arch/ia64/mm/init.c

   1 /*
   2  * Initialize MMU support.
   3  *
   4  * Copyright (C) 1998-2003 Hewlett-Packard Co
   5  *      David Mosberger-Tang <davidm@hpl.hp.com>
   6  */
   7 #include <linux/kernel.h>
   8 #include <linux/init.h>
   9
  10 #include <linux/bootmem.h>
  11 #include <linux/efi.h>
  12 #include <linux/elf.h>
  13 #include <linux/mm.h>
  14 #include <linux/mmzone.h>
  15 #include <linux/module.h>
  16 #include <linux/personality.h>
  17 #include <linux/reboot.h>
  18 #include <linux/slab.h>
  19 #include <linux/swap.h>
  20 #include <linux/proc_fs.h>
  21 #include <linux/bitops.h>
  22 #include <linux/kexec.h>
  23
  24 #include <asm/a.out.h>
  25 #include <asm/dma.h>
  26 #include <asm/ia32.h>
  27 #include <asm/io.h>
  28 #include <asm/machvec.h>
  29 #include <asm/numa.h>
  30 #include <asm/patch.h>
  31 #include <asm/pgalloc.h>
  32 #include <asm/sal.h>
  33 #include <asm/sections.h>
  34 #include <asm/system.h>
  35 #include <asm/tlb.h>
  36 #include <asm/uaccess.h>
  37 #include <asm/unistd.h>
  38 #include <asm/mca.h>
  39
  40 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
  41
  42 DEFINE_PER_CPU(unsigned long *, __pgtable_quicklist);
  43 DEFINE_PER_CPU(long, __pgtable_quicklist_size);
  44
  45 extern void ia64_tlb_init (void);
  46
  47 unsigned long MAX_DMA_ADDRESS = PAGE_OFFSET + 0x100000000UL;
  48
  49 #ifdef CONFIG_VIRTUAL_MEM_MAP
  50 unsigned long vmalloc_end = VMALLOC_END_INIT;
  51 EXPORT_SYMBOL(vmalloc_end);
  52 struct page *vmem_map;
  53 EXPORT_SYMBOL(vmem_map);
  54 #endif
  55
  56 struct page *zero_page_memmap_ptr;      /* map entry for zero page */
  57 EXPORT_SYMBOL(zero_page_memmap_ptr);
  58
  59 #define MIN_PGT_PAGES                   25UL
  60 #define MAX_PGT_FREES_PER_PASS          16L
  61 #define PGT_FRACTION_OF_NODE_MEM        16
  62
  63 static inline long
  64 max_pgt_pages(void)
  65 {
  66         u64 node_free_pages, max_pgt_pages;
  67
  68 #ifndef CONFIG_NUMA
  69         node_free_pages = nr_free_pages();
  70 #else
  71         node_free_pages = node_page_state(numa_node_id(), NR_FREE_PAGES);
  72 #endif
  73         max_pgt_pages = node_free_pages / PGT_FRACTION_OF_NODE_MEM;
  74         max_pgt_pages = max(max_pgt_pages, MIN_PGT_PAGES);
  75         return max_pgt_pages;
  76 }
  77
  78 static inline long
  79 min_pages_to_free(void)
  80 {
  81         long pages_to_free;
  82
  83         pages_to_free = pgtable_quicklist_size - max_pgt_pages();
  84         pages_to_free = min(pages_to_free, MAX_PGT_FREES_PER_PASS);
  85         return pages_to_free;
  86 }
  87
  88 void
  89 check_pgt_cache(void)
  90 {
  91         long pages_to_free;
  92
  93         if (unlikely(pgtable_quicklist_size <= MIN_PGT_PAGES))
  94                 return;
  95
  96         preempt_disable();
  97         while (unlikely((pages_to_free = min_pages_to_free()) > 0)) {
  98                 while (pages_to_free--) {
  99                         free_page((unsigned long)pgtable_quicklist_alloc());
 100                 }
 101                 preempt_enable();
 102                 preempt_disable();
 103         }
 104         preempt_enable();
 105 }
 106
 107 void
 108 lazy_mmu_prot_update (pte_t pte)
 109 {
 110         unsigned long addr;
 111         struct page *page;
 112         unsigned long order;
 113
 114         if (!pte_exec(pte))
 115                 return;                         /* not an executable page... */
 116
 117         page = pte_page(pte);
 118         addr = (unsigned long) page_address(page);
 119
 120         if (test_bit(PG_arch_1, &page->flags))
 121                 return;                         /* i-cache is already coherent with d-cache */
 122
 123         if (PageCompound(page)) {
 124                 order = compound_order(page);
 125                 flush_icache_range(addr, addr + (1UL << order << PAGE_SHIFT));
 126         }
 127         else
 128                 flush_icache_range(addr, addr + PAGE_SIZE);
 129         set_bit(PG_arch_1, &page->flags);       /* mark page as clean */
 130 }
 131
 132 /*
 133  * Since DMA is i-cache coherent, any (complete) pages that were written via
 134  * DMA can be marked as "clean" so that lazy_mmu_prot_update() doesn't have to
 135  * flush them when they get mapped into an executable vm-area.
