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
  49 const struct dma_mapping_ops *dma_ops;
  50 EXPORT_SYMBOL(dma_ops);
  51
  52 static unsigned long dma_reserve __initdata;
  53
  54 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
  55
  56 /*
  57  * NOTE: pagetable_init alloc all the fixmap pagetables contiguous on the
  58  * physical space so we can cache the place of the first one and move
  59  * around without checking the pgd every time.
  60  */
  61
  62 void show_mem(void)
  63 {
  64         long i, total = 0, reserved = 0;
  65         long shared = 0, cached = 0;
  66         struct page *page;
  67         pg_data_t *pgdat;
  68
  69         printk(KERN_INFO "Mem-info:\n");
  70         show_free_areas();
  71         printk(KERN_INFO "Free swap:       %6ldkB\n",
  72                 nr_swap_pages << (PAGE_SHIFT-10));
  73
  74         for_each_online_pgdat(pgdat) {
  75                 for (i = 0; i < pgdat->node_spanned_pages; ++i) {
  76                         /*
  77                          * This loop can take a while with 256 GB and
  78                          * 4k pages so defer the NMI watchdog:
  79                          */
  80                         if (unlikely(i % MAX_ORDER_NR_PAGES == 0))
  81                                 touch_nmi_watchdog();
  82
  83                         if (!pfn_valid(pgdat->node_start_pfn + i))
  84                                 continue;
  85
  86                         page = pfn_to_page(pgdat->node_start_pfn + i);
  87                         total++;
  88                         if (PageReserved(page))
  89                                 reserved++;
  90                         else if (PageSwapCache(page))
  91                                 cached++;
  92                         else if (page_count(page))
  93                                 shared += page_count(page) - 1;
  94                 }
  95         }
  96         printk(KERN_INFO "%lu pages of RAM\n",          total);
  97         printk(KERN_INFO "%lu reserved pages\n",        reserved);
  98         printk(KERN_INFO "%lu pages shared\n",          shared);
  99         printk(KERN_INFO "%lu pages swap cached\n",     cached);
 100 }
 101
 102 int after_bootmem;
 103
 104 static __init void *spp_getpage(void)
 105 {
 106         void *ptr;
 107
 108         if (after_bootmem)
 109                 ptr = (void *) get_zeroed_page(GFP_ATOMIC);
 110         else
 111                 ptr = alloc_bootmem_pages(PAGE_SIZE);
 112
 113         if (!ptr || ((unsigned long)ptr & ~PAGE_MASK)) {
 114                 panic("set_pte_phys: cannot allocate page data %s\n",
 115                         after_bootmem ? "after bootmem" : "");
 116         }
 117
 118         pr_debug("spp_getpage %p\n", ptr);
 119
 120         return ptr;
 121 }
 122
 123 static __init void
 124 set_pte_phys(unsigned long vaddr, unsigned long phys, pgprot_t prot)
 125 {
 126         pgd_t *pgd;
 127         pud_t *pud;
 128         pmd_t *pmd;
 129         pte_t *pte, new_pte;
 130
 131         pr_debug("set_pte_phys %lx to %lx\n", vaddr, phys);
 132
 133         pgd = pgd_offset_k(vaddr);
 134         if (pgd_none(*pgd)) {
 135                 printk(KERN_ERR
 136                         "PGD FIXMAP MISSING, it should be setup in head.S!\n");
 137                 return;
 138         }
 139         pud = pud_offset(pgd, vaddr);
 140         if (pud_none(*pud)) {
 141                 pmd = (pmd_t *) spp_getpage();
 142                 set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE | _PAGE_USER));
 143                 if (pmd != pmd_offset(pud, 0)) {
 144                         printk(KERN_ERR "PAGETABLE BUG #01! %p <-> %p\n",
 145                                 pmd, pmd_offset(pud, 0));
 146                         return;
 147                 }
 148         }
 149         pmd = pmd_offset(pud, vaddr);
 150         if (pmd_none(*pmd)) {
 151                 pte = (pte_t *) spp_getpage();
 152                 set_pmd(pmd, __pmd(__pa(pte) | _KERNPG_TABLE | _PAGE_USER));
 153                 if (pte != pte_offset_kernel(pmd, 0)) {
 154                         printk(KERN_ERR "PAGETABLE BUG #02!\n");
 155                         return;
 156                 }
 157         }
 158         new_pte = pfn_pte(phys >> PAGE_SHIFT, prot);
 159
 160         pte = pte_offset_kernel(pmd, vaddr);
 161         if (!pte_none(*pte) &&
 162             pte_val(*pte) != (pte_val(new_pte) & __supported_pte_mask))
 163                 pte_ERROR(*pte);
 164         set_pte(pte, new_pte);
 165
 166         /*
 167          * It's enough to flush this one mapping.
