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/pagemap.h>
  22 #include <linux/bootmem.h>
  23 #include <linux/proc_fs.h>
  24 #include <linux/pci.h>
  25 #include <linux/pfn.h>
  26 #include <linux/poison.h>
  27 #include <linux/dma-mapping.h>
  28 #include <linux/module.h>
  29 #include <linux/memory_hotplug.h>
  30 #include <linux/nmi.h>
  31
  32 #include <asm/processor.h>
  33 #include <asm/system.h>
  34 #include <asm/uaccess.h>
  35 #include <asm/pgtable.h>
  36 #include <asm/pgalloc.h>
  37 #include <asm/dma.h>
  38 #include <asm/fixmap.h>
  39 #include <asm/e820.h>
  40 #include <asm/apic.h>
  41 #include <asm/tlb.h>
  42 #include <asm/mmu_context.h>
  43 #include <asm/proto.h>
  44 #include <asm/smp.h>
  45 #include <asm/sections.h>
  46 #include <asm/kdebug.h>
  47 #include <asm/numa.h>
  48 #include <asm/cacheflush.h>
  49
  50 const struct dma_mapping_ops *dma_ops;
  51 EXPORT_SYMBOL(dma_ops);
  52
  53 static unsigned long dma_reserve __initdata;
  54
  55 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
  56
  57 int direct_gbpages __meminitdata
  58 #ifdef CONFIG_DIRECT_GBPAGES
  59                                 = 1
  60 #endif
  61 ;
  62
  63 static int __init parse_direct_gbpages_off(char *arg)
  64 {
  65         direct_gbpages = 0;
  66         return 0;
  67 }
  68 early_param("nogbpages", parse_direct_gbpages_off);
  69
  70 static int __init parse_direct_gbpages_on(char *arg)
  71 {
  72         direct_gbpages = 1;
  73         return 0;
  74 }
  75 early_param("gbpages", parse_direct_gbpages_on);
  76
  77 /*
  78  * NOTE: pagetable_init alloc all the fixmap pagetables contiguous on the
  79  * physical space so we can cache the place of the first one and move
  80  * around without checking the pgd every time.
  81  */
  82
  83 void show_mem(void)
  84 {
  85         long i, total = 0, reserved = 0;
  86         long shared = 0, cached = 0;
  87         struct page *page;
  88         pg_data_t *pgdat;
  89
  90         printk(KERN_INFO "Mem-info:\n");
  91         show_free_areas();
  92         for_each_online_pgdat(pgdat) {
  93                 for (i = 0; i < pgdat->node_spanned_pages; ++i) {
  94                         /*
  95                          * This loop can take a while with 256 GB and
  96                          * 4k pages so defer the NMI watchdog:
  97                          */
  98                         if (unlikely(i % MAX_ORDER_NR_PAGES == 0))
  99                                 touch_nmi_watchdog();
 100
 101                         if (!pfn_valid(pgdat->node_start_pfn + i))
 102                                 continue;
 103
 104                         page = pfn_to_page(pgdat->node_start_pfn + i);
 105                         total++;
 106                         if (PageReserved(page))
 107                                 reserved++;
 108                         else if (PageSwapCache(page))
 109                                 cached++;
 110                         else if (page_count(page))
 111                                 shared += page_count(page) - 1;
 112                 }
 113         }
 114         printk(KERN_INFO "%lu pages of RAM\n",          total);
 115         printk(KERN_INFO "%lu reserved pages\n",        reserved);
 116         printk(KERN_INFO "%lu pages shared\n",          shared);
 117         printk(KERN_INFO "%lu pages swap cached\n",     cached);
 118 }
 119
 120 int after_bootmem;
 121
 122 static __init void *spp_getpage(void)
 123 {
 124         void *ptr;
 125
 126         if (after_bootmem)
 127                 ptr = (void *) get_zeroed_page(GFP_ATOMIC);
 128         else
 129                 ptr = alloc_bootmem_pages(PAGE_SIZE);
 130
 131         if (!ptr || ((unsigned long)ptr & ~PAGE_MASK)) {
 132                 panic("set_pte_phys: cannot allocate page data %s\n",
 133                         after_bootmem ? "after bootmem" : "");
 134         }
 135
 136         pr_debug("spp_getpage %p\n", ptr);
 137
 138         return ptr;
 139 }
 140
 141 static __init void
 142 set_pte_phys(unsigned long vaddr, unsigned long phys, pgprot_t prot)
 143 {
 144         pgd_t *pgd;
 145         pud_t *pud;
 146         pmd_t *pmd;
 147         pte_t *pte, new_pte;
 148
 149         pr_debug("set_pte_phys %lx to %lx\n", vaddr, phys);
 150
 151         pgd = pgd_offset_k(vaddr);
 152         if (pgd_none(*pgd)) {
 153                 printk(KERN_ERR
