1// SPDX-License-Identifier: GPL-2.0-only 2/* Common code for 32 and 64-bit NUMA */ 3#include <linux/acpi.h> 4#include <linux/kernel.h> 5#include <linux/mm.h> 6#include <linux/string.h> 7#include <linux/init.h> 8#include <linux/memblock.h> 9#include <linux/mmzone.h> 10#include <linux/ctype.h> 11#include <linux/nodemask.h> 12#include <linux/sched.h> 13#include <linux/topology.h> 14#include <linux/sort.h> 15 16#include <asm/e820/api.h> 17#include <asm/proto.h> 18#include <asm/dma.h> 19#include <asm/amd_nb.h> 20 21#include "numa_internal.h" 22 23int numa_off; 24nodemask_t numa_nodes_parsed __initdata; 25 26struct pglist_data *node_data[MAX_NUMNODES] __read_mostly; 27EXPORT_SYMBOL(node_data); 28 29static struct numa_meminfo numa_meminfo __initdata_or_meminfo; 30static struct numa_meminfo numa_reserved_meminfo __initdata_or_meminfo; 31 32static int numa_distance_cnt; 33static u8 *numa_distance; 34 35static __init int numa_setup(char *opt) 36{ 37 if (!opt) 38 return -EINVAL; 39 if (!strncmp(opt, "off", 3)) 40 numa_off = 1; 41 if (!strncmp(opt, "fake=", 5)) 42 return numa_emu_cmdline(opt + 5); 43 if (!strncmp(opt, "noacpi", 6)) 44 disable_srat(); 45 if (!strncmp(opt, "nohmat", 6)) 46 disable_hmat(); 47 return 0; 48} 49early_param("numa", numa_setup); 50 51/* 52 * apicid, cpu, node mappings 53 */ 54s16 __apicid_to_node[MAX_LOCAL_APIC] = { 55 [0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE 56}; 57 58int numa_cpu_node(int cpu) 59{ 60 int apicid = early_per_cpu(x86_cpu_to_apicid, cpu); 61 62 if (apicid != BAD_APICID) 63 return __apicid_to_node[apicid]; 64 return NUMA_NO_NODE; 65} 66 67cpumask_var_t node_to_cpumask_map[MAX_NUMNODES]; 68EXPORT_SYMBOL(node_to_cpumask_map); 69 70/* 71 * Map cpu index to node index 72 */ 73DEFINE_EARLY_PER_CPU(int, x86_cpu_to_node_map, NUMA_NO_NODE); 74EXPORT_EARLY_PER_CPU_SYMBOL(x86_cpu_to_node_map); 75 76void numa_set_node(int cpu, int node) 77{ 78 int *cpu_to_node_map = early_per_cpu_ptr(x86_cpu_to_node_map); 79 80 /* early setting, no percpu area yet */ 81 if (cpu_to_node_map) { 82 cpu_to_node_map[cpu] = node; 83 return; 84 } 85 86#ifdef CONFIG_DEBUG_PER_CPU_MAPS 87 if (cpu >= nr_cpu_ids || !cpu_possible(cpu)) { 88 printk(KERN_ERR "numa_set_node: invalid cpu# (%d)\n", cpu); 89 dump_stack(); 90 return; 91 } 92#endif 93 per_cpu(x86_cpu_to_node_map, cpu) = node; 94 95 set_cpu_numa_node(cpu, node); 96} 97 98void numa_clear_node(int cpu) 99{ 100 numa_set_node(cpu, NUMA_NO_NODE); 101} 102 103/* 104 * Allocate node_to_cpumask_map based on number of available nodes 105 * Requires node_possible_map to be valid. 106 * 107 * Note: cpumask_of_node() is not valid until after this is done. 108 * (Use CONFIG_DEBUG_PER_CPU_MAPS to check this.) 109 */ 110void __init setup_node_to_cpumask_map(void) 111{ 112 unsigned int node; 113 114 /* setup nr_node_ids if not done yet */ 115 if (nr_node_ids == MAX_NUMNODES) 116 setup_nr_node_ids(); 117 118 /* allocate the map */ 119 for (node = 0; node < nr_node_ids; node++) 120 alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]); 121 122 /* cpumask_of_node() will now work */ 123 pr_debug("Node to cpumask map for %u nodes\n", nr_node_ids); 124} 125 126static int __init numa_add_memblk_to(int nid, u64 start, u64 end, 127 struct numa_meminfo *mi) 128{ 129 /* ignore zero length blks */ 130 if (start == end) 131 return 0; 132 133 /* whine about and ignore invalid blks */ 134 if (start > end || nid < 0 || nid >= MAX_NUMNODES) { 135 pr_warn("Warning: invalid memblk node %d [mem %#010Lx-%#010Lx]\n", 136 nid, start, end - 1); 137 return 0; 138 } 139 140 if (mi->nr_blks >= NR_NODE_MEMBLKS) { 141 pr_err("too many memblk ranges\n"); 142 return -EINVAL; 143 } 144 145 mi->blk[mi->nr_blks].start = start; 146 mi->blk[mi->nr_blks].end = end; 147 mi->blk[mi->nr_blks].nid = nid; 148 mi->nr_blks++; 149 return 0; 150} 151 152/** 153 * numa_remove_memblk_from - Remove one numa_memblk from a numa_meminfo 154 * @idx: Index of memblk to remove 155 * @mi: numa_meminfo to remove memblk from 156 * 157 * Remove @idx'th numa_memblk from @mi by shifting @mi->blk[] and 158 * decrementing @mi->nr_blks. 159 */ 160void __init numa_remove_memblk_from(int idx, struct numa_meminfo *mi) 161{ 162 mi->nr_blks--; 163 memmove(&mi->blk[idx], &mi->blk[idx + 1], 164 (mi->nr_blks - idx) * sizeof(mi->blk[0])); 165} 166 167/** 168 * numa_move_tail_memblk - Move a numa_memblk from one numa_meminfo to another 169 * @dst: numa_meminfo to append block to 170 * @idx: Index of memblk to remove 171 * @src: numa_meminfo to remove memblk from 172 */ 173static void __init numa_move_tail_memblk(struct numa_meminfo *dst, int idx, 174 struct numa_meminfo *src) 175{ 176 dst->blk[dst->nr_blks++] = src->blk[idx]; 177 numa_remove_memblk_from(idx, src); 178} 179 180/** 181 * numa_add_memblk - Add one numa_memblk to numa_meminfo 182 * @nid: NUMA node ID of the new memblk 183 * @start: Start address of the new memblk 184 * @end: End address of the new memblk 185 * 186 * Add a new memblk to the default numa_meminfo. 187 * 188 * RETURNS: 189 * 0 on success, -errno on failure. 190 */ 191int __init numa_add_memblk(int nid, u64 start, u64 end) 192{ 193 return numa_add_memblk_to(nid, start, end, &numa_meminfo); 194} 195 196/* Allocate NODE_DATA for a node on the local memory */ 197static void __init alloc_node_data(int nid) 198{ 199 const size_t nd_size = roundup(sizeof(pg_data_t), PAGE_SIZE); 200 u64 nd_pa; 201 void *nd; 202 int tnid; 203 204 /* 205 * Allocate node data. Try node-local memory and then any node. 206 * Never allocate in DMA zone. 207 */ 208 nd_pa = memblock_phys_alloc_try_nid(nd_size, SMP_CACHE_BYTES, nid); 209 if (!nd_pa) { 210 pr_err("Cannot find %zu bytes in any node (initial node: %d)\n", 211 nd_size, nid); 212 return; 213 } 214 nd = __va(nd_pa); 215 216 /* report and initialize */ 217 printk(KERN_INFO "NODE_DATA(%d) allocated [mem %#010Lx-%#010Lx]\n", nid, 218 nd_pa, nd_pa + nd_size - 1); 219 tnid = early_pfn_to_nid(nd_pa >> PAGE_SHIFT); 220 if (tnid != nid) 221 printk(KERN_INFO " NODE_DATA(%d) on node %d\n", nid, tnid); 222 223 node_data[nid] = nd; 224 memset(NODE_DATA(nid), 0, sizeof(pg_data_t)); 225 226 node_set_online(nid); 227} 228 229/** 230 * numa_cleanup_meminfo - Cleanup a numa_meminfo 231 * @mi: numa_meminfo to clean up 232 * 233 * Sanitize @mi by merging and removing unnecessary memblks. Also check for 234 * conflicts and clear unused memblks. 235 * 236 * RETURNS: 237 * 0 on success, -errno on failure. 238 */ 239int __init numa_cleanup_meminfo(struct numa_meminfo *mi) 240{ 241 const u64 low = 0; 242 const u64 high = PFN_PHYS(max_pfn); 243 int i, j, k; 244 245 /* first, trim all entries */ 246 for (i = 0; i < mi->nr_blks; i++) { 247 struct numa_memblk *bi = &mi->blk[i]; 248 249 /* move / save reserved memory ranges */ 250 if (!memblock_overlaps_region(&memblock.memory, 251 bi->start, bi->end - bi->start)) { 252 numa_move_tail_memblk(&numa_reserved_meminfo, i--, mi); 253 continue; 254 } 255 256 /* make sure all non-reserved blocks are inside the limits */ 257 bi->start = max(bi->start, low); 258 259 /* preserve info for non-RAM areas above 'max_pfn': */ 260 if (bi->end > high) { 261 numa_add_memblk_to(bi->nid, high, bi->end, 262 &numa_reserved_meminfo); 263 bi->end = high; 264 } 265 266 /* and there's no empty block */ 267 if (bi->start >= bi->end) 268 numa_remove_memblk_from(i--, mi); 269 } 270 271 /* merge neighboring / overlapping entries */ 272 for (i = 0; i < mi->nr_blks; i++) { 273 struct numa_memblk *bi = &mi->blk[i]; 274 275 for (j = i + 1; j < mi->nr_blks; j++) { 276 struct numa_memblk *bj = &mi->blk[j]; 277 u64 start, end; 278 279 /* 280 * See whether there are overlapping blocks. Whine 281 * about but allow overlaps of the same nid. They 282 * will be merged below. 