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#include <asm/k8.h> 26 27static struct bootnode __initdata nodes[8]; 28static nodemask_t __initdata nodes_parsed = NODE_MASK_NONE; 29 30static __init int find_northbridge(void) 31{ 32 int num; 33 34 for (num = 0; num < 32; num++) { 35 u32 header; 36 37 header = read_pci_config(0, num, 0, 0x00); 38 if (header != (PCI_VENDOR_ID_AMD | (0x1100<<16)) && 39 header != (PCI_VENDOR_ID_AMD | (0x1200<<16)) && 40 header != (PCI_VENDOR_ID_AMD | (0x1300<<16))) 41 continue; 42 43 header = read_pci_config(0, num, 1, 0x00); 44 if (header != (PCI_VENDOR_ID_AMD | (0x1101<<16)) && 45 header != (PCI_VENDOR_ID_AMD | (0x1201<<16)) && 46 header != (PCI_VENDOR_ID_AMD | (0x1301<<16))) 47 continue; 48 return num; 49 } 50 51 return -1; 52} 53 54static __init void early_get_boot_cpu_id(void) 55{ 56 /* 57 * need to get boot_cpu_id so can use that to create apicid_to_node 58 * in k8_scan_nodes() 59 */ 60#ifdef CONFIG_X86_MPPARSE 61 /* 62 * get boot-time SMP configuration: 63 */ 64 if (smp_found_config) 65 early_get_smp_config(); 66#endif 67 early_init_lapic_mapping(); 68} 69 70int __init k8_get_nodes(struct bootnode *physnodes) 71{ 72 int i; 73 int ret = 0; 74 75 for_each_node_mask(i, nodes_parsed) { 76 physnodes[ret].start = nodes[i].start; 77 physnodes[ret].end = nodes[i].end; 78 ret++; 79 } 80 return ret; 81} 82 83int __init k8_numa_init(unsigned long start_pfn, unsigned long end_pfn) 84{ 85 unsigned long start = PFN_PHYS(start_pfn); 86 unsigned long end = PFN_PHYS(end_pfn); 87 unsigned numnodes; 88 unsigned long prevbase; 89 int i, nb, found = 0; 90 u32 nodeid, reg; 91 92 if (!early_pci_allowed()) 93 return -1; 94 95 nb = find_northbridge(); 96 if (nb < 0) 97 return nb; 98 99 pr_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 pr_info("Number of physical nodes %d\n", numnodes); 107 108 prevbase = 0; 109 for (i = 0; i < 8; i++) { 110 unsigned long base, limit; 111 112 base = read_pci_config(0, nb, 1, 0x40 + i*8); 113 limit = read_pci_config(0, nb, 1, 0x44 + i*8); 114 115 nodeid = limit & 7; 116 if ((base & 3) == 0) { 117 if (i < numnodes) 118 pr_info("Skipping disabled node %d\n", i); 119 continue; 120 } 121 if (nodeid >= numnodes) { 122 pr_info("Ignoring excess node %d (%lx:%lx)\n", nodeid, 123 base, limit); 124 continue; 125 } 126 127 if (!limit) { 128 pr_info("Skipping node entry %d (base %lx)\n", 129 i, base); 130 continue; 131 } 132 if ((base >> 8) & 3 || (limit >> 8) & 3) { 133 pr_err("Node %d using interleaving mode %lx/%lx\n", 134 nodeid, (base >> 8) & 3, (limit >> 8) & 3); 135 return -1; 136 } 137 if (node_isset(nodeid, nodes_parsed)) { 138 pr_info("Node %d already present, skipping\n", 139 nodeid); 140 continue; 141 } 142 143 limit >>= 16; 144 limit <<= 24; 145 limit |= (1<<24)-1; 146 limit++; 147 148 if (limit > end) 149 limit = end; 150 if (limit <= base) 151 continue; 152 153 base >>= 16; 154 base <<= 24; 155 156 if (base < start) 157 base = start; 158 if (limit > end) 159 limit = end; 160 if (limit == base) { 161 pr_err("Empty node %d\n", nodeid); 162 continue; 163 } 164 if (limit < base) { 165 pr_err("Node %d bogus settings %lx-%lx.\n", 166 nodeid, base, limit); 167 continue; 168 } 169 170 /* Could sort here, but pun for now. Should not happen anyroads. */ 171 if (prevbase > base) { 172 pr_err("Node map not sorted %lx,%lx\n", 173 prevbase, base); 174 return -1; 175 } 176 177 pr_info("Node %d MemBase %016lx Limit %016lx\n", 178 nodeid, base, limit); 179 180 found++; 181 182 nodes[nodeid].start = base; 183 nodes[nodeid].end = limit; 184 185 prevbase = base; 186 187 node_set(nodeid, nodes_parsed); 188 } 189 190 if (!found) 191 return -1; 192 return 0; 193} 194 195int __init k8_scan_nodes(void) 196{ 197 unsigned int bits; 198 unsigned int cores; 199 unsigned int apicid_base; 200 int i; 201 202 BUG_ON(nodes_empty(nodes_parsed)); 203 node_possible_map = nodes_parsed; 204 memnode_shift = compute_hash_shift(nodes, 8, NULL); 205 if (memnode_shift < 0) { 206 pr_err("No NUMA node hash function found. Contact maintainer\n"); 207 return -1; 208 } 209 pr_info("Using node hash shift of %d\n", memnode_shift); 210 211 /* use the coreid bits from early_identify_cpu */ 212 bits = boot_cpu_data.x86_coreid_bits; 213 cores = (1<<bits); 214 apicid_base = 0; 215 /* need to get boot_cpu_id early for system with apicid lifting */ 216 early_get_boot_cpu_id(); 217 if (boot_cpu_physical_apicid > 0) { 218 pr_info("BSP APIC ID: %02x\n", boot_cpu_physical_apicid); 219 apicid_base = boot_cpu_physical_apicid; 220 } 221 222 for_each_node_mask(i, node_possible_map) { 223 int j; 224 225 e820_register_active_regions(i, 226 nodes[i].start >> PAGE_SHIFT, 227 nodes[i].end >> PAGE_SHIFT); 228 for (j = apicid_base; j < cores + apicid_base; j++) 229 apicid_to_node[(i << bits) + j] = i; 230 setup_node_bootmem(i, nodes[i].start, nodes[i].end); 231 } 232 233 numa_init_array(); 234 return 0; 235} 236