 154                         "PGD FIXMAP MISSING, it should be setup in head.S!\n");
 155                 return;
 156         }
 157         pud = pud_offset(pgd, vaddr);
 158         if (pud_none(*pud)) {
 159                 pmd = (pmd_t *) spp_getpage();
 160                 set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE | _PAGE_USER));
 161                 if (pmd != pmd_offset(pud, 0)) {
 162                         printk(KERN_ERR "PAGETABLE BUG #01! %p <-> %p\n",
 163                                 pmd, pmd_offset(pud, 0));
 164                         return;
 165                 }
 166         }
 167         pmd = pmd_offset(pud, vaddr);
 168         if (pmd_none(*pmd)) {
 169                 pte = (pte_t *) spp_getpage();
 170                 set_pmd(pmd, __pmd(__pa(pte) | _KERNPG_TABLE | _PAGE_USER));
 171                 if (pte != pte_offset_kernel(pmd, 0)) {
 172                         printk(KERN_ERR "PAGETABLE BUG #02!\n");
 173                         return;
 174                 }
 175         }
 176         new_pte = pfn_pte(phys >> PAGE_SHIFT, prot);
 177
 178         pte = pte_offset_kernel(pmd, vaddr);
 179         if (!pte_none(*pte) &&
 180             pte_val(*pte) != (pte_val(new_pte) & __supported_pte_mask))
 181                 pte_ERROR(*pte);
 182         set_pte(pte, new_pte);
 183
 184         /*
 185          * It's enough to flush this one mapping.
 186          * (PGE mappings get flushed as well)
 187          */
 188         __flush_tlb_one(vaddr);
 189 }
 190
 191 /*
 192  * The head.S code sets up the kernel high mapping:
 193  *
 194  *   from __START_KERNEL_map to __START_KERNEL_map + size (== _end-_text)
 195  *
 196  * phys_addr holds the negative offset to the kernel, which is added
 197  * to the compile time generated pmds. This results in invalid pmds up
 198  * to the point where we hit the physaddr 0 mapping.
 199  *
 200  * We limit the mappings to the region from _text to _end.  _end is
 201  * rounded up to the 2MB boundary. This catches the invalid pmds as
 202  * well, as they are located before _text:
 203  */
 204 void __init cleanup_highmap(void)
 205 {
 206         unsigned long vaddr = __START_KERNEL_map;
 207         unsigned long end = round_up((unsigned long)_end, PMD_SIZE) - 1;
 208         pmd_t *pmd = level2_kernel_pgt;
 209         pmd_t *last_pmd = pmd + PTRS_PER_PMD;
 210
 211         for (; pmd < last_pmd; pmd++, vaddr += PMD_SIZE) {
 212                 if (!pmd_present(*pmd))
 213                         continue;
 214                 if (vaddr < (unsigned long) _text || vaddr > end)
 215                         set_pmd(pmd, __pmd(0));
 216         }
 217 }
 218
 219 /* NOTE: this is meant to be run only at boot */
 220 void __init
 221 __set_fixmap(enum fixed_addresses idx, unsigned long phys, pgprot_t prot)
 222 {
 223         unsigned long address = __fix_to_virt(idx);
 224
 225         if (idx >= __end_of_fixed_addresses) {
 226                 printk(KERN_ERR "Invalid __set_fixmap\n");
 227                 return;
 228         }
 229         set_pte_phys(address, phys, prot);
 230 }
 231
 232 static unsigned long __initdata table_start;
 233 static unsigned long __meminitdata table_end;
 234
 235 static __meminit void *alloc_low_page(unsigned long *phys)
 236 {
 237         unsigned long pfn = table_end++;
 238         void *adr;
 239
 240         if (after_bootmem) {
 241                 adr = (void *)get_zeroed_page(GFP_ATOMIC);
 242                 *phys = __pa(adr);
 243
 244                 return adr;
 245         }
 246
 247         if (pfn >= end_pfn)
 248                 panic("alloc_low_page: ran out of memory");
 249
 250         adr = early_ioremap(pfn * PAGE_SIZE, PAGE_SIZE);
 251         memset(adr, 0, PAGE_SIZE);
 252         *phys  = pfn * PAGE_SIZE;
 253         return adr;
 254 }
 255
 256 static __meminit void unmap_low_page(void *adr)
 257 {
 258         if (after_bootmem)
 259                 return;
 260
 261         early_iounmap(adr, PAGE_SIZE);
 262 }
 263
 264 /* Must run before zap_low_mappings */
 265 __meminit void *early_ioremap(unsigned long addr, unsigned long size)
 266 {
 267         pmd_t *pmd, *last_pmd;
 268         unsigned long vaddr;
 269         int i, pmds;
 270
 271         pmds = ((addr & ~PMD_MASK) + size + ~PMD_MASK) / PMD_SIZE;
 272         vaddr = __START_KERNEL_map;
 273         pmd = level2_kernel_pgt;
