1// SPDX-License-Identifier: GPL-2.0-or-later 2/* 3 * pSeries NUMA support 4 * 5 * Copyright (C) 2002 Anton Blanchard <anton@au.ibm.com>, IBM 6 */ 7#define pr_fmt(fmt) "numa: " fmt 8 9#include <linux/threads.h> 10#include <linux/memblock.h> 11#include <linux/init.h> 12#include <linux/mm.h> 13#include <linux/mmzone.h> 14#include <linux/export.h> 15#include <linux/nodemask.h> 16#include <linux/cpu.h> 17#include <linux/notifier.h> 18#include <linux/of.h> 19#include <linux/of_address.h> 20#include <linux/pfn.h> 21#include <linux/cpuset.h> 22#include <linux/node.h> 23#include <linux/stop_machine.h> 24#include <linux/proc_fs.h> 25#include <linux/seq_file.h> 26#include <linux/uaccess.h> 27#include <linux/slab.h> 28#include <asm/cputhreads.h> 29#include <asm/sparsemem.h> 30#include <asm/smp.h> 31#include <asm/topology.h> 32#include <asm/firmware.h> 33#include <asm/paca.h> 34#include <asm/hvcall.h> 35#include <asm/setup.h> 36#include <asm/vdso.h> 37#include <asm/vphn.h> 38#include <asm/drmem.h> 39 40static int numa_enabled = 1; 41 42static char *cmdline __initdata; 43 44int numa_cpu_lookup_table[NR_CPUS]; 45cpumask_var_t node_to_cpumask_map[MAX_NUMNODES]; 46struct pglist_data *node_data[MAX_NUMNODES]; 47 48EXPORT_SYMBOL(numa_cpu_lookup_table); 49EXPORT_SYMBOL(node_to_cpumask_map); 50EXPORT_SYMBOL(node_data); 51 52static int primary_domain_index; 53static int n_mem_addr_cells, n_mem_size_cells; 54 55#define FORM0_AFFINITY 0 56#define FORM1_AFFINITY 1 57#define FORM2_AFFINITY 2 58static int affinity_form; 59 60#define MAX_DISTANCE_REF_POINTS 4 61static int distance_ref_points_depth; 62static const __be32 *distance_ref_points; 63static int distance_lookup_table[MAX_NUMNODES][MAX_DISTANCE_REF_POINTS]; 64static int numa_distance_table[MAX_NUMNODES][MAX_NUMNODES] = { 65 [0 ... MAX_NUMNODES - 1] = { [0 ... MAX_NUMNODES - 1] = -1 } 66}; 67static int numa_id_index_table[MAX_NUMNODES] = { [0 ... MAX_NUMNODES - 1] = NUMA_NO_NODE }; 68 69/* 70 * Allocate node_to_cpumask_map based on number of available nodes 71 * Requires node_possible_map to be valid. 72 * 73 * Note: cpumask_of_node() is not valid until after this is done. 74 */ 75static void __init setup_node_to_cpumask_map(void) 76{ 77 unsigned int node; 78 79 /* setup nr_node_ids if not done yet */ 80 if (nr_node_ids == MAX_NUMNODES) 81 setup_nr_node_ids(); 82 83 /* allocate the map */ 84 for_each_node(node) 85 alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]); 86 87 /* cpumask_of_node() will now work */ 88 pr_debug("Node to cpumask map for %u nodes\n", nr_node_ids); 89} 90 91static int __init fake_numa_create_new_node(unsigned long end_pfn, 92 unsigned int *nid) 93{ 94 unsigned long long mem; 95 char *p = cmdline; 96 static unsigned int fake_nid; 97 static unsigned long long curr_boundary; 98 99 /* 100 * Modify node id, iff we started creating NUMA nodes 101 * We want to continue from where we left of the last time 102 */ 103 if (fake_nid) 104 *nid = fake_nid; 105 /* 106 * In case there are no more arguments to parse, the 107 * node_id should be the same as the last fake node id 108 * (we've handled this above). 109 */ 110 if (!p) 111 return 0; 112 113 mem = memparse(p, &p); 114 if (!mem) 115 return 0; 116 117 if (mem < curr_boundary) 118 return 0; 119 120 curr_boundary = mem; 121 122 if ((end_pfn << PAGE_SHIFT) > mem) { 123 /* 124 * Skip commas and spaces 125 */ 126 while (*p == ',' || *p == ' ' || *p == '\t') 127 p++; 128 129 cmdline = p; 130 fake_nid++; 131 *nid = fake_nid; 132 pr_debug("created new fake_node with id %d\n", fake_nid); 133 return 1; 134 } 135 return 0; 136} 137 138static void __init reset_numa_cpu_lookup_table(void) 139{ 140 unsigned int cpu; 141 142 for_each_possible_cpu(cpu) 143 numa_cpu_lookup_table[cpu] = -1; 144} 145 146void map_cpu_to_node(int cpu, int node) 147{ 148 update_numa_cpu_lookup_table(cpu, node); 149 150 if (!(cpumask_test_cpu(cpu, node_to_cpumask_map[node]))) { 151 pr_debug("adding cpu %d to node %d\n", cpu, node); 152 cpumask_set_cpu(cpu, node_to_cpumask_map[node]); 153 } 154} 155 156#if defined(CONFIG_HOTPLUG_CPU) || defined(CONFIG_PPC_SPLPAR) 157void unmap_cpu_from_node(unsigned long cpu) 158{ 159 int node = numa_cpu_lookup_table[cpu]; 160 161 if (cpumask_test_cpu(cpu, node_to_cpumask_map[node])) { 162 cpumask_clear_cpu(cpu, node_to_cpumask_map[node]); 163 pr_debug("removing cpu %lu from node %d\n", cpu, node); 164 } else { 165 pr_warn("Warning: cpu %lu not found in node %d\n", cpu, node); 166 } 167} 168#endif /* CONFIG_HOTPLUG_CPU || CONFIG_PPC_SPLPAR */ 169 170static int __associativity_to_nid(const __be32 *associativity, 171 int max_array_sz) 172{ 173 int nid; 174 /* 175 * primary_domain_index is 1 based array index. 176 */ 177 int index = primary_domain_index - 1; 178 179 if (!numa_enabled || index >= max_array_sz) 180 return NUMA_NO_NODE; 181 182 nid = of_read_number(&associativity[index], 1); 183 184 /* POWER4 LPAR uses 0xffff as invalid node */ 185 if (nid == 0xffff || nid >= nr_node_ids) 186 nid = NUMA_NO_NODE; 187 return nid; 188} 189/* 190 * Returns nid in the range [0..nr_node_ids], or -1 if no useful NUMA 191 * info is found. 