kernel/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/memblock.h>
  14 #include <linux/mm.h>
  15 #include <linux/mmzone.h>
  16 #include <linux/module.h>
  17 #include <linux/personality.h>
  18 #include <linux/reboot.h>
  19 #include <linux/slab.h>
  20 #include <linux/swap.h>
  21 #include <linux/proc_fs.h>
  22 #include <linux/bitops.h>
  23 #include <linux/kexec.h>
  24
  25 #include <asm/dma.h>
  26 #include <asm/io.h>
  27 #include <asm/machvec.h>
  28 #include <asm/numa.h>
  29 #include <asm/patch.h>
  30 #include <asm/pgalloc.h>
  31 #include <asm/sal.h>
  32 #include <asm/sections.h>
  33 #include <asm/tlb.h>
  34 #include <asm/uaccess.h>
  35 #include <asm/unistd.h>
  36 #include <asm/mca.h>
  37
  38 extern void ia64_tlb_init (void);
  39
  40 unsigned long MAX_DMA_ADDRESS = PAGE_OFFSET + 0x100000000UL;
  41
  42 #ifdef CONFIG_VIRTUAL_MEM_MAP
  43 unsigned long VMALLOC_END = VMALLOC_END_INIT;
  44 EXPORT_SYMBOL(VMALLOC_END);
  45 struct page *vmem_map;
  46 EXPORT_SYMBOL(vmem_map);
  47 #endif
  48
  49 struct page *zero_page_memmap_ptr;      /* map entry for zero page */
  50 EXPORT_SYMBOL(zero_page_memmap_ptr);
  51
  52 void
  53 __ia64_sync_icache_dcache (pte_t pte)
  54 {
  55         unsigned long addr;
  56         struct page *page;
  57
  58         page = pte_page(pte);
  59         addr = (unsigned long) page_address(page);
  60
  61         if (test_bit(PG_arch_1, &page->flags))
  62                 return;                         /* i-cache is already coherent with d-cache */
  63
  64         flush_icache_range(addr, addr + (PAGE_SIZE << compound_order(page)));
  65         set_bit(PG_arch_1, &page->flags);       /* mark page as clean */
  66 }
  67
  68 /*
  69  * Since DMA is i-cache coherent, any (complete) pages that were written via
  70  * DMA can be marked as "clean" so that lazy_mmu_prot_update() doesn't have to
  71  * flush them when they get mapped into an executable vm-area.
  72  */
  73 void
  74 dma_mark_clean(void *addr, size_t size)
  75 {
  76         unsigned long pg_addr, end;
  77
  78         pg_addr = PAGE_ALIGN((unsigned long) addr);
  79         end = (unsigned long) addr + size;
  80         while (pg_addr + PAGE_SIZE <= end) {
  81                 struct page *page = virt_to_page(pg_addr);
  82                 set_bit(PG_arch_1, &page->flags);
  83                 pg_addr += PAGE_SIZE;
  84         }
  85 }
  86
  87 inline void
  88 ia64_set_rbs_bot (void)
  89 {
  90         unsigned long stack_size = rlimit_max(RLIMIT_STACK) & -16;
  91
  92         if (stack_size > MAX_USER_STACK_SIZE)
  93                 stack_size = MAX_USER_STACK_SIZE;
  94         current->thread.rbs_bot = PAGE_ALIGN(current->mm->start_stack - stack_size);
  95 }
  96
  97 /*
  98  * This performs some platform-dependent address space initialization.
  99  * On IA-64, we want to setup the VM area for the register backing
 100  * store (which grows upwards) and install the gateway page which is
 101  * used for signal trampolines, etc.
 102  */
 103 void
 104 ia64_init_addr_space (void)
 105 {
 106         struct vm_area_struct *vma;
 107
 108         ia64_set_rbs_bot();
 109
 110         /*
 111          * If we're out of memory and kmem_cache_alloc() returns NULL, we simply ignore
 112          * the problem.  When the process attempts to write to the register backing store
 113          * for the first time, it will get a SEGFAULT in this case.