 274         last_pmd = level2_kernel_pgt + PTRS_PER_PMD - 1;
 275
 276         for (; pmd <= last_pmd; pmd++, vaddr += PMD_SIZE) {
 277                 for (i = 0; i < pmds; i++) {
 278                         if (pmd_present(pmd[i]))
 279                                 goto continue_outer_loop;
 280                 }
 281                 vaddr += addr & ~PMD_MASK;
 282                 addr &= PMD_MASK;
 283
 284                 for (i = 0; i < pmds; i++, addr += PMD_SIZE)
 285                         set_pmd(pmd+i, __pmd(addr | __PAGE_KERNEL_LARGE_EXEC));
 286                 __flush_tlb_all();
 287
 288                 return (void *)vaddr;
 289 continue_outer_loop:
 290                 ;
 291         }
 292         printk(KERN_ERR "early_ioremap(0x%lx, %lu) failed\n", addr, size);
 293
 294         return NULL;
 295 }
 296
 297 /*
 298  * To avoid virtual aliases later:
 299  */
 300 __meminit void early_iounmap(void *addr, unsigned long size)
 301 {
 302         unsigned long vaddr;
 303         pmd_t *pmd;
 304         int i, pmds;
 305
 306         vaddr = (unsigned long)addr;
 307         pmds = ((vaddr & ~PMD_MASK) + size + ~PMD_MASK) / PMD_SIZE;
 308         pmd = level2_kernel_pgt + pmd_index(vaddr);
 309
 310         for (i = 0; i < pmds; i++)
 311                 pmd_clear(pmd + i);
 312
 313         __flush_tlb_all();
 314 }
 315
 316 static unsigned long __meminit
 317 phys_pmd_init(pmd_t *pmd_page, unsigned long address, unsigned long end)
 318 {
 319         int i = pmd_index(address);
 320
 321         for (; i < PTRS_PER_PMD; i++, address += PMD_SIZE) {
 322                 pmd_t *pmd = pmd_page + pmd_index(address);
 323
 324                 if (address >= end) {
 325                         if (!after_bootmem) {
 326                                 for (; i < PTRS_PER_PMD; i++, pmd++)
 327                                         set_pmd(pmd, __pmd(0));
 328                         }
 329                         break;
 330                 }
 331
 332                 if (pmd_val(*pmd))
 333                         continue;
 334
 335                 set_pte((pte_t *)pmd,
 336                         pfn_pte(address >> PAGE_SHIFT, PAGE_KERNEL_LARGE));
 337         }
 338         return address;
 339 }
 340
 341 static unsigned long __meminit
 342 phys_pmd_update(pud_t *pud, unsigned long address, unsigned long end)
 343 {
 344         pmd_t *pmd = pmd_offset(pud, 0);
 345         unsigned long last_map_addr;
 346
 347         spin_lock(&init_mm.page_table_lock);
 348         last_map_addr = phys_pmd_init(pmd, address, end);
 349         spin_unlock(&init_mm.page_table_lock);
 350         __flush_tlb_all();
 351         return last_map_addr;
 352 }
 353
 354 static unsigned long __meminit
 355 phys_pud_init(pud_t *pud_page, unsigned long addr, unsigned long end)
 356 {
 357         unsigned long last_map_addr = end;
 358         int i = pud_index(addr);
 359
 360         for (; i < PTRS_PER_PUD; i++, addr = (addr & PUD_MASK) + PUD_SIZE) {
 361                 unsigned long pmd_phys;
 362                 pud_t *pud = pud_page + pud_index(addr);
 363                 pmd_t *pmd;
 364
 365                 if (addr >= end)
 366                         break;
 367
 368                 if (!after_bootmem &&
 369                                 !e820_any_mapped(addr, addr+PUD_SIZE, 0)) {
 370                         set_pud(pud, __pud(0));
 371                         continue;
 372                 }
 373
 374                 if (pud_val(*pud)) {
 375                         if (!pud_large(*pud))
 376                                 last_map_addr = phys_pmd_update(pud, addr, end);
 377                         continue;
 378                 }
 379
 380                 if (direct_gbpages) {
 381                         set_pte((pte_t *)pud,
 382                                 pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL_LARGE));
 383                         last_map_addr = (addr & PUD_MASK) + PUD_SIZE;
 384                         continue;
 385                 }
 386
 387                 pmd = alloc_low_page(&pmd_phys);
 388
 389                 spin_lock(&init_mm.page_table_lock);
 390                 set_pud(pud, __pud(pmd_phys | _KERNPG_TABLE));
 391                 last_map_addr = phys_pmd_init(pmd, addr, end);
 392                 spin_unlock(&init_mm.page_table_lock);
 393
 394                 unmap_low_page(pmd);