192 */ 193static int associativity_to_nid(const __be32 *associativity) 194{ 195 int array_sz = of_read_number(associativity, 1); 196 197 /* Skip the first element in the associativity array */ 198 return __associativity_to_nid((associativity + 1), array_sz); 199} 200 201static int __cpu_form2_relative_distance(__be32 *cpu1_assoc, __be32 *cpu2_assoc) 202{ 203 int dist; 204 int node1, node2; 205 206 node1 = associativity_to_nid(cpu1_assoc); 207 node2 = associativity_to_nid(cpu2_assoc); 208 209 dist = numa_distance_table[node1][node2]; 210 if (dist <= LOCAL_DISTANCE) 211 return 0; 212 else if (dist <= REMOTE_DISTANCE) 213 return 1; 214 else 215 return 2; 216} 217 218static int __cpu_form1_relative_distance(__be32 *cpu1_assoc, __be32 *cpu2_assoc) 219{ 220 int dist = 0; 221 222 int i, index; 223 224 for (i = 0; i < distance_ref_points_depth; i++) { 225 index = be32_to_cpu(distance_ref_points[i]); 226 if (cpu1_assoc[index] == cpu2_assoc[index]) 227 break; 228 dist++; 229 } 230 231 return dist; 232} 233 234int cpu_relative_distance(__be32 *cpu1_assoc, __be32 *cpu2_assoc) 235{ 236 /* We should not get called with FORM0 */ 237 VM_WARN_ON(affinity_form == FORM0_AFFINITY); 238 if (affinity_form == FORM1_AFFINITY) 239 return __cpu_form1_relative_distance(cpu1_assoc, cpu2_assoc); 240 return __cpu_form2_relative_distance(cpu1_assoc, cpu2_assoc); 241} 242 243/* must hold reference to node during call */ 244static const __be32 *of_get_associativity(struct device_node *dev) 245{ 246 return of_get_property(dev, "ibm,associativity", NULL); 247} 248 249int __node_distance(int a, int b) 250{ 251 int i; 252 int distance = LOCAL_DISTANCE; 253 254 if (affinity_form == FORM2_AFFINITY) 255 return numa_distance_table[a][b]; 256 else if (affinity_form == FORM0_AFFINITY) 257 return ((a == b) ? LOCAL_DISTANCE : REMOTE_DISTANCE); 258 259 for (i = 0; i < distance_ref_points_depth; i++) { 260 if (distance_lookup_table[a][i] == distance_lookup_table[b][i]) 261 break; 262 263 /* Double the distance for each NUMA level */ 264 distance *= 2; 265 } 266 267 return distance; 268} 269EXPORT_SYMBOL(__node_distance); 270 271/* Returns the nid associated with the given device tree node, 272 * or -1 if not found. 273 */ 274static int of_node_to_nid_single(struct device_node *device) 275{ 276 int nid = NUMA_NO_NODE; 277 const __be32 *tmp; 278 279 tmp = of_get_associativity(device); 280 if (tmp) 281 nid = associativity_to_nid(tmp); 282 return nid; 283} 284 285/* Walk the device tree upwards, looking for an associativity id */ 286int of_node_to_nid(struct device_node *device) 287{ 288 int nid = NUMA_NO_NODE; 289 290 of_node_get(device); 291 while (device) { 292 nid = of_node_to_nid_single(device); 293 if (nid != -1) 294 break; 295 296 device = of_get_next_parent(device); 297 } 298 of_node_put(device); 299 300 return nid; 301} 302EXPORT_SYMBOL(of_node_to_nid); 303 304static void __initialize_form1_numa_distance(const __be32 *associativity, 305 int max_array_sz) 306{ 307 int i, nid; 308 309 if (affinity_form != FORM1_AFFINITY) 310 return; 311 312 nid = __associativity_to_nid(associativity, max_array_sz); 313 if (nid != NUMA_NO_NODE) { 314 for (i = 0; i < distance_ref_points_depth; i++) { 315 const __be32 *entry; 316 int index = be32_to_cpu(distance_ref_points[i]) - 1; 317 318 /* 319 * broken hierarchy, return with broken distance table 320 */ 321 if (WARN(index >= max_array_sz, "Broken ibm,associativity property")) 322 return; 323 324 entry = &associativity[index]; 325 distance_lookup_table[nid][i] = of_read_number(entry, 1); 326 } 327 } 328} 329 330static void initialize_form1_numa_distance(const __be32 *associativity) 331{ 332 int array_sz; 333 334 array_sz = of_read_number(associativity, 1); 335 /* Skip the first element in the associativity array */ 336 __initialize_form1_numa_distance(associativity + 1, array_sz); 337} 338 339/* 340 * Used to update distance information w.r.t newly added node. 341 */ 342void update_numa_distance(struct device_node *node) 343{ 344 int nid; 345 346 if (affinity_form == FORM0_AFFINITY) 347 return; 348 else if (affinity_form == FORM1_AFFINITY) { 349 const __be32 *associativity; 350 351 associativity = of_get_associativity(node); 352 if (!associativity) 353 return; 354 355 initialize_form1_numa_distance(associativity); 356 return; 357 } 358 359 /* FORM2 affinity */ 360 nid = of_node_to_nid_single(node); 361 if (nid == NUMA_NO_NODE) 362 return; 363 364 /* 365 * With FORM2 we expect NUMA distance of all possible NUMA 366 * nodes to be provided during boot. 367 */ 368 WARN(numa_distance_table[nid][nid] == -1, 369 "NUMA distance details for node %d not provided\n", nid); 370} 371EXPORT_SYMBOL_GPL(update_numa_distance); 372 373/* 374 * ibm,numa-lookup-index-table= {N, domainid1, domainid2, ..... domainidN} 375 * ibm,numa-distance-table = { N, 1, 2, 4, 5, 1, 6, .... N elements} 376 */ 377static void __init initialize_form2_numa_distance_lookup_table(void) 378{ 379 int i, j; 380 struct device_node *root; 381 const __u8 *form2_distances; 382 const __be32 *numa_lookup_index; 383 int form2_distances_length; 384 int max_numa_index, distance_index; 