 114          */
 115         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
 116         if (vma) {
 117                 INIT_LIST_HEAD(&vma->anon_vma_chain);
 118                 vma->vm_mm = current->mm;
 119                 vma->vm_start = current->thread.rbs_bot & PAGE_MASK;
 120                 vma->vm_end = vma->vm_start + PAGE_SIZE;
 121                 vma->vm_flags = VM_DATA_DEFAULT_FLAGS|VM_GROWSUP|VM_ACCOUNT;
 122                 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
 123                 down_write(&current->mm->mmap_sem);
 124                 if (insert_vm_struct(current->mm, vma)) {
 125                         up_write(&current->mm->mmap_sem);
 126                         kmem_cache_free(vm_area_cachep, vma);
 127                         return;
 128                 }
 129                 up_write(&current->mm->mmap_sem);
 130         }
 131
 132         /* map NaT-page at address zero to speed up speculative dereferencing of NULL: */
 133         if (!(current->personality & MMAP_PAGE_ZERO)) {
 134                 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
 135                 if (vma) {
 136                         INIT_LIST_HEAD(&vma->anon_vma_chain);
 137                         vma->vm_mm = current->mm;
 138                         vma->vm_end = PAGE_SIZE;
 139                         vma->vm_page_prot = __pgprot(pgprot_val(PAGE_READONLY) | _PAGE_MA_NAT);
 140                         vma->vm_flags = VM_READ | VM_MAYREAD | VM_IO |
 141                                         VM_DONTEXPAND | VM_DONTDUMP;
 142                         down_write(&current->mm->mmap_sem);
 143                         if (insert_vm_struct(current->mm, vma)) {
 144                                 up_write(&current->mm->mmap_sem);
 145                                 kmem_cache_free(vm_area_cachep, vma);
 146                                 return;
 147                         }
 148                         up_write(&current->mm->mmap_sem);
 149                 }
 150         }
 151 }
 152
 153 void
 154 free_initmem (void)
 155 {
 156         free_reserved_area(ia64_imva(__init_begin), ia64_imva(__init_end),
 157                            -1, "unused kernel");
 158 }
 159
 160 void __init
 161 free_initrd_mem (unsigned long start, unsigned long end)
 162 {
 163         /*
 164          * EFI uses 4KB pages while the kernel can use 4KB or bigger.
 165          * Thus EFI and the kernel may have different page sizes. It is
 166          * therefore possible to have the initrd share the same page as
 167          * the end of the kernel (given current setup).
 168          *
 169          * To avoid freeing/using the wrong page (kernel sized) we:
 170          *      - align up the beginning of initrd
 171          *      - align down the end of initrd
 172          *
 173          *  |             |
 174          *  |=============| a000
 175          *  |             |
 176          *  |             |
 177          *  |             | 9000
 178          *  |/////////////|
 179          *  |/////////////|
 180          *  |=============| 8000
 181          *  |///INITRD////|
 182          *  |/////////////|
 183          *  |/////////////| 7000
 184          *  |             |
 185          *  |KKKKKKKKKKKKK|
 186          *  |=============| 6000
 187          *  |KKKKKKKKKKKKK|
 188          *  |KKKKKKKKKKKKK|
 189          *  K=kernel using 8KB pages
 190          *
 191          * In this example, we must free page 8000 ONLY. So we must align up
 192          * initrd_start and keep initrd_end as is.
 193          */
 194         start = PAGE_ALIGN(start);
 195         end = end & PAGE_MASK;
 196
 197         if (start < end)
 198                 printk(KERN_INFO "Freeing initrd memory: %ldkB freed\n", (end - start) >> 10);
 199
 200         for (; start < end; start += PAGE_SIZE) {
 201                 if (!virt_addr_valid(start))
 202                         continue;
 203                 free_reserved_page(virt_to_page(start));
 204         }
 205 }
 206
 207 /*
 208  * This installs a clean page in the kernel's page table.