283 */ 284 if (bi->end > bj->start && bi->start < bj->end) { 285 if (bi->nid != bj->nid) { 286 pr_err("node %d [mem %#010Lx-%#010Lx] overlaps with node %d [mem %#010Lx-%#010Lx]\n", 287 bi->nid, bi->start, bi->end - 1, 288 bj->nid, bj->start, bj->end - 1); 289 return -EINVAL; 290 } 291 pr_warn("Warning: node %d [mem %#010Lx-%#010Lx] overlaps with itself [mem %#010Lx-%#010Lx]\n", 292 bi->nid, bi->start, bi->end - 1, 293 bj->start, bj->end - 1); 294 } 295 296 /* 297 * Join together blocks on the same node, holes 298 * between which don't overlap with memory on other 299 * nodes. 300 */ 301 if (bi->nid != bj->nid) 302 continue; 303 start = min(bi->start, bj->start); 304 end = max(bi->end, bj->end); 305 for (k = 0; k < mi->nr_blks; k++) { 306 struct numa_memblk *bk = &mi->blk[k]; 307 308 if (bi->nid == bk->nid) 309 continue; 310 if (start < bk->end && end > bk->start) 311 break; 312 } 313 if (k < mi->nr_blks) 314 continue; 315 printk(KERN_INFO "NUMA: Node %d [mem %#010Lx-%#010Lx] + [mem %#010Lx-%#010Lx] -> [mem %#010Lx-%#010Lx]\n", 316 bi->nid, bi->start, bi->end - 1, bj->start, 317 bj->end - 1, start, end - 1); 318 bi->start = start; 319 bi->end = end; 320 numa_remove_memblk_from(j--, mi); 321 } 322 } 323 324 /* clear unused ones */ 325 for (i = mi->nr_blks; i < ARRAY_SIZE(mi->blk); i++) { 326 mi->blk[i].start = mi->blk[i].end = 0; 327 mi->blk[i].nid = NUMA_NO_NODE; 328 } 329 330 return 0; 331} 332 333/* 334 * Set nodes, which have memory in @mi, in *@nodemask. 335 */ 336static void __init numa_nodemask_from_meminfo(nodemask_t *nodemask, 337 const struct numa_meminfo *mi) 338{ 339 int i; 340 341 for (i = 0; i < ARRAY_SIZE(mi->blk); i++) 342 if (mi->blk[i].start != mi->blk[i].end && 343 mi->blk[i].nid != NUMA_NO_NODE) 344 node_set(mi->blk[i].nid, *nodemask); 345} 346 347/** 348 * numa_reset_distance - Reset NUMA distance table 349 * 350 * The current table is freed. The next numa_set_distance() call will 351 * create a new one. 352 */ 353void __init numa_reset_distance(void) 354{ 355 size_t size = numa_distance_cnt * numa_distance_cnt * sizeof(numa_distance[0]); 356 357 /* numa_distance could be 1LU marking allocation failure, test cnt */ 358 if (numa_distance_cnt) 359 memblock_free(numa_distance, size); 360 numa_distance_cnt = 0; 361 numa_distance = NULL; /* enable table creation */ 362} 363 364static int __init numa_alloc_distance(void) 365{ 366 nodemask_t nodes_parsed; 367 size_t size; 368 int i, j, cnt = 0; 369 u64 phys; 370 371 /* size the new table and allocate it */ 372 nodes_parsed = numa_nodes_parsed; 373 numa_nodemask_from_meminfo(&nodes_parsed, &numa_meminfo); 374 375 for_each_node_mask(i, nodes_parsed) 376 cnt = i; 377 cnt++; 378 size = cnt * cnt * sizeof(numa_distance[0]); 379 380 phys = memblock_phys_alloc_range(size, PAGE_SIZE, 0, 381 PFN_PHYS(max_pfn_mapped)); 382 if (!phys) { 383 pr_warn("Warning: can't allocate distance table!\n"); 384 /* don't retry until explicitly reset */ 385 numa_distance = (void *)1LU; 386 return -ENOMEM; 387 } 388 389 numa_distance = __va(phys); 390 numa_distance_cnt = cnt; 391 392 /* fill with the default distances */ 393 for (i = 0; i < cnt; i++) 394 for (j = 0; j < cnt; j++) 395 numa_distance[i * cnt + j] = i == j ? 