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, int use_pse,
 460                                           int use_gbpages)
 461 {
 462         unsigned long puds, pmds, ptes, tables, start;
 463
 464         puds = (end + PUD_SIZE - 1) >> PUD_SHIFT;
 465         tables = round_up(puds * sizeof(pud_t), PAGE_SIZE);
 466         if (use_gbpages) {
 467                 unsigned long extra;
 468                 extra = end - ((end>>PUD_SHIFT) << PUD_SHIFT);
 469                 pmds = (extra + PMD_SIZE - 1) >> PMD_SHIFT;
 470         } else
 471                 pmds = (end + PMD_SIZE - 1) >> PMD_SHIFT;
 472         tables += round_up(pmds * sizeof(pmd_t), PAGE_SIZE);
 473
 474         if (use_pse) {
 475                 unsigned long extra;
 476                 extra = end - ((end>>PMD_SHIFT) << PMD_SHIFT);
 477                 ptes = (extra + PAGE_SIZE - 1) >> PAGE_SHIFT;
 478         } else
 479                 ptes = (end + PAGE_SIZE - 1) >> PAGE_SHIFT;
 480         tables += round_up(ptes * sizeof(pte_t), PAGE_SIZE);
 481
 482         /*
 483          * RED-PEN putting page tables only on node 0 could
 484          * cause a hotspot and fill up ZONE_DMA. The page tables
 485          * need roughly 0.5KB per GB.
 486          */
 487         start = 0x8000;
 488         table_start = find_e820_area(start, end, tables, PAGE_SIZE);
 489         if (table_start == -1UL)
 490                 panic("Cannot find space for the kernel page tables");
 491
 492         table_start >>= PAGE_SHIFT;
 493         table_end = table_start;
 494         table_top = table_start + (tables >> PAGE_SHIFT);
 495
 496         printk(KERN_DEBUG "kernel direct mapping tables up to %lx @ %lx-%lx\n",
 497                 end, table_start << PAGE_SHIFT, table_top << PAGE_SHIFT);
 498 }
 499
 500 static void __init init_gbpages(void)
 501 {
 502         if (direct_gbpages && cpu_has_gbpages)
 503                 printk(KERN_INFO "Using GB pages for direct mapping\n");
 504         else
 505                 direct_gbpages = 0;
 506 }
 507
 508 static int is_kernel(unsigned long pfn)
 509 {
 510         unsigned long pg_addresss = pfn << PAGE_SHIFT;
 511
 512         if (pg_addresss >= (unsigned long) __pa(_text) &&
 513             pg_addresss <= (unsigned long) __pa(_end))
 514                 return 1;
 515
 516         return 0;
 517 }
 518
 519 static unsigned long __init kernel_physical_mapping_init(unsigned long start,
 520                                                 unsigned long end,
 521                                                 unsigned long page_size_mask)
 522 {
 523
 524         unsigned long next, last_map_addr;
 525         u64 cached_supported_pte_mask = __supported_pte_mask;
 526         unsigned long cache_start = start;
 527         unsigned long cache_end = end;
 528
 529         /*
 530          * First iteration will setup identity mapping using large/small pages
 531          * based on page_size_mask, with other attributes same as set by
 532          * the early code in head_64.S
 533          *
 534          * Second iteration will setup the appropriate attributes
 535          * as desired for the kernel identity mapping.
 536          *
 537          * This two pass mechanism conforms to the TLB app note which says:
 538          *
 539          *     "Software should not write to a paging-structure entry in a way
 540          *      that would change, for any linear address, both the page size
 541          *      and either the page frame or attributes."
 542          *
 543          * For now, only difference between very early PTE attributes used in
 544          * head_64.S and here is _PAGE_NX.
