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 static bootmem_data_t plat_node_bdata[MAX_NUMNODES];
  31
  32 struct memnode memnode;
  33
  34 s16 apicid_to_node[MAX_LOCAL_APIC] __cpuinitdata = {
  35         [0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE
  36 };
  37
  38 int numa_off __initdata;
  39 static unsigned long __initdata nodemap_addr;
  40 static unsigned long __initdata nodemap_size;
  41
  42 /*
  43  * Given a shift value, try to populate memnodemap[]
  44  * Returns :
  45  * 1 if OK
  46  * 0 if memnodmap[] too small (of shift too small)
  47  * -1 if node overlap or lost ram (shift too big)
  48  */
  49 static int __init populate_memnodemap(const struct bootnode *nodes,
  50                                       int numnodes, int shift, int *nodeids)
  51 {
  52         unsigned long addr, end;
  53         int i, res = -1;
  54
  55         memset(memnodemap, 0xff, sizeof(s16)*memnodemapsize);
  56         for (i = 0; i < numnodes; i++) {
  57                 addr = nodes[i].start;
  58                 end = nodes[i].end;
  59                 if (addr >= end)
  60                         continue;
  61                 if ((end >> shift) >= memnodemapsize)
  62                         return 0;
  63                 do {
  64                         if (memnodemap[addr >> shift] != NUMA_NO_NODE)
  65                                 return -1;
  66
  67                         if (!nodeids)
  68                                 memnodemap[addr >> shift] = i;
  69                         else
  70                                 memnodemap[addr >> shift] = nodeids[i];
  71
  72                         addr += (1UL << shift);
  73                 } while (addr < end);
  74                 res = 1;
  75         }
  76         return res;
  77 }
  78
  79 static int __init allocate_cachealigned_memnodemap(void)
  80 {
  81         unsigned long addr;
  82
  83         memnodemap = memnode.embedded_map;
  84         if (memnodemapsize <= ARRAY_SIZE(memnode.embedded_map))
  85                 return 0;
  86
  87         addr = 0x8000;
  88         nodemap_size = round_up(sizeof(s16) * memnodemapsize, L1_CACHE_BYTES);
  89         nodemap_addr = find_e820_area(addr, end_pfn<<PAGE_SHIFT,
  90                                       nodemap_size, L1_CACHE_BYTES);
  91         if (nodemap_addr == -1UL) {
  92                 printk(KERN_ERR
  93                        "NUMA: Unable to allocate Memory to Node hash map\n");
  94                 nodemap_addr = nodemap_size = 0;
  95                 return -1;
  96         }
  97         memnodemap = phys_to_virt(nodemap_addr);
  98         reserve_early(nodemap_addr, nodemap_addr + nodemap_size, "MEMNODEMAP");
  99
 100         printk(KERN_DEBUG "NUMA: Allocated memnodemap from %lx - %lx\n",
 101                nodemap_addr, nodemap_addr + nodemap_size);
 102         return 0;
 103 }
 104
 105 /*
 106  * The LSB of all start and end addresses in the node map is the value of the
 107  * maximum possible shift.