 136  */
 137 void
 138 dma_mark_clean(void *addr, size_t size)
 139 {
 140         unsigned long pg_addr, end;
 141
 142         pg_addr = PAGE_ALIGN((unsigned long) addr);
 143         end = (unsigned long) addr + size;
 144         while (pg_addr + PAGE_SIZE <= end) {
 145                 struct page *page = virt_to_page(pg_addr);
 146                 set_bit(PG_arch_1, &page->flags);
 147                 pg_addr += PAGE_SIZE;
 148         }
 149 }
 150
 151 inline void
 152 ia64_set_rbs_bot (void)
 153 {
 154         unsigned long stack_size = current->signal->rlim[RLIMIT_STACK].rlim_max & -16;
 155
 156         if (stack_size > MAX_USER_STACK_SIZE)
 157                 stack_size = MAX_USER_STACK_SIZE;
 158         current->thread.rbs_bot = PAGE_ALIGN(current->mm->start_stack - stack_size);
 159 }
 160
 161 /*
 162  * This performs some platform-dependent address space initialization.
 163  * On IA-64, we want to setup the VM area for the register backing
 164  * store (which grows upwards) and install the gateway page which is
 165  * used for signal trampolines, etc.
 166  */
 167 void
 168 ia64_init_addr_space (void)
 169 {
 170         struct vm_area_struct *vma;
 171
 172         ia64_set_rbs_bot();
 173
 174         /*
 175          * If we're out of memory and kmem_cache_alloc() returns NULL, we simply ignore
 176          * the problem.  When the process attempts to write to the register backing store
 177          * for the first time, it will get a SEGFAULT in this case.
