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 #define pr_fmt(fmt) "numa: " fmt 12 13 #include <linux/threads.h> 14 #include <linux/memblock.h> 15 #include <linux/init.h> 16 #include <linux/mm.h> 17 #include <linux/mmzone.h> 18 #include <linux/export.h> 19 #include <linux/nodemask.h> 20 #include <linux/cpu.h> 21 #include <linux/notifier.h> 22 #include <linux/of.h> 23 #include <linux/pfn.h> 24 #include <linux/cpuset.h> 25 #include <linux/node.h> 26 #include <linux/stop_machine.h> 27 #include <linux/proc_fs.h> 28 #include <linux/seq_file.h> 29 #include <linux/uaccess.h> 30 #include <linux/slab.h> 31 #include <asm/cputhreads.h> 32 #include <asm/sparsemem.h> 33 #include <asm/prom.h> 34 #include <asm/smp.h> 35 #include <asm/cputhreads.h> 36 #include <asm/topology.h> 37 #include <asm/firmware.h> 38 #include <asm/paca.h> 39 #include <asm/hvcall.h> 40 #include <asm/setup.h> 41 #include <asm/vdso.h> 42 #include <asm/drmem.h> 43 44 static int numa_enabled = 1; 45 46 static char *cmdline __initdata; 47 48 static int numa_debug; 49 #define dbg(args...) if (numa_debug) { printk(KERN_INFO args); } 50 51 int numa_cpu_lookup_table[NR_CPUS]; 52 cpumask_var_t node_to_cpumask_map[MAX_NUMNODES]; 53 struct pglist_data *node_data[MAX_NUMNODES]; 54 55 EXPORT_SYMBOL(numa_cpu_lookup_table); 56 EXPORT_SYMBOL(node_to_cpumask_map); 57 EXPORT_SYMBOL(node_data); 58 59 static int min_common_depth; 60 static int n_mem_addr_cells, n_mem_size_cells; 61 static int form1_affinity; 62 63 #define MAX_DISTANCE_REF_POINTS 4 64 static int distance_ref_points_depth; 65 static const __be32 *distance_ref_points; 66 static int distance_lookup_table[MAX_NUMNODES][MAX_DISTANCE_REF_POINTS]; 67 68 /* 69 * Allocate node_to_cpumask_map based on number of available nodes 70 * Requires node_possible_map to be valid. 71 * 72 * Note: cpumask_of_node() is not valid until after this is done. 73 */ 74 static void __init setup_node_to_cpumask_map(void) 75 { 76 unsigned int node; 77 78 /* setup nr_node_ids if not done yet */ 79 if (nr_node_ids == MAX_NUMNODES) 80 setup_nr_node_ids(); 81 82 /* allocate the map */ 83 for_each_node(node) 84 alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]); 85 86 /* cpumask_of_node() will now work */ 87 dbg("Node to cpumask map for %u nodes\n", nr_node_ids); 88 } 89 90 static int __init fake_numa_create_new_node(unsigned long end_pfn, 91 unsigned int *nid) 92 { 93 unsigned long long mem; 94 char *p = cmdline; 95 static unsigned int fake_nid; 96 static unsigned long long curr_boundary; 97 98 /* 99 * Modify node id, iff we started creating NUMA nodes 100 * We want to continue from where we left of the last time 101 */ 102 if (fake_nid) 103 *nid = fake_nid; 104 /* 105 * In case there are no more arguments to parse, the 106 * node_id should be the same as the last fake node id 107 * (we've handled this above). 108 */ 109 if (!p) 110 return 0; 111 112 mem = memparse(p, &p); 113 if (!mem) 114 return 0; 115 116 if (mem < curr_boundary) 117 return 0; 118 119 curr_boundary = mem; 120 121 if ((end_pfn << PAGE_SHIFT) > mem) { 122 /* 123 * Skip commas and spaces 124 */ 125 while (*p == ',' || *p == ' ' || *p == '\t') 126 p++; 127 128 cmdline = p; 129 fake_nid++; 130 *nid = fake_nid; 131 dbg("created new fake_node with id %d\n", fake_nid); 132 return 1; 133 } 134 return 0; 135 } 136 137 static void reset_numa_cpu_lookup_table(void) 138 { 139 unsigned int cpu; 140 141 for_each_possible_cpu(cpu) 142 numa_cpu_lookup_table[cpu] = -1; 143 } 144 145 static void map_cpu_to_node(int cpu, int node) 146 { 147 update_numa_cpu_lookup_table(cpu, node); 148 149 dbg("adding cpu %d to node %d\n", cpu, node); 150 151 if (!