| 1 | /* |
| 2 | * Generic VM initialization for x86-64 NUMA setups. |
| 3 | * Copyright 2002,2003 Andi Kleen, SuSE Labs. |
| 4 | */ |
| 5 | #include <linux/kernel.h> |
| 6 | #include <linux/mm.h> |
| 7 | #include <linux/string.h> |
| 8 | #include <linux/init.h> |
| 9 | #include <linux/bootmem.h> |
| 10 | #include <linux/mmzone.h> |
| 11 | #include <linux/ctype.h> |
| 12 | #include <linux/module.h> |
| 13 | #include <linux/nodemask.h> |
| 14 | #include <linux/sched.h> |
| 15 | |
| 16 | #include <asm/e820.h> |
| 17 | #include <asm/proto.h> |
| 18 | #include <asm/dma.h> |
| 19 | #include <asm/numa.h> |
| 20 | #include <asm/acpi.h> |
| 21 | #include <asm/k8.h> |
| 22 | |
| 23 | struct pglist_data *node_data[MAX_NUMNODES] __read_mostly; |
| 24 | EXPORT_SYMBOL(node_data); |
| 25 | |
| 26 | struct memnode memnode; |
| 27 | |
| 28 | s16 apicid_to_node[MAX_LOCAL_APIC] __cpuinitdata = { |
| 29 | [0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE |
| 30 | }; |
| 31 | |
| 32 | int numa_off __initdata; |
| 33 | static unsigned long __initdata nodemap_addr; |
| 34 | static unsigned long __initdata nodemap_size; |
| 35 | |
| 36 | DEFINE_PER_CPU(int, node_number) = 0; |
| 37 | EXPORT_PER_CPU_SYMBOL(node_number); |
| 38 | |
| 39 | /* |
| 40 | * Map cpu index to node index |
| 41 | */ |
| 42 | DEFINE_EARLY_PER_CPU(int, x86_cpu_to_node_map, NUMA_NO_NODE); |
| 43 | EXPORT_EARLY_PER_CPU_SYMBOL(x86_cpu_to_node_map); |
| 44 | |
| 45 | /* |
| 46 | * Given a shift value, try to populate memnodemap[] |
| 47 | * Returns : |
| 48 | * 1 if OK |
| 49 | * 0 if memnodmap[] too small (of shift too small) |
| 50 | * -1 if node overlap or lost ram (shift too big) |
| 51 | */ |
| 52 | static int __init populate_memnodemap(const struct bootnode *nodes, |
| 53 | int numnodes, int shift, int *nodeids) |
| 54 | { |
| 55 | unsigned long addr, end; |
| 56 | int i, res = -1; |
| 57 | |
| 58 | memset(memnodemap, 0xff, sizeof(s16)*memnodemapsize); |
| 59 | for (i = 0; i < numnodes; i++) { |
| 60 | addr = nodes[i].start; |
| 61 | end = nodes[i].end; |
| 62 | if (addr >= end) |
| 63 | continue; |
| 64 | if ((end >> shift) >= memnodemapsize) |
| 65 | return 0; |
| 66 | do { |
| 67 | if (memnodemap[addr >> shift] != NUMA_NO_NODE) |
| 68 | return -1; |
| 69 | |
| 70 | if (!nodeids) |
| 71 | memnodemap[addr >> shift] = i; |
| 72 | else |
| 73 | memnodemap[addr >> shift] = nodeids[i]; |
| 74 | |
| 75 | addr += (1UL << shift); |
| 76 | } while (addr < end); |
| 77 | res = 1; |
| 78 | } |
| 79 | return res; |
| 80 | } |
| 81 | |
| 82 | static int __init allocate_cachealigned_memnodemap(void) |
| 83 | { |
| 84 | unsigned long addr; |
| 85 | |
| 86 | memnodemap = memnode.embedded_map; |
| 87 | if (memnodemapsize <= ARRAY_SIZE(memnode.embedded_map)) |
| 88 | return 0; |
| 89 | |
| 90 | addr = 0x8000; |
| 91 | nodemap_size = roundup(sizeof(s16) * memnodemapsize, L1_CACHE_BYTES); |
| 92 | nodemap_addr = find_e820_area(addr, max_pfn<<PAGE_SHIFT, |
| 93 | nodemap_size, L1_CACHE_BYTES); |
| 94 | if (nodemap_addr == -1UL) { |
| 95 | printk(KERN_ERR |
| 96 | "NUMA: Unable to allocate Memory to Node hash map\n"); |
| 97 | nodemap_addr = nodemap_size = 0; |
| 98 | return -1; |
| 99 | } |
| 100 | memnodemap = phys_to_virt(nodemap_addr); |