 395         }
 396         __flush_tlb_all();
 397
 398         return last_map_addr >> PAGE_SHIFT;
 399 }
 400
 401 static void __init find_early_table_space(unsigned long end)
 402 {
 403         unsigned long puds, pmds, tables, start;
 404
 405         puds = (end + PUD_SIZE - 1) >> PUD_SHIFT;
 406         tables = round_up(puds * sizeof(pud_t), PAGE_SIZE);
 407         if (!direct_gbpages) {
 408                 pmds = (end + PMD_SIZE - 1) >> PMD_SHIFT;
 409                 tables += round_up(pmds * sizeof(pmd_t), PAGE_SIZE);
 410         }
 411
 412         /*
 413          * RED-PEN putting page tables only on node 0 could
 414          * cause a hotspot and fill up ZONE_DMA. The page tables
 415          * need roughly 0.5KB per GB.
 416          */
 417         start = 0x8000;
 418         table_start = find_e820_area(start, end, tables, PAGE_SIZE);
 419         if (table_start == -1UL)
 420                 panic("Cannot find space for the kernel page tables");
 421
 422         table_start >>= PAGE_SHIFT;
 423         table_end = table_start;
 424
 425         early_printk("kernel direct mapping tables up to %lx @ %lx-%lx\n",
 426                 end, table_start << PAGE_SHIFT,
 427                 (table_start << PAGE_SHIFT) + tables);
 428 }
 429
 430 static void __init init_gbpages(void)
 431 {
 432         if (direct_gbpages && cpu_has_gbpages)
 433                 printk(KERN_INFO "Using GB pages for direct mapping\n");
 434         else
 435                 direct_gbpages = 0;
 436 }
 437
 438 #ifdef CONFIG_MEMTEST_BOOTPARAM
 439
 440 static void __init memtest(unsigned long start_phys, unsigned long size,
 441                                  unsigned pattern)
 442 {
 443         unsigned long i;
 444         unsigned long *start;
 445         unsigned long start_bad;
 446         unsigned long last_bad;
 447         unsigned long val;
 448         unsigned long start_phys_aligned;
 449         unsigned long count;
 450         unsigned long incr;
 451
 452         switch (pattern) {
 453         case 0:
 454                 val = 0UL;
 455                 break;
 456         case 1:
 457                 val = -1UL;
 458                 break;
 459         case 2:
 460                 val = 0x5555555555555555UL;
 461                 break;
 462         case 3:
 463                 val = 0xaaaaaaaaaaaaaaaaUL;
 464                 break;
 465         default:
 466                 return;
 467         }
 468
 469         incr = sizeof(unsigned long);
 470         start_phys_aligned = ALIGN(start_phys, incr);
 471         count = (size - (start_phys_aligned - start_phys))/incr;
 472         start = __va(start_phys_aligned);
 473         start_bad = 0;
 474         last_bad = 0;
 475
 476         for (i = 0; i < count; i++)
 477                 start[i] = val;
 478         for (i = 0; i < count; i++, start++, start_phys_aligned += incr) {
 479                 if (*start != val) {
 480                         if (start_phys_aligned == last_bad + incr) {
 481                                 last_bad += incr;
 482                         } else {
 483                                 if (start_bad) {
 484                                         printk(KERN_CONT "\n  %016lx bad mem addr %016lx - %016lx reserved",
 485                                                 val, start_bad, last_bad + incr);
 486                                         reserve_early(start_bad, last_bad - start_bad, "BAD RAM");
 487                                 }
 488                                 start_bad = last_bad = start_phys_aligned;
 489                         }
 490                 }
 491         }
 492         if (start_bad) {
 493                 printk(KERN_CONT "\n  %016lx bad mem addr %016lx - %016lx reserved",
 494                         val, start_bad, last_bad + incr);
 495                 reserve_early(start_bad, last_bad - start_bad, "BAD RAM");
 496         }
 497
 498 }
 499
 500 static int memtest_pattern __initdata = CONFIG_MEMTEST_BOOTPARAM_VALUE;
 501
 502 static int __init parse_memtest(char *arg)
 503 {
 504         if (arg)
 505                 memtest_pattern = simple_strtoul(arg, NULL, 0);
 506         return 0;
 507 }
 508
 509 early_param("memtest", parse_memtest);
 510
 511 static void __init early_memtest(unsigned long start, unsigned long end)
 512 {
 513         unsigned long t_start, t_size;
 514         unsigned pattern;
