drivers/acpi/numa/srat.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  *  acpi_numa.c - ACPI NUMA support
   4  *
   5  *  Copyright (C) 2002 Takayoshi Kochi <[email protected]>
   6  */
   7
   8 #define pr_fmt(fmt) "ACPI: " fmt
   9
  10 #include <linux/module.h>
  11 #include <linux/init.h>
  12 #include <linux/kernel.h>
  13 #include <linux/types.h>
  14 #include <linux/errno.h>
  15 #include <linux/acpi.h>
  16 #include <linux/memblock.h>
  17 #include <linux/numa.h>
  18 #include <linux/nodemask.h>
  19 #include <linux/topology.h>
  20
  21 static nodemask_t nodes_found_map = NODE_MASK_NONE;
  22
  23 /* maps to convert between proximity domain and logical node ID */
  24 static int pxm_to_node_map[MAX_PXM_DOMAINS]
  25                         = { [0 ... MAX_PXM_DOMAINS - 1] = NUMA_NO_NODE };
  26 static int node_to_pxm_map[MAX_NUMNODES]
  27                         = { [0 ... MAX_NUMNODES - 1] = PXM_INVAL };
  28
  29 unsigned char acpi_srat_revision __initdata;
  30 int acpi_numa __initdata;
  31
  32 int pxm_to_node(int pxm)
  33 {
  34         if (pxm < 0)
  35                 return NUMA_NO_NODE;
  36         return pxm_to_node_map[pxm];
  37 }
  38 EXPORT_SYMBOL(pxm_to_node);
  39
  40 int node_to_pxm(int node)
  41 {
  42         if (node < 0)
  43                 return PXM_INVAL;
  44         return node_to_pxm_map[node];
  45 }
  46
  47 static void __acpi_map_pxm_to_node(int pxm, int node)
  48 {
  49         if (pxm_to_node_map[pxm] == NUMA_NO_NODE || node < pxm_to_node_map[pxm])
  50                 pxm_to_node_map[pxm] = node;
  51         if (node_to_pxm_map[node] == PXM_INVAL || pxm < node_to_pxm_map[node])
  52                 node_to_pxm_map[node] = pxm;
  53 }
  54
  55 int acpi_map_pxm_to_node(int pxm)
  56 {
  57         int node;
  58
  59         if (pxm < 0 || pxm >= MAX_PXM_DOMAINS || numa_off)
  60                 return NUMA_NO_NODE;
  61
  62         node = pxm_to_node_map[pxm];
  63
  64         if (node == NUMA_NO_NODE) {
  65                 if (nodes_weight(nodes_found_map) >= MAX_NUMNODES)
  66                         return NUMA_NO_NODE;
  67                 node = first_unset_node(nodes_found_map);
  68                 __acpi_map_pxm_to_node(pxm, node);
  69                 node_set(node, nodes_found_map);
  70         }
  71
  72         return node;
  73 }
  74 EXPORT_SYMBOL(acpi_map_pxm_to_node);
  75
  76 static void __init
  77 acpi_table_print_srat_entry(struct acpi_subtable_header *header)
  78 {
  79         switch (header->type) {
  80         case ACPI_SRAT_TYPE_CPU_AFFINITY:
  81                 {
  82                         struct acpi_srat_cpu_affinity *p =
  83                             (struct acpi_srat_cpu_affinity *)header;
  84                         pr_debug("SRAT Processor (id[0x%02x] eid[0x%02x]) in proximity domain %d %s\n",
  85                                  p->apic_id, p->local_sapic_eid,
  86                                  p->proximity_domain_lo,
  87                                  (p->flags & ACPI_SRAT_CPU_ENABLED) ?
  88                                  "enabled" : "disabled");
  89                 }
  90                 break;
  91
  92         case ACPI_SRAT_TYPE_MEMORY_AFFINITY:
  93                 {
  94                         struct acpi_srat_mem_affinity *p =
  95                             (struct acpi_srat_mem_affinity *)header;
  96                         pr_debug("SRAT Memory (0x%llx length 0x%llx) in proximity domain %d %s%s%s\n",
  97                                  (unsigned long long)p->base_address,
  98                                  (unsigned long long)p->length,
  99                                  p->proximity_domain,
 100                                  (p->flags & ACPI_SRAT_MEM_ENABLED) ?
 101                                  "enabled" : "disabled",
 102                                  (p->flags & ACPI_SRAT_MEM_HOT_PLUGGABLE) ?