| 101 | reserve_early(nodemap_addr, nodemap_addr + nodemap_size, "MEMNODEMAP"); |
| 102 | |
| 103 | printk(KERN_DEBUG "NUMA: Allocated memnodemap from %lx - %lx\n", |
| 104 | nodemap_addr, nodemap_addr + nodemap_size); |
| 105 | return 0; |
| 106 | } |
| 107 | |
| 108 | /* |
| 109 | * The LSB of all start and end addresses in the node map is the value of the |
| 110 | * maximum possible shift. |
| 111 | */ |
| 112 | static int __init extract_lsb_from_nodes(const struct bootnode *nodes, |
| 113 | int numnodes) |
| 114 | { |
| 115 | int i, nodes_used = 0; |
| 116 | unsigned long start, end; |
| 117 | unsigned long bitfield = 0, memtop = 0; |
| 118 | |
| 119 | for (i = 0; i < numnodes; i++) { |
| 120 | start = nodes[i].start; |
| 121 | end = nodes[i].end; |
| 122 | if (start >= end) |
| 123 | continue; |
| 124 | bitfield |= start; |
| 125 | nodes_used++; |
| 126 | if (end > memtop) |
| 127 | memtop = end; |
| 128 | } |
| 129 | if (nodes_used <= 1) |
| 130 | i = 63; |
| 131 | else |
| 132 | i = find_first_bit(&bitfield, sizeof(unsigned long)*8); |
| 133 | memnodemapsize = (memtop >> i)+1; |
| 134 | return i; |
| 135 | } |
| 136 | |
| 137 | int __init compute_hash_shift(struct bootnode *nodes, int numnodes, |
| 138 | int *nodeids) |
| 139 | { |
| 140 | int shift; |
| 141 | |
| 142 | shift = extract_lsb_from_nodes(nodes, numnodes); |
| 143 | if (allocate_cachealigned_memnodemap()) |
| 144 | return -1; |
| 145 | printk(KERN_DEBUG "NUMA: Using %d for the hash shift.\n", |
| 146 | shift); |
| 147 | |
| 148 | if (populate_memnodemap(nodes, numnodes, shift, nodeids) != 1) { |
| 149 | printk(KERN_INFO "Your memory is not aligned you need to " |
| 150 | "rebuild your kernel with a bigger NODEMAPSIZE " |
| 151 | "shift=%d\n", shift); |
| 152 | return -1; |
| 153 | } |
| 154 | return shift; |
| 155 | } |
| 156 | |
| 157 | int __meminit __early_pfn_to_nid(unsigned long pfn) |
| 158 | { |
| 159 | return phys_to_nid(pfn << PAGE_SHIFT); |
| 160 | } |
| 161 | |
| 162 | static void * __init early_node_mem(int nodeid, unsigned long start, |
| 163 | unsigned long end, unsigned long size, |
| 164 | unsigned long align) |
| 165 | { |
| 166 | unsigned long mem = find_e820_area(start, end, size, align); |
| 167 | void *ptr; |
| 168 | |
| 169 | if (mem != -1L) |
| 170 | return __va(mem); |
| 171 | |
| 172 | ptr = __alloc_bootmem_nopanic(size, align, __pa(MAX_DMA_ADDRESS)); |
| 173 | if (ptr == NULL) { |
| 174 | printk(KERN_ERR "Cannot find %lu bytes in node %d\n", |
| 175 | size, nodeid); |
| 176 | return NULL; |
| 177 | } |
| 178 | return ptr; |
| 179 | } |
| 180 | |
| 181 | /* Initialize bootmem allocator for a node */ |
| 182 | void __init setup_node_bootmem(int nodeid, unsigned long start, |
| 183 | unsigned long end) |
| 184 | { |
| 185 | unsigned long start_pfn, last_pfn, bootmap_pages, bootmap_size; |
| 186 | unsigned long bootmap_start, nodedata_phys; |
| 187 | void *bootmap; |
| 188 | const int pgdat_size = roundup(sizeof(pg_data_t), PAGE_SIZE); |
| 189 | int nid; |
| 190 | |
| 191 | if (!end) |
| 192 | return; |
| 193 | |
| 194 | start = roundup(start, ZONE_ALIGN); |
| 195 | |
| 196 | printk(KERN_INFO "Bootmem setup node %d %016lx-%016lx\n", nodeid, |
| 197 | start, end); |
| 198 | |
| 199 | start_pfn = start >> PAGE_SHIFT; |