(cpumask_test_cpu(cpu, node_to_cpumask_map[node]))) 152 cpumask_set_cpu(cpu, node_to_cpumask_map[node]); 153 } 154 155 #if defined(CONFIG_HOTPLUG_CPU) || defined(CONFIG_PPC_SPLPAR) 156 static void unmap_cpu_from_node(unsigned long cpu) 157 { 158 int node = numa_cpu_lookup_table[cpu]; 159 160 dbg("removing cpu %lu from node %d\n", cpu, node); 161 162 if (cpumask_test_cpu(cpu, node_to_cpumask_map[node])) { 163 cpumask_clear_cpu(cpu, node_to_cpumask_map[node]); 164 } else { 165 printk(KERN_ERR "WARNING: cpu %lu not found in node %d\n", 166 cpu, node); 167 } 168 } 169 #endif /* CONFIG_HOTPLUG_CPU || CONFIG_PPC_SPLPAR */ 170 171 /* must hold reference to node during call */ 172 static const __be32 *of_get_associativity(struct device_node *dev) 173 { 174 return of_get_property(dev, "ibm,associativity", NULL); 175 } 176 177 int __node_distance(int a, int b) 178 { 179 int i; 180 int distance = LOCAL_DISTANCE; 181 182 if (!form1_affinity) 183 return ((a == b) ? LOCAL_DISTANCE : REMOTE_DISTANCE); 184 185 for (i = 0; i < distance_ref_points_depth; i++) { 186 if (distance_lookup_table[a][i] == distance_lookup_table[b][i]) 187 break; 188 189 /* Double the distance for each NUMA level */ 190 distance *= 2; 191 } 192 193 return distance; 194 } 195 EXPORT_SYMBOL(__node_distance); 196 197 static void initialize_distance_lookup_table(int nid, 198 const __be32 *associativity) 199 { 200 int i; 201 202 if (!form1_affinity) 203 return; 204 205 for (i = 0; i < distance_ref_points_depth; i++) { 206 const __be32 *entry; 207 208 entry = &associativity[be32_to_cpu(distance_ref_points[i]) - 1]; 209 distance_lookup_table[nid][i] = of_read_number(entry, 1); 210 } 211 } 212 213 /* Returns nid in the range [0..MAX_NUMNODES-1], or -1 if no useful numa 214 * info is found. 215 */ 216 static int associativity_to_nid(const __be32 *associativity) 217 { 218 int nid = NUMA_NO_NODE; 219 220 if (min_common_depth == -1) 221 goto out; 222 223 if (of_read_number(associativity, 1) >= min_common_depth) 224 nid = of_read_number(&associativity[min_common_depth], 1); 225 226 /* POWER4 LPAR uses 0xffff as invalid node */ 227 if (nid == 0xffff || nid >= MAX_NUMNODES) 228 nid = NUMA_NO_NODE; 229 230 if (nid > 0 && 231 of_read_number(associativity, 1) >= distance_ref_points_depth) { 232 /* 233 * Skip the length field and send start of associativity array 234 */ 235 initialize_distance_lookup_table(nid, associativity + 1); 236 } 237 238 out: 239 return nid; 240 } 241 242 /* Returns the nid associated with the given device tree node, 243 * or -1 if not found. 244 */ 245 static int of_node_to_nid_single(struct device_node *device) 246 { 247 int nid = NUMA_NO_NODE; 248 const __be32 *tmp; 249 250 tmp = of_get_associativity(device); 251 if (tmp) 252 nid = associativity_to_nid(tmp); 253 return nid; 254 } 255 256 /* Walk the device tree upwards, looking for an associativity id */ 257 int of_node_to_nid(struct device_node *device) 258 { 259 int nid = NUMA_NO_NODE; 260 261 of_node_get(device); 262 while (device) { 263 nid = of_node_to_nid_single(device); 264 if (nid != -1) 265 break; 266 267 device = of_get_next_parent(device); 268 } 269 of_node_put(device); 270 271 return nid; 272 } 273 EXPORT_SYMBOL(of_node_to_nid); 274 275 static int __init find_min_common_depth(void) 276 { 277 int depth; 278 struct device_node *root; 279 280 if (firmware_has_feature(FW_FEATURE_OPAL)) 281 root = of_find_node_by_path("/ibm,opal"); 282 else 283 root = of_find_node_by_path("/rtas"); 284 if (!root) 285 root = of_find_node_by_path("/"); 286 287 /* 288 * This property is a set of 32-bit integers, each representing 289 * an index into the ibm,associativity nodes. 290 * 291 * With form 0 affinity the first integer is for an SMP configuration 292 * (should be all 0's) and the second is for a normal NUMA 293 * configuration. We have only one level of NUMA. 294 * 295 * With form 1 affinity the first integer is the most significant 296 * NUMA boundary and the following are progressively less significant 297 * boundaries. There can be more than one level of NUMA. 298 */ 299 distance_ref_points = of_get_property(root, 300 "ibm,associativity-reference-points", 301 &distance_ref_points_depth); 302 303 if (!distance_ref_points) { 304 dbg("NUMA: ibm,associativity-reference-points not found.