1 /* 2 * pSeries NUMA support 3 * 4 * Copyright (C) 2002 Anton Blanchard <anton@au.ibm.com>, IBM 5 * 6 * This program is free software; you can redistribute it and/or 7 * modify it under the terms of the GNU General Public License 8 * as published by the Free Software Foundation; either version 9 * 2 of the License, or (at your option) any later version. 10 */ 11 #include <linux/threads.h> 12 #include <linux/bootmem.h> 13 #include <linux/init.h> 14 #include <linux/mm.h> 15 #include <linux/mmzone.h> 16 #include <linux/module.h> 17 #include <linux/nodemask.h> 18 #include <linux/cpu.h> 19 #include <linux/notifier.h> 20 #include <linux/memblock.h> 21 #include <linux/of.h> 22 #include <linux/pfn.h> 23 #include <asm/sparsemem.h> 24 #include <asm/prom.h> 25 #include <asm/system.h> 26 #include <asm/smp.h> 27 28 static int numa_enabled = 1; 29 30 static char *cmdline __initdata; 31 32 static int numa_debug; 33 #define dbg(args...) if (numa_debug) { printk(KERN_INFO args); } 34 35 int numa_cpu_lookup_table[NR_CPUS]; 36 cpumask_var_t node_to_cpumask_map[MAX_NUMNODES]; 37 struct pglist_data *node_data[MAX_NUMNODES]; 38 39 EXPORT_SYMBOL(numa_cpu_lookup_table); 40 EXPORT_SYMBOL(node_to_cpumask_map); 41 EXPORT_SYMBOL(node_data); 42 43 static int min_common_depth; 44 static int n_mem_addr_cells, n_mem_size_cells; 45 static int form1_affinity; 46 47 #define MAX_DISTANCE_REF_POINTS 4 48 static int distance_ref_points_depth; 49 static const unsigned int *distance_ref_points; 50 static int distance_lookup_table[MAX_NUMNODES][MAX_DISTANCE_REF_POINTS]; 51 52 /* 53 * Allocate node_to_cpumask_map based on number of available nodes 54 * Requires node_possible_map to be valid. 55 * 56 * Note: node_to_cpumask() is not valid until after this is done. 57 */ 58 static void __init setup_node_to_cpumask_map(void) 59 { 60 unsigned int node, num = 0; 61 62 /* setup nr_node_ids if not done yet */ 63 if (nr_node_ids == MAX_NUMNODES) { 64 for_each_node_mask(node, node_possible_map) 65 num = node; 66 nr_node_ids = num + 1; 67 } 68 69 /* allocate the map */ 70 for (node = 0; node < nr_node_ids; node++) 71 alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]); 72 73 /* cpumask_of_node() will now work */ 74 dbg("Node to cpumask map for %d nodes\n", nr_node_ids); 75 } 76 77 static int __cpuinit fake_numa_create_new_node(unsigned long end_pfn, 78 unsigned int *nid) 79 { 80 unsigned long long mem; 81 char *p = cmdline; 82 static unsigned int fake_nid; 83 static unsigned long long curr_boundary; 84 85 /* 86 * Modify node id, iff we started creating NUMA nodes 87 * We want to continue from where we left of the last time 88 */ 89 if (fake_nid) 90 *nid = fake_nid; 91 /* 92 * In case there are no more arguments to parse, the 93 * node_id should be the same as the last fake node id 94 * (we've handled this above). 95 */ 96 if (!p) 97 return 0; 98 99 mem = memparse(p, &p); 100 if (!mem) 101 return 0; 102 103 if (mem < curr_boundary) 104 return 0; 105 106 curr_boundary = mem; 107 108 if ((end_pfn << PAGE_SHIFT) > mem) { 109 /* 110 * Skip commas and spaces 111 */ 112 while (*p == ',' || *p == ' ' || *p == '\t') 113 p++; 114 115 cmdline = p; 116 fake_nid++; 117 *nid = fake_nid; 118 dbg("created new fake_node with id %d\n", fake_nid); 119 return 1; 120 } 121 return 0; 122 } 123 124 /* 125 * get_active_region_work_fn - A helper function for get_node_active_region 126 * Returns datax set to the start_pfn and end_pfn if they contain 127 * the initial value of datax->start_pfn between them 128 * @start_pfn: start page(inclusive) of region to check 129 * @end_pfn: end page(exclusive) of region to check 130 * @datax: comes in with ->start_pfn set to value to search for and 131 * goes out with active range if it contains it 132 * Returns 1 if search value is in range else 0 133 */ 134 static int __init get_active_region_work_fn(unsigned long start_pfn, 135 unsigned long end_pfn, void *datax) 136 { 137 struct node_active_region *data; 138 data = (struct node_active_region *)datax; 139 140 if (start_pfn <= data->start_pfn && end_pfn > data->start_pfn) { 141 data->start_pfn = start_pfn; 142 data->end_pfn = end_pfn; 143 return 1; 144 } 145 return 0; 146 147 } 148 149 /* 150 * get_node_active_region - Return active region containing start_pfn 151 * Active range returned is empty if none found. 