 515
 516         if (!memtest_pattern)
 517                 return;
 518
 519         printk(KERN_INFO "early_memtest: pattern num %d", memtest_pattern);
 520         for (pattern = 0; pattern < memtest_pattern; pattern++) {
 521                 t_start = start;
 522                 t_size = 0;
 523                 while (t_start < end) {
 524                         t_start = find_e820_area_size(t_start, &t_size, 1);
 525
 526                         /* done ? */
 527                         if (t_start >= end)
 528                                 break;
 529                         if (t_start + t_size > end)
 530                                 t_size = end - t_start;
 531
 532                         printk(KERN_CONT "\n  %016lx - %016lx pattern %d",
 533                                 t_start, t_start + t_size, pattern);
 534
 535                         memtest(t_start, t_size, pattern);
 536
 537                         t_start += t_size;
 538                 }
 539         }
 540         printk(KERN_CONT "\n");
 541 }
 542 #else
 543 static void __init early_memtest(unsigned long start, unsigned long end)
 544 {
 545 }
 546 #endif
 547
 548 /*
 549  * Setup the direct mapping of the physical memory at PAGE_OFFSET.
 550  * This runs before bootmem is initialized and gets pages directly from
 551  * the physical memory. To access them they are temporarily mapped.
 552  */
 553 unsigned long __init_refok init_memory_mapping(unsigned long start, unsigned long end)
 554 {
 555         unsigned long next, last_map_addr = end;
 556         unsigned long start_phys = start, end_phys = end;
 557
 558         printk(KERN_INFO "init_memory_mapping\n");
 559
 560         /*
 561          * Find space for the kernel direct mapping tables.
 562          *
 563          * Later we should allocate these tables in the local node of the
 564          * memory mapped. Unfortunately this is done currently before the
 565          * nodes are discovered.
 566          */
 567         if (!after_bootmem) {
 568                 init_gbpages();
 569                 find_early_table_space(end);
 570         }
 571
 572         start = (unsigned long)__va(start);
 573         end = (unsigned long)__va(end);
 574
 575         for (; start < end; start = next) {
 576                 pgd_t *pgd = pgd_offset_k(start);
 577                 unsigned long pud_phys;
 578                 pud_t *pud;
 579
 580                 if (after_bootmem)
 581                         pud = pud_offset(pgd, start & PGDIR_MASK);
 582                 else
 583                         pud = alloc_low_page(&pud_phys);
 584
 585                 next = start + PGDIR_SIZE;
 586                 if (next > end)
 587                         next = end;
 588                 last_map_addr = phys_pud_init(pud, __pa(start), __pa(next));
 589                 if (!after_bootmem)
 590                         set_pgd(pgd_offset_k(start), mk_kernel_pgd(pud_phys));
 591                 unmap_low_page(pud);
 592         }
 593
 594         if (!after_bootmem)
 595                 mmu_cr4_features = read_cr4();
 596         __flush_tlb_all();
 597
 598         if (!after_bootmem)
 599                 reserve_early(table_start << PAGE_SHIFT,
 600                                  table_end << PAGE_SHIFT, "PGTABLE");
 601
 602         if (!after_bootmem)
 603                 early_memtest(start_phys, end_phys);
 604
 605         return last_map_addr;
 606 }
 607
 608 #ifndef CONFIG_NUMA
 609 void __init paging_init(void)
 610 {
 611         unsigned long max_zone_pfns[MAX_NR_ZONES];
 612
 613         memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
 614         max_zone_pfns[ZONE_DMA] = MAX_DMA_PFN;
 615         max_zone_pfns[ZONE_DMA32] = MAX_DMA32_PFN;
 616         max_zone_pfns[ZONE_NORMAL] = end_pfn;
 617
 618         memory_present(0, 0, end_pfn);
 619         sparse_init();
 620         free_area_init_nodes(max_zone_pfns);
 621 }
 622 #endif
 623
 624 /*
 625  * Memory hotplug specific functions
 626  */
 627 void online_page(struct page *page)
 628 {
 629         ClearPageReserved(page);
 630         init_page_count(page);
 631         __free_page(page);
 632         totalram_pages++;
 633         num_physpages++;
 634 }
 635
 636 #ifdef CONFIG_MEMORY_HOTPLUG
 637 /*
 638  * Memory is added always to NORMAL zone. This means you will never get
 639  * additional DMA/DMA32 memory.