 168          * (PGE mappings get flushed as well)
 169          */
 170         __flush_tlb_one(vaddr);
 171 }
 172
 173 /* NOTE: this is meant to be run only at boot */
 174 void __init
 175 __set_fixmap(enum fixed_addresses idx, unsigned long phys, pgprot_t prot)
 176 {
 177         unsigned long address = __fix_to_virt(idx);
 178
 179         if (idx >= __end_of_fixed_addresses) {
 180                 printk(KERN_ERR "Invalid __set_fixmap\n");
 181                 return;
 182         }
 183         set_pte_phys(address, phys, prot);
 184 }
 185
 186 static unsigned long __initdata table_start;
 187 static unsigned long __meminitdata table_end;
 188
 189 static __meminit void *alloc_low_page(unsigned long *phys)
 190 {
 191         unsigned long pfn = table_end++;
 192         void *adr;
 193
 194         if (after_bootmem) {
 195                 adr = (void *)get_zeroed_page(GFP_ATOMIC);
 196                 *phys = __pa(adr);
 197
 198                 return adr;
 199         }
 200
 201         if (pfn >= end_pfn)
 202                 panic("alloc_low_page: ran out of memory");
 203
 204         adr = early_ioremap(pfn * PAGE_SIZE, PAGE_SIZE);
 205         memset(adr, 0, PAGE_SIZE);
 206         *phys  = pfn * PAGE_SIZE;
 207         return adr;
 208 }
 209
 210 static __meminit void unmap_low_page(void *adr)
 211 {
 212         if (after_bootmem)
 213                 return;
 214
 215         early_iounmap(adr, PAGE_SIZE);
 216 }
 217
 218 /* Must run before zap_low_mappings */
 219 __meminit void *early_ioremap(unsigned long addr, unsigned long size)
 220 {
 221         pmd_t *pmd, *last_pmd;
 222         unsigned long vaddr;
 223         int i, pmds;
 224
 225         pmds = ((addr & ~PMD_MASK) + size + ~PMD_MASK) / PMD_SIZE;
 226         vaddr = __START_KERNEL_map;
 227         pmd = level2_kernel_pgt;
 228         last_pmd = level2_kernel_pgt + PTRS_PER_PMD - 1;
 229
 230         for (; pmd <= last_pmd; pmd++, vaddr += PMD_SIZE) {
 231                 for (i = 0; i < pmds; i++) {
 232                         if (pmd_present(pmd[i]))
 233                                 goto continue_outer_loop;
 234                 }
 235                 vaddr += addr & ~PMD_MASK;
 236                 addr &= PMD_MASK;
 237
 238                 for (i = 0; i < pmds; i++, addr += PMD_SIZE)
 239                         set_pmd(pmd+i, __pmd(addr | __PAGE_KERNEL_LARGE_EXEC));
 240                 __flush_tlb_all();
 241
 242                 return (void *)vaddr;
 243 continue_outer_loop:
 244                 ;
 245         }
 246         printk(KERN_ERR "early_ioremap(0x%lx, %lu) failed\n", addr, size);
 247
 248         return NULL;
 249 }
 250
 251 /*
 252  * To avoid virtual aliases later:
 253  */
 254 __meminit void early_iounmap(void *addr, unsigned long size)
 255 {
 256         unsigned long vaddr;
 257         pmd_t *pmd;
 258         int i, pmds;
 259
 260         vaddr = (unsigned long)addr;
 261         pmds = ((vaddr & ~PMD_MASK) + size + ~PMD_MASK) / PMD_SIZE;
 262         pmd = level2_kernel_pgt + pmd_index(vaddr);
 263
 264         for (i = 0; i < pmds; i++)