| 200 | last_pfn = end >> PAGE_SHIFT; |
| 201 | |
| 202 | node_data[nodeid] = early_node_mem(nodeid, start, end, pgdat_size, |
| 203 | SMP_CACHE_BYTES); |
| 204 | if (node_data[nodeid] == NULL) |
| 205 | return; |
| 206 | nodedata_phys = __pa(node_data[nodeid]); |
| 207 | printk(KERN_INFO " NODE_DATA [%016lx - %016lx]\n", nodedata_phys, |
| 208 | nodedata_phys + pgdat_size - 1); |
| 209 | |
| 210 | memset(NODE_DATA(nodeid), 0, sizeof(pg_data_t)); |
| 211 | NODE_DATA(nodeid)->bdata = &bootmem_node_data[nodeid]; |
| 212 | NODE_DATA(nodeid)->node_start_pfn = start_pfn; |
| 213 | NODE_DATA(nodeid)->node_spanned_pages = last_pfn - start_pfn; |
| 214 | |
| 215 | /* |
| 216 | * Find a place for the bootmem map |
| 217 | * nodedata_phys could be on other nodes by alloc_bootmem, |
| 218 | * so need to sure bootmap_start not to be small, otherwise |
| 219 | * early_node_mem will get that with find_e820_area instead |
| 220 | * of alloc_bootmem, that could clash with reserved range |
| 221 | */ |
| 222 | bootmap_pages = bootmem_bootmap_pages(last_pfn - start_pfn); |
| 223 | nid = phys_to_nid(nodedata_phys); |
| 224 | if (nid == nodeid) |
| 225 | bootmap_start = roundup(nodedata_phys + pgdat_size, PAGE_SIZE); |
| 226 | else |
| 227 | bootmap_start = roundup(start, PAGE_SIZE); |
| 228 | /* |
| 229 | * SMP_CACHE_BYTES could be enough, but init_bootmem_node like |
| 230 | * to use that to align to PAGE_SIZE |
| 231 | */ |
| 232 | bootmap = early_node_mem(nodeid, bootmap_start, end, |
| 233 | bootmap_pages<<PAGE_SHIFT, PAGE_SIZE); |
| 234 | if (bootmap == NULL) { |
| 235 | if (nodedata_phys < start || nodedata_phys >= end) |
| 236 | free_bootmem(nodedata_phys, pgdat_size); |
| 237 | node_data[nodeid] = NULL; |
| 238 | return; |
| 239 | } |
| 240 | bootmap_start = __pa(bootmap); |
| 241 | |
| 242 | bootmap_size = init_bootmem_node(NODE_DATA(nodeid), |
| 243 | bootmap_start >> PAGE_SHIFT, |
| 244 | start_pfn, last_pfn); |
| 245 | |
| 246 | printk(KERN_INFO " bootmap [%016lx - %016lx] pages %lx\n", |
| 247 | bootmap_start, bootmap_start + bootmap_size - 1, |
| 248 | bootmap_pages); |
| 249 | |
| 250 | free_bootmem_with_active_regions(nodeid, end); |
| 251 | |
| 252 | /* |
| 253 | * convert early reserve to bootmem reserve earlier |
| 254 | * otherwise early_node_mem could use early reserved mem |
| 255 | * on previous node |
| 256 | */ |
| 257 | early_res_to_bootmem(start, end); |
| 258 | |
| 259 | /* |
| 260 | * in some case early_node_mem could use alloc_bootmem |
| 261 | * to get range on other node, don't reserve that again |
| 262 | */ |
| 263 | if (nid != nodeid) |
| 264 | printk(KERN_INFO " NODE_DATA(%d) on node %d\n", nodeid, nid); |
| 265 | else |
| 266 | reserve_bootmem_node(NODE_DATA(nodeid), nodedata_phys, |
| 267 | pgdat_size, BOOTMEM_DEFAULT); |
| 268 | nid = phys_to_nid(bootmap_start); |
| 269 | if (nid != nodeid) |
| 270 | printk(KERN_INFO " bootmap(%d) on node %d\n", nodeid, nid); |
| 271 | else |
| 272 | reserve_bootmem_node(NODE_DATA(nodeid), bootmap_start, |
| 273 | bootmap_pages<<PAGE_SHIFT, BOOTMEM_DEFAULT); |
| 274 | |
| 275 | #ifdef CONFIG_ACPI_NUMA |
| 276 | srat_reserve_add_area(nodeid); |
| 277 | #endif |
| 278 | node_set_online(nodeid); |
| 279 | } |
| 280 | |
| 281 | /* |
