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 = { [0 ... NR_CPUS-1] = NUMA_NO_NODE };
32 cpumask_t node_to_cpumask[MAX_NUMNODES] __read_mostly;
34 int numa_off __initdata;
36 int __init compute_hash_shift(struct node *nodes, int numnodes)
40 unsigned long addr,maxend=0;
42 for (i = 0; i < numnodes; i++)
43 if ((nodes[i].start != nodes[i].end) && (nodes[i].end > maxend))
44 maxend = nodes[i].end;
46 while ((1UL << shift) < (maxend / NODEMAPSIZE))
49 printk (KERN_DEBUG"Using %d for the hash shift. Max adder is %lx \n",
51 memset(memnodemap,0xff,sizeof(*memnodemap) * NODEMAPSIZE);
52 for (i = 0; i < numnodes; i++) {
53 if (nodes[i].start == nodes[i].end)
55 for (addr = nodes[i].start;
57 addr += (1UL << shift)) {
58 if (memnodemap[addr >> shift] != 0xff) {
60 "Your memory is not aligned you need to rebuild your kernel "
61 "with a bigger NODEMAPSIZE shift=%d adder=%lu\n",
65 memnodemap[addr >> shift] = i;
71 #ifdef CONFIG_SPARSEMEM
72 int early_pfn_to_nid(unsigned long pfn)
74 return phys_to_nid(pfn << PAGE_SHIFT);
78 /* Initialize bootmem allocator for a node */
79 void __init setup_node_bootmem(int nodeid, unsigned long start, unsigned long end)
81 unsigned long start_pfn, end_pfn, bootmap_pages, bootmap_size, bootmap_start;
82 unsigned long nodedata_phys;
83 const int pgdat_size = round_up(sizeof(pg_data_t), PAGE_SIZE);
85 start = round_up(start, ZONE_ALIGN);
87 printk("Bootmem setup node %d %016lx-%016lx\n", nodeid, start, end);
89 start_pfn = start >> PAGE_SHIFT;
90 end_pfn = end >> PAGE_SHIFT;
92 memory_present(nodeid, start_pfn, end_pfn);
93 nodedata_phys = find_e820_area(start, end, pgdat_size);
94 if (nodedata_phys == -1L)
95 panic("Cannot find memory pgdat in node %d\n", nodeid);
97 Dprintk("nodedata_phys %lx\n", nodedata_phys);
99 node_data[nodeid] = phys_to_virt(nodedata_phys);
100 memset(NODE_DATA(nodeid), 0, sizeof(pg_data_t));
101 NODE_DATA(nodeid)->bdata = &plat_node_bdata[nodeid];
102 NODE_DATA(nodeid)->node_start_pfn = start_pfn;
103 NODE_DATA(nodeid)->node_spanned_pages = end_pfn - start_pfn;
105 /* Find a place for the bootmem map */
106 bootmap_pages = bootmem_bootmap_pages(end_pfn - start_pfn);
107 bootmap_start = round_up(nodedata_phys + pgdat_size, PAGE_SIZE);
108 bootmap_start = find_e820_area(bootmap_start, end, bootmap_pages<<PAGE_SHIFT);
109 if (bootmap_start == -1L)
110 panic("Not enough continuous space for bootmap on node %d", nodeid);
111 Dprintk("bootmap start %lu pages %lu\n", bootmap_start, bootmap_pages);
113 bootmap_size = init_bootmem_node(NODE_DATA(nodeid),
114 bootmap_start >> PAGE_SHIFT,
117 e820_bootmem_free(NODE_DATA(nodeid), start, end);
119 reserve_bootmem_node(NODE_DATA(nodeid), nodedata_phys, pgdat_size);
120 reserve_bootmem_node(NODE_DATA(nodeid), bootmap_start, bootmap_pages<<PAGE_SHIFT);
121 node_set_online(nodeid);
124 /* Initialize final allocator for a zone */
125 void __init setup_node_zones(int nodeid)
127 unsigned long start_pfn, end_pfn;
128 unsigned long zones[MAX_NR_ZONES];
129 unsigned long holes[MAX_NR_ZONES];
130 unsigned long dma_end_pfn;
132 memset(zones, 0, sizeof(unsigned long) * MAX_NR_ZONES);
133 memset(holes, 0, sizeof(unsigned long) * MAX_NR_ZONES);
135 start_pfn = node_start_pfn(nodeid);
136 end_pfn = node_end_pfn(nodeid);
138 Dprintk(KERN_INFO "setting up node %d %lx-%lx\n", nodeid, start_pfn, end_pfn);
140 /* All nodes > 0 have a zero length zone DMA */
141 dma_end_pfn = __pa(MAX_DMA_ADDRESS) >> PAGE_SHIFT;
142 if (start_pfn < dma_end_pfn) {
143 zones[ZONE_DMA] = dma_end_pfn - start_pfn;
144 holes[ZONE_DMA] = e820_hole_size(start_pfn, dma_end_pfn);
145 zones[ZONE_NORMAL] = end_pfn - dma_end_pfn;
146 holes[ZONE_NORMAL] = e820_hole_size(dma_end_pfn, end_pfn);
149 zones[ZONE_NORMAL] = end_pfn - start_pfn;
150 holes[ZONE_NORMAL] = e820_hole_size(start_pfn, end_pfn);
153 free_area_init_node(nodeid, NODE_DATA(nodeid), zones,
157 void __init numa_init_array(void)
