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
  15 #include <asm/e820.h>
  16 #include <asm/proto.h>
  17 #include <asm/dma.h>
  18 #include <asm/numa.h>
  19 #include <asm/acpi.h>
  20 #include <asm/k8.h>
  21
  22 #ifndef Dprintk
  23 #define Dprintk(x...)
  24 #endif
  25
  26 struct pglist_data *node_data[MAX_NUMNODES] __read_mostly;
  27 EXPORT_SYMBOL(node_data);
  28
  29 bootmem_data_t plat_node_bdata[MAX_NUMNODES];
  30
  31 struct memnode memnode;
  32
  33 unsigned char cpu_to_node[NR_CPUS] __read_mostly = {
  34         [0 ... NR_CPUS-1] = NUMA_NO_NODE
  35 };
  36 EXPORT_SYMBOL(cpu_to_node);
  37
  38 unsigned char apicid_to_node[MAX_LOCAL_APIC] __cpuinitdata = {
  39         [0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE
  40 };
  41
  42 cpumask_t node_to_cpumask[MAX_NUMNODES] __read_mostly;
  43 EXPORT_SYMBOL(node_to_cpumask);
  44
  45 int numa_off __initdata;
  46 unsigned long __initdata nodemap_addr;
  47 unsigned long __initdata nodemap_size;
  48
  49 /*
  50  * Given a shift value, try to populate memnodemap[]
  51  * Returns :
  52  * 1 if OK
  53  * 0 if memnodmap[] too small (of shift too small)
  54  * -1 if node overlap or lost ram (shift too big)
  55  */
  56 static int __init populate_memnodemap(const struct bootnode *nodes,
  57                                       int numnodes, int shift)
  58 {
  59         unsigned long addr, end;
  60         int i, res = -1;
  61
  62         memset(memnodemap, 0xff, memnodemapsize);
  63         for (i = 0; i < numnodes; i++) {
  64                 addr = nodes[i].start;
  65                 end = nodes[i].end;
  66                 if (addr >= end)
  67                         continue;
  68                 if ((end >> shift) >= memnodemapsize)
  69                         return 0;
  70                 do {
  71                         if (memnodemap[addr >> shift] != 0xff)
  72                                 return -1;
  73                         memnodemap[addr >> shift] = i;
  74                         addr += (1UL << shift);
  75                 } while (addr < end);
  76                 res = 1;
  77         }
  78         return res;
  79 }
  80
  81 static int __init allocate_cachealigned_memnodemap(void)
  82 {
  83         unsigned long pad, pad_addr;
  84
  85         memnodemap = memnode.embedded_map;
  86         if (memnodemapsize <= 48)
  87                 return 0;
  88
  89         pad = L1_CACHE_BYTES - 1;
  90         pad_addr = 0x8000;
  91         nodemap_size = pad + memnodemapsize;
  92         nodemap_addr = find_e820_area(pad_addr, end_pfn<<PAGE_SHIFT,
  93                                       nodemap_size);
  94         if (nodemap_addr == -1UL) {
  95                 printk(KERN_ERR
  96                        "NUMA: Unable to allocate Memory to Node hash map\n");
  97                 nodemap_addr = nodemap_size = 0;
  98                 return -1;
  99         }
 100         pad_addr = (nodemap_addr + pad) & ~pad;
 101         memnodemap = phys_to_virt(pad_addr);
 102
 103         printk(KERN_DEBUG "NUMA: Allocated memnodemap from %lx - %lx\n",
 104                nodemap_addr, nodemap_addr + nodemap_size);
 105         return 0;
 106 }
 107
 108 /*
 109  * The LSB of all start and end addresses in the node map is the value of the
 110  * maximum possible shift.
