Commit | Line | Data |
---|---|---|
3aa88cdf | 1 | /* |
eec1d4fa | 2 | * AMD NUMA support. |
1da177e4 | 3 | * Discover the memory map and associated nodes. |
3aa88cdf | 4 | * |
eec1d4fa | 5 | * This version reads it directly from the AMD northbridge. |
3aa88cdf | 6 | * |
1da177e4 LT |
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> | |
a9ce6bc1 YL |
14 | #include <linux/memblock.h> |
15 | ||
1da177e4 LT |
16 | #include <asm/io.h> |
17 | #include <linux/pci_ids.h> | |
cbf9bd60 | 18 | #include <linux/acpi.h> |
1da177e4 LT |
19 | #include <asm/types.h> |
20 | #include <asm/mmzone.h> | |
21 | #include <asm/proto.h> | |
22 | #include <asm/e820.h> | |
23 | #include <asm/pci-direct.h> | |
24 | #include <asm/numa.h> | |
cbf9bd60 YL |
25 | #include <asm/mpspec.h> |
26 | #include <asm/apic.h> | |
23ac4ae8 | 27 | #include <asm/amd_nb.h> |
1da177e4 | 28 | |
8ee2debc DR |
29 | static struct bootnode __initdata nodes[8]; |
30 | static nodemask_t __initdata nodes_parsed = NODE_MASK_NONE; | |
31 | ||
1da177e4 LT |
32 | static __init int find_northbridge(void) |
33 | { | |
3aa88cdf | 34 | int num; |
1da177e4 | 35 | |
3aa88cdf | 36 | for (num = 0; num < 32; num++) { |
1da177e4 | 37 | u32 header; |
3aa88cdf CM |
38 | |
39 | header = read_pci_config(0, num, 0, 0x00); | |
bb4a1d64 JD |
40 | if (header != (PCI_VENDOR_ID_AMD | (0x1100<<16)) && |
41 | header != (PCI_VENDOR_ID_AMD | (0x1200<<16)) && | |
42 | header != (PCI_VENDOR_ID_AMD | (0x1300<<16))) | |
3aa88cdf | 43 | continue; |
1da177e4 | 44 | |
3aa88cdf | 45 | header = read_pci_config(0, num, 1, 0x00); |
bb4a1d64 JD |
46 | if (header != (PCI_VENDOR_ID_AMD | (0x1101<<16)) && |
47 | header != (PCI_VENDOR_ID_AMD | (0x1201<<16)) && | |
48 | header != (PCI_VENDOR_ID_AMD | (0x1301<<16))) | |
3aa88cdf CM |
49 | continue; |
50 | return num; | |
51 | } | |
1da177e4 | 52 | |
3aa88cdf | 53 | return -1; |
1da177e4 LT |
54 | } |
55 | ||
cbf9bd60 YL |
56 | static __init void early_get_boot_cpu_id(void) |
57 | { | |
58 | /* | |
f6e9456c | 59 | * need to get the APIC ID of the BSP so can use that to |
eec1d4fa | 60 | * create apicid_to_node in amd_scan_nodes() |
cbf9bd60 | 61 | */ |
a4caa18e | 62 | #ifdef CONFIG_X86_MPPARSE |
cbf9bd60 YL |
63 | /* |
64 | * get boot-time SMP configuration: | |
65 | */ | |
66 | if (smp_found_config) | |
67 | early_get_smp_config(); | |
a4caa18e | 68 | #endif |
cbf9bd60 YL |
69 | early_init_lapic_mapping(); |
70 | } | |
71 | ||
eec1d4fa | 72 | int __init amd_get_nodes(struct bootnode *physnodes) |
3aa88cdf | 73 | { |
8ee2debc DR |
74 | int i; |
75 | int ret = 0; | |
76 | ||
77 | for_each_node_mask(i, nodes_parsed) { | |
78 | physnodes[ret].start = nodes[i].start; | |
79 | physnodes[ret].end = nodes[i].end; | |
80 | ret++; | |
81 | } | |
82 | return ret; | |
83 | } | |
84 | ||
