arch/x86/mm/k8topology_64.c

   1 /*
   2  * AMD K8 NUMA support.
   3  * Discover the memory map and associated nodes.
   4  *
   5  * This version reads it directly from the K8 northbridge.
   6  *
   7  * Copyright 2002,2003 Andi Kleen, SuSE Labs.
   8  */
   9 #include <linux/kernel.h>
  10 #include <linux/init.h>
  11 #include <linux/string.h>
  12 #include <linux/module.h>
  13 #include <linux/nodemask.h>
  14 #include <asm/io.h>
  15 #include <linux/pci_ids.h>
  16 #include <linux/acpi.h>
  17 #include <asm/types.h>
  18 #include <asm/mmzone.h>
  19 #include <asm/proto.h>
  20 #include <asm/e820.h>
  21 #include <asm/pci-direct.h>
  22 #include <asm/numa.h>
  23 #include <asm/mpspec.h>
  24 #include <asm/apic.h>
  25
  26 static __init int find_northbridge(void)
  27 {
  28         int num;
  29
  30         for (num = 0; num < 32; num++) {
  31                 u32 header;
  32
  33                 header = read_pci_config(0, num, 0, 0x00);
  34                 if (header != (PCI_VENDOR_ID_AMD | (0x1100<<16)) &&
  35                         header != (PCI_VENDOR_ID_AMD | (0x1200<<16)) &&
  36                         header != (PCI_VENDOR_ID_AMD | (0x1300<<16)))
  37                         continue;
  38
  39                 header = read_pci_config(0, num, 1, 0x00);
  40                 if (header != (PCI_VENDOR_ID_AMD | (0x1101<<16)) &&
  41                         header != (PCI_VENDOR_ID_AMD | (0x1201<<16)) &&
  42                         header != (PCI_VENDOR_ID_AMD | (0x1301<<16)))
  43                         continue;
  44                 return num;
  45         }
  46
  47         return -1;
  48 }
  49
  50 static __init void early_get_boot_cpu_id(void)
  51 {
  52         /*
  53          * need to get boot_cpu_id so can use that to create apicid_to_node
  54          * in k8_scan_nodes()
  55          */
  56         /*
  57          * Find possible boot-time SMP configuration:
  58          */
  59 #ifdef CONFIG_X86_MPPARSE
  60         early_find_smp_config();
  61 #endif
  62 #ifdef CONFIG_ACPI
  63         /*
  64          * Read APIC information from ACPI tables.
  65          */
  66         early_acpi_boot_init();
  67 #endif
  68 #ifdef CONFIG_X86_MPPARSE
  69         /*
  70          * get boot-time SMP configuration:
  71          */
  72         if (smp_found_config)
  73                 early_get_smp_config();
  74 #endif
  75         early_init_lapic_mapping();
  76 }
  77
  78 int __init k8_scan_nodes(unsigned long start, unsigned long end)
  79 {
  80         unsigned long prevbase;
  81         struct bootnode nodes[8];
  82         int nodeid, i, nb;
  83         unsigned char nodeids[8];
  84         int found = 0;
  85         u32 reg;
  86         unsigned numnodes;
  87         unsigned cores;
  88         unsigned bits;
  89         int j;
  90         unsigned apicid_base;
  91
  92         if (!early_pci_allowed())
  93                 return -1;
  94
  95         nb = find_northbridge();
  96         if (nb < 0)
  97                 return nb;
  98
  99         printk(KERN_INFO "Scanning NUMA topology in Northbridge %d\n", nb);
 100
 101         reg = read_pci_config(0, nb, 0, 0x60);
 102         numnodes = ((reg >> 4) & 0xF) + 1;
 103         if (numnodes <= 1)
 104                 return -1;
 105
 106         printk(KERN_INFO "Number of nodes %d\n", numnodes);
 107
 108         memset(&nodes, 0, sizeof(nodes));
 109         prevbase = 0;
 110         for (i = 0; i < 8; i++) {
 111                 unsigned long base, limit;
 112                 u32 nodeid;
 113
 114                 base = read_pci_config(0, nb, 1, 0x40 + i*8);
 115                 limit = read_pci_config(0, nb, 1, 0x44 + i*8);
 116
 117                 nodeid = limit & 7;
 118                 nodeids[i] = nodeid;
 119                 if ((base & 3) == 0) {
 120                         if (i < numnodes)
 121                                 printk("Skipping disabled node %d\n", i);
 122                         continue;
 123                 }
 124                 if (nodeid >= numnodes) {
 125                         printk("Ignoring excess node %d (%lx:%lx)\n", nodeid,
 126                                base, limit);