 640  */
 641 int arch_add_memory(int nid, u64 start, u64 size)
 642 {
 643         struct pglist_data *pgdat = NODE_DATA(nid);
 644         struct zone *zone = pgdat->node_zones + ZONE_NORMAL;
 645         unsigned long last_mapped_pfn, start_pfn = start >> PAGE_SHIFT;
 646         unsigned long nr_pages = size >> PAGE_SHIFT;
 647         int ret;
 648
 649         last_mapped_pfn = init_memory_mapping(start, start + size-1);
 650         if (last_mapped_pfn > max_pfn_mapped)
 651                 max_pfn_mapped = last_mapped_pfn;
 652
 653         ret = __add_pages(zone, start_pfn, nr_pages);
 654         WARN_ON(1);
 655
 656         return ret;
 657 }
 658 EXPORT_SYMBOL_GPL(arch_add_memory);
 659
 660 #if !defined(CONFIG_ACPI_NUMA) && defined(CONFIG_NUMA)
 661 int memory_add_physaddr_to_nid(u64 start)
 662 {
 663         return 0;
 664 }
 665 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
 666 #endif
 667
 668 #endif /* CONFIG_MEMORY_HOTPLUG */
 669
 670 static struct kcore_list kcore_mem, kcore_vmalloc, kcore_kernel,
 671                          kcore_modules, kcore_vsyscall;
 672
 673 void __init mem_init(void)
 674 {
 675         long codesize, reservedpages, datasize, initsize;
 676
 677         pci_iommu_alloc();
 678
 679         /* clear_bss() already clear the empty_zero_page */
 680
 681         reservedpages = 0;
 682
 683         /* this will put all low memory onto the freelists */
 684 #ifdef CONFIG_NUMA
 685         totalram_pages = numa_free_all_bootmem();
 686 #else
 687         totalram_pages = free_all_bootmem();
 688 #endif
 689         reservedpages = end_pfn - totalram_pages -
 690                                         absent_pages_in_range(0, end_pfn);
 691         after_bootmem = 1;
 692
 693         codesize =  (unsigned long) &_etext - (unsigned long) &_text;
 694         datasize =  (unsigned long) &_edata - (unsigned long) &_etext;
 695         initsize =  (unsigned long) &__init_end - (unsigned long) &__init_begin;
 696
 697         /* Register memory areas for /proc/kcore */
 698         kclist_add(&kcore_mem, __va(0), max_low_pfn << PAGE_SHIFT);
 699         kclist_add(&kcore_vmalloc, (void *)VMALLOC_START,
 700                    VMALLOC_END-VMALLOC_START);
 701         kclist_add(&kcore_kernel, &_stext, _end - _stext);
 702         kclist_add(&kcore_modules, (void *)MODULES_VADDR, MODULES_LEN);
 703         kclist_add(&kcore_vsyscall, (void *)VSYSCALL_START,
 704                                  VSYSCALL_END - VSYSCALL_START);
 705
 706         printk(KERN_INFO "Memory: %luk/%luk available (%ldk kernel code, "
 707                                 "%ldk reserved, %ldk data, %ldk init)\n",
 708                 (unsigned long) nr_free_pages() << (PAGE_SHIFT-10),
 709                 end_pfn << (PAGE_SHIFT-10),
 710                 codesize >> 10,
 711                 reservedpages << (PAGE_SHIFT-10),
 712                 datasize >> 10,
 713                 initsize >> 10);
 714
 715         cpa_init();
 716 }
 717
 718 void free_init_pages(char *what, unsigned long begin, unsigned long end)
 719 {
 720         unsigned long addr = begin;
 721
 722         if (addr >= end)
 723                 return;
 724
 725         /*
 726          * If debugging page accesses then do not free this memory but
 727          * mark them not present - any buggy init-section access will
 728          * create a kernel page fault:
 729          */
 730 #ifdef CONFIG_DEBUG_PAGEALLOC
 731         printk(KERN_INFO "debug: unmapping init memory %08lx..%08lx\n",
 732                 begin, PAGE_ALIGN(end));
 733         set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
 734 #else
 735         printk(KERN_INFO "Freeing %s: %luk freed\n", what, (end - begin) >> 10);