152 * @start_pfn: The page to return the region for. 153 * @node_ar: Returned set to the active region containing start_pfn 154 */ 155 static void __init get_node_active_region(unsigned long start_pfn, 156 struct node_active_region *node_ar) 157 { 158 int nid = early_pfn_to_nid(start_pfn); 159 160 node_ar->nid = nid; 161 node_ar->start_pfn = start_pfn; 162 node_ar->end_pfn = start_pfn; 163 work_with_active_regions(nid, get_active_region_work_fn, node_ar); 164 } 165 166 static void __cpuinit map_cpu_to_node(int cpu, int node) 167 { 168 numa_cpu_lookup_table[cpu] = node; 169 170 dbg("adding cpu %d to node %d\n", cpu, node); 171 172 if (!(cpumask_test_cpu(cpu, node_to_cpumask_map[node]))) 173 cpumask_set_cpu(cpu, node_to_cpumask_map[node]); 174 } 175 176 #ifdef CONFIG_HOTPLUG_CPU 177 static void unmap_cpu_from_node(unsigned long cpu) 178 { 179 int node = numa_cpu_lookup_table[cpu]; 180 181 dbg("removing cpu %lu from node %d\n", cpu, node); 182 183 if (cpumask_test_cpu(cpu, node_to_cpumask_map[node])) { 184 cpumask_set_cpu(cpu, node_to_cpumask_map[node]); 185 } else { 186 printk(KERN_ERR "WARNING: cpu %lu not found in node %d\n", 187 cpu, node); 188 } 189 } 190 #endif /* CONFIG_HOTPLUG_CPU */ 191 192 /* must hold reference to node during call */ 193 static const int *of_get_associativity(struct device_node *dev) 194 { 195 return of_get_property(dev, "ibm,associativity", NULL); 196 } 197 198 /* 199 * Returns the property linux,drconf-usable-memory if 200 * it exists (the property exists only in kexec/kdump kernels, 201 * added by kexec-tools) 202 */ 203 static const u32 *of_get_usable_memory(struct device_node *memory) 204 { 205 const u32 *prop; 206 u32 len; 207 prop = of_get_property(memory, "linux,drconf-usable-memory", &len); 208 if (!prop || len < sizeof(unsigned int)) 209 return 0; 210 return prop; 211 } 212 213 int __node_distance(int a, int b) 214 { 215 int i; 216 int distance = LOCAL_DISTANCE; 217 218 if (!form1_affinity) 219 return distance; 220 221 for (i = 0; i < distance_ref_points_depth; i++) { 222 if (distance_lookup_table[a][i] == distance_lookup_table[b][i]) 223 break; 224 225 /* Double the distance for each NUMA level */ 226 distance *= 2; 227 } 228 229 return distance; 230 } 231 232 static void initialize_distance_lookup_table(int nid, 233 const unsigned int *associativity) 234 { 235 int i; 236 237 if (!form1_affinity) 238 return; 239 240 for (i = 0; i < distance_ref_points_depth; i++) { 241 distance_lookup_table[nid][i] = 242 associativity[distance_ref_points[i]]; 243 } 244 } 245 246 /* Returns nid in the range [0..MAX_NUMNODES-1], or -1 if no useful numa 247 * info is found. 248 */ 249 static int of_node_to_nid_single(struct device_node *device) 250 { 251 int nid = -1; 252 const unsigned int *tmp; 253 254 if (min_common_depth == -1) 255 goto out; 256 257 tmp = of_get_associativity(device); 258 if (!tmp) 259 goto out; 260 261 if (tmp[0] >= min_common_depth) 262 nid = tmp[min_common_depth]; 263 264 /* POWER4 LPAR uses 0xffff as invalid node */ 265 if (nid == 0xffff || nid >= MAX_NUMNODES) 266 nid = -1; 267 268 if (nid > 0 && tmp[0] >= distance_ref_points_depth) 269 initialize_distance_lookup_table(nid, tmp); 270 271 out: 272 return nid; 273 } 274 275 /* Walk the device tree upwards, looking for an associativity id */ 276 int of_node_to_nid(struct device_node *device) 277 { 278 struct device_node *tmp; 279 int nid = -1; 280 281 of_node_get(device); 282 while (device) { 283 nid = of_node_to_nid_single(device); 284 if (nid != -1) 285 break; 286 287 tmp = device; 288 device = of_get_parent(tmp); 289 of_node_put(tmp); 290 } 291 of_node_put(device); 292 293 return nid; 294 } 295 EXPORT_SYMBOL_GPL(of_node_to_nid); 296 297 static int __init find_min_common_depth(void) 298 { 299 int depth; 300 struct device_node *rtas_root; 301 struct device_node *chosen; 302 const char *vec5; 303 304 rtas_root = of_find_node_by_path("/rtas"); 305 306 if (!rtas_root) 307 return -1; 308 309 /* 310 * This property is a set of 32-bit integers, each representing 311 * an index into the ibm,associativity nodes. 