\n"); 305 goto err; 306 } 307 308 distance_ref_points_depth /= sizeof(int); 309 310 if (firmware_has_feature(FW_FEATURE_OPAL) || 311 firmware_has_feature(FW_FEATURE_TYPE1_AFFINITY)) { 312 dbg("Using form 1 affinity\n"); 313 form1_affinity = 1; 314 } 315 316 if (form1_affinity) { 317 depth = of_read_number(distance_ref_points, 1); 318 } else { 319 if (distance_ref_points_depth < 2) { 320 printk(KERN_WARNING "NUMA: " 321 "short ibm,associativity-reference-points\n"); 322 goto err; 323 } 324 325 depth = of_read_number(&distance_ref_points[1], 1); 326 } 327 328 /* 329 * Warn and cap if the hardware supports more than 330 * MAX_DISTANCE_REF_POINTS domains. 331 */ 332 if (distance_ref_points_depth > MAX_DISTANCE_REF_POINTS) { 333 printk(KERN_WARNING "NUMA: distance array capped at " 334 "%d entries\n", MAX_DISTANCE_REF_POINTS); 335 distance_ref_points_depth = MAX_DISTANCE_REF_POINTS; 336 } 337 338 of_node_put(root); 339 return depth; 340 341 err: 342 of_node_put(root); 343 return -1; 344 } 345 346 static void __init get_n_mem_cells(int *n_addr_cells, int *n_size_cells) 347 { 348 struct device_node *memory = NULL; 349 350 memory = of_find_node_by_type(memory, "memory"); 351 if (!memory) 352 panic("numa.c: No memory nodes found!"); 353 354 *n_addr_cells = of_n_addr_cells(memory); 355 *n_size_cells = of_n_size_cells(memory); 356 of_node_put(memory); 357 } 358 359 static unsigned long read_n_cells(int n, const __be32 **buf) 360 { 361 unsigned long result = 0; 362 363 while (n--) { 364 result = (result << 32) | of_read_number(*buf, 1); 365 (*buf)++; 366 } 367 return result; 368 } 369 370 struct assoc_arrays { 371 u32 n_arrays; 372 u32 array_sz; 373 const __be32 *arrays; 374 }; 375 376 /* 377 * Retrieve and validate the list of associativity arrays for drconf 378 * memory from the ibm,associativity-lookup-arrays property of the 379 * device tree.. 380 * 381 * The layout of the ibm,associativity-lookup-arrays property is a number N 382 * indicating the number of associativity arrays, followed by a number M 383 * indicating the size of each associativity array, followed by a list 384 * of N associativity arrays. 385 */ 386 static int of_get_assoc_arrays(struct assoc_arrays *aa) 387 { 388 struct device_node *memory; 389 const __be32 *prop; 390 u32 len; 391 392 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory"); 393 if (!memory) 394 return -1; 395 396 prop = of_get_property(memory, "ibm,associativity-lookup-arrays", &len); 397 if (!prop || len < 2 * sizeof(unsigned int)) { 398 of_node_put(memory); 399 return -1; 400 } 401 402 aa->n_arrays = of_read_number(prop++, 1); 403 aa->array_sz = of_read_number(prop++, 1); 404 405 of_node_put(memory); 406 407 /* Now that we know the number of arrays and size of each array, 408 * revalidate the size of the property read in. 409 */ 410 if (len < (aa->n_arrays * aa->array_sz + 2) * sizeof(unsigned int)) 411 return -1; 412 413 aa->arrays = prop; 414 return 0; 415 } 416 417 /* 418 * This is like of_node_to_nid_single() for memory represented in the 419 * ibm,dynamic-reconfiguration-memory node. 420 */ 421 static int of_drconf_to_nid_single(struct drmem_lmb *lmb) 422 { 423 struct assoc_arrays aa = { .arrays = NULL }; 424 int default_nid = 0; 425 int nid = default_nid; 426 int rc, index; 427 428 rc = of_get_assoc_arrays(&aa); 429 if (rc) 430 return default_nid; 431 432 if (min_common_depth > 0 && min_common_depth <= aa.array_sz && 433 !(lmb->flags & DRCONF_MEM_AI_INVALID) && 434 lmb->aa_index < aa.n_arrays) { 435 index = lmb->aa_index * aa.array_sz + min_common_depth - 1; 436 nid = of_read_number(&aa.arrays[index], 1); 437 438 if (nid == 0xffff || nid >= MAX_NUMNODES) 439 nid = default_nid; 440 441 if (nid > 0) { 442 index = lmb->aa_index * aa.array_sz; 443 initialize_distance_lookup_table(nid, 444 &aa.arrays[index]); 445 } 446 } 447 448 return nid; 449 } 450 451 /* 452 * Figure out to which domain a cpu belongs and stick it there. 