 209  */
 210 static struct page * __init
 211 put_kernel_page (struct page *page, unsigned long address, pgprot_t pgprot)
 212 {
 213         pgd_t *pgd;
 214         pud_t *pud;
 215         pmd_t *pmd;
 216         pte_t *pte;
 217
 218         pgd = pgd_offset_k(address);            /* note: this is NOT pgd_offset()! */
 219
 220         {
 221                 pud = pud_alloc(&init_mm, pgd, address);
 222                 if (!pud)
 223                         goto out;
 224                 pmd = pmd_alloc(&init_mm, pud, address);
 225                 if (!pmd)
 226                         goto out;
 227                 pte = pte_alloc_kernel(pmd, address);
 228                 if (!pte)
 229                         goto out;
 230                 if (!pte_none(*pte))
 231                         goto out;
 232                 set_pte(pte, mk_pte(page, pgprot));
 233         }
 234   out:
 235         /* no need for flush_tlb */
 236         return page;
 237 }
 238
 239 static void __init
 240 setup_gate (void)
 241 {
 242         struct page *page;
 243
 244         /*
 245          * Map the gate page twice: once read-only to export the ELF
 246          * headers etc. and once execute-only page to enable
 247          * privilege-promotion via "epc":
 248          */
 249         page = virt_to_page(ia64_imva(__start_gate_section));
 250         put_kernel_page(page, GATE_ADDR, PAGE_READONLY);
 251 #ifdef HAVE_BUGGY_SEGREL
 252         page = virt_to_page(ia64_imva(__start_gate_section + PAGE_SIZE));
 253         put_kernel_page(page, GATE_ADDR + PAGE_SIZE, PAGE_GATE);
 254 #else
 255         put_kernel_page(page, GATE_ADDR + PERCPU_PAGE_SIZE, PAGE_GATE);
 256         /* Fill in the holes (if any) with read-only zero pages: */
 257         {
 258                 unsigned long addr;
 259
 260                 for (addr = GATE_ADDR + PAGE_SIZE;
 261                      addr < GATE_ADDR + PERCPU_PAGE_SIZE;
 262                      addr += PAGE_SIZE)
 263                 {
 264                         put_kernel_page(ZERO_PAGE(0), addr,
 265                                         PAGE_READONLY);
 266                         put_kernel_page(ZERO_PAGE(0), addr + PERCPU_PAGE_SIZE,
 267                                         PAGE_READONLY);
 268                 }
 269         }
 270 #endif
 271         ia64_patch_gate();
 272 }
 273
 274 static struct vm_area_struct gate_vma;
 275
 276 static int __init gate_vma_init(void)
 277 {
 278         gate_vma.vm_mm = NULL;
 279         gate_vma.vm_start = FIXADDR_USER_START;
 280         gate_vma.vm_end = FIXADDR_USER_END;
 281         gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
 282         gate_vma.vm_page_prot = __P101;
 283
 284         return 0;
 285 }
 286 __initcall(gate_vma_init);
 287
 288 struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
 289 {
 290         return &gate_vma;
 291 }
 292
 293 int in_gate_area_no_mm(unsigned long addr)
 294 {
 295         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
 296                 return 1;
 297         return 0;
 298 }
 299
 300 int in_gate_area(struct mm_struct *mm, unsigned long addr)
 301 {
 302         return in_gate_area_no_mm(addr);
 303 }
 304
 305 void ia64_mmu_init(void *my_cpu_data)
 306 {
 307         unsigned long pta, impl_va_bits;
 308         extern void tlb_init(void);
 309
 310 #ifdef CONFIG_DISABLE_VHPT
 311 #       define VHPT_ENABLE_BIT  0
 312 #else
 313 #       define VHPT_ENABLE_BIT  1
 314 #endif
 315
 316         /*
 317          * Check if the virtually mapped linear page table (VMLPT) overlaps with a mapped
 318          * address space.  The IA-64 architecture guarantees that at least 50 bits of
 319          * virtual address space are implemented but if we pick a large enough page size
 320          * (e.g., 64KB), the mapped address space is big enough that it will overlap with
 321          * VMLPT.  I assume that once we run on machines big enough to warrant 64KB pages,
 322          * IMPL_VA_MSB will be significantly bigger, so this is unlikely to become a
 323          * problem in practice.  Alternatively, we could truncate the top of the mapped
 324          * address space to not permit mappings that would overlap with the VMLPT.