312 * 313 * With form 0 affinity the first integer is for an SMP configuration 314 * (should be all 0's) and the second is for a normal NUMA 315 * configuration. We have only one level of NUMA. 316 * 317 * With form 1 affinity the first integer is the most significant 318 * NUMA boundary and the following are progressively less significant 319 * boundaries. There can be more than one level of NUMA. 320 */ 321 distance_ref_points = of_get_property(rtas_root, 322 "ibm,associativity-reference-points", 323 &distance_ref_points_depth); 324 325 if (!distance_ref_points) { 326 dbg("NUMA: ibm,associativity-reference-points not found.\n"); 327 goto err; 328 } 329 330 distance_ref_points_depth /= sizeof(int); 331 332 #define VEC5_AFFINITY_BYTE 5 333 #define VEC5_AFFINITY 0x80 334 chosen = of_find_node_by_path("/chosen"); 335 if (chosen) { 336 vec5 = of_get_property(chosen, "ibm,architecture-vec-5", NULL); 337 if (vec5 && (vec5[VEC5_AFFINITY_BYTE] & VEC5_AFFINITY)) { 338 dbg("Using form 1 affinity\n"); 339 form1_affinity = 1; 340 } 341 } 342 343 if (form1_affinity) { 344 depth = distance_ref_points[0]; 345 } else { 346 if (distance_ref_points_depth < 2) { 347 printk(KERN_WARNING "NUMA: " 348 "short ibm,associativity-reference-points\n"); 349 goto err; 350 } 351 352 depth = distance_ref_points[1]; 353 } 354 355 /* 356 * Warn and cap if the hardware supports more than 357 * MAX_DISTANCE_REF_POINTS domains. 358 */ 359 if (distance_ref_points_depth > MAX_DISTANCE_REF_POINTS) { 360 printk(KERN_WARNING "NUMA: distance array capped at " 361 "%d entries\n", MAX_DISTANCE_REF_POINTS); 362 distance_ref_points_depth = MAX_DISTANCE_REF_POINTS; 363 } 364 365 of_node_put(rtas_root); 366 return depth; 367 368 err: 369 of_node_put(rtas_root); 370 return -1; 371 } 372 373 static void __init get_n_mem_cells(int *n_addr_cells, int *n_size_cells) 374 { 375 struct device_node *memory = NULL; 376 377 memory = of_find_node_by_type(memory, "memory"); 378 if (!memory) 379 panic("numa.c: No memory nodes found!"); 380 381 *n_addr_cells = of_n_addr_cells(memory); 382 *n_size_cells = of_n_size_cells(memory); 383 of_node_put(memory); 384 } 385 386 static unsigned long __devinit read_n_cells(int n, const unsigned int **buf) 387 { 388 unsigned long result = 0; 389 390 while (n--) { 391 result = (result << 32) | **buf; 392 (*buf)++; 393 } 394 return result; 395 } 396 397 struct of_drconf_cell { 398 u64 base_addr; 399 u32 drc_index; 400 u32 reserved; 401 u32 aa_index; 402 u32 flags; 403 }; 404 405 #define DRCONF_MEM_ASSIGNED 0x00000008 406 #define DRCONF_MEM_AI_INVALID 0x00000040 407 #define DRCONF_MEM_RESERVED 0x00000080 408 409 /* 410 * Read the next memblock list entry from the ibm,dynamic-memory property 411 * and return the information in the provided of_drconf_cell structure. 412 */ 413 static void read_drconf_cell(struct of_drconf_cell *drmem, const u32 **cellp) 414 { 415 const u32 *cp; 416 417 drmem->base_addr = read_n_cells(n_mem_addr_cells, cellp); 418 419 cp = *cellp; 420 drmem->drc_index = cp[0]; 421 drmem->reserved = cp[1]; 422 drmem->aa_index = cp[2]; 423 drmem->flags = cp[3]; 424 425 *cellp = cp + 4; 426 } 427 428 /* 429 * Retreive and validate the ibm,dynamic-memory property of the device tree. 430 * 431 * The layout of the ibm,dynamic-memory property is a number N of memblock 432 * list entries followed by N memblock list entries. Each memblock list entry 433 * contains information as layed out in the of_drconf_cell struct above. 434 */ 435 static int of_get_drconf_memory(struct device_node *memory, const u32 **dm) 436 { 437 const u32 *prop; 438 u32 len, entries; 439 440 prop = of_get_property(memory, "ibm,dynamic-memory", &len); 441 if (!prop || len < sizeof(unsigned int)) 442 return 0; 443 444 entries = *prop++; 445 446 /* Now that we know the number of entries, revalidate the size 447 * of the property read in to ensure we have everything 448 */ 449 if (len < (entries * (n_mem_addr_cells + 4) + 1) * sizeof(unsigned int)) 450 return 0; 451 452 *dm = prop; 453 return entries; 454 } 455 456 /* 457 * Retreive and validate the ibm,lmb-size property for drconf memory 458 * from the device tree. 459 */ 460 static u64 of_get_lmb_size(struct device_node *memory) 461 { 462 const u32 *prop; 463 u32 len; 464 465 prop = of_get_property(memory, "ibm,lmb-size", &len); 466 if (!prop || len < sizeof(unsigned int)) 467 return 0; 468 469 return read_n_cells(n_mem_size_cells, &prop); 470 } 471 472 struct assoc_arrays { 473 u32 n_arrays; 474 u32 array_sz; 475 const u32 *arrays; 476 }; 477 478 /* 479 * Retreive and validate the list of associativity arrays for drconf 480 * memory from the ibm,associativity-lookup-arrays property of the 481 * device tree.. 482 * 483 * The layout of the ibm,associativity-lookup-arrays property is a number N 484 * indicating the number of associativity arrays, followed by a number M 485 * indicating the size of each associativity array, followed by a list 486 * of N associativity arrays. 