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 static int acpi_numa __initdata;
  31
  32 static int last_real_pxm;
  33
  34 void __init disable_srat(void)
  35 {
  36         acpi_numa = -1;
  37 }
  38
  39 int pxm_to_node(int pxm)
  40 {
  41         if (pxm < 0 || pxm >= MAX_PXM_DOMAINS || numa_off)
  42                 return NUMA_NO_NODE;
  43         return pxm_to_node_map[pxm];
  44 }
  45 EXPORT_SYMBOL(pxm_to_node);
  46
  47 int node_to_pxm(int node)
  48 {
  49         if (node < 0)
  50                 return PXM_INVAL;
  51         return node_to_pxm_map[node];
  52 }
  53
  54 static void __acpi_map_pxm_to_node(int pxm, int node)
  55 {
  56         if (pxm_to_node_map[pxm] == NUMA_NO_NODE || node < pxm_to_node_map[pxm])
  57                 pxm_to_node_map[pxm] = node;
  58         if (node_to_pxm_map[node] == PXM_INVAL || pxm < node_to_pxm_map[node])
  59                 node_to_pxm_map[node] = pxm;
  60 }
  61
  62 int acpi_map_pxm_to_node(int pxm)
  63 {
  64         int node;
  65
  66         if (pxm < 0 || pxm >= MAX_PXM_DOMAINS || numa_off)
  67                 return NUMA_NO_NODE;
  68
  69         node = pxm_to_node_map[pxm];
  70
  71         if (node == NUMA_NO_NODE) {
  72                 node = first_unset_node(nodes_found_map);
  73                 if (node >= MAX_NUMNODES)
  74                         return NUMA_NO_NODE;
  75                 __acpi_map_pxm_to_node(pxm, node);
  76                 node_set(node, nodes_found_map);
  77         }
  78
  79         return node;
  80 }
  81 EXPORT_SYMBOL(acpi_map_pxm_to_node);
  82
  83 static void __init
  84 acpi_table_print_srat_entry(struct acpi_subtable_header *header)
  85 {
  86         switch (header->type) {
  87         case ACPI_SRAT_TYPE_CPU_AFFINITY:
  88                 {
  89                         struct acpi_srat_cpu_affinity *p =
  90                             (struct acpi_srat_cpu_affinity *)header;
  91                         pr_debug("SRAT Processor (id[0x%02x] eid[0x%02x]) in proximity domain %d %s\n",
  92                                  p->apic_id, p->local_sapic_eid,
  93                                  p->proximity_domain_lo,
  94                                  (p->flags & ACPI_SRAT_CPU_ENABLED) ?
  95                                  "enabled" : "disabled");
  96                 }
  97                 break;
  98
  99         case ACPI_SRAT_TYPE_MEMORY_AFFINITY:
 100                 {
 101                         struct acpi_srat_mem_affinity *p =
 102                             (struct acpi_srat_mem_affinity *)header;
 103                         pr_debug("SRAT Memory (0x%llx length 0x%llx) in proximity domain %d %s%s%s\n",
 104                                  (unsigned long long)p->base_address,
 105                                  (unsigned long long)p->length,
 106                                  p->proximity_domain,
 107                                  (p->flags & ACPI_SRAT_MEM_ENABLED) ?
 108                                  "enabled" : "disabled",
 109                                  (p->flags & ACPI_SRAT_MEM_HOT_PLUGGABLE) ?
 110                                  " hot-pluggable" : "",
 111                                  (p->flags & ACPI_SRAT_MEM_NON_VOLATILE) ?
 112                                  " non-volatile" : "");
 113                 }
 114                 break;
 115
 116         case ACPI_SRAT_TYPE_X2APIC_CPU_AFFINITY:
 117                 {
 118                         struct acpi_srat_x2apic_cpu_affinity *p =
 119                             (struct acpi_srat_x2apic_cpu_affinity *)header;
 120                         pr_debug("SRAT Processor (x2apicid[0x%08x]) in proximity domain %d %s\n",
 121                                  p->apic_id,
 122                                  p->proximity_domain,
 123                                  (p->flags & ACPI_SRAT_CPU_ENABLED) ?
