arch/x86/mm/numa_64.c

   1 /*
   2  * Generic VM initialization for x86-64 NUMA setups.
   3  * Copyright 2002,2003 Andi Kleen, SuSE Labs.
   4  */
   5 #include <linux/kernel.h>
   6 #include <linux/mm.h>
   7 #include <linux/string.h>
   8 #include <linux/init.h>
   9 #include <linux/bootmem.h>
  10 #include <linux/memblock.h>
  11 #include <linux/mmzone.h>
  12 #include <linux/ctype.h>
  13 #include <linux/module.h>
  14 #include <linux/nodemask.h>
  15 #include <linux/sched.h>
  16 #include <linux/acpi.h>
  17
  18 #include <asm/e820.h>
  19 #include <asm/proto.h>
  20 #include <asm/dma.h>
  21 #include <asm/numa.h>
  22 #include <asm/acpi.h>
  23 #include <asm/amd_nb.h>
  24
  25 struct pglist_data *node_data[MAX_NUMNODES] __read_mostly;
  26 EXPORT_SYMBOL(node_data);
  27
  28 struct memnode memnode;
  29
  30 static unsigned long __initdata nodemap_addr;
  31 static unsigned long __initdata nodemap_size;
  32
  33 /*
  34  * Given a shift value, try to populate memnodemap[]
  35  * Returns :
  36  * 1 if OK
  37  * 0 if memnodmap[] too small (of shift too small)
  38  * -1 if node overlap or lost ram (shift too big)
  39  */
  40 static int __init populate_memnodemap(const struct bootnode *nodes,
  41                                       int numnodes, int shift, int *nodeids)
  42 {
  43         unsigned long addr, end;
  44         int i, res = -1;
  45
  46         memset(memnodemap, 0xff, sizeof(s16)*memnodemapsize);
  47         for (i = 0; i < numnodes; i++) {
  48                 addr = nodes[i].start;
  49                 end = nodes[i].end;
  50                 if (addr >= end)
  51                         continue;
  52                 if ((end >> shift) >= memnodemapsize)
  53                         return 0;
  54                 do {
  55                         if (memnodemap[addr >> shift] != NUMA_NO_NODE)
  56                                 return -1;
  57
  58                         if (!nodeids)
  59                                 memnodemap[addr >> shift] = i;
  60                         else
  61                                 memnodemap[addr >> shift] = nodeids[i];
  62
  63                         addr += (1UL << shift);
  64                 } while (addr < end);
  65                 res = 1;
  66         }
  67         return res;
  68 }
  69
  70 static int __init allocate_cachealigned_memnodemap(void)
  71 {
  72         unsigned long addr;
  73
  74         memnodemap = memnode.embedded_map;
  75         if (memnodemapsize <= ARRAY_SIZE(memnode.embedded_map))
  76                 return 0;
  77
  78         addr = 0x8000;
  79         nodemap_size = roundup(sizeof(s16) * memnodemapsize, L1_CACHE_BYTES);
  80         nodemap_addr = memblock_find_in_range(addr, get_max_mapped(),
  81                                       nodemap_size, L1_CACHE_BYTES);
  82         if (nodemap_addr == MEMBLOCK_ERROR) {
  83                 printk(KERN_ERR
  84                        "NUMA: Unable to allocate Memory to Node hash map\n");
  85                 nodemap_addr = nodemap_size = 0;
  86                 return -1;
  87         }
  88         memnodemap = phys_to_virt(nodemap_addr);
  89         memblock_x86_reserve_range(nodemap_addr, nodemap_addr + nodemap_size, "MEMNODEMAP");
  90
  91         printk(KERN_DEBUG "NUMA: Allocated memnodemap from %lx - %lx\n",
  92                nodemap_addr, nodemap_addr + nodemap_size);
  93         return 0;
  94 }
  95
  96 /*
  97  * The LSB of all start and end addresses in the node map is the value of the
  98  * maximum possible shift.
