_INITRD
 846 void free_initrd_mem(unsigned long start, unsigned long end)
 847 {
 848         free_init_pages("initrd memory", start, end);
 849 }
 850 #endif
 851
 852 int __init reserve_bootmem_generic(unsigned long phys, unsigned long len,
 853                                    int flags)
 854 {
 855 #ifdef CONFIG_NUMA
 856         int nid, next_nid;
 857         int ret;
 858 #endif
 859         unsigned long pfn = phys >> PAGE_SHIFT;
 860
 861         if (pfn >= max_pfn) {
 862                 /*
 863                  * This can happen with kdump kernels when accessing
 864                  * firmware tables:
 865                  */
 866                 if (pfn < max_pfn_mapped)
 867                         return -EFAULT;
 868
 869                 printk(KERN_ERR "reserve_bootmem: illegal reserve %lx %lu\n",
 870                                 phys, len);
 871                 return -EFAULT;
 872         }
 873
 874         /* Should check here against the e820 map to avoid double free */
 875 #ifdef CONFIG_NUMA
 876         nid = phys_to_nid(phys);
 877         next_nid = phys_to_nid(phys + len - 1);
 878         if (nid == next_nid)
 879                 ret = reserve_bootmem_node(NODE_DATA(nid), phys, len, flags);
 880         else
 881                 ret = reserve_bootmem(phys, len, flags);
 882
 883         if (ret != 0)
 884                 return ret;
 885
 886 #else
 887         reserve_bootmem(phys, len, BOOTMEM_DEFAULT);
 888 #endif
 889
 890         if (phys+len <= MAX_DMA_PFN*PAGE_SIZE) {
 891                 dma_reserve += len / PAGE_SIZE;
 892                 set_dma_reserve(dma_reserve);
 893         }
 894
 895         return 0;
 896 }
 897
 898 int kern_addr_valid(unsigned long addr)
 899 {
 900         unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
 901         pgd_t *pgd;
 902         pud_t *pud;
 903         pmd_t *pmd;
 904         pte_t *pte;
 905
 906         if (above != 0 && above != -1UL)
 907                 return 0;
 908
 909         pgd = pgd_offset_k(addr);
 910         if (pgd_none(*pgd))
 911                 return 0;
 912
 913         pud = pud_offset(pgd, addr);
 914         if (pud_none(*pud))
 915                 return 0;
 916
 917         pmd = pmd_offset(pud, addr);
 918         if (pmd_none(*pmd))
 919                 return 0;
 920
 921         if (pmd_large(*pmd))
 922                 return pfn_valid(pmd_pfn(*pmd));
 923
 924         pte = pte_offset_kernel(pmd, addr);
 925         if (pte_none(*pte))
 926                 return 0;
 927
 928         return pfn_valid(pte_pfn(*pte));
 929 }
 930
 931 /*
 932  * A pseudo VMA to allow ptrace access for the vsyscall page.  This only
 933  * covers the 64bit vsyscall page now. 32bit has a real VMA now and does
 934  * not need special handling anymore:
 935  */
 936 static struct vm_area_struct gate_vma = {
 937         .vm_start       = VSYSCALL_START,
 938         .vm_end         = VSYSCALL_START + (VSYSCALL_MAPPED_PAGES * PAGE_SIZE),
 939         .vm_page_prot   = PAGE_READONLY_EXEC,
 940         .vm_flags       = VM_READ | VM_EXEC
 941 };
 942
 943 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
 944 {
 945 #ifdef CONFIG_IA32_EMULATION
 946         if (test_tsk_thread_flag(tsk, TIF_IA32))
 947                 return NULL;
 948 #endif
 949         return &gate_vma;
 950 }
 951
 952 int in_gate_area(struct task_struct *task, unsigned long addr)
 953 {
 954         struct vm_area_struct *vma = get_gate_vma(task);
 955
 956         if (!vma)
 957                 return 0;
 958
 959         return (addr >= vma->vm_start) && (addr < vma->vm_end);
 960 }
 961
 962 /*
 963  * Use this when you have no reliable task/vma, typically from interrupt
 964  * context. It is less reliable than using the task's vma and may give
 965  * false positives:
 966  */
 967 int in_gate_area_no_task(unsigned long addr)
 968 {
 969         return (addr >= VSYSCALL_START) && (addr < VSYSCALL_END);
 970 }
 971
 972 const char *arch_vma_name(struct vm_area_struct *vma)
 973 {
 974         if (vma->vm_mm && vma->vm_start == (long)vma->vm_mm->context.vdso)
 975                 return "[vdso]";
 976         if (vma == &gate_vma)
 977                 return "[vsyscall]";
 978         return NULL;
 979 }
 980
 981 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 982 /*
 983  * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
 984  */
 985 static long __meminitdata addr_start, addr_end;
 986 static void __meminitdata *p_start, *p_end;
 987 static int __meminitdata node_start;
 988
 989 int __meminit
 990 vmemmap_populate(struct page *start_page, unsigned long size, int node)
 991 {
 992         unsigned long addr = (unsigned long)start_page;
 993         unsigned long end = (unsigned long)(start_page + size);
 994         unsigned long next;
 995         pgd_t *pgd;
 996         pud_t *pud;
 997         pmd_t *pmd;
 998
 999         for (; addr < end; addr = next) {
1000                 void *p = NULL;
1001
1002                 pgd = vmemmap_pgd_populate(addr, node);
1003                 if (!pgd)
1004                         return -ENOMEM;
1005
1006                 pud = vmemmap_pud_populate(pgd, addr, node);
1007                 if (!pud)
1008                         return -ENOMEM;
1009
1010                 if (!cpu_has_pse) {
1011                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1012                         pmd = vmemmap_pmd_populate(pud, addr, node);
1013
1014                         if (!pmd)
1015                                 return -ENOMEM;
1016
1017                         p = vmemmap_pte_populate(pmd, addr, node);
1018
1019                         if (!p)
1020                                 return -ENOMEM;
1021
1022                         addr_end = addr + PAGE_SIZE;
1023                         p_end = p + PAGE_SIZE;
1024                 } else {
1025                         next = pmd_addr_end(addr, end);
1026
1027                         pmd = pmd_offset(pud, addr);
1028                         if (pmd_none(*pmd)) {
1029                                 pte_t entry;
1030
1031                                 p = vmemmap_alloc_block(PMD_SIZE, node);
1032                                 if (!p)
1033                                         return -ENOMEM;
1034
1035                                 entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
1036                                                 PAGE_KERNEL_LARGE);
1037                                 set_pmd(pmd, __pmd(pte_val(entry)));
1038
1039                                 addr_end = addr + PMD_SIZE;
1040                                 p_end = p + PMD_SIZE;
1041
1042                                 /* check to see if we have contiguous blocks */
1043                                 if (p_end != p || node_start != node) {
1044                                         if (p_start)
1045                                                 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1046                                                        addr_start, addr_end-1, p_start, p_end-1, node_start);
1047                                         addr_start = addr;
1048                                         node_start = node;
1049                                         p_start = p;
1050                                 }
1051                         } else
1052                                 vmemmap_verify((pte_t *)pmd, node, addr, next);
1053                 }
1054
1055         }
1056         return 0;
1057 }
1058
1059 void __meminit vmemmap_populate_print_last(void)
1060 {
1061         if (p_start) {
1062                 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1063                         addr_start, addr_end-1, p_start, p_end-1, node_start);
1064                 p_start = NULL;
1065                 p_end = NULL;
1066                 node_start = 0;
1067         }
1068 }
1069 #endif