 108  */
 109 static int __init extract_lsb_from_nodes(const struct bootnode *nodes,
 110                                          int numnodes)
 111 {
 112         int i, nodes_used = 0;
 113         unsigned long start, end;
 114         unsigned long bitfield = 0, memtop = 0;
 115
 116         for (i = 0; i < numnodes; i++) {
 117                 start = nodes[i].start;
 118                 end = nodes[i].end;
 119                 if (start >= end)
 120                         continue;
 121                 bitfield |= start;
 122                 nodes_used++;
 123                 if (end > memtop)
 124                         memtop = end;
 125         }
 126         if (nodes_used <= 1)
 127                 i = 63;
 128         else
 129                 i = find_first_bit(&bitfield, sizeof(unsigned long)*8);
 130         memnodemapsize = (memtop >> i)+1;
 131         return i;
 132 }
 133
 134 int __init compute_hash_shift(struct bootnode *nodes, int numnodes,
 135                               int *nodeids)
 136 {
 137         int shift;
 138
 139         shift = extract_lsb_from_nodes(nodes, numnodes);
 140         if (allocate_cachealigned_memnodemap())
 141                 return -1;
 142         printk(KERN_DEBUG "NUMA: Using %d for the hash shift.\n",
 143                 shift);
 144
 145         if (populate_memnodemap(nodes, numnodes, shift, nodeids) != 1) {
 146                 printk(KERN_INFO "Your memory is not aligned you need to "
 147                        "rebuild your kernel with a bigger NODEMAPSIZE "
 148                        "shift=%d\n", shift);
 149                 return -1;
 150         }
 151         return shift;
 152 }
 153
 154 int early_pfn_to_nid(unsigned long pfn)
 155 {
 156         return phys_to_nid(pfn << PAGE_SHIFT);
 157 }
 158
 159 static void * __init early_node_mem(int nodeid, unsigned long start,
 160                                     unsigned long end, unsigned long size,
 161                                     unsigned long align)
 162 {
 163         unsigned long mem = find_e820_area(start, end, size, align);
 164         void *ptr;
 165
 166         if (mem != -1L)
 167                 return __va(mem);
 168
 169         ptr = __alloc_bootmem_nopanic(size, align, __pa(MAX_DMA_ADDRESS));
 170         if (ptr == NULL) {
 171                 printk(KERN_ERR "Cannot find %lu bytes in node %d\n",
 172                        size, nodeid);
 173                 return NULL;
 174         }
 175         return ptr;
 176 }
 177
 178 /* Initialize bootmem allocator for a node */
 179 void __init setup_node_bootmem(int nodeid, unsigned long start,
 180                                unsigned long end)
 181 {
 182         unsigned long start_pfn, last_pfn, bootmap_pages, bootmap_size;
 183         unsigned long bootmap_start, nodedata_phys;
 184         void *bootmap;
 185         const int pgdat_size = round_up(sizeof(pg_data_t), PAGE_SIZE);
 186         int nid;
 187
 188         start = round_up(start, ZONE_ALIGN);
 189
 190         printk(KERN_INFO "Bootmem setup node %d %016lx-%016lx\n", nodeid,
 191                start, end);
 192
 193         start_pfn = start >> PAGE_SHIFT;
 194         last_pfn = end >> PAGE_SHIFT;
 195
 196         node_data[nodeid] = early_node_mem(nodeid, start, end, pgdat_size,
 197                                            SMP_CACHE_BYTES);
 198         if (node_data[nodeid] == NULL)
 199                 return;
 200         nodedata_phys = __pa(node_data[nodeid]);
 201         printk(KERN_INFO "  NODE_DATA [%016lx - %016lx]\n", nodedata_phys,
 202                 nodedata_phys + pgdat_size - 1);
 203
 204         memset(NODE_DATA(nodeid), 0, sizeof(pg_data_t));
 205         NODE_DATA(nodeid)->bdata = &plat_node_bdata[nodeid];
 206         NODE_DATA(nodeid)->node_start_pfn = start_pfn;
 207         NODE_DATA(nodeid)->node_spanned_pages = last_pfn - start_pfn;
 208
 209         /*
 210          * Find a place for the bootmem map
 211          * nodedata_phys could be on other nodes by alloc_bootmem,
 212          * so need to sure bootmap_start not to be small, otherwise
 213          * early_node_mem will get that with find_e820_area instead
 214          * of alloc_bootmem, that could clash with reserved range
 215          */
 216         bootmap_pages = bootmem_bootmap_pages(last_pfn - start_pfn);
 217         nid = phys_to_nid(nodedata_phys);
 218         if (nid == nodeid)
 219                 bootmap_start = round_up(nodedata_phys + pgdat_size, PAGE_SIZE);