 736
 737         for (; addr < end; addr += PAGE_SIZE) {
 738                 ClearPageReserved(virt_to_page(addr));
 739                 init_page_count(virt_to_page(addr));
 740                 memset((void *)(addr & ~(PAGE_SIZE-1)),
 741                         POISON_FREE_INITMEM, PAGE_SIZE);
 742                 free_page(addr);
 743                 totalram_pages++;
 744         }
 745 #endif
 746 }
 747
 748 void free_initmem(void)
 749 {
 750         free_init_pages("unused kernel memory",
 751                         (unsigned long)(&__init_begin),
 752                         (unsigned long)(&__init_end));
 753 }
 754
 755 #ifdef CONFIG_DEBUG_RODATA
 756 const int rodata_test_data = 0xC3;
 757 EXPORT_SYMBOL_GPL(rodata_test_data);
 758
 759 void mark_rodata_ro(void)
 760 {
 761         unsigned long start = PFN_ALIGN(_stext), end = PFN_ALIGN(__end_rodata);
 762
 763         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
 764                (end - start) >> 10);
 765         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
 766
 767         /*
 768          * The rodata section (but not the kernel text!) should also be
 769          * not-executable.
 770          */
 771         start = ((unsigned long)__start_rodata + PAGE_SIZE - 1) & PAGE_MASK;
 772         set_memory_nx(start, (end - start) >> PAGE_SHIFT);
 773
 774         rodata_test();
 775
 776 #ifdef CONFIG_CPA_DEBUG
 777         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
 778         set_memory_rw(start, (end-start) >> PAGE_SHIFT);
 779
 780         printk(KERN_INFO "Testing CPA: again\n");
 781         set_memory_ro(start, (end-start) >> PAGE_SHIFT);
 782 #endif
 783 }
 784
 785 #endif
 786
 787 #ifdef CONFIG_BLK_DEV_INITRD
 788 void free_initrd_mem(unsigned long start, unsigned long end)
 789 {
 790         free_init_pages("initrd memory", start, end);
 791 }
 792 #endif
 793
 794 void __init reserve_bootmem_generic(unsigned long phys, unsigned len)
 795 {
 796 #ifdef CONFIG_NUMA
 797         int nid = phys_to_nid(phys);
 798 #endif
 799         unsigned long pfn = phys >> PAGE_SHIFT;
 800
 801         if (pfn >= end_pfn) {
 802                 /*
 803                  * This can happen with kdump kernels when accessing
 804                  * firmware tables:
 805                  */
 806                 if (pfn < max_pfn_mapped)
 807                         return;
 808
 809                 printk(KERN_ERR "reserve_bootmem: illegal reserve %lx %u\n",
 810                                 phys, len);
 811                 return;
 812         }
 813
 814         /* Should check here against the e820 map to avoid double free */
 815 #ifdef CONFIG_NUMA
 816         reserve_bootmem_node(NODE_DATA(nid), phys, len, BOOTMEM_DEFAULT);
 817 #else
 818         reserve_bootmem(phys, len, BOOTMEM_DEFAULT);
 819 #endif
 820         if (phys+len <= MAX_DMA_PFN*PAGE_SIZE) {
 821                 dma_reserve += len / PAGE_SIZE;
 822                 set_dma_reserve(dma_reserve);
 823         }
 824 }
 825
 826 int kern_addr_valid(unsigned long addr)
 827 {
 828         unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
 829         pgd_t *pgd;
 830         pud_t *pud;
 831         pmd_t *pmd;
 832         pte_t *pte;
 833
 834         if (above != 0 && above != -1UL)
 835                 return 0;
 836
 837         pgd = pgd_offset_k(addr);
 838         if (pgd_none(*pgd))
 839                 return 0;
 840
 841         pud = pud_offset(pgd, addr);
 842         if (pud_none(*pud))
 843                 return 0;
 844
 845         pmd = pmd_offset(pud, addr);
 846         if (pmd_none(*pmd))
 847                 return 0;
 848