 545          */
 546         BUILD_BUG_ON((__PAGE_KERNEL_LARGE & ~__PAGE_KERNEL_IDENT_LARGE_EXEC)
 547                      != _PAGE_NX);
 548         __supported_pte_mask &= ~(_PAGE_NX);
 549         physical_mapping_iter = 1;
 550
 551 repeat:
 552         last_map_addr = cache_end;
 553
 554         start = (unsigned long)__va(cache_start);
 555         end = (unsigned long)__va(cache_end);
 556
 557         for (; start < end; start = next) {
 558                 pgd_t *pgd = pgd_offset_k(start);
 559                 unsigned long pud_phys;
 560                 pud_t *pud;
 561
 562                 next = (start + PGDIR_SIZE) & PGDIR_MASK;
 563                 if (next > end)
 564                         next = end;
 565
 566                 if (pgd_val(*pgd)) {
 567                         /*
 568                          * Static identity mappings will be overwritten
 569                          * with run-time mappings. For example, this allows
 570                          * the static 0-1GB identity mapping to be mapped
 571                          * non-executable with this.
 572                          */
 573                         if (is_kernel(pte_pfn(*((pte_t *) pgd))))
 574                                 goto realloc;
 575
 576                         last_map_addr = phys_pud_update(pgd, __pa(start),
 577                                                  __pa(end), page_size_mask);
 578                         continue;
 579                 }
 580
 581 realloc:
 582                 pud = alloc_low_page(&pud_phys);
 583                 last_map_addr = phys_pud_init(pud, __pa(start), __pa(next),
 584                                                  page_size_mask);
 585                 unmap_low_page(pud);
 586
 587                 spin_lock(&init_mm.page_table_lock);
 588                 pgd_populate(&init_mm, pgd, __va(pud_phys));
 589                 spin_unlock(&init_mm.page_table_lock);
 590         }
 591         __flush_tlb_all();
 592
 593         if (physical_mapping_iter == 1) {
 594                 physical_mapping_iter = 2;
 595                 /*
 596                  * Second iteration will set the actual desired PTE attributes.
 597                  */
 598                 __supported_pte_mask = cached_supported_pte_mask;
 599                 goto repeat;
 600         }
 601
 602         return last_map_addr;
 603 }
 604
 605 struct map_range {
 606         unsigned long start;
 607         unsigned long end;
 608         unsigned page_size_mask;
 609 };
 610
 611 #define NR_RANGE_MR 5
 612
 613 static int save_mr(struct map_range *mr, int nr_range,
 614                    unsigned long start_pfn, unsigned long end_pfn,
 615                    unsigned long page_size_mask)
 616 {
 617
 618         if (start_pfn < end_pfn) {
 619                 if (nr_range >= NR_RANGE_MR)
 620                         panic("run out of range for init_memory_mapping\n");
 621                 mr[nr_range].start = start_pfn<<PAGE_SHIFT;
 622                 mr[nr_range].end   = end_pfn<<PAGE_SHIFT;
 623                 mr[nr_range].page_size_mask = page_size_mask;
 624                 nr_range++;
 625         }
 626
 627         return nr_range;
 628 }
 629
 630 /*
 631  * Setup the direct mapping of the physical memory at PAGE_OFFSET.
 632  * This runs before bootmem is initialized and gets pages directly from
 633  * the physical memory. To access them they are temporarily mapped.
 634  */
 635 unsigned long __init_refok init_memory_mapping(unsigned long start,
 636                                                unsigned long end)
 637 {
 638         unsigned long last_map_addr = 0;
 639         unsigned long page_size_mask = 0;
 640         unsigned long start_pfn, end_pfn;
 641
 642         struct map_range mr[NR_RANGE_MR];
 643         int nr_range, i;
 644         int use_pse, use_gbpages;
 645
 646         printk(KERN_INFO "init_memory_mapping\n");
 647
 648         /*
 649          * Find space for the kernel direct mapping tables.
 650          *
 651          * Later we should allocate these tables in the local node of the
 652          * memory mapped. Unfortunately this is done currently before the
 653          * nodes are discovered.
 654          */
 655         if (!after_bootmem)
 656                 init_gbpages();
 657
 658 #ifdef CONFIG_DEBUG_PAGEALLOC
 659         /*
 660          * For CONFIG_DEBUG_PAGEALLOC, identity mapping will use small pages.
 661          * This will simplify cpa(), which otherwise needs to support splitting
 662          * large pages into small in interrupt context, etc.
