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/mmzone.h>
  11 #include <linux/ctype.h>
  12 #include <linux/module.h>
  13 #include <linux/nodemask.h>
  14 #include <linux/sched.h>
  15
  16 #include <asm/e820.h>
  17 #include <asm/proto.h>
  18 #include <asm/dma.h>
  19 #include <asm/numa.h>
  20 #include <asm/acpi.h>
  21 #include <asm/k8.h>
  22
  23 #ifndef Dprintk
  24 #define Dprintk(x...)
  25 #endif
  26
  27 struct pglist_data *node_data[MAX_NUMNODES] __read_mostly;
  28 EXPORT_SYMBOL(node_data);
  29
  30 bootmem_data_t plat_node_bdata[MAX_NUMNODES];
  31
  32 struct memnode memnode;
  33
  34 u16 x86_cpu_to_node_map_init[NR_CPUS] = {
  35         [0 ... NR_CPUS-1] = NUMA_NO_NODE
  36 };
  37 void *x86_cpu_to_node_map_early_ptr;
  38 DEFINE_PER_CPU(u16, x86_cpu_to_node_map) = NUMA_NO_NODE;
  39 EXPORT_PER_CPU_SYMBOL(x86_cpu_to_node_map);
  40 #ifdef  CONFIG_DEBUG_PER_CPU_MAPS
  41 EXPORT_SYMBOL(x86_cpu_to_node_map_early_ptr);
  42 #endif
  43
  44 u16 apicid_to_node[MAX_LOCAL_APIC] __cpuinitdata = {
  45         [0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE
  46 };
  47
  48 cpumask_t node_to_cpumask_map[MAX_NUMNODES] __read_mostly;
  49 EXPORT_SYMBOL(node_to_cpumask_map);
  50
  51 int numa_off __initdata;
  52 unsigned long __initdata nodemap_addr;
  53 unsigned long __initdata nodemap_size;
  54
  55 /*
  56  * Given a shift value, try to populate memnodemap[]
  57  * Returns :
  58  * 1 if OK
  59  * 0 if memnodmap[] too small (of shift too small)
  60  * -1 if node overlap or lost ram (shift too big)
  61  */
  62 static int __init populate_memnodemap(const struct bootnode *nodes,
  63                                       int numnodes, int shift)
  64 {
  65         unsigned long addr, end;
  66         int i, res = -1;
  67
  68         memset(memnodemap, 0xff, memnodemapsize);
  69         for (i = 0; i < numnodes; i++) {
  70                 addr = nodes[i].start;
  71                 end = nodes[i].end;
  72                 if (addr >= end)
  73                         continue;
  74                 if ((end >> shift) >= memnodemapsize)
  75                         return 0;
  76                 do {
  77                         if (memnodemap[addr >> shift] != 0xff)
  78                                 return -1;
  79                         memnodemap[addr >> shift] = i;
  80                         addr += (1UL << shift);
  81                 } while (addr < end);
  82                 res = 1;
  83         }
  84         return res;
  85 }
  86
  87 static int __init allocate_cachealigned_memnodemap(void)
  88 {
  89         unsigned long pad, pad_addr;
  90
  91         memnodemap = memnode.embedded_map;
  92         if (memnodemapsize <= ARRAY_SIZE(memnode.embedded_map))
  93                 return 0;
  94
  95         pad = L1_CACHE_BYTES - 1;
  96         pad_addr = 0x8000;
  97         nodemap_size = pad + memnodemapsize;
  98         nodemap_addr = find_e820_area(pad_addr, end_pfn<<PAGE_SHIFT,
  99                                       nodemap_size);
 100         if (nodemap_addr == -1UL) {
 101                 printk(KERN_ERR
 102                        "NUMA: Unable to allocate Memory to Node hash map\n");
 103                 nodemap_addr = nodemap_size = 0;
 104                 return -1;
 105         }
 106         pad_addr = (nodemap_addr + pad) & ~pad;
 107         memnodemap = phys_to_virt(pad_addr);
 108         reserve_early(nodemap_addr, nodemap_addr + nodemap_size);
 109
 110         printk(KERN_DEBUG "NUMA: Allocated memnodemap from %lx - %lx\n",
 111                nodemap_addr, nodemap_addr + nodemap_size);
 112         return 0;
 113 }
 114
 115 /*
 116  * The LSB of all start and end addresses in the node map is the value of the
 117  * maximum possible shift.