 124                                  "enabled" : "disabled");
 125                 }
 126                 break;
 127
 128         case ACPI_SRAT_TYPE_GICC_AFFINITY:
 129                 {
 130                         struct acpi_srat_gicc_affinity *p =
 131                             (struct acpi_srat_gicc_affinity *)header;
 132                         pr_debug("SRAT Processor (acpi id[0x%04x]) in proximity domain %d %s\n",
 133                                  p->acpi_processor_uid,
 134                                  p->proximity_domain,
 135                                  (p->flags & ACPI_SRAT_GICC_ENABLED) ?
 136                                  "enabled" : "disabled");
 137                 }
 138                 break;
 139
 140         case ACPI_SRAT_TYPE_GENERIC_AFFINITY:
 141         {
 142                 struct acpi_srat_generic_affinity *p =
 143                         (struct acpi_srat_generic_affinity *)header;
 144
 145                 if (p->device_handle_type == 0) {
 146                         /*
 147                          * For pci devices this may be the only place they
 148                          * are assigned a proximity domain
 149                          */
 150                         pr_debug("SRAT Generic Initiator(Seg:%u BDF:%u) in proximity domain %d %s\n",
 151                                  *(u16 *)(&p->device_handle[0]),
 152                                  *(u16 *)(&p->device_handle[2]),
 153                                  p->proximity_domain,
 154                                  (p->flags & ACPI_SRAT_GENERIC_AFFINITY_ENABLED) ?
 155                                 "enabled" : "disabled");
 156                 } else {
 157                         /*
 158                          * In this case we can rely on the device having a
 159                          * proximity domain reference
 160                          */
 161                         pr_debug("SRAT Generic Initiator(HID=%.8s UID=%.4s) in proximity domain %d %s\n",
 162                                 (char *)(&p->device_handle[0]),
 163                                 (char *)(&p->device_handle[8]),
 164                                 p->proximity_domain,
 165                                 (p->flags & ACPI_SRAT_GENERIC_AFFINITY_ENABLED) ?
 166                                 "enabled" : "disabled");
 167                 }
 168         }
 169         break;
 170         default:
 171                 pr_warn("Found unsupported SRAT entry (type = 0x%x)\n",
 172                         header->type);
 173                 break;
 174         }
 175 }
 176
 177 /*
 178  * A lot of BIOS fill in 10 (= no distance) everywhere. This messes
 179  * up the NUMA heuristics which wants the local node to have a smaller
 180  * distance than the others.
 181  * Do some quick checks here and only use the SLIT if it passes.
 182  */
 183 static int __init slit_valid(struct acpi_table_slit *slit)
 184 {
 185         int i, j;
 186         int d = slit->locality_count;
 187         for (i = 0; i < d; i++) {
 188                 for (j = 0; j < d; j++) {
 189                         u8 val = slit->entry[d*i + j];
 190                         if (i == j) {
 191                                 if (val != LOCAL_DISTANCE)
 192                                         return 0;
 193                         } else if (val <= LOCAL_DISTANCE)
 194                                 return 0;
 195                 }
 196         }
 197         return 1;
 198 }
 199
 200 void __init bad_srat(void)
 201 {
 202         pr_err("SRAT: SRAT not used.\n");
 203         disable_srat();
 204 }
 205
 206 int __init srat_disabled(void)
 207 {
 208         return acpi_numa < 0;
 209 }
 210
 211 __weak int __init numa_fill_memblks(u64 start, u64 end)
 212 {
 213         return NUMA_NO_MEMBLK;
 214 }
 215
 216 /*
 217  * Callback for SLIT parsing.  pxm_to_node() returns NUMA_NO_NODE for
 218  * I/O localities since SRAT does not list them.  I/O localities are
 219  * not supported at this point.