 325          * --davidm 00/12/06
 326          */
 327 #       define pte_bits                 3
 328 #       define mapped_space_bits        (3*(PAGE_SHIFT - pte_bits) + PAGE_SHIFT)
 329         /*
 330          * The virtual page table has to cover the entire implemented address space within
 331          * a region even though not all of this space may be mappable.  The reason for
 332          * this is that the Access bit and Dirty bit fault handlers perform
 333          * non-speculative accesses to the virtual page table, so the address range of the
 334          * virtual page table itself needs to be covered by virtual page table.
 335          */
 336 #       define vmlpt_bits               (impl_va_bits - PAGE_SHIFT + pte_bits)
 337 #       define POW2(n)                  (1ULL << (n))
 338
 339         impl_va_bits = ffz(~(local_cpu_data->unimpl_va_mask | (7UL << 61)));
 340
 341         if (impl_va_bits < 51 || impl_va_bits > 61)
 342                 panic("CPU has bogus IMPL_VA_MSB value of %lu!\n", impl_va_bits - 1);
 343         /*
 344          * mapped_space_bits - PAGE_SHIFT is the total number of ptes we need,
 345          * which must fit into "vmlpt_bits - pte_bits" slots. Second half of
 346          * the test makes sure that our mapped space doesn't overlap the
 347          * unimplemented hole in the middle of the region.
 348          */
 349         if ((mapped_space_bits - PAGE_SHIFT > vmlpt_bits - pte_bits) ||
 350             (mapped_space_bits > impl_va_bits - 1))
 351                 panic("Cannot build a big enough virtual-linear page table"
 352                       " to cover mapped address space.\n"
 353                       " Try using a smaller page size.\n");
 354
 355
 356         /* place the VMLPT at the end of each page-table mapped region: */
 357         pta = POW2(61) - POW2(vmlpt_bits);
 358
 359         /*
 360          * Set the (virtually mapped linear) page table address.  Bit
 361          * 8 selects between the short and long format, bits 2-7 the
 362          * size of the table, and bit 0 whether the VHPT walker is
 363          * enabled.
 364          */
 365         ia64_set_pta(pta | (0 << 8) | (vmlpt_bits << 2) | VHPT_ENABLE_BIT);
 366
 367         ia64_tlb_init();
 368
 369 #ifdef  CONFIG_HUGETLB_PAGE
 370         ia64_set_rr(HPAGE_REGION_BASE, HPAGE_SHIFT << 2);
 371         ia64_srlz_d();
 372 #endif
 373 }
 374
 375 #ifdef CONFIG_VIRTUAL_MEM_MAP
 376 int vmemmap_find_next_valid_pfn(int node, int i)
 377 {
 378         unsigned long end_address, hole_next_pfn;
 379         unsigned long stop_address;
 380         pg_data_t *pgdat = NODE_DATA(node);
 381
 382         end_address = (unsigned long) &vmem_map[pgdat->node_start_pfn + i];
 383         end_address = PAGE_ALIGN(end_address);
 384         stop_address = (unsigned long) &vmem_map[pgdat_end_pfn(pgdat)];
 385
 386         do {
 387                 pgd_t *pgd;
 388                 pud_t *pud;
 389                 pmd_t *pmd;
 390                 pte_t *pte;
 391
 392                 pgd = pgd_offset_k(end_address);
 393                 if (pgd_none(*pgd)) {
 394                         end_address += PGDIR_SIZE;
 395                         continue;
 396                 }
 397
 398                 pud = pud_offset(pgd, end_address);
 399                 if (pud_none(*pud)) {
 400                         end_address += PUD_SIZE;
 401                         continue;
 402                 }
 403
 404                 pmd = pmd_offset(pud, end_address);
 405                 if (pmd_none(*pmd)) {
 406                         end_address += PMD_SIZE;