 663          */
 664         use_pse = use_gbpages = 0;
 665 #else
 666         use_pse = cpu_has_pse;
 667         use_gbpages = direct_gbpages;
 668 #endif
 669
 670         if (use_gbpages)
 671                 page_size_mask |= 1 << PG_LEVEL_1G;
 672         if (use_pse)
 673                 page_size_mask |= 1 << PG_LEVEL_2M;
 674
 675         memset(mr, 0, sizeof(mr));
 676         nr_range = 0;
 677
 678         /* head if not big page alignment ?*/
 679         start_pfn = start >> PAGE_SHIFT;
 680         end_pfn = ((start + (PMD_SIZE - 1)) >> PMD_SHIFT)
 681                         << (PMD_SHIFT - PAGE_SHIFT);
 682         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
 683
 684         /* big page (2M) range*/
 685         start_pfn = ((start + (PMD_SIZE - 1))>>PMD_SHIFT)
 686                          << (PMD_SHIFT - PAGE_SHIFT);
 687         end_pfn = ((start + (PUD_SIZE - 1))>>PUD_SHIFT)
 688                          << (PUD_SHIFT - PAGE_SHIFT);
 689         if (end_pfn > ((end>>PUD_SHIFT)<<(PUD_SHIFT - PAGE_SHIFT)))
 690                 end_pfn = ((end>>PUD_SHIFT)<<(PUD_SHIFT - PAGE_SHIFT));
 691         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 692                         page_size_mask & (1<<PG_LEVEL_2M));
 693
 694         /* big page (1G) range */
 695         start_pfn = end_pfn;
 696         end_pfn = (end>>PUD_SHIFT) << (PUD_SHIFT - PAGE_SHIFT);
 697         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 698                                 page_size_mask &
 699                                  ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G)));
 700
 701         /* tail is not big page (1G) alignment */
 702         start_pfn = end_pfn;
 703         end_pfn = (end>>PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
 704         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 705                         page_size_mask & (1<<PG_LEVEL_2M));
 706
 707         /* tail is not big page (2M) alignment */
 708         start_pfn = end_pfn;
 709         end_pfn = end>>PAGE_SHIFT;
 710         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
 711
 712         /* try to merge same page size and continuous */
 713         for (i = 0; nr_range > 1 && i < nr_range - 1; i++) {
 714                 unsigned long old_start;
 715                 if (mr[i].end != mr[i+1].start ||
 716                     mr[i].page_size_mask != mr[i+1].page_size_mask)
 717                         continue;
 718                 /* move it */
 719                 old_start = mr[i].start;
 720                 memmove(&mr[i], &mr[i+1],
 721                          (nr_range - 1 - i) * sizeof (struct map_range));
 722                 mr[i].start = old_start;
 723                 nr_range--;
 724         }
 725
 726         for (i = 0; i < nr_range; i++)
 727                 printk(KERN_DEBUG " %010lx - %010lx page %s\n",
 728                                 mr[i].start, mr[i].end,
 729                         (mr[i].page_size_mask & (1<<PG_LEVEL_1G))?"1G":(
 730                          (mr[i].page_size_mask & (1<<PG_LEVEL_2M))?"2M":"4k"));
 731
 732         if (!after_bootmem)
 733                 find_early_table_space(end, use_pse, use_gbpages);
 734
 735         for (i = 0; i < nr_range; i++)
 736                 last_map_addr = kernel_physical_mapping_init(
 737                                         mr[i].start, mr[i].end,
 738                                         mr[i].page_size_mask);
 739
 740         if (!after_bootmem)
 741                 mmu_cr4_features = read_cr4();
