1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * NUMA support, based on the x86 implementation. 4 * 5 * Copyright (C) 2015 Cavium Inc. 6 * Author: Ganapatrao Kulkarni <gkulkarni@cavium.com> 7 */ 8 9 #define pr_fmt(fmt) "NUMA: " fmt 10 11 #include <linux/acpi.h> 12 #include <linux/memblock.h> 13 #include <linux/module.h> 14 #include <linux/of.h> 15 16 #include <asm/sections.h> 17 #include <asm/pgalloc.h> 18 19 struct pglist_data *node_data[MAX_NUMNODES] __read_mostly; 20 EXPORT_SYMBOL(node_data); 21 nodemask_t numa_nodes_parsed __initdata; 22 static int cpu_to_node_map[NR_CPUS] = { [0 ... NR_CPUS-1] = NUMA_NO_NODE }; 23 24 static int numa_distance_cnt; 25 static u8 *numa_distance; 26 bool numa_off; 27 28 static __init int numa_parse_early_param(char *opt) 29 { 30 if (!opt) 31 return -EINVAL; 32 if (str_has_prefix(opt, "off")) 33 numa_off = true; 34 35 return 0; 36 } 37 early_param("numa", numa_parse_early_param); 38 39 cpumask_var_t node_to_cpumask_map[MAX_NUMNODES]; 40 EXPORT_SYMBOL(node_to_cpumask_map); 41 42 #ifdef CONFIG_DEBUG_PER_CPU_MAPS 43 44 /* 45 * Returns a pointer to the bitmask of CPUs on Node 'node'. 46 */ 47 const struct cpumask *cpumask_of_node(int node) 48 { 49 50 if (node == NUMA_NO_NODE) 51 return cpu_all_mask; 52 53 if (WARN_ON(node < 0 || node >= nr_node_ids)) 54 return cpu_none_mask; 55 56 if (WARN_ON(node_to_cpumask_map[node] == NULL)) 57 return cpu_online_mask; 58 59 return node_to_cpumask_map[node]; 60 } 61 EXPORT_SYMBOL(cpumask_of_node); 62 63 #endif 64 65 static void numa_update_cpu(unsigned int cpu, bool remove) 66 { 67 int nid = cpu_to_node(cpu); 68 69 if (nid == NUMA_NO_NODE) 70 return; 71 72 if (remove) 73 cpumask_clear_cpu(cpu, node_to_cpumask_map[nid]); 74 else 75 cpumask_set_cpu(cpu, node_to_cpumask_map[nid]); 76 } 77 78 void numa_add_cpu(unsigned int cpu) 79 { 80 numa_update_cpu(cpu, false); 81 } 82 83 void numa_remove_cpu(unsigned int cpu) 84 { 85 numa_update_cpu(cpu, true); 86 } 87 88 void numa_clear_node(unsigned int cpu) 89 { 90 numa_remove_cpu(cpu); 91 set_cpu_numa_node(cpu, NUMA_NO_NODE); 92 } 93 94 /* 95 * Allocate node_to_cpumask_map based on number of available nodes 96 * Requires node_possible_map to be valid. 97 * 98 * Note: cpumask_of_node() is not valid until after this is done. 99 * (Use CONFIG_DEBUG_PER_CPU_MAPS to check this.) 100 */ 101 static void __init setup_node_to_cpumask_map(void) 102 { 103 int node; 104 105 /* setup nr_node_ids if not done yet */ 106 if (nr_node_ids == MAX_NUMNODES) 107 setup_nr_node_ids(); 108 109 /* allocate and clear the mapping */ 110 for (node = 0; node < nr_node_ids; node++) { 111 alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]); 112 cpumask_clear(node_to_cpumask_map[node]); 113 } 114 115 /* cpumask_of_node() will now work */ 116 pr_debug("Node to cpumask map for %u nodes\n", nr_node_ids); 117 } 118 119 /* 120 * Set the cpu to node and mem mapping 121 */ 122 void numa_store_cpu_info(unsigned int cpu) 123 { 124 set_cpu_numa_node(cpu, cpu_to_node_map[cpu]); 125 } 126 127 void __init early_map_cpu_to_node(unsigned int cpu, int nid) 128 { 129 /* fallback to node 0 */ 130 if (nid < 0 || nid >= MAX_NUMNODES || numa_off) 131 nid = 0; 132 133 cpu_to_node_map[cpu] = nid; 134 135 /* 136 * We should set the numa node of cpu0 as soon as possible, because it 137 * has already been set up online before. cpu_to_node(0) will soon be 138 * called. 