arch/x86_64/mm/numa.c

   1 /*
   2  * Generic VM initialization for x86-64 NUMA setups.
   3  * Copyright 2002,2003 Andi Kleen, SuSE Labs.
   4  */
   5 #include <linux/kernel.h>
   6 #include <linux/mm.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>
  14
  15 #include <asm/e820.h>
  16 #include <asm/proto.h>
  17 #include <asm/dma.h>
  18 #include <asm/numa.h>
  19 #include <asm/acpi.h>
  20
  21 #ifndef Dprintk
  22 #define Dprintk(x...)
  23 #endif
  24
  25 struct pglist_data *node_data[MAX_NUMNODES] __read_mostly;
  26 bootmem_data_t plat_node_bdata[MAX_NUMNODES];
  27
  28 int memnode_shift;
  29 u8  memnodemap[NODEMAPSIZE];
  30
  31 unsigned char cpu_to_node[NR_CPUS] __read_mostly = {
  32         [0 ... NR_CPUS-1] = NUMA_NO_NODE
  33 };
  34 unsigned char apicid_to_node[MAX_LOCAL_APIC] __cpuinitdata = {
  35         [0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE
  36 };
  37 cpumask_t node_to_cpumask[MAX_NUMNODES] __read_mostly;
  38
  39 int numa_off __initdata;
  40
  41 int __init compute_hash_shift(struct node *nodes, int numnodes)
  42 {
  43         int i;
  44         int shift = 20;
  45         unsigned long addr,maxend=0;
  46
  47         for (i = 0; i < numnodes; i++)
  48                 if ((nodes[i].start != nodes[i].end) && (nodes[i].end > maxend))
  49                                 maxend = nodes[i].end;
  50
  51         while ((1UL << shift) <  (maxend / NODEMAPSIZE))
  52                 shift++;
  53
  54         printk (KERN_DEBUG"Using %d for the hash shift. Max adder is %lx \n",
  55                         shift,maxend);
  56         memset(memnodemap,0xff,sizeof(*memnodemap) * NODEMAPSIZE);
  57         for (i = 0; i < numnodes; i++) {
  58                 if (nodes[i].start == nodes[i].end)
  59                         continue;
  60                 for (addr = nodes[i].start;
  61                      addr < nodes[i].end;
  62                      addr += (1UL << shift)) {
  63                         if (memnodemap[addr >> shift] != 0xff) {
  64                                 printk(KERN_INFO
  65         "Your memory is not aligned you need to rebuild your kernel "
  66         "with a bigger NODEMAPSIZE shift=%d adder=%lu\n",
  67                                         shift,addr);
  68                                 return -1;
  69                         }
  70                         memnodemap[addr >> shift] = i;
  71                 }
  72         }
  73         return shift;
  74 }
  75
  76 #ifdef CONFIG_SPARSEMEM
  77 int early_pfn_to_nid(unsigned long pfn)
  78 {
  79         return phys_to_nid(pfn << PAGE_SHIFT);
  80 }
  81 #endif
  82
  83 /* Initialize bootmem allocator for a node */
  84 void __init setup_node_bootmem(int nodeid, unsigned long start, unsigned long end)
  85 {
  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);
  89
  90         start = round_up(start, ZONE_ALIGN);
  91
  92         printk("Bootmem setup node %d %016lx-%016lx\n", nodeid, start, end);
  93
  94         start_pfn = start >> PAGE_SHIFT;
  95         end_pfn = end >> PAGE_SHIFT;
  96
  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);
 101
 102         Dprintk("nodedata_phys %lx\n", nodedata_phys);
 103
 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;
 109
 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);
 117
 118         bootmap_size = init_bootmem_node(NODE_DATA(nodeid),
 119                                          bootmap_start >> PAGE_SHIFT,
 120                                          start_pfn, end_pfn);
 121
 122         e820_bootmem_free(NODE_DATA(nodeid), start, end);
 123
 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);
 127 }
 128
 129 /* Initialize final allocator for a zone */
 130 void __init setup_node_zones(int nodeid)
 131 {
 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;
 136
 137         memset(zones, 0, sizeof(unsigned long) * MAX_NR_ZONES);
 138         memset(holes, 0, sizeof(unsigned long) * MAX_NR_ZONES);
 139
 140         start_pfn = node_start_pfn(nodeid);
 141         end_pfn = node_end_pfn(nodeid);
 142
 143         Dprintk(KERN_INFO "setting up node %d %lx-%lx\n", nodeid, start_pfn, end_pfn);
 144
 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);
 152
 153         } else {
 154                 zones[ZONE_NORMAL] = end_pfn - start_pfn;
 155                 holes[ZONE_NORMAL] = e820_hole_size(start_pfn, end_pfn);
 156         }
 157
 158         free_area_init_node(nodeid, NODE_DATA(nodeid), zones,
 159                             start_pfn, holes);
 160 }
 161
 162 void __init numa_init_array(void)
 163 {
 164         int rr, i;
 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 = 0;
 171         for (i = 0; i < NR_CPUS; i++) {
 172                 if (cpu_to_node[i] != NUMA_NO_NODE)
 173                         continue;