 103                                  " hot-pluggable" : "",
 104                                  (p->flags & ACPI_SRAT_MEM_NON_VOLATILE) ?
 105                                  " non-volatile" : "");
 106                 }
 107                 break;
 108
 109         case ACPI_SRAT_TYPE_X2APIC_CPU_AFFINITY:
 110                 {
 111                         struct acpi_srat_x2apic_cpu_affinity *p =
 112                             (struct acpi_srat_x2apic_cpu_affinity *)header;
 113                         pr_debug("SRAT Processor (x2apicid[0x%08x]) in proximity domain %d %s\n",
 114                                  p->apic_id,
 115                                  p->proximity_domain,
 116                                  (p->flags & ACPI_SRAT_CPU_ENABLED) ?
 117                                  "enabled" : "disabled");
 118                 }
 119                 break;
 120
 121         case ACPI_SRAT_TYPE_GICC_AFFINITY:
 122                 {
 123                         struct acpi_srat_gicc_affinity *p =
 124                             (struct acpi_srat_gicc_affinity *)header;
 125                         pr_debug("SRAT Processor (acpi id[0x%04x]) in proximity domain %d %s\n",
 126                                  p->acpi_processor_uid,
 127                                  p->proximity_domain,
 128                                  (p->flags & ACPI_SRAT_GICC_ENABLED) ?
 129                                  "enabled" : "disabled");
 130                 }
 131                 break;
 132
 133         default:
 134                 pr_warn("Found unsupported SRAT entry (type = 0x%x)\n",
 135                         header->type);
 136                 break;
 137         }
 138 }
 139
 140 /*
 141  * A lot of BIOS fill in 10 (= no distance) everywhere. This messes
 142  * up the NUMA heuristics which wants the local node to have a smaller
 143  * distance than the others.
 144  * Do some quick checks here and only use the SLIT if it passes.
 145  */
 146 static int __init slit_valid(struct acpi_table_slit *slit)
 147 {
 148         int i, j;
 149         int d = slit->locality_count;
 150         for (i = 0; i < d; i++) {
 151                 for (j = 0; j < d; j++)  {
 152                         u8 val = slit->entry[d*i + j];
 153                         if (i == j) {
 154                                 if (val != LOCAL_DISTANCE)
 155                                         return 0;
 156                         } else if (val <= LOCAL_DISTANCE)
 157                                 return 0;
 158                 }
 159         }
 160         return 1;
 161 }
 162
 163 void __init bad_srat(void)
 164 {
 165         pr_err("SRAT: SRAT not used.\n");
 166         acpi_numa = -1;
 167 }
 168
 169 int __init srat_disabled(void)
 170 {
 171         return acpi_numa < 0;
 172 }
 173
 174 #if defined(CONFIG_X86) || defined(CONFIG_ARM64)
 175 /*
 176  * Callback for SLIT parsing.  pxm_to_node() returns NUMA_NO_NODE for
 177  * I/O localities since SRAT does not list them.  I/O localities are
 178  * not supported at this point.
 179  */
 180 void __init acpi_numa_slit_init(struct acpi_table_slit *slit)
 181 {
 182         int i, j;
 183
 184         for (i = 0; i < slit->locality_count; i++) {
 185                 const int from_node = pxm_to_node(i);
 186
 187                 if (from_node == NUMA_NO_NODE)
 188                         continue;
 189
 190                 for (j = 0; j < slit->locality_count; j++) {
 191                         const int to_node = pxm_to_node(j);
 192
 193                         if (to_node == NUMA_NO_NODE)
 194                                 continue;
 195
 196                         numa_set_distance(from_node, to_node,
 197                                 slit->entry[slit->locality_count * i + j]);
 198                 }
 199         }
 200 }
 201
 202 /*
 203  * Default callback for parsing of the Proximity Domain <-> Memory
 204  * Area mappings
 205  */
 206 int __init
 207 acpi_numa_memory_affinity_init(struct acpi_srat_mem_affinity *ma)
 208 {
 209         u64 start, end;
 210         u32 hotpluggable;
 211         int node, pxm;
 212
 213         if (srat_disabled())
 214                 goto out_err;