139 */ 140 if (!cpu) 141 set_cpu_numa_node(cpu, nid); 142 } 143 144 #ifdef CONFIG_HAVE_SETUP_PER_CPU_AREA 145 unsigned long __per_cpu_offset[NR_CPUS] __read_mostly; 146 EXPORT_SYMBOL(__per_cpu_offset); 147 148 static int __init early_cpu_to_node(int cpu) 149 { 150 return cpu_to_node_map[cpu]; 151 } 152 153 static int __init pcpu_cpu_distance(unsigned int from, unsigned int to) 154 { 155 return node_distance(early_cpu_to_node(from), early_cpu_to_node(to)); 156 } 157 158 static void * __init pcpu_fc_alloc(unsigned int cpu, size_t size, 159 size_t align) 160 { 161 int nid = early_cpu_to_node(cpu); 162 163 return memblock_alloc_try_nid(size, align, 164 __pa(MAX_DMA_ADDRESS), MEMBLOCK_ALLOC_ACCESSIBLE, nid); 165 } 166 167 static void __init pcpu_fc_free(void *ptr, size_t size) 168 { 169 memblock_free(ptr, size); 170 } 171 172 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK 173 static void __init pcpu_populate_pte(unsigned long addr) 174 { 175 pgd_t *pgd = pgd_offset_k(addr); 176 p4d_t *p4d; 177 pud_t *pud; 178 pmd_t *pmd; 179 180 p4d = p4d_offset(pgd, addr); 181 if (p4d_none(*p4d)) { 182 pud_t *new; 183 184 new = memblock_alloc(PAGE_SIZE, PAGE_SIZE); 185 if (!new) 186 goto err_alloc; 187 p4d_populate(&init_mm, p4d, new); 188 } 189 190 pud = pud_offset(p4d, addr); 191 if (pud_none(*pud)) { 192 pmd_t *new; 193 194 new = memblock_alloc(PAGE_SIZE, PAGE_SIZE); 195 if (!new) 196 goto err_alloc; 197 pud_populate(&init_mm, pud, new); 198 } 199 200 pmd = pmd_offset(pud, addr); 201 if (!pmd_present(*pmd)) { 202 pte_t *new; 203 204 new = memblock_alloc(PAGE_SIZE, PAGE_SIZE); 205 if (!new) 206 goto err_alloc; 207 pmd_populate_kernel(&init_mm, pmd, new); 208 } 209 210 return; 211 212 err_alloc: 213 panic("%s: Failed to allocate %lu bytes align=%lx from=%lx\n", 214 __func__, PAGE_SIZE, PAGE_SIZE, PAGE_SIZE); 215 } 216 #endif 217 218 void __init setup_per_cpu_areas(void) 219 { 220 unsigned long delta; 221 unsigned int cpu; 222 int rc = -EINVAL; 223 224 if (pcpu_chosen_fc != PCPU_FC_PAGE) { 225 /* 226 * Always reserve area for module percpu variables. That's 227 * what the legacy allocator did. 228 */ 229 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE, 230 PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, 231 pcpu_cpu_distance, 232 pcpu_fc_alloc, pcpu_fc_free); 233 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK 234 if (rc < 0) 235 pr_warn("PERCPU: %s allocator failed (%d), falling back to page size\n", 236 pcpu_fc_names[pcpu_chosen_fc], rc); 237 #endif 238 } 239 240 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK 241 if (rc < 0) 242 rc = pcpu_page_first_chunk(PERCPU_MODULE_RESERVE, 243 pcpu_fc_alloc, 244 pcpu_fc_free, 245 pcpu_populate_pte); 246 #endif 247 if (rc < 0) 248 panic("Failed to initialize percpu areas (err=%d).", rc); 249 250 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start; 251 for_each_possible_cpu(cpu) 252 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu]; 253 } 254 #endif 255 256 /** 257 * numa_add_memblk() - Set node id to memblk 258 * @nid: NUMA node ID of the new memblk 259 * @start: Start address of the new memblk 260 * @end: End address of the new memblk 261 * 262 * RETURNS: 263 * 0 on success, -errno on failure. 