487 */ 488 static int of_get_assoc_arrays(struct device_node *memory, 489 struct assoc_arrays *aa) 490 { 491 const u32 *prop; 492 u32 len; 493 494 prop = of_get_property(memory, "ibm,associativity-lookup-arrays", &len); 495 if (!prop || len < 2 * sizeof(unsigned int)) 496 return -1; 497 498 aa->n_arrays = *prop++; 499 aa->array_sz = *prop++; 500 501 /* Now that we know the number of arrrays and size of each array, 502 * revalidate the size of the property read in. 503 */ 504 if (len < (aa->n_arrays * aa->array_sz + 2) * sizeof(unsigned int)) 505 return -1; 506 507 aa->arrays = prop; 508 return 0; 509 } 510 511 /* 512 * This is like of_node_to_nid_single() for memory represented in the 513 * ibm,dynamic-reconfiguration-memory node. 514 */ 515 static int of_drconf_to_nid_single(struct of_drconf_cell *drmem, 516 struct assoc_arrays *aa) 517 { 518 int default_nid = 0; 519 int nid = default_nid; 520 int index; 521 522 if (min_common_depth > 0 && min_common_depth <= aa->array_sz && 523 !(drmem->flags & DRCONF_MEM_AI_INVALID) && 524 drmem->aa_index < aa->n_arrays) { 525 index = drmem->aa_index * aa->array_sz + min_common_depth - 1; 526 nid = aa->arrays[index]; 527 528 if (nid == 0xffff || nid >= MAX_NUMNODES) 529 nid = default_nid; 530 } 531 532 return nid; 533 } 534 535 /* 536 * Figure out to which domain a cpu belongs and stick it there. 537 * Return the id of the domain used. 538 */ 539 static int __cpuinit numa_setup_cpu(unsigned long lcpu) 540 { 541 int nid = 0; 542 struct device_node *cpu = of_get_cpu_node(lcpu, NULL); 543 544 if (!cpu) { 545 WARN_ON(1); 546 goto out; 547 } 548 549 nid = of_node_to_nid_single(cpu); 550 551 if (nid < 0 || !node_online(nid)) 552 nid = first_online_node; 553 out: 554 map_cpu_to_node(lcpu, nid); 555 556 of_node_put(cpu); 557 558 return nid; 559 } 560 561 static int __cpuinit cpu_numa_callback(struct notifier_block *nfb, 562 unsigned long action, 563 void *hcpu) 564 { 565 unsigned long lcpu = (unsigned long)hcpu; 566 int ret = NOTIFY_DONE; 567 568 switch (action) { 569 case CPU_UP_PREPARE: 570 case CPU_UP_PREPARE_FROZEN: 571 numa_setup_cpu(lcpu); 572 ret = NOTIFY_OK; 573 break; 574 #ifdef CONFIG_HOTPLUG_CPU 575 case CPU_DEAD: 576 case CPU_DEAD_FROZEN: 577 case CPU_UP_CANCELED: 578 case CPU_UP_CANCELED_FROZEN: 579 unmap_cpu_from_node(lcpu); 580 break; 581 ret = NOTIFY_OK; 582 #endif 583 } 584 return ret; 585 } 586 587 /* 588 * Check and possibly modify a memory region to enforce the memory limit. 589 * 590 * Returns the size the region should have to enforce the memory limit. 591 * This will either be the original value of size, a truncated value, 592 * or zero. If the returned value of size is 0 the region should be 593 * discarded as it lies wholy above the memory limit. 594 */ 595 static unsigned long __init numa_enforce_memory_limit(unsigned long start, 596 unsigned long size) 597 { 598 /* 599 * We use memblock_end_of_DRAM() in here instead of memory_limit because 600 * we've already adjusted it for the limit and it takes care of 601 * having memory holes below the limit. Also, in the case of 602 * iommu_is_off, memory_limit is not set but is implicitly enforced. 603 */ 604 605 if (start + size <= memblock_end_of_DRAM()) 606 return size; 607 608 if (start >= memblock_end_of_DRAM()) 609 return 0; 610 611 return memblock_end_of_DRAM() - start; 612 } 613 614 /* 615 * Reads the counter for a given entry in 616 * linux,drconf-usable-memory property 617 */ 618 static inline int __init read_usm_ranges(const u32 **usm) 619 { 620 /* 621 * For each lmb in ibm,dynamic-memory a corresponding 622 * entry in linux,drconf-usable-memory property contains 623 * a counter followed by that many (base, size) duple. 