 849         if (pmd_large(*pmd))
 850                 return pfn_valid(pmd_pfn(*pmd));
 851
 852         pte = pte_offset_kernel(pmd, addr);
 853         if (pte_none(*pte))
 854                 return 0;
 855
 856         return pfn_valid(pte_pfn(*pte));
 857 }
 858
 859 /*
 860  * A pseudo VMA to allow ptrace access for the vsyscall page.  This only
 861  * covers the 64bit vsyscall page now. 32bit has a real VMA now and does
 862  * not need special handling anymore:
 863  */
 864 static struct vm_area_struct gate_vma = {
 865         .vm_start       = VSYSCALL_START,
 866         .vm_end         = VSYSCALL_START + (VSYSCALL_MAPPED_PAGES * PAGE_SIZE),
 867         .vm_page_prot   = PAGE_READONLY_EXEC,
 868         .vm_flags       = VM_READ | VM_EXEC
 869 };
 870
 871 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
 872 {
 873 #ifdef CONFIG_IA32_EMULATION
 874         if (test_tsk_thread_flag(tsk, TIF_IA32))
 875                 return NULL;
 876 #endif
 877         return &gate_vma;
 878 }
 879
 880 int in_gate_area(struct task_struct *task, unsigned long addr)
 881 {
 882         struct vm_area_struct *vma = get_gate_vma(task);
 883
 884         if (!vma)
 885                 return 0;
 886
 887         return (addr >= vma->vm_start) && (addr < vma->vm_end);
 888 }
 889
 890 /*
 891  * Use this when you have no reliable task/vma, typically from interrupt
 892  * context. It is less reliable than using the task's vma and may give
 893  * false positives:
 894  */
 895 int in_gate_area_no_task(unsigned long addr)
 896 {
 897         return (addr >= VSYSCALL_START) && (addr < VSYSCALL_END);
 898 }
 899
 900 const char *arch_vma_name(struct vm_area_struct *vma)
 901 {
 902         if (vma->vm_mm && vma->vm_start == (long)vma->vm_mm->context.vdso)
 903                 return "[vdso]";
 904         if (vma == &gate_vma)
 905                 return "[vsyscall]";
 906         return NULL;
 907 }
 908
 909 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 910 /*
 911  * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
 912  */
 913 int __meminit
 914 vmemmap_populate(struct page *start_page, unsigned long size, int node)
 915 {
 916         unsigned long addr = (unsigned long)start_page;
 917         unsigned long end = (unsigned long)(start_page + size);
 918         unsigned long next;
 919         pgd_t *pgd;
 920         pud_t *pud;
 921         pmd_t *pmd;
 922
 923         for (; addr < end; addr = next) {
 924                 next = pmd_addr_end(addr, end);
 925
 926                 pgd = vmemmap_pgd_populate(addr, node);
 927                 if (!pgd)
 928                         return -ENOMEM;
 929
 930                 pud = vmemmap_pud_populate(pgd, addr, node);
 931                 if (!pud)
 932                         return -ENOMEM;
 933
 934                 pmd = pmd_offset(pud, addr);
 935                 if (pmd_none(*pmd)) {
 936                         pte_t entry;
 937                         void *p;
 938
 939                         p = vmemmap_alloc_block(PMD_SIZE, node);
 940                         if (!p)
 941                                 return -ENOMEM;
 942
 943                         entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
 944                                                         PAGE_KERNEL_LARGE);
 945                         set_pmd(pmd, __pmd(pte_val(entry)));
 946
 947                         printk(KERN_DEBUG " [%lx-%lx] PMD ->%p on node %d\n",
 948                                 addr, addr + PMD_SIZE - 1, p, node);
 949                 } else {
 950                         vmemmap_verify((pte_t *)pmd, node, addr, next);
 951                 }
 952         }
 953         return 0;
 954 }
 955 #endif