385 386 if (firmware_has_feature(FW_FEATURE_OPAL)) 387 root = of_find_node_by_path("/ibm,opal"); 388 else 389 root = of_find_node_by_path("/rtas"); 390 if (!root) 391 root = of_find_node_by_path("/"); 392 393 numa_lookup_index = of_get_property(root, "ibm,numa-lookup-index-table", NULL); 394 max_numa_index = of_read_number(&numa_lookup_index[0], 1); 395 396 /* first element of the array is the size and is encode-int */ 397 form2_distances = of_get_property(root, "ibm,numa-distance-table", NULL); 398 form2_distances_length = of_read_number((const __be32 *)&form2_distances[0], 1); 399 /* Skip the size which is encoded int */ 400 form2_distances += sizeof(__be32); 401 402 pr_debug("form2_distances_len = %d, numa_dist_indexes_len = %d\n", 403 form2_distances_length, max_numa_index); 404 405 for (i = 0; i < max_numa_index; i++) 406 /* +1 skip the max_numa_index in the property */ 407 numa_id_index_table[i] = of_read_number(&numa_lookup_index[i + 1], 1); 408 409 410 if (form2_distances_length != max_numa_index * max_numa_index) { 411 WARN(1, "Wrong NUMA distance information\n"); 412 form2_distances = NULL; // don't use it 413 } 414 distance_index = 0; 415 for (i = 0; i < max_numa_index; i++) { 416 for (j = 0; j < max_numa_index; j++) { 417 int nodeA = numa_id_index_table[i]; 418 int nodeB = numa_id_index_table[j]; 419 int dist; 420 421 if (form2_distances) 422 dist = form2_distances[distance_index++]; 423 else if (nodeA == nodeB) 424 dist = LOCAL_DISTANCE; 425 else 426 dist = REMOTE_DISTANCE; 427 numa_distance_table[nodeA][nodeB] = dist; 428 pr_debug("dist[%d][%d]=%d ", nodeA, nodeB, dist); 429 } 430 } 431 432 of_node_put(root); 433} 434 435static int __init find_primary_domain_index(void) 436{ 437 int index; 438 struct device_node *root; 439 440 /* 441 * Check for which form of affinity. 442 */ 443 if (firmware_has_feature(FW_FEATURE_OPAL)) { 444 affinity_form = FORM1_AFFINITY; 445 } else if (firmware_has_feature(FW_FEATURE_FORM2_AFFINITY)) { 446 pr_debug("Using form 2 affinity\n"); 447 affinity_form = FORM2_AFFINITY; 448 } else if (firmware_has_feature(FW_FEATURE_FORM1_AFFINITY)) { 449 pr_debug("Using form 1 affinity\n"); 450 affinity_form = FORM1_AFFINITY; 451 } else 452 affinity_form = FORM0_AFFINITY; 453 454 if (firmware_has_feature(FW_FEATURE_OPAL)) 455 root = of_find_node_by_path("/ibm,opal"); 456 else 457 root = of_find_node_by_path("/rtas"); 458 if (!root) 459 root = of_find_node_by_path("/"); 460 461 /* 462 * This property is a set of 32-bit integers, each representing 463 * an index into the ibm,associativity nodes. 464 * 465 * With form 0 affinity the first integer is for an SMP configuration 466 * (should be all 0's) and the second is for a normal NUMA 467 * configuration. We have only one level of NUMA. 468 * 469 * With form 1 affinity the first integer is the most significant 470 * NUMA boundary and the following are progressively less significant 471 * boundaries. There can be more than one level of NUMA. 472 */ 473 distance_ref_points = of_get_property(root, 474 "ibm,associativity-reference-points", 475 &distance_ref_points_depth); 476 477 if (!distance_ref_points) { 478 pr_debug("ibm,associativity-reference-points not found.\n"); 479 goto err; 480 } 481 482 distance_ref_points_depth /= sizeof(int); 483 if (affinity_form == FORM0_AFFINITY) { 484 if (distance_ref_points_depth < 2) { 485 pr_warn("short ibm,associativity-reference-points\n"); 486 goto err; 487 } 488 489 index = of_read_number(&distance_ref_points[1], 1); 490 } else { 491 /* 492 * Both FORM1 and FORM2 affinity find the primary domain details 493 * at the same offset. 494 */ 495 index = of_read_number(distance_ref_points, 1); 496 } 497 /* 498 * Warn and cap if the hardware supports more than 499 * MAX_DISTANCE_REF_POINTS domains. 500 */ 501 if (distance_ref_points_depth > MAX_DISTANCE_REF_POINTS) { 502 pr_warn("distance array capped at %d entries\n", 503 MAX_DISTANCE_REF_POINTS); 504 distance_ref_points_depth = MAX_DISTANCE_REF_POINTS; 505 } 506 507 of_node_put(root); 508 return index; 509 510err: 511 of_node_put(root); 512 return -1; 513} 514 515static void __init get_n_mem_cells(int *n_addr_cells, int *n_size_cells) 516{ 517 struct device_node *memory = NULL; 518 519 memory = of_find_node_by_type(memory, "memory"); 520 if (!memory) 521 panic("numa.c: No memory nodes found!"); 522 523 *n_addr_cells = of_n_addr_cells(memory); 524 *n_size_cells = of_n_size_cells(memory); 525 of_node_put(memory); 526} 527 528static unsigned long read_n_cells(int n, const __be32 **buf) 529{ 530 unsigned long result = 0; 531 532 while (n--) { 533 result = (result << 32) | of_read_number(*buf, 1); 534 (*buf)++; 535 } 536 return result; 537} 538 539struct assoc_arrays { 540 u32 n_arrays; 541 u32 array_sz; 542 const __be32 *arrays; 543}; 544 545/* 546 * Retrieve and validate the list of associativity arrays for drconf 547 * memory from the ibm,associativity-lookup-arrays property of the 548 * device tree.. 549 * 550 * The layout of the ibm,associativity-lookup-arrays property is a number N 551 * indicating the number of associativity arrays, followed by a number M 552 * indicating the size of each associativity array, followed by a list 553 * of N associativity arrays. 