  99  */
 100 static int __init extract_lsb_from_nodes(const struct bootnode *nodes,
 101                                          int numnodes)
 102 {
 103         int i, nodes_used = 0;
 104         unsigned long start, end;
 105         unsigned long bitfield = 0, memtop = 0;
 106
 107         for (i = 0; i < numnodes; i++) {
 108                 start = nodes[i].start;
 109                 end = nodes[i].end;
 110                 if (start >= end)
 111                         continue;
 112                 bitfield |= start;
 113                 nodes_used++;
 114                 if (end > memtop)
 115                         memtop = end;
 116         }
 117         if (nodes_used <= 1)
 118                 i = 63;
 119         else
 120                 i = find_first_bit(&bitfield, sizeof(unsigned long)*8);
 121         memnodemapsize = (memtop >> i)+1;
 122         return i;
 123 }
 124
 125 int __init compute_hash_shift(struct bootnode *nodes, int numnodes,
 126                               int *nodeids)
 127 {
 128         int shift;
 129
 130         shift = extract_lsb_from_nodes(nodes, numnodes);
 131         if (allocate_cachealigned_memnodemap())
 132                 return -1;
 133         printk(KERN_DEBUG "NUMA: Using %d for the hash shift.\n",
 134                 shift);
 135
 136         if (populate_memnodemap(nodes, numnodes, shift, nodeids) != 1) {
 137                 printk(KERN_INFO "Your memory is not aligned you need to "
 138                        "rebuild your kernel with a bigger NODEMAPSIZE "
 139                        "shift=%d\n", shift);
 140                 return -1;
 141         }
 142         return shift;
 143 }
 144
 145 int __meminit  __early_pfn_to_nid(unsigned long pfn)
 146 {
 147         return phys_to_nid(pfn << PAGE_SHIFT);
 148 }
 149
 150 static void * __init early_node_mem(int nodeid, unsigned long start,
 151                                     unsigned long end, unsigned long size,
 152                                     unsigned long align)
 153 {
 154         unsigned long mem;
 155
 156         /*
 157          * put it on high as possible
 158          * something will go with NODE_DATA
 159          */
 160         if (start < (MAX_DMA_PFN<<PAGE_SHIFT))
 161                 start = MAX_DMA_PFN<<PAGE_SHIFT;
 162         if (start < (MAX_DMA32_PFN<<PAGE_SHIFT) &&
 163             end > (MAX_DMA32_PFN<<PAGE_SHIFT))
 164                 start = MAX_DMA32_PFN<<PAGE_SHIFT;
 165         mem = memblock_x86_find_in_range_node(nodeid, start, end, size, align);
 166         if (mem != MEMBLOCK_ERROR)
 167                 return __va(mem);
 168
 169         /* extend the search scope */
 170         end = max_pfn_mapped << PAGE_SHIFT;
 171         start = MAX_DMA_PFN << PAGE_SHIFT;
 172         mem = memblock_find_in_range(start, end, size, align);
 173         if (mem != MEMBLOCK_ERROR)
 174                 return __va(mem);
 175
 176         printk(KERN_ERR "Cannot find %lu bytes in node %d\n",
 177                        size, nodeid);
 178
 179         return NULL;
 180 }
 181
 182 /* Initialize bootmem allocator for a node */
 183 void __init
 184 setup_node_bootmem(int nodeid, unsigned long start, unsigned long end)
 185 {
 186         unsigned long start_pfn, last_pfn, nodedata_phys;
 187         const int pgdat_size = roundup(sizeof(pg_data_t), PAGE_SIZE);
 188         int nid;
 189
 190         if (!end)
 191                 return;
 192
 193         /*
 194          * Don't confuse VM with a node that doesn't have the
 195          * minimum amount of memory:
 196          */
 197         if (end && (end - start) < NODE_MIN_SIZE)
 198                 return;
 199
 200         start = roundup(start, ZONE_ALIGN);
 201
 202         printk(KERN_INFO "Initmem setup node %d %016lx-%016lx\n", nodeid,
 203                start, end);
 204
 205         start_pfn = start >> PAGE_SHIFT;
 206         last_pfn = end >> PAGE_SHIFT;
 207
 208         node_data[nodeid] = early_node_mem(nodeid, start, end, pgdat_size,
 209                                            SMP_CACHE_BYTES);
 210         if (node_data[nodeid] == NULL)
 211                 return;
 212         nodedata_phys = __pa(node_data[nodeid]);
 213         memblock_x86_reserve_range(nodedata_phys, nodedata_phys + pgdat_size, "NODE_DATA");
 214         printk(KERN_INFO "  NODE_DATA [%016lx - %016lx]\n", nodedata_phys,