624 * read the counter from linux,drconf-usable-memory 625 */ 626 return read_n_cells(n_mem_size_cells, usm); 627 } 628 629 /* 630 * Extract NUMA information from the ibm,dynamic-reconfiguration-memory 631 * node. This assumes n_mem_{addr,size}_cells have been set. 632 */ 633 static void __init parse_drconf_memory(struct device_node *memory) 634 { 635 const u32 *dm, *usm; 636 unsigned int n, rc, ranges, is_kexec_kdump = 0; 637 unsigned long lmb_size, base, size, sz; 638 int nid; 639 struct assoc_arrays aa; 640 641 n = of_get_drconf_memory(memory, &dm); 642 if (!n) 643 return; 644 645 lmb_size = of_get_lmb_size(memory); 646 if (!lmb_size) 647 return; 648 649 rc = of_get_assoc_arrays(memory, &aa); 650 if (rc) 651 return; 652 653 /* check if this is a kexec/kdump kernel */ 654 usm = of_get_usable_memory(memory); 655 if (usm != NULL) 656 is_kexec_kdump = 1; 657 658 for (; n != 0; --n) { 659 struct of_drconf_cell drmem; 660 661 read_drconf_cell(&drmem, &dm); 662 663 /* skip this block if the reserved bit is set in flags (0x80) 664 or if the block is not assigned to this partition (0x8) */ 665 if ((drmem.flags & DRCONF_MEM_RESERVED) 666 || !(drmem.flags & DRCONF_MEM_ASSIGNED)) 667 continue; 668 669 base = drmem.base_addr; 670 size = lmb_size; 671 ranges = 1; 672 673 if (is_kexec_kdump) { 674 ranges = read_usm_ranges(&usm); 675 if (!ranges) /* there are no (base, size) duple */ 676 continue; 677 } 678 do { 679 if (is_kexec_kdump) { 680 base = read_n_cells(n_mem_addr_cells, &usm); 681 size = read_n_cells(n_mem_size_cells, &usm); 682 } 683 nid = of_drconf_to_nid_single(&drmem, &aa); 684 fake_numa_create_new_node( 685 ((base + size) >> PAGE_SHIFT), 686 &nid); 687 node_set_online(nid); 688 sz = numa_enforce_memory_limit(base, size); 689 if (sz) 690 add_active_range(nid, base >> PAGE_SHIFT, 691 (base >> PAGE_SHIFT) 692 + (sz >> PAGE_SHIFT)); 693 } while (--ranges); 694 } 695 } 696 697 static int __init parse_numa_properties(void) 698 { 699 struct device_node *cpu = NULL; 700 struct device_node *memory = NULL; 701 int default_nid = 0; 702 unsigned long i; 703 704 if (numa_enabled == 0) { 705 printk(KERN_WARNING "NUMA disabled by user\n"); 706 return -1; 707 } 708 709 min_common_depth = find_min_common_depth(); 710 711 if (min_common_depth < 0) 712 return min_common_depth; 713 714 dbg("NUMA associativity depth for CPU/Memory: %d\n", min_common_depth); 715 716 /* 717 * Even though we connect cpus to numa domains later in SMP 718 * init, we need to know the node ids now. This is because 719 * each node to be onlined must have NODE_DATA etc backing it. 720 */ 721 for_each_present_cpu(i) { 722 int nid; 723 724 cpu = of_get_cpu_node(i, NULL); 725 BUG_ON(!cpu); 726 nid = of_node_to_nid_single(cpu); 727 of_node_put(cpu); 728 729 /* 730 * Don't fall back to default_nid yet -- we will plug 731 * cpus into nodes once the memory scan has discovered 732 * the topology. 733 */ 734 if (nid < 0) 735 continue; 736 node_set_online(nid); 737 } 738 739 get_n_mem_cells(&n_mem_addr_cells, &n_mem_size_cells); 740 memory = NULL; 741 while ((memory = of_find_node_by_type(memory, "memory")) != NULL) { 742 unsigned long start; 743 unsigned long size; 744 int nid; 745 int ranges; 746 const unsigned int *memcell_buf; 747 unsigned int len; 748 749 memcell_buf = of_get_property(memory, 750 "linux,usable-memory", &len); 751 if (!memcell_buf || len <= 0) 752 memcell_buf = of_get_property(memory, "reg", &len); 753 if (!memcell_buf || len <= 0) 754 continue; 755 756 /* ranges in cell */ 757 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells); 758 new_range: 759 /* these are order-sensitive, and modify the buffer pointer */ 760 start = read_n_cells(n_mem_addr_cells, &memcell_buf); 761 size = read_n_cells(n_mem_size_cells, &memcell_buf); 762 763 /* 764 * Assumption: either all memory nodes or none will 765 * have associativity properties. If none, then 766 * everything goes to default_nid. 767 */ 768 nid = of_node_to_nid_single(memory); 769 if (nid < 0) 770 nid = default_nid; 771 772 fake_numa_create_new_node(((start + size) >> PAGE_SHIFT), &nid); 773 node_set_online(nid); 774 775 if (!