eec1d4fa | 85 | int __init amd_numa_init(unsigned long start_pfn, unsigned long end_pfn) |
8ee2debc DR |
86 | { |
87 | unsigned long start = PFN_PHYS(start_pfn); | |
88 | unsigned long end = PFN_PHYS(end_pfn); | |
89 | unsigned numnodes; | |
1da177e4 | 90 | unsigned long prevbase; |
8ee2debc | 91 | int i, nb, found = 0; |
d34c0895 | 92 | u32 nodeid, reg; |
1da177e4 | 93 | |
0637a70a AK |
94 | if (!early_pci_allowed()) |
95 | return -1; | |
96 | ||
3aa88cdf CM |
97 | nb = find_northbridge(); |
98 | if (nb < 0) | |
1da177e4 LT |
99 | return nb; |
100 | ||
1af5ba51 | 101 | pr_info("Scanning NUMA topology in Northbridge %d\n", nb); |
1da177e4 | 102 | |
3aa88cdf | 103 | reg = read_pci_config(0, nb, 0, 0x60); |
1da177e4 | 104 | numnodes = ((reg >> 4) & 0xF) + 1; |
3bea9c97 AK |
105 | if (numnodes <= 1) |
106 | return -1; | |
1da177e4 | 107 | |
8ee2debc | 108 | pr_info("Number of physical nodes %d\n", numnodes); |
1da177e4 | 109 | |
1da177e4 | 110 | prevbase = 0; |
3aa88cdf CM |
111 | for (i = 0; i < 8; i++) { |
112 | unsigned long base, limit; | |
3aa88cdf | 113 | |
1da177e4 LT |
114 | base = read_pci_config(0, nb, 1, 0x40 + i*8); |
115 | limit = read_pci_config(0, nb, 1, 0x44 + i*8); | |
116 | ||
3aa88cdf | 117 | nodeid = limit & 7; |
3aa88cdf | 118 | if ((base & 3) == 0) { |
1da177e4 | 119 | if (i < numnodes) |
1af5ba51 | 120 | pr_info("Skipping disabled node %d\n", i); |
1da177e4 | 121 | continue; |
3aa88cdf | 122 | } |
1da177e4 | 123 | if (nodeid >= numnodes) { |
1af5ba51 DR |
124 | pr_info("Ignoring excess node %d (%lx:%lx)\n", nodeid, |
125 | base, limit); | |
1da177e4 | 126 | continue; |
3aa88cdf | 127 | } |
1da177e4 | 128 | |
3aa88cdf | 129 | if (!limit) { |
1af5ba51 DR |
130 | pr_info("Skipping node entry %d (base %lx)\n", |
131 | i, base); | |
1da177e4 LT |
132 | continue; |
133 | } | |
134 | if ((base >> 8) & 3 || (limit >> 8) & 3) { | |
1af5ba51 DR |
135 | pr_err("Node %d using interleaving mode %lx/%lx\n", |
136 | nodeid, (base >> 8) & 3, (limit >> 8) & 3); | |
3aa88cdf CM |
137 | return -1; |
138 | } | |
8ee2debc | 139 | if (node_isset(nodeid, nodes_parsed)) { |
1af5ba51 DR |
140 | pr_info("Node %d already present, skipping\n", |
141 | nodeid); | |
1da177e4 LT |
142 | continue; |
143 | } | |
144 | ||
3aa88cdf CM |
145 | limit >>= 16; |
146 | limit <<= 24; | |
1da177e4 | 147 | limit |= (1<<24)-1; |
ffd10a2b | 148 | limit++; |
1da177e4 | 149 | |
8ee2debc DR |
150 | if (limit > end) |
151 | limit = end; | |
1da177e4 | 152 | if (limit <= base) |
3aa88cdf CM |
153 | continue; |
154 | ||
1da177e4 | 155 | base >>= 16; |
3aa88cdf CM |
156 | base <<= 24; |
157 | ||
158 | if (base < start) | |
159 | base = start; | |
160 | if (limit > end) | |
161 | limit = end; | |
162 | if (limit == base) { | |
1af5ba51 | 163 | pr_err("Empty node %d\n", nodeid); |
3aa88cdf | 164 | continue; |