 174                 rr = next_node(rr, node_online_map);
 175                 if (rr == MAX_NUMNODES)
 176                         rr = first_node(node_online_map);
 177                 cpu_to_node[i] = rr;
 178                 rr++;
 179         }
 180
 181 }
 182
 183 #ifdef CONFIG_NUMA_EMU
 184 int numa_fake __initdata = 0;
 185
 186 /* Numa emulation */
 187 static int numa_emulation(unsigned long start_pfn, unsigned long end_pfn)
 188 {
 189         int i;
 190         struct node nodes[MAX_NUMNODES];
 191         unsigned long sz = ((end_pfn - start_pfn)<<PAGE_SHIFT) / numa_fake;
 192
 193         /* Kludge needed for the hash function */
 194         if (hweight64(sz) > 1) {
 195                 unsigned long x = 1;
 196                 while ((x << 1) < sz)
 197                         x <<= 1;
 198                 if (x < sz/2)
 199                         printk("Numa emulation unbalanced. Complain to maintainer\n");
 200                 sz = x;
 201         }
 202
 203         memset(&nodes,0,sizeof(nodes));
 204         for (i = 0; i < numa_fake; i++) {
 205                 nodes[i].start = (start_pfn<<PAGE_SHIFT) + i*sz;
 206                 if (i == numa_fake-1)
 207                         sz = (end_pfn<<PAGE_SHIFT) - nodes[i].start;
 208                 nodes[i].end = nodes[i].start + sz;
 209                 if (i != numa_fake-1)
 210                         nodes[i].end--;
 211                 printk(KERN_INFO "Faking node %d at %016Lx-%016Lx (%LuMB)\n",
 212                        i,
 213                        nodes[i].start, nodes[i].end,
 214                        (nodes[i].end - nodes[i].start) >> 20);
 215                 node_set_online(i);
 216         }
 217         memnode_shift = compute_hash_shift(nodes, numa_fake);
 218         if (memnode_shift < 0) {
 219                 memnode_shift = 0;
 220                 printk(KERN_ERR "No NUMA hash function found. Emulation disabled.\n");
 221                 return -1;
 222         }
 223         for_each_online_node(i)
 224                 setup_node_bootmem(i, nodes[i].start, nodes[i].end);
 225         numa_init_array();
 226         return 0;
 227 }
 228 #endif
 229
 230 void __init numa_initmem_init(unsigned long start_pfn, unsigned long end_pfn)
 231 {
 232         int i;
 233
 234 #ifdef CONFIG_NUMA_EMU
 235         if (numa_fake && !numa_emulation(start_pfn, end_pfn))
 236                 return;
 237 #endif
 238
 239 #ifdef CONFIG_ACPI_NUMA
 240         if (!numa_off && !acpi_scan_nodes(start_pfn << PAGE_SHIFT,
 241                                           end_pfn << PAGE_SHIFT))
 242                 return;
 243 #endif
 244
 245 #ifdef CONFIG_K8_NUMA
 246         if (!numa_off && !k8_scan_nodes(start_pfn<<PAGE_SHIFT, end_pfn<<PAGE_SHIFT))
 247                 return;
 248 #endif
 249         printk(KERN_INFO "%s\n",
 250                numa_off ? "NUMA turned off" : "No NUMA configuration found");
 251
 252         printk(KERN_INFO "Faking a node at %016lx-%016lx\n",
 253                start_pfn << PAGE_SHIFT,
 254                end_pfn << PAGE_SHIFT);
 255                 /* setup dummy node covering all memory */
 256         memnode_shift = 63;
 257         memnodemap[0] = 0;
 258         nodes_clear(node_online_map);
 259         node_set_online(0);
 260         for (i = 0; i < NR_CPUS; i++)
 261                 cpu_to_node[i] = 0;
 262         node_to_cpumask[0] = cpumask_of_cpu(0);
 263         setup_node_bootmem(0, start_pfn << PAGE_SHIFT, end_pfn << PAGE_SHIFT);
 264 }
 265
 266 __cpuinit void numa_add_cpu(int cpu)
 267 {
 268         set_bit(cpu, &node_to_cpumask[cpu_to_node(cpu)]);
 269 }
 270
 271 unsigned long __init numa_free_all_bootmem(void)
 272 {
 273         int i;
 274         unsigned long pages = 0;
 275         for_each_online_node(i) {
 276                 pages += free_all_bootmem_node(NODE_DATA(i));
 277         }
 278         return pages;
 279 }
 280
 281 void __init paging_init(void)
 282 {
 283         int i;
 284         for_each_online_node(i) {
 285                 setup_node_zones(i);
 286         }
 287 }
 288
 289 /* [numa=off] */
 290 __init int numa_setup(char *opt)
 291 {
 292         if (!strncmp(opt,"off",3))
 293                 numa_off = 1;
 294 #ifdef CONFIG_NUMA_EMU
 295         if(!strncmp(opt, "fake=", 5)) {
 296                 numa_fake = simple_strtoul(opt+5,NULL,0); ;
 297                 if (numa_fake >= MAX_NUMNODES)
 298                         numa_fake = MAX_NUMNODES;
 299         }
 300 #endif
 301 #ifdef CONFIG_ACPI_NUMA
 302         if (!strncmp(opt,"noacpi",6))
 303                 acpi_numa = -1;
 304 #endif
 305         return 1;
 306 }
 307
 308 EXPORT_SYMBOL(cpu_to_node);
 309 EXPORT_SYMBOL(node_to_cpumask);
 310 EXPORT_SYMBOL(memnode_shift);
 311 EXPORT_SYMBOL(memnodemap);
 312 EXPORT_SYMBOL(node_data);