| 282 | * There are unfortunately some poorly designed mainboards around that |
| 283 | * only connect memory to a single CPU. This breaks the 1:1 cpu->node |
| 284 | * mapping. To avoid this fill in the mapping for all possible CPUs, |
| 285 | * as the number of CPUs is not known yet. We round robin the existing |
| 286 | * nodes. |
| 287 | */ |
| 288 | void __init numa_init_array(void) |
| 289 | { |
| 290 | int rr, i; |
| 291 | |
| 292 | rr = first_node(node_online_map); |
| 293 | for (i = 0; i < nr_cpu_ids; i++) { |
| 294 | if (early_cpu_to_node(i) != NUMA_NO_NODE) |
| 295 | continue; |
| 296 | numa_set_node(i, rr); |
| 297 | rr = next_node(rr, node_online_map); |
| 298 | if (rr == MAX_NUMNODES) |
| 299 | rr = first_node(node_online_map); |
| 300 | } |
| 301 | } |
| 302 | |
| 303 | #ifdef CONFIG_NUMA_EMU |
| 304 | /* Numa emulation */ |
| 305 | static char *cmdline __initdata; |
| 306 | |
| 307 | /* |
| 308 | * Setups up nid to range from addr to addr + size. If the end |
| 309 | * boundary is greater than max_addr, then max_addr is used instead. |
| 310 | * The return value is 0 if there is additional memory left for |
| 311 | * allocation past addr and -1 otherwise. addr is adjusted to be at |
| 312 | * the end of the node. |
| 313 | */ |
| 314 | static int __init setup_node_range(int nid, struct bootnode *nodes, u64 *addr, |
| 315 | u64 size, u64 max_addr) |
| 316 | { |
| 317 | int ret = 0; |
| 318 | |
| 319 | nodes[nid].start = *addr; |
| 320 | *addr += size; |
| 321 | if (*addr >= max_addr) { |
| 322 | *addr = max_addr; |
| 323 | ret = -1; |
| 324 | } |
| 325 | nodes[nid].end = *addr; |
| 326 | node_set(nid, node_possible_map); |
| 327 | printk(KERN_INFO "Faking node %d at %016Lx-%016Lx (%LuMB)\n", nid, |
| 328 | nodes[nid].start, nodes[nid].end, |
| 329 | (nodes[nid].end - nodes[nid].start) >> 20); |
| 330 | return ret; |
| 331 | } |
| 332 | |
| 333 | /* |
| 334 | * Splits num_nodes nodes up equally starting at node_start. The return value |
| 335 | * is the number of nodes split up and addr is adjusted to be at the end of the |
| 336 | * last node allocated. |
| 337 | */ |
| 338 | static int __init split_nodes_equally(struct bootnode *nodes, u64 *addr, |
| 339 | u64 max_addr, int node_start, |
| 340 | int num_nodes) |
| 341 | { |
| 342 | unsigned int big; |
| 343 | u64 size; |
| 344 | int i; |
| 345 | |
| 346 | if (num_nodes <= 0) |
| 347 | return -1; |
| 348 | if (num_nodes > MAX_NUMNODES) |
| 349 | num_nodes = MAX_NUMNODES; |
| 350 | size = (max_addr - *addr - e820_hole_size(*addr, max_addr)) / |
| 351 | num_nodes; |
| 352 | /* |
| 353 | * Calculate the number of big nodes that can be allocated as a result |
| 354 | * of consolidating the leftovers. |
| 355 | */ |
| 356 | big = ((size & ~FAKE_NODE_MIN_HASH_MASK) * num_nodes) / |
| 357 | FAKE_NODE_MIN_SIZE; |
| 358 | |
| 359 | /* Round down to nearest FAKE_NODE_MIN_SIZE. */ |
| 360 | size &= FAKE_NODE_MIN_HASH_MASK; |
| 361 | if (!size) { |
| 362 | printk(KERN_ERR "Not enough memory for each node. " |
| 363 | "NUMA emulation disabled.\n"); |
| 364 | return -1; |
| 365 | } |
| 366 | |
| 367 | for (i = node_start; i < num_nodes + node_start; i++) { |
| 368 | u64 end = *addr + size; |
| 369 | |
| 370 | if (i < big) |
| 371 | end += FAKE_NODE_MIN_SIZE; |
| 372 | /* |