264 */ 265 int __init numa_add_memblk(int nid, u64 start, u64 end) 266 { 267 int ret; 268 269 ret = memblock_set_node(start, (end - start), &memblock.memory, nid); 270 if (ret < 0) { 271 pr_err("memblock [0x%llx - 0x%llx] failed to add on node %d\n", 272 start, (end - 1), nid); 273 return ret; 274 } 275 276 node_set(nid, numa_nodes_parsed); 277 return ret; 278 } 279 280 /* 281 * Initialize NODE_DATA for a node on the local memory 282 */ 283 static void __init setup_node_data(int nid, u64 start_pfn, u64 end_pfn) 284 { 285 const size_t nd_size = roundup(sizeof(pg_data_t), SMP_CACHE_BYTES); 286 u64 nd_pa; 287 void *nd; 288 int tnid; 289 290 if (start_pfn >= end_pfn) 291 pr_info("Initmem setup node %d [<memory-less node>]\n", nid); 292 293 nd_pa = memblock_phys_alloc_try_nid(nd_size, SMP_CACHE_BYTES, nid); 294 if (!nd_pa) 295 panic("Cannot allocate %zu bytes for node %d data\n", 296 nd_size, nid); 297 298 nd = __va(nd_pa); 299 300 /* report and initialize */ 301 pr_info("NODE_DATA [mem %#010Lx-%#010Lx]\n", 302 nd_pa, nd_pa + nd_size - 1); 303 tnid = early_pfn_to_nid(nd_pa >> PAGE_SHIFT); 304 if (tnid != nid) 305 pr_info("NODE_DATA(%d) on node %d\n", nid, tnid); 306 307 node_data[nid] = nd; 308 memset(NODE_DATA(nid), 0, sizeof(pg_data_t)); 309 NODE_DATA(nid)->node_id = nid; 310 NODE_DATA(nid)->node_start_pfn = start_pfn; 311 NODE_DATA(nid)->node_spanned_pages = end_pfn - start_pfn; 312 } 313 314 /* 315 * numa_free_distance 316 * 317 * The current table is freed. 318 */ 319 void __init numa_free_distance(void) 320 { 321 size_t size; 322 323 if (!numa_distance) 324 return; 325 326 size = numa_distance_cnt * numa_distance_cnt * 327 sizeof(numa_distance[0]); 328 329 memblock_free(numa_distance, size); 330 numa_distance_cnt = 0; 331 numa_distance = NULL; 332 } 333 334 /* 335 * Create a new NUMA distance table. 336 */ 337 static int __init numa_alloc_distance(void) 338 { 339 size_t size; 340 int i, j; 341 342 size = nr_node_ids * nr_node_ids * sizeof(numa_distance[0]); 343 numa_distance = memblock_alloc(size, PAGE_SIZE); 344 if (WARN_ON(!numa_distance)) 345 return -ENOMEM; 346 347 numa_distance_cnt = nr_node_ids; 348 349 /* fill with the default distances */ 350 for (i = 0; i < numa_distance_cnt; i++) 351 for (j = 0; j < numa_distance_cnt; j++) 352 numa_distance[i * numa_distance_cnt + j] = i == j ? 353 LOCAL_DISTANCE : REMOTE_DISTANCE; 354 355 pr_debug("Initialized distance table, cnt=%d\n", numa_distance_cnt); 356 357 return 0; 358 } 359 360 /** 361 * numa_set_distance() - Set inter node NUMA distance from node to node. 362 * @from: the 'from' node to set distance 363 * @to: the 'to' node to set distance 364 * @distance: NUMA distance 365 * 366 * Set the distance from node @from to @to to @distance. 367 * If distance table doesn't exist, a warning is printed. 368 * 369 * If @from or @to is higher than the highest known node or lower than zero 370 * or @distance doesn't make sense, the call is ignored. 371 */ 372 void __init numa_set_distance(int from, int to, int distance) 373 { 374 if (!numa_distance) { 375 pr_warn_once("Warning: distance table not allocated yet\n"); 376 return; 377 } 378 379 if (from >= numa_distance_cnt || to >= numa_distance_cnt || 380 from < 0 || to < 0) { 381 pr_warn_once("Warning: node ids are out of bound, from=%d to=%d distance=%d\n", 382 from, to, distance); 383 return; 384 } 385 386 if ((u8)distance != distance || 387 (from == to && distance != LOCAL_DISTANCE)) { 388 pr_warn_once("Warning: invalid distance parameter, from=%d to=%d distance=%d\n", 389 from, to, distance); 390 return; 391 } 392 393 numa_distance[from * numa_distance_cnt + to] = distance; 394 } 395 396 /* 397 * Return NUMA distance @from to @to 398 */ 399 int __node_distance(int from, int to) 400 { 401 if (from >= numa_distance_cnt || to >= numa_distance_cnt) 402 