 742         __flush_tlb_all();
 743
 744         if (!after_bootmem && table_end > table_start)
 745                 reserve_early(table_start << PAGE_SHIFT,
 746                                  table_end << PAGE_SHIFT, "PGTABLE");
 747
 748         printk(KERN_INFO "last_map_addr: %lx end: %lx\n",
 749                          last_map_addr, end);
 750
 751         if (!after_bootmem)
 752                 early_memtest(start, end);
 753
 754         return last_map_addr >> PAGE_SHIFT;
 755 }
 756
 757 #ifndef CONFIG_NUMA
 758 void __init initmem_init(unsigned long start_pfn, unsigned long end_pfn)
 759 {
 760         unsigned long bootmap_size, bootmap;
 761
 762         bootmap_size = bootmem_bootmap_pages(end_pfn)<<PAGE_SHIFT;
 763         bootmap = find_e820_area(0, end_pfn<<PAGE_SHIFT, bootmap_size,
 764                                  PAGE_SIZE);
 765         if (bootmap == -1L)
 766                 panic("Cannot find bootmem map of size %ld\n", bootmap_size);
 767         /* don't touch min_low_pfn */
 768         bootmap_size = init_bootmem_node(NODE_DATA(0), bootmap >> PAGE_SHIFT,
 769                                          0, end_pfn);
 770         e820_register_active_regions(0, start_pfn, end_pfn);
 771         free_bootmem_with_active_regions(0, end_pfn);
 772         early_res_to_bootmem(0, end_pfn<<PAGE_SHIFT);
 773         reserve_bootmem(bootmap, bootmap_size, BOOTMEM_DEFAULT);
 774 }
 775
 776 void __init paging_init(void)
 777 {
 778         unsigned long max_zone_pfns[MAX_NR_ZONES];
 779
 780         memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
 781         max_zone_pfns[ZONE_DMA] = MAX_DMA_PFN;
 782         max_zone_pfns[ZONE_DMA32] = MAX_DMA32_PFN;
 783         max_zone_pfns[ZONE_NORMAL] = max_pfn;
 784
 785         memory_present(0, 0, max_pfn);
 786         sparse_init();
 787         free_area_init_nodes(max_zone_pfns);
 788 }
 789 #endif
 790
 791 /*
 792  * Memory hotplug specific functions
 793  */
 794 #ifdef CONFIG_MEMORY_HOTPLUG
 795 /*
 796  * Memory is added always to NORMAL zone. This means you will never get
 797  * additional DMA/DMA32 memory.
 798  */
 799 int arch_add_memory(int nid, u64 start, u64 size)
 800 {
 801         struct pglist_data *pgdat = NODE_DATA(nid);
 802         struct zone *zone = pgdat->node_zones + ZONE_NORMAL;
 803         unsigned long last_mapped_pfn, start_pfn = start >> PAGE_SHIFT;
 804         unsigned long nr_pages = size >> PAGE_SHIFT;
 805         int ret;
 806
 807         last_mapped_pfn = init_memory_mapping(start, start + size-1);
 808         if (last_mapped_pfn > max_pfn_mapped)
 809                 max_pfn_mapped = last_mapped_pfn;
 810
 811         ret = __add_pages(zone, start_pfn, nr_pages);
 812         WARN_ON(1);
 813
 814         return ret;
 815 }
 816 EXPORT_SYMBOL_GPL(arch_add_memory);
 817
 818 #if !defined(CONFIG_ACPI_NUMA) && defined(CONFIG_NUMA)
 819 int memory_add_physaddr_to_nid(u64 start)
 820 {
 821         return 0;
 822 }
 823 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
 824 #endif
 825
 826 #endif /* CONFIG_MEMORY_HOTPLUG */
 827
 828 /*
 829  * devmem_is_allowed() checks to see if /dev/mem access to a certain address
 830  * is valid. The argument is a physical page number.
 831  *
 832  *
 833  * On x86, access has to be given to the first megabyte of ram because that area
 834  * contains bios code and data regions used by X and dosemu and similar apps.
 835  * Access has to be given to non-kernel-ram areas as well, these contain the PCI
 836  * mmio resources as well as potential bios/acpi data regions.