 407                         continue;
 408                 }
 409
 410                 pte = pte_offset_kernel(pmd, end_address);
 411 retry_pte:
 412                 if (pte_none(*pte)) {
 413                         end_address += PAGE_SIZE;
 414                         pte++;
 415                         if ((end_address < stop_address) &&
 416                             (end_address != ALIGN(end_address, 1UL << PMD_SHIFT)))
 417                                 goto retry_pte;
 418                         continue;
 419                 }
 420                 /* Found next valid vmem_map page */
 421                 break;
 422         } while (end_address < stop_address);
 423
 424         end_address = min(end_address, stop_address);
 425         end_address = end_address - (unsigned long) vmem_map + sizeof(struct page) - 1;
 426         hole_next_pfn = end_address / sizeof(struct page);
 427         return hole_next_pfn - pgdat->node_start_pfn;
 428 }
 429
 430 int __init create_mem_map_page_table(u64 start, u64 end, void *arg)
 431 {
 432         unsigned long address, start_page, end_page;
 433         struct page *map_start, *map_end;
 434         int node;
 435         pgd_t *pgd;
 436         pud_t *pud;
 437         pmd_t *pmd;
 438         pte_t *pte;
 439
 440         map_start = vmem_map + (__pa(start) >> PAGE_SHIFT);
 441         map_end   = vmem_map + (__pa(end) >> PAGE_SHIFT);
 442
 443         start_page = (unsigned long) map_start & PAGE_MASK;
 444         end_page = PAGE_ALIGN((unsigned long) map_end);
 445         node = paddr_to_nid(__pa(start));
 446
 447         for (address = start_page; address < end_page; address += PAGE_SIZE) {
 448                 pgd = pgd_offset_k(address);
 449                 if (pgd_none(*pgd))
 450                         pgd_populate(&init_mm, pgd, alloc_bootmem_pages_node(NODE_DATA(node), PAGE_SIZE));
 451                 pud = pud_offset(pgd, address);
 452
 453                 if (pud_none(*pud))
 454                         pud_populate(&init_mm, pud, alloc_bootmem_pages_node(NODE_DATA(node), PAGE_SIZE));
 455                 pmd = pmd_offset(pud, address);
 456
 457                 if (pmd_none(*pmd))
 458                         pmd_populate_kernel(&init_mm, pmd, alloc_bootmem_pages_node(NODE_DATA(node), PAGE_SIZE));
 459                 pte = pte_offset_kernel(pmd, address);
 460
 461                 if (pte_none(*pte))
 462                         set_pte(pte, pfn_pte(__pa(alloc_bootmem_pages_node(NODE_DATA(node), PAGE_SIZE)) >> PAGE_SHIFT,
 463                                              PAGE_KERNEL));
 464         }
 465         return 0;
 466 }
 467
 468 struct memmap_init_callback_data {
 469         struct page *start;
 470         struct page *end;
 471         int nid;
 472         unsigned long zone;
 473 };
 474
 475 static int __meminit
 476 virtual_memmap_init(u64 start, u64 end, void *arg)
 477 {
 478         struct memmap_init_callback_data *args;
 479         struct page *map_start, *map_end;
 480
 481         args = (struct memmap_init_callback_data *) arg;
 482         map_start = vmem_map + (__pa(start) >> PAGE_SHIFT);
 483         map_end   = vmem_map + (__pa(end) >> PAGE_SHIFT);
 484
 485         if (map_start < args->start)
 486                 map_start = args->start;
 487         if (map_end > args->end)
 488                 map_end = args->end;
 489
 490         /*
 491          * We have to initialize "out of bounds" struct page elements that fit completely
 492          * on the same pages that were allocated for the "in bounds" elements because they
 493          * may be referenced later (and found to be "reserved").