 178          */
 179         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
 180         if (vma) {
 181                 vma->vm_mm = current->mm;
 182                 vma->vm_start = current->thread.rbs_bot & PAGE_MASK;
 183                 vma->vm_end = vma->vm_start + PAGE_SIZE;
 184                 vma->vm_page_prot = protection_map[VM_DATA_DEFAULT_FLAGS & 0x7];
 185                 vma->vm_flags = VM_DATA_DEFAULT_FLAGS|VM_GROWSUP|VM_ACCOUNT;
 186                 down_write(&current->mm->mmap_sem);
 187                 if (insert_vm_struct(current->mm, vma)) {
 188                         up_write(&current->mm->mmap_sem);
 189                         kmem_cache_free(vm_area_cachep, vma);
 190                         return;
 191                 }
 192                 up_write(&current->mm->mmap_sem);
 193         }
 194
 195         /* map NaT-page at address zero to speed up speculative dereferencing of NULL: */
 196         if (!(current->personality & MMAP_PAGE_ZERO)) {
 197                 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
 198                 if (vma) {
 199                         vma->vm_mm = current->mm;
 200                         vma->vm_end = PAGE_SIZE;
 201                         vma->vm_page_prot = __pgprot(pgprot_val(PAGE_READONLY) | _PAGE_MA_NAT);
 202                         vma->vm_flags = VM_READ | VM_MAYREAD | VM_IO | VM_RESERVED;
 203                         down_write(&current->mm->mmap_sem);
 204                         if (insert_vm_struct(current->mm, vma)) {
 205                                 up_write(&current->mm->mmap_sem);
 206                                 kmem_cache_free(vm_area_cachep, vma);
 207                                 return;
 208                         }
 209                         up_write(&current->mm->mmap_sem);
 210                 }
 211         }
 212 }
 213
 214 void
 215 free_initmem (void)
 216 {
 217         unsigned long addr, eaddr;
 218
 219         addr = (unsigned long) ia64_imva(__init_begin);
 220         eaddr = (unsigned long) ia64_imva(__init_end);
 221         while (addr < eaddr) {
 222                 ClearPageReserved(virt_to_page(addr));
 223                 init_page_count(virt_to_page(addr));
 224                 free_page(addr);
 225                 ++totalram_pages;
 226                 addr += PAGE_SIZE;
 227         }
 228         printk(KERN_INFO "Freeing unused kernel memory: %ldkB freed\n",
 229                (__init_end - __init_begin) >> 10);
 230 }
 231
 232 void __init
 233 free_initrd_mem (unsigned long start, unsigned long end)
 234 {
 235         struct page *page;
 236         /*
 237          * EFI uses 4KB pages while the kernel can use 4KB or bigger.
 238          * Thus EFI and the kernel may have different page sizes. It is
 239          * therefore possible to have the initrd share the same page as
 240          * the end of the kernel (given current setup).
 241          *
 242          * To avoid freeing/using the wrong page (kernel sized) we:
 243          *      - align up the beginning of initrd
 244          *      - align down the end of initrd
 245          *
 246          *  |             |
 247          *  |=============| a000
 248          *  |             |
 249          *  |             |
 250          *  |             | 9000
 251          *  |/////////////|
 252          *  |/////////////|
 253          *  |=============| 8000
 254          *  |///INITRD////|
 255          *  |/////////////|
 256          *  |/////////////| 7000
 257          *  |             |
 258          *  |KKKKKKKKKKKKK|
 259          *  |=============| 6000
 260          *  |KKKKKKKKKKKKK|
 261          *  |KKKKKKKKKKKKK|
 262          *  K=kernel using 8KB pages
 263          *
 264          * In this example, we must free page 8000 ONLY. So we must align up
 265          * initrd_start and keep initrd_end as is.
 266          */
 267         start = PAGE_ALIGN(start);
 268         end = end & PAGE_MASK;
 269
 270         if (start < end)
 271                 printk(KERN_INFO "Freeing initrd memory: %ldkB freed\n", (end - start) >> 10);
 272
 273         for (; start < end; start += PAGE_SIZE) {
 274                 if (!virt_addr_valid(start))
 275                         continue;
 276                 page = virt_to_page(start);
 277                 ClearPageReserved(page);
 278                 init_page_count(page);
 279                 free_page(start);
 280                 ++totalram_pages;
 281         }
 282 }
 283
 284 /*
 285  * This installs a clean page in the kernel's page table.