453 * Return the id of the domain used. 454 */ 455 static int numa_setup_cpu(unsigned long lcpu) 456 { 457 int nid = NUMA_NO_NODE; 458 struct device_node *cpu; 459 460 /* 461 * If a valid cpu-to-node mapping is already available, use it 462 * directly instead of querying the firmware, since it represents 463 * the most recent mapping notified to us by the platform (eg: VPHN). 464 */ 465 if ((nid = numa_cpu_lookup_table[lcpu]) >= 0) { 466 map_cpu_to_node(lcpu, nid); 467 return nid; 468 } 469 470 cpu = of_get_cpu_node(lcpu, NULL); 471 472 if (!cpu) { 473 WARN_ON(1); 474 if (cpu_present(lcpu)) 475 goto out_present; 476 else 477 goto out; 478 } 479 480 nid = of_node_to_nid_single(cpu); 481 482 out_present: 483 if (nid < 0 || !node_possible(nid)) 484 nid = first_online_node; 485 486 map_cpu_to_node(lcpu, nid); 487 of_node_put(cpu); 488 out: 489 return nid; 490 } 491 492 static void verify_cpu_node_mapping(int cpu, int node) 493 { 494 int base, sibling, i; 495 496 /* Verify that all the threads in the core belong to the same node */ 497 base = cpu_first_thread_sibling(cpu); 498 499 for (i = 0; i < threads_per_core; i++) { 500 sibling = base + i; 501 502 if (sibling == cpu || cpu_is_offline(sibling)) 503 continue; 504 505 if (cpu_to_node(sibling) != node) { 506 WARN(1, "CPU thread siblings %d and %d don't belong" 507 " to the same node!\n", cpu, sibling); 508 break; 509 } 510 } 511 } 512 513 /* Must run before sched domains notifier. */ 514 static int ppc_numa_cpu_prepare(unsigned int cpu) 515 { 516 int nid; 517 518 nid = numa_setup_cpu(cpu); 519 verify_cpu_node_mapping(cpu, nid); 520 return 0; 521 } 522 523 static int ppc_numa_cpu_dead(unsigned int cpu) 524 { 525 #ifdef CONFIG_HOTPLUG_CPU 526 unmap_cpu_from_node(cpu); 527 #endif 528 return 0; 529 } 530 531 /* 532 * Check and possibly modify a memory region to enforce the memory limit. 533 * 534 * Returns the size the region should have to enforce the memory limit. 535 * This will either be the original value of size, a truncated value, 536 * or zero. If the returned value of size is 0 the region should be 537 * discarded as it lies wholly above the memory limit. 538 */ 539 static unsigned long __init numa_enforce_memory_limit(unsigned long start, 540 unsigned long size) 541 { 542 /* 543 * We use memblock_end_of_DRAM() in here instead of memory_limit because 544 * we've already adjusted it for the limit and it takes care of 545 * having memory holes below the limit. Also, in the case of 546 * iommu_is_off, memory_limit is not set but is implicitly enforced. 547 */ 548 549 if (start + size <= memblock_end_of_DRAM()) 550 return size; 551 552 if (start >= memblock_end_of_DRAM()) 553 return 0; 554 555 return memblock_end_of_DRAM() - start; 556 } 557 558 /* 559 * Reads the counter for a given entry in 560 * linux,drconf-usable-memory property 561 */ 562 static inline int __init read_usm_ranges(const __be32 **usm) 563 { 564 /* 565 * For each lmb in ibm,dynamic-memory a corresponding 566 * entry in linux,drconf-usable-memory property contains 567 * a counter followed by that many (base, size) duple. 568 * read the counter from linux,drconf-usable-memory 569 */ 570 return read_n_cells(n_mem_size_cells, usm); 571 } 572 573 /* 574 * Extract NUMA information from the ibm,dynamic-reconfiguration-memory 575 * node. This assumes n_mem_{addr,size}_cells have been set. 576 */ 577 static void __init numa_setup_drmem_lmb(struct drmem_lmb *lmb, 578 const __be32 **usm) 579 { 580 unsigned int ranges, is_kexec_kdump = 0; 581 unsigned long base, size, sz; 582 int nid; 583 584 /* 585 * Skip this block if the reserved bit is set in flags (0x80) 586 * or if the block is not assigned to this partition (0x8) 587 */ 588 if ((lmb->flags & DRCONF_MEM_RESERVED) 589 || !