 215                 nodedata_phys + pgdat_size - 1);
 216         nid = phys_to_nid(nodedata_phys);
 217         if (nid != nodeid)
 218                 printk(KERN_INFO "    NODE_DATA(%d) on node %d\n", nodeid, nid);
 219
 220         memset(NODE_DATA(nodeid), 0, sizeof(pg_data_t));
 221         NODE_DATA(nodeid)->node_id = nodeid;
 222         NODE_DATA(nodeid)->node_start_pfn = start_pfn;
 223         NODE_DATA(nodeid)->node_spanned_pages = last_pfn - start_pfn;
 224
 225         node_set_online(nodeid);
 226 }
 227
 228 #ifdef CONFIG_NUMA_EMU
 229 /* Numa emulation */
 230 static struct bootnode nodes[MAX_NUMNODES] __initdata;
 231 static struct bootnode physnodes[MAX_NUMNODES] __cpuinitdata;
 232 static char *cmdline __initdata;
 233
 234 void __init numa_emu_cmdline(char *str)
 235 {
 236         cmdline = str;
 237 }
 238
 239 static int __init setup_physnodes(unsigned long start, unsigned long end,
 240                                         int acpi, int amd)
 241 {
 242         int ret = 0;
 243         int i;
 244
 245         memset(physnodes, 0, sizeof(physnodes));
 246 #ifdef CONFIG_ACPI_NUMA
 247         if (acpi)
 248                 acpi_get_nodes(physnodes, start, end);
 249 #endif
 250 #ifdef CONFIG_AMD_NUMA
 251         if (amd)
 252                 amd_get_nodes(physnodes);
 253 #endif
 254         /*
 255          * Basic sanity checking on the physical node map: there may be errors
 256          * if the SRAT or AMD code incorrectly reported the topology or the mem=
 257          * kernel parameter is used.
 258          */
 259         for (i = 0; i < MAX_NUMNODES; i++) {
 260                 if (physnodes[i].start == physnodes[i].end)
 261                         continue;
 262                 if (physnodes[i].start > end) {
 263                         physnodes[i].end = physnodes[i].start;
 264                         continue;
 265                 }
 266                 if (physnodes[i].end < start) {
 267                         physnodes[i].start = physnodes[i].end;
 268                         continue;
 269                 }
 270                 if (physnodes[i].start < start)
 271                         physnodes[i].start = start;
 272                 if (physnodes[i].end > end)
 273                         physnodes[i].end = end;
 274                 ret++;
 275         }
 276
 277         /*
 278          * If no physical topology was detected, a single node is faked to cover
 279          * the entire address space.
 280          */
 281         if (!ret) {
 282                 physnodes[ret].start = start;
 283                 physnodes[ret].end = end;
 284                 ret = 1;
 285         }
 286         return ret;
 287 }
 288
 289 static void __init fake_physnodes(int acpi, int amd, int nr_nodes)
 290 {
 291         int i;
 292
 293         BUG_ON(acpi && amd);
 294 #ifdef CONFIG_ACPI_NUMA
 295         if (acpi)
 296                 acpi_fake_nodes(nodes, nr_nodes);
 297 #endif
 298 #ifdef CONFIG_AMD_NUMA
 299         if (amd)
 300                 amd_fake_nodes(nodes, nr_nodes);
 301 #endif
 302         if (!acpi && !amd)
 303                 for (i = 0; i < nr_cpu_ids; i++)
 304                         numa_set_node(i, 0);
 305 }
 306
 307 /*
 308  * Setups up nid to range from addr to addr + size.  If the end
 309  * boundary is greater than max_addr, then max_addr is used instead.
 310  * The return value is 0 if there is additional memory left for
 311  * allocation past addr and -1 otherwise.  addr is adjusted to be at
 312  * the end of the node.
 313  */
 314 static int __init setup_node_range(int nid, u64 *addr, u64 size, u64 max_addr)
 315 {
 316         int ret = 0;
 317         nodes[nid].start = *addr;
 318         *addr += size;
 319         if (*addr >= max_addr) {
 320                 *addr = max_addr;
 321                 ret = -1;
 322         }
 323         nodes[nid].end = *addr;
 324         node_set(nid, node_possible_map);
 325         printk(KERN_INFO "Faking node %d at %016Lx-%016Lx (%LuMB)\n", nid,
 326                nodes[nid].start, nodes[nid].end,
 327                (nodes[nid].end - nodes[nid].start) >> 20);
 328         return ret;
 329 }
 330
 331 /*
 332  * Sets up nr_nodes fake nodes interleaved over physical nodes ranging from addr
 333  * to max_addr.  The return value is the number of nodes allocated.