1da177e4 | 165 | } |
3aa88cdf | 166 | if (limit < base) { |
1af5ba51 | 167 | pr_err("Node %d bogus settings %lx-%lx.\n", |
3aa88cdf | 168 | nodeid, base, limit); |
1da177e4 | 169 | continue; |
3aa88cdf CM |
170 | } |
171 | ||
1da177e4 | 172 | /* Could sort here, but pun for now. Should not happen anyroads. */ |
3aa88cdf | 173 | if (prevbase > base) { |
1af5ba51 | 174 | pr_err("Node map not sorted %lx,%lx\n", |
3aa88cdf | 175 | prevbase, base); |
1da177e4 LT |
176 | return -1; |
177 | } | |
3aa88cdf | 178 | |
1af5ba51 DR |
179 | pr_info("Node %d MemBase %016lx Limit %016lx\n", |
180 | nodeid, base, limit); | |
3aa88cdf | 181 | |
1da177e4 | 182 | found++; |
3aa88cdf CM |
183 | |
184 | nodes[nodeid].start = base; | |
1da177e4 LT |
185 | nodes[nodeid].end = limit; |
186 | ||
187 | prevbase = base; | |
188 | ||
8ee2debc | 189 | node_set(nodeid, nodes_parsed); |
3aa88cdf | 190 | } |
1da177e4 LT |
191 | |
192 | if (!found) | |
3aa88cdf | 193 | return -1; |
8ee2debc DR |
194 | return 0; |
195 | } | |
1da177e4 | 196 | |
eec1d4fa | 197 | int __init amd_scan_nodes(void) |
8ee2debc DR |
198 | { |
199 | unsigned int bits; | |
200 | unsigned int cores; | |
201 | unsigned int apicid_base; | |
202 | int i; | |
203 | ||
204 | BUG_ON(nodes_empty(nodes_parsed)); | |
205 | node_possible_map = nodes_parsed; | |
6ec6e0d9 | 206 | memnode_shift = compute_hash_shift(nodes, 8, NULL); |
3aa88cdf | 207 | if (memnode_shift < 0) { |
1af5ba51 | 208 | pr_err("No NUMA node hash function found. Contact maintainer\n"); |
3aa88cdf CM |
209 | return -1; |
210 | } | |
1af5ba51 | 211 | pr_info("Using node hash shift of %d\n", memnode_shift); |
1da177e4 | 212 | |
3f6e5a12 YL |
213 | /* use the coreid bits from early_identify_cpu */ |
214 | bits = boot_cpu_data.x86_coreid_bits; | |
215 | cores = (1<<bits); | |
cbf9bd60 | 216 | apicid_base = 0; |
f6e9456c | 217 | /* get the APIC ID of the BSP early for systems with apicid lifting */ |
cbf9bd60 YL |
218 | early_get_boot_cpu_id(); |
219 | if (boot_cpu_physical_apicid > 0) { | |
1af5ba51 | 220 | pr_info("BSP APIC ID: %02x\n", boot_cpu_physical_apicid); |
cbf9bd60 YL |
221 | apicid_base = boot_cpu_physical_apicid; |
222 | } | |
3f6e5a12 | 223 | |
1411e0ec | 224 | for_each_node_mask(i, node_possible_map) |
a9ce6bc1 | 225 | memblock_x86_register_active_regions(i, |
40bcc69b YL |
226 | nodes[i].start >> PAGE_SHIFT, |
227 | nodes[i].end >> PAGE_SHIFT); | |
1411e0ec YL |
228 | init_memory_mapping_high(); |
229 | for_each_node_mask(i, node_possible_map) { | |
230 | int j; | |
231 | ||
40bcc69b YL |
232 | for (j = apicid_base; j < cores + apicid_base; j++) |
233 | apicid_to_node[(i << bits) + j] = i; | |
234 | setup_node_bootmem(i, nodes[i].start, nodes[i].end); | |
1da177e4 LT |
235 | } |
236 | ||
237 | numa_init_array(); | |
238 | return 0; | |
3aa88cdf | 239 | } |