 265                 pmd_clear(pmd + i);
 266
 267         __flush_tlb_all();
 268 }
 269
 270 static void __meminit
 271 phys_pmd_init(pmd_t *pmd_page, unsigned long address, unsigned long end)
 272 {
 273         int i = pmd_index(address);
 274
 275         for (; i < PTRS_PER_PMD; i++, address += PMD_SIZE) {
 276                 pmd_t *pmd = pmd_page + pmd_index(address);
 277
 278                 if (address >= end) {
 279                         if (!after_bootmem) {
 280                                 for (; i < PTRS_PER_PMD; i++, pmd++)
 281                                         set_pmd(pmd, __pmd(0));
 282                         }
 283                         break;
 284                 }
 285
 286                 if (pmd_val(*pmd))
 287                         continue;
 288
 289                 set_pte((pte_t *)pmd,
 290                         pfn_pte(address >> PAGE_SHIFT, PAGE_KERNEL_LARGE));
 291         }
 292 }
 293
 294 static void __meminit
 295 phys_pmd_update(pud_t *pud, unsigned long address, unsigned long end)
 296 {
 297         pmd_t *pmd = pmd_offset(pud, 0);
 298         spin_lock(&init_mm.page_table_lock);
 299         phys_pmd_init(pmd, address, end);
 300         spin_unlock(&init_mm.page_table_lock);
 301         __flush_tlb_all();
 302 }
 303
 304 static void __meminit
 305 phys_pud_init(pud_t *pud_page, unsigned long addr, unsigned long end)
 306 {
 307         int i = pud_index(addr);
 308
 309         for (; i < PTRS_PER_PUD; i++, addr = (addr & PUD_MASK) + PUD_SIZE) {
 310                 unsigned long pmd_phys;
 311                 pud_t *pud = pud_page + pud_index(addr);
 312                 pmd_t *pmd;
 313
 314                 if (addr >= end)
 315                         break;
 316
 317                 if (!after_bootmem &&
 318                                 !e820_any_mapped(addr, addr+PUD_SIZE, 0)) {
 319                         set_pud(pud, __pud(0));
 320                         continue;
 321                 }
 322
 323                 if (pud_val(*pud)) {
 324                         phys_pmd_update(pud, addr, end);
 325                         continue;
 326                 }
 327
 328                 pmd = alloc_low_page(&pmd_phys);
 329
 330                 spin_lock(&init_mm.page_table_lock);
 331                 set_pud(pud, __pud(pmd_phys | _KERNPG_TABLE));
 332                 phys_pmd_init(pmd, addr, end);
 333                 spin_unlock(&init_mm.page_table_lock);
 334
 335                 unmap_low_page(pmd);
 336         }
 337         __flush_tlb_all();
 338 }
 339
 340 static void __init find_early_table_space(unsigned long end)
 341 {
 342         unsigned long puds, pmds, tables, start;
 343
 344         puds = (end + PUD_SIZE - 1) >> PUD_SHIFT;
 345         pmds = (end + PMD_SIZE - 1) >> PMD_SHIFT;
 346         tables = round_up(puds * sizeof(pud_t), PAGE_SIZE) +
 347                  round_up(pmds * sizeof(pmd_t), PAGE_SIZE);
 348
 349         /*
 350          * RED-PEN putting page tables only on node 0 could
 351          * cause a hotspot and fill up ZONE_DMA. The page tables
 352          * need roughly 0.5KB per GB.
