2 * Generic VM initialization for x86-64 NUMA setups.
3 * Copyright 2002,2003 Andi Kleen, SuSE Labs.
5 #include <linux/kernel.h>
7 #include <linux/string.h>
8 #include <linux/init.h>
9 #include <linux/bootmem.h>
10 #include <linux/mmzone.h>
11 #include <linux/ctype.h>
12 #include <linux/module.h>
13 #include <linux/nodemask.h>
16 #include <asm/proto.h>
25 struct pglist_data *node_data[MAX_NUMNODES] __read_mostly;
26 bootmem_data_t plat_node_bdata[MAX_NUMNODES];
29 u8 memnodemap[NODEMAPSIZE];
31 unsigned char cpu_to_node[NR_CPUS] __read_mostly = {
32 [0 ... NR_CPUS-1] = NUMA_NO_NODE
34 unsigned char apicid_to_node[MAX_LOCAL_APIC] __cpuinitdata = {
35 [0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE
37 cpumask_t node_to_cpumask[MAX_NUMNODES] __read_mostly;
39 int numa_off __initdata;
41 int __init compute_hash_shift(struct node *nodes, int numnodes)
45 unsigned long addr,maxend=0;
47 for (i = 0; i < numnodes; i++)
48 if ((nodes[i].start != nodes[i].end) && (nodes[i].end > maxend))
49 maxend = nodes[i].end;
51 while ((1UL << shift) < (maxend / NODEMAPSIZE))
54 printk (KERN_DEBUG"Using %d for the hash shift. Max adder is %lx \n",
56 memset(memnodemap,0xff,sizeof(*memnodemap) * NODEMAPSIZE);
57 for (i = 0; i < numnodes; i++) {
58 if (nodes[i].start == nodes[i].end)
60 for (addr = nodes[i].start;
62 addr += (1UL << shift)) {
63 if (memnodemap[addr >> shift] != 0xff) {
65 "Your memory is not aligned you need to rebuild your kernel "
66 "with a bigger NODEMAPSIZE shift=%d adder=%lu\n",
70 memnodemap[addr >> shift] = i;
76 #ifdef CONFIG_SPARSEMEM
77 int early_pfn_to_nid(unsigned long pfn)
79 return phys_to_nid(pfn << PAGE_SHIFT);
83 /* Initialize bootmem allocator for a node */
84 void __init setup_node_bootmem(int nodeid, unsigned long start, unsigned long end)
86 unsigned long start_pfn, end_pfn, bootmap_pages, bootmap_size, bootmap_start;
87 unsigned long nodedata_phys;
88 const int pgdat_size = round_up(sizeof(pg_data_t), PAGE_SIZE);
90 start = round_up(start, ZONE_ALIGN);
92 printk("Bootmem setup node %d %016lx-%016lx\n", nodeid, start, end);
94 start_pfn = start >> PAGE_SHIFT;
95 end_pfn = end >> PAGE_SHIFT;
97 memory_present(nodeid, start_pfn, end_pfn);
98 nodedata_phys = find_e820_area(start, end, pgdat_size);
99 if (nodedata_phys == -1L)
100 panic("Cannot find memory pgdat in node %d\n", nodeid);
102 Dprintk("nodedata_phys %lx\n", nodedata_phys);
104 node_data[nodeid] = phys_to_virt(nodedata_phys);
105 memset(NODE_DATA(nodeid), 0, sizeof(pg_data_t));
106 NODE_DATA(nodeid)->bdata = &plat_node_bdata[nodeid];
107 NODE_DATA(nodeid)->node_start_pfn = start_pfn;
108 NODE_DATA(nodeid)->node_spanned_pages = end_pfn - start_pfn;
110 /* Find a place for the bootmem map */
111 bootmap_pages = bootmem_bootmap_pages(end_pfn - start_pfn);
112 bootmap_start = round_up(nodedata_phys + pgdat_size, PAGE_SIZE);
113 bootmap_start = find_e820_area(bootmap_start, end, bootmap_pages<<PAGE_SHIFT);
114 if (bootmap_start == -1L)
115 panic("Not enough continuous space for bootmap on node %d", nodeid);
116 Dprintk("bootmap start %lu pages %lu\n", bootmap_start, bootmap_pages);
118 bootmap_size = init_bootmem_node(NODE_DATA(nodeid),
119 bootmap_start >> PAGE_SHIFT,
122 e820_bootmem_free(NODE_DATA(nodeid), start, end);
124 reserve_bootmem_node(NODE_DATA(nodeid), nodedata_phys, pgdat_size);
125 reserve_bootmem_node(NODE_DATA(nodeid), bootmap_start, bootmap_pages<<PAGE_SHIFT);
126 node_set_online(nodeid);
129 /* Initialize final allocator for a zone */
130 void __init setup_node_zones(int nodeid)
132 unsigned long start_pfn, end_pfn;
133 unsigned long zones[MAX_NR_ZONES];
134 unsigned long holes[MAX_NR_ZONES];
135 unsigned long dma_end_pfn;
137 memset(zones, 0, sizeof(unsigned long) * MAX_NR_ZONES);
138 memset(holes, 0, sizeof(unsigned long) * MAX_NR_ZONES);
140 start_pfn = node_start_pfn(nodeid);
141 end_pfn = node_end_pfn(nodeid);
143 Dprintk(KERN_INFO "setting up node %d %lx-%lx\n", nodeid, start_pfn, end_pfn);
145 /* All nodes > 0 have a zero length zone DMA */
146 dma_end_pfn = __pa(MAX_DMA_ADDRESS) >> PAGE_SHIFT;
147 if (start_pfn < dma_end_pfn) {
148 zones[ZONE_DMA] = dma_end_pfn - start_pfn;
149 holes[ZONE_DMA] = e820_hole_size(start_pfn, dma_end_pfn);
150 zones[ZONE_NORMAL] = end_pfn - dma_end_pfn;
151 holes[ZONE_NORMAL] = e820_hole_size(dma_end_pfn, end_pfn);
154 zones[ZONE_NORMAL] = end_pfn - start_pfn;
155 holes[ZONE_NORMAL] = e820_hole_size(start_pfn, end_pfn);
158 free_area_init_node(nodeid, NODE_DATA(nodeid), zones,
162 void __init numa_init_array(void)