 215         if (ma->header.length < sizeof(struct acpi_srat_mem_affinity)) {
 216                 pr_err("SRAT: Unexpected header length: %d\n",
 217                        ma->header.length);
 218                 goto out_err_bad_srat;
 219         }
 220         if ((ma->flags & ACPI_SRAT_MEM_ENABLED) == 0)
 221                 goto out_err;
 222         hotpluggable = ma->flags & ACPI_SRAT_MEM_HOT_PLUGGABLE;
 223         if (hotpluggable && !IS_ENABLED(CONFIG_MEMORY_HOTPLUG))
 224                 goto out_err;
 225
 226         start = ma->base_address;
 227         end = start + ma->length;
 228         pxm = ma->proximity_domain;
 229         if (acpi_srat_revision <= 1)
 230                 pxm &= 0xff;
 231
 232         node = acpi_map_pxm_to_node(pxm);
 233         if (node == NUMA_NO_NODE) {
 234                 pr_err("SRAT: Too many proximity domains.\n");
 235                 goto out_err_bad_srat;
 236         }
 237
 238         if (numa_add_memblk(node, start, end) < 0) {
 239                 pr_err("SRAT: Failed to add memblk to node %u [mem %#010Lx-%#010Lx]\n",
 240                        node, (unsigned long long) start,
 241                        (unsigned long long) end - 1);
 242                 goto out_err_bad_srat;
 243         }
 244
 245         node_set(node, numa_nodes_parsed);
 246
 247         pr_info("SRAT: Node %u PXM %u [mem %#010Lx-%#010Lx]%s%s\n",
 248                 node, pxm,
 249                 (unsigned long long) start, (unsigned long long) end - 1,
 250                 hotpluggable ? " hotplug" : "",
 251                 ma->flags & ACPI_SRAT_MEM_NON_VOLATILE ? " non-volatile" : "");
 252
 253         /* Mark hotplug range in memblock. */
 254         if (hotpluggable && memblock_mark_hotplug(start, ma->length))
 255                 pr_warn("SRAT: Failed to mark hotplug range [mem %#010Lx-%#010Lx] in memblock\n",
 256                         (unsigned long long)start, (unsigned long long)end - 1);
 257
 258         max_possible_pfn = max(max_possible_pfn, PFN_UP(end - 1));
 259
 260         return 0;
 261 out_err_bad_srat:
 262         bad_srat();
 263 out_err:
 264         return -EINVAL;
 265 }
 266 #endif /* defined(CONFIG_X86) || defined (CONFIG_ARM64) */
 267
 268 static int __init acpi_parse_slit(struct acpi_table_header *table)
 269 {
 270         struct acpi_table_slit *slit = (struct acpi_table_slit *)table;
 271
 272         if (!slit_valid(slit)) {
 273                 pr_info("SLIT table looks invalid. Not used.\n");
 274                 return -EINVAL;
 275         }
 276         acpi_numa_slit_init(slit);
 277
 278         return 0;
 279 }
 280
 281 void __init __weak
 282 acpi_numa_x2apic_affinity_init(struct acpi_srat_x2apic_cpu_affinity *pa)
 283 {
 284         pr_warn("Found unsupported x2apic [0x%08x] SRAT entry\n", pa->apic_id);
 285 }
 286
 287 static int __init
 288 acpi_parse_x2apic_affinity(union acpi_subtable_headers *header,
 289                            const unsigned long end)
 290 {
 291         struct acpi_srat_x2apic_cpu_affinity *processor_affinity;
 292
 293         processor_affinity = (struct acpi_srat_x2apic_cpu_affinity *)header;
 294
 295         acpi_table_print_srat_entry(&header->common);
 296
 297         /* let architecture-dependent part to do it */
 298         acpi_numa_x2apic_affinity_init(processor_affinity);
 299
 300         return 0;
 301 }
 302
 303 static int __init
 304 acpi_parse_processor_affinity(union acpi_subtable_headers *header,
 305                               const unsigned long end)
 306 {
 307         struct acpi_srat_cpu_affinity *processor_affinity;
 308
 309         processor_affinity = (struct acpi_srat_cpu_affinity *)header;
 310
 311         acpi_table_print_srat_entry(&header->common);
 312
 313         /* let architecture-dependent part to do it */
 314         acpi_numa_processor_affinity_init(processor_affinity);
 315
 316         return 0;
 317 }
 318
 319 static int __init
 320 acpi_parse_gicc_affinity(union acpi_subtable_headers *header,
 321                          const unsigned long end)
 322 {
 323         struct acpi_srat_gicc_affinity *processor_affinity;
 324
 325         processor_affinity = (struct acpi_srat_gicc_affinity *)header;