 837  */
 838 int devmem_is_allowed(unsigned long pagenr)
 839 {
 840         if (pagenr <= 256)
 841                 return 1;
 842         if (!page_is_ram(pagenr))
 843                 return 1;
 844         return 0;
 845 }
 846
 847
 848 static struct kcore_list kcore_mem, kcore_vmalloc, kcore_kernel,
 849                          kcore_modules, kcore_vsyscall;
 850
 851 void __init mem_init(void)
 852 {
 853         long codesize, reservedpages, datasize, initsize;
 854
 855         pci_iommu_alloc();
 856
 857         /* clear_bss() already clear the empty_zero_page */
 858
 859         reservedpages = 0;
 860
 861         /* this will put all low memory onto the freelists */
 862 #ifdef CONFIG_NUMA
 863         totalram_pages = numa_free_all_bootmem();
 864 #else
 865         totalram_pages = free_all_bootmem();
 866 #endif
 867         reservedpages = max_pfn - totalram_pages -
 868                                         absent_pages_in_range(0, max_pfn);
 869         after_bootmem = 1;
 870
 871         codesize =  (unsigned long) &_etext - (unsigned long) &_text;
 872         datasize =  (unsigned long) &_edata - (unsigned long) &_etext;
 873         initsize =  (unsigned long) &__init_end - (unsigned long) &__init_begin;
 874
 875         /* Register memory areas for /proc/kcore */
 876         kclist_add(&kcore_mem, __va(0), max_low_pfn << PAGE_SHIFT);
 877         kclist_add(&kcore_vmalloc, (void *)VMALLOC_START,
 878                    VMALLOC_END-VMALLOC_START);
 879         kclist_add(&kcore_kernel, &_stext, _end - _stext);
 880         kclist_add(&kcore_modules, (void *)MODULES_VADDR, MODULES_LEN);
 881         kclist_add(&kcore_vsyscall, (void *)VSYSCALL_START,
 882                                  VSYSCALL_END - VSYSCALL_START);
 883
 884         printk(KERN_INFO "Memory: %luk/%luk available (%ldk kernel code, "
 885                                 "%ldk reserved, %ldk data, %ldk init)\n",
 886                 (unsigned long) nr_free_pages() << (PAGE_SHIFT-10),
 887                 max_pfn << (PAGE_SHIFT-10),
 888                 codesize >> 10,
 889                 reservedpages << (PAGE_SHIFT-10),
 890                 datasize >> 10,
 891                 initsize >> 10);
 892
 893         cpa_init();
 894 }
 895
 896 void free_init_pages(char *what, unsigned long begin, unsigned long end)
 897 {
 898         unsigned long addr = begin;
 899
 900         if (addr >= end)
 901                 return;
 902
 903         /*
 904          * If debugging page accesses then do not free this memory but
 905          * mark them not present - any buggy init-section access will
 906          * create a kernel page fault:
 907          */
 908 #ifdef CONFIG_DEBUG_PAGEALLOC
 909         printk(KERN_INFO "debug: unmapping init memory %08lx..%08lx\n",
 910                 begin, PAGE_ALIGN(end));
 911         set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
 912 #else
 913         printk(KERN_INFO "Freeing %s: %luk freed\n", what, (end - begin) >> 10);
 914
 915         for (; addr < end; addr += PAGE_SIZE) {
 916                 ClearPageReserved(virt_to_page(addr));
 917                 init_page_count(virt_to_page(addr));
 918                 memset((void *)(addr & ~(PAGE_SIZE-1)),
 919                         POISON_FREE_INITMEM, PAGE_SIZE);
 920                 free_page(addr);
 921                 totalram_pages++;
 922         }
 923 #endif
 924 }
 925
 926 void free_initmem(void)
 927 {
 928         free_init_pages("unused kernel memory",
 929                         (unsigned long)(&__init_begin),
 930                         (unsigned long)(&__init_end));
 931 }
 932
 933 #ifdef CONFIG_DEBUG_RODATA
 934 const int rodata_test_data = 0xC3;
 935 EXPORT_SYMBOL_GPL(rodata_test_data);
 936
 937 void mark_rodata_ro(void)
 938 {
 939         unsigned long start = PFN_ALIGN(_stext), end = PFN_ALIGN(__end_rodata);
 940         unsigned long rodata_start =
 941                 ((unsigned long)__start_rodata + PAGE_SIZE - 1) & PAGE_MASK;
 942
 943 #ifdef CONFIG_DYNAMIC_FTRACE
 944         /* Dynamic tracing modifies the kernel text section */
 945         start = rodata_start;
 946 #endif
 947
 948         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
 949                (end - start) >> 10);
 950         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
 951
 952         /*
 953          * The rodata section (but not the kernel text!) should also be
 954          * not-executable.