554 */ 555static int of_get_assoc_arrays(struct assoc_arrays *aa) 556{ 557 struct device_node *memory; 558 const __be32 *prop; 559 u32 len; 560 561 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory"); 562 if (!memory) 563 return -1; 564 565 prop = of_get_property(memory, "ibm,associativity-lookup-arrays", &len); 566 if (!prop || len < 2 * sizeof(unsigned int)) { 567 of_node_put(memory); 568 return -1; 569 } 570 571 aa->n_arrays = of_read_number(prop++, 1); 572 aa->array_sz = of_read_number(prop++, 1); 573 574 of_node_put(memory); 575 576 /* Now that we know the number of arrays and size of each array, 577 * revalidate the size of the property read in. 578 */ 579 if (len < (aa->n_arrays * aa->array_sz + 2) * sizeof(unsigned int)) 580 return -1; 581 582 aa->arrays = prop; 583 return 0; 584} 585 586static int __init get_nid_and_numa_distance(struct drmem_lmb *lmb) 587{ 588 struct assoc_arrays aa = { .arrays = NULL }; 589 int default_nid = NUMA_NO_NODE; 590 int nid = default_nid; 591 int rc, index; 592 593 if ((primary_domain_index < 0) || !numa_enabled) 594 return default_nid; 595 596 rc = of_get_assoc_arrays(&aa); 597 if (rc) 598 return default_nid; 599 600 if (primary_domain_index <= aa.array_sz && 601 !(lmb->flags & DRCONF_MEM_AI_INVALID) && lmb->aa_index < aa.n_arrays) { 602 const __be32 *associativity; 603 604 index = lmb->aa_index * aa.array_sz; 605 associativity = &aa.arrays[index]; 606 nid = __associativity_to_nid(associativity, aa.array_sz); 607 if (nid > 0 && affinity_form == FORM1_AFFINITY) { 608 /* 609 * lookup array associativity entries have 610 * no length of the array as the first element. 611 */ 612 __initialize_form1_numa_distance(associativity, aa.array_sz); 613 } 614 } 615 return nid; 616} 617 618/* 619 * This is like of_node_to_nid_single() for memory represented in the 620 * ibm,dynamic-reconfiguration-memory node. 621 */ 622int of_drconf_to_nid_single(struct drmem_lmb *lmb) 623{ 624 struct assoc_arrays aa = { .arrays = NULL }; 625 int default_nid = NUMA_NO_NODE; 626 int nid = default_nid; 627 int rc, index; 628 629 if ((primary_domain_index < 0) || !numa_enabled) 630 return default_nid; 631 632 rc = of_get_assoc_arrays(&aa); 633 if (rc) 634 return default_nid; 635 636 if (primary_domain_index <= aa.array_sz && 637 !(lmb->flags & DRCONF_MEM_AI_INVALID) && lmb->aa_index < aa.n_arrays) { 638 const __be32 *associativity; 639 640 index = lmb->aa_index * aa.array_sz; 641 associativity = &aa.arrays[index]; 642 nid = __associativity_to_nid(associativity, aa.array_sz); 643 } 644 return nid; 645} 646 647#ifdef CONFIG_PPC_SPLPAR 648 649static int __vphn_get_associativity(long lcpu, __be32 *associativity) 650{ 651 long rc, hwid; 652 653 /* 654 * On a shared lpar, device tree will not have node associativity. 655 * At this time lppaca, or its __old_status field may not be 656 * updated. Hence kernel cannot detect if its on a shared lpar. So 657 * request an explicit associativity irrespective of whether the 658 * lpar is shared or dedicated. Use the device tree property as a 659 * fallback. cpu_to_phys_id is only valid between 660 * smp_setup_cpu_maps() and smp_setup_pacas(). 661 */ 662 if (firmware_has_feature(FW_FEATURE_VPHN)) { 663 if (cpu_to_phys_id) 664 hwid = cpu_to_phys_id[lcpu]; 665 else 666 hwid = get_hard_smp_processor_id(lcpu); 667 668 rc = hcall_vphn(hwid, VPHN_FLAG_VCPU, associativity); 669 if (rc == H_SUCCESS) 670 return 0; 671 } 672 673 return -1; 674} 675 676static int vphn_get_nid(long lcpu) 677{ 678 __be32 associativity[VPHN_ASSOC_BUFSIZE] = {0}; 679 680 681 if (!__vphn_get_associativity(lcpu, associativity)) 682 return associativity_to_nid(associativity); 683 684 return NUMA_NO_NODE; 685 686} 687#else 688 689static int __vphn_get_associativity(long lcpu, __be32 *associativity) 690{ 691 return -1; 692} 693 694static int vphn_get_nid(long unused) 695{ 696 return NUMA_NO_NODE; 697} 698#endif /* CONFIG_PPC_SPLPAR */ 699 700/* 701 * Figure out to which domain a cpu belongs and stick it there. 702 * Return the id of the domain used. 703 */ 704static int numa_setup_cpu(unsigned long lcpu) 705{ 706 struct device_node *cpu; 707 int fcpu = cpu_first_thread_sibling(lcpu); 708 int nid = NUMA_NO_NODE; 709 710 if (!cpu_present(lcpu)) { 711 set_cpu_numa_node(lcpu, first_online_node); 712 return first_online_node; 713 } 714 715 /* 716 * If a valid cpu-to-node mapping is already available, use it 717 * directly instead of querying the firmware, since it represents 718 * the most recent mapping notified to us by the platform (eg: VPHN). 