1/* 2 * ACPI 3.0 based NUMA setup 3 * Copyright 2004 Andi Kleen, SuSE Labs. 4 * 5 * Reads the ACPI SRAT table to figure out what memory belongs to which CPUs. 6 * 7 * Called from acpi_numa_init while reading the SRAT and SLIT tables. 8 * Assumes all memory regions belonging to a single proximity domain 9 * are in one chunk. Holes between them will be included in the node. 10 */ 11 12#include <linux/kernel.h> 13#include <linux/acpi.h> 14#include <linux/mmzone.h> 15#include <linux/bitmap.h> 16#include <linux/module.h> 17#include <linux/topology.h> 18#include <linux/bootmem.h> 19#include <linux/mm.h> 20#include <asm/proto.h> 21#include <asm/numa.h> 22#include <asm/e820.h> 23#include <asm/apic.h> 24#include <asm/uv/uv.h> 25 26int acpi_numa __initdata; 27 28static struct acpi_table_slit *acpi_slit; 29 30static nodemask_t nodes_parsed __initdata; 31static nodemask_t cpu_nodes_parsed __initdata; 32static struct bootnode nodes[MAX_NUMNODES] __initdata; 33static struct bootnode nodes_add[MAX_NUMNODES]; 34 35static int num_node_memblks __initdata; 36static struct bootnode node_memblk_range[NR_NODE_MEMBLKS] __initdata; 37static int memblk_nodeid[NR_NODE_MEMBLKS] __initdata; 38 39static __init int setup_node(int pxm) 40{ 41 return acpi_map_pxm_to_node(pxm); 42} 43 44static __init int conflicting_memblks(unsigned long start, unsigned long end) 45{ 46 int i; 47 for (i = 0; i < num_node_memblks; i++) { 48 struct bootnode *nd = &node_memblk_range[i]; 49 if (nd->start == nd->end) 50 continue; 51 if (nd->end > start && nd->start < end) 52 return memblk_nodeid[i]; 53 if (nd->end == end && nd->start == start) 54 return memblk_nodeid[i]; 55 } 56 return -1; 57} 58 59static __init void cutoff_node(int i, unsigned long start, unsigned long end) 60{ 61 struct bootnode *nd = &nodes[i]; 62 63 if (nd->start < start) { 64 nd->start = start; 65 if (nd->end < nd->start) 66 nd->start = nd->end; 67 } 68 if (nd->end > end) { 69 nd->end = end; 70 if (nd->start > nd->end) 71 nd->start = nd->end; 72 } 73} 74 75static __init void bad_srat(void) 76{ 77 int i; 78 printk(KERN_ERR "SRAT: SRAT not used.\n"); 79 acpi_numa = -1; 80 for (i = 0; i < MAX_LOCAL_APIC; i++) 81 apicid_to_node[i] = NUMA_NO_NODE; 82 for (i = 0; i < MAX_NUMNODES; i++) { 83 nodes[i].start = nodes[i].end = 0; 84 nodes_add[i].start = nodes_add[i].end = 0; 85 } 86 remove_all_active_ranges(); 87} 88 89static __init inline int srat_disabled(void) 90{ 91 return numa_off || acpi_numa < 0; 92} 93 94/* Callback for SLIT parsing */ 95void __init acpi_numa_slit_init(struct acpi_table_slit *slit) 96{ 97 unsigned length; 98 unsigned long phys; 99 100 length = slit->header.length; 101 phys = find_e820_area(0, max_pfn_mapped<<PAGE_SHIFT, length, 102 PAGE_SIZE); 103 104 if (phys == -1L) 105 panic(" Can not save slit!