 118  */
 119 static int __init extract_lsb_from_nodes(const struct bootnode *nodes,
 120                                          int numnodes)
 121 {
 122         int i, nodes_used = 0;
 123         unsigned long start, end;
 124         unsigned long bitfield = 0, memtop = 0;
 125
 126         for (i = 0; i < numnodes; i++) {
 127                 start = nodes[i].start;
 128                 end = nodes[i].end;
 129                 if (start >= end)
 130                         continue;
 131                 bitfield |= start;
 132                 nodes_used++;
 133                 if (end > memtop)
 134                         memtop = end;
 135         }
 136         if (nodes_used <= 1)
 137                 i = 63;
 138         else
 139                 i = find_first_bit(&bitfield, sizeof(unsigned long)*8);
 140         memnodemapsize = (memtop >> i)+1;
 141         return i;
 142 }
 143
 144 int __init compute_hash_shift(struct bootnode *nodes, int numnodes)
 145 {
 146         int shift;
 147
 148         shift = extract_lsb_from_nodes(nodes, numnodes);
 149         if (allocate_cachealigned_memnodemap())
 150                 return -1;
 151         printk(KERN_DEBUG "NUMA: Using %d for the hash shift.\n",
 152                 shift);
 153
 154         if (populate_memnodemap(nodes, numnodes, shift) != 1) {
 155                 printk(KERN_INFO "Your memory is not aligned you need to "
 156                        "rebuild your kernel with a bigger NODEMAPSIZE "
 157                        "shift=%d\n", shift);
 158                 return -1;
 159         }
 160         return shift;
 161 }
 162
 163 int early_pfn_to_nid(unsigned long pfn)
 164 {
 165         return phys_to_nid(pfn << PAGE_SHIFT);
 166 }
 167
 168 static void * __init early_node_mem(int nodeid, unsigned long start,
 169                                     unsigned long end, unsigned long size)
 170 {
 171         unsigned long mem = find_e820_area(start, end, size);
 172         void *ptr;
 173
 174         if (mem != -1L)
 175                 return __va(mem);
 176         ptr = __alloc_bootmem_nopanic(size,
 177                                 SMP_CACHE_BYTES, __pa(MAX_DMA_ADDRESS));
 178         if (ptr == NULL) {
 179                 printk(KERN_ERR "Cannot find %lu bytes in node %d\n",
 180                        size, nodeid);
 181                 return NULL;
 182         }
 183         return ptr;
 184 }
 185
 186 /* Initialize bootmem allocator for a node */
 187 void __init setup_node_bootmem(int nodeid, unsigned long start,
 188                                unsigned long end)
 189 {
 190         unsigned long start_pfn, end_pfn, bootmap_pages, bootmap_size;
 191         unsigned long bootmap_start, nodedata_phys;
 192         void *bootmap;
 193         const int pgdat_size = round_up(sizeof(pg_data_t), PAGE_SIZE);
 194
 195         start = round_up(start, ZONE_ALIGN);
 196
 197         printk(KERN_INFO "Bootmem setup node %d %016lx-%016lx\n", nodeid,
 198                start, end);
 199
 200         start_pfn = start >> PAGE_SHIFT;
 201         end_pfn = end >> PAGE_SHIFT;
 202
 203         node_data[nodeid] = early_node_mem(nodeid, start, end, pgdat_size);
 204         if (node_data[nodeid] == NULL)
 205                 return;
 206         nodedata_phys = __pa(node_data[nodeid]);
 207
 208         memset(NODE_DATA(nodeid), 0, sizeof(pg_data_t));
 209         NODE_DATA(nodeid)->bdata = &plat_node_bdata[nodeid];
 210         NODE_DATA(nodeid)->node_start_pfn = start_pfn;
 211         NODE_DATA(nodeid)->node_spanned_pages = end_pfn - start_pfn;
 212
 213         /* Find a place for the bootmem map */
 214         bootmap_pages = bootmem_bootmap_pages(end_pfn - start_pfn);
 215         bootmap_start = round_up(nodedata_phys + pgdat_size, PAGE_SIZE);
 216         bootmap = early_node_mem(nodeid, bootmap_start, end,
 217                                         bootmap_pages<<PAGE_SHIFT);
 218         if (bootmap == NULL)  {
 219                 if (nodedata_phys < start || nodedata_phys >= end)
 220                         free_bootmem((unsigned long)node_data[nodeid],