return from == to ? LOCAL_DISTANCE : REMOTE_DISTANCE; 403 return numa_distance[from * numa_distance_cnt + to]; 404 } 405 EXPORT_SYMBOL(__node_distance); 406 407 static int __init numa_register_nodes(void) 408 { 409 int nid; 410 struct memblock_region *mblk; 411 412 /* Check that valid nid is set to memblks */ 413 for_each_mem_region(mblk) { 414 int mblk_nid = memblock_get_region_node(mblk); 415 phys_addr_t start = mblk->base; 416 phys_addr_t end = mblk->base + mblk->size - 1; 417 418 if (mblk_nid == NUMA_NO_NODE || mblk_nid >= MAX_NUMNODES) { 419 pr_warn("Warning: invalid memblk node %d [mem %pap-%pap]\n", 420 mblk_nid, &start, &end); 421 return -EINVAL; 422 } 423 } 424 425 /* Finally register nodes. */ 426 for_each_node_mask(nid, numa_nodes_parsed) { 427 unsigned long start_pfn, end_pfn; 428 429 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn); 430 setup_node_data(nid, start_pfn, end_pfn); 431 node_set_online(nid); 432 } 433 434 /* Setup online nodes to actual nodes*/ 435 node_possible_map = numa_nodes_parsed; 436 437 return 0; 438 } 439 440 static int __init numa_init(int (*init_func)(void)) 441 { 442 int ret; 443 444 nodes_clear(numa_nodes_parsed); 445 nodes_clear(node_possible_map); 446 nodes_clear(node_online_map); 447 448 ret = numa_alloc_distance(); 449 if (ret < 0) 450 return ret; 451 452 ret = init_func(); 453 if (ret < 0) 454 goto out_free_distance; 455 456 if (nodes_empty(numa_nodes_parsed)) { 457 pr_info("No NUMA configuration found\n"); 458 ret = -EINVAL; 459 goto out_free_distance; 460 } 461 462 ret = numa_register_nodes(); 463 if (ret < 0) 464 goto out_free_distance; 465 466 setup_node_to_cpumask_map(); 467 468 return 0; 469 out_free_distance: 470 numa_free_distance(); 471 return ret; 472 } 473 474 /** 475 * dummy_numa_init() - Fallback dummy NUMA init 476 * 477 * Used if there's no underlying NUMA architecture, NUMA initialization 478 * fails, or NUMA is disabled on the command line. 479 * 480 * Must online at least one node (node 0) and add memory blocks that cover all 481 * allowed memory. It is unlikely that this function fails. 482 * 483 * Return: 0 on success, -errno on failure. 484 */ 485 static int __init dummy_numa_init(void) 486 { 487 phys_addr_t start = memblock_start_of_DRAM(); 488 phys_addr_t end = memblock_end_of_DRAM() - 1; 489 int ret; 490 491 if (numa_off) 492 pr_info("NUMA disabled\n"); /* Forced off on command line. */ 493 pr_info("Faking a node at [mem %pap-%pap]\n", &start, &end); 494 495 ret = numa_add_memblk(0, start, end + 1); 496 if (ret) { 497 pr_err("NUMA init failed\n"); 498 return ret; 499 } 500 501 numa_off = true; 502 return 0; 503 } 504 505 #ifdef CONFIG_ACPI_NUMA 506 static int __init arch_acpi_numa_init(void) 507 { 508 int ret; 509 510 ret = acpi_numa_init(); 511 if (ret) { 512 pr_info("Failed to initialise from firmware\n"); 513 return ret; 514 } 515 516 return srat_disabled() ? -EINVAL : 0; 517 } 518 #else 519 static int __init arch_acpi_numa_init(void) 520 { 521 return -EOPNOTSUPP; 522 } 523 #endif 524 525 /** 526 * arch_numa_init() - Initialize NUMA 527 * 528 * Try each configured NUMA initialization method until one succeeds. The 529 * last fallback is dummy single node config encompassing whole memory. 530 */ 531 void __init arch_numa_init(void) 532 { 533 if (!numa_off) { 534 if (!acpi_disabled && !numa_init(arch_acpi_numa_init)) 535 return; 536 if (acpi_disabled && !numa_init(of_numa_init)) 537 return; 538 } 539 540 numa_init(dummy_numa_init); 541 } 542