719 * Since cpu_to_node binding remains the same for all threads in the 720 * core. If a valid cpu-to-node mapping is already available, for 721 * the first thread in the core, use it. 722 */ 723 nid = numa_cpu_lookup_table[fcpu]; 724 if (nid >= 0) { 725 map_cpu_to_node(lcpu, nid); 726 return nid; 727 } 728 729 nid = vphn_get_nid(lcpu); 730 if (nid != NUMA_NO_NODE) 731 goto out_present; 732 733 cpu = of_get_cpu_node(lcpu, NULL); 734 735 if (!cpu) { 736 WARN_ON(1); 737 if (cpu_present(lcpu)) 738 goto out_present; 739 else 740 goto out; 741 } 742 743 nid = of_node_to_nid_single(cpu); 744 of_node_put(cpu); 745 746out_present: 747 if (nid < 0 || !node_possible(nid)) 748 nid = first_online_node; 749 750 /* 751 * Update for the first thread of the core. All threads of a core 752 * have to be part of the same node. This not only avoids querying 753 * for every other thread in the core, but always avoids a case 754 * where virtual node associativity change causes subsequent threads 755 * of a core to be associated with different nid. However if first 756 * thread is already online, expect it to have a valid mapping. 757 */ 758 if (fcpu != lcpu) { 759 WARN_ON(cpu_online(fcpu)); 760 map_cpu_to_node(fcpu, nid); 761 } 762 763 map_cpu_to_node(lcpu, nid); 764out: 765 return nid; 766} 767 768static void verify_cpu_node_mapping(int cpu, int node) 769{ 770 int base, sibling, i; 771 772 /* Verify that all the threads in the core belong to the same node */ 773 base = cpu_first_thread_sibling(cpu); 774 775 for (i = 0; i < threads_per_core; i++) { 776 sibling = base + i; 777 778 if (sibling == cpu || cpu_is_offline(sibling)) 779 continue; 780 781 if (cpu_to_node(sibling) != node) { 782 WARN(1, "CPU thread siblings %d and %d don't belong" 783 " to the same node!\n", cpu, sibling); 784 break; 785 } 786 } 787} 788 789/* Must run before sched domains notifier. */ 790static int ppc_numa_cpu_prepare(unsigned int cpu) 791{ 792 int nid; 793 794 nid = numa_setup_cpu(cpu); 795 verify_cpu_node_mapping(cpu, nid); 796 return 0; 797} 798 799static int ppc_numa_cpu_dead(unsigned int cpu) 800{ 801 return 0; 802} 803 804/* 805 * Check and possibly modify a memory region to enforce the memory limit. 806 * 807 * Returns the size the region should have to enforce the memory limit. 808 * This will either be the original value of size, a truncated value, 809 * or zero. If the returned value of size is 0 the region should be 810 * discarded as it lies wholly above the memory limit. 811 */ 812static unsigned long __init numa_enforce_memory_limit(unsigned long start, 813 unsigned long size) 814{ 815 /* 816 * We use memblock_end_of_DRAM() in here instead of memory_limit because 817 * we've already adjusted it for the limit and it takes care of 818 * having memory holes below the limit. Also, in the case of 819 * iommu_is_off, memory_limit is not set but is implicitly enforced. 820 */ 821 822 if (start + size <= memblock_end_of_DRAM()) 823 return size; 824 825 if (start >= memblock_end_of_DRAM()) 826 return 0; 827 828 return memblock_end_of_DRAM() - start; 829} 830 831/* 832 * Reads the counter for a given entry in 833 * linux,drconf-usable-memory property 834 */ 835static inline int __init read_usm_ranges(const __be32 **usm) 836{ 837 /* 838 * For each lmb in ibm,dynamic-memory a corresponding 839 * entry in linux,drconf-usable-memory property contains 840 * a counter followed by that many (base, size) duple. 841 * read the counter from linux,drconf-usable-memory 842 */ 843 return read_n_cells(n_mem_size_cells, usm); 844} 845 846/* 847 * Extract NUMA information from the ibm,dynamic-reconfiguration-memory 848 * node. This assumes n_mem_{addr,size}_cells have been set. 849 */ 850static int __init numa_setup_drmem_lmb(struct drmem_lmb *lmb, 851 const __be32 **usm, 852 void *data) 853{ 854 unsigned int ranges, is_kexec_kdump = 0; 855 unsigned long base, size, sz; 856 int nid; 857 858 /* 859 * Skip this block if the reserved bit is set in flags (0x80) 860 * or if the block is not assigned to this partition (0x8) 861 */ 862 if ((lmb->flags & DRCONF_MEM_RESERVED) 863 || !(lmb->flags & DRCONF_MEM_ASSIGNED)) 864 return 0; 865 866 if (*usm) 867 is_kexec_kdump = 1; 868 869 base = lmb->base_addr; 870 size = drmem_lmb_size(); 871 ranges = 1; 872 873 if (is_kexec_kdump) { 874 ranges = read_usm_ranges(usm); 875 if (!ranges) /* there are no (base, size) duple */ 876 return 0; 877 } 878 879 do { 880 if (is_kexec_kdump) { 881 base = read_n_cells(n_mem_addr_cells, usm); 882 size = read_n_cells(n_mem_size_cells, usm); 883 } 884 885 nid = get_nid_and_numa_distance(lmb); 886 fake_numa_create_new_node(((base + size) >> PAGE_SHIFT), 887 &nid); 888 node_set_online(nid); 889 sz = numa_enforce_memory_limit(base, size); 890 if (sz) 891 memblock_set_node(base, sz, &memblock.memory, nid); 892 } while (--ranges); 893 894 return 0; 895} 896 897static int __init parse_numa_properties(void) 898{ 899 struct device_node *memory; 900 int default_nid = 0; 901 unsigned long i; 902 const __be32 *associativity; 903 904 if (numa_enabled == 0) { 905 pr_warn("disabled by user\n"); 906 return -1; 907 } 908 909 primary_domain_index = find_primary_domain_index(); 910 911 if (primary_domain_index < 0) { 912 /* 913 * if we fail to parse primary_domain_index from device tree 914 * mark the numa disabled, boot with numa disabled. 