396 LOCAL_DISTANCE : REMOTE_DISTANCE; 397 printk(KERN_DEBUG "NUMA: Initialized distance table, cnt=%d\n", cnt); 398 399 return 0; 400} 401 402/** 403 * numa_set_distance - Set NUMA distance from one NUMA to another 404 * @from: the 'from' node to set distance 405 * @to: the 'to' node to set distance 406 * @distance: NUMA distance 407 * 408 * Set the distance from node @from to @to to @distance. If distance table 409 * doesn't exist, one which is large enough to accommodate all the currently 410 * known nodes will be created. 411 * 412 * If such table cannot be allocated, a warning is printed and further 413 * calls are ignored until the distance table is reset with 414 * numa_reset_distance(). 415 * 416 * If @from or @to is higher than the highest known node or lower than zero 417 * at the time of table creation or @distance doesn't make sense, the call 418 * is ignored. 419 * This is to allow simplification of specific NUMA config implementations. 420 */ 421void __init numa_set_distance(int from, int to, int distance) 422{ 423 if (!numa_distance && numa_alloc_distance() < 0) 424 return; 425 426 if (from >= numa_distance_cnt || to >= numa_distance_cnt || 427 from < 0 || to < 0) { 428 pr_warn_once("Warning: node ids are out of bound, from=%d to=%d distance=%d\n", 429 from, to, distance); 430 return; 431 } 432 433 if ((u8)distance != distance || 434 (from == to && distance != LOCAL_DISTANCE)) { 435 pr_warn_once("Warning: invalid distance parameter, from=%d to=%d distance=%d\n", 436 from, to, distance); 437 return; 438 } 439 440 numa_distance[from * numa_distance_cnt + to] = distance; 441} 442 443int __node_distance(int from, int to) 444{ 445 if (from >= numa_distance_cnt || to >= numa_distance_cnt) 446 return from == to ? LOCAL_DISTANCE : REMOTE_DISTANCE; 447 return numa_distance[from * numa_distance_cnt + to]; 448} 449EXPORT_SYMBOL(__node_distance); 450 451/* 452 * Sanity check to catch more bad NUMA configurations (they are amazingly 453 * common). Make sure the nodes cover all memory. 454 */ 455static bool __init numa_meminfo_cover_memory(const struct numa_meminfo *mi) 456{ 457 u64 numaram, e820ram; 458 int i; 459 460 numaram = 0; 461 for (i = 0; i < mi->nr_blks; i++) { 462 u64 s = mi->blk[i].start >> PAGE_SHIFT; 463 u64 e = mi->blk[i].end >> PAGE_SHIFT; 464 numaram += e - s; 465 numaram -= __absent_pages_in_range(mi->blk[i].nid, s, e); 466 if ((s64)numaram < 0) 467 numaram = 0; 468 } 469 470 e820ram = max_pfn - absent_pages_in_range(0, max_pfn); 471 472 /* We seem to lose 3 pages somewhere. Allow 1M of slack. */ 473 if ((s64)(e820ram - numaram) >= (1 << (20 - PAGE_SHIFT))) { 474 printk(KERN_ERR "NUMA: nodes only cover %LuMB of your %LuMB e820 RAM. Not used.\n", 475 (numaram << PAGE_SHIFT) >> 20, 476 (e820ram << PAGE_SHIFT) >> 20); 477 return false; 478 } 479 return true; 480} 481 482/* 483 * Mark all currently memblock-reserved physical memory (which covers the 484 * kernel's own memory ranges) as hot-unswappable. 485 */ 486static void __init numa_clear_kernel_node_hotplug(void) 487{ 488 nodemask_t reserved_nodemask = NODE_MASK_NONE; 489 struct memblock_region *mb_region; 490 int i; 491 492 /* 493 * We have to do some preprocessing of memblock regions, to 494 * make them suitable for reservation. 495 * 496 * At this time, all memory regions reserved by memblock are 497 * used by the kernel, but those regions are not split up 498 * along node boundaries yet, and don't necessarily have their 499 * node ID set yet either. 500 * 501 * So iterate over all memory known to the x86 architecture, 502 * and use those ranges to set the nid in memblock.reserved. 