(lmb->flags & DRCONF_MEM_ASSIGNED)) 590 return; 591 592 if (*usm) 593 is_kexec_kdump = 1; 594 595 base = lmb->base_addr; 596 size = drmem_lmb_size(); 597 ranges = 1; 598 599 if (is_kexec_kdump) { 600 ranges = read_usm_ranges(usm); 601 if (!ranges) /* there are no (base, size) duple */ 602 return; 603 } 604 605 do { 606 if (is_kexec_kdump) { 607 base = read_n_cells(n_mem_addr_cells, usm); 608 size = read_n_cells(n_mem_size_cells, usm); 609 } 610 611 nid = of_drconf_to_nid_single(lmb); 612 fake_numa_create_new_node(((base + size) >> PAGE_SHIFT), 613 &nid); 614 node_set_online(nid); 615 sz = numa_enforce_memory_limit(base, size); 616 if (sz) 617 memblock_set_node(base, sz, &memblock.memory, nid); 618 } while (--ranges); 619 } 620 621 static int __init parse_numa_properties(void) 622 { 623 struct device_node *memory; 624 int default_nid = 0; 625 unsigned long i; 626 627 if (numa_enabled == 0) { 628 printk(KERN_WARNING "NUMA disabled by user\n"); 629 return -1; 630 } 631 632 min_common_depth = find_min_common_depth(); 633 634 if (min_common_depth < 0) 635 return min_common_depth; 636 637 dbg("NUMA associativity depth for CPU/Memory: %d\n", min_common_depth); 638 639 /* 640 * Even though we connect cpus to numa domains later in SMP 641 * init, we need to know the node ids now. This is because 642 * each node to be onlined must have NODE_DATA etc backing it. 643 */ 644 for_each_present_cpu(i) { 645 struct device_node *cpu; 646 int nid; 647 648 cpu = of_get_cpu_node(i, NULL); 649 BUG_ON(!cpu); 650 nid = of_node_to_nid_single(cpu); 651 of_node_put(cpu); 652 653 /* 654 * Don't fall back to default_nid yet -- we will plug 655 * cpus into nodes once the memory scan has discovered 656 * the topology. 657 */ 658 if (nid < 0) 659 continue; 660 node_set_online(nid); 661 } 662 663 get_n_mem_cells(&n_mem_addr_cells, &n_mem_size_cells); 664 665 for_each_node_by_type(memory, "memory") { 666 unsigned long start; 667 unsigned long size; 668 int nid; 669 int ranges; 670 const __be32 *memcell_buf; 671 unsigned int len; 672 673 memcell_buf = of_get_property(memory, 674 "linux,usable-memory", &len); 675 if (!memcell_buf || len <= 0) 676 memcell_buf = of_get_property(memory, "reg", &len); 677 if (!memcell_buf || len <= 0) 678 continue; 679 680 /* ranges in cell */ 681 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells); 682 new_range: 683 /* these are order-sensitive, and modify the buffer pointer */ 684 start = read_n_cells(n_mem_addr_cells, &memcell_buf); 685 size = read_n_cells(n_mem_size_cells, &memcell_buf); 686 687 /* 688 * Assumption: either all memory nodes or none will 689 * have associativity properties. If none, then 690 * everything goes to default_nid. 691 */ 692 nid = of_node_to_nid_single(memory); 693 if (nid < 0) 694 nid = default_nid; 695 696 fake_numa_create_new_node(((start + size) >> PAGE_SHIFT), &nid); 697 node_set_online(nid); 698 699 size = numa_enforce_memory_limit(start, size); 700 if (size) 701 memblock_set_node(start, size, &memblock.memory, nid); 702 703 if (--ranges) 704 goto new_range; 705 } 706 707 /* 708 * Now do the same thing for each MEMBLOCK listed in the 709 * ibm,dynamic-memory property in the 710 * ibm,dynamic-reconfiguration-memory node. 711 */ 712 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory"); 713 if (memory) { 714 walk_drmem_lmbs(memory, numa_setup_drmem_lmb); 715 of_node_put(memory); 716 } 717 718 return 0; 719 } 720 721 static void __init setup_nonnuma(void) 722 { 723 unsigned long top_of_ram = memblock_end_of_DRAM(); 724 unsigned long total_ram = memblock_phys_mem_size(); 725 unsigned long start_pfn, end_pfn; 726 unsigned int nid = 0; 727 struct memblock_region *reg; 728 729 printk(KERN_DEBUG "Top of RAM: 0x%lx, Total RAM: 0x%lx\n", 730 top_of_ram, total_ram); 731 printk(KERN_DEBUG "Memory hole size: %ldMB\n", 732 (top_of_ram - total_ram) >> 20); 733 734 for_each_memblock(memory, reg) { 735 start_pfn = memblock_region_memory_base_pfn(reg); 736 end_pfn = memblock_region_memory_end_pfn(reg); 737 738 fake_numa_create_new_node(end_pfn, &nid); 739 memblock_set_node(PFN_PHYS(start_pfn), 740 PFN_PHYS(end_pfn - start_pfn), 741 &memblock.memory, nid); 742 node_set_online(nid); 743 } 744 } 745 746 void __init dump_numa_cpu_topology(void) 747 { 748 unsigned int node; 749 unsigned int cpu, count; 750 751 if (min_common_depth == -1 || !numa_enabled) 752 return; 753 754 for_each_online_node(node) { 755 pr_info("Node %d CPUs:", node); 756 757 count = 0; 758 /* 759 * If we used a CPU iterator here we would miss printing 760 * the holes in the cpumap. 