 220  */
 221 static int __init acpi_parse_slit(struct acpi_table_header *table)
 222 {
 223         struct acpi_table_slit *slit = (struct acpi_table_slit *)table;
 224         int i, j;
 225
 226         if (!slit_valid(slit)) {
 227                 pr_info("SLIT table looks invalid. Not used.\n");
 228                 return -EINVAL;
 229         }
 230
 231         for (i = 0; i < slit->locality_count; i++) {
 232                 const int from_node = pxm_to_node(i);
 233
 234                 if (from_node == NUMA_NO_NODE)
 235                         continue;
 236
 237                 for (j = 0; j < slit->locality_count; j++) {
 238                         const int to_node = pxm_to_node(j);
 239
 240                         if (to_node == NUMA_NO_NODE)
 241                                 continue;
 242
 243                         numa_set_distance(from_node, to_node,
 244                                 slit->entry[slit->locality_count * i + j]);
 245                 }
 246         }
 247
 248         return 0;
 249 }
 250
 251 static int parsed_numa_memblks __initdata;
 252
 253 static int __init
 254 acpi_parse_memory_affinity(union acpi_subtable_headers *header,
 255                            const unsigned long table_end)
 256 {
 257         struct acpi_srat_mem_affinity *ma;
 258         u64 start, end;
 259         u32 hotpluggable;
 260         int node, pxm;
 261
 262         ma = (struct acpi_srat_mem_affinity *)header;
 263
 264         acpi_table_print_srat_entry(&header->common);
 265
 266         if (srat_disabled())
 267                 return 0;
 268         if (ma->header.length < sizeof(struct acpi_srat_mem_affinity)) {
 269                 pr_err("SRAT: Unexpected header length: %d\n",
 270                        ma->header.length);
 271                 goto out_err_bad_srat;
 272         }
 273         if ((ma->flags & ACPI_SRAT_MEM_ENABLED) == 0)
 274                 return 0;
 275         hotpluggable = IS_ENABLED(CONFIG_MEMORY_HOTPLUG) &&
 276                 (ma->flags & ACPI_SRAT_MEM_HOT_PLUGGABLE);
 277
 278         start = ma->base_address;
 279         end = start + ma->length;
 280         pxm = ma->proximity_domain;
 281         if (acpi_srat_revision <= 1)
 282                 pxm &= 0xff;
 283
 284         node = acpi_map_pxm_to_node(pxm);
 285         if (node == NUMA_NO_NODE) {
 286                 pr_err("SRAT: Too many proximity domains.\n");
 287                 goto out_err_bad_srat;
 288         }
 289
 290         if (numa_add_memblk(node, start, end) < 0) {
 291                 pr_err("SRAT: Failed to add memblk to node %u [mem %#010Lx-%#010Lx]\n",
 292                        node, (unsigned long long) start,
 293                        (unsigned long long) end - 1);
 294                 goto out_err_bad_srat;
 295         }
 296
 297         node_set(node, numa_nodes_parsed);
 298
 299         pr_info("SRAT: Node %u PXM %u [mem %#010Lx-%#010Lx]%s%s\n",
 300                 node, pxm,
 301                 (unsigned long long) start, (unsigned long long) end - 1,
 302                 hotpluggable ? " hotplug" : "",
 303                 ma->flags & ACPI_SRAT_MEM_NON_VOLATILE ? " non-volatile" : "");
 304
 305         /* Mark hotplug range in memblock. */
 306         if (hotpluggable && memblock_mark_hotplug(start, ma->length))
 307                 pr_warn("SRAT: Failed to mark hotplug range [mem %#010Lx-%#010Lx] in memblock\n",
 308                         (unsigned long long)start, (unsigned long long)end - 1);
 309
 310         max_possible_pfn = max(max_possible_pfn, PFN_UP(end - 1));
 311
 312         parsed_numa_memblks++;
 313
 314         return 0;
 315
 316 out_err_bad_srat:
 317         /* Just disable SRAT, but do not fail and ignore errors. */
 318         bad_srat();
 319
 320         return 0;
 321 }
 322
 323 static int __init acpi_parse_cfmws(union acpi_subtable_headers *header,
 324                                    void *arg, const unsigned long table_end)
 325 {
 326         struct acpi_cedt_cfmws *cfmws;
 327         int *fake_pxm = arg;
 328         u64 start, end;
 329         int node;
 330
 331         cfmws = (struct acpi_cedt_cfmws *)header;
 332         start = cfmws->base_hpa;
 333         end = cfmws->base_hpa + cfmws->window_size;
 334
 335         /*
 336          * The SRAT may have already described NUMA details for all,
 337          * or a portion of, this CFMWS HPA range. Extend the memblks
 338          * found for any portion of the window to cover the entire
 339          * window.
