drivers/acpi/numa.c

   1 /*
   2  *  acpi_numa.c - ACPI NUMA support
   3  *
   4  *  Copyright (C) 2002 Takayoshi Kochi <t-kochi@bq.jp.nec.com>
   5  *
   6  * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
   7  *
   8  *  This program is free software; you can redistribute it and/or modify
   9  *  it under the terms of the GNU General Public License as published by
  10  *  the Free Software Foundation; either version 2 of the License, or
  11  *  (at your option) any later version.
  12  *
  13  *  This program is distributed in the hope that it will be useful,
  14  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  16  *  GNU General Public License for more details.
  17  *
  18  * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
  19  *
  20  */
  21 #include <linux/module.h>
  22 #include <linux/init.h>
  23 #include <linux/kernel.h>
  24 #include <linux/types.h>
  25 #include <linux/errno.h>
  26 #include <linux/acpi.h>
  27 #include <linux/numa.h>
  28 #include <linux/nodemask.h>
  29 #include <linux/topology.h>
  30
  31 #define PREFIX "ACPI: "
  32
  33 #define ACPI_NUMA       0x80000000
  34 #define _COMPONENT      ACPI_NUMA
  35 ACPI_MODULE_NAME("numa");
  36
  37 static nodemask_t nodes_found_map = NODE_MASK_NONE;
  38
  39 /* maps to convert between proximity domain and logical node ID */
  40 static int pxm_to_node_map[MAX_PXM_DOMAINS]
  41                         = { [0 ... MAX_PXM_DOMAINS - 1] = NUMA_NO_NODE };
  42 static int node_to_pxm_map[MAX_NUMNODES]
  43                         = { [0 ... MAX_NUMNODES - 1] = PXM_INVAL };
  44
  45 unsigned char acpi_srat_revision __initdata;
  46
  47 int pxm_to_node(int pxm)
  48 {
  49         if (pxm < 0)
  50                 return NUMA_NO_NODE;
  51         return pxm_to_node_map[pxm];
  52 }
  53
  54 int node_to_pxm(int node)
  55 {
  56         if (node < 0)
  57                 return PXM_INVAL;
  58         return node_to_pxm_map[node];
  59 }
  60
  61 static void __acpi_map_pxm_to_node(int pxm, int node)
  62 {
  63         if (pxm_to_node_map[pxm] == NUMA_NO_NODE || node < pxm_to_node_map[pxm])
  64                 pxm_to_node_map[pxm] = node;
  65         if (node_to_pxm_map[node] == PXM_INVAL || pxm < node_to_pxm_map[node])
  66                 node_to_pxm_map[node] = pxm;
  67 }
  68
  69 int acpi_map_pxm_to_node(int pxm)
  70 {
  71         int node;
  72
  73         if (pxm < 0 || pxm >= MAX_PXM_DOMAINS)
  74                 return NUMA_NO_NODE;
  75
  76         node = pxm_to_node_map[pxm];
  77
  78         if (node == NUMA_NO_NODE) {
  79                 if (nodes_weight(nodes_found_map) >= MAX_NUMNODES)
  80                         return NUMA_NO_NODE;
  81                 node = first_unset_node(nodes_found_map);
  82                 __acpi_map_pxm_to_node(pxm, node);
  83                 node_set(node, nodes_found_map);
  84         }
  85
  86         return node;
  87 }
  88
  89 /**
  90  * acpi_map_pxm_to_online_node - Map proximity ID to online node
  91  * @pxm: ACPI proximity ID
  92  *
  93  * This is similar to acpi_map_pxm_to_node(), but always returns an online
  94  * node.  When the mapped node from a given proximity ID is offline, it
  95  * looks up the node distance table and returns the nearest online node.
  96  *
  97  * ACPI device drivers, which are called after the NUMA initialization has
  98  * completed in the kernel, can call this interface to obtain their device
  99  * NUMA topology from ACPI tables.  Such drivers do not have to deal with
 100  * offline nodes.  A node may be offline when a device proximity ID is
 101  * unique, SRAT memory entry does not exist, or NUMA is disabled, ex.
 102  * "numa=off" on x86.
 103  */
 104 int acpi_map_pxm_to_online_node(int pxm)
 105 {
 106         int node, min_node;
 107
 108         node = acpi_map_pxm_to_node(pxm);
 109
 110         if (node == NUMA_NO_NODE)
 111                 node = 0;
 112
 113         min_node = node;
 114         if (!node_online(node)) {
 115                 int min_dist = INT_MAX, dist, n;
 116
 117                 for_each_online_node(n) {
 118                         dist = node_distance(node, n);
 119                         if (dist < min_dist) {
 120                                 min_dist = dist;
 121                                 min_node = n;
 122                         }
 123                 }
 124         }
 125
 126         return min_node;
 127 }
 128 EXPORT_SYMBOL(acpi_map_pxm_to_online_node);
 129
 130 static void __init
 131 acpi_table_print_srat_entry(struct acpi_subtable_header *header)
 132 {
 133
 134         ACPI_FUNCTION_NAME("acpi_table_print_srat_entry");
 135
 136         if (!header)
 137                 return;
 138
 139         switch (header->type) {
 140
 141         case ACPI_SRAT_TYPE_CPU_AFFINITY:
 142 #ifdef ACPI_DEBUG_OUTPUT
 143                 {
 144                         struct acpi_srat_cpu_affinity *p =
 145                             (struct acpi_srat_cpu_affinity *)header;
 146                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 147                                           "SRAT Processor (id[0x%02x] eid[0x%02x]) in proximity domain %d %s\n",
 148                                           p->apic_id, p->local_sapic_eid,
 149                                           p->proximity_domain_lo,
 150                                           (p->flags & ACPI_SRAT_CPU_ENABLED)?