 286  */
 287 static struct page * __init
 288 put_kernel_page (struct page *page, unsigned long address, pgprot_t pgprot)
 289 {
 290         pgd_t *pgd;
 291         pud_t *pud;
 292         pmd_t *pmd;
 293         pte_t *pte;
 294
 295         if (!PageReserved(page))
 296                 printk(KERN_ERR "put_kernel_page: page at 0x%p not in reserved memory\n",
 297                        page_address(page));
 298
 299         pgd = pgd_offset_k(address);            /* note: this is NOT pgd_offset()! */
 300
 301         {
 302                 pud = pud_alloc(&init_mm, pgd, address);
 303                 if (!pud)
 304                         goto out;
 305                 pmd = pmd_alloc(&init_mm, pud, address);
 306                 if (!pmd)
 307                         goto out;
 308                 pte = pte_alloc_kernel(pmd, address);
 309                 if (!pte)
 310                         goto out;
 311                 if (!pte_none(*pte))
 312                         goto out;
 313                 set_pte(pte, mk_pte(page, pgprot));
 314         }
 315   out:
 316         /* no need for flush_tlb */
 317         return page;
 318 }
 319
 320 static void __init
 321 setup_gate (void)
 322 {
 323         struct page *page;
 324
 325         /*
 326          * Map the gate page twice: once read-only to export the ELF
 327          * headers etc. and once execute-only page to enable
 328          * privilege-promotion via "epc":
 329          */
 330         page = virt_to_page(ia64_imva(__start_gate_section));
 331         put_kernel_page(page, GATE_ADDR, PAGE_READONLY);
 332 #ifdef HAVE_BUGGY_SEGREL
 333         page = virt_to_page(ia64_imva(__start_gate_section + PAGE_SIZE));
 334         put_kernel_page(page, GATE_ADDR + PAGE_SIZE, PAGE_GATE);
 335 #else
 336         put_kernel_page(page, GATE_ADDR + PERCPU_PAGE_SIZE, PAGE_GATE);
 337         /* Fill in the holes (if any) with read-only zero pages: */
 338         {
 339                 unsigned long addr;
 340
 341                 for (addr = GATE_ADDR + PAGE_SIZE;
 342                      addr < GATE_ADDR + PERCPU_PAGE_SIZE;
 343                      addr += PAGE_SIZE)
 344                 {
 345                         put_kernel_page(ZERO_PAGE(0), addr,
 346                                         PAGE_READONLY);
 347                         put_kernel_page(ZERO_PAGE(0), addr + PERCPU_PAGE_SIZE,
 348                                         PAGE_READONLY);
 349                 }
 350         }
 351 #endif
 352         ia64_patch_gate();
 353 }
 354
 355 void __devinit
 356 ia64_mmu_init (void *my_cpu_data)
 357 {
 358         unsigned long pta, impl_va_bits;
 359         extern void __devinit tlb_init (void);
 360
 361 #ifdef CONFIG_DISABLE_VHPT
 362 #       define VHPT_ENABLE_BIT  0
 363 #else
 364 #       define VHPT_ENABLE_BIT  1
 365 #endif
 366
 367         /*
 368          * Check if the virtually mapped linear page table (VMLPT) overlaps with a mapped
 369          * address space.  The IA-64 architecture guarantees that at least 50 bits of
 370          * virtual address space are implemented but if we pick a large enough page size
 371          * (e.g., 64KB), the mapped address space is big enough that it will overlap with
 372          * VMLPT.  I assume that once we run on machines big enough to warrant 64KB pages,
 373          * IMPL_VA_MSB will be significantly bigger, so this is unlikely to become a
 374          * problem in practice.  Alternatively, we could truncate the top of the mapped
 375          * address space to not permit mappings that would overlap with the VMLPT.
 376          * --davidm 00/12/06
 377          */
 378 #       define pte_bits                 3
 379 #       define mapped_space_bits        (3*(PAGE_SHIFT - pte_bits) + PAGE_SHIFT)
 380         /*
 381          * The virtual page table has to cover the entire implemented address space within
 382          * a region even though not all of this space may be mappable.  The reason for
 383          * this is that the Access bit and Dirty bit fault handlers perform
 384          * non-speculative accesses to the virtual page table, so the address range of the
 385          * virtual page table itself needs to be covered by virtual page table.
 386          */
 387 #       define vmlpt_bits               (impl_va_bits - PAGE_SHIFT + pte_bits)
 388 #       define POW2(n)                  (1ULL << (n))
 389
 390         impl_va_bits = ffz(~(local_cpu_data->unimpl_va_mask | (7UL << 61)));
 391
 392         if (impl_va_bits < 51 || impl_va_bits > 61)
 393                 panic("CPU has bogus IMPL_VA_MSB value of %lu!\n", impl_va_bits - 1);
 394         /*
 395          * mapped_space_bits - PAGE_SHIFT is the total number of ptes we need,
 396          * which must fit into "vmlpt_bits - pte_bits" slots. Second half of
 397          * the test makes sure that our mapped space doesn't overlap the
 398          * unimplemented hole in the middle of the region.
