3 * Discover the memory map and associated nodes.
5 * This version reads it directly from the AMD northbridge.
7 * Copyright 2002,2003 Andi Kleen, SuSE Labs.
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 <linux/memblock.h>
17 #include <linux/pci_ids.h>
18 #include <linux/acpi.h>
19 #include <asm/types.h>
20 #include <asm/mmzone.h>
21 #include <asm/proto.h>
23 #include <asm/pci-direct.h>
25 #include <asm/mpspec.h>
27 #include <asm/amd_nb.h>
29 static unsigned char __initdata nodeids
[8];
31 static __init
int find_northbridge(void)
35 for (num
= 0; num
< 32; num
++) {
38 header
= read_pci_config(0, num
, 0, 0x00);
39 if (header
!= (PCI_VENDOR_ID_AMD
| (0x1100<<16)) &&
40 header
!= (PCI_VENDOR_ID_AMD
| (0x1200<<16)) &&
41 header
!= (PCI_VENDOR_ID_AMD
| (0x1300<<16)))
44 header
= read_pci_config(0, num
, 1, 0x00);
45 if (header
!= (PCI_VENDOR_ID_AMD
| (0x1101<<16)) &&
46 header
!= (PCI_VENDOR_ID_AMD
| (0x1201<<16)) &&
47 header
!= (PCI_VENDOR_ID_AMD
| (0x1301<<16)))
55 static __init
void early_get_boot_cpu_id(void)
58 * need to get the APIC ID of the BSP so can use that to
59 * create apicid_to_node in amd_scan_nodes()
61 #ifdef CONFIG_X86_MPPARSE
63 * get boot-time SMP configuration:
66 early_get_smp_config();
70 int __init
amd_numa_init(void)
72 unsigned long start
= PFN_PHYS(0);
73 unsigned long end
= PFN_PHYS(max_pfn
);
75 unsigned long prevbase
;
78 unsigned int bits
, cores
, apicid_base
;
80 if (!early_pci_allowed())
83 nb
= find_northbridge();
87 pr_info("Scanning NUMA topology in Northbridge %d\n", nb
);
89 reg
= read_pci_config(0, nb
, 0, 0x60);
90 numnodes
= ((reg
>> 4) & 0xF) + 1;
94 pr_info("Number of physical nodes %d\n", numnodes
);
97 for (i
= 0; i
< 8; i
++) {
98 unsigned long base
, limit
;
100 base
= read_pci_config(0, nb
, 1, 0x40 + i
*8);
101 limit
= read_pci_config(0, nb
, 1, 0x44 + i
*8);
103 nodeids
[i
] = nodeid
= limit
& 7;
104 if ((base
& 3) == 0) {
106 pr_info("Skipping disabled node %d\n", i
);
109 if (nodeid
>= numnodes
) {
110 pr_info("Ignoring excess node %d (%lx:%lx)\n", nodeid
,
116 pr_info("Skipping node entry %d (base %lx)\n",
120 if ((base
>> 8) & 3 || (limit
>> 8) & 3) {
121 pr_err("Node %d using interleaving mode %lx/%lx\n",
122 nodeid
, (base
>> 8) & 3, (limit
>> 8) & 3);
125 if (node_isset(nodeid
, mem_nodes_parsed
)) {
126 pr_info("Node %d already present, skipping\n",
149 pr_err("Empty node %d\n", nodeid
);
153 pr_err("Node %d bogus settings %lx-%lx.\n",
154 nodeid
, base
, limit
);
158 /* Could sort here, but pun for now. Should not happen anyroads. */
159 if (prevbase
> base
) {
160 pr_err("Node map not sorted %lx,%lx\n",
165 pr_info("Node %d MemBase %016lx Limit %016lx\n",
166 nodeid
, base
, limit
);
168 numa_nodes
[nodeid
].start
= base
;
169 numa_nodes
[nodeid
].end
= limit
;
173 node_set(nodeid
, mem_nodes_parsed
);
174 node_set(nodeid
, cpu_nodes_parsed
);
177 if (!nodes_weight(mem_nodes_parsed
))
181 * We seem to have valid NUMA configuration. Map apicids to nodes
182 * using the coreid bits from early_identify_cpu.