 334  */
 335 static int __init split_nodes_interleave(u64 addr, u64 max_addr, int nr_nodes)
 336 {
 337         nodemask_t physnode_mask = NODE_MASK_NONE;
 338         u64 size;
 339         int big;
 340         int ret = 0;
 341         int i;
 342
 343         if (nr_nodes <= 0)
 344                 return -1;
 345         if (nr_nodes > MAX_NUMNODES) {
 346                 pr_info("numa=fake=%d too large, reducing to %d\n",
 347                         nr_nodes, MAX_NUMNODES);
 348                 nr_nodes = MAX_NUMNODES;
 349         }
 350
 351         size = (max_addr - addr - memblock_x86_hole_size(addr, max_addr)) / nr_nodes;
 352         /*
 353          * Calculate the number of big nodes that can be allocated as a result
 354          * of consolidating the remainder.
 355          */
 356         big = ((size & ~FAKE_NODE_MIN_HASH_MASK) * nr_nodes) /
 357                 FAKE_NODE_MIN_SIZE;
 358
 359         size &= FAKE_NODE_MIN_HASH_MASK;
 360         if (!size) {
 361                 pr_err("Not enough memory for each node.  "
 362                         "NUMA emulation disabled.\n");
 363                 return -1;
 364         }
 365
 366         for (i = 0; i < MAX_NUMNODES; i++)
 367                 if (physnodes[i].start != physnodes[i].end)
 368                         node_set(i, physnode_mask);
 369
 370         /*
 371          * Continue to fill physical nodes with fake nodes until there is no
 372          * memory left on any of them.
 373          */
 374         while (nodes_weight(physnode_mask)) {
 375                 for_each_node_mask(i, physnode_mask) {
 376                         u64 end = physnodes[i].start + size;
 377                         u64 dma32_end = PFN_PHYS(MAX_DMA32_PFN);
 378
 379                         if (ret < big)
 380                                 end += FAKE_NODE_MIN_SIZE;
 381
 382                         /*
 383                          * Continue to add memory to this fake node if its
 384                          * non-reserved memory is less than the per-node size.
 385                          */
 386                         while (end - physnodes[i].start -
 387                                 memblock_x86_hole_size(physnodes[i].start, end) < size) {
 388                                 end += FAKE_NODE_MIN_SIZE;
 389                                 if (end > physnodes[i].end) {
 390                                         end = physnodes[i].end;
 391                                         break;
 392                                 }
 393                         }
 394
 395                         /*
 396                          * If there won't be at least FAKE_NODE_MIN_SIZE of
 397                          * non-reserved memory in ZONE_DMA32 for the next node,
 398                          * this one must extend to the boundary.
 399                          */
 400                         if (end < dma32_end && dma32_end - end -
 401                             memblock_x86_hole_size(end, dma32_end) < FAKE_NODE_MIN_SIZE)
 402                                 end = dma32_end;
 403
 404                         /*
 405                          * If there won't be enough non-reserved memory for the
 406                          * next node, this one must extend to the end of the
 407                          * physical node.
 408                          */
 409                         if (physnodes[i].end - end -
 410                             memblock_x86_hole_size(end, physnodes[i].end) < size)
 411                                 end = physnodes[i].end;
 412
 413                         /*
 414                          * Avoid allocating more nodes than requested, which can
 415                          * happen as a result of rounding down each node's size
 416                          * to FAKE_NODE_MIN_SIZE.
 417                          */
 418                         if (nodes_weight(physnode_mask) + ret >= nr_nodes)
 419                                 end = physnodes[i].end;
 420
 421                         if (setup_node_range(ret++, &physnodes[i].start,
 422                                                 end - physnodes[i].start,
 423                                                 physnodes[i].end) < 0)
 424                                 node_clear(i, physnode_mask);
 425                 }
 426         }
 427         return ret;
 428 }
 429
 430 /*
 431  * Returns the end address of a node so that there is at least `size' amount of
 432  * non-reserved memory or `max_addr' is reached.