| 373 | * The final node can have the remaining system RAM. Other |
| 374 | * nodes receive roughly the same amount of available pages. |
| 375 | */ |
| 376 | if (i == num_nodes + node_start - 1) |
| 377 | end = max_addr; |
| 378 | else |
| 379 | while (end - *addr - e820_hole_size(*addr, end) < |
| 380 | size) { |
| 381 | end += FAKE_NODE_MIN_SIZE; |
| 382 | if (end > max_addr) { |
| 383 | end = max_addr; |
| 384 | break; |
| 385 | } |
| 386 | } |
| 387 | if (setup_node_range(i, nodes, addr, end - *addr, max_addr) < 0) |
| 388 | break; |
| 389 | } |
| 390 | return i - node_start + 1; |
| 391 | } |
| 392 | |
| 393 | /* |
| 394 | * Splits the remaining system RAM into chunks of size. The remaining memory is |
| 395 | * always assigned to a final node and can be asymmetric. Returns the number of |
| 396 | * nodes split. |
| 397 | */ |
| 398 | static int __init split_nodes_by_size(struct bootnode *nodes, u64 *addr, |
| 399 | u64 max_addr, int node_start, u64 size) |
| 400 | { |
| 401 | int i = node_start; |
| 402 | size = (size << 20) & FAKE_NODE_MIN_HASH_MASK; |
| 403 | while (!setup_node_range(i++, nodes, addr, size, max_addr)) |
| 404 | ; |
| 405 | return i - node_start; |
| 406 | } |
| 407 | |
| 408 | /* |
| 409 | * Sets up the system RAM area from start_pfn to last_pfn according to the |
| 410 | * numa=fake command-line option. |
| 411 | */ |
| 412 | static struct bootnode nodes[MAX_NUMNODES] __initdata; |
| 413 | |
| 414 | static int __init numa_emulation(unsigned long start_pfn, unsigned long last_pfn) |
| 415 | { |
| 416 | u64 size, addr = start_pfn << PAGE_SHIFT; |
| 417 | u64 max_addr = last_pfn << PAGE_SHIFT; |
| 418 | int num_nodes = 0, num = 0, coeff_flag, coeff = -1, i; |
| 419 | |
| 420 | memset(&nodes, 0, sizeof(nodes)); |
| 421 | /* |
| 422 | * If the numa=fake command-line is just a single number N, split the |
| 423 | * system RAM into N fake nodes. |
| 424 | */ |
| 425 | if (!strchr(cmdline, '*') && !strchr(cmdline, ',')) { |
| 426 | long n = simple_strtol(cmdline, NULL, 0); |
| 427 | |
| 428 | num_nodes = split_nodes_equally(nodes, &addr, max_addr, 0, n); |
| 429 | if (num_nodes < 0) |
| 430 | return num_nodes; |
| 431 | goto out; |
| 432 | } |
| 433 | |
| 434 | /* Parse the command line. */ |
| 435 | for (coeff_flag = 0; ; cmdline++) { |
| 436 | if (*cmdline && isdigit(*cmdline)) { |
| 437 | num = num * 10 + *cmdline - '0'; |
| 438 | continue; |
| 439 | } |
| 440 | if (*cmdline == '*') { |
| 441 | if (num > 0) |
| 442 | coeff = num; |
| 443 | coeff_flag = 1; |
| 444 | } |
| 445 | if (!*cmdline || *cmdline == ',') { |
| 446 | if (!coeff_flag) |
| 447 | coeff = 1; |
| 448 | /* |
| 449 | * Round down to the nearest FAKE_NODE_MIN_SIZE. |
| 450 | * Command-line coefficients are in megabytes. |
| 451 | */ |
| 452 | size = ((u64)num << 20) & FAKE_NODE_MIN_HASH_MASK; |
| 453 | if (size) |
| 454 | for (i = 0; i < coeff; i++, num_nodes++) |
| 455 | if (setup_node_range(num_nodes, nodes, |
| 456 | &addr, size, max_addr) < 0) |
| 457 | goto done; |
| 458 | if (!*cmdline) |
| 459 | break; |
| 460 | coeff_flag = 0; |
| 461 | coeff = -1; |
| 462 | } |
| 463 | num = 0; |
| 464 | } |
| 465 | done: |
| 466 | if (!num_nodes) |
| 467 | return -1; |
| 468 | /* Fill remainder of system RAM, if appropriate. */ |
