arch/x86/mm/srat.c

   1 /*
   2  * ACPI 3.0 based NUMA setup
   3  * Copyright 2004 Andi Kleen, SuSE Labs.
   4  *
   5  * Reads the ACPI SRAT table to figure out what memory belongs to which CPUs.
   6  *
   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.
  10  */
  11
  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 <linux/bootmem.h>
  19 #include <linux/memblock.h>
  20 #include <linux/mm.h>
  21 #include <asm/proto.h>
  22 #include <asm/numa.h>
  23 #include <asm/e820.h>
  24 #include <asm/apic.h>
  25 #include <asm/uv/uv.h>
  26
  27 int acpi_numa __initdata;
  28
  29 static __init int setup_node(int pxm)
  30 {
  31         return acpi_map_pxm_to_node(pxm);
  32 }
  33
  34 static __init void bad_srat(void)
  35 {
  36         printk(KERN_ERR "SRAT: SRAT not used.\n");
  37         acpi_numa = -1;
  38 }
  39
  40 static __init inline int srat_disabled(void)
  41 {
  42         return acpi_numa < 0;
  43 }
  44
  45 /* Callback for SLIT parsing */
  46 void __init acpi_numa_slit_init(struct acpi_table_slit *slit)
  47 {
  48         int i, j;
  49
  50         for (i = 0; i < slit->locality_count; i++)
  51                 for (j = 0; j < slit->locality_count; j++)
  52                         numa_set_distance(pxm_to_node(i), pxm_to_node(j),
  53                                 slit->entry[slit->locality_count * i + j]);
  54 }
  55
  56 /* Callback for Proximity Domain -> x2APIC mapping */
  57 void __init
  58 acpi_numa_x2apic_affinity_init(struct acpi_srat_x2apic_cpu_affinity *pa)
  59 {
  60         int pxm, node;
  61         int apic_id;
  62
  63         if (srat_disabled())
  64                 return;
  65         if (pa->header.length < sizeof(struct acpi_srat_x2apic_cpu_affinity)) {
  66                 bad_srat();
  67                 return;
  68         }
  69         if ((pa->flags & ACPI_SRAT_CPU_ENABLED) == 0)
  70                 return;
  71         pxm = pa->proximity_domain;
  72         apic_id = pa->apic_id;
  73         if (!apic->apic_id_valid(apic_id)) {
  74                 printk(KERN_INFO "SRAT: PXM %u -> X2APIC 0x%04x ignored\n",
  75                          pxm, apic_id);
  76                 return;
  77         }
  78         node = setup_node(pxm);
  79         if (node < 0) {
  80                 printk(KERN_ERR "SRAT: Too many proximity domains %x\n", pxm);
  81                 bad_srat();
  82                 return;
  83         }
  84
  85         if (apic_id >= MAX_LOCAL_APIC) {
  86                 printk(KERN_INFO "SRAT: PXM %u -> APIC 0x%04x -> Node %u skipped apicid that is too big\n", pxm, apic_id, node);
  87                 return;
  88         }
  89         set_apicid_to_node(apic_id, node);
  90         node_set(node, numa_nodes_parsed);
  91         acpi_numa = 1;
  92         printk(KERN_INFO "SRAT: PXM %u -> APIC 0x%04x -> Node %u\n",
  93                pxm, apic_id, node);
  94 }
  95
  96 /* Callback for Proximity Domain -> LAPIC mapping */
  97 void __init
  98 acpi_numa_processor_affinity_init(struct acpi_srat_cpu_affinity *pa)
  99 {
 100         int pxm, node;
 101         int apic_id;
 102
 103         if (srat_disabled())
 104                 return;
 105         if (pa->header.length != sizeof(struct acpi_srat_cpu_affinity)) {
 106                 bad_srat();
 107                 return;
 108         }
 109         if ((pa->flags & ACPI_SRAT_CPU_ENABLED) == 0)
 110                 return;
 111         pxm = pa->proximity_domain_lo;
 112         if (acpi_srat_revision >= 2)
 113                 pxm |= *((unsigned int*)pa->proximity_domain_hi) << 8;
 114         node = setup_node(pxm);
 115         if (node < 0) {
 116                 printk(KERN_ERR "SRAT: Too many proximity domains %x\n", pxm);
 117                 bad_srat();
 118                 return;
 119         }
 120
 121         if (get_uv_system_type() >= UV_X2APIC)