 399          */
 400         if ((mapped_space_bits - PAGE_SHIFT > vmlpt_bits - pte_bits) ||
 401             (mapped_space_bits > impl_va_bits - 1))
 402                 panic("Cannot build a big enough virtual-linear page table"
 403                       " to cover mapped address space.\n"
 404                       " Try using a smaller page size.\n");
 405
 406
 407         /* place the VMLPT at the end of each page-table mapped region: */
 408         pta = POW2(61) - POW2(vmlpt_bits);
 409
 410         /*
 411          * Set the (virtually mapped linear) page table address.  Bit
 412          * 8 selects between the short and long format, bits 2-7 the
 413          * size of the table, and bit 0 whether the VHPT walker is
 414          * enabled.
 415          */
 416         ia64_set_pta(pta | (0 << 8) | (vmlpt_bits << 2) | VHPT_ENABLE_BIT);
 417
 418         ia64_tlb_init();
 419
 420 #ifdef  CONFIG_HUGETLB_PAGE
 421         ia64_set_rr(HPAGE_REGION_BASE, HPAGE_SHIFT << 2);
 422         ia64_srlz_d();
 423 #endif
 424 }
 425
 426 #ifdef CONFIG_VIRTUAL_MEM_MAP
 427 int vmemmap_find_next_valid_pfn(int node, int i)
 428 {
 429         unsigned long end_address, hole_next_pfn;
 430         unsigned long stop_address;
 431         pg_data_t *pgdat = NODE_DATA(node);
 432
 433         end_address = (unsigned long) &vmem_map[pgdat->node_start_pfn + i];
 434         end_address = PAGE_ALIGN(end_address);
 435
 436         stop_address = (unsigned long) &vmem_map[
 437                 pgdat->node_start_pfn + pgdat->node_spanned_pages];
 438
 439         do {
 440                 pgd_t *pgd;
 441                 pud_t *pud;
 442                 pmd_t *pmd;
 443                 pte_t *pte;
 444
 445                 pgd = pgd_offset_k(end_address);
 446                 if (pgd_none(*pgd)) {
 447                         end_address += PGDIR_SIZE;
 448                         continue;
 449                 }
 450
 451                 pud = pud_offset(pgd, end_address);
 452                 if (pud_none(*pud)) {
 453                         end_address += PUD_SIZE;
 454                         continue;
 455                 }
 456
 457                 pmd = pmd_offset(pud, end_address);
 458                 if (pmd_none(*pmd)) {
 459                         end_address += PMD_SIZE;
 460                         continue;
 461                 }
 462
 463                 pte = pte_offset_kernel(pmd, end_address);
 464 retry_pte:
 465                 if (pte_none(*pte)) {
 466                         end_address += PAGE_SIZE;
 467                         pte++;
 468                         if ((end_address < stop_address) &&
 469                             (end_address != ALIGN(end_address, 1UL << PMD_SHIFT)))
 470                                 goto retry_pte;
 471                         continue;
 472                 }
 473                 /* Found next valid vmem_map page */
 474                 break;
 475         } while (end_address < stop_address);
 476
 477         end_address = min(end_address, stop_address);
 478         end_address = end_address - (unsigned long) vmem_map + sizeof(struct page) - 1;
 479         hole_next_pfn = end_address / sizeof(struct page);
 480         return hole_next_pfn - pgdat->node_start_pfn;
 481 }
 482
 483 int __init
 484 create_mem_map_page_table (u64 start, u64 end, void *arg)
 485 {
 486         unsigned long address, start_page, end_page;
 487         struct page *map_start, *map_end;
 488         int node;
 489         pgd_t *pgd;
 490         pud_t *pud;
 491         pmd_t *pmd;
 492         pte_t *pte;
 493
 494         map_start = vmem_map + (__pa(start) >> PAGE_SHIFT);