| 469 | if (addr < max_addr) { |
| 470 | if (coeff_flag && coeff < 0) { |
| 471 | /* Split remaining nodes into num-sized chunks */ |
| 472 | num_nodes += split_nodes_by_size(nodes, &addr, max_addr, |
| 473 | num_nodes, num); |
| 474 | goto out; |
| 475 | } |
| 476 | switch (*(cmdline - 1)) { |
| 477 | case '*': |
| 478 | /* Split remaining nodes into coeff chunks */ |
| 479 | if (coeff <= 0) |
| 480 | break; |
| 481 | num_nodes += split_nodes_equally(nodes, &addr, max_addr, |
| 482 | num_nodes, coeff); |
| 483 | break; |
| 484 | case ',': |
| 485 | /* Do not allocate remaining system RAM */ |
| 486 | break; |
| 487 | default: |
| 488 | /* Give one final node */ |
| 489 | setup_node_range(num_nodes, nodes, &addr, |
| 490 | max_addr - addr, max_addr); |
| 491 | num_nodes++; |
| 492 | } |
| 493 | } |
| 494 | out: |
| 495 | memnode_shift = compute_hash_shift(nodes, num_nodes, NULL); |
| 496 | if (memnode_shift < 0) { |
| 497 | memnode_shift = 0; |
| 498 | printk(KERN_ERR "No NUMA hash function found. NUMA emulation " |
| 499 | "disabled.\n"); |
| 500 | return -1; |
| 501 | } |
| 502 | |
| 503 | /* |
| 504 | * We need to vacate all active ranges that may have been registered by |
| 505 | * SRAT and set acpi_numa to -1 so that srat_disabled() always returns |
| 506 | * true. NUMA emulation has succeeded so we will not scan ACPI nodes. |
| 507 | */ |
| 508 | remove_all_active_ranges(); |
| 509 | #ifdef CONFIG_ACPI_NUMA |
| 510 | acpi_numa = -1; |
| 511 | #endif |
| 512 | for_each_node_mask(i, node_possible_map) { |
| 513 | e820_register_active_regions(i, nodes[i].start >> PAGE_SHIFT, |
| 514 | nodes[i].end >> PAGE_SHIFT); |
| 515 | setup_node_bootmem(i, nodes[i].start, nodes[i].end); |
| 516 | } |
| 517 | acpi_fake_nodes(nodes, num_nodes); |
| 518 | numa_init_array(); |
| 519 | return 0; |
| 520 | } |
| 521 | #endif /* CONFIG_NUMA_EMU */ |
| 522 | |
| 523 | void __init initmem_init(unsigned long start_pfn, unsigned long last_pfn) |
| 524 | { |
| 525 | int i; |
| 526 | |
| 527 | nodes_clear(node_possible_map); |
| 528 | nodes_clear(node_online_map); |
| 529 | |
| 530 | #ifdef CONFIG_NUMA_EMU |
| 531 | if (cmdline && !numa_emulation(start_pfn, last_pfn)) |
| 532 | return; |
| 533 | nodes_clear(node_possible_map); |
| 534 | nodes_clear(node_online_map); |
| 535 | #endif |
| 536 | |
| 537 | #ifdef CONFIG_ACPI_NUMA |
| 538 | if (!numa_off && !acpi_scan_nodes(start_pfn << PAGE_SHIFT, |
| 539 | last_pfn << PAGE_SHIFT)) |
| 540 | return; |
| 541 | nodes_clear(node_possible_map); |
| 542 | nodes_clear(node_online_map); |
| 543 | #endif |
| 544 | |
| 545 | #ifdef CONFIG_K8_NUMA |
| 546 | if (!numa_off && !k8_scan_nodes(start_pfn<<PAGE_SHIFT, |
| 547 | last_pfn<<PAGE_SHIFT)) |
| 548 | return; |
| 549 | nodes_clear(node_possible_map); |
| 550 | nodes_clear(node_online_map); |
| 551 | #endif |
| 552 | printk(KERN_INFO "%s\n", |
| 553 | numa_off ? "NUMA turned off" : "No NUMA configuration found"); |
| 554 | |
| 555 | printk(KERN_INFO "Faking a node at %016lx-%016lx\n", |
| 556 | start_pfn << PAGE_SHIFT, |
| 557 | last_pfn << PAGE_SHIFT); |
| 558 | /* setup dummy node covering all memory */ |
| 559 | memnode_shift = 63; |
| 560 | memnodemap = memnode.embedded_map; |
| 561 | memnodemap[0] = 0; |
| 562 | node_set_online(0); |