 127                         continue;
 128                 }
 129
 130                 if (!limit) {
 131                         printk(KERN_INFO "Skipping node entry %d (base %lx)\n",
 132                                i, base);
 133                         continue;
 134                 }
 135                 if ((base >> 8) & 3 || (limit >> 8) & 3) {
 136                         printk(KERN_ERR "Node %d using interleaving mode %lx/%lx\n",
 137                                nodeid, (base>>8)&3, (limit>>8) & 3);
 138                         return -1;
 139                 }
 140                 if (node_isset(nodeid, node_possible_map)) {
 141                         printk(KERN_INFO "Node %d already present. Skipping\n",
 142                                nodeid);
 143                         continue;
 144                 }
 145
 146                 limit >>= 16;
 147                 limit <<= 24;
 148                 limit |= (1<<24)-1;
 149                 limit++;
 150
 151                 if (limit > end_pfn << PAGE_SHIFT)
 152                         limit = end_pfn << PAGE_SHIFT;
 153                 if (limit <= base)
 154                         continue;
 155
 156                 base >>= 16;
 157                 base <<= 24;
 158
 159                 if (base < start)
 160                         base = start;
 161                 if (limit > end)
 162                         limit = end;
 163                 if (limit == base) {
 164                         printk(KERN_ERR "Empty node %d\n", nodeid);
 165                         continue;
 166                 }
 167                 if (limit < base) {
 168                         printk(KERN_ERR "Node %d bogus settings %lx-%lx.\n",
 169                                nodeid, base, limit);
 170                         continue;
 171                 }
 172
 173                 /* Could sort here, but pun for now. Should not happen anyroads. */
 174                 if (prevbase > base) {
 175                         printk(KERN_ERR "Node map not sorted %lx,%lx\n",
 176                                prevbase, base);
 177                         return -1;
 178                 }
 179
 180                 printk(KERN_INFO "Node %d MemBase %016lx Limit %016lx\n",
 181                        nodeid, base, limit);
 182
 183                 found++;
 184
 185                 nodes[nodeid].start = base;
 186                 nodes[nodeid].end = limit;
 187                 e820_register_active_regions(nodeid,
 188                                 nodes[nodeid].start >> PAGE_SHIFT,
 189                                 nodes[nodeid].end >> PAGE_SHIFT);
 190
 191                 prevbase = base;
 192
 193                 node_set(nodeid, node_possible_map);
 194         }
 195
 196         if (!found)
 197                 return -1;
 198
 199         memnode_shift = compute_hash_shift(nodes, 8, NULL);
 200         if (memnode_shift < 0) {
 201                 printk(KERN_ERR "No NUMA node hash function found. Contact maintainer\n");
 202                 return -1;
 203         }
 204         printk(KERN_INFO "Using node hash shift of %d\n", memnode_shift);
 205
 206         /* use the coreid bits from early_identify_cpu */
 207         bits = boot_cpu_data.x86_coreid_bits;
 208         cores = (1<<bits);
 209         apicid_base = 0;
 210         /* need to get boot_cpu_id early for system with apicid lifting */
 211         early_get_boot_cpu_id();
 212         if (boot_cpu_physical_apicid > 0) {
 213                 printk(KERN_INFO "BSP APIC ID: %02x\n",
 214                                  boot_cpu_physical_apicid);
 215                 apicid_base = boot_cpu_physical_apicid;
 216         }
 217
 218         for (i = 0; i < 8; i++) {
 219                 if (nodes[i].start != nodes[i].end) {
 220                         nodeid = nodeids[i];
 221                         for (j = apicid_base; j < cores + apicid_base; j++)
 222                                 apicid_to_node[(nodeid << bits) + j] = i;
 223                         setup_node_bootmem(i, nodes[i].start, nodes[i].end);
 224                 }
 225         }
 226
 227         numa_init_array();
 228         return 0;
 229 }