 495         map_end   = vmem_map + (__pa(end) >> PAGE_SHIFT);
 496
 497         start_page = (unsigned long) map_start & PAGE_MASK;
 498         end_page = PAGE_ALIGN((unsigned long) map_end);
 499         node = paddr_to_nid(__pa(start));
 500
 501         for (address = start_page; address < end_page; address += PAGE_SIZE) {
 502                 pgd = pgd_offset_k(address);
 503                 if (pgd_none(*pgd))
 504                         pgd_populate(&init_mm, pgd, alloc_bootmem_pages_node(NODE_DATA(node), PAGE_SIZE));
 505                 pud = pud_offset(pgd, address);
 506
 507                 if (pud_none(*pud))
 508                         pud_populate(&init_mm, pud, alloc_bootmem_pages_node(NODE_DATA(node), PAGE_SIZE));
 509                 pmd = pmd_offset(pud, address);
 510
 511                 if (pmd_none(*pmd))
 512                         pmd_populate_kernel(&init_mm, pmd, alloc_bootmem_pages_node(NODE_DATA(node), PAGE_SIZE));
 513                 pte = pte_offset_kernel(pmd, address);
 514
 515                 if (pte_none(*pte))
 516                         set_pte(pte, pfn_pte(__pa(alloc_bootmem_pages_node(NODE_DATA(node), PAGE_SIZE)) >> PAGE_SHIFT,
 517                                              PAGE_KERNEL));
 518         }
 519         return 0;
 520 }
 521
 522 struct memmap_init_callback_data {
 523         struct page *start;
 524         struct page *end;
 525         int nid;
 526         unsigned long zone;
 527 };
 528
 529 static int
 530 virtual_memmap_init (u64 start, u64 end, void *arg)
 531 {
 532         struct memmap_init_callback_data *args;
 533         struct page *map_start, *map_end;
 534
 535         args = (struct memmap_init_callback_data *) arg;
 536         map_start = vmem_map + (__pa(start) >> PAGE_SHIFT);
 537         map_end   = vmem_map + (__pa(end) >> PAGE_SHIFT);
 538
 539         if (map_start < args->start)
 540                 map_start = args->start;
 541         if (map_end > args->end)
 542                 map_end = args->end;
 543
 544         /*
 545          * We have to initialize "out of bounds" struct page elements that fit completely
 546          * on the same pages that were allocated for the "in bounds" elements because they
 547          * may be referenced later (and found to be "reserved").
 548          */
 549         map_start -= ((unsigned long) map_start & (PAGE_SIZE - 1)) / sizeof(struct page);
 550         map_end += ((PAGE_ALIGN((unsigned long) map_end) - (unsigned long) map_end)
 551                     / sizeof(struct page));
 552
 553         if (map_start < map_end)
 554                 memmap_init_zone((unsigned long)(map_end - map_start),
 555                                  args->nid, args->zone, page_to_pfn(map_start),
 556                                  MEMMAP_EARLY);
 557         return 0;
 558 }
 559
 560 void
 561 memmap_init (unsigned long size, int nid, unsigned long zone,
 562              unsigned long start_pfn)
 563 {
 564         if (!vmem_map)
 565                 memmap_init_zone(size, nid, zone, start_pfn, MEMMAP_EARLY);
 566         else {
 567                 struct page *start;
 568                 struct memmap_init_callback_data args;
 569
 570                 start = pfn_to_page(start_pfn);
 571                 args.start = start;
 572                 args.end = start + size;
 573                 args.nid = nid;
 574                 args.zone = zone;
 575
 576                 efi_memmap_walk(virtual_memmap_init, &args);
 577         }
 578 }
 579
 580 int
 581 ia64_pfn_valid (unsigned long pfn)
 582 {
 583         char byte;
 584         struct page *pg = pfn_to_page(pfn);
 585
 586         return     (__get_user(byte, (char __user *) pg) == 0)