 494          */
 495         map_start -= ((unsigned long) map_start & (PAGE_SIZE - 1)) / sizeof(struct page);
 496         map_end += ((PAGE_ALIGN((unsigned long) map_end) - (unsigned long) map_end)
 497                     / sizeof(struct page));
 498
 499         if (map_start < map_end)
 500                 memmap_init_zone((unsigned long)(map_end - map_start),
 501                                  args->nid, args->zone, page_to_pfn(map_start),
 502                                  MEMMAP_EARLY);
 503         return 0;
 504 }
 505
 506 void __meminit
 507 memmap_init (unsigned long size, int nid, unsigned long zone,
 508              unsigned long start_pfn)
 509 {
 510         if (!vmem_map)
 511                 memmap_init_zone(size, nid, zone, start_pfn, MEMMAP_EARLY);
 512         else {
 513                 struct page *start;
 514                 struct memmap_init_callback_data args;
 515
 516                 start = pfn_to_page(start_pfn);
 517                 args.start = start;
 518                 args.end = start + size;
 519                 args.nid = nid;
 520                 args.zone = zone;
 521
 522                 efi_memmap_walk(virtual_memmap_init, &args);
 523         }
 524 }
 525
 526 int
 527 ia64_pfn_valid (unsigned long pfn)
 528 {
 529         char byte;
 530         struct page *pg = pfn_to_page(pfn);
 531
 532         return     (__get_user(byte, (char __user *) pg) == 0)
 533                 && ((((u64)pg & PAGE_MASK) == (((u64)(pg + 1) - 1) & PAGE_MASK))
 534                         || (__get_user(byte, (char __user *) (pg + 1) - 1) == 0));
 535 }
 536 EXPORT_SYMBOL(ia64_pfn_valid);
 537
 538 int __init find_largest_hole(u64 start, u64 end, void *arg)
 539 {
 540         u64 *max_gap = arg;
 541
 542         static u64 last_end = PAGE_OFFSET;
 543
 544         /* NOTE: this algorithm assumes efi memmap table is ordered */
 545
 546         if (*max_gap < (start - last_end))
 547                 *max_gap = start - last_end;
 548         last_end = end;
 549         return 0;
 550 }
 551
 552 #endif /* CONFIG_VIRTUAL_MEM_MAP */
 553
 554 int __init register_active_ranges(u64 start, u64 len, int nid)
 555 {
 556         u64 end = start + len;
 557
 558 #ifdef CONFIG_KEXEC
 559         if (start > crashk_res.start && start < crashk_res.end)
 560                 start = crashk_res.end;
 561         if (end > crashk_res.start && end < crashk_res.end)
 562                 end = crashk_res.start;
 563 #endif
 564
 565         if (start < end)
 566                 memblock_add_node(__pa(start), end - start, nid);
 567         return 0;
 568 }
 569
 570 int
 571 find_max_min_low_pfn (u64 start, u64 end, void *arg)
 572 {
 573         unsigned long pfn_start, pfn_end;
 574 #ifdef CONFIG_FLATMEM
 575         pfn_start = (PAGE_ALIGN(__pa(start))) >> PAGE_SHIFT;
 576         pfn_end = (PAGE_ALIGN(__pa(end - 1))) >> PAGE_SHIFT;
 577 #else
 578         pfn_start = GRANULEROUNDDOWN(__pa(start)) >> PAGE_SHIFT;
 579         pfn_end = GRANULEROUNDUP(__pa(end - 1)) >> PAGE_SHIFT;
 580 #endif
 581         min_low_pfn = min(min_low_pfn, pfn_start);
 582         max_low_pfn = max(max_low_pfn, pfn_end);
 583         return 0;
 584 }
 585
 586 /*
 587  * Boot command-line option "nolwsys" can be used to disable the use of any light-weight
 588  * system call handler.  When this option is in effect, all fsyscalls will end up bubbling
 589  * down into the kernel and calling the normal (heavy-weight) syscall handler.  This is
 590  * useful for performance testing, but conceivably could also come in handy for debugging
 591  * purposes.
 592  */
 593
 594 static int nolwsys __initdata;
 595
 596 static int __init
 597 nolwsys_setup (char *s)
 598 {
 599         nolwsys = 1;
 600         return 1;
 601 }
 602
 603 __setup("nolwsys", nolwsys_setup);
 604
 605 void __init
 606 mem_init (void)
 607 {
 608         int i;
 609
 610         BUG_ON(PTRS_PER_PGD * sizeof(pgd_t) != PAGE_SIZE);
 611         BUG_ON(PTRS_PER_PMD * sizeof(pmd_t) != PAGE_SIZE);
 612         BUG_ON(PTRS_PER_PTE * sizeof(pte_t) != PAGE_SIZE);
 613
 614 #ifdef CONFIG_PCI
 615         /*
 616          * This needs to be called _after_ the command line has been parsed but _before_
 617          * any drivers that may need the PCI DMA interface are initialized or bootmem has
 618          * been freed.
 619          */
 620         platform_dma_init();
 621 #endif
 622
 623 #ifdef CONFIG_FLATMEM
 624         BUG_ON(!mem_map);
 625 #endif
 626
 627         set_max_mapnr(max_low_pfn);
 628         high_memory = __va(max_low_pfn * PAGE_SIZE);
 629         free_all_bootmem();
 630         mem_init_print_info(NULL);
 631
 632         /*
 633          * For fsyscall entrpoints with no light-weight handler, use the ordinary
 634          * (heavy-weight) handler, but mark it by setting bit 0, so the fsyscall entry
 635          * code can tell them apart.