\n"); 106 107 acpi_slit = __va(phys); 108 memcpy(acpi_slit, slit, length); 109 reserve_early(phys, phys + length, "ACPI SLIT"); 110} 111 112/* Callback for Proximity Domain -> x2APIC mapping */ 113void __init 114acpi_numa_x2apic_affinity_init(struct acpi_srat_x2apic_cpu_affinity *pa) 115{ 116 int pxm, node; 117 int apic_id; 118 119 if (srat_disabled()) 120 return; 121 if (pa->header.length < sizeof(struct acpi_srat_x2apic_cpu_affinity)) { 122 bad_srat(); 123 return; 124 } 125 if ((pa->flags & ACPI_SRAT_CPU_ENABLED) == 0) 126 return; 127 pxm = pa->proximity_domain; 128 node = setup_node(pxm); 129 if (node < 0) { 130 printk(KERN_ERR "SRAT: Too many proximity domains %x\n", pxm); 131 bad_srat(); 132 return; 133 } 134 135 apic_id = pa->apic_id; 136 apicid_to_node[apic_id] = node; 137 node_set(node, cpu_nodes_parsed); 138 acpi_numa = 1; 139 printk(KERN_INFO "SRAT: PXM %u -> APIC 0x%04x -> Node %u\n", 140 pxm, apic_id, node); 141} 142 143/* Callback for Proximity Domain -> LAPIC mapping */ 144void __init 145acpi_numa_processor_affinity_init(struct acpi_srat_cpu_affinity *pa) 146{ 147 int pxm, node; 148 int apic_id; 149 150 if (srat_disabled()) 151 return; 152 if (pa->header.length != sizeof(struct acpi_srat_cpu_affinity)) { 153 bad_srat(); 154 return; 155 } 156 if ((pa->flags & ACPI_SRAT_CPU_ENABLED) == 0) 157 return; 158 pxm = pa->proximity_domain_lo; 159 node = setup_node(pxm); 160 if (node < 0) { 161 printk(KERN_ERR "SRAT: Too many proximity domains %x\n", pxm); 162 bad_srat(); 163 return; 164 } 165 166 if (get_uv_system_type() >= UV_X2APIC) 167 apic_id = (pa->apic_id << 8) | pa->local_sapic_eid; 168 else 169 apic_id = pa->apic_id; 170 apicid_to_node[apic_id] = node; 171 node_set(node, cpu_nodes_parsed); 172 acpi_numa = 1; 173 printk(KERN_INFO "SRAT: PXM %u -> APIC 0x%02x -> Node %u\n", 174 pxm, apic_id, node); 175} 176 177#ifdef CONFIG_MEMORY_HOTPLUG_SPARSE 178static inline int save_add_info(void) {return 1;} 179#else 180static inline int save_add_info(void) {return 0;} 181#endif 182/* 183 * Update nodes_add[] 184 * This code supports one contiguous hot add area per node 185 */ 186static void __init 187update_nodes_add(int node, unsigned long start, unsigned long end) 188{ 189 unsigned long s_pfn = start >> PAGE_SHIFT; 190 unsigned long e_pfn = end >> PAGE_SHIFT; 191 int changed = 0; 192 struct bootnode *nd = &nodes_add[node]; 193 194 /* I had some trouble with strange memory hotadd regions breaking 195 the boot. Be very strict here and reject anything unexpected. 196 If you want working memory hotadd write correct SRATs. 197 198 The node size check is a basic sanity check to guard against 199 mistakes */ 200 if ((signed long)(end - start) < NODE_MIN_SIZE) { 201 printk(KERN_ERR "SRAT: Hotplug area too small\n"); 202 return; 203 } 204 205 /* This check might be a bit too strict, but I'm keeping it for now. */ 206 if (absent_pages_in_range(s_pfn, e_pfn) != e_pfn - s_pfn) { 207 printk(KERN_ERR 208 "SRAT: Hotplug area %lu -> %lu has existing memory\n", 209 s_pfn, e_pfn); 210 return; 211 } 212 213 /* Looks good */ 214 215 if (nd->start == nd->end) { 216 nd->start = start; 217 nd->end = end; 218 changed = 1; 219 } else { 220 if (nd->start == end) { 221 nd->start = start; 222 changed = 1; 223 } 224 if (nd->end == start) { 225 nd->end = end; 226 changed = 1; 227 } 228 if (!changed) 229 printk(KERN_ERR "SRAT: Hotplug zone not continuous. Partly ignored\n"); 230 } 231 232 if (changed) { 233 node_set(node, cpu_nodes_parsed); 234 printk(KERN_INFO "SRAT: hot plug zone found %Lx - %Lx\n", 235 nd->start, nd->end); 236 } 237} 238 239/* Callback for parsing of the Proximity Domain <-> Memory Area mappings */ 240void __init 241acpi_numa_memory_affinity_init(struct acpi_srat_mem_affinity *ma) 242{ 243 struct bootnode *nd, oldnode; 244 unsigned long start, end; 245 int node, pxm; 246 int i; 247 248 if (srat_disabled()) 249 return; 250 if (ma->header.length != sizeof(struct acpi_srat_mem_affinity)) { 251 bad_srat(); 252 return; 253 } 254 if ((ma->flags & ACPI_SRAT_MEM_ENABLED) == 0) 255 return; 256 257 if ((ma->flags & ACPI_SRAT_MEM_HOT_PLUGGABLE) && !save_add_info()) 258 return; 259 start = ma->base_address; 260 end = start + ma->length; 261 pxm = ma->proximity_domain; 262 node = setup_node(pxm); 263 if (node < 0) { 264 printk(KERN_ERR "SRAT: Too many proximity domains.\n"); 265 bad_srat(); 266 return; 267 } 268 i = conflicting_memblks(start, end); 269 if (i == node) { 270 printk(KERN_WARNING 271 "SRAT: Warning: PXM %d (%lx-%lx) overlaps with itself (%Lx-%Lx)\n", 272 pxm, start, end, nodes[i].start, nodes[i].end); 273 } else if (i >= 0) { 274 printk(KERN_ERR 275 "SRAT: PXM %d (%lx-%lx) overlaps with PXM %d (%Lx-%Lx)\n", 276 pxm, start, end, node_to_pxm(i), 277 nodes[i].start, nodes[i].end); 278 bad_srat(); 279 return; 280 } 281 nd = &nodes[node]; 282 oldnode = *nd; 283 if (!node_test_and_set(node, nodes_parsed)) { 284 nd->start = start; 285 nd->end = end; 286 } else { 287 if (start < nd->start) 288 nd->start = start; 289 if (nd->end < end) 290 nd->end = end; 291 } 292 293 printk(KERN_INFO "SRAT: Node %u PXM %u %lx-%lx\n", node, pxm, 294 start, end); 295 296 if (ma->flags & ACPI_SRAT_MEM_HOT_PLUGGABLE) { 297 update_nodes_add(node, start, end); 298 /* restore nodes[node] */ 299 *nd = oldnode; 300 if ((nd->start | nd->end) == 0) 301 node_clear(node, nodes_parsed); 302 } 303 304 node_memblk_range[num_node_memblks].start = start; 305 node_memblk_range[num_node_memblks].end = end; 306 memblk_nodeid[num_node_memblks] = node; 307 num_node_memblks++; 308} 309 310/* Sanity check to catch more bad SRATs (they are amazingly common). 311 Make sure the PXMs cover all memory. */ 312static int __init nodes_cover_memory(const struct bootnode *nodes) 313{ 314 int i; 315 unsigned long pxmram, e820ram; 316 317 pxmram = 0; 318 for_each_node_mask(i, nodes_parsed) { 319 unsigned long s = nodes[i].start >> PAGE_SHIFT; 320 unsigned long e = nodes[i].end >> PAGE_SHIFT; 321 pxmram += e - s; 322 pxmram -= __absent_pages_in_range(i, s, e); 323 if ((long)pxmram < 0) 324 pxmram = 0; 325 } 326 327 e820ram = max_pfn - (e820_hole_size(0, max_pfn<<PAGE_SHIFT)>>PAGE_SHIFT); 328 /* We seem to lose 3 pages somewhere. Allow 1M of slack. */ 329 if ((long)(e820ram - pxmram) >= (1<<(20 - PAGE_SHIFT))) { 330 printk(KERN_ERR 331 "SRAT: PXMs only cover %luMB of your %luMB e820 RAM. Not used.