 151                                           "enabled" : "disabled"));
 152                 }
 153 #endif                          /* ACPI_DEBUG_OUTPUT */
 154                 break;
 155
 156         case ACPI_SRAT_TYPE_MEMORY_AFFINITY:
 157 #ifdef ACPI_DEBUG_OUTPUT
 158                 {
 159                         struct acpi_srat_mem_affinity *p =
 160                             (struct acpi_srat_mem_affinity *)header;
 161                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 162                                           "SRAT Memory (0x%lx length 0x%lx) in proximity domain %d %s%s%s\n",
 163                                           (unsigned long)p->base_address,
 164                                           (unsigned long)p->length,
 165                                           p->proximity_domain,
 166                                           (p->flags & ACPI_SRAT_MEM_ENABLED)?
 167                                           "enabled" : "disabled",
 168                                           (p->flags & ACPI_SRAT_MEM_HOT_PLUGGABLE)?
 169                                           " hot-pluggable" : "",
 170                                           (p->flags & ACPI_SRAT_MEM_NON_VOLATILE)?
 171                                           " non-volatile" : ""));
 172                 }
 173 #endif                          /* ACPI_DEBUG_OUTPUT */
 174                 break;
 175
 176         case ACPI_SRAT_TYPE_X2APIC_CPU_AFFINITY:
 177 #ifdef ACPI_DEBUG_OUTPUT
 178                 {
 179                         struct acpi_srat_x2apic_cpu_affinity *p =
 180                             (struct acpi_srat_x2apic_cpu_affinity *)header;
 181                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 182                                           "SRAT Processor (x2apicid[0x%08x]) in"
 183                                           " proximity domain %d %s\n",
 184                                           p->apic_id,
 185                                           p->proximity_domain,
 186                                           (p->flags & ACPI_SRAT_CPU_ENABLED) ?
 187                                           "enabled" : "disabled"));
 188                 }
 189 #endif                          /* ACPI_DEBUG_OUTPUT */
 190                 break;
 191         default:
 192                 printk(KERN_WARNING PREFIX
 193                        "Found unsupported SRAT entry (type = 0x%x)\n",
 194                        header->type);
 195                 break;
 196         }
 197 }
 198
 199 /*
 200  * A lot of BIOS fill in 10 (= no distance) everywhere. This messes
 201  * up the NUMA heuristics which wants the local node to have a smaller
 202  * distance than the others.
 203  * Do some quick checks here and only use the SLIT if it passes.
 204  */
 205 static int __init slit_valid(struct acpi_table_slit *slit)
 206 {
 207         int i, j;
 208         int d = slit->locality_count;
 209         for (i = 0; i < d; i++) {
 210                 for (j = 0; j < d; j++)  {
 211                         u8 val = slit->entry[d*i + j];
 212                         if (i == j) {
 213                                 if (val != LOCAL_DISTANCE)
 214                                         return 0;
 215                         } else if (val <= LOCAL_DISTANCE)
 216                                 return 0;
 217                 }
 218         }
 219         return 1;
 220 }
 221
 222 static int __init acpi_parse_slit(struct acpi_table_header *table)
 223 {
 224         struct acpi_table_slit *slit = (struct acpi_table_slit *)table;
 225
 226         if (!slit_valid(slit)) {
 227                 printk(KERN_INFO "ACPI: SLIT table looks invalid. Not used.\n");
 228                 return -EINVAL;
 229         }
 230         acpi_numa_slit_init(slit);
 231
 232         return 0;
 233 }
 234
 235 void __init __weak
 236 acpi_numa_x2apic_affinity_init(struct acpi_srat_x2apic_cpu_affinity *pa)
 237 {
 238         printk(KERN_WARNING PREFIX
 239                "Found unsupported x2apic [0x%08x] SRAT entry\n", pa->apic_id);
 240         return;
 241 }
 242
 243
 244 static int __init
 245 acpi_parse_x2apic_affinity(struct acpi_subtable_header *header,
 246                            const unsigned long end)
 247 {
 248         struct acpi_srat_x2apic_cpu_affinity *processor_affinity;
 249
 250         processor_affinity = (struct acpi_srat_x2apic_cpu_affinity *)header;
 251         if (!processor_affinity)
 252                 return -EINVAL;
 253
 254         acpi_table_print_srat_entry(header);
 255
 256         /* let architecture-dependent part to do it */