184 bits
= boot_cpu_data
.x86_coreid_bits
;
188 /* get the APIC ID of the BSP early for systems with apicid lifting */
189 early_get_boot_cpu_id();
190 if (boot_cpu_physical_apicid
> 0) {
191 pr_info("BSP APIC ID: %02x\n", boot_cpu_physical_apicid
);
192 apicid_base
= boot_cpu_physical_apicid
;
195 for_each_node_mask(i
, cpu_nodes_parsed
)
196 for (j
= apicid_base
; j
< cores
+ apicid_base
; j
++)
197 set_apicid_to_node((i
<< bits
) + j
, i
);
202 #ifdef CONFIG_NUMA_EMU
203 static s16 fake_apicid_to_node
[MAX_LOCAL_APIC
] __initdata
= {
204 [0 ... MAX_LOCAL_APIC
-1] = NUMA_NO_NODE
207 static int __init
find_node_by_addr(unsigned long addr
)
209 int ret
= NUMA_NO_NODE
;
212 for (i
= 0; i
< 8; i
++)
213 if (addr
>= numa_nodes
[i
].start
&& addr
< numa_nodes
[i
].end
) {
221 * For NUMA emulation, fake proximity domain (_PXM) to node id mappings must be
222 * setup to represent the physical topology but reflect the emulated
223 * environment. For each emulated node, the real node which it appears on is
224 * found and a fake pxm to nid mapping is created which mirrors the actual
225 * locality. node_distance() then represents the correct distances between
226 * emulated nodes by using the fake acpi mappings to pxms.
228 void __init
amd_fake_nodes(const struct bootnode
*nodes
, int nr_nodes
)
232 unsigned int apicid_base
= 0;
235 bits
= boot_cpu_data
.x86_coreid_bits
;
237 early_get_boot_cpu_id();
238 if (boot_cpu_physical_apicid
> 0)
239 apicid_base
= boot_cpu_physical_apicid
;
241 for (i
= 0; i
< nr_nodes
; i
++) {
246 nid
= find_node_by_addr(nodes
[i
].start
);
247 if (nid
== NUMA_NO_NODE
)
250 index
= nodeids
[nid
] << bits
;
251 if (fake_apicid_to_node
[index
+ apicid_base
] == NUMA_NO_NODE
)
252 for (j
= apicid_base
; j
< cores
+ apicid_base
; j
++)
253 fake_apicid_to_node
[index
+ j
] = i
;
254 #ifdef CONFIG_ACPI_NUMA
255 __acpi_map_pxm_to_node(nid
, i
);
258 memcpy(__apicid_to_node
, fake_apicid_to_node
, sizeof(__apicid_to_node
));
260 #endif /* CONFIG_NUMA_EMU */
262 int __init
amd_scan_nodes(void)
266 memnode_shift
= compute_hash_shift(numa_nodes
, 8, NULL
);
267 if (memnode_shift
< 0) {
268 pr_err("No NUMA node hash function found. Contact maintainer\n");
271 pr_info("Using node hash shift of %d\n", memnode_shift
);
273 /* use the coreid bits from early_identify_cpu */
274 for_each_node_mask(i
, node_possible_map
)
275 memblock_x86_register_active_regions(i
,
276 numa_nodes
[i
].start
>> PAGE_SHIFT
,
277 numa_nodes
[i
].end
>> PAGE_SHIFT
);
278 init_memory_mapping_high();
279 for_each_node_mask(i
, node_possible_map
)
280 setup_node_bootmem(i
, numa_nodes
[i
].start
, numa_nodes
[i
].end
);