761 */ 762 for (cpu = 0; cpu < nr_cpu_ids; cpu++) { 763 if (cpumask_test_cpu(cpu, 764 node_to_cpumask_map[node])) { 765 if (count == 0) 766 pr_cont(" %u", cpu); 767 ++count; 768 } else { 769 if (count > 1) 770 pr_cont("-%u", cpu - 1); 771 count = 0; 772 } 773 } 774 775 if (count > 1) 776 pr_cont("-%u", nr_cpu_ids - 1); 777 pr_cont("\n"); 778 } 779 } 780 781 /* Initialize NODE_DATA for a node on the local memory */ 782 static void __init setup_node_data(int nid, u64 start_pfn, u64 end_pfn) 783 { 784 u64 spanned_pages = end_pfn - start_pfn; 785 const size_t nd_size = roundup(sizeof(pg_data_t), SMP_CACHE_BYTES); 786 u64 nd_pa; 787 void *nd; 788 int tnid; 789 790 nd_pa = memblock_phys_alloc_try_nid(nd_size, SMP_CACHE_BYTES, nid); 791 if (!nd_pa) 792 panic("Cannot allocate %zu bytes for node %d data\n", 793 nd_size, nid); 794 795 nd = __va(nd_pa); 796 797 /* report and initialize */ 798 pr_info(" NODE_DATA [mem %#010Lx-%#010Lx]\n", 799 nd_pa, nd_pa + nd_size - 1); 800 tnid = early_pfn_to_nid(nd_pa >> PAGE_SHIFT); 801 if (tnid != nid) 802 pr_info(" NODE_DATA(%d) on node %d\n", nid, tnid); 803 804 node_data[nid] = nd; 805 memset(NODE_DATA(nid), 0, sizeof(pg_data_t)); 806 NODE_DATA(nid)->node_id = nid; 807 NODE_DATA(nid)->node_start_pfn = start_pfn; 808 NODE_DATA(nid)->node_spanned_pages = spanned_pages; 809 } 810 811 static void __init find_possible_nodes(void) 812 { 813 struct device_node *rtas; 814 u32 numnodes, i; 815 816 if (min_common_depth <= 0) 817 return; 818 819 rtas = of_find_node_by_path("/rtas"); 820 if (!rtas) 821 return; 822 823 if (of_property_read_u32_index(rtas, 824 "ibm,max-associativity-domains", 825 min_common_depth, &numnodes)) 826 goto out; 827 828 for (i = 0; i < numnodes; i++) { 829 if (!node_possible(i)) 830 node_set(i, node_possible_map); 831 } 832 833 out: 834 of_node_put(rtas); 835 } 836 837 void __init mem_topology_setup(void) 838 { 839 int cpu; 840 841 if (parse_numa_properties()) 842 setup_nonnuma(); 843 844 /* 845 * Modify the set of possible NUMA nodes to reflect information 846 * available about the set of online nodes, and the set of nodes 847 * that we expect to make use of for this platform's affinity 848 * calculations. 849 */ 850 nodes_and(node_possible_map, node_possible_map, node_online_map); 851 852 find_possible_nodes(); 853 854 setup_node_to_cpumask_map(); 855 856 reset_numa_cpu_lookup_table(); 857 858 for_each_present_cpu(cpu) 859 numa_setup_cpu(cpu); 860 } 861 862 void __init initmem_init(void) 863 { 864 int nid; 865 866 max_low_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT; 867 max_pfn = max_low_pfn; 868 869 memblock_dump_all(); 870 871 for_each_online_node(nid) { 872 unsigned long start_pfn, end_pfn; 873 874 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn); 875 setup_node_data(nid, start_pfn, end_pfn); 876 sparse_memory_present_with_active_regions(nid); 877 } 878 879 sparse_init(); 880 881 /* 882 * We need the numa_cpu_lookup_table to be accurate for all CPUs, 883 * even before we online them, so that we can use cpu_to_{node,mem} 884 * early in boot, cf. smp_prepare_cpus(). 885 * _nocalls() + manual invocation is used because cpuhp is not yet 886 * initialized for the boot CPU. 887 */ 888 cpuhp_setup_state_nocalls(CPUHP_POWER_NUMA_PREPARE, "powerpc/numa:prepare", 889 ppc_numa_cpu_prepare, ppc_numa_cpu_dead); 890 } 891 892 static int __init early_numa(char *p) 893 { 894 if (!p) 895 return 0; 896 897 if (strstr(p, "off")) 898 numa_enabled = 0; 899 900 if (strstr(p, "debug")) 901 numa_debug = 1; 902 903 p = strstr(p, "fake="); 904 if (p) 905 cmdline = p + strlen("fake="); 906 907 return 0; 908 } 909 early_param("numa", early_numa); 910 911 static bool topology_updates_enabled = true; 912 913 static int __init early_topology_updates(char *p) 914 { 915 if (!p) 916 return 0; 917 918 if (!strcmp(p, "off")) { 919 pr_info("Disabling topology updates\n"); 920 topology_updates_enabled = false; 921 } 922 923 return 0; 924 } 925 early_param("topology_updates", early_topology_updates); 926 927 #ifdef CONFIG_MEMORY_HOTPLUG 928 /* 929 * Find the node associated with a hot added memory section for 930 * memory represented in the device tree by the property 931 * ibm,dynamic-reconfiguration-memory/ibm,dynamic-memory. 