 257         acpi_numa_x2apic_affinity_init(processor_affinity);
 258
 259         return 0;
 260 }
 261
 262 static int __init
 263 acpi_parse_processor_affinity(struct acpi_subtable_header *header,
 264                               const unsigned long end)
 265 {
 266         struct acpi_srat_cpu_affinity *processor_affinity;
 267
 268         processor_affinity = (struct acpi_srat_cpu_affinity *)header;
 269         if (!processor_affinity)
 270                 return -EINVAL;
 271
 272         acpi_table_print_srat_entry(header);
 273
 274         /* let architecture-dependent part to do it */
 275         acpi_numa_processor_affinity_init(processor_affinity);
 276
 277         return 0;
 278 }
 279
 280 static int __initdata parsed_numa_memblks;
 281
 282 static int __init
 283 acpi_parse_memory_affinity(struct acpi_subtable_header * header,
 284                            const unsigned long end)
 285 {
 286         struct acpi_srat_mem_affinity *memory_affinity;
 287
 288         memory_affinity = (struct acpi_srat_mem_affinity *)header;
 289         if (!memory_affinity)
 290                 return -EINVAL;
 291
 292         acpi_table_print_srat_entry(header);
 293
 294         /* let architecture-dependent part to do it */
 295         if (!acpi_numa_memory_affinity_init(memory_affinity))
 296                 parsed_numa_memblks++;
 297         return 0;
 298 }
 299
 300 static int __init acpi_parse_srat(struct acpi_table_header *table)
 301 {
 302         struct acpi_table_srat *srat = (struct acpi_table_srat *)table;
 303
 304         acpi_srat_revision = srat->header.revision;
 305
 306         /* Real work done in acpi_table_parse_srat below. */
 307
 308         return 0;
 309 }
 310
 311 static int __init
 312 acpi_table_parse_srat(enum acpi_srat_type id,
 313                       acpi_tbl_entry_handler handler, unsigned int max_entries)
 314 {
 315         return acpi_table_parse_entries(ACPI_SIG_SRAT,
 316                                             sizeof(struct acpi_table_srat), id,
 317                                             handler, max_entries);
 318 }
 319
 320 int __init acpi_numa_init(void)
 321 {
 322         int cnt = 0;
 323
 324         /*
 325          * Should not limit number with cpu num that is from NR_CPUS or nr_cpus=
 326          * SRAT cpu entries could have different order with that in MADT.
 327          * So go over all cpu entries in SRAT to get apicid to node mapping.
 328          */
 329
 330         /* SRAT: Static Resource Affinity Table */
 331         if (!acpi_table_parse(ACPI_SIG_SRAT, acpi_parse_srat)) {
 332                 struct acpi_subtable_proc srat_proc[2];
 333
 334                 memset(srat_proc, 0, sizeof(srat_proc));
 335                 srat_proc[0].id = ACPI_SRAT_TYPE_CPU_AFFINITY;
 336                 srat_proc[0].handler = acpi_parse_processor_affinity;
 337                 srat_proc[1].id = ACPI_SRAT_TYPE_X2APIC_CPU_AFFINITY;
 338                 srat_proc[1].handler = acpi_parse_x2apic_affinity;
 339
 340                 acpi_table_parse_entries_array(ACPI_SIG_SRAT,
 341                                         sizeof(struct acpi_table_srat),
 342                                         srat_proc, ARRAY_SIZE(srat_proc), 0);
 343
 344                 cnt = acpi_table_parse_srat(ACPI_SRAT_TYPE_MEMORY_AFFINITY,
 345                                             acpi_parse_memory_affinity,
 346                                             NR_NODE_MEMBLKS);
 347         }
 348
 349         /* SLIT: System Locality Information Table */
 350         acpi_table_parse(ACPI_SIG_SLIT, acpi_parse_slit);
 351
 352         acpi_numa_arch_fixup();
 353
 354         if (cnt < 0)
 355                 return cnt;
 356         else if (!parsed_numa_memblks)
 357                 return -ENOENT;
 358         return 0;
 359 }
 360
 361 static int acpi_get_pxm(acpi_handle h)
 362 {
 363         unsigned long long pxm;
 364         acpi_status status;
 365         acpi_handle handle;
 366         acpi_handle phandle = h;
 367
 368         do {
 369                 handle = phandle;
 370                 status = acpi_evaluate_integer(handle, "_PXM", NULL, &pxm);
 371                 if (ACPI_SUCCESS(status))
 372                         return pxm;
 373                 status = acpi_get_parent(handle, &phandle);
 374         } while (ACPI_SUCCESS(status));
 375         return -1;
 376 }
 377
 378 int acpi_get_node(acpi_handle handle)
 379 {
 380         int pxm;
 381
 382         pxm = acpi_get_pxm(handle);
 383
 384         return acpi_map_pxm_to_node(pxm);
 385 }
 386 EXPORT_SYMBOL(acpi_get_node);