 587                 && ((((u64)pg & PAGE_MASK) == (((u64)(pg + 1) - 1) & PAGE_MASK))
 588                         || (__get_user(byte, (char __user *) (pg + 1) - 1) == 0));
 589 }
 590 EXPORT_SYMBOL(ia64_pfn_valid);
 591
 592 int __init
 593 find_largest_hole (u64 start, u64 end, void *arg)
 594 {
 595         u64 *max_gap = arg;
 596
 597         static u64 last_end = PAGE_OFFSET;
 598
 599         /* NOTE: this algorithm assumes efi memmap table is ordered */
 600
 601         if (*max_gap < (start - last_end))
 602                 *max_gap = start - last_end;
 603         last_end = end;
 604         return 0;
 605 }
 606
 607 #endif /* CONFIG_VIRTUAL_MEM_MAP */
 608
 609 int __init
 610 register_active_ranges(u64 start, u64 end, void *arg)
 611 {
 612         int nid = paddr_to_nid(__pa(start));
 613
 614         if (nid < 0)
 615                 nid = 0;
 616 #ifdef CONFIG_KEXEC
 617         if (start > crashk_res.start && start < crashk_res.end)
 618                 start = crashk_res.end;
 619         if (end > crashk_res.start && end < crashk_res.end)
 620                 end = crashk_res.start;
 621 #endif
 622
 623         if (start < end)
 624                 add_active_range(nid, __pa(start) >> PAGE_SHIFT,
 625                         __pa(end) >> PAGE_SHIFT);
 626         return 0;
 627 }
 628
 629 static int __init
 630 count_reserved_pages (u64 start, u64 end, void *arg)
 631 {
 632         unsigned long num_reserved = 0;
 633         unsigned long *count = arg;
 634
 635         for (; start < end; start += PAGE_SIZE)
 636                 if (PageReserved(virt_to_page(start)))
 637                         ++num_reserved;
 638         *count += num_reserved;
 639         return 0;
 640 }
 641
 642 int
 643 find_max_min_low_pfn (unsigned long start, unsigned long end, void *arg)
 644 {
 645         unsigned long pfn_start, pfn_end;
 646 #ifdef CONFIG_FLATMEM
 647         pfn_start = (PAGE_ALIGN(__pa(start))) >> PAGE_SHIFT;
 648         pfn_end = (PAGE_ALIGN(__pa(end - 1))) >> PAGE_SHIFT;
 649 #else
 650         pfn_start = GRANULEROUNDDOWN(__pa(start)) >> PAGE_SHIFT;
 651         pfn_end = GRANULEROUNDUP(__pa(end - 1)) >> PAGE_SHIFT;
 652 #endif
 653         min_low_pfn = min(min_low_pfn, pfn_start);
 654         max_low_pfn = max(max_low_pfn, pfn_end);
 655         return 0;
 656 }
 657
 658 /*
 659  * Boot command-line option "nolwsys" can be used to disable the use of any light-weight
 660  * system call handler.  When this option is in effect, all fsyscalls will end up bubbling
 661  * down into the kernel and calling the normal (heavy-weight) syscall handler.  This is
 662  * useful for performance testing, but conceivably could also come in handy for debugging
 663  * purposes.
 664  */
 665
 666 static int nolwsys __initdata;
 667
 668 static int __init
 669 nolwsys_setup (char *s)
 670 {
 671         nolwsys = 1;
 672         return 1;
 673 }
 674
 675 __setup("nolwsys", nolwsys_setup);
 676
 677 void __init
 678 mem_init (void)
 679 {
 680         long reserved_pages, codesize, datasize, initsize;
 681         pg_data_t *pgdat;
 682         int i;
 683         static struct kcore_list kcore_mem, kcore_vmem, kcore_kernel;
 684
 685         BUG_ON(PTRS_PER_PGD * sizeof(pgd_t) != PAGE_SIZE);
 686         BUG_ON(PTRS_PER_PMD * sizeof(pmd_t) != PAGE_SIZE);
 687         BUG_ON(PTRS_PER_PTE * sizeof(pte_t) != PAGE_SIZE);
 688
 689 #ifdef CONFIG_PCI
 690         /*
 691          * This needs to be called _after_ the command line has been parsed but _before_
 692          * any drivers that may need the PCI DMA interface are initialized or bootmem has
 693          * been freed.