 433  */
 434 static u64 __init find_end_of_node(u64 start, u64 max_addr, u64 size)
 435 {
 436         u64 end = start + size;
 437
 438         while (end - start - memblock_x86_hole_size(start, end) < size) {
 439                 end += FAKE_NODE_MIN_SIZE;
 440                 if (end > max_addr) {
 441                         end = max_addr;
 442                         break;
 443                 }
 444         }
 445         return end;
 446 }
 447
 448 /*
 449  * Sets up fake nodes of `size' interleaved over physical nodes ranging from
 450  * `addr' to `max_addr'.  The return value is the number of nodes allocated.
 451  */
 452 static int __init split_nodes_size_interleave(u64 addr, u64 max_addr, u64 size)
 453 {
 454         nodemask_t physnode_mask = NODE_MASK_NONE;
 455         u64 min_size;
 456         int ret = 0;
 457         int i;
 458
 459         if (!size)
 460                 return -1;
 461         /*
 462          * The limit on emulated nodes is MAX_NUMNODES, so the size per node is
 463          * increased accordingly if the requested size is too small.  This
 464          * creates a uniform distribution of node sizes across the entire
 465          * machine (but not necessarily over physical nodes).
 466          */
 467         min_size = (max_addr - addr - memblock_x86_hole_size(addr, max_addr)) /
 468                                                 MAX_NUMNODES;
 469         min_size = max(min_size, FAKE_NODE_MIN_SIZE);
 470         if ((min_size & FAKE_NODE_MIN_HASH_MASK) < min_size)
 471                 min_size = (min_size + FAKE_NODE_MIN_SIZE) &
 472                                                 FAKE_NODE_MIN_HASH_MASK;
 473         if (size < min_size) {
 474                 pr_err("Fake node size %LuMB too small, increasing to %LuMB\n",
 475                         size >> 20, min_size >> 20);
 476                 size = min_size;
 477         }
 478         size &= FAKE_NODE_MIN_HASH_MASK;
 479
 480         for (i = 0; i < MAX_NUMNODES; i++)
 481                 if (physnodes[i].start != physnodes[i].end)
 482                         node_set(i, physnode_mask);
 483         /*
 484          * Fill physical nodes with fake nodes of size until there is no memory
 485          * left on any of them.
 486          */
 487         while (nodes_weight(physnode_mask)) {
 488                 for_each_node_mask(i, physnode_mask) {
 489                         u64 dma32_end = MAX_DMA32_PFN << PAGE_SHIFT;
 490                         u64 end;
 491
 492                         end = find_end_of_node(physnodes[i].start,
 493                                                 physnodes[i].end, size);
 494                         /*
 495                          * If there won't be at least FAKE_NODE_MIN_SIZE of
 496                          * non-reserved memory in ZONE_DMA32 for the next node,
 497                          * this one must extend to the boundary.
 498                          */
 499                         if (end < dma32_end && dma32_end - end -
 500                             memblock_x86_hole_size(end, dma32_end) < FAKE_NODE_MIN_SIZE)
 501                                 end = dma32_end;
 502
 503                         /*
 504                          * If there won't be enough non-reserved memory for the
 505                          * next node, this one must extend to the end of the
 506                          * physical node.
 507                          */
 508                         if (physnodes[i].end - end -
 509                             memblock_x86_hole_size(end, physnodes[i].end) < size)
 510                                 end = physnodes[i].end;
 511
 512                         /*
 513                          * Setup the fake node that will be allocated as bootmem
 514                          * later.  If setup_node_range() returns non-zero, there
 515                          * is no more memory available on this physical node.
 516                          */
 517                         if (setup_node_range(ret++, &physnodes[i].start,
 518                                                 end - physnodes[i].start,
 519                                                 physnodes[i].end) < 0)
 520                                 node_clear(i, physnode_mask);
 521                 }
 522         }
 523         return ret;
 524 }
 525
 526 /*
 527  * Sets up the system RAM area from start_pfn to last_pfn according to the
 528  * numa=fake command-line option.
 529  */
 530 static int __init numa_emulation(unsigned long start_pfn,
 531                         unsigned long last_pfn, int acpi, int amd)
 532 {
 533         u64 addr = start_pfn << PAGE_SHIFT;
 534         u64 max_addr = last_pfn << PAGE_SHIFT;
 535         int num_nodes;
 536         int i;
 537
 538         /*
 539          * If the numa=fake command-line contains a 'M' or 'G', it represents
 540          * the fixed node size.  Otherwise, if it is just a single number N,
 541          * split the system RAM into N fake nodes.