 122                 apic_id = (pa->apic_id << 8) | pa->local_sapic_eid;
 123         else
 124                 apic_id = pa->apic_id;
 125
 126         if (apic_id >= MAX_LOCAL_APIC) {
 127                 printk(KERN_INFO "SRAT: PXM %u -> APIC 0x%02x -> Node %u skipped apicid that is too big\n", pxm, apic_id, node);
 128                 return;
 129         }
 130
 131         set_apicid_to_node(apic_id, node);
 132         node_set(node, numa_nodes_parsed);
 133         acpi_numa = 1;
 134         printk(KERN_INFO "SRAT: PXM %u -> APIC 0x%02x -> Node %u\n",
 135                pxm, apic_id, node);
 136 }
 137
 138 #ifdef CONFIG_MEMORY_HOTPLUG
 139 static inline int save_add_info(void) {return 1;}
 140 #else
 141 static inline int save_add_info(void) {return 0;}
 142 #endif
 143
 144 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
 145 static void __init
 146 handle_movablemem(int node, u64 start, u64 end, u32 hotpluggable)
 147 {
 148         int overlap, i;
 149         unsigned long start_pfn, end_pfn;
 150
 151         start_pfn = PFN_DOWN(start);
 152         end_pfn = PFN_UP(end);
 153
 154         /*
 155          * For movablemem_map=acpi:
 156          *
 157          * SRAT:                |_____| |_____| |_________| |_________| ......
 158          * node id:                0       1         1           2
 159          * hotpluggable:           n       y         y           n
 160          * movablemem_map:              |_____| |_________|
 161          *
 162          * Using movablemem_map, we can prevent memblock from allocating memory
 163          * on ZONE_MOVABLE at boot time.
 164          *
 165          * Before parsing SRAT, memblock has already reserve some memory ranges
 166          * for other purposes, such as for kernel image. We cannot prevent
 167          * kernel from using these memory, so we need to exclude these memory
 168          * even if it is hotpluggable.
 169          * Furthermore, to ensure the kernel has enough memory to boot, we make
 170          * all the memory on the node which the kernel resides in
 171          * un-hotpluggable.
 172          */
 173         if (hotpluggable && movablemem_map.acpi) {
 174                 /* Exclude ranges reserved by memblock. */
 175                 struct memblock_type *rgn = &memblock.reserved;
 176
 177                 for (i = 0; i < rgn->cnt; i++) {
 178                         if (end <= rgn->regions[i].base ||
 179                             start >= rgn->regions[i].base +
 180                             rgn->regions[i].size)
 181                                 continue;
 182
 183                         /*
 184                          * If the memory range overlaps the memory reserved by
 185                          * memblock, then the kernel resides in this node.
 186                          */
 187                         node_set(node, movablemem_map.numa_nodes_kernel);
 188
 189                         goto out;
 190                 }
 191
 192                 /*
 193                  * If the kernel resides in this node, then the whole node
 194                  * should not be hotpluggable.
 195                  */
 196                 if (node_isset(node, movablemem_map.numa_nodes_kernel))
 197                         goto out;
 198
 199                 insert_movablemem_map(start_pfn, end_pfn);
 200
 201                 /*
 202                  * numa_nodes_hotplug nodemask represents which nodes are put
 203                  * into movablemem_map.map[].
 204                  */
 205                 node_set(node, movablemem_map.numa_nodes_hotplug);
 206                 goto out;
 207         }
 208
 209         /*
 210          * For movablemem_map=nn[KMG]@ss[KMG]:
 211          *
 212          * SRAT:                |_____| |_____| |_________| |_________| ......