503 * This will split up the memblock regions along node 504 * boundaries and will set the node IDs as well. 505 */ 506 for (i = 0; i < numa_meminfo.nr_blks; i++) { 507 struct numa_memblk *mb = numa_meminfo.blk + i; 508 int ret; 509 510 ret = memblock_set_node(mb->start, mb->end - mb->start, &memblock.reserved, mb->nid); 511 WARN_ON_ONCE(ret); 512 } 513 514 /* 515 * Now go over all reserved memblock regions, to construct a 516 * node mask of all kernel reserved memory areas. 517 * 518 * [ Note, when booting with mem=nn[kMG] or in a kdump kernel, 519 * numa_meminfo might not include all memblock.reserved 520 * memory ranges, because quirks such as trim_snb_memory() 521 * reserve specific pages for Sandy Bridge graphics. ] 522 */ 523 for_each_reserved_mem_region(mb_region) { 524 int nid = memblock_get_region_node(mb_region); 525 526 if (nid != MAX_NUMNODES) 527 node_set(nid, reserved_nodemask); 528 } 529 530 /* 531 * Finally, clear the MEMBLOCK_HOTPLUG flag for all memory 532 * belonging to the reserved node mask. 533 * 534 * Note that this will include memory regions that reside 535 * on nodes that contain kernel memory - entire nodes 536 * become hot-unpluggable: 537 */ 538 for (i = 0; i < numa_meminfo.nr_blks; i++) { 539 struct numa_memblk *mb = numa_meminfo.blk + i; 540 541 if (!node_isset(mb->nid, reserved_nodemask)) 542 continue; 543 544 memblock_clear_hotplug(mb->start, mb->end - mb->start); 545 } 546} 547 548static int __init numa_register_memblks(struct numa_meminfo *mi) 549{ 550 int i, nid; 551 552 /* Account for nodes with cpus and no memory */ 553 node_possible_map = numa_nodes_parsed; 554 numa_nodemask_from_meminfo(&node_possible_map, mi); 555 if (WARN_ON(nodes_empty(node_possible_map))) 556 return -EINVAL; 557 558 for (i = 0; i < mi->nr_blks; i++) { 559 struct numa_memblk *mb = &mi->blk[i]; 560 memblock_set_node(mb->start, mb->end - mb->start, 561 &memblock.memory, mb->nid); 562 } 563 564 /* 565 * At very early time, the kernel have to use some memory such as 566 * loading the kernel image. We cannot prevent this anyway. So any 567 * node the kernel resides in should be un-hotpluggable. 568 * 569 * And when we come here, alloc node data won't fail. 570 */ 571 numa_clear_kernel_node_hotplug(); 572 573 /* 574 * If sections array is gonna be used for pfn -> nid mapping, check 575 * whether its granularity is fine enough. 576 */ 577 if (IS_ENABLED(NODE_NOT_IN_PAGE_FLAGS)) { 578 unsigned long pfn_align = node_map_pfn_alignment(); 579 580 if (pfn_align && pfn_align < PAGES_PER_SECTION) { 581 pr_warn("Node alignment %LuMB < min %LuMB, rejecting NUMA config\n", 582 PFN_PHYS(pfn_align) >> 20, 583 PFN_PHYS(PAGES_PER_SECTION) >> 20); 584 return -EINVAL; 585 } 586 } 587 if (!numa_meminfo_cover_memory(mi)) 588 return -EINVAL; 589 590 /* Finally register nodes. */ 591 for_each_node_mask(nid, node_possible_map) { 592 u64 start = PFN_PHYS(max_pfn); 593 u64 end = 0; 594 595 for (i = 0; i < mi->nr_blks; i++) { 596 if (nid != mi->blk[i].nid) 597 continue; 598 start = min(mi->blk[i].start, start); 599 end = max(mi->blk[i].end, end); 600 } 601 602 if (start >= end) 603 continue; 604 605 alloc_node_data(nid); 606 } 607 608 /* Dump memblock with node info and return. */ 609 memblock_dump_all(); 610 return 0; 611} 612 613/* 614 * There are unfortunately some poorly designed mainboards around that 615 * only connect memory to a single CPU. This breaks the 1:1 cpu->node 616 * mapping. To avoid this fill in the mapping for all possible CPUs, 617 * as the number of CPUs is not known yet. We round robin the existing 618 * nodes. 