 111  */
 112 static int __init extract_lsb_from_nodes(const struct bootnode *nodes,
 113                                          int numnodes)
 114 {
 115         int i, nodes_used = 0;
 116         unsigned long start, end;
 117         unsigned long bitfield = 0, memtop = 0;
 118
 119         for (i = 0; i < numnodes; i++) {
 120                 start = nodes[i].start;
 121                 end = nodes[i].end;
 122                 if (start >= end)
 123                         continue;
 124                 bitfield |= start;
 125                 nodes_used++;
 126                 if (end > memtop)
 127                         memtop = end;
 128         }
 129         if (nodes_used <= 1)
 130                 i = 63;
 131         else
 132                 i = find_first_bit(&bitfield, sizeof(unsigned long)*8);
 133         memnodemapsize = (memtop >> i)+1;
 134         return i;
 135 }
 136
 137 int __init compute_hash_shift(struct bootnode *nodes, int numnodes)
 138 {
 139         int shift;
 140
 141         shift = extract_lsb_from_nodes(nodes, numnodes);
 142         if (allocate_cachealigned_memnodemap())
 143                 return -1;
 144         printk(KERN_DEBUG "NUMA: Using %d for the hash shift.\n",
 145                 shift);
 146
 147         if (populate_memnodemap(nodes, numnodes, shift) != 1) {
 148                 printk(KERN_INFO "Your memory is not aligned you need to "
 149                        "rebuild your kernel with a bigger NODEMAPSIZE "
 150                        "shift=%d\n", shift);
 151                 return -1;
 152         }
 153         return shift;
 154 }
 155
 156 #ifdef CONFIG_SPARSEMEM
 157 int early_pfn_to_nid(unsigned long pfn)
 158 {
 159         return phys_to_nid(pfn << PAGE_SHIFT);
 160 }
 161 #endif
 162
 163 static void * __init early_node_mem(int nodeid, unsigned long start,
 164                                     unsigned long end, unsigned long size)
 165 {
 166         unsigned long mem = find_e820_area(start, end, size);
 167         void *ptr;
 168
 169         if (mem != -1L)
 170                 return __va(mem);
 171         ptr = __alloc_bootmem_nopanic(size,
 172                                 SMP_CACHE_BYTES, __pa(MAX_DMA_ADDRESS));
 173         if (ptr == NULL) {
 174                 printk(KERN_ERR "Cannot find %lu bytes in node %d\n",
 175                        size, nodeid);
 176                 return NULL;
 177         }
 178         return ptr;
 179 }
 180
 181 /* Initialize bootmem allocator for a node */
 182 void __init setup_node_bootmem(int nodeid, unsigned long start,
 183                                unsigned long end)
 184 {
 185         unsigned long start_pfn, end_pfn, bootmap_pages, bootmap_size;
 186         unsigned long bootmap_start, nodedata_phys;
 187         void *bootmap;
 188         const int pgdat_size = round_up(sizeof(pg_data_t), PAGE_SIZE);
 189
 190         start = round_up(start, ZONE_ALIGN);
 191
 192         printk(KERN_INFO "Bootmem setup node %d %016lx-%016lx\n", nodeid,
 193                start, end);
 194
 195         start_pfn = start >> PAGE_SHIFT;
 196         end_pfn = end >> PAGE_SHIFT;
 197
 198         node_data[nodeid] = early_node_mem(nodeid, start, end, pgdat_size);
 199         if (node_data[nodeid] == NULL)
 200                 return;
 201         nodedata_phys = __pa(node_data[nodeid]);
 202
 203         memset(NODE_DATA(nodeid), 0, sizeof(pg_data_t));
 204         NODE_DATA(nodeid)->bdata = &plat_node_bdata[nodeid];
 205         NODE_DATA(nodeid)->node_start_pfn = start_pfn;
 206         NODE_DATA(nodeid)->node_spanned_pages = end_pfn - start_pfn;
 207
 208         /* Find a place for the bootmem map */
 209         bootmap_pages = bootmem_bootmap_pages(end_pfn - start_pfn);
 210         bootmap_start = round_up(nodedata_phys + pgdat_size, PAGE_SIZE);
 211         bootmap = early_node_mem(nodeid, bootmap_start, end,
 212                                         bootmap_pages<<PAGE_SHIFT);
 213         if (bootmap == NULL)  {
 214                 if (nodedata_phys < start || nodedata_phys >= end)
 215                         free_bootmem((unsigned long)node_data[nodeid],
 216                                      pgdat_size);
 217                 node_data[nodeid] = NULL;
 218                 return;
 219         }
 220         bootmap_start = __pa(bootmap);
 221         Dprintk("bootmap start %lu pages %lu\n", bootmap_start, bootmap_pages);
 222
 223         bootmap_size = init_bootmem_node(NODE_DATA(nodeid),
 224                                          bootmap_start >> PAGE_SHIFT,
 225                                          start_pfn, end_pfn);
 226
 227         free_bootmem_with_active_regions(nodeid, end);
 228
 229         reserve_bootmem_node(NODE_DATA(nodeid), nodedata_phys, pgdat_size);
 230         reserve_bootmem_node(NODE_DATA(nodeid), bootmap_start,
 231                              bootmap_pages<<PAGE_SHIFT);
 232 #ifdef CONFIG_ACPI_NUMA
 233         srat_reserve_add_area(nodeid);
 234 #endif
 235         node_set_online(nodeid);
 236 }
 237
 238 /* Initialize final allocator for a zone */
 239 void __init setup_node_zones(int nodeid)
 240 {
 241         unsigned long start_pfn, end_pfn, memmapsize, limit;
 242
 243         start_pfn = node_start_pfn(nodeid);
 244         end_pfn = node_end_pfn(nodeid);
 245
 246         Dprintk(KERN_INFO "Setting up memmap for node %d %lx-%lx\n",
 247                 nodeid, start_pfn, end_pfn);
 248
 249         /*
 250          * Try to allocate mem_map at end to not fill up precious <4GB
 251          * memory.