 340          */
 341         if (!numa_fill_memblks(start, end))
 342                 return 0;
 343
 344         /* No SRAT description. Create a new node. */
 345         node = acpi_map_pxm_to_node(*fake_pxm);
 346
 347         if (node == NUMA_NO_NODE) {
 348                 pr_err("ACPI NUMA: Too many proximity domains while processing CFMWS.\n");
 349                 return -EINVAL;
 350         }
 351
 352         if (numa_add_memblk(node, start, end) < 0) {
 353                 /* CXL driver must handle the NUMA_NO_NODE case */
 354                 pr_warn("ACPI NUMA: Failed to add memblk for CFMWS node %d [mem %#llx-%#llx]\n",
 355                         node, start, end);
 356         }
 357         node_set(node, numa_nodes_parsed);
 358
 359         /* Set the next available fake_pxm value */
 360         (*fake_pxm)++;
 361         return 0;
 362 }
 363
 364 void __init __weak
 365 acpi_numa_x2apic_affinity_init(struct acpi_srat_x2apic_cpu_affinity *pa)
 366 {
 367         pr_warn("Found unsupported x2apic [0x%08x] SRAT entry\n", pa->apic_id);
 368 }
 369
 370 static int __init
 371 acpi_parse_x2apic_affinity(union acpi_subtable_headers *header,
 372                            const unsigned long end)
 373 {
 374         struct acpi_srat_x2apic_cpu_affinity *processor_affinity;
 375
 376         processor_affinity = (struct acpi_srat_x2apic_cpu_affinity *)header;
 377
 378         acpi_table_print_srat_entry(&header->common);
 379
 380         /* let architecture-dependent part to do it */
 381         acpi_numa_x2apic_affinity_init(processor_affinity);
 382
 383         return 0;
 384 }
 385
 386 static int __init
 387 acpi_parse_processor_affinity(union acpi_subtable_headers *header,
 388                               const unsigned long end)
 389 {
 390         struct acpi_srat_cpu_affinity *processor_affinity;
 391
 392         processor_affinity = (struct acpi_srat_cpu_affinity *)header;
 393
 394         acpi_table_print_srat_entry(&header->common);
 395
 396         /* let architecture-dependent part to do it */
 397         acpi_numa_processor_affinity_init(processor_affinity);
 398
 399         return 0;
 400 }
 401
 402 static int __init
 403 acpi_parse_gicc_affinity(union acpi_subtable_headers *header,
 404                          const unsigned long end)
 405 {
 406         struct acpi_srat_gicc_affinity *processor_affinity;
 407
 408         processor_affinity = (struct acpi_srat_gicc_affinity *)header;
 409
 410         acpi_table_print_srat_entry(&header->common);
 411
 412         /* let architecture-dependent part to do it */
 413         acpi_numa_gicc_affinity_init(processor_affinity);
 414
 415         return 0;
 416 }
 417
 418 #if defined(CONFIG_X86) || defined(CONFIG_ARM64)
 419 static int __init
 420 acpi_parse_gi_affinity(union acpi_subtable_headers *header,
 421                        const unsigned long end)
 422 {
 423         struct acpi_srat_generic_affinity *gi_affinity;
 424         int node;
 425
 426         gi_affinity = (struct acpi_srat_generic_affinity *)header;
 427         if (!gi_affinity)
 428                 return -EINVAL;
 429         acpi_table_print_srat_entry(&header->common);
 430
 431         if (!(gi_affinity->flags & ACPI_SRAT_GENERIC_AFFINITY_ENABLED))
 432                 return -EINVAL;
 433
 434         node = acpi_map_pxm_to_node(gi_affinity->proximity_domain);
 435         if (node == NUMA_NO_NODE) {
 436                 pr_err("SRAT: Too many proximity domains.\n");
 437                 return -EINVAL;
 438         }
 439         node_set(node, numa_nodes_parsed);
 440         node_set_state(node, N_GENERIC_INITIATOR);
 441
 442         return 0;
 443 }
 444 #else
 445 static int __init
 446 acpi_parse_gi_affinity(union acpi_subtable_headers *header,
 447                        const unsigned long end)
 448 {
 449         return 0;
 450 }
 451 #endif /* defined(CONFIG_X86) || defined (CONFIG_ARM64) */
 452
 453 static int __init acpi_parse_srat(struct acpi_table_header *table)
 454 {
 455         struct acpi_table_srat *srat = (struct acpi_table_srat *)table;
 456
 457         acpi_srat_revision = srat->header.revision;
 458
 459         /* Real work done in acpi_table_parse_srat below. */
 460
 461         return 0;
 462 }
 463
 464 static int __init
 465 acpi_table_parse_srat(enum acpi_srat_type id,
 466                       acpi_tbl_entry_handler handler, unsigned int max_entries)
 467 {
 468         return acpi_table_parse_entries(ACPI_SIG_SRAT,
 469                                             sizeof(struct acpi_table_srat), id,
 470                                             handler, max_entries);
 471 }
 472
 473 int __init acpi_numa_init(void)
 474 {
 475         int i, fake_pxm, cnt = 0;
 476
 477         if (acpi_disabled)
 478                 return -EINVAL;
 479
 480         /*
 481          * Should not limit number with cpu num that is from NR_CPUS or nr_cpus=
 482          * SRAT cpu entries could have different order with that in MADT.