 220         else
 221                 bootmap_start = round_up(start, PAGE_SIZE);
 222         /*
 223          * SMP_CACHE_BYTES could be enough, but init_bootmem_node like
 224          * to use that to align to PAGE_SIZE
 225          */
 226         bootmap = early_node_mem(nodeid, bootmap_start, end,
 227                                  bootmap_pages<<PAGE_SHIFT, PAGE_SIZE);
 228         if (bootmap == NULL)  {
 229                 if (nodedata_phys < start || nodedata_phys >= end)
 230                         free_bootmem(nodedata_phys, pgdat_size);
 231                 node_data[nodeid] = NULL;
 232                 return;
 233         }
 234         bootmap_start = __pa(bootmap);
 235
 236         bootmap_size = init_bootmem_node(NODE_DATA(nodeid),
 237                                          bootmap_start >> PAGE_SHIFT,
 238                                          start_pfn, last_pfn);
 239
 240         printk(KERN_INFO "  bootmap [%016lx -  %016lx] pages %lx\n",
 241                  bootmap_start, bootmap_start + bootmap_size - 1,
 242                  bootmap_pages);
 243
 244         free_bootmem_with_active_regions(nodeid, end);
 245
 246         /*
 247          * convert early reserve to bootmem reserve earlier
 248          * otherwise early_node_mem could use early reserved mem
 249          * on previous node
 250          */
 251         early_res_to_bootmem(start, end);
 252
 253         /*
 254          * in some case early_node_mem could use alloc_bootmem
 255          * to get range on other node, don't reserve that again
 256          */
 257         if (nid != nodeid)
 258                 printk(KERN_INFO "    NODE_DATA(%d) on node %d\n", nodeid, nid);
 259         else
 260                 reserve_bootmem_node(NODE_DATA(nodeid), nodedata_phys,
 261                                         pgdat_size, BOOTMEM_DEFAULT);
 262         nid = phys_to_nid(bootmap_start);
 263         if (nid != nodeid)
 264                 printk(KERN_INFO "    bootmap(%d) on node %d\n", nodeid, nid);
 265         else
 266                 reserve_bootmem_node(NODE_DATA(nodeid), bootmap_start,
 267                                  bootmap_pages<<PAGE_SHIFT, BOOTMEM_DEFAULT);
 268
 269 #ifdef CONFIG_ACPI_NUMA
 270         srat_reserve_add_area(nodeid);
 271 #endif
 272         node_set_online(nodeid);
 273 }
 274
 275 /*
 276  * There are unfortunately some poorly designed mainboards around that
 277  * only connect memory to a single CPU. This breaks the 1:1 cpu->node
 278  * mapping. To avoid this fill in the mapping for all possible CPUs,
 279  * as the number of CPUs is not known yet. We round robin the existing
 280  * nodes.
 281  */
 282 void __init numa_init_array(void)
 283 {
 284         int rr, i;
 285
 286         rr = first_node(node_online_map);
 287         for (i = 0; i < NR_CPUS; i++) {
 288                 if (early_cpu_to_node(i) != NUMA_NO_NODE)
 289                         continue;
 290                 numa_set_node(i, rr);
 291                 rr = next_node(rr, node_online_map);
 292                 if (rr == MAX_NUMNODES)
 293                         rr = first_node(node_online_map);
 294         }
 295 }
 296
 297 #ifdef CONFIG_NUMA_EMU
 298 /* Numa emulation */
 299 static char *cmdline __initdata;
 300
 301 /*
 302  * Setups up nid to range from addr to addr + size.  If the end
 303  * boundary is greater than max_addr, then max_addr is used instead.
 304  * The return value is 0 if there is additional memory left for
 305  * allocation past addr and -1 otherwise.  addr is adjusted to be at
 306  * the end of the node.
 307  */
 308 static int __init setup_node_range(int nid, struct bootnode *nodes, u64 *addr,
 309                                    u64 size, u64 max_addr)
 310 {
 311         int ret = 0;
 312
 313         nodes[nid].start = *addr;
 314         *addr += size;
 315         if (*addr >= max_addr) {
 316                 *addr = max_addr;
 317                 ret = -1;
 318         }
 319         nodes[nid].end = *addr;
 320         node_set(nid, node_possible_map);
 321         printk(KERN_INFO "Faking node %d at %016Lx-%016Lx (%LuMB)\n", nid,
 322                nodes[nid].start, nodes[nid].end,
 323                (nodes[nid].end - nodes[nid].start) >> 20);
 324         return ret;
 325 }
 326
 327 /*
 328  * Splits num_nodes nodes up equally starting at node_start.  The return value
 329  * is the number of nodes split up and addr is adjusted to be at the end of the
 330  * last node allocated.