 221                                      pgdat_size);
 222                 node_data[nodeid] = NULL;
 223                 return;
 224         }
 225         bootmap_start = __pa(bootmap);
 226         Dprintk("bootmap start %lu pages %lu\n", bootmap_start, bootmap_pages);
 227
 228         bootmap_size = init_bootmem_node(NODE_DATA(nodeid),
 229                                          bootmap_start >> PAGE_SHIFT,
 230                                          start_pfn, end_pfn);
 231
 232         free_bootmem_with_active_regions(nodeid, end);
 233
 234         reserve_bootmem_node(NODE_DATA(nodeid), nodedata_phys, pgdat_size);
 235         reserve_bootmem_node(NODE_DATA(nodeid), bootmap_start,
 236                              bootmap_pages<<PAGE_SHIFT);
 237 #ifdef CONFIG_ACPI_NUMA
 238         srat_reserve_add_area(nodeid);
 239 #endif
 240         node_set_online(nodeid);
 241 }
 242
 243 #ifdef CONFIG_FLAT_NODE_MEM_MAP
 244 /* Initialize final allocator for a zone */
 245 static void __init flat_setup_node_zones(int nodeid)
 246 {
 247         unsigned long start_pfn, end_pfn, memmapsize, limit;
 248
 249         start_pfn = node_start_pfn(nodeid);
 250         end_pfn = node_end_pfn(nodeid);
 251
 252         Dprintk(KERN_INFO "Setting up memmap for node %d %lx-%lx\n",
 253                 nodeid, start_pfn, end_pfn);
 254
 255         /*
 256          * Try to allocate mem_map at end to not fill up precious <4GB
 257          * memory.
 258          */
 259         memmapsize = sizeof(struct page) * (end_pfn-start_pfn);
 260         limit = end_pfn << PAGE_SHIFT;
 261
 262         NODE_DATA(nodeid)->node_mem_map =
 263                 __alloc_bootmem_core(NODE_DATA(nodeid)->bdata,
 264                                      memmapsize, SMP_CACHE_BYTES,
 265                                      round_down(limit - memmapsize, PAGE_SIZE),
 266                                      limit);
 267 }
 268 #else
 269 #define flat_setup_node_zones(i) do {} while (0)
 270 #endif
 271
 272 /*
 273  * There are unfortunately some poorly designed mainboards around that
 274  * only connect memory to a single CPU. This breaks the 1:1 cpu->node
 275  * mapping. To avoid this fill in the mapping for all possible CPUs,
 276  * as the number of CPUs is not known yet. We round robin the existing
 277  * nodes.
 278  */
 279 void __init numa_init_array(void)
 280 {
 281         int rr, i;
 282
 283         rr = first_node(node_online_map);
 284         for (i = 0; i < NR_CPUS; i++) {
 285                 if (cpu_to_node(i) != NUMA_NO_NODE)
 286                         continue;
 287                 numa_set_node(i, rr);
 288                 rr = next_node(rr, node_online_map);
 289                 if (rr == MAX_NUMNODES)
 290                         rr = first_node(node_online_map);
 291         }
 292 }
 293
 294 #ifdef CONFIG_NUMA_EMU
 295 /* Numa emulation */
 296 char *cmdline __initdata;
 297
 298 /*
 299  * Setups up nid to range from addr to addr + size.  If the end
 300  * boundary is greater than max_addr, then max_addr is used instead.
 301  * The return value is 0 if there is additional memory left for
 302  * allocation past addr and -1 otherwise.  addr is adjusted to be at
 303  * the end of the node.
 304  */
 305 static int __init setup_node_range(int nid, struct bootnode *nodes, u64 *addr,
 306                                    u64 size, u64 max_addr)
 307 {
 308         int ret = 0;
 309
 310         nodes[nid].start = *addr;
 311         *addr += size;
 312         if (*addr >= max_addr) {
 313                 *addr = max_addr;
 314                 ret = -1;
 315         }
 316         nodes[nid].end = *addr;
 317         node_set(nid, node_possible_map);
 318         printk(KERN_INFO "Faking node %d at %016Lx-%016Lx (%LuMB)\n", nid,
 319                nodes[nid].start, nodes[nid].end,
 320                (nodes[nid].end - nodes[nid].start) >> 20);
 321         return ret;
 322 }
 323
 324 /*
 325  * Splits num_nodes nodes up equally starting at node_start.  The return value
 326  * is the number of nodes split up and addr is adjusted to be at the end of the
 327  * last node allocated.