915 */ 916 numa_enabled = false; 917 return primary_domain_index; 918 } 919 920 pr_debug("associativity depth for CPU/Memory: %d\n", primary_domain_index); 921 922 /* 923 * If it is FORM2 initialize the distance table here. 924 */ 925 if (affinity_form == FORM2_AFFINITY) 926 initialize_form2_numa_distance_lookup_table(); 927 928 /* 929 * Even though we connect cpus to numa domains later in SMP 930 * init, we need to know the node ids now. This is because 931 * each node to be onlined must have NODE_DATA etc backing it. 932 */ 933 for_each_present_cpu(i) { 934 __be32 vphn_assoc[VPHN_ASSOC_BUFSIZE]; 935 struct device_node *cpu; 936 int nid = NUMA_NO_NODE; 937 938 memset(vphn_assoc, 0, VPHN_ASSOC_BUFSIZE * sizeof(__be32)); 939 940 if (__vphn_get_associativity(i, vphn_assoc) == 0) { 941 nid = associativity_to_nid(vphn_assoc); 942 initialize_form1_numa_distance(vphn_assoc); 943 } else { 944 945 /* 946 * Don't fall back to default_nid yet -- we will plug 947 * cpus into nodes once the memory scan has discovered 948 * the topology. 949 */ 950 cpu = of_get_cpu_node(i, NULL); 951 BUG_ON(!cpu); 952 953 associativity = of_get_associativity(cpu); 954 if (associativity) { 955 nid = associativity_to_nid(associativity); 956 initialize_form1_numa_distance(associativity); 957 } 958 of_node_put(cpu); 959 } 960 961 /* node_set_online() is an UB if 'nid' is negative */ 962 if (likely(nid >= 0)) 963 node_set_online(nid); 964 } 965 966 get_n_mem_cells(&n_mem_addr_cells, &n_mem_size_cells); 967 968 for_each_node_by_type(memory, "memory") { 969 unsigned long start; 970 unsigned long size; 971 int nid; 972 int ranges; 973 const __be32 *memcell_buf; 974 unsigned int len; 975 976 memcell_buf = of_get_property(memory, 977 "linux,usable-memory", &len); 978 if (!memcell_buf || len <= 0) 979 memcell_buf = of_get_property(memory, "reg", &len); 980 if (!memcell_buf || len <= 0) 981 continue; 982 983 /* ranges in cell */ 984 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells); 985new_range: 986 /* these are order-sensitive, and modify the buffer pointer */ 987 start = read_n_cells(n_mem_addr_cells, &memcell_buf); 988 size = read_n_cells(n_mem_size_cells, &memcell_buf); 989 990 /* 991 * Assumption: either all memory nodes or none will 992 * have associativity properties. If none, then 993 * everything goes to default_nid. 994 */ 995 associativity = of_get_associativity(memory); 996 if (associativity) { 997 nid = associativity_to_nid(associativity); 998 initialize_form1_numa_distance(associativity); 999 } else 1000 nid = default_nid; 1001 1002 fake_numa_create_new_node(((start + size) >> PAGE_SHIFT), &nid); 1003 node_set_online(nid); 1004 1005 size = numa_enforce_memory_limit(start, size); 1006 if (size) 1007 memblock_set_node(start, size, &memblock.memory, nid); 1008 1009 if (--ranges) 1010 goto new_range; 1011 } 1012 1013 /* 1014 * Now do the same thing for each MEMBLOCK listed in the 1015 * ibm,dynamic-memory property in the 1016 * ibm,dynamic-reconfiguration-memory node. 1017 */ 1018 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory"); 1019 if (memory) { 1020 walk_drmem_lmbs(memory, NULL, numa_setup_drmem_lmb); 1021 of_node_put(memory); 1022 } 1023 1024 return 0; 1025} 1026 1027static void __init setup_nonnuma(void) 1028{ 1029 unsigned long top_of_ram = memblock_end_of_DRAM(); 1030 unsigned long total_ram = memblock_phys_mem_size(); 1031 unsigned long start_pfn, end_pfn; 1032 unsigned int nid = 0; 1033 int i; 1034 1035 pr_debug("Top of RAM: 0x%lx, Total RAM: 0x%lx\n", top_of_ram, total_ram); 1036 pr_debug("Memory hole size: %ldMB\n", (top_of_ram - total_ram) >> 20); 1037 1038 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, NULL) { 1039 fake_numa_create_new_node(end_pfn, &nid); 1040 memblock_set_node(PFN_PHYS(start_pfn), 1041 PFN_PHYS(end_pfn - start_pfn), 1042 &memblock.memory, nid); 1043 node_set_online(nid); 1044 } 1045} 1046 1047void __init dump_numa_cpu_topology(void) 1048{ 1049 unsigned int node; 1050 unsigned int cpu, count; 1051 1052 if (!numa_enabled) 1053 return; 1054 1055 for_each_online_node(node) { 1056 pr_info("Node %d CPUs:", node); 1057 1058 count = 0; 1059 /* 1060 * If we used a CPU iterator here we would miss printing 1061 * the holes in the cpumap. 1062 */ 1063 for (cpu = 0; cpu < nr_cpu_ids; cpu++) { 1064 if (cpumask_test_cpu(cpu, 1065 node_to_cpumask_map[node])) { 1066 if (count == 0) 1067 pr_cont(" %u", cpu); 1068 ++count; 1069 } else { 1070 if (count > 1) 1071 pr_cont("-%u", cpu - 1); 1072 count = 0; 1073 } 1074 } 1075 1076 if (count > 1) 1077 pr_cont("-%u", nr_cpu_ids - 1); 1078 pr_cont("\n"); 1079 } 1080} 1081 1082/* Initialize NODE_DATA for a node on the local memory */ 1083static void __init setup_node_data(int nid, u64 start_pfn, u64 end_pfn) 1084{ 1085 u64 spanned_pages = end_pfn - start_pfn; 1086 const size_t nd_size = roundup(sizeof(pg_data_t), SMP_CACHE_BYTES); 1087 u64 nd_pa; 1088 void *nd; 1089 int tnid; 1090 1091 nd_pa = memblock_phys_alloc_try_nid(nd_size, SMP_CACHE_BYTES, nid); 1092 if (!nd_pa) 1093 panic("Cannot allocate %zu bytes for node %d data\n", 1094 nd_size, nid); 1095 1096 nd = __va(nd_pa); 1097 1098 /* report and initialize */ 1099 pr_info(" NODE_DATA [mem %#010Lx-%#010Lx]\n", 1100 nd_pa, nd_pa + nd_size - 1); 1101 tnid = early_pfn_to_nid(nd_pa >> PAGE_SHIFT); 1102 if (tnid != nid) 1103 pr_info(" NODE_DATA(%d) on node %d\n", nid, tnid); 1104 1105 node_data[nid] = nd; 1106 memset(NODE_DATA(nid), 0, sizeof(pg_data_t)); 1107 NODE_DATA(nid)->node_id = nid; 1108 NODE_DATA(nid)->node_start_pfn = start_pfn; 1109 NODE_DATA(nid)->node_spanned_pages = spanned_pages; 1110} 1111 1112static void __init find_possible_nodes(void) 1113{ 1114 struct device_node *rtas, *root; 1115 const __be32 *domains = NULL; 1116 int prop_length, max_nodes; 1117 u32 i; 1118 1119 if (!numa_enabled) 1120 return; 1121 1122 rtas = of_find_node_by_path("/rtas"); 1123 if (!rtas) 1124 return; 1125 1126 /* 1127 * ibm,current-associativity-domains is a fairly recent property. If 1128 * it doesn't exist, then fallback on ibm,max-associativity-domains. 