 331  */
 332 static int __init split_nodes_equally(struct bootnode *nodes, u64 *addr,
 333                                       u64 max_addr, int node_start,
 334                                       int num_nodes)
 335 {
 336         unsigned int big;
 337         u64 size;
 338         int i;
 339
 340         if (num_nodes <= 0)
 341                 return -1;
 342         if (num_nodes > MAX_NUMNODES)
 343                 num_nodes = MAX_NUMNODES;
 344         size = (max_addr - *addr - e820_hole_size(*addr, max_addr)) /
 345                num_nodes;
 346         /*
 347          * Calculate the number of big nodes that can be allocated as a result
 348          * of consolidating the leftovers.
 349          */
 350         big = ((size & ~FAKE_NODE_MIN_HASH_MASK) * num_nodes) /
 351               FAKE_NODE_MIN_SIZE;
 352
 353         /* Round down to nearest FAKE_NODE_MIN_SIZE. */
 354         size &= FAKE_NODE_MIN_HASH_MASK;
 355         if (!size) {
 356                 printk(KERN_ERR "Not enough memory for each node.  "
 357                        "NUMA emulation disabled.\n");
 358                 return -1;
 359         }
 360
 361         for (i = node_start; i < num_nodes + node_start; i++) {
 362                 u64 end = *addr + size;
 363
 364                 if (i < big)
 365                         end += FAKE_NODE_MIN_SIZE;
 366                 /*
 367                  * The final node can have the remaining system RAM.  Other
 368                  * nodes receive roughly the same amount of available pages.
 369                  */
 370                 if (i == num_nodes + node_start - 1)
 371                         end = max_addr;
 372                 else
 373                         while (end - *addr - e820_hole_size(*addr, end) <
 374                                size) {
 375                                 end += FAKE_NODE_MIN_SIZE;
 376                                 if (end > max_addr) {
 377                                         end = max_addr;
 378                                         break;
 379                                 }
 380                         }
 381                 if (setup_node_range(i, nodes, addr, end - *addr, max_addr) < 0)
 382                         break;
 383         }
 384         return i - node_start + 1;
 385 }
 386
 387 /*
 388  * Splits the remaining system RAM into chunks of size.  The remaining memory is
 389  * always assigned to a final node and can be asymmetric.  Returns the number of
 390  * nodes split.
 391  */
 392 static int __init split_nodes_by_size(struct bootnode *nodes, u64 *addr,
 393                                       u64 max_addr, int node_start, u64 size)
 394 {
 395         int i = node_start;
 396         size = (size << 20) & FAKE_NODE_MIN_HASH_MASK;
 397         while (!setup_node_range(i++, nodes, addr, size, max_addr))
 398                 ;
 399         return i - node_start;
 400 }
 401
 402 /*
 403  * Sets up the system RAM area from start_pfn to last_pfn according to the
 404  * numa=fake command-line option.
 405  */
 406 static struct bootnode nodes[MAX_NUMNODES] __initdata;
 407
 408 static int __init numa_emulation(unsigned long start_pfn, unsigned long last_pfn)
 409 {
 410         u64 size, addr = start_pfn << PAGE_SHIFT;
 411         u64 max_addr = last_pfn << PAGE_SHIFT;
 412         int num_nodes = 0, num = 0, coeff_flag, coeff = -1, i;
 413
 414         memset(&nodes, 0, sizeof(nodes));
 415         /*
 416          * If the numa=fake command-line is just a single number N, split the
 417          * system RAM into N fake nodes.
 418          */
 419         if (!strchr(cmdline, '*') && !strchr(cmdline, ',')) {
 420                 long n = simple_strtol(cmdline, NULL, 0);
 421
 422                 num_nodes = split_nodes_equally(nodes, &addr, max_addr, 0, n);
 423                 if (num_nodes < 0)
 424                         return num_nodes;
 425                 goto out;
 426         }
 427
 428         /* Parse the command line. */
 429         for (coeff_flag = 0; ; cmdline++) {
 430                 if (*cmdline && isdigit(*cmdline)) {
 431                         num = num * 10 + *cmdline - '0';
 432                         continue;
 433                 }
 434                 if (*cmdline == '*') {
 435                         if (num > 0)
 436                                 coeff = num;
 437                         coeff_flag = 1;
 438                 }
 439                 if (!*cmdline || *cmdline == ',') {
 440                         if (!coeff_flag)
 441                                 coeff = 1;
 442                         /*
 443                          * Round down to the nearest FAKE_NODE_MIN_SIZE.