| 563 | node_set(0, node_possible_map); |
| 564 | for (i = 0; i < nr_cpu_ids; i++) |
| 565 | numa_set_node(i, 0); |
| 566 | e820_register_active_regions(0, start_pfn, last_pfn); |
| 567 | setup_node_bootmem(0, start_pfn << PAGE_SHIFT, last_pfn << PAGE_SHIFT); |
| 568 | } |
| 569 | |
| 570 | unsigned long __init numa_free_all_bootmem(void) |
| 571 | { |
| 572 | unsigned long pages = 0; |
| 573 | int i; |
| 574 | |
| 575 | for_each_online_node(i) |
| 576 | pages += free_all_bootmem_node(NODE_DATA(i)); |
| 577 | |
| 578 | return pages; |
| 579 | } |
| 580 | |
| 581 | void __init paging_init(void) |
| 582 | { |
| 583 | unsigned long max_zone_pfns[MAX_NR_ZONES]; |
| 584 | |
| 585 | memset(max_zone_pfns, 0, sizeof(max_zone_pfns)); |
| 586 | max_zone_pfns[ZONE_DMA] = MAX_DMA_PFN; |
| 587 | max_zone_pfns[ZONE_DMA32] = MAX_DMA32_PFN; |
| 588 | max_zone_pfns[ZONE_NORMAL] = max_pfn; |
| 589 | |
| 590 | sparse_memory_present_with_active_regions(MAX_NUMNODES); |
| 591 | sparse_init(); |
| 592 | |
| 593 | free_area_init_nodes(max_zone_pfns); |
| 594 | } |
| 595 | |
| 596 | static __init int numa_setup(char *opt) |
| 597 | { |
| 598 | if (!opt) |
| 599 | return -EINVAL; |
| 600 | if (!strncmp(opt, "off", 3)) |
| 601 | numa_off = 1; |
| 602 | #ifdef CONFIG_NUMA_EMU |
| 603 | if (!strncmp(opt, "fake=", 5)) |
| 604 | cmdline = opt + 5; |
| 605 | #endif |
| 606 | #ifdef CONFIG_ACPI_NUMA |
| 607 | if (!strncmp(opt, "noacpi", 6)) |
| 608 | acpi_numa = -1; |
| 609 | if (!strncmp(opt, "hotadd=", 7)) |
| 610 | hotadd_percent = simple_strtoul(opt+7, NULL, 10); |
| 611 | #endif |
| 612 | return 0; |
| 613 | } |
| 614 | early_param("numa", numa_setup); |
| 615 | |
| 616 | #ifdef CONFIG_NUMA |
| 617 | /* |
| 618 | * Setup early cpu_to_node. |
| 619 | * |
| 620 | * Populate cpu_to_node[] only if x86_cpu_to_apicid[], |
| 621 | * and apicid_to_node[] tables have valid entries for a CPU. |
| 622 | * This means we skip cpu_to_node[] initialisation for NUMA |
| 623 | * emulation and faking node case (when running a kernel compiled |
| 624 | * for NUMA on a non NUMA box), which is OK as cpu_to_node[] |
| 625 | * is already initialized in a round robin manner at numa_init_array, |
| 626 | * prior to this call, and this initialization is good enough |
| 627 | * for the fake NUMA cases. |
| 628 | * |
| 629 | * Called before the per_cpu areas are setup. |
| 630 | */ |
| 631 | void __init init_cpu_to_node(void) |
| 632 | { |
| 633 | int cpu; |
| 634 | u16 *cpu_to_apicid = early_per_cpu_ptr(x86_cpu_to_apicid); |
| 635 | |
| 636 | BUG_ON(cpu_to_apicid == NULL); |
| 637 | |
| 638 | for_each_possible_cpu(cpu) { |
| 639 | int node; |
| 640 | u16 apicid = cpu_to_apicid[cpu]; |
| 641 | |
| 642 | if (apicid == BAD_APICID) |
| 643 | continue; |
| 644 | node = apicid_to_node[apicid]; |
| 645 | if (node == NUMA_NO_NODE) |
| 646 | continue; |
| 647 | if (!node_online(node)) |
| 648 | continue; |
| 649 | numa_set_node(cpu, node); |
| 650 | } |
| 651 | } |
| 652 | #endif |
| 653 | |
| 654 | |
| 655 | void __cpuinit numa_set_node(int cpu, int node) |
| 656 | { |
| 657 | int *cpu_to_node_map = early_per_cpu_ptr(x86_cpu_to_node_map); |
| 658 | |
| 659 | /* early setting, no percpu area yet */ |
| 660 | if (cpu_to_node_map) { |