619 */ 620static void __init numa_init_array(void) 621{ 622 int rr, i; 623 624 rr = first_node(node_online_map); 625 for (i = 0; i < nr_cpu_ids; i++) { 626 if (early_cpu_to_node(i) != NUMA_NO_NODE) 627 continue; 628 numa_set_node(i, rr); 629 rr = next_node_in(rr, node_online_map); 630 } 631} 632 633static int __init numa_init(int (*init_func)(void)) 634{ 635 int i; 636 int ret; 637 638 for (i = 0; i < MAX_LOCAL_APIC; i++) 639 set_apicid_to_node(i, NUMA_NO_NODE); 640 641 nodes_clear(numa_nodes_parsed); 642 nodes_clear(node_possible_map); 643 nodes_clear(node_online_map); 644 memset(&numa_meminfo, 0, sizeof(numa_meminfo)); 645 WARN_ON(memblock_set_node(0, ULLONG_MAX, &memblock.memory, 646 MAX_NUMNODES)); 647 WARN_ON(memblock_set_node(0, ULLONG_MAX, &memblock.reserved, 648 MAX_NUMNODES)); 649 /* In case that parsing SRAT failed. */ 650 WARN_ON(memblock_clear_hotplug(0, ULLONG_MAX)); 651 numa_reset_distance(); 652 653 ret = init_func(); 654 if (ret < 0) 655 return ret; 656 657 /* 658 * We reset memblock back to the top-down direction 659 * here because if we configured ACPI_NUMA, we have 660 * parsed SRAT in init_func(). It is ok to have the 661 * reset here even if we did't configure ACPI_NUMA 662 * or acpi numa init fails and fallbacks to dummy 663 * numa init. 664 */ 665 memblock_set_bottom_up(false); 666 667 ret = numa_cleanup_meminfo(&numa_meminfo); 668 if (ret < 0) 669 return ret; 670 671 numa_emulation(&numa_meminfo, numa_distance_cnt); 672 673 ret = numa_register_memblks(&numa_meminfo); 674 if (ret < 0) 675 return ret; 676 677 for (i = 0; i < nr_cpu_ids; i++) { 678 int nid = early_cpu_to_node(i); 679 680 if (nid == NUMA_NO_NODE) 681 continue; 682 if (!node_online(nid)) 683 numa_clear_node(i); 684 } 685 numa_init_array(); 686 687 return 0; 688} 689 690/** 691 * dummy_numa_init - Fallback dummy NUMA init 692 * 693 * Used if there's no underlying NUMA architecture, NUMA initialization 694 * fails, or NUMA is disabled on the command line. 695 * 696 * Must online at least one node and add memory blocks that cover all 697 * allowed memory. This function must not fail. 698 */ 699static int __init dummy_numa_init(void) 700{ 701 printk(KERN_INFO "%s\n", 702 numa_off ? "NUMA turned off" : "No NUMA configuration found"); 703 printk(KERN_INFO "Faking a node at [mem %#018Lx-%#018Lx]\n", 704 0LLU, PFN_PHYS(max_pfn) - 1); 705 706 node_set(0, numa_nodes_parsed); 707 numa_add_memblk(0, 0, PFN_PHYS(max_pfn)); 708 709 return 0; 710} 711 712/** 713 * x86_numa_init - Initialize NUMA 714 * 715 * Try each configured NUMA initialization method until one succeeds. The 716 * last fallback is dummy single node config encompassing whole memory and 717 * never fails. 718 */ 719void __init x86_numa_init(void) 720{ 721 if (!numa_off) { 722#ifdef CONFIG_ACPI_NUMA 723 if (!numa_init(x86_acpi_numa_init)) 724 return; 725#endif 726#ifdef CONFIG_AMD_NUMA 727 if (!numa_init(amd_numa_init)) 728 return; 729#endif 730 } 731 732 numa_init(dummy_numa_init); 733} 734 735 736/* 737 * A node may exist which has one or more Generic Initiators but no CPUs and no 738 * memory. 739 * 740 * This function must be called after init_cpu_to_node(), to ensure that any 741 * memoryless CPU nodes have already been brought online, and before the 742 * node_data[nid] is needed for zone list setup in build_all_zonelists(). 743 * 744 * When this function is called, any nodes containing either memory and/or CPUs 745 * will already be online and there is no need to do anything extra, even if 746 * they also contain one or more Generic Initiators. 