 444                          * Command-line coefficients are in megabytes.
 445                          */
 446                         size = ((u64)num << 20) & FAKE_NODE_MIN_HASH_MASK;
 447                         if (size)
 448                                 for (i = 0; i < coeff; i++, num_nodes++)
 449                                         if (setup_node_range(num_nodes, nodes,
 450                                                 &addr, size, max_addr) < 0)
 451                                                 goto done;
 452                         if (!*cmdline)
 453                                 break;
 454                         coeff_flag = 0;
 455                         coeff = -1;
 456                 }
 457                 num = 0;
 458         }
 459 done:
 460         if (!num_nodes)
 461                 return -1;
 462         /* Fill remainder of system RAM, if appropriate. */
 463         if (addr < max_addr) {
 464                 if (coeff_flag && coeff < 0) {
 465                         /* Split remaining nodes into num-sized chunks */
 466                         num_nodes += split_nodes_by_size(nodes, &addr, max_addr,
 467                                                          num_nodes, num);
 468                         goto out;
 469                 }
 470                 switch (*(cmdline - 1)) {
 471                 case '*':
 472                         /* Split remaining nodes into coeff chunks */
 473                         if (coeff <= 0)
 474                                 break;
 475                         num_nodes += split_nodes_equally(nodes, &addr, max_addr,
 476                                                          num_nodes, coeff);
 477                         break;
 478                 case ',':
 479                         /* Do not allocate remaining system RAM */
 480                         break;
 481                 default:
 482                         /* Give one final node */
 483                         setup_node_range(num_nodes, nodes, &addr,
 484                                          max_addr - addr, max_addr);
 485                         num_nodes++;
 486                 }
 487         }
 488 out:
 489         memnode_shift = compute_hash_shift(nodes, num_nodes, NULL);
 490         if (memnode_shift < 0) {
 491                 memnode_shift = 0;
 492                 printk(KERN_ERR "No NUMA hash function found.  NUMA emulation "
 493                        "disabled.\n");
 494                 return -1;
 495         }
 496
 497         /*
 498          * We need to vacate all active ranges that may have been registered by
 499          * SRAT and set acpi_numa to -1 so that srat_disabled() always returns
 500          * true.  NUMA emulation has succeeded so we will not scan ACPI nodes.
 501          */
 502         remove_all_active_ranges();
 503 #ifdef CONFIG_ACPI_NUMA
 504         acpi_numa = -1;
 505 #endif
 506         for_each_node_mask(i, node_possible_map) {
 507                 e820_register_active_regions(i, nodes[i].start >> PAGE_SHIFT,
 508                                                 nodes[i].end >> PAGE_SHIFT);
 509                 setup_node_bootmem(i, nodes[i].start, nodes[i].end);
 510         }
 511         acpi_fake_nodes(nodes, num_nodes);
 512         numa_init_array();
 513         return 0;
 514 }
 515 #endif /* CONFIG_NUMA_EMU */
 516
 517 void __init numa_initmem_init(unsigned long start_pfn, unsigned long last_pfn)
 518 {
 519         int i;
 520
 521         nodes_clear(node_possible_map);
 522         nodes_clear(node_online_map);
 523
 524 #ifdef CONFIG_NUMA_EMU
 525         if (cmdline && !numa_emulation(start_pfn, last_pfn))
 526                 return;
 527         nodes_clear(node_possible_map);
 528         nodes_clear(node_online_map);
 529 #endif
 530
 531 #ifdef CONFIG_ACPI_NUMA
 532         if (!numa_off && !acpi_scan_nodes(start_pfn << PAGE_SHIFT,
 533                                           last_pfn << PAGE_SHIFT))