 353          */
 354         start = 0x8000;
 355         table_start = find_e820_area(start, end, tables, PAGE_SIZE);
 356         if (table_start == -1UL)
 357                 panic("Cannot find space for the kernel page tables");
 358
 359         table_start >>= PAGE_SHIFT;
 360         table_end = table_start;
 361
 362         early_printk("kernel direct mapping tables up to %lx @ %lx-%lx\n",
 363                 end, table_start << PAGE_SHIFT,
 364                 (table_start << PAGE_SHIFT) + tables);
 365 }
 366
 367 /*
 368  * Setup the direct mapping of the physical memory at PAGE_OFFSET.
 369  * This runs before bootmem is initialized and gets pages directly from
 370  * the physical memory. To access them they are temporarily mapped.
 371  */
 372 void __init_refok init_memory_mapping(unsigned long start, unsigned long end)
 373 {
 374         unsigned long next;
 375
 376         pr_debug("init_memory_mapping\n");
 377
 378         /*
 379          * Find space for the kernel direct mapping tables.
 380          *
 381          * Later we should allocate these tables in the local node of the
 382          * memory mapped. Unfortunately this is done currently before the
 383          * nodes are discovered.
 384          */
 385         if (!after_bootmem)
 386                 find_early_table_space(end);
 387
 388         start = (unsigned long)__va(start);
 389         end = (unsigned long)__va(end);
 390
 391         for (; start < end; start = next) {
 392                 pgd_t *pgd = pgd_offset_k(start);
 393                 unsigned long pud_phys;
 394                 pud_t *pud;
 395
 396                 if (after_bootmem)
 397                         pud = pud_offset(pgd, start & PGDIR_MASK);
 398                 else
 399                         pud = alloc_low_page(&pud_phys);
 400
 401                 next = start + PGDIR_SIZE;
 402                 if (next > end)
 403                         next = end;
 404                 phys_pud_init(pud, __pa(start), __pa(next));
 405                 if (!after_bootmem)
 406                         set_pgd(pgd_offset_k(start), mk_kernel_pgd(pud_phys));
 407                 unmap_low_page(pud);
 408         }
 409
 410         if (!after_bootmem)
 411                 mmu_cr4_features = read_cr4();
 412         __flush_tlb_all();
 413
 414         if (!after_bootmem)
 415                 reserve_early(table_start << PAGE_SHIFT,
 416                                  table_end << PAGE_SHIFT, "PGTABLE");
 417 }
 418
 419 #ifndef CONFIG_NUMA
 420 void __init paging_init(void)
 421 {
 422         unsigned long max_zone_pfns[MAX_NR_ZONES];
 423
 424         memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
 425         max_zone_pfns[ZONE_DMA] = MAX_DMA_PFN;
 426         max_zone_pfns[ZONE_DMA32] = MAX_DMA32_PFN;
 427         max_zone_pfns[ZONE_NORMAL] = end_pfn;
 428
 429         memory_present(0, 0, end_pfn);
 430         sparse_init();
 431         free_area_init_nodes(max_zone_pfns);
 432 }
 433 #endif
 434
 435 /*
 436  * Memory hotplug specific functions
 437  */
 438 void online_page(struct page *page)
 439 {
 440         ClearPageReserved(page);
 441         init_page_count(page);
 442         __free_page(page);
 443         totalram_pages++;
 444         num_physpages++;
 445 }
 446
 447 #ifdef CONFIG_MEMORY_HOTPLUG
 448 /*
 449  * Memory is added always to NORMAL zone. This means you will never get
 450  * additional DMA/DMA32 memory.