 326
 327         acpi_table_print_srat_entry(&header->common);
 328
 329         /* let architecture-dependent part to do it */
 330         acpi_numa_gicc_affinity_init(processor_affinity);
 331
 332         return 0;
 333 }
 334
 335 static int __initdata parsed_numa_memblks;
 336
 337 static int __init
 338 acpi_parse_memory_affinity(union acpi_subtable_headers * header,
 339                            const unsigned long end)
 340 {
 341         struct acpi_srat_mem_affinity *memory_affinity;
 342
 343         memory_affinity = (struct acpi_srat_mem_affinity *)header;
 344
 345         acpi_table_print_srat_entry(&header->common);
 346
 347         /* let architecture-dependent part to do it */
 348         if (!acpi_numa_memory_affinity_init(memory_affinity))
 349                 parsed_numa_memblks++;
 350         return 0;
 351 }
 352
 353 static int __init acpi_parse_srat(struct acpi_table_header *table)
 354 {
 355         struct acpi_table_srat *srat = (struct acpi_table_srat *)table;
 356
 357         acpi_srat_revision = srat->header.revision;
 358
 359         /* Real work done in acpi_table_parse_srat below. */
 360
 361         return 0;
 362 }
 363
 364 static int __init
 365 acpi_table_parse_srat(enum acpi_srat_type id,
 366                       acpi_tbl_entry_handler handler, unsigned int max_entries)
 367 {
 368         return acpi_table_parse_entries(ACPI_SIG_SRAT,
 369                                             sizeof(struct acpi_table_srat), id,
 370                                             handler, max_entries);
 371 }
 372
 373 int __init acpi_numa_init(void)
 374 {
 375         int cnt = 0;
 376
 377         if (acpi_disabled)
 378                 return -EINVAL;
 379
 380         /*
 381          * Should not limit number with cpu num that is from NR_CPUS or nr_cpus=
 382          * SRAT cpu entries could have different order with that in MADT.
 383          * So go over all cpu entries in SRAT to get apicid to node mapping.
 384          */
 385
 386         /* SRAT: System Resource Affinity Table */
 387         if (!acpi_table_parse(ACPI_SIG_SRAT, acpi_parse_srat)) {
 388                 struct acpi_subtable_proc srat_proc[3];
 389
 390                 memset(srat_proc, 0, sizeof(srat_proc));
 391                 srat_proc[0].id = ACPI_SRAT_TYPE_CPU_AFFINITY;
 392                 srat_proc[0].handler = acpi_parse_processor_affinity;
 393                 srat_proc[1].id = ACPI_SRAT_TYPE_X2APIC_CPU_AFFINITY;
 394                 srat_proc[1].handler = acpi_parse_x2apic_affinity;
 395                 srat_proc[2].id = ACPI_SRAT_TYPE_GICC_AFFINITY;
 396                 srat_proc[2].handler = acpi_parse_gicc_affinity;
 397
 398                 acpi_table_parse_entries_array(ACPI_SIG_SRAT,
 399                                         sizeof(struct acpi_table_srat),
 400                                         srat_proc, ARRAY_SIZE(srat_proc), 0);
 401
 402                 cnt = acpi_table_parse_srat(ACPI_SRAT_TYPE_MEMORY_AFFINITY,
 403                                             acpi_parse_memory_affinity, 0);
 404         }
 405
 406         /* SLIT: System Locality Information Table */
 407         acpi_table_parse(ACPI_SIG_SLIT, acpi_parse_slit);
 408
 409         if (cnt < 0)
 410                 return cnt;
 411         else if (!parsed_numa_memblks)
 412                 return -ENOENT;
 413         return 0;
 414 }
 415
 416 static int acpi_get_pxm(acpi_handle h)
 417 {
 418         unsigned long long pxm;
 419         acpi_status status;
 420         acpi_handle handle;
 421         acpi_handle phandle = h;
 422
 423         do {
 424                 handle = phandle;
 425                 status = acpi_evaluate_integer(handle, "_PXM", NULL, &pxm);
 426                 if (ACPI_SUCCESS(status))
 427                         return pxm;
 428                 status = acpi_get_parent(handle, &phandle);
 429         } while (ACPI_SUCCESS(status));
 430         return -1;
 431 }
 432
 433 int acpi_get_node(acpi_handle handle)
 434 {
 435         int pxm;
 436
 437         pxm = acpi_get_pxm(handle);
 438
 439         return acpi_map_pxm_to_node(pxm);
 440 }
 441 EXPORT_SYMBOL(acpi_get_node);