(size = numa_enforce_memory_limit(start, size))) { 776 if (--ranges) 777 goto new_range; 778 else 779 continue; 780 } 781 782 add_active_range(nid, start >> PAGE_SHIFT, 783 (start >> PAGE_SHIFT) + (size >> PAGE_SHIFT)); 784 785 if (--ranges) 786 goto new_range; 787 } 788 789 /* 790 * Now do the same thing for each MEMBLOCK listed in the ibm,dynamic-memory 791 * property in the ibm,dynamic-reconfiguration-memory node. 792 */ 793 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory"); 794 if (memory) 795 parse_drconf_memory(memory); 796 797 return 0; 798 } 799 800 static void __init setup_nonnuma(void) 801 { 802 unsigned long top_of_ram = memblock_end_of_DRAM(); 803 unsigned long total_ram = memblock_phys_mem_size(); 804 unsigned long start_pfn, end_pfn; 805 unsigned int nid = 0; 806 struct memblock_region *reg; 807 808 printk(KERN_DEBUG "Top of RAM: 0x%lx, Total RAM: 0x%lx\n", 809 top_of_ram, total_ram); 810 printk(KERN_DEBUG "Memory hole size: %ldMB\n", 811 (top_of_ram - total_ram) >> 20); 812 813 for_each_memblock(memory, reg) { 814 start_pfn = memblock_region_memory_base_pfn(reg); 815 end_pfn = memblock_region_memory_end_pfn(reg); 816 817 fake_numa_create_new_node(end_pfn, &nid); 818 add_active_range(nid, start_pfn, end_pfn); 819 node_set_online(nid); 820 } 821 } 822 823 void __init dump_numa_cpu_topology(void) 824 { 825 unsigned int node; 826 unsigned int cpu, count; 827 828 if (min_common_depth == -1 || !numa_enabled) 829 return; 830 831 for_each_online_node(node) { 832 printk(KERN_DEBUG "Node %d CPUs:", node); 833 834 count = 0; 835 /* 836 * If we used a CPU iterator here we would miss printing 837 * the holes in the cpumap. 838 */ 839 for (cpu = 0; cpu < nr_cpu_ids; cpu++) { 840 if (cpumask_test_cpu(cpu, 841 node_to_cpumask_map[node])) { 842 if (count == 0) 843 printk(" %u", cpu); 844 ++count; 845 } else { 846 if (count > 1) 847 printk("-%u", cpu - 1); 848 count = 0; 849 } 850 } 851 852 if (count > 1) 853 printk("-%u", nr_cpu_ids - 1); 854 printk("\n"); 855 } 856 } 857 858 static void __init dump_numa_memory_topology(void) 859 { 860 unsigned int node; 861 unsigned int count; 862 863 if (min_common_depth == -1 || !numa_enabled) 864 return; 865 866 for_each_online_node(node) { 867 unsigned long i; 868 869 printk(KERN_DEBUG "Node %d Memory:", node); 870 871 count = 0; 872 873 for (i = 0; i < memblock_end_of_DRAM(); 874 i += (1 << SECTION_SIZE_BITS)) { 875 if (early_pfn_to_nid(i >> PAGE_SHIFT) == node) { 876 if (count == 0) 877 printk(" 0x%lx", i); 878 ++count; 879 } else { 880 if (count > 0) 881 printk("-0x%lx", i); 882 count = 0; 883 } 884 } 885 886 if (count > 0) 887 printk("-0x%lx", i); 888 printk("\n"); 889 } 890 } 891 892 /* 893 * Allocate some memory, satisfying the memblock or bootmem allocator where 894 * required. nid is the preferred node and end is the physical address of 895 * the highest address in the node. 896 * 897 * Returns the virtual address of the memory. 898 */ 899 static void __init *careful_zallocation(int nid, unsigned long size, 900 unsigned long align, 901 unsigned long end_pfn) 902 { 903 void *ret; 904 int new_nid; 905 unsigned long ret_paddr; 906 907 ret_paddr = __memblock_alloc_base(size, align, end_pfn << PAGE_SHIFT); 908 909 /* retry over all memory */ 910 if (!ret_paddr) 911 ret_paddr = __memblock_alloc_base(size, align, memblock_end_of_DRAM()); 912 913 if (!ret_paddr) 914 panic("numa.c: cannot allocate %lu bytes for node %d", 915 size, nid); 916 917 ret = __va(ret_paddr); 918 919 /* 920 * We initialize the nodes in numeric order: 0, 1, 2... 921 * and hand over control from the MEMBLOCK allocator to the 922 * bootmem allocator. If this function is called for 923 * node 5, then we know that all nodes <5 are using the 924 * bootmem allocator instead of the MEMBLOCK allocator. 925 * 926 * So, check the nid from which this allocation came 927 * and double check to see if we need to use bootmem 928 * instead of the MEMBLOCK. We don't free the MEMBLOCK memory 929 * since it would be useless. 