 534                 return;
 535         nodes_clear(node_possible_map);
 536         nodes_clear(node_online_map);
 537 #endif
 538
 539 #ifdef CONFIG_K8_NUMA
 540         if (!numa_off && !k8_scan_nodes(start_pfn<<PAGE_SHIFT,
 541                                         last_pfn<<PAGE_SHIFT))
 542                 return;
 543         nodes_clear(node_possible_map);
 544         nodes_clear(node_online_map);
 545 #endif
 546         printk(KERN_INFO "%s\n",
 547                numa_off ? "NUMA turned off" : "No NUMA configuration found");
 548
 549         printk(KERN_INFO "Faking a node at %016lx-%016lx\n",
 550                start_pfn << PAGE_SHIFT,
 551                last_pfn << PAGE_SHIFT);
 552         /* setup dummy node covering all memory */
 553         memnode_shift = 63;
 554         memnodemap = memnode.embedded_map;
 555         memnodemap[0] = 0;
 556         node_set_online(0);
 557         node_set(0, node_possible_map);
 558         for (i = 0; i < NR_CPUS; i++)
 559                 numa_set_node(i, 0);
 560         e820_register_active_regions(0, start_pfn, last_pfn);
 561         setup_node_bootmem(0, start_pfn << PAGE_SHIFT, last_pfn << PAGE_SHIFT);
 562 }
 563
 564 unsigned long __init numa_free_all_bootmem(void)
 565 {
 566         unsigned long pages = 0;
 567         int i;
 568
 569         for_each_online_node(i)
 570                 pages += free_all_bootmem_node(NODE_DATA(i));
 571
 572         return pages;
 573 }
 574
 575 void __init paging_init(void)
 576 {
 577         unsigned long max_zone_pfns[MAX_NR_ZONES];
 578
 579         memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
 580         max_zone_pfns[ZONE_DMA] = MAX_DMA_PFN;
 581         max_zone_pfns[ZONE_DMA32] = MAX_DMA32_PFN;
 582         max_zone_pfns[ZONE_NORMAL] = end_pfn;
 583
 584         sparse_memory_present_with_active_regions(MAX_NUMNODES);
 585         sparse_init();
 586
 587         free_area_init_nodes(max_zone_pfns);
 588 }
 589
 590 static __init int numa_setup(char *opt)
 591 {
 592         if (!opt)
 593                 return -EINVAL;
 594         if (!strncmp(opt, "off", 3))
 595                 numa_off = 1;
 596 #ifdef CONFIG_NUMA_EMU
 597         if (!strncmp(opt, "fake=", 5))
 598                 cmdline = opt + 5;
 599 #endif
 600 #ifdef CONFIG_ACPI_NUMA
 601         if (!strncmp(opt, "noacpi", 6))
 602                 acpi_numa = -1;
 603         if (!strncmp(opt, "hotadd=", 7))
 604                 hotadd_percent = simple_strtoul(opt+7, NULL, 10);
 605 #endif
 606         return 0;
 607 }
 608 early_param("numa", numa_setup);
 609
 610 #ifdef CONFIG_NUMA
 611 /*
 612  * Setup early cpu_to_node.
 613  *
 614  * Populate cpu_to_node[] only if x86_cpu_to_apicid[],
 615  * and apicid_to_node[] tables have valid entries for a CPU.
 616  * This means we skip cpu_to_node[] initialisation for NUMA
 617  * emulation and faking node case (when running a kernel compiled
 618  * for NUMA on a non NUMA box), which is OK as cpu_to_node[]
 619  * is already initialized in a round robin manner at numa_init_array,
 620  * prior to this call, and this initialization is good enough
 621  * for the fake NUMA cases.
 622  *
 623  * Called before the per_cpu areas are setup.
 624  */
 625 void __init init_cpu_to_node(void)
 626 {
 627         int cpu;
 628         u16 *cpu_to_apicid = early_per_cpu_ptr(x86_cpu_to_apicid);
 629
 630         BUG_ON(cpu_to_apicid == NULL);
 631
 632         for_each_possible_cpu(cpu) {
 633                 int node;
 634                 u16 apicid = cpu_to_apicid[cpu];
 635
 636                 if (apicid == BAD_APICID)
 637                         continue;
 638                 node = apicid_to_node[apicid];
 639                 if (node == NUMA_NO_NODE)
 640                         continue;
 641                 if (!node_online(node))
 642                         continue;
 643                 numa_set_node(cpu, node);
 644         }
 645 }
 646 #endif
 647
 648