 542          */
 543         if (strchr(cmdline, 'M') || strchr(cmdline, 'G')) {
 544                 u64 size;
 545
 546                 size = memparse(cmdline, &cmdline);
 547                 num_nodes = split_nodes_size_interleave(addr, max_addr, size);
 548         } else {
 549                 unsigned long n;
 550
 551                 n = simple_strtoul(cmdline, NULL, 0);
 552                 num_nodes = split_nodes_interleave(addr, max_addr, n);
 553         }
 554
 555         if (num_nodes < 0)
 556                 return num_nodes;
 557         memnode_shift = compute_hash_shift(nodes, num_nodes, NULL);
 558         if (memnode_shift < 0) {
 559                 memnode_shift = 0;
 560                 printk(KERN_ERR "No NUMA hash function found.  NUMA emulation "
 561                        "disabled.\n");
 562                 return -1;
 563         }
 564
 565         /*
 566          * We need to vacate all active ranges that may have been registered for
 567          * the e820 memory map.
 568          */
 569         remove_all_active_ranges();
 570         for_each_node_mask(i, node_possible_map)
 571                 memblock_x86_register_active_regions(i, nodes[i].start >> PAGE_SHIFT,
 572                                                 nodes[i].end >> PAGE_SHIFT);
 573         init_memory_mapping_high();
 574         for_each_node_mask(i, node_possible_map)
 575                 setup_node_bootmem(i, nodes[i].start, nodes[i].end);
 576         setup_physnodes(addr, max_addr, acpi, amd);
 577         fake_physnodes(acpi, amd, num_nodes);
 578         numa_init_array();
 579         return 0;
 580 }
 581 #endif /* CONFIG_NUMA_EMU */
 582
 583 void __init initmem_init(void)
 584 {
 585         int acpi = 0, amd = 0;
 586         int i;
 587
 588 #ifdef CONFIG_ACPI_NUMA
 589         /*
 590          * Parse SRAT to discover nodes.
 591          */
 592         acpi = !x86_acpi_numa_init();
 593 #endif
 594
 595 #ifdef CONFIG_AMD_NUMA
 596         if (!acpi)
 597                 amd = !amd_numa_init();
 598 #endif
 599
 600         nodes_clear(node_possible_map);
 601         nodes_clear(node_online_map);
 602
 603 #ifdef CONFIG_NUMA_EMU
 604         setup_physnodes(0, max_pfn << PAGE_SHIFT, acpi, amd);
 605         if (cmdline && !numa_emulation(0, max_pfn, acpi, amd))
 606                 return;
 607         setup_physnodes(0, max_pfn << PAGE_SHIFT, acpi, amd);
 608         nodes_clear(node_possible_map);
 609         nodes_clear(node_online_map);
 610 #endif
 611
 612 #ifdef CONFIG_ACPI_NUMA
 613         if (!numa_off && acpi && !acpi_scan_nodes())
 614                 return;
 615         nodes_clear(node_possible_map);
 616         nodes_clear(node_online_map);
 617 #endif
 618
 619 #ifdef CONFIG_AMD_NUMA
 620         if (!numa_off && amd && !amd_scan_nodes())
 621                 return;
 622         nodes_clear(node_possible_map);
 623         nodes_clear(node_online_map);
 624 #endif
 625         printk(KERN_INFO "%s\n",
 626                numa_off ? "NUMA turned off" : "No NUMA configuration found");
 627
 628         printk(KERN_INFO "Faking a node at %016lx-%016lx\n",
 629                0LU, max_pfn << PAGE_SHIFT);
 630         /* setup dummy node covering all memory */
 631         memnode_shift = 63;
 632         memnodemap = memnode.embedded_map;
 633         memnodemap[0] = 0;
 634         node_set_online(0);
 635         node_set(0, node_possible_map);
 636         for (i = 0; i < MAX_LOCAL_APIC; i++)
 637                 set_apicid_to_node(i, NUMA_NO_NODE);
 638         memblock_x86_register_active_regions(0, 0, max_pfn);
 639         init_memory_mapping_high();
 640         setup_node_bootmem(0, 0, max_pfn << PAGE_SHIFT);
 641         numa_init_array();
 642 }
 643
 644 unsigned long __init numa_free_all_bootmem(void)
 645 {
 646         unsigned long pages = 0;
 647         int i;
 648
 649         for_each_online_node(i)
 650                 pages += free_all_bootmem_node(NODE_DATA(i));
 651
 652         pages += free_all_memory_core_early(MAX_NUMNODES);
 653
 654         return pages;
 655 }
 656
 657 int __cpuinit numa_cpu_node(int cpu)
 658 {
 659         int apicid = early_per_cpu(x86_cpu_to_apicid, cpu);
 660
 661         if (apicid != BAD_APICID)
 662                 return __apicid_to_node[apicid];
 663         return NUMA_NO_NODE;
 664 }
 665
 666 /*
 667  * UGLINESS AHEAD: Currently, CONFIG_NUMA_EMU is 64bit only and makes use
 668  * of 64bit specific data structures.  The distinction is artificial and
 669  * should be removed.  numa_{add|remove}_cpu() are implemented in numa.c
 670  * for both 32 and 64bit when CONFIG_NUMA_EMU is disabled but here when
 671  * enabled.