930 */ 931 new_nid = early_pfn_to_nid(ret_paddr >> PAGE_SHIFT); 932 if (new_nid < nid) { 933 ret = __alloc_bootmem_node(NODE_DATA(new_nid), 934 size, align, 0); 935 936 dbg("alloc_bootmem %p %lx\n", ret, size); 937 } 938 939 memset(ret, 0, size); 940 return ret; 941 } 942 943 static struct notifier_block __cpuinitdata ppc64_numa_nb = { 944 .notifier_call = cpu_numa_callback, 945 .priority = 1 /* Must run before sched domains notifier. */ 946 }; 947 948 static void mark_reserved_regions_for_nid(int nid) 949 { 950 struct pglist_data *node = NODE_DATA(nid); 951 struct memblock_region *reg; 952 953 for_each_memblock(reserved, reg) { 954 unsigned long physbase = reg->base; 955 unsigned long size = reg->size; 956 unsigned long start_pfn = physbase >> PAGE_SHIFT; 957 unsigned long end_pfn = PFN_UP(physbase + size); 958 struct node_active_region node_ar; 959 unsigned long node_end_pfn = node->node_start_pfn + 960 node->node_spanned_pages; 961 962 /* 963 * Check to make sure that this memblock.reserved area is 964 * within the bounds of the node that we care about. 965 * Checking the nid of the start and end points is not 966 * sufficient because the reserved area could span the 967 * entire node. 968 */ 969 if (end_pfn <= node->node_start_pfn || 970 start_pfn >= node_end_pfn) 971 continue; 972 973 get_node_active_region(start_pfn, &node_ar); 974 while (start_pfn < end_pfn && 975 node_ar.start_pfn < node_ar.end_pfn) { 976 unsigned long reserve_size = size; 977 /* 978 * if reserved region extends past active region 979 * then trim size to active region 980 */ 981 if (end_pfn > node_ar.end_pfn) 982 reserve_size = (node_ar.end_pfn << PAGE_SHIFT) 983 - physbase; 984 /* 985 * Only worry about *this* node, others may not 986 * yet have valid NODE_DATA(). 987 */ 988 if (node_ar.nid == nid) { 989 dbg("reserve_bootmem %lx %lx nid=%d\n", 990 physbase, reserve_size, node_ar.nid); 991 reserve_bootmem_node(NODE_DATA(node_ar.nid), 992 physbase, reserve_size, 993 BOOTMEM_DEFAULT); 994 } 995 /* 996 * if reserved region is contained in the active region 997 * then done. 998 */ 999 if (end_pfn <= node_ar.end_pfn) 1000 break; 1001 1002 /* 1003 * reserved region extends past the active region 1004 * get next active region that contains this 1005 * reserved region 1006 */ 1007 start_pfn = node_ar.end_pfn; 1008 physbase = start_pfn << PAGE_SHIFT; 1009 size = size - reserve_size; 1010 get_node_active_region(start_pfn, &node_ar); 1011 } 1012 } 1013 } 1014 1015 1016 void __init do_init_bootmem(void) 1017 { 1018 int nid; 1019 1020 min_low_pfn = 0; 1021 max_low_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT; 1022 max_pfn = max_low_pfn; 1023 1024 if (parse_numa_properties()) 1025 setup_nonnuma(); 1026 else 1027 dump_numa_memory_topology(); 1028 1029 for_each_online_node(nid) { 1030 unsigned long start_pfn, end_pfn; 1031 void *bootmem_vaddr; 1032 unsigned long bootmap_pages; 1033 1034 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn); 1035 1036 /* 1037 * Allocate the node structure node local if possible 1038 * 1039 * Be careful moving this around, as it relies on all 1040 * previous nodes' bootmem to be initialized and have 1041 * all reserved areas marked. 1042 */ 1043 NODE_DATA(nid) = careful_zallocation(nid, 1044 sizeof(struct pglist_data), 1045 SMP_CACHE_BYTES, end_pfn); 1046 1047 dbg("node %d\n", nid); 1048 dbg("NODE_DATA() = %p\n", NODE_DATA(nid)); 1049 1050 NODE_DATA(nid)->bdata = &bootmem_node_data[nid]; 1051 NODE_DATA(nid)->node_start_pfn = start_pfn; 1052 NODE_DATA(nid)->node_spanned_pages = end_pfn - start_pfn; 1053 1054 if (NODE_DATA(nid)->node_spanned_pages == 0) 1055 continue; 1056 1057 dbg("start_paddr = %lx\n", start_pfn << PAGE_SHIFT); 1058 dbg("end_paddr = %lx\n", end_pfn << PAGE_SHIFT); 1059 1060 bootmap_pages = bootmem_bootmap_pages(end_pfn - start_pfn); 1061 bootmem_vaddr = careful_zallocation(nid, 1062 bootmap_pages << PAGE_SHIFT, 1063 PAGE_SIZE, end_pfn); 1064 1065 dbg("bootmap_vaddr = %p\n", bootmem_vaddr); 1066 1067 init_bootmem_node(NODE_DATA(nid), 1068 __pa(bootmem_vaddr) >> PAGE_SHIFT, 1069 start_pfn, end_pfn); 1070 1071 free_bootmem_with_active_regions(nid, end_pfn); 1072 /* 1073 * Be very careful about moving this around. Future 1074 * calls to careful_zallocation() depend on this getting 1075 * done correctly. 