 252          */
 253         memmapsize = sizeof(struct page) * (end_pfn-start_pfn);
 254         limit = end_pfn << PAGE_SHIFT;
 255 #ifdef CONFIG_FLAT_NODE_MEM_MAP
 256         NODE_DATA(nodeid)->node_mem_map =
 257                 __alloc_bootmem_core(NODE_DATA(nodeid)->bdata,
 258                                      memmapsize, SMP_CACHE_BYTES,
 259                                      round_down(limit - memmapsize, PAGE_SIZE),
 260                                      limit);
 261 #endif
 262 }
 263
 264 /*
 265  * There are unfortunately some poorly designed mainboards around that
 266  * only connect memory to a single CPU. This breaks the 1:1 cpu->node
 267  * mapping. To avoid this fill in the mapping for all possible CPUs,
 268  * as the number of CPUs is not known yet. We round robin the existing
 269  * nodes.
 270  */
 271 void __init numa_init_array(void)
 272 {
 273         int rr, i;
 274
 275         rr = first_node(node_online_map);
 276         for (i = 0; i < NR_CPUS; i++) {
 277                 if (cpu_to_node(i) != NUMA_NO_NODE)
 278                         continue;
 279                 numa_set_node(i, rr);
 280                 rr = next_node(rr, node_online_map);
 281                 if (rr == MAX_NUMNODES)
 282                         rr = first_node(node_online_map);
 283         }
 284 }
 285
 286 #ifdef CONFIG_NUMA_EMU
 287 /* Numa emulation */
 288 char *cmdline __initdata;
 289
 290 /*
 291  * Setups up nid to range from addr to addr + size.  If the end
 292  * boundary is greater than max_addr, then max_addr is used instead.
 293  * The return value is 0 if there is additional memory left for
 294  * allocation past addr and -1 otherwise.  addr is adjusted to be at
 295  * the end of the node.
 296  */
 297 static int __init setup_node_range(int nid, struct bootnode *nodes, u64 *addr,
 298                                    u64 size, u64 max_addr)
 299 {
 300         int ret = 0;
 301
 302         nodes[nid].start = *addr;
 303         *addr += size;
 304         if (*addr >= max_addr) {
 305                 *addr = max_addr;
 306                 ret = -1;
 307         }
 308         nodes[nid].end = *addr;
 309         node_set(nid, node_possible_map);
 310         printk(KERN_INFO "Faking node %d at %016Lx-%016Lx (%LuMB)\n", nid,
 311                nodes[nid].start, nodes[nid].end,
 312                (nodes[nid].end - nodes[nid].start) >> 20);
 313         return ret;
 314 }
 315
 316 /*
 317  * Splits num_nodes nodes up equally starting at node_start.  The return value
 318  * is the number of nodes split up and addr is adjusted to be at the end of the
 319  * last node allocated.