747 */ 748void __init init_gi_nodes(void) 749{ 750 int nid; 751 752 /* 753 * Exclude this node from 754 * bringup_nonboot_cpus 755 * cpu_up 756 * __try_online_node 757 * register_one_node 758 * because node_subsys is not initialized yet. 759 * TODO remove dependency on node_online 760 */ 761 for_each_node_state(nid, N_GENERIC_INITIATOR) 762 if (!node_online(nid)) 763 node_set_online(nid); 764} 765 766/* 767 * Setup early cpu_to_node. 768 * 769 * Populate cpu_to_node[] only if x86_cpu_to_apicid[], 770 * and apicid_to_node[] tables have valid entries for a CPU. 771 * This means we skip cpu_to_node[] initialisation for NUMA 772 * emulation and faking node case (when running a kernel compiled 773 * for NUMA on a non NUMA box), which is OK as cpu_to_node[] 774 * is already initialized in a round robin manner at numa_init_array, 775 * prior to this call, and this initialization is good enough 776 * for the fake NUMA cases. 777 * 778 * Called before the per_cpu areas are setup. 779 */ 780void __init init_cpu_to_node(void) 781{ 782 int cpu; 783 u16 *cpu_to_apicid = early_per_cpu_ptr(x86_cpu_to_apicid); 784 785 BUG_ON(cpu_to_apicid == NULL); 786 787 for_each_possible_cpu(cpu) { 788 int node = numa_cpu_node(cpu); 789 790 if (node == NUMA_NO_NODE) 791 continue; 792 793 /* 794 * Exclude this node from 795 * bringup_nonboot_cpus 796 * cpu_up 797 * __try_online_node 798 * register_one_node 799 * because node_subsys is not initialized yet. 800 * TODO remove dependency on node_online 801 */ 802 if (!node_online(node)) 803 node_set_online(node); 804 805 numa_set_node(cpu, node); 806 } 807} 808 809#ifndef CONFIG_DEBUG_PER_CPU_MAPS 810 811# ifndef CONFIG_NUMA_EMU 812void numa_add_cpu(int cpu) 813{ 814 cpumask_set_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]); 815} 816 817void numa_remove_cpu(int cpu) 818{ 819 cpumask_clear_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]); 820} 821# endif /* !CONFIG_NUMA_EMU */ 822 823#else /* !CONFIG_DEBUG_PER_CPU_MAPS */ 824 825int __cpu_to_node(int cpu) 826{ 827 if (early_per_cpu_ptr(x86_cpu_to_node_map)) { 828 printk(KERN_WARNING 829 "cpu_to_node(%d): usage too early!\n", cpu); 830 dump_stack(); 831 return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu]; 832 } 833 return per_cpu(x86_cpu_to_node_map, cpu); 834} 835EXPORT_SYMBOL(__cpu_to_node); 836 837/* 838 * Same function as cpu_to_node() but used if called before the 839 * per_cpu areas are setup. 840 */ 841int early_cpu_to_node(int cpu) 842{ 843 if (early_per_cpu_ptr(x86_cpu_to_node_map)) 844 return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu]; 845 846 if (!cpu_possible(cpu)) { 847 printk(KERN_WARNING 848 "early_cpu_to_node(%d): no per_cpu area!\n", cpu); 849 dump_stack(); 850 return NUMA_NO_NODE; 851 } 852 return per_cpu(x86_cpu_to_node_map, cpu); 853} 854 855void debug_cpumask_set_cpu(int cpu, int node, bool enable) 856{ 857 struct cpumask *mask; 858 859 if (node == NUMA_NO_NODE) { 860 /* early_cpu_to_node() already emits a warning and trace */ 861 return; 862 } 863 mask = node_to_cpumask_map[node]; 864 if (!cpumask_available(mask)) { 865 pr_err("node_to_cpumask_map[%i] NULL\n", node); 866 dump_stack(); 867 return; 868 } 869 870 if (enable) 871 cpumask_set_cpu(cpu, mask); 872 else 873 cpumask_clear_cpu(cpu, mask); 874 875 printk(KERN_DEBUG "%s cpu %d node %d: mask now %*pbl\n", 876 enable ? "numa_add_cpu" : "numa_remove_cpu", 877 cpu, node, cpumask_pr_args(mask)); 878 return; 879} 880 881# ifndef CONFIG_NUMA_EMU 882static void numa_set_cpumask(int cpu, bool enable) 883{ 884 debug_cpumask_set_cpu(cpu, early_cpu_to_node(cpu), enable); 885} 886 887void numa_add_cpu(int cpu) 888{ 889 numa_set_cpumask(cpu, true); 890} 891 892void numa_remove_cpu(int cpu) 893{ 894 numa_set_cpumask(cpu, false); 895} 896# endif /* !CONFIG_NUMA_EMU */ 897 898/* 899 * Returns a pointer to the bitmask of CPUs on Node 'node'. 