160 /* There are unfortunately some poorly designed mainboards around
161 that only connect memory to a single CPU. This breaks the 1:1 cpu->node
162 mapping. To avoid this fill in the mapping for all possible
163 CPUs, as the number of CPUs is not known yet.
164 We round robin the existing nodes. */
166 for (i = 0; i < NR_CPUS; i++) {
167 if (cpu_to_node[i] != NUMA_NO_NODE)
169 rr = next_node(rr, node_online_map);
170 if (rr == MAX_NUMNODES)
171 rr = first_node(node_online_map);
176 set_bit(0, &node_to_cpumask[cpu_to_node(0)]);
179 #ifdef CONFIG_NUMA_EMU
180 int numa_fake __initdata = 0;
183 static int numa_emulation(unsigned long start_pfn, unsigned long end_pfn)
186 struct node nodes[MAX_NUMNODES];
187 unsigned long sz = ((end_pfn - start_pfn)<<PAGE_SHIFT) / numa_fake;
189 /* Kludge needed for the hash function */
190 if (hweight64(sz) > 1) {
192 while ((x << 1) < sz)
195 printk("Numa emulation unbalanced. Complain to maintainer\n");
199 memset(&nodes,0,sizeof(nodes));
200 for (i = 0; i < numa_fake; i++) {
201 nodes[i].start = (start_pfn<<PAGE_SHIFT) + i*sz;
202 if (i == numa_fake-1)
203 sz = (end_pfn<<PAGE_SHIFT) - nodes[i].start;
204 nodes[i].end = nodes[i].start + sz;
205 if (i != numa_fake-1)
207 printk(KERN_INFO "Faking node %d at %016Lx-%016Lx (%LuMB)\n",
209 nodes[i].start, nodes[i].end,
210 (nodes[i].end - nodes[i].start) >> 20);
213 memnode_shift = compute_hash_shift(nodes, numa_fake);
214 if (memnode_shift < 0) {
216 printk(KERN_ERR "No NUMA hash function found. Emulation disabled.\n");
219 for_each_online_node(i)
220 setup_node_bootmem(i, nodes[i].start, nodes[i].end);
226 void __init numa_initmem_init(unsigned long start_pfn, unsigned long end_pfn)
230 #ifdef CONFIG_NUMA_EMU
231 if (numa_fake && !numa_emulation(start_pfn, end_pfn))
235 #ifdef CONFIG_ACPI_NUMA
236 if (!numa_off && !acpi_scan_nodes(start_pfn << PAGE_SHIFT,
237 end_pfn << PAGE_SHIFT))
241 #ifdef CONFIG_K8_NUMA
242 if (!numa_off && !k8_scan_nodes(start_pfn<<PAGE_SHIFT, end_pfn<<PAGE_SHIFT))
245 printk(KERN_INFO "%s\n",
246 numa_off ? "NUMA turned off" : "No NUMA configuration found");
248 printk(KERN_INFO "Faking a node at %016lx-%016lx\n",
249 start_pfn << PAGE_SHIFT,
250 end_pfn << PAGE_SHIFT);
251 /* setup dummy node covering all memory */
254 nodes_clear(node_online_map);
256 for (i = 0; i < NR_CPUS; i++)
258 node_to_cpumask[0] = cpumask_of_cpu(0);
259 setup_node_bootmem(0, start_pfn << PAGE_SHIFT, end_pfn << PAGE_SHIFT);
262 __cpuinit void numa_add_cpu(int cpu)
264 /* BP is initialized elsewhere */
266 set_bit(cpu, &node_to_cpumask[cpu_to_node(cpu)]);
269 unsigned long __init numa_free_all_bootmem(void)
272 unsigned long pages = 0;
273 for_each_online_node(i) {
274 pages += free_all_bootmem_node(NODE_DATA(i));
279 void __init paging_init(void)
282 for_each_online_node(i) {
288 __init int numa_setup(char *opt)
290 if (!strncmp(opt,"off",3))
292 #ifdef CONFIG_NUMA_EMU
293 if(!strncmp(opt, "fake=", 5)) {
294 numa_fake = simple_strtoul(opt+5,NULL,0); ;
295 if (numa_fake >= MAX_NUMNODES)
296 numa_fake = MAX_NUMNODES;
299 #ifdef CONFIG_ACPI_NUMA
300 if (!strncmp(opt,"noacpi",6))
306 EXPORT_SYMBOL(cpu_to_node);
307 EXPORT_SYMBOL(node_to_cpumask);
308 EXPORT_SYMBOL(memnode_shift);
309 EXPORT_SYMBOL(memnodemap);
310 EXPORT_SYMBOL(node_data);