 694          */
 695         platform_dma_init();
 696 #endif
 697
 698 #ifdef CONFIG_FLATMEM
 699         if (!mem_map)
 700                 BUG();
 701         max_mapnr = max_low_pfn;
 702 #endif
 703
 704         high_memory = __va(max_low_pfn * PAGE_SIZE);
 705
 706         kclist_add(&kcore_mem, __va(0), max_low_pfn * PAGE_SIZE);
 707         kclist_add(&kcore_vmem, (void *)VMALLOC_START, VMALLOC_END-VMALLOC_START);
 708         kclist_add(&kcore_kernel, _stext, _end - _stext);
 709
 710         for_each_online_pgdat(pgdat)
 711                 if (pgdat->bdata->node_bootmem_map)
 712                         totalram_pages += free_all_bootmem_node(pgdat);
 713
 714         reserved_pages = 0;
 715         efi_memmap_walk(count_reserved_pages, &reserved_pages);
 716
 717         codesize =  (unsigned long) _etext - (unsigned long) _stext;
 718         datasize =  (unsigned long) _edata - (unsigned long) _etext;
 719         initsize =  (unsigned long) __init_end - (unsigned long) __init_begin;
 720
 721         printk(KERN_INFO "Memory: %luk/%luk available (%luk code, %luk reserved, "
 722                "%luk data, %luk init)\n", (unsigned long) nr_free_pages() << (PAGE_SHIFT - 10),
 723                num_physpages << (PAGE_SHIFT - 10), codesize >> 10,
 724                reserved_pages << (PAGE_SHIFT - 10), datasize >> 10, initsize >> 10);
 725
 726
 727         /*
 728          * For fsyscall entrpoints with no light-weight handler, use the ordinary
 729          * (heavy-weight) handler, but mark it by setting bit 0, so the fsyscall entry
 730          * code can tell them apart.
 731          */
 732         for (i = 0; i < NR_syscalls; ++i) {
 733                 extern unsigned long fsyscall_table[NR_syscalls];
 734                 extern unsigned long sys_call_table[NR_syscalls];
 735
 736                 if (!fsyscall_table[i] || nolwsys)
 737                         fsyscall_table[i] = sys_call_table[i] | 1;
 738         }
 739         setup_gate();
 740
 741 #ifdef CONFIG_IA32_SUPPORT
 742         ia32_mem_init();
 743 #endif
 744 }
 745
 746 #ifdef CONFIG_MEMORY_HOTPLUG
 747 void online_page(struct page *page)
 748 {
 749         ClearPageReserved(page);
 750         init_page_count(page);
 751         __free_page(page);
 752         totalram_pages++;
 753         num_physpages++;
 754 }
 755
 756 int arch_add_memory(int nid, u64 start, u64 size)
 757 {
 758         pg_data_t *pgdat;
 759         struct zone *zone;
 760         unsigned long start_pfn = start >> PAGE_SHIFT;
 761         unsigned long nr_pages = size >> PAGE_SHIFT;
 762         int ret;
 763
 764         pgdat = NODE_DATA(nid);
 765
 766         zone = pgdat->node_zones + ZONE_NORMAL;
 767         ret = __add_pages(zone, start_pfn, nr_pages);
 768
 769         if (ret)
 770                 printk("%s: Problem encountered in __add_pages() as ret=%d\n",
 771                        __FUNCTION__,  ret);
 772
 773         return ret;
 774 }
 775
 776 int remove_memory(u64 start, u64 size)
 777 {
 778         return -EINVAL;
 779 }
 780 EXPORT_SYMBOL_GPL(remove_memory);
 781 #endif