 320  */
 321 static int __init split_nodes_equally(struct bootnode *nodes, u64 *addr,
 322                                       u64 max_addr, int node_start,
 323                                       int num_nodes)
 324 {
 325         unsigned int big;
 326         u64 size;
 327         int i;
 328
 329         if (num_nodes <= 0)
 330                 return -1;
 331         if (num_nodes > MAX_NUMNODES)
 332                 num_nodes = MAX_NUMNODES;
 333         size = (max_addr - *addr - e820_hole_size(*addr, max_addr)) /
 334                num_nodes;
 335         /*
 336          * Calculate the number of big nodes that can be allocated as a result
 337          * of consolidating the leftovers.
 338          */
 339         big = ((size & ~FAKE_NODE_MIN_HASH_MASK) * num_nodes) /
 340               FAKE_NODE_MIN_SIZE;
 341
 342         /* Round down to nearest FAKE_NODE_MIN_SIZE. */
 343         size &= FAKE_NODE_MIN_HASH_MASK;
 344         if (!size) {
 345                 printk(KERN_ERR "Not enough memory for each node.  "
 346                        "NUMA emulation disabled.\n");
 347                 return -1;
 348         }
 349
 350         for (i = node_start; i < num_nodes + node_start; i++) {
 351                 u64 end = *addr + size;
 352
 353                 if (i < big)
 354                         end += FAKE_NODE_MIN_SIZE;
 355                 /*
 356                  * The final node can have the remaining system RAM.  Other
 357                  * nodes receive roughly the same amount of available pages.
 358                  */
 359                 if (i == num_nodes + node_start - 1)
 360                         end = max_addr;
 361                 else
 362                         while (end - *addr - e820_hole_size(*addr, end) <
 363                                size) {
 364                                 end += FAKE_NODE_MIN_SIZE;
 365                                 if (end > max_addr) {
 366                                         end = max_addr;
 367                                         break;
 368                                 }
 369                         }
 370                 if (setup_node_range(i, nodes, addr, end - *addr, max_addr) < 0)
 371                         break;
 372         }
 373         return i - node_start + 1;
 374 }
 375
 376 /*
 377  * Splits the remaining system RAM into chunks of size.  The remaining memory is
 378  * always assigned to a final node and can be asymmetric.  Returns the number of
 379  * nodes split.
 380  */
 381 static int __init split_nodes_by_size(struct bootnode *nodes, u64 *addr,
 382                                       u64 max_addr, int node_start, u64 size)
 383 {
 384         int i = node_start;
 385         size = (size << 20) & FAKE_NODE_MIN_HASH_MASK;
 386         while (!setup_node_range(i++, nodes, addr, size, max_addr))
 387                 ;
 388         return i - node_start;
 389 }
 390
 391 /*
 392  * Sets up the system RAM area from start_pfn to end_pfn according to the
 393  * numa=fake command-line option.
 394  */
 395 static int __init numa_emulation(unsigned long start_pfn, unsigned long end_pfn)
 396 {
 397         struct bootnode nodes[MAX_NUMNODES];
 398         u64 size, addr = start_pfn << PAGE_SHIFT;
 399         u64 max_addr = end_pfn << PAGE_SHIFT;
 400         int num_nodes = 0, num = 0, coeff_flag, coeff = -1, i;
 401
 402         memset(&nodes, 0, sizeof(nodes));
 403         /*
 404          * If the numa=fake command-line is just a single number N, split the
 405          * system RAM into N fake nodes.
 406          */
 407         if (!strchr(cmdline, '*') && !strchr(cmdline, ',')) {
 408                 long n = simple_strtol(cmdline, NULL, 0);
 409
 410                 num_nodes = split_nodes_equally(nodes, &addr, max_addr, 0, n);
 411                 if (num_nodes < 0)
 412                         return num_nodes;
 413                 goto out;
 414         }
 415
 416         /* Parse the command line. */
 417         for (coeff_flag = 0; ; cmdline++) {
 418                 if (*cmdline && isdigit(*cmdline)) {
 419                         num = num * 10 + *cmdline - '0';
 420                         continue;
 421                 }
 422                 if (*cmdline == '*') {
 423                         if (num > 0)
 424                                 coeff = num;
 425                         coeff_flag = 1;
 426                 }
 427                 if (!*cmdline || *cmdline == ',') {
 428                         if (!coeff_flag)
 429                                 coeff = 1;
 430                         /*
 431                          * Round down to the nearest FAKE_NODE_MIN_SIZE.