 213          * node id:                0       1         1           2
 214          * user specified:                |__|                 |___|
 215          * movablemem_map:                |___| |_________|    |______| ......
 216          *
 217          * Using movablemem_map, we can prevent memblock from allocating memory
 218          * on ZONE_MOVABLE at boot time.
 219          *
 220          * NOTE: In this case, SRAT info will be ingored.
 221          */
 222         overlap = movablemem_map_overlap(start_pfn, end_pfn);
 223         if (overlap >= 0) {
 224                 /*
 225                  * If part of this range is in movablemem_map, we need to
 226                  * add the range after it to extend the range to the end
 227                  * of the node, because from the min address specified to
 228                  * the end of the node will be ZONE_MOVABLE.
 229                  */
 230                 start_pfn = max(start_pfn,
 231                             movablemem_map.map[overlap].start_pfn);
 232                 insert_movablemem_map(start_pfn, end_pfn);
 233
 234                 /*
 235                  * Set the nodemask, so that if the address range on one node
 236                  * is not continuse, we can add the subsequent ranges on the
 237                  * same node into movablemem_map.
 238                  */
 239                 node_set(node, movablemem_map.numa_nodes_hotplug);
 240         } else {
 241                 if (node_isset(node, movablemem_map.numa_nodes_hotplug))
 242                         /*
 243                          * Insert the range if we already have movable ranges
 244                          * on the same node.
 245                          */
 246                         insert_movablemem_map(start_pfn, end_pfn);
 247         }
 248 out:
 249         return;
 250 }
 251 #else           /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
 252 static inline void
 253 handle_movablemem(int node, u64 start, u64 end, u32 hotpluggable)
 254 {
 255 }
 256 #endif          /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
 257
 258 /* Callback for parsing of the Proximity Domain <-> Memory Area mappings */
 259 int __init
 260 acpi_numa_memory_affinity_init(struct acpi_srat_mem_affinity *ma)
 261 {
 262         u64 start, end;
 263         u32 hotpluggable;
 264         int node, pxm;
 265
 266         if (srat_disabled())
 267                 goto out_err;
 268         if (ma->header.length != sizeof(struct acpi_srat_mem_affinity))
 269                 goto out_err_bad_srat;
 270         if ((ma->flags & ACPI_SRAT_MEM_ENABLED) == 0)
 271                 goto out_err;
 272         hotpluggable = ma->flags & ACPI_SRAT_MEM_HOT_PLUGGABLE;
 273         if (hotpluggable && !save_add_info())
 274                 goto out_err;
 275
 276         start = ma->base_address;
 277         end = start + ma->length;
 278         pxm = ma->proximity_domain;
 279         if (acpi_srat_revision <= 1)
 280                 pxm &= 0xff;
 281
 282         node = setup_node(pxm);
 283         if (node < 0) {
 284                 printk(KERN_ERR "SRAT: Too many proximity domains.\n");
 285                 goto out_err_bad_srat;
 286         }
 287
 288         if (numa_add_memblk(node, start, end) < 0)
 289                 goto out_err_bad_srat;
 290
 291         node_set(node, numa_nodes_parsed);
 292
 293         printk(KERN_INFO "SRAT: Node %u PXM %u [mem %#010Lx-%#010Lx] %s\n",
 294                node, pxm,
 295                (unsigned long long) start, (unsigned long long) end - 1,
 296                hotpluggable ? "Hot Pluggable": "");
 297
 298         handle_movablemem(node, start, end, hotpluggable);
 299
 300         return 0;
 301 out_err_bad_srat:
 302         bad_srat();
 303 out_err:
 304         return -1;
 305 }
 306
 307 void __init acpi_numa_arch_fixup(void) {}
 308
 309 int __init x86_acpi_numa_init(void)
 310 {
 311         int ret;
 312
 313         ret = acpi_numa_init();
 314         if (ret < 0)
 315                 return ret;
 316         return srat_disabled() ? -EINVAL : 0;
 317 }