b9e2dbfe55c36feda159e6c97c102d546d38328c
[linux-2.6.git] / 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 #include <asm/k8.h>
26
27 static struct bootnode __initdata nodes[8];
28 static nodemask_t __initdata nodes_parsed = NODE_MASK_NONE;
29
30 static __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
54 static __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         /*
61          * Find possible boot-time SMP configuration:
62          */
63 #ifdef CONFIG_X86_MPPARSE
64         early_find_smp_config();
65 #endif
66 #ifdef CONFIG_ACPI
67         /*
68          * Read APIC information from ACPI tables.
69          */
70         early_acpi_boot_init();
71 #endif
72 #ifdef CONFIG_X86_MPPARSE
73         /*
74          * get boot-time SMP configuration:
75          */
76         if (smp_found_config)
77                 early_get_smp_config();
78 #endif
79         early_init_lapic_mapping();
80 }
81
82 int __init k8_get_nodes(struct bootnode *physnodes)
83 {
84         int i;
85         int ret = 0;
86
87         for_each_node_mask(i, nodes_parsed) {
88                 physnodes[ret].start = nodes[i].start;
89                 physnodes[ret].end = nodes[i].end;
90                 ret++;
91         }
92         return ret;
93 }
94
95 int __init k8_numa_init(unsigned long start_pfn, unsigned long end_pfn)
96 {
97         unsigned long start = PFN_PHYS(start_pfn);
98         unsigned long end = PFN_PHYS(end_pfn);
99         unsigned numnodes;
100         unsigned long prevbase;
101         int i, nb, found = 0;
102         u32 nodeid, reg;
103
104         if (!early_pci_allowed())
105                 return -1;
106
107         nb = find_northbridge();
108         if (nb < 0)
109                 return nb;
110
111         pr_info("Scanning NUMA topology in Northbridge %d\n", nb);
112
113         reg = read_pci_config(0, nb, 0, 0x60);
114         numnodes = ((reg >> 4) & 0xF) + 1;
115         if (numnodes <= 1)
116                 return -1;
117
118         pr_info("Number of physical nodes %d\n", numnodes);
119
120         prevbase = 0;
121         for (i = 0; i < 8; i++) {
122                 unsigned long base, limit;
123
124                 base = read_pci_config(0, nb, 1, 0x40 + i*8);
125                 limit = read_pci_config(0, nb, 1, 0x44 + i*8);
126
127                 nodeid = limit & 7;
128                 if ((base & 3) == 0) {
129                         if (i < numnodes)
130                                 pr_info("Skipping disabled node %d\n", i);
131                         continue;
132                 }
133                 if (nodeid >= numnodes) {
134                         pr_info("Ignoring excess node %d (%lx:%lx)\n", nodeid,
135                                 base, limit);
136                         continue;
137                 }
138
139                 if (!limit) {
140                         pr_info("Skipping node entry %d (base %lx)\n",
141                                 i, base);
142                         continue;
143                 }
144                 if ((base >> 8) & 3 || (limit >> 8) & 3) {
145                         pr_err("Node %d using interleaving mode %lx/%lx\n",
146                                nodeid, (base >> 8) & 3, (limit >> 8) & 3);
147                         return -1;
148                 }
149                 if (node_isset(nodeid, nodes_parsed)) {
150                         pr_info("Node %d already present, skipping\n",
151                                 nodeid);
152                         continue;
153                 }
154
155                 limit >>= 16;
156                 limit <<= 24;
157                 limit |= (1<<24)-1;
158                 limit++;
159
160                 if (limit > end)
161                         limit = end;
162                 if (limit <= base)
163                         continue;
164
165                 base >>= 16;
166                 base <<= 24;
167
168                 if (base < start)
169                         base = start;
170                 if (limit > end)
171                         limit = end;
172                 if (limit == base) {
173                         pr_err("Empty node %d\n", nodeid);
174                         continue;
175                 }
176                 if (limit < base) {
177                         pr_err("Node %d bogus settings %lx-%lx.\n",
178                                nodeid, base, limit);
179                         continue;
180                 }
181
182                 /* Could sort here, but pun for now. Should not happen anyroads. */
183                 if (prevbase > base) {
184                         pr_err("Node map not sorted %lx,%lx\n",
185                                prevbase, base);
186                         return -1;
187                 }
188
189                 pr_info("Node %d MemBase %016lx Limit %016lx\n",
190                         nodeid, base, limit);
191
192                 found++;
193
194                 nodes[nodeid].start = base;
195                 nodes[nodeid].end = limit;
196
197                 prevbase = base;
198
199                 node_set(nodeid, nodes_parsed);
200         }
201
202         if (!found)
203                 return -1;
204         return 0;
205 }
206
207 int __init k8_scan_nodes(void)
208 {
209         unsigned int bits;
210         unsigned int cores;
211         unsigned int apicid_base;
212         int i;
213
214         BUG_ON(nodes_empty(nodes_parsed));
215         node_possible_map = nodes_parsed;
216         memnode_shift = compute_hash_shift(nodes, 8, NULL);
217         if (memnode_shift < 0) {
218                 pr_err("No NUMA node hash function found. Contact maintainer\n");
219                 return -1;
220         }
221         pr_info("Using node hash shift of %d\n", memnode_shift);
222
223         /* use the coreid bits from early_identify_cpu */
224         bits = boot_cpu_data.x86_coreid_bits;
225         cores = (1<<bits);
226         apicid_base = 0;
227         /* need to get boot_cpu_id early for system with apicid lifting */
228         early_get_boot_cpu_id();
229         if (boot_cpu_physical_apicid > 0) {
230                 pr_info("BSP APIC ID: %02x\n", boot_cpu_physical_apicid);
231                 apicid_base = boot_cpu_physical_apicid;
232         }
233
234         for_each_node_mask(i, node_possible_map) {
235                 int j;
236
237                 e820_register_active_regions(i,
238                                 nodes[i].start >> PAGE_SHIFT,
239                                 nodes[i].end >> PAGE_SHIFT);
240                 for (j = apicid_base; j < cores + apicid_base; j++)
241                         apicid_to_node[(i << bits) + j] = i;
242                 setup_node_bootmem(i, nodes[i].start, nodes[i].end);
243         }
244
245         numa_init_array();
246         return 0;
247 }