1129 * Current denotes what the platform can support compared to max 1130 * which denotes what the Hypervisor can support. 1131 * 1132 * If the LPAR is migratable, new nodes might be activated after a LPM, 1133 * so we should consider the max number in that case. 1134 */ 1135 root = of_find_node_by_path("/"); 1136 if (!of_get_property(root, "ibm,migratable-partition", NULL)) 1137 domains = of_get_property(rtas, 1138 "ibm,current-associativity-domains", 1139 &prop_length); 1140 of_node_put(root); 1141 if (!domains) { 1142 domains = of_get_property(rtas, "ibm,max-associativity-domains", 1143 &prop_length); 1144 if (!domains) 1145 goto out; 1146 } 1147 1148 max_nodes = of_read_number(&domains[primary_domain_index], 1); 1149 pr_info("Partition configured for %d NUMA nodes.\n", max_nodes); 1150 1151 for (i = 0; i < max_nodes; i++) { 1152 if (!node_possible(i)) 1153 node_set(i, node_possible_map); 1154 } 1155 1156 prop_length /= sizeof(int); 1157 if (prop_length > primary_domain_index + 2) 1158 coregroup_enabled = 1; 1159 1160out: 1161 of_node_put(rtas); 1162} 1163 1164void __init mem_topology_setup(void) 1165{ 1166 int cpu; 1167 1168 max_low_pfn = max_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT; 1169 min_low_pfn = MEMORY_START >> PAGE_SHIFT; 1170 1171 /* 1172 * Linux/mm assumes node 0 to be online at boot. However this is not 1173 * true on PowerPC, where node 0 is similar to any other node, it 1174 * could be cpuless, memoryless node. So force node 0 to be offline 1175 * for now. This will prevent cpuless, memoryless node 0 showing up 1176 * unnecessarily as online. If a node has cpus or memory that need 1177 * to be online, then node will anyway be marked online. 1178 */ 1179 node_set_offline(0); 1180 1181 if (parse_numa_properties()) 1182 setup_nonnuma(); 1183 1184 /* 1185 * Modify the set of possible NUMA nodes to reflect information 1186 * available about the set of online nodes, and the set of nodes 1187 * that we expect to make use of for this platform's affinity 1188 * calculations. 1189 */ 1190 nodes_and(node_possible_map, node_possible_map, node_online_map); 1191 1192 find_possible_nodes(); 1193 1194 setup_node_to_cpumask_map(); 1195 1196 reset_numa_cpu_lookup_table(); 1197 1198 for_each_possible_cpu(cpu) { 1199 /* 1200 * Powerpc with CONFIG_NUMA always used to have a node 0, 1201 * even if it was memoryless or cpuless. For all cpus that 1202 * are possible but not present, cpu_to_node() would point 1203 * to node 0. To remove a cpuless, memoryless dummy node, 1204 * powerpc need to make sure all possible but not present 1205 * cpu_to_node are set to a proper node. 1206 */ 1207 numa_setup_cpu(cpu); 1208 } 1209} 1210 1211void __init initmem_init(void) 1212{ 1213 int nid; 1214 1215 memblock_dump_all(); 1216 1217 for_each_online_node(nid) { 1218 unsigned long start_pfn, end_pfn; 1219 1220 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn); 1221 setup_node_data(nid, start_pfn, end_pfn); 1222 } 1223 1224 sparse_init(); 1225 1226 /* 1227 * We need the numa_cpu_lookup_table to be accurate for all CPUs, 1228 * even before we online them, so that we can use cpu_to_{node,mem} 1229 * early in boot, cf. smp_prepare_cpus(). 1230 * _nocalls() + manual invocation is used because cpuhp is not yet 1231 * initialized for the boot CPU. 1232 */ 1233 cpuhp_setup_state_nocalls(CPUHP_POWER_NUMA_PREPARE, "powerpc/numa:prepare", 1234 ppc_numa_cpu_prepare, ppc_numa_cpu_dead); 1235} 1236 1237static int __init early_numa(char *p) 1238{ 1239 if (!p) 1240 return 0; 1241 1242 if (strstr(p, "off")) 1243 numa_enabled = 0; 1244 1245 p = strstr(p, "fake="); 1246 if (p) 1247 cmdline = p + strlen("fake="); 1248 1249 return 0; 1250} 1251early_param("numa", early_numa); 1252 1253#ifdef CONFIG_MEMORY_HOTPLUG 1254/* 1255 * Find the node associated with a hot added memory section for 1256 * memory represented in the device tree by the property 1257 * ibm,dynamic-reconfiguration-memory/ibm,dynamic-memory. 1258 */ 1259static int hot_add_drconf_scn_to_nid(unsigned long scn_addr) 1260{ 1261 struct drmem_lmb *lmb; 1262 unsigned long lmb_size; 1263 int nid = NUMA_NO_NODE; 1264 1265 lmb_size = drmem_lmb_size(); 1266 1267 for_each_drmem_lmb(lmb) { 1268 /* skip this block if it is reserved or not assigned to 1269 * this partition */ 1270 if ((lmb->flags & DRCONF_MEM_RESERVED) 1271 || !