 328  */
 329 static int __init split_nodes_equally(struct bootnode *nodes, u64 *addr,
 330                                       u64 max_addr, int node_start,
 331                                       int num_nodes)
 332 {
 333         unsigned int big;
 334         u64 size;
 335         int i;
 336
 337         if (num_nodes <= 0)
 338                 return -1;
 339         if (num_nodes > MAX_NUMNODES)
 340                 num_nodes = MAX_NUMNODES;
 341         size = (max_addr - *addr - e820_hole_size(*addr, max_addr)) /
 342                num_nodes;
 343         /*
 344          * Calculate the number of big nodes that can be allocated as a result
 345          * of consolidating the leftovers.
 346          */
 347         big = ((size & ~FAKE_NODE_MIN_HASH_MASK) * num_nodes) /
 348               FAKE_NODE_MIN_SIZE;
 349
 350         /* Round down to nearest FAKE_NODE_MIN_SIZE. */
 351         size &= FAKE_NODE_MIN_HASH_MASK;
 352         if (!size) {
 353                 printk(KERN_ERR "Not enough memory for each node.  "
 354                        "NUMA emulation disabled.\n");
 355                 return -1;
 356         }
 357
 358         for (i = node_start; i < num_nodes + node_start; i++) {
 359                 u64 end = *addr + size;
 360
 361                 if (i < big)
 362                         end += FAKE_NODE_MIN_SIZE;
 363                 /*
 364                  * The final node can have the remaining system RAM.  Other
 365                  * nodes receive roughly the same amount of available pages.
 366                  */
 367                 if (i == num_nodes + node_start - 1)
 368                         end = max_addr;
 369                 else
 370                         while (end - *addr - e820_hole_size(*addr, end) <
 371                                size) {
 372                                 end += FAKE_NODE_MIN_SIZE;
 373                                 if (end > max_addr) {
 374                                         end = max_addr;
 375                                         break;
 376                                 }
 377                         }
 378                 if (setup_node_range(i, nodes, addr, end - *addr, max_addr) < 0)
 379                         break;
 380         }
 381         return i - node_start + 1;
 382 }
 383
 384 /*
 385  * Splits the remaining system RAM into chunks of size.  The remaining memory is
 386  * always assigned to a final node and can be asymmetric.  Returns the number of
 387  * nodes split.
 388  */
 389 static int __init split_nodes_by_size(struct bootnode *nodes, u64 *addr,
 390                                       u64 max_addr, int node_start, u64 size)
 391 {
 392         int i = node_start;
 393         size = (size << 20) & FAKE_NODE_MIN_HASH_MASK;
 394         while (!setup_node_range(i++, nodes, addr, size, max_addr))
 395                 ;
 396         return i - node_start;
 397 }
 398
 399 /*
 400  * Sets up the system RAM area from start_pfn to end_pfn according to the
 401  * numa=fake command-line option.
 402  */
 403 static int __init numa_emulation(unsigned long start_pfn, unsigned long end_pfn)
 404 {
 405         struct bootnode nodes[MAX_NUMNODES];
 406         u64 size, addr = start_pfn << PAGE_SHIFT;
 407         u64 max_addr = end_pfn << PAGE_SHIFT;
 408         int num_nodes = 0, num = 0, coeff_flag, coeff = -1, i;
 409
 410         memset(&nodes, 0, sizeof(nodes));
 411         /*
 412          * If the numa=fake command-line is just a single number N, split the
 413          * system RAM into N fake nodes.
 414          */
 415         if (!strchr(cmdline, '*') && !strchr(cmdline, ',')) {
 416                 long n = simple_strtol(cmdline, NULL, 0);
 417
 418                 num_nodes = split_nodes_equally(nodes, &addr, max_addr, 0, n);
 419                 if (num_nodes < 0)
 420                         return num_nodes;
 421                 goto out;
 422         }
 423
 424         /* Parse the command line. */
 425         for (coeff_flag = 0; ; cmdline++) {
 426                 if (*cmdline && isdigit(*cmdline)) {
 427                         num = num * 10 + *cmdline - '0';
 428                         continue;
 429                 }
 430                 if (*cmdline == '*') {
 431                         if (num > 0)
 432                                 coeff = num;
 433                         coeff_flag = 1;
 434                 }
 435                 if (!*cmdline || *cmdline == ',') {
 436                         if (!coeff_flag)
 437                                 coeff = 1;
 438                         /*
 439                          * Round down to the nearest FAKE_NODE_MIN_SIZE.