 636          */
 637         for (i = 0; i < NR_syscalls; ++i) {
 638                 extern unsigned long fsyscall_table[NR_syscalls];
 639                 extern unsigned long sys_call_table[NR_syscalls];
 640
 641                 if (!fsyscall_table[i] || nolwsys)
 642                         fsyscall_table[i] = sys_call_table[i] | 1;
 643         }
 644         setup_gate();
 645 }
 646
 647 #ifdef CONFIG_MEMORY_HOTPLUG
 648 int arch_add_memory(int nid, u64 start, u64 size, bool for_device)
 649 {
 650         pg_data_t *pgdat;
 651         struct zone *zone;
 652         unsigned long start_pfn = start >> PAGE_SHIFT;
 653         unsigned long nr_pages = size >> PAGE_SHIFT;
 654         int ret;
 655
 656         pgdat = NODE_DATA(nid);
 657
 658         zone = pgdat->node_zones +
 659                 zone_for_memory(nid, start, size, ZONE_NORMAL, for_device);
 660         ret = __add_pages(nid, zone, start_pfn, nr_pages);
 661
 662         if (ret)
 663                 printk("%s: Problem encountered in __add_pages() as ret=%d\n",
 664                        __func__,  ret);
 665
 666         return ret;
 667 }
 668
 669 #ifdef CONFIG_MEMORY_HOTREMOVE
 670 int arch_remove_memory(u64 start, u64 size)
 671 {
 672         unsigned long start_pfn = start >> PAGE_SHIFT;
 673         unsigned long nr_pages = size >> PAGE_SHIFT;
 674         struct zone *zone;
 675         int ret;
 676
 677         zone = page_zone(pfn_to_page(start_pfn));
 678         ret = __remove_pages(zone, start_pfn, nr_pages);
 679         if (ret)
 680                 pr_warn("%s: Problem encountered in __remove_pages() as"
 681                         " ret=%d\n", __func__,  ret);
 682
 683         return ret;
 684 }
 685 #endif
 686 #endif
 687
 688 /**
 689  * show_mem - give short summary of memory stats
 690  *
 691  * Shows a simple page count of reserved and used pages in the system.
 692  * For discontig machines, it does this on a per-pgdat basis.
 693  */
 694 void show_mem(unsigned int filter)
 695 {
 696         int total_reserved = 0;
 697         unsigned long total_present = 0;
 698         pg_data_t *pgdat;
 699
 700         printk(KERN_INFO "Mem-info:\n");
 701         show_free_areas(filter);
 702         printk(KERN_INFO "Node memory in pages:\n");
 703         for_each_online_pgdat(pgdat) {
 704                 unsigned long present;
 705                 unsigned long flags;
 706                 int reserved = 0;
 707                 int nid = pgdat->node_id;
 708                 int zoneid;
 709
 710                 if (skip_free_areas_node(filter, nid))
 711                         continue;
 712                 pgdat_resize_lock(pgdat, &flags);
 713
 714                 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
 715                         struct zone *zone = &pgdat->node_zones[zoneid];
 716                         if (!populated_zone(zone))
 717                                 continue;
 718
 719                         reserved += zone->present_pages - zone->managed_pages;
 720                 }
 721                 present = pgdat->node_present_pages;
 722
 723                 pgdat_resize_unlock(pgdat, &flags);
 724                 total_present += present;
 725                 total_reserved += reserved;
 726                 printk(KERN_INFO "Node %4d:  RAM: %11ld, rsvd: %8d, ",
 727                        nid, present, reserved);
 728         }
 729         printk(KERN_INFO "%ld pages of RAM\n", total_present);
 730         printk(KERN_INFO "%d reserved pages\n", total_reserved);
 731         printk(KERN_INFO "Total of %ld pages in page table cache\n",
 732                quicklist_total_size());
 733         printk(KERN_INFO "%ld free buffer pages\n", nr_free_buffer_pages());
 734 }