| 661 | cpu_to_node_map[cpu] = node; |
| 662 | return; |
| 663 | } |
| 664 | |
| 665 | #ifdef CONFIG_DEBUG_PER_CPU_MAPS |
| 666 | if (cpu >= nr_cpu_ids || !cpu_possible(cpu)) { |
| 667 | printk(KERN_ERR "numa_set_node: invalid cpu# (%d)\n", cpu); |
| 668 | dump_stack(); |
| 669 | return; |
| 670 | } |
| 671 | #endif |
| 672 | per_cpu(x86_cpu_to_node_map, cpu) = node; |
| 673 | |
| 674 | if (node != NUMA_NO_NODE) |
| 675 | per_cpu(node_number, cpu) = node; |
| 676 | } |
| 677 | |
| 678 | void __cpuinit numa_clear_node(int cpu) |
| 679 | { |
| 680 | numa_set_node(cpu, NUMA_NO_NODE); |
| 681 | } |
| 682 | |
| 683 | #ifndef CONFIG_DEBUG_PER_CPU_MAPS |
| 684 | |
| 685 | void __cpuinit numa_add_cpu(int cpu) |
| 686 | { |
| 687 | cpumask_set_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]); |
| 688 | } |
| 689 | |
| 690 | void __cpuinit numa_remove_cpu(int cpu) |
| 691 | { |
| 692 | cpumask_clear_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]); |
| 693 | } |
| 694 | |
| 695 | #else /* CONFIG_DEBUG_PER_CPU_MAPS */ |
| 696 | |
| 697 | /* |
| 698 | * --------- debug versions of the numa functions --------- |
| 699 | */ |
| 700 | static void __cpuinit numa_set_cpumask(int cpu, int enable) |
| 701 | { |
| 702 | int node = early_cpu_to_node(cpu); |
| 703 | struct cpumask *mask; |
| 704 | char buf[64]; |
| 705 | |
| 706 | mask = node_to_cpumask_map[node]; |
| 707 | if (mask == NULL) { |
| 708 | printk(KERN_ERR "node_to_cpumask_map[%i] NULL\n", node); |
| 709 | dump_stack(); |
| 710 | return; |
| 711 | } |
| 712 | |
| 713 | if (enable) |
| 714 | cpumask_set_cpu(cpu, mask); |
| 715 | else |
| 716 | cpumask_clear_cpu(cpu, mask); |
| 717 | |
| 718 | cpulist_scnprintf(buf, sizeof(buf), mask); |
| 719 | printk(KERN_DEBUG "%s cpu %d node %d: mask now %s\n", |
| 720 | enable ? "numa_add_cpu" : "numa_remove_cpu", cpu, node, buf); |
| 721 | } |
| 722 | |
| 723 | void __cpuinit numa_add_cpu(int cpu) |
| 724 | { |
| 725 | numa_set_cpumask(cpu, 1); |
| 726 | } |
| 727 | |
| 728 | void __cpuinit numa_remove_cpu(int cpu) |
| 729 | { |
| 730 | numa_set_cpumask(cpu, 0); |
| 731 | } |
| 732 | |
| 733 | int cpu_to_node(int cpu) |
| 734 | { |
| 735 | if (early_per_cpu_ptr(x86_cpu_to_node_map)) { |
| 736 | printk(KERN_WARNING |
| 737 | "cpu_to_node(%d): usage too early!\n", cpu); |
| 738 | dump_stack(); |
| 739 | return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu]; |
| 740 | } |
| 741 | return per_cpu(x86_cpu_to_node_map, cpu); |
| 742 | } |
| 743 | EXPORT_SYMBOL(cpu_to_node); |
| 744 | |
| 745 | /* |
| 746 | * Same function as cpu_to_node() but used if called before the |
| 747 | * per_cpu areas are setup. |
| 748 | */ |
| 749 | int early_cpu_to_node(int cpu) |
| 750 | { |
| 751 | if (early_per_cpu_ptr(x86_cpu_to_node_map)) |
| 752 | return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu]; |
| 753 | |
| 754 | if (!cpu_possible(cpu)) { |
| 755 | printk(KERN_WARNING |
| 756 | "early_cpu_to_node(%d): no per_cpu area!\n", cpu); |
| 757 | dump_stack(); |
| 758 | return NUMA_NO_NODE; |
| 759 | } |
| 760 | return per_cpu(x86_cpu_to_node_map, cpu); |
| 761 | } |
| 762 | |
| 763 | /* |
| 764 | * --------- end of debug versions of the numa functions --------- |
| 765 | */ |
| 766 | |
| 767 | #endif /* CONFIG_DEBUG_PER_CPU_MAPS */ |