165 /* There are unfortunately some poorly designed mainboards around
166 that only connect memory to a single CPU. This breaks the 1:1 cpu->node
167 mapping. To avoid this fill in the mapping for all possible
168 CPUs, as the number of CPUs is not known yet.
169 We round robin the existing nodes. */
170 rr = first_node(node_online_map);
171 for (i = 0; i < NR_CPUS; i++) {
172 if (cpu_to_node[i] != NUMA_NO_NODE)
175 rr = next_node(rr, node_online_map);
176 if (rr == MAX_NUMNODES)
177 rr = first_node(node_online_map);
182 #ifdef CONFIG_NUMA_EMU
183 int numa_fake __initdata = 0;
186 static int numa_emulation(unsigned long start_pfn, unsigned long end_pfn)
189 struct node nodes[MAX_NUMNODES];
190 unsigned long sz = ((end_pfn - start_pfn)<<PAGE_SHIFT) / numa_fake;
192 /* Kludge needed for the hash function */
193 if (hweight64(sz) > 1) {
195 while ((x << 1) < sz)
198 printk("Numa emulation unbalanced. Complain to maintainer\n");
202 memset(&nodes,0,sizeof(nodes));
203 for (i = 0; i < numa_fake; i++) {
204 nodes[i].start = (start_pfn<<PAGE_SHIFT) + i*sz;
205 if (i == numa_fake-1)
206 sz = (end_pfn<<PAGE_SHIFT) - nodes[i].start;
207 nodes[i].end = nodes[i].start + sz;
208 if (i != numa_fake-1)
210 printk(KERN_INFO "Faking node %d at %016Lx-%016Lx (%LuMB)\n",
212 nodes[i].start, nodes[i].end,
213 (nodes[i].end - nodes[i].start) >> 20);
216 memnode_shift = compute_hash_shift(nodes, numa_fake);
217 if (memnode_shift < 0) {
219 printk(KERN_ERR "No NUMA hash function found. Emulation disabled.\n");
222 for_each_online_node(i)
223 setup_node_bootmem(i, nodes[i].start, nodes[i].end);
229 void __init numa_initmem_init(unsigned long start_pfn, unsigned long end_pfn)
233 #ifdef CONFIG_NUMA_EMU
234 if (numa_fake && !numa_emulation(start_pfn, end_pfn))
238 #ifdef CONFIG_ACPI_NUMA
239 if (!numa_off && !acpi_scan_nodes(start_pfn << PAGE_SHIFT,
240 end_pfn << PAGE_SHIFT))
244 #ifdef CONFIG_K8_NUMA
245 if (!numa_off && !k8_scan_nodes(start_pfn<<PAGE_SHIFT, end_pfn<<PAGE_SHIFT))
248 printk(KERN_INFO "%s\n",
249 numa_off ? "NUMA turned off" : "No NUMA configuration found");
251 printk(KERN_INFO "Faking a node at %016lx-%016lx\n",
252 start_pfn << PAGE_SHIFT,
253 end_pfn << PAGE_SHIFT);
254 /* setup dummy node covering all memory */
257 nodes_clear(node_online_map);
259 for (i = 0; i < NR_CPUS; i++)
261 node_to_cpumask[0] = cpumask_of_cpu(0);
262 setup_node_bootmem(0, start_pfn << PAGE_SHIFT, end_pfn << PAGE_SHIFT);
265 __cpuinit void numa_add_cpu(int cpu)
267 set_bit(cpu, &node_to_cpumask[cpu_to_node(cpu)]);
270 unsigned long __init numa_free_all_bootmem(void)
273 unsigned long pages = 0;
274 for_each_online_node(i) {
275 pages += free_all_bootmem_node(NODE_DATA(i));
280 void __init paging_init(void)
283 for_each_online_node(i) {
289 __init int numa_setup(char *opt)
291 if (!strncmp(opt,"off",3))
293 #ifdef CONFIG_NUMA_EMU
294 if(!strncmp(opt, "fake=", 5)) {
295 numa_fake = simple_strtoul(opt+5,NULL,0); ;
296 if (numa_fake >= MAX_NUMNODES)
297 numa_fake = MAX_NUMNODES;
300 #ifdef CONFIG_ACPI_NUMA
301 if (!strncmp(opt,"noacpi",6))
307 EXPORT_SYMBOL(cpu_to_node);
308 EXPORT_SYMBOL(node_to_cpumask);
309 EXPORT_SYMBOL(memnode_shift);
310 EXPORT_SYMBOL(memnodemap);
311 EXPORT_SYMBOL(node_data);