2 * ACPI 3.0 based NUMA setup
3 * Copyright 2004 Andi Kleen, SuSE Labs.
5 * Reads the ACPI SRAT table to figure out what memory belongs to which CPUs.
7 * Called from acpi_numa_init while reading the SRAT and SLIT tables.
8 * Assumes all memory regions belonging to a single proximity domain
9 * are in one chunk. Holes between them will be included in the node.
12 #include <linux/kernel.h>
13 #include <linux/acpi.h>
14 #include <linux/mmzone.h>
15 #include <linux/bitmap.h>
16 #include <linux/module.h>
17 #include <linux/topology.h>
18 #include <asm/proto.h>
22 static struct acpi_table_slit *acpi_slit;
24 static nodemask_t nodes_parsed __initdata;
25 static nodemask_t nodes_found __initdata;
26 static struct node nodes[MAX_NUMNODES] __initdata;
27 static u8 pxm2node[256] = { [0 ... 255] = 0xff };
29 static int node_to_pxm(int n);
31 int pxm_to_node(int pxm)
33 if ((unsigned)pxm >= 256)
35 /* Extend 0xff to (int)-1 */
36 return (signed char)pxm2node[pxm];
39 static __init int setup_node(int pxm)
41 unsigned node = pxm2node[pxm];
43 if (nodes_weight(nodes_found) >= MAX_NUMNODES)
45 node = first_unset_node(nodes_found);
46 node_set(node, nodes_found);
52 static __init int conflicting_nodes(unsigned long start, unsigned long end)
55 for_each_node_mask(i, nodes_parsed) {
56 struct node *nd = &nodes[i];
57 if (nd->start == nd->end)
59 if (nd->end > start && nd->start < end)
61 if (nd->end == end && nd->start == start)
67 static __init void cutoff_node(int i, unsigned long start, unsigned long end)
69 struct node *nd = &nodes[i];
70 if (nd->start < start) {
72 if (nd->end < nd->start)
77 if (nd->start > nd->end)
82 static __init void bad_srat(void)
85 printk(KERN_ERR "SRAT: SRAT not used.\n");
87 for (i = 0; i < MAX_LOCAL_APIC; i++)
88 apicid_to_node[i] = NUMA_NO_NODE;
91 static __init inline int srat_disabled(void)
93 return numa_off || acpi_numa < 0;
97 * A lot of BIOS fill in 10 (= no distance) everywhere. This messes
98 * up the NUMA heuristics which wants the local node to have a smaller
99 * distance than the others.
100 * Do some quick checks here and only use the SLIT if it passes.
102 static __init int slit_valid(struct acpi_table_slit *slit)
105 int d = slit->localities;
106 for (i = 0; i < d; i++) {
107 for (j = 0; j < d; j++) {
108 u8 val = slit->entry[d*i + j];
112 } else if (val <= 10)
119 /* Callback for SLIT parsing */
120 void __init acpi_numa_slit_init(struct acpi_table_slit *slit)
122 if (!slit_valid(slit)) {
123 printk(KERN_INFO "ACPI: SLIT table looks invalid. Not used.\n");
129 /* Callback for Proximity Domain -> LAPIC mapping */
131 acpi_numa_processor_affinity_init(struct acpi_table_processor_affinity *pa)
134 if (srat_disabled() || pa->flags.enabled == 0)
136 pxm = pa->proximity_domain;
137 node = setup_node(pxm);
139 printk(KERN_ERR "SRAT: Too many proximity domains %x\n", pxm);
143 apicid_to_node[pa->apic_id] = node;
145 printk(KERN_INFO "SRAT: PXM %u -> APIC %u -> Node %u\n",
146 pxm, pa->apic_id, node);
149 /* Callback for parsing of the Proximity Domain <-> Memory Area mappings */
151 acpi_numa_memory_affinity_init(struct acpi_table_memory_affinity *ma)
154 unsigned long start, end;
158 if (srat_disabled() || ma->flags.enabled == 0)
160 pxm = ma->proximity_domain;
161 node = setup_node(pxm);
163 printk(KERN_ERR "SRAT: Too many proximity domains.\n");
167 start = ma->base_addr_lo | ((u64)ma->base_addr_hi << 32);
168 end = start + (ma->length_lo | ((u64)ma->length_hi << 32));
169 /* It is fine to add this area to the nodes data it will be used later*/
170 if (ma->flags.hot_pluggable == 1)
171 printk(KERN_INFO "SRAT: hot plug zone found %lx - %lx \n",
173 i = conflicting_nodes(start, end);
176 "SRAT: Warning: PXM %d (%lx-%lx) overlaps with itself (%Lx-%Lx)\n",
177 pxm, start, end, nodes[i].start, nodes[i].end);
180 "SRAT: PXM %d (%lx-%lx) overlaps with PXM %d (%Lx-%Lx)\n",
181 pxm, start, end, node_to_pxm(i),
182 nodes[i].start, nodes[i].end);
187 if (!node_test_and_set(node, nodes_parsed)) {
191 if (start < nd->start)
196 printk(KERN_INFO "SRAT: Node %u PXM %u %Lx-%Lx\n", node, pxm,
200 /* Sanity check to catch more bad SRATs (they are amazingly common).
201 Make sure the PXMs cover all memory. */
202 static int nodes_cover_memory(void)
205 unsigned long pxmram, e820ram;
208 for_each_node_mask(i, nodes_parsed) {
209 unsigned long s = nodes[i].start >> PAGE_SHIFT;
210 unsigned long e = nodes[i].end >> PAGE_SHIFT;
212 pxmram -= e820_hole_size(s, e);
215 e820ram = end_pfn - e820_hole_size(0, end_pfn);
216 if (pxmram < e820ram) {
218 "SRAT: PXMs only cover %luMB of your %luMB e820 RAM. Not used.\n",
219 (pxmram << PAGE_SHIFT) >> 20,
220 (e820ram << PAGE_SHIFT) >> 20);
226 void __init acpi_numa_arch_fixup(void) {}
228 /* Use the information discovered above to actually set up the nodes. */
229 int __init acpi_scan_nodes(unsigned long start, unsigned long end)
236 /* First clean up the node list */
237 for_each_node_mask(i, nodes_parsed) {
238 cutoff_node(i, start, end);
239 if (nodes[i].start == nodes[i].end)
240 node_clear(i, nodes_parsed);
243 if (!nodes_cover_memory()) {
248 memnode_shift = compute_hash_shift(nodes, nodes_weight(nodes_parsed));
249 if (memnode_shift < 0) {
251 "SRAT: No NUMA node hash function found. Contact maintainer\n");
256 /* Finally register nodes */
257 for_each_node_mask(i, nodes_parsed)
258 setup_node_bootmem(i, nodes[i].start, nodes[i].end);
259 for (i = 0; i < NR_CPUS; i++) {
260 if (cpu_to_node[i] == NUMA_NO_NODE)
262 if (!node_isset(cpu_to_node[i], nodes_parsed))
263 numa_set_node(i, NUMA_NO_NODE);
269 static int node_to_pxm(int n)
272 if (pxm2node[n] == n)
274 for (i = 0; i < 256; i++)
275 if (pxm2node[i] == n)
280 int __node_distance(int a, int b)
285 return a == b ? 10 : 20;
286 index = acpi_slit->localities * node_to_pxm(a);
287 return acpi_slit->entry[index + node_to_pxm(b)];
290 EXPORT_SYMBOL(__node_distance);