(lmb->flags & DRCONF_MEM_ASSIGNED)) 1272 continue; 1273 1274 if ((scn_addr < lmb->base_addr) 1275 || (scn_addr >= (lmb->base_addr + lmb_size))) 1276 continue; 1277 1278 nid = of_drconf_to_nid_single(lmb); 1279 break; 1280 } 1281 1282 return nid; 1283} 1284 1285/* 1286 * Find the node associated with a hot added memory section for memory 1287 * represented in the device tree as a node (i.e. memory@XXXX) for 1288 * each memblock. 1289 */ 1290static int hot_add_node_scn_to_nid(unsigned long scn_addr) 1291{ 1292 struct device_node *memory; 1293 int nid = NUMA_NO_NODE; 1294 1295 for_each_node_by_type(memory, "memory") { 1296 int i = 0; 1297 1298 while (1) { 1299 struct resource res; 1300 1301 if (of_address_to_resource(memory, i++, &res)) 1302 break; 1303 1304 if ((scn_addr < res.start) || (scn_addr > res.end)) 1305 continue; 1306 1307 nid = of_node_to_nid_single(memory); 1308 break; 1309 } 1310 1311 if (nid >= 0) 1312 break; 1313 } 1314 1315 of_node_put(memory); 1316 1317 return nid; 1318} 1319 1320/* 1321 * Find the node associated with a hot added memory section. Section 1322 * corresponds to a SPARSEMEM section, not an MEMBLOCK. It is assumed that 1323 * sections are fully contained within a single MEMBLOCK. 1324 */ 1325int hot_add_scn_to_nid(unsigned long scn_addr) 1326{ 1327 struct device_node *memory = NULL; 1328 int nid; 1329 1330 if (!numa_enabled) 1331 return first_online_node; 1332 1333 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory"); 1334 if (memory) { 1335 nid = hot_add_drconf_scn_to_nid(scn_addr); 1336 of_node_put(memory); 1337 } else { 1338 nid = hot_add_node_scn_to_nid(scn_addr); 1339 } 1340 1341 if (nid < 0 || !node_possible(nid)) 1342 nid = first_online_node; 1343 1344 return nid; 1345} 1346 1347static u64 hot_add_drconf_memory_max(void) 1348{ 1349 struct device_node *memory = NULL; 1350 struct device_node *dn = NULL; 1351 const __be64 *lrdr = NULL; 1352 1353 dn = of_find_node_by_path("/rtas"); 1354 if (dn) { 1355 lrdr = of_get_property(dn, "ibm,lrdr-capacity", NULL); 1356 of_node_put(dn); 1357 if (lrdr) 1358 return be64_to_cpup(lrdr); 1359 } 1360 1361 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory"); 1362 if (memory) { 1363 of_node_put(memory); 1364 return drmem_lmb_memory_max(); 1365 } 1366 return 0; 1367} 1368 1369/* 1370 * memory_hotplug_max - return max address of memory that may be added 1371 * 1372 * This is currently only used on systems that support drconfig memory 1373 * hotplug. 1374 */ 1375u64 memory_hotplug_max(void) 1376{ 1377 return max(hot_add_drconf_memory_max(), memblock_end_of_DRAM()); 1378} 1379#endif /* CONFIG_MEMORY_HOTPLUG */ 1380 1381/* Virtual Processor Home Node (VPHN) support */ 1382#ifdef CONFIG_PPC_SPLPAR 1383static int topology_inited; 1384 1385/* 1386 * Retrieve the new associativity information for a virtual processor's 1387 * home node. 1388 */ 1389static long vphn_get_associativity(unsigned long cpu, 1390 __be32 *associativity) 1391{ 1392 long rc; 1393 1394 rc = hcall_vphn(get_hard_smp_processor_id(cpu), 1395 VPHN_FLAG_VCPU, associativity); 1396 1397 switch (rc) { 1398 case H_SUCCESS: 1399 pr_debug("VPHN hcall succeeded. Reset polling...\n"); 1400 goto out; 1401 1402 case H_FUNCTION: 1403 pr_err_ratelimited("VPHN unsupported. Disabling polling...\n"); 1404 break; 1405 case H_HARDWARE: 1406 pr_err_ratelimited("hcall_vphn() experienced a hardware fault " 1407 "preventing VPHN. Disabling polling...\n"); 1408 break; 1409 case H_PARAMETER: 1410 pr_err_ratelimited("hcall_vphn() was passed an invalid parameter. " 1411 "Disabling polling...\n"); 1412 break; 1413 default: 1414 pr_err_ratelimited("hcall_vphn() returned %ld. Disabling polling...\n" 1415 , rc); 1416 break; 1417 } 1418out: 1419 return rc; 1420} 1421 1422void find_and_update_cpu_nid(int cpu) 1423{ 1424 __be32 associativity[VPHN_ASSOC_BUFSIZE] = {0}; 1425 int new_nid; 1426 1427 /* Use associativity from first thread for all siblings */ 1428 if (vphn_get_associativity(cpu, associativity)) 1429 return; 1430 1431 /* Do not have previous associativity, so find it now. */ 1432 new_nid = associativity_to_nid(associativity); 1433 1434 if (new_nid < 0 || !node_possible(new_nid)) 1435 new_nid = first_online_node; 1436 else 1437 // Associate node <-> cpu, so cpu_up() calls 1438 // try_online_node() on the right node. 1439 set_cpu_numa_node(cpu, new_nid); 1440 1441 pr_debug("%s:%d cpu %d nid %d\n", __func__, __LINE__, cpu, new_nid); 1442} 1443 1444int cpu_to_coregroup_id(int cpu) 1445{ 1446 __be32 associativity[VPHN_ASSOC_BUFSIZE] = {0}; 1447 int index; 1448 1449 if (cpu < 0 || cpu > nr_cpu_ids) 1450 return -1; 1451 1452 if (!coregroup_enabled) 1453 goto out; 1454 1455 if (!firmware_has_feature(FW_FEATURE_VPHN)) 1456 goto out; 1457 1458 if (vphn_get_associativity(cpu, associativity)) 1459 goto out; 1460 1461 index = of_read_number(associativity, 1); 1462 if (index > primary_domain_index + 1) 1463 return of_read_number(&associativity[index - 1], 1); 1464 1465out: 1466 return cpu_to_core_id(cpu); 1467} 1468 1469static int topology_update_init(void) 1470{ 1471 topology_inited = 1; 1472 return 0; 1473} 1474device_initcall(topology_update_init); 1475#endif /* CONFIG_PPC_SPLPAR */ 1476