 955          */
 956         set_memory_nx(rodata_start, (end - rodata_start) >> PAGE_SHIFT);
 957
 958         rodata_test();
 959
 960 #ifdef CONFIG_CPA_DEBUG
 961         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
 962         set_memory_rw(start, (end-start) >> PAGE_SHIFT);
 963
 964         printk(KERN_INFO "Testing CPA: again\n");
 965         set_memory_ro(start, (end-start) >> PAGE_SHIFT);
 966 #endif
 967 }
 968
 969 #endif
 970
 971 #ifdef CONFIG_BLK_DEV_INITRD
 972 void free_initrd_mem(unsigned long start, unsigned long end)
 973 {
 974         free_init_pages("initrd memory", start, end);
 975 }
 976 #endif
 977
 978 int __init reserve_bootmem_generic(unsigned long phys, unsigned long len,
 979                                    int flags)
 980 {
 981 #ifdef CONFIG_NUMA
 982         int nid, next_nid;
 983         int ret;
 984 #endif
 985         unsigned long pfn = phys >> PAGE_SHIFT;
 986
 987         if (pfn >= max_pfn) {
 988                 /*
 989                  * This can happen with kdump kernels when accessing
 990                  * firmware tables:
 991                  */
 992                 if (pfn < max_pfn_mapped)
 993                         return -EFAULT;
 994
 995                 printk(KERN_ERR "reserve_bootmem: illegal reserve %lx %lu\n",
 996                                 phys, len);
 997                 return -EFAULT;
 998         }
 999
1000         /* Should check here against the e820 map to avoid double free */
1001 #ifdef CONFIG_NUMA
1002         nid = phys_to_nid(phys);
1003         next_nid = phys_to_nid(phys + len - 1);
1004         if (nid == next_nid)
1005                 ret = reserve_bootmem_node(NODE_DATA(nid), phys, len, flags);
1006         else
1007                 ret = reserve_bootmem(phys, len, flags);
1008
1009         if (ret != 0)
1010                 return ret;
1011
1012 #else
1013         reserve_bootmem(phys, len, BOOTMEM_DEFAULT);
1014 #endif
1015
1016         if (phys+len <= MAX_DMA_PFN*PAGE_SIZE) {
1017                 dma_reserve += len / PAGE_SIZE;
1018                 set_dma_reserve(dma_reserve);
1019         }
1020
1021         return 0;
1022 }
1023
1024 int kern_addr_valid(unsigned long addr)
1025 {
1026         unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
1027         pgd_t *pgd;
1028         pud_t *pud;
1029         pmd_t *pmd;
1030         pte_t *pte;
1031
1032         if (above != 0 && above != -1UL)
1033                 return 0;
1034
1035         pgd = pgd_offset_k(addr);
1036         if (pgd_none(*pgd))
1037                 return 0;
1038
1039         pud = pud_offset(pgd, addr);
1040         if (pud_none(*pud))
1041                 return 0;
1042
1043         pmd = pmd_offset(pud, addr);
1044         if (pmd_none(*pmd))
1045                 return 0;
1046
1047         if (pmd_large(*pmd))
1048                 return pfn_valid(pmd_pfn(*pmd));
1049
1050         pte = pte_offset_kernel(pmd, addr);
1051         if (pte_none(*pte))
1052                 return 0;
1053
1054         return pfn_valid(pte_pfn(*pte));
1055 }
1056
1057 /*
1058  * A pseudo VMA to allow ptrace access for the vsyscall page.  This only
1059  * covers the 64bit vsyscall page now. 32bit has a real VMA now and does
1060  * not need special handling anymore:
1061  */
1062 static struct vm_area_struct gate_vma = {
1063         .vm_start       = VSYSCALL_START,
1064         .vm_end         = VSYSCALL_START + (VSYSCALL_MAPPED_PAGES * PAGE_SIZE),
1065         .vm_page_prot   = PAGE_READONLY_EXEC,
1066         .vm_flags       = VM_READ | VM_EXEC
1067 };
1068
1069 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
1070 {
1071 #ifdef CONFIG_IA32_EMULATION
1072         if (test_tsk_thread_flag(tsk, TIF_IA32))
1073                 return NULL;
1074 #endif
1075         return &gate_vma;
1076 }
1077
1078 int in_gate_area(struct task_struct *task, unsigned long addr)
1079 {
1080         struct vm_area_struct *vma = get_gate_vma(task);
1081
1082         if (!vma)
1083                 return 0;
1084
1085         return (addr >= vma->vm_start) && (addr < vma->vm_end);
1086 }
1087
1088 /*
1089  * Use this when you have no reliable task/vma, typically from interrupt