900 */ 901const struct cpumask *cpumask_of_node(int node) 902{ 903 if ((unsigned)node >= nr_node_ids) { 904 printk(KERN_WARNING 905 "cpumask_of_node(%d): (unsigned)node >= nr_node_ids(%u)\n", 906 node, nr_node_ids); 907 dump_stack(); 908 return cpu_none_mask; 909 } 910 if (!cpumask_available(node_to_cpumask_map[node])) { 911 printk(KERN_WARNING 912 "cpumask_of_node(%d): no node_to_cpumask_map!\n", 913 node); 914 dump_stack(); 915 return cpu_online_mask; 916 } 917 return node_to_cpumask_map[node]; 918} 919EXPORT_SYMBOL(cpumask_of_node); 920 921#endif /* !CONFIG_DEBUG_PER_CPU_MAPS */ 922 923#ifdef CONFIG_NUMA_KEEP_MEMINFO 924static int meminfo_to_nid(struct numa_meminfo *mi, u64 start) 925{ 926 int i; 927 928 for (i = 0; i < mi->nr_blks; i++) 929 if (mi->blk[i].start <= start && mi->blk[i].end > start) 930 return mi->blk[i].nid; 931 return NUMA_NO_NODE; 932} 933 934int phys_to_target_node(phys_addr_t start) 935{ 936 int nid = meminfo_to_nid(&numa_meminfo, start); 937 938 /* 939 * Prefer online nodes, but if reserved memory might be 940 * hot-added continue the search with reserved ranges. 941 */ 942 if (nid != NUMA_NO_NODE) 943 return nid; 944 945 return meminfo_to_nid(&numa_reserved_meminfo, start); 946} 947EXPORT_SYMBOL_GPL(phys_to_target_node); 948 949int memory_add_physaddr_to_nid(u64 start) 950{ 951 int nid = meminfo_to_nid(&numa_meminfo, start); 952 953 if (nid == NUMA_NO_NODE) 954 nid = numa_meminfo.blk[0].nid; 955 return nid; 956} 957EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid); 958 959static int __init cmp_memblk(const void *a, const void *b) 960{ 961 const struct numa_memblk *ma = *(const struct numa_memblk **)a; 962 const struct numa_memblk *mb = *(const struct numa_memblk **)b; 963 964 return (ma->start > mb->start) - (ma->start < mb->start); 965} 966 967static struct numa_memblk *numa_memblk_list[NR_NODE_MEMBLKS] __initdata; 968 969/** 970 * numa_fill_memblks - Fill gaps in numa_meminfo memblks 971 * @start: address to begin fill 972 * @end: address to end fill 973 * 974 * Find and extend numa_meminfo memblks to cover the physical 975 * address range @start-@end 976 * 977 * RETURNS: 978 * 0 : Success 979 * NUMA_NO_MEMBLK : No memblks exist in address range @start-@end 980 */ 981 982int __init numa_fill_memblks(u64 start, u64 end) 983{ 984 struct numa_memblk **blk = &numa_memblk_list[0]; 985 struct numa_meminfo *mi = &numa_meminfo; 986 int count = 0; 987 u64 prev_end; 988 989 /* 990 * Create a list of pointers to numa_meminfo memblks that 991 * overlap start, end. The list is used to make in-place 992 * changes that fill out the numa_meminfo memblks. 993 */ 994 for (int i = 0; i < mi->nr_blks; i++) { 995 struct numa_memblk *bi = &mi->blk[i]; 996 997 if (memblock_addrs_overlap(start, end - start, bi->start, 998 bi->end - bi->start)) { 999 blk[count] = &mi->blk[i]; 1000 count++; 1001 } 1002 } 1003 if (!count) 1004 return NUMA_NO_MEMBLK; 1005 1006 /* Sort the list of pointers in memblk->start order */ 1007 sort(&blk[0], count, sizeof(blk[0]), cmp_memblk, NULL); 1008 1009 /* Make sure the first/last memblks include start/end */ 1010 blk[0]->start = min(blk[0]->start, start); 1011 blk[count - 1]->end = max(blk[count - 1]->end, end); 1012 1013 /* 1014 * Fill any gaps by tracking the previous memblks 1015 * end address and backfilling to it if needed. 1016 */ 1017 prev_end = blk[0]->end; 1018 for (int i = 1; i < count; i++) { 1019 struct numa_memblk *curr = blk[i]; 1020 1021 if (prev_end >= curr->start) { 1022 if (prev_end < curr->end) 1023 prev_end = curr->end; 1024 } else { 1025 curr->start = prev_end; 1026 prev_end = curr->end; 1027 } 1028 } 1029 return 0; 1030} 1031 1032#endif 1033