 451  */
 452 int arch_add_memory(int nid, u64 start, u64 size)
 453 {
 454         struct pglist_data *pgdat = NODE_DATA(nid);
 455         struct zone *zone = pgdat->node_zones + ZONE_NORMAL;
 456         unsigned long start_pfn = start >> PAGE_SHIFT;
 457         unsigned long nr_pages = size >> PAGE_SHIFT;
 458         int ret;
 459
 460         init_memory_mapping(start, start + size-1);
 461
 462         ret = __add_pages(zone, start_pfn, nr_pages);
 463         WARN_ON(1);
 464
 465         return ret;
 466 }
 467 EXPORT_SYMBOL_GPL(arch_add_memory);
 468
 469 #if !defined(CONFIG_ACPI_NUMA) && defined(CONFIG_NUMA)
 470 int memory_add_physaddr_to_nid(u64 start)
 471 {
 472         return 0;
 473 }
 474 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
 475 #endif
 476
 477 #endif /* CONFIG_MEMORY_HOTPLUG */
 478
 479 static struct kcore_list kcore_mem, kcore_vmalloc, kcore_kernel,
 480                          kcore_modules, kcore_vsyscall;
 481
 482 void __init mem_init(void)
 483 {
 484         long codesize, reservedpages, datasize, initsize;
 485
 486         pci_iommu_alloc();
 487
 488         /* clear_bss() already clear the empty_zero_page */
 489
 490         /* temporary debugging - double check it's true: */
 491         {
 492                 int i;
 493
 494                 for (i = 0; i < 1024; i++)
 495                         WARN_ON_ONCE(empty_zero_page[i]);
 496         }
 497
 498         reservedpages = 0;
 499
 500         /* this will put all low memory onto the freelists */
 501 #ifdef CONFIG_NUMA
 502         totalram_pages = numa_free_all_bootmem();
 503 #else
 504         totalram_pages = free_all_bootmem();
 505 #endif
 506         reservedpages = end_pfn - totalram_pages -
 507                                         absent_pages_in_range(0, end_pfn);
 508         after_bootmem = 1;
 509
 510         codesize =  (unsigned long) &_etext - (unsigned long) &_text;
 511         datasize =  (unsigned long) &_edata - (unsigned long) &_etext;
 512         initsize =  (unsigned long) &__init_end - (unsigned long) &__init_begin;
 513
 514         /* Register memory areas for /proc/kcore */
 515         kclist_add(&kcore_mem, __va(0), max_low_pfn << PAGE_SHIFT);
 516         kclist_add(&kcore_vmalloc, (void *)VMALLOC_START,
 517                    VMALLOC_END-VMALLOC_START);
 518         kclist_add(&kcore_kernel, &_stext, _end - _stext);
 519         kclist_add(&kcore_modules, (void *)MODULES_VADDR, MODULES_LEN);
 520         kclist_add(&kcore_vsyscall, (void *)VSYSCALL_START,
 521                                  VSYSCALL_END - VSYSCALL_START);
 522
 523         printk(KERN_INFO "Memory: %luk/%luk available (%ldk kernel code, "
 524                                 "%ldk reserved, %ldk data, %ldk init)\n",
 525                 (unsigned long) nr_free_pages() << (PAGE_SHIFT-10),
 526                 end_pfn << (PAGE_SHIFT-10),
 527                 codesize >> 10,
 528                 reservedpages << (PAGE_SHIFT-10),
 529                 datasize >> 10,
 530                 initsize >> 10);
 531 }
 532
 533 void free_init_pages(char *what, unsigned long begin, unsigned long end)
 534 {
 535         unsigned long addr;
 536
 537         if (begin >= end)
 538                 return;
 539
 540         /*
 541          * If debugging page accesses then do not free this memory but
 542          * mark them not present - any buggy init-section access will
 543          * create a kernel page fault:
 544          */
 545 #ifdef CONFIG_DEBUG_PAGEALLOC
 546         printk(KERN_INFO "debug: unmapping init memory %08lx..%08lx\n",
 547                 begin, PAGE_ALIGN(end));
 548         set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
 549 #else
 550         printk(KERN_INFO "Freeing %s: %luk freed\n", what, (end - begin) >> 10);
 551
 552         for (addr = begin; addr < end; addr += PAGE_SIZE) {