 432                          * Command-line coefficients are in megabytes.
 433                          */
 434                         size = ((u64)num << 20) & FAKE_NODE_MIN_HASH_MASK;
 435                         if (size)
 436                                 for (i = 0; i < coeff; i++, num_nodes++)
 437                                         if (setup_node_range(num_nodes, nodes,
 438                                                 &addr, size, max_addr) < 0)
 439                                                 goto done;
 440                         if (!*cmdline)
 441                                 break;
 442                         coeff_flag = 0;
 443                         coeff = -1;
 444                 }
 445                 num = 0;
 446         }
 447 done:
 448         if (!num_nodes)
 449                 return -1;
 450         /* Fill remainder of system RAM, if appropriate. */
 451         if (addr < max_addr) {
 452                 if (coeff_flag && coeff < 0) {
 453                         /* Split remaining nodes into num-sized chunks */
 454                         num_nodes += split_nodes_by_size(nodes, &addr, max_addr,
 455                                                          num_nodes, num);
 456                         goto out;
 457                 }
 458                 switch (*(cmdline - 1)) {
 459                 case '*':
 460                         /* Split remaining nodes into coeff chunks */
 461                         if (coeff <= 0)
 462                                 break;
 463                         num_nodes += split_nodes_equally(nodes, &addr, max_addr,
 464                                                          num_nodes, coeff);
 465                         break;
 466                 case ',':
 467                         /* Do not allocate remaining system RAM */
 468                         break;
 469                 default:
 470                         /* Give one final node */
 471                         setup_node_range(num_nodes, nodes, &addr,
 472                                          max_addr - addr, max_addr);
 473                         num_nodes++;
 474                 }
 475         }
 476 out:
 477         memnode_shift = compute_hash_shift(nodes, num_nodes);
 478         if (memnode_shift < 0) {
 479                 memnode_shift = 0;
 480                 printk(KERN_ERR "No NUMA hash function found.  NUMA emulation "
 481                        "disabled.\n");
 482                 return -1;
 483         }
 484
 485         /*
 486          * We need to vacate all active ranges that may have been registered by
 487          * SRAT and set acpi_numa to -1 so that srat_disabled() always returns
 488          * true.  NUMA emulation has succeeded so we will not scan ACPI nodes.
 489          */
 490         remove_all_active_ranges();
 491 #ifdef CONFIG_ACPI_NUMA
 492         acpi_numa = -1;
 493 #endif
 494         for_each_node_mask(i, node_possible_map) {
 495                 e820_register_active_regions(i, nodes[i].start >> PAGE_SHIFT,
 496                                                 nodes[i].end >> PAGE_SHIFT);
 497                 setup_node_bootmem(i, nodes[i].start, nodes[i].end);
 498         }
 499         acpi_fake_nodes(nodes, num_nodes);
 500         numa_init_array();
 501         return 0;
 502 }
 503 #endif /* CONFIG_NUMA_EMU */
 504
 505 void __init numa_initmem_init(unsigned long start_pfn, unsigned long end_pfn)
 506 {
 507         int i;
 508
 509         nodes_clear(node_possible_map);
 510
 511 #ifdef CONFIG_NUMA_EMU
 512         if (cmdline && !numa_emulation(start_pfn, end_pfn))
 513                 return;
 514         nodes_clear(node_possible_map);
 515 #endif
 516
 517 #ifdef CONFIG_ACPI_NUMA
 518         if (!numa_off && !acpi_scan_nodes(start_pfn << PAGE_SHIFT,
 519                                           end_pfn << PAGE_SHIFT))
 520                 return;
 521         nodes_clear(node_possible_map);
 522 #endif
 523
 524 #ifdef CONFIG_K8_NUMA