 440                          * Command-line coefficients are in megabytes.
 441                          */
 442                         size = ((u64)num << 20) & FAKE_NODE_MIN_HASH_MASK;
 443                         if (size)
 444                                 for (i = 0; i < coeff; i++, num_nodes++)
 445                                         if (setup_node_range(num_nodes, nodes,
 446                                                 &addr, size, max_addr) < 0)
 447                                                 goto done;
 448                         if (!*cmdline)
 449                                 break;
 450                         coeff_flag = 0;
 451                         coeff = -1;
 452                 }
 453                 num = 0;
 454         }
 455 done:
 456         if (!num_nodes)
 457                 return -1;
 458         /* Fill remainder of system RAM, if appropriate. */
 459         if (addr < max_addr) {
 460                 if (coeff_flag && coeff < 0) {
 461                         /* Split remaining nodes into num-sized chunks */
 462                         num_nodes += split_nodes_by_size(nodes, &addr, max_addr,
 463                                                          num_nodes, num);
 464                         goto out;
 465                 }
 466                 switch (*(cmdline - 1)) {
 467                 case '*':
 468                         /* Split remaining nodes into coeff chunks */
 469                         if (coeff <= 0)
 470                                 break;
 471                         num_nodes += split_nodes_equally(nodes, &addr, max_addr,
 472                                                          num_nodes, coeff);
 473                         break;
 474                 case ',':
 475                         /* Do not allocate remaining system RAM */
 476                         break;
 477                 default:
 478                         /* Give one final node */
 479                         setup_node_range(num_nodes, nodes, &addr,
 480                                          max_addr - addr, max_addr);
 481                         num_nodes++;
 482                 }
 483         }
 484 out:
 485         memnode_shift = compute_hash_shift(nodes, num_nodes);
 486         if (memnode_shift < 0) {
 487                 memnode_shift = 0;
 488                 printk(KERN_ERR "No NUMA hash function found.  NUMA emulation "
 489                        "disabled.\n");
 490                 return -1;
 491         }
 492
 493         /*
 494          * We need to vacate all active ranges that may have been registered by
 495          * SRAT and set acpi_numa to -1 so that srat_disabled() always returns
 496          * true.  NUMA emulation has succeeded so we will not scan ACPI nodes.
 497          */
 498         remove_all_active_ranges();
 499 #ifdef CONFIG_ACPI_NUMA
 500         acpi_numa = -1;
 501 #endif
 502         for_each_node_mask(i, node_possible_map) {
 503                 e820_register_active_regions(i, nodes[i].start >> PAGE_SHIFT,
 504                                                 nodes[i].end >> PAGE_SHIFT);
 505                 setup_node_bootmem(i, nodes[i].start, nodes[i].end);
 506         }
 507         acpi_fake_nodes(nodes, num_nodes);
 508         numa_init_array();
 509         return 0;
 510 }
 511 #endif /* CONFIG_NUMA_EMU */
 512
 513 void __init numa_initmem_init(unsigned long start_pfn, unsigned long end_pfn)
 514 {
 515         int i;
 516
 517         nodes_clear(node_possible_map);
 518
 519 #ifdef CONFIG_NUMA_EMU
 520         if (cmdline && !numa_emulation(start_pfn, end_pfn))
 521                 return;
 522         nodes_clear(node_possible_map);
 523 #endif
 524
 525 #ifdef CONFIG_ACPI_NUMA
 526         if (!numa_off && !acpi_scan_nodes(start_pfn << PAGE_SHIFT,
 527                                           end_pfn << PAGE_SHIFT))
 528                 return;
 529         nodes_clear(node_possible_map);
 530 #endif
 531
 532 #ifdef CONFIG_K8_NUMA
 533         if (!numa_off && !k8_scan_nodes(start_pfn<<PAGE_SHIFT,
 534                                         end_pfn<<PAGE_SHIFT))
 535                 return;
 536         nodes_clear(node_possible_map);