 553                 ClearPageReserved(virt_to_page(addr));
 554                 init_page_count(virt_to_page(addr));
 555                 memset((void *)(addr & ~(PAGE_SIZE-1)),
 556                         POISON_FREE_INITMEM, PAGE_SIZE);
 557                 free_page(addr);
 558                 totalram_pages++;
 559         }
 560 #endif
 561 }
 562
 563 void free_initmem(void)
 564 {
 565         free_init_pages("unused kernel memory",
 566                         (unsigned long)(&__init_begin),
 567                         (unsigned long)(&__init_end));
 568 }
 569
 570 #ifdef CONFIG_DEBUG_RODATA
 571 const int rodata_test_data = 0xC3;
 572 EXPORT_SYMBOL_GPL(rodata_test_data);
 573
 574 void mark_rodata_ro(void)
 575 {
 576         unsigned long start = (unsigned long)_stext, end;
 577
 578 #ifdef CONFIG_HOTPLUG_CPU
 579         /* It must still be possible to apply SMP alternatives. */
 580         if (num_possible_cpus() > 1)
 581                 start = (unsigned long)_etext;
 582 #endif
 583
 584 #ifdef CONFIG_KPROBES
 585         start = (unsigned long)__start_rodata;
 586 #endif
 587
 588         end = (unsigned long)__end_rodata;
 589         start = (start + PAGE_SIZE - 1) & PAGE_MASK;
 590         end &= PAGE_MASK;
 591         if (end <= start)
 592                 return;
 593
 594
 595         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
 596                (end - start) >> 10);
 597         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
 598
 599         /*
 600          * The rodata section (but not the kernel text!) should also be
 601          * not-executable.
 602          */
 603         start = ((unsigned long)__start_rodata + PAGE_SIZE - 1) & PAGE_MASK;
 604         set_memory_nx(start, (end - start) >> PAGE_SHIFT);
 605
 606         rodata_test();
 607
 608 #ifdef CONFIG_CPA_DEBUG
 609         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
 610         set_memory_rw(start, (end-start) >> PAGE_SHIFT);
 611
 612         printk(KERN_INFO "Testing CPA: again\n");
 613         set_memory_ro(start, (end-start) >> PAGE_SHIFT);
 614 #endif
 615 }
 616 #endif
 617
 618 #ifdef CONFIG_BLK_DEV_INITRD
 619 void free_initrd_mem(unsigned long start, unsigned long end)
 620 {
 621         free_init_pages("initrd memory", start, end);
 622 }
 623 #endif
 624
 625 void __init reserve_bootmem_generic(unsigned long phys, unsigned len)
 626 {
 627 #ifdef CONFIG_NUMA
 628         int nid = phys_to_nid(phys);
 629 #endif
 630         unsigned long pfn = phys >> PAGE_SHIFT;
 631
 632         if (pfn >= end_pfn) {
 633                 /*
 634                  * This can happen with kdump kernels when accessing
 635                  * firmware tables:
 636                  */
 637                 if (pfn < end_pfn_map)
 638                         return;
 639
 640                 printk(KERN_ERR "reserve_bootmem: illegal reserve %lx %u\n",
 641                                 phys, len);
 642                 return;
 643         }
 644
 645         /* Should check here against the e820 map to avoid double free */
 646 #ifdef CONFIG_NUMA
 647         reserve_bootmem_node(NODE_DATA(nid), phys, len);
 648 #else
 649         reserve_bootmem(phys, len);
 650 #endif
 651         if (phys+len <= MAX_DMA_PFN*PAGE_SIZE) {
 652                 dma_reserve += len / PAGE_SIZE;
 653                 set_dma_reserve(dma_reserve);
 654         }
 655 }
 656
 657 int kern_addr_valid(unsigned long addr)
 658 {
 659         unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
 660         pgd_t *pgd;
 661         pud_t *pud;
 662         pmd_t *pmd;
 663         pte_t *pte;
 664
 665         if (above != 0 && above != -1UL)
 666                 return 0;
 667
 668         pgd = pgd_offset_k(addr);
 669         if (pgd_none(*pgd))
 670                 return 0;
 671
 672         pud = pud_offset(pgd, addr);
 673         if (pud_none(*pud))