\n", 332 (pxmram << PAGE_SHIFT) >> 20, 333 (e820ram << PAGE_SHIFT) >> 20); 334 return 0; 335 } 336 return 1; 337} 338 339void __init acpi_numa_arch_fixup(void) {} 340 341int __init acpi_get_nodes(struct bootnode *physnodes) 342{ 343 int i; 344 int ret = 0; 345 346 for_each_node_mask(i, nodes_parsed) { 347 physnodes[ret].start = nodes[i].start; 348 physnodes[ret].end = nodes[i].end; 349 ret++; 350 } 351 return ret; 352} 353 354/* Use the information discovered above to actually set up the nodes. */ 355int __init acpi_scan_nodes(unsigned long start, unsigned long end) 356{ 357 int i; 358 359 if (acpi_numa <= 0) 360 return -1; 361 362 /* First clean up the node list */ 363 for (i = 0; i < MAX_NUMNODES; i++) 364 cutoff_node(i, start, end); 365 366 /* 367 * Join together blocks on the same node, holes between 368 * which don't overlap with memory on other nodes. 369 */ 370 for (i = 0; i < num_node_memblks; ++i) { 371 int j, k; 372 373 for (j = i + 1; j < num_node_memblks; ++j) { 374 unsigned long start, end; 375 376 if (memblk_nodeid[i] != memblk_nodeid[j]) 377 continue; 378 start = min(node_memblk_range[i].end, 379 node_memblk_range[j].end); 380 end = max(node_memblk_range[i].start, 381 node_memblk_range[j].start); 382 for (k = 0; k < num_node_memblks; ++k) { 383 if (memblk_nodeid[i] == memblk_nodeid[k]) 384 continue; 385 if (start < node_memblk_range[k].end && 386 end > node_memblk_range[k].start) 387 break; 388 } 389 if (k < num_node_memblks) 390 continue; 391 start = min(node_memblk_range[i].start, 392 node_memblk_range[j].start); 393 end = max(node_memblk_range[i].end, 394 node_memblk_range[j].end); 395 printk(KERN_INFO "SRAT: Node %d " 396 "[%Lx,%Lx) + [%Lx,%Lx) -> [%lx,%lx)\n", 397 memblk_nodeid[i], 398 node_memblk_range[i].start, 399 node_memblk_range[i].end, 400 node_memblk_range[j].start, 401 node_memblk_range[j].end, 402 start, end); 403 node_memblk_range[i].start = start; 404 node_memblk_range[i].end = end; 405 k = --num_node_memblks - j; 406 memmove(memblk_nodeid + j, memblk_nodeid + j+1, 407 k * sizeof(*memblk_nodeid)); 408 memmove(node_memblk_range + j, node_memblk_range + j+1, 409 k * sizeof(*node_memblk_range)); 410 --j; 411 } 412 } 413 414 memnode_shift = compute_hash_shift(node_memblk_range, num_node_memblks, 415 memblk_nodeid); 416 if (memnode_shift < 0) { 417 printk(KERN_ERR 418 "SRAT: No NUMA node hash function found. Contact maintainer\n"); 419 bad_srat(); 420 return -1; 421 } 422 423 for (i = 0; i < num_node_memblks; i++) 424 e820_register_active_regions(memblk_nodeid[i], 425 node_memblk_range[i].start >> PAGE_SHIFT, 426 node_memblk_range[i].end >> PAGE_SHIFT); 427 428 /* for out of order entries in SRAT */ 429 sort_node_map(); 430 if (!nodes_cover_memory(nodes)) { 431 bad_srat(); 432 return -1; 433 } 434 435 /* Account for nodes with cpus and no memory */ 436 nodes_or(node_possible_map, nodes_parsed, cpu_nodes_parsed); 437 438 /* Finally register nodes */ 439 for_each_node_mask(i, node_possible_map) 440 setup_node_bootmem(i, nodes[i].start, nodes[i].end); 441 /* Try again in case setup_node_bootmem missed one due 442 to missing bootmem */ 443 for_each_node_mask(i, node_possible_map) 444 if (!node_online(i)) 445 setup_node_bootmem(i, nodes[i].start, nodes[i].end); 446 447 for (i = 0; i < nr_cpu_ids; i++) { 448 int node = early_cpu_to_node(i); 449 450 if (node == NUMA_NO_NODE) 451 continue; 452 if (!node_online(node)) 453 numa_clear_node(i); 454 } 455 numa_init_array(); 456 return 0; 457} 458 