932 */ 933 static int hot_add_drconf_scn_to_nid(unsigned long scn_addr) 934 { 935 struct drmem_lmb *lmb; 936 unsigned long lmb_size; 937 int nid = NUMA_NO_NODE; 938 939 lmb_size = drmem_lmb_size(); 940 941 for_each_drmem_lmb(lmb) { 942 /* skip this block if it is reserved or not assigned to 943 * this partition */ 944 if ((lmb->flags & DRCONF_MEM_RESERVED) 945 || !(lmb->flags & DRCONF_MEM_ASSIGNED)) 946 continue; 947 948 if ((scn_addr < lmb->base_addr) 949 || (scn_addr >= (lmb->base_addr + lmb_size))) 950 continue; 951 952 nid = of_drconf_to_nid_single(lmb); 953 break; 954 } 955 956 return nid; 957 } 958 959 /* 960 * Find the node associated with a hot added memory section for memory 961 * represented in the device tree as a node (i.e. memory@XXXX) for 962 * each memblock. 963 */ 964 static int hot_add_node_scn_to_nid(unsigned long scn_addr) 965 { 966 struct device_node *memory; 967 int nid = NUMA_NO_NODE; 968 969 for_each_node_by_type(memory, "memory") { 970 unsigned long start, size; 971 int ranges; 972 const __be32 *memcell_buf; 973 unsigned int len; 974 975 memcell_buf = of_get_property(memory, "reg", &len); 976 if (!memcell_buf || len <= 0) 977 continue; 978 979 /* ranges in cell */ 980 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells); 981 982 while (ranges--) { 983 start = read_n_cells(n_mem_addr_cells, &memcell_buf); 984 size = read_n_cells(n_mem_size_cells, &memcell_buf); 985 986 if ((scn_addr < start) || (scn_addr >= (start + size))) 987 continue; 988 989 nid = of_node_to_nid_single(memory); 990 break; 991 } 992 993 if (nid >= 0) 994 break; 995 } 996 997 of_node_put(memory); 998 999 return nid; 1000 } 1001 1002 /* 1003 * Find the node associated with a hot added memory section. Section 1004 * corresponds to a SPARSEMEM section, not an MEMBLOCK. It is assumed that 1005 * sections are fully contained within a single MEMBLOCK. 1006 */ 1007 int hot_add_scn_to_nid(unsigned long scn_addr) 1008 { 1009 struct device_node *memory = NULL; 1010 int nid; 1011 1012 if (!numa_enabled || (min_common_depth < 0)) 1013 return first_online_node; 1014 1015 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory"); 1016 if (memory) { 1017 nid = hot_add_drconf_scn_to_nid(scn_addr); 1018 of_node_put(memory); 1019 } else { 1020 nid = hot_add_node_scn_to_nid(scn_addr); 1021 } 1022 1023 if (nid < 0 || !node_possible(nid)) 1024 nid = first_online_node; 1025 1026 return nid; 1027 } 1028 1029 static u64 hot_add_drconf_memory_max(void) 1030 { 1031 struct device_node *memory = NULL; 1032 struct device_node *dn = NULL; 1033 const __be64 *lrdr = NULL; 1034 1035 dn = of_find_node_by_path("/rtas"); 1036 if (dn) { 1037 lrdr = of_get_property(dn, "ibm,lrdr-capacity", NULL); 1038 of_node_put(dn); 1039 if (lrdr) 1040 return be64_to_cpup(lrdr); 1041 } 1042 1043 memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory"); 1044 if (memory) { 1045 of_node_put(memory); 1046 return drmem_lmb_memory_max(); 1047 } 1048 return 0; 1049 } 1050 1051 /* 1052 * memory_hotplug_max - return max address of memory that may be added 1053 * 1054 * This is currently only used on systems that support drconfig memory 1055 * hotplug. 1056 */ 1057 u64 memory_hotplug_max(void) 1058 { 1059 return max(hot_add_drconf_memory_max(), memblock_end_of_DRAM()); 1060 } 1061 #endif /* CONFIG_MEMORY_HOTPLUG */ 1062 1063 /* Virtual Processor Home Node (VPHN) support */ 1064 #ifdef CONFIG_PPC_SPLPAR 1065 1066 #include "vphn.h" 1067 1068 struct topology_update_data { 1069 struct topology_update_data *next; 1070 unsigned int cpu; 1071 int old_nid; 1072 int new_nid; 1073 }; 1074 1075 #define TOPOLOGY_DEF_TIMER_SECS 60 1076 1077 static u8 vphn_cpu_change_counts[NR_CPUS][MAX_DISTANCE_REF_POINTS]; 1078 static cpumask_t cpu_associativity_changes_mask; 1079 static int vphn_enabled; 1080 static int prrn_enabled; 1081 static void reset_topology_timer(void); 1082 static int topology_timer_secs = 1; 1083 static int topology_inited; 1084 1085 /* 1086 * Change polling interval for associativity changes. 