 537 #endif
 538         printk(KERN_INFO "%s\n",
 539                numa_off ? "NUMA turned off" : "No NUMA configuration found");
 540
 541         printk(KERN_INFO "Faking a node at %016lx-%016lx\n",
 542                start_pfn << PAGE_SHIFT,
 543                end_pfn << PAGE_SHIFT);
 544         /* setup dummy node covering all memory */
 545         memnode_shift = 63;
 546         memnodemap = memnode.embedded_map;
 547         memnodemap[0] = 0;
 548         nodes_clear(node_online_map);
 549         node_set_online(0);
 550         node_set(0, node_possible_map);
 551         for (i = 0; i < NR_CPUS; i++)
 552                 numa_set_node(i, 0);
 553         /* cpumask_of_cpu() may not be available during early startup */
 554         memset(&node_to_cpumask_map[0], 0, sizeof(node_to_cpumask_map[0]));
 555         cpu_set(0, node_to_cpumask_map[0]);
 556         e820_register_active_regions(0, start_pfn, end_pfn);
 557         setup_node_bootmem(0, start_pfn << PAGE_SHIFT, end_pfn << PAGE_SHIFT);
 558 }
 559
 560 __cpuinit void numa_add_cpu(int cpu)
 561 {
 562         set_bit(cpu, (unsigned long *)&node_to_cpumask_map[cpu_to_node(cpu)]);
 563 }
 564
 565 void __cpuinit numa_set_node(int cpu, int node)
 566 {
 567         u16 *cpu_to_node_map = x86_cpu_to_node_map_early_ptr;
 568
 569         cpu_pda(cpu)->nodenumber = node;
 570
 571         if(cpu_to_node_map)
 572                 cpu_to_node_map[cpu] = node;
 573         else if(per_cpu_offset(cpu))
 574                 per_cpu(x86_cpu_to_node_map, cpu) = node;
 575         else
 576                 Dprintk(KERN_INFO "Setting node for non-present cpu %d\n", cpu);
 577 }
 578
 579 unsigned long __init numa_free_all_bootmem(void)
 580 {
 581         unsigned long pages = 0;
 582         int i;
 583
 584         for_each_online_node(i)
 585                 pages += free_all_bootmem_node(NODE_DATA(i));
 586
 587         return pages;
 588 }
 589
 590 void __init paging_init(void)
 591 {
 592         unsigned long max_zone_pfns[MAX_NR_ZONES];
 593         int i;
 594
 595         memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
 596         max_zone_pfns[ZONE_DMA] = MAX_DMA_PFN;
 597         max_zone_pfns[ZONE_DMA32] = MAX_DMA32_PFN;
 598         max_zone_pfns[ZONE_NORMAL] = end_pfn;
 599
 600         sparse_memory_present_with_active_regions(MAX_NUMNODES);
 601         sparse_init();
 602
 603         for_each_online_node(i)
 604                 flat_setup_node_zones(i);
 605
 606         free_area_init_nodes(max_zone_pfns);
 607 }
 608
 609 static __init int numa_setup(char *opt)
 610 {
 611         if (!opt)
 612                 return -EINVAL;
 613         if (!strncmp(opt, "off", 3))
 614                 numa_off = 1;
 615 #ifdef CONFIG_NUMA_EMU
 616         if (!strncmp(opt, "fake=", 5))
 617                 cmdline = opt + 5;
 618 #endif
 619 #ifdef CONFIG_ACPI_NUMA
 620         if (!strncmp(opt, "noacpi", 6))
 621                 acpi_numa = -1;
 622         if (!strncmp(opt, "hotadd=", 7))
 623                 hotadd_percent = simple_strtoul(opt+7, NULL, 10);
 624 #endif
 625         return 0;
 626 }
 627 early_param("numa", numa_setup);
 628
 629 /*
 630  * Setup early cpu_to_node.
 631  *
 632  * Populate cpu_to_node[] only if x86_cpu_to_apicid[],
 633  * and apicid_to_node[] tables have valid entries for a CPU.
 634  * This means we skip cpu_to_node[] initialisation for NUMA
 635  * emulation and faking node case (when running a kernel compiled
 636  * for NUMA on a non NUMA box), which is OK as cpu_to_node[]
 637  * is already initialized in a round robin manner at numa_init_array,
 638  * prior to this call, and this initialization is good enough
 639  * for the fake NUMA cases.
 640  */
 641 void __init init_cpu_to_node(void)
 642 {
 643         int i;
 644
 645         for (i = 0; i < NR_CPUS; i++) {
 646                 u16 apicid = x86_cpu_to_apicid_init[i];
 647
 648                 if (apicid == BAD_APICID)
 649                         continue;
 650                 if (apicid_to_node[apicid] == NUMA_NO_NODE)
 651                         continue;
 652                 numa_set_node(i, apicid_to_node[apicid]);
 653         }
 654 }
 655
 656