459#ifdef CONFIG_NUMA_EMU 460static int fake_node_to_pxm_map[MAX_NUMNODES] __initdata = { 461 [0 ... MAX_NUMNODES-1] = PXM_INVAL 462}; 463static s16 fake_apicid_to_node[MAX_LOCAL_APIC] __initdata = { 464 [0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE 465}; 466static int __init find_node_by_addr(unsigned long addr) 467{ 468 int ret = NUMA_NO_NODE; 469 int i; 470 471 for_each_node_mask(i, nodes_parsed) { 472 /* 473 * Find the real node that this emulated node appears on. For 474 * the sake of simplicity, we only use a real node's starting 475 * address to determine which emulated node it appears on. 476 */ 477 if (addr >= nodes[i].start && addr < nodes[i].end) { 478 ret = i; 479 break; 480 } 481 } 482 return ret; 483} 484 485/* 486 * In NUMA emulation, we need to setup proximity domain (_PXM) to node ID 487 * mappings that respect the real ACPI topology but reflect our emulated 488 * environment. For each emulated node, we find which real node it appears on 489 * and create PXM to NID mappings for those fake nodes which mirror that 490 * locality. SLIT will now represent the correct distances between emulated 491 * nodes as a result of the real topology. 492 */ 493void __init acpi_fake_nodes(const struct bootnode *fake_nodes, int num_nodes) 494{ 495 int i, j; 496 497 printk(KERN_INFO "Faking PXM affinity for fake nodes on real " 498 "topology.\n"); 499 for (i = 0; i < num_nodes; i++) { 500 int nid, pxm; 501 502 nid = find_node_by_addr(fake_nodes[i].start); 503 if (nid == NUMA_NO_NODE) 504 continue; 505 pxm = node_to_pxm(nid); 506 if (pxm == PXM_INVAL) 507 continue; 508 fake_node_to_pxm_map[i] = pxm; 509 /* 510 * For each apicid_to_node mapping that exists for this real 511 * node, it must now point to the fake node ID. 512 */ 513 for (j = 0; j < MAX_LOCAL_APIC; j++) 514 if (apicid_to_node[j] == nid && 515 fake_apicid_to_node[j] == NUMA_NO_NODE) 516 fake_apicid_to_node[j] = i; 517 } 518 for (i = 0; i < num_nodes; i++) 519 __acpi_map_pxm_to_node(fake_node_to_pxm_map[i], i); 520 memcpy(apicid_to_node, fake_apicid_to_node, sizeof(apicid_to_node)); 521 522 nodes_clear(nodes_parsed); 523 for (i = 0; i < num_nodes; i++) 524 if (fake_nodes[i].start != fake_nodes[i].end) 525 node_set(i, nodes_parsed); 526} 527 528static int null_slit_node_compare(int a, int b) 529{ 530 return node_to_pxm(a) == node_to_pxm(b); 531} 532#else 533static int null_slit_node_compare(int a, int b) 534{ 535 return a == b; 536} 537#endif /* CONFIG_NUMA_EMU */ 538 539int __node_distance(int a, int b) 540{ 541 int index; 542 543 if (!acpi_slit) 544 return null_slit_node_compare(a, b) ? LOCAL_DISTANCE : 545 REMOTE_DISTANCE; 546 index = acpi_slit->locality_count * node_to_pxm(a); 547 return acpi_slit->entry[index + node_to_pxm(b)]; 548} 549 550EXPORT_SYMBOL(__node_distance); 551 552#if defined(CONFIG_MEMORY_HOTPLUG_SPARSE) || defined(CONFIG_ACPI_HOTPLUG_MEMORY) 553int memory_add_physaddr_to_nid(u64 start) 554{ 555 int i, ret = 0; 556 557 for_each_node(i) 558 if (nodes_add[i].start <= start && nodes_add[i].end > start) 559 ret = i; 560 561 return ret; 562} 563EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid); 564#endif 565