 674                 return 0;
 675
 676         pmd = pmd_offset(pud, addr);
 677         if (pmd_none(*pmd))
 678                 return 0;
 679
 680         if (pmd_large(*pmd))
 681                 return pfn_valid(pmd_pfn(*pmd));
 682
 683         pte = pte_offset_kernel(pmd, addr);
 684         if (pte_none(*pte))
 685                 return 0;
 686
 687         return pfn_valid(pte_pfn(*pte));
 688 }
 689
 690 /*
 691  * A pseudo VMA to allow ptrace access for the vsyscall page.  This only
 692  * covers the 64bit vsyscall page now. 32bit has a real VMA now and does
 693  * not need special handling anymore:
 694  */
 695 static struct vm_area_struct gate_vma = {
 696         .vm_start       = VSYSCALL_START,
 697         .vm_end         = VSYSCALL_START + (VSYSCALL_MAPPED_PAGES * PAGE_SIZE),
 698         .vm_page_prot   = PAGE_READONLY_EXEC,
 699         .vm_flags       = VM_READ | VM_EXEC
 700 };
 701
 702 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
 703 {
 704 #ifdef CONFIG_IA32_EMULATION
 705         if (test_tsk_thread_flag(tsk, TIF_IA32))
 706                 return NULL;
 707 #endif
 708         return &gate_vma;
 709 }
 710
 711 int in_gate_area(struct task_struct *task, unsigned long addr)
 712 {
 713         struct vm_area_struct *vma = get_gate_vma(task);
 714
 715         if (!vma)
 716                 return 0;
 717
 718         return (addr >= vma->vm_start) && (addr < vma->vm_end);
 719 }
 720
 721 /*
 722  * Use this when you have no reliable task/vma, typically from interrupt
 723  * context. It is less reliable than using the task's vma and may give
 724  * false positives:
 725  */
 726 int in_gate_area_no_task(unsigned long addr)
 727 {
 728         return (addr >= VSYSCALL_START) && (addr < VSYSCALL_END);
 729 }
 730
 731 const char *arch_vma_name(struct vm_area_struct *vma)
 732 {
 733         if (vma->vm_mm && vma->vm_start == (long)vma->vm_mm->context.vdso)
 734                 return "[vdso]";
 735         if (vma == &gate_vma)
 736                 return "[vsyscall]";
 737         return NULL;
 738 }
 739
 740 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 741 /*
 742  * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
 743  */
 744 int __meminit
 745 vmemmap_populate(struct page *start_page, unsigned long size, int node)
 746 {
 747         unsigned long addr = (unsigned long)start_page;
 748         unsigned long end = (unsigned long)(start_page + size);
 749         unsigned long next;
 750         pgd_t *pgd;
 751         pud_t *pud;
 752         pmd_t *pmd;
 753
 754         for (; addr < end; addr = next) {
 755                 next = pmd_addr_end(addr, end);
 756
 757                 pgd = vmemmap_pgd_populate(addr, node);
 758                 if (!pgd)
 759                         return -ENOMEM;
 760
 761                 pud = vmemmap_pud_populate(pgd, addr, node);
 762                 if (!pud)
 763                         return -ENOMEM;
 764
 765                 pmd = pmd_offset(pud, addr);
 766                 if (pmd_none(*pmd)) {
 767                         pte_t entry;
 768                         void *p;
 769
 770                         p = vmemmap_alloc_block(PMD_SIZE, node);
 771                         if (!p)
 772                                 return -ENOMEM;
 773
 774                         entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
 775                                                         PAGE_KERNEL_LARGE);
 776                         set_pmd(pmd, __pmd(pte_val(entry)));
 777
 778                         printk(KERN_DEBUG " [%lx-%lx] PMD ->%p on node %d\n",
 779                                 addr, addr + PMD_SIZE - 1, p, node);
 780                 } else {
 781                         vmemmap_verify((pte_t *)pmd, node, addr, next);
 782                 }
 783         }
 784         return 0;
 785 }
 786 #endif