 483          * So go over all cpu entries in SRAT to get apicid to node mapping.
 484          */
 485
 486         /* SRAT: System Resource Affinity Table */
 487         if (!acpi_table_parse(ACPI_SIG_SRAT, acpi_parse_srat)) {
 488                 struct acpi_subtable_proc srat_proc[4];
 489
 490                 memset(srat_proc, 0, sizeof(srat_proc));
 491                 srat_proc[0].id = ACPI_SRAT_TYPE_CPU_AFFINITY;
 492                 srat_proc[0].handler = acpi_parse_processor_affinity;
 493                 srat_proc[1].id = ACPI_SRAT_TYPE_X2APIC_CPU_AFFINITY;
 494                 srat_proc[1].handler = acpi_parse_x2apic_affinity;
 495                 srat_proc[2].id = ACPI_SRAT_TYPE_GICC_AFFINITY;
 496                 srat_proc[2].handler = acpi_parse_gicc_affinity;
 497                 srat_proc[3].id = ACPI_SRAT_TYPE_GENERIC_AFFINITY;
 498                 srat_proc[3].handler = acpi_parse_gi_affinity;
 499
 500                 acpi_table_parse_entries_array(ACPI_SIG_SRAT,
 501                                         sizeof(struct acpi_table_srat),
 502                                         srat_proc, ARRAY_SIZE(srat_proc), 0);
 503
 504                 cnt = acpi_table_parse_srat(ACPI_SRAT_TYPE_MEMORY_AFFINITY,
 505                                             acpi_parse_memory_affinity, 0);
 506         }
 507
 508         /* SLIT: System Locality Information Table */
 509         acpi_table_parse(ACPI_SIG_SLIT, acpi_parse_slit);
 510
 511         /*
 512          * CXL Fixed Memory Window Structures (CFMWS) must be parsed
 513          * after the SRAT. Create NUMA Nodes for CXL memory ranges that
 514          * are defined in the CFMWS and not already defined in the SRAT.
 515          * Initialize a fake_pxm as the first available PXM to emulate.
 516          */
 517
 518         /* fake_pxm is the next unused PXM value after SRAT parsing */
 519         for (i = 0, fake_pxm = -1; i < MAX_NUMNODES; i++) {
 520                 if (node_to_pxm_map[i] > fake_pxm)
 521                         fake_pxm = node_to_pxm_map[i];
 522         }
 523         last_real_pxm = fake_pxm;
 524         fake_pxm++;
 525         acpi_table_parse_cedt(ACPI_CEDT_TYPE_CFMWS, acpi_parse_cfmws,
 526                               &fake_pxm);
 527
 528         if (cnt < 0)
 529                 return cnt;
 530         else if (!parsed_numa_memblks)
 531                 return -ENOENT;
 532         return 0;
 533 }
 534
 535 bool acpi_node_backed_by_real_pxm(int nid)
 536 {
 537         int pxm = node_to_pxm(nid);
 538
 539         return pxm <= last_real_pxm;
 540 }
 541 EXPORT_SYMBOL_GPL(acpi_node_backed_by_real_pxm);
 542
 543 static int acpi_get_pxm(acpi_handle h)
 544 {
 545         unsigned long long pxm;
 546         acpi_status status;
 547         acpi_handle handle;
 548         acpi_handle phandle = h;
 549
 550         do {
 551                 handle = phandle;
 552                 status = acpi_evaluate_integer(handle, "_PXM", NULL, &pxm);
 553                 if (ACPI_SUCCESS(status))
 554                         return pxm;
 555                 status = acpi_get_parent(handle, &phandle);
 556         } while (ACPI_SUCCESS(status));
 557         return -1;
 558 }
 559
 560 int acpi_get_node(acpi_handle handle)
 561 {
 562         int pxm;
 563
 564         pxm = acpi_get_pxm(handle);
 565
 566         return pxm_to_node(pxm);
 567 }
 568 EXPORT_SYMBOL(acpi_get_node);