 672  *
 673  * NUMA emulation is planned to be made generic and the following and other
 674  * related code should be moved to numa.c.
 675  */
 676 #ifdef CONFIG_NUMA_EMU
 677 # ifndef CONFIG_DEBUG_PER_CPU_MAPS
 678 void __cpuinit numa_add_cpu(int cpu)
 679 {
 680         unsigned long addr;
 681         int physnid, nid;
 682
 683         nid = numa_cpu_node(cpu);
 684         if (nid == NUMA_NO_NODE)
 685                 nid = early_cpu_to_node(cpu);
 686         BUG_ON(nid == NUMA_NO_NODE || !node_online(nid));
 687
 688         /*
 689          * Use the starting address of the emulated node to find which physical
 690          * node it is allocated on.
 691          */
 692         addr = node_start_pfn(nid) << PAGE_SHIFT;
 693         for (physnid = 0; physnid < MAX_NUMNODES; physnid++)
 694                 if (addr >= physnodes[physnid].start &&
 695                     addr < physnodes[physnid].end)
 696                         break;
 697
 698         /*
 699          * Map the cpu to each emulated node that is allocated on the physical
 700          * node of the cpu's apic id.
 701          */
 702         for_each_online_node(nid) {
 703                 addr = node_start_pfn(nid) << PAGE_SHIFT;
 704                 if (addr >= physnodes[physnid].start &&
 705                     addr < physnodes[physnid].end)
 706                         cpumask_set_cpu(cpu, node_to_cpumask_map[nid]);
 707         }
 708 }
 709
 710 void __cpuinit numa_remove_cpu(int cpu)
 711 {
 712         int i;
 713
 714         for_each_online_node(i)
 715                 cpumask_clear_cpu(cpu, node_to_cpumask_map[i]);
 716 }
 717 # else  /* !CONFIG_DEBUG_PER_CPU_MAPS */
 718 static void __cpuinit numa_set_cpumask(int cpu, int enable)
 719 {
 720         int node = early_cpu_to_node(cpu);
 721         struct cpumask *mask;
 722         int i;
 723
 724         if (node == NUMA_NO_NODE) {
 725                 /* early_cpu_to_node() already emits a warning and trace */
 726                 return;
 727         }
 728         for_each_online_node(i) {
 729                 unsigned long addr;
 730
 731                 addr = node_start_pfn(i) << PAGE_SHIFT;
 732                 if (addr < physnodes[node].start ||
 733                                         addr >= physnodes[node].end)
 734                         continue;
 735                 mask = debug_cpumask_set_cpu(cpu, enable);
 736                 if (!mask)
 737                         return;
 738
 739                 if (enable)
 740                         cpumask_set_cpu(cpu, mask);
 741                 else
 742                         cpumask_clear_cpu(cpu, mask);
 743         }
 744 }
 745
 746 void __cpuinit numa_add_cpu(int cpu)
 747 {
 748         numa_set_cpumask(cpu, 1);
 749 }
 750
 751 void __cpuinit numa_remove_cpu(int cpu)
 752 {
 753         numa_set_cpumask(cpu, 0);
 754 }
 755 # endif /* !CONFIG_DEBUG_PER_CPU_MAPS */
 756 #endif  /* CONFIG_NUMA_EMU */