1090  * context. It is less reliable than using the task's vma and may give
1091  * false positives:
1092  */
1093 int in_gate_area_no_task(unsigned long addr)
1094 {
1095         return (addr >= VSYSCALL_START) && (addr < VSYSCALL_END);
1096 }
1097
1098 const char *arch_vma_name(struct vm_area_struct *vma)
1099 {
1100         if (vma->vm_mm && vma->vm_start == (long)vma->vm_mm->context.vdso)
1101                 return "[vdso]";
1102         if (vma == &gate_vma)
1103                 return "[vsyscall]";
1104         return NULL;
1105 }
1106
1107 #ifdef CONFIG_SPARSEMEM_VMEMMAP
1108 /*
1109  * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
1110  */
1111 static long __meminitdata addr_start, addr_end;
1112 static void __meminitdata *p_start, *p_end;
1113 static int __meminitdata node_start;
1114
1115 int __meminit
1116 vmemmap_populate(struct page *start_page, unsigned long size, int node)
1117 {
1118         unsigned long addr = (unsigned long)start_page;
1119         unsigned long end = (unsigned long)(start_page + size);
1120         unsigned long next;
1121         pgd_t *pgd;
1122         pud_t *pud;
1123         pmd_t *pmd;
1124
1125         for (; addr < end; addr = next) {
1126                 void *p = NULL;
1127
1128                 pgd = vmemmap_pgd_populate(addr, node);
1129                 if (!pgd)
1130                         return -ENOMEM;
1131
1132                 pud = vmemmap_pud_populate(pgd, addr, node);
1133                 if (!pud)
1134                         return -ENOMEM;
1135
1136                 if (!cpu_has_pse) {
1137                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1138                         pmd = vmemmap_pmd_populate(pud, addr, node);
1139
1140                         if (!pmd)
1141                                 return -ENOMEM;
1142
1143                         p = vmemmap_pte_populate(pmd, addr, node);
1144
1145                         if (!p)
1146                                 return -ENOMEM;
1147
1148                         addr_end = addr + PAGE_SIZE;
1149                         p_end = p + PAGE_SIZE;
1150                 } else {
1151                         next = pmd_addr_end(addr, end);
1152
1153                         pmd = pmd_offset(pud, addr);
1154                         if (pmd_none(*pmd)) {
1155                                 pte_t entry;
1156
1157                                 p = vmemmap_alloc_block(PMD_SIZE, node);
1158                                 if (!p)
1159                                         return -ENOMEM;
1160
1161                                 entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
1162                                                 PAGE_KERNEL_LARGE);
1163                                 set_pmd(pmd, __pmd(pte_val(entry)));
1164
1165                                 /* check to see if we have contiguous blocks */
1166                                 if (p_end != p || node_start != node) {
1167                                         if (p_start)
1168                                                 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1169                                                        addr_start, addr_end-1, p_start, p_end-1, node_start);
1170                                         addr_start = addr;
1171                                         node_start = node;
1172                                         p_start = p;
1173                                 }
1174
1175                                 addr_end = addr + PMD_SIZE;
1176                                 p_end = p + PMD_SIZE;
1177                         } else
1178                                 vmemmap_verify((pte_t *)pmd, node, addr, next);
1179                 }
1180
1181         }
1182         return 0;
1183 }
1184
1185 void __meminit vmemmap_populate_print_last(void)
1186 {
1187         if (p_start) {
1188                 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1189                         addr_start, addr_end-1, p_start, p_end-1, node_start);
1190                 p_start = NULL;
1191                 p_end = NULL;
1192                 node_start = 0;
1193         }
1194 }
1195 #endif