1076 */ 1077 mark_reserved_regions_for_nid(nid); 1078 sparse_memory_present_with_active_regions(nid); 1079 } 1080 1081 init_bootmem_done = 1; 1082 1083 /* 1084 * Now bootmem is initialised we can create the node to cpumask 1085 * lookup tables and setup the cpu callback to populate them. 1086 */ 1087 setup_node_to_cpumask_map(); 1088 1089 register_cpu_notifier(&ppc64_numa_nb); 1090 cpu_numa_callback(&ppc64_numa_nb, CPU_UP_PREPARE, 1091 (void *)(unsigned long)boot_cpuid); 1092 } 1093 1094 void __init paging_init(void) 1095 { 1096 unsigned long max_zone_pfns[MAX_NR_ZONES]; 1097 memset(max_zone_pfns, 0, sizeof(max_zone_pfns)); 1098 max_zone_pfns[ZONE_DMA] = memblock_end_of_DRAM() >> PAGE_SHIFT; 1099 free_area_init_nodes(max_zone_pfns); 1100 } 1101 1102 static int __init early_numa(char *p) 1103 { 1104 if (!p) 1105 return 0; 1106 1107 if (strstr(p, "off")) 1108 numa_enabled = 0; 1109 1110 if (strstr(p, "debug")) 1111 numa_debug = 1; 1112 1113 p = strstr(p, "fake="); 1114 if (p) 1115 cmdline = p + strlen("fake="); 1116 1117 return 0; 1118 } 1119 early_param("numa", early_numa); 1120 1121 #ifdef CONFIG_MEMORY_HOTPLUG 1122 /* 1123 * Find the node associated with a hot added memory section for 1124 * memory represented in the device tree by the property 1125 * ibm,dynamic-reconfiguration-memory/ibm,dynamic-memory. 1126 */ 1127 static int hot_add_drconf_scn_to_nid(struct device_node *memory, 1128 unsigned long scn_addr) 1129 { 1130 const u32 *dm; 1131 unsigned int drconf_cell_cnt, rc; 1132 unsigned long lmb_size; 1133 struct assoc_arrays aa; 1134 int nid = -1; 1135 1136 drconf_cell_cnt = of_get_drconf_memory(memory, &dm); 1137 if (!drconf_cell_cnt) 1138 return -1; 1139 1140 lmb_size = of_get_lmb_size(memory); 1141 if (!lmb_size) 1142 return -1; 1143 1144 rc = of_get_assoc_arrays(memory, &aa); 1145 if (rc) 1146 return -1; 1147 1148 for (; drconf_cell_cnt != 0; --drconf_cell_cnt) { 1149 struct of_drconf_cell drmem; 1150 1151 read_drconf_cell(&drmem, &dm); 1152 1153 /* skip this block if it is reserved or not assigned to 1154 * this partition */ 1155 if ((drmem.flags & DRCONF_MEM_RESERVED) 1156 || !(drmem.flags & DRCONF_MEM_ASSIGNED)) 1157 continue; 1158 1159 if ((scn_addr < drmem.base_addr) 1160 || (scn_addr >= (drmem.base_addr + lmb_size))) 1161 continue; 1162 1163 nid = of_drconf_to_nid_single(&drmem, &aa); 1164 break; 1165 } 1166 1167 return nid; 1168 } 1169 1170 /* 1171 * Find the node associated with a hot added memory section for memory 1172 * represented in the device tree as a node (i.e. memory@XXXX) for 1173 * each memblock. 1174 */ 1175 int hot_add_node_scn_to_nid(unsigned long scn_addr) 1176 { 1177 struct device_node *memory = NULL; 1178 int nid = -1; 1179 1180 while ((memory = of_find_node_by_type(memory, "memory")) != NULL) { 1181 unsigned long start, size; 1182 int ranges; 1183 const unsigned int *memcell_buf; 1184 unsigned int len; 1185 1186 memcell_buf = of_get_property(memory, "reg", &len); 1187 if (!memcell_buf || len <= 0) 1188 continue; 1189 1190 /* ranges in cell */ 1191 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells); 1192 1193 while (ranges--) { 1194 start = read_n_cells(n_mem_addr_cells, &memcell_buf); 1195 size = read_n_cells(n_mem_size_cells, &memcell_buf); 1196 1197 if ((scn_addr < start) || (scn_addr >= (start + size))) 1198 continue; 1199 1200 nid = of_node_to_nid_single(memory); 1201 break; 1202 } 1203 1204 of_node_put(memory); 1205 if (nid >= 0) 1206 break; 1207 } 1208 1209 return nid; 1210 } 1211 1212 /* 1213 * Find the node associated with a hot added memory section. Section 1214 * corresponds to a SPARSEMEM section, not an MEMBLOCK. It is assumed that 1215 * sections are fully contained within a single MEMBLOCK. 1216 */ 1217 int hot_add_scn_to_nid(unsigned long scn_addr) 1218 { 1219 struct device_node *memory = NULL; 1220 int nid, found = 0; 1221 1222 if (!numa_enabled || (min_common_depth < 0)) 1223 return first_online_node; 1224 1225 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory"); 1226 if (memory) { 1227 nid = hot_add_drconf_scn_to_nid(memory, scn_addr); 1228 of_node_put(memory); 1229 } else { 1230 nid = hot_add_node_scn_to_nid(scn_addr); 1231 } 1232 1233 if (nid < 0 || !node_online(nid)) 1234 nid = first_online_node; 1235 1236 if (NODE_DATA(nid)->node_spanned_pages) 1237 return nid; 1238 1239 for_each_online_node(nid) { 1240 if (NODE_DATA(nid)->node_spanned_pages) { 1241 found = 1; 1242 break; 1243 } 1244 } 1245 1246 BUG_ON(!found); 1247 return nid; 1248 } 1249 1250 #endif /* CONFIG_MEMORY_HOTPLUG */ 1251