1087 */ 1088 int timed_topology_update(int nsecs) 1089 { 1090 if (vphn_enabled) { 1091 if (nsecs > 0) 1092 topology_timer_secs = nsecs; 1093 else 1094 topology_timer_secs = TOPOLOGY_DEF_TIMER_SECS; 1095 1096 reset_topology_timer(); 1097 } 1098 1099 return 0; 1100 } 1101 1102 /* 1103 * Store the current values of the associativity change counters in the 1104 * hypervisor. 1105 */ 1106 static void setup_cpu_associativity_change_counters(void) 1107 { 1108 int cpu; 1109 1110 /* The VPHN feature supports a maximum of 8 reference points */ 1111 BUILD_BUG_ON(MAX_DISTANCE_REF_POINTS > 8); 1112 1113 for_each_possible_cpu(cpu) { 1114 int i; 1115 u8 *counts = vphn_cpu_change_counts[cpu]; 1116 volatile u8 *hypervisor_counts = lppaca_of(cpu).vphn_assoc_counts; 1117 1118 for (i = 0; i < distance_ref_points_depth; i++) 1119 counts[i] = hypervisor_counts[i]; 1120 } 1121 } 1122 1123 /* 1124 * The hypervisor maintains a set of 8 associativity change counters in 1125 * the VPA of each cpu that correspond to the associativity levels in the 1126 * ibm,associativity-reference-points property. When an associativity 1127 * level changes, the corresponding counter is incremented. 1128 * 1129 * Set a bit in cpu_associativity_changes_mask for each cpu whose home 1130 * node associativity levels have changed. 1131 * 1132 * Returns the number of cpus with unhandled associativity changes. 1133 */ 1134 static int update_cpu_associativity_changes_mask(void) 1135 { 1136 int cpu; 1137 cpumask_t *changes = &cpu_associativity_changes_mask; 1138 1139 for_each_possible_cpu(cpu) { 1140 int i, changed = 0; 1141 u8 *counts = vphn_cpu_change_counts[cpu]; 1142 volatile u8 *hypervisor_counts = lppaca_of(cpu).vphn_assoc_counts; 1143 1144 for (i = 0; i < distance_ref_points_depth; i++) { 1145 if (hypervisor_counts[i] != counts[i]) { 1146 counts[i] = hypervisor_counts[i]; 1147 changed = 1; 1148 } 1149 } 1150 if (changed) { 1151 cpumask_or(changes, changes, cpu_sibling_mask(cpu)); 1152 cpu = cpu_last_thread_sibling(cpu); 1153 } 1154 } 1155 1156 return cpumask_weight(changes); 1157 } 1158 1159 /* 1160 * Retrieve the new associativity information for a virtual processor's 1161 * home node. 1162 */ 1163 static long hcall_vphn(unsigned long cpu, __be32 *associativity) 1164 { 1165 long rc; 1166 long retbuf[PLPAR_HCALL9_BUFSIZE] = {0}; 1167 u64 flags = 1; 1168 int hwcpu = get_hard_smp_processor_id(cpu); 1169 1170 rc = plpar_hcall9(H_HOME_NODE_ASSOCIATIVITY, retbuf, flags, hwcpu); 1171 vphn_unpack_associativity(retbuf, associativity); 1172 1173 return rc; 1174 } 1175 1176 static long vphn_get_associativity(unsigned long cpu, 1177 __be32 *associativity) 1178 { 1179 long rc; 1180 1181 rc = hcall_vphn(cpu, associativity); 1182 1183 switch (rc) { 1184 case H_FUNCTION: 1185 printk_once(KERN_INFO 1186 "VPHN is not supported. Disabling polling...\n"); 1187 stop_topology_update(); 1188 break; 1189 case H_HARDWARE: 1190 printk(KERN_ERR 1191 "hcall_vphn() experienced a hardware fault " 1192 "preventing VPHN. Disabling polling...\n"); 1193 stop_topology_update(); 1194 break; 1195 case H_SUCCESS: 1196 dbg("VPHN hcall succeeded. Reset polling...\n"); 1197 timed_topology_update(0); 1198 break; 1199 } 1200 1201 return rc; 1202 } 1203 1204 int find_and_online_cpu_nid(int cpu) 1205 { 1206 __be32 associativity[VPHN_ASSOC_BUFSIZE] = {0}; 1207 int new_nid; 1208 1209 /* Use associativity from first thread for all siblings */ 1210 if (vphn_get_associativity(cpu, associativity)) 1211 return cpu_to_node(cpu); 1212 1213 new_nid = associativity_to_nid(associativity); 1214 if (new_nid < 0 || !node_possible(new_nid