 525         if (!numa_off && !k8_scan_nodes(start_pfn<<PAGE_SHIFT,
 526                                         end_pfn<<PAGE_SHIFT))
 527                 return;
 528         nodes_clear(node_possible_map);
 529 #endif
 530         printk(KERN_INFO "%s\n",
 531                numa_off ? "NUMA turned off" : "No NUMA configuration found");
 532
 533         printk(KERN_INFO "Faking a node at %016lx-%016lx\n",
 534                start_pfn << PAGE_SHIFT,
 535                end_pfn << PAGE_SHIFT);
 536         /* setup dummy node covering all memory */
 537         memnode_shift = 63;
 538         memnodemap = memnode.embedded_map;
 539         memnodemap[0] = 0;
 540         nodes_clear(node_online_map);
 541         node_set_online(0);
 542         node_set(0, node_possible_map);
 543         for (i = 0; i < NR_CPUS; i++)
 544                 numa_set_node(i, 0);
 545         node_to_cpumask[0] = cpumask_of_cpu(0);
 546         e820_register_active_regions(0, start_pfn, end_pfn);
 547         setup_node_bootmem(0, start_pfn << PAGE_SHIFT, end_pfn << PAGE_SHIFT);
 548 }
 549
 550 __cpuinit void numa_add_cpu(int cpu)
 551 {
 552         set_bit(cpu, &node_to_cpumask[cpu_to_node(cpu)]);
 553 }
 554
 555 void __cpuinit numa_set_node(int cpu, int node)
 556 {
 557         cpu_pda(cpu)->nodenumber = node;
 558         cpu_to_node(cpu) = node;
 559 }
 560
 561 unsigned long __init numa_free_all_bootmem(void)
 562 {
 563         unsigned long pages = 0;
 564         int i;
 565
 566         for_each_online_node(i)
 567                 pages += free_all_bootmem_node(NODE_DATA(i));
 568
 569         return pages;
 570 }
 571
 572 void __init paging_init(void)
 573 {
 574         unsigned long max_zone_pfns[MAX_NR_ZONES];
 575         int i;
 576
 577         memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
 578         max_zone_pfns[ZONE_DMA] = MAX_DMA_PFN;
 579         max_zone_pfns[ZONE_DMA32] = MAX_DMA32_PFN;
 580         max_zone_pfns[ZONE_NORMAL] = end_pfn;
 581
 582         sparse_memory_present_with_active_regions(MAX_NUMNODES);
 583         sparse_init();
 584
 585         for_each_online_node(i)
 586                 setup_node_zones(i);
 587
 588         free_area_init_nodes(max_zone_pfns);
 589 }
 590
 591 static __init int numa_setup(char *opt)
 592 {
 593         if (!opt)
 594                 return -EINVAL;
 595         if (!strncmp(opt, "off", 3))
 596                 numa_off = 1;
 597 #ifdef CONFIG_NUMA_EMU
 598         if (!strncmp(opt, "fake=", 5))
 599                 cmdline = opt + 5;
 600 #endif
 601 #ifdef CONFIG_ACPI_NUMA
 602         if (!strncmp(opt, "noacpi", 6))
 603                 acpi_numa = -1;
 604         if (!strncmp(opt, "hotadd=", 7))
 605                 hotadd_percent = simple_strtoul(opt+7, NULL, 10);
 606 #endif
 607         return 0;
 608 }
 609 early_param("numa", numa_setup);
 610
 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 void __init init_cpu_to_node(void)
 624 {
 625         int i;
 626
 627         for (i = 0; i < NR_CPUS; i++) {
 628                 u8 apicid = x86_cpu_to_apicid_init[i];
 629
 630                 if (apicid == BAD_APICID)
 631                         continue;
 632                 if (apicid_to_node[apicid] == NUMA_NO_NODE)
 633                         continue;
 634                 numa_set_node(i, apicid_to_node[apicid]);
 635         }
 636 }
 637
 638 #ifdef CONFIG_DISCONTIGMEM
 639 /*
 640  * Functions to convert PFNs from/to per node page addresses.
 641  * These are out of line because they are quite big.
 642  * They could be all tuned by pre caching more state.
 643  * Should do that.
 644  */
 645
 646 int pfn_valid(unsigned long pfn)
 647 {
 648         unsigned nid;
 649         if (pfn >= num_physpages)
 650                 return 0;
 651         nid = pfn_to_nid(pfn);
 652         if (nid == 0xff)
 653                 return 0;
 654         return pfn >= node_start_pfn(nid) && (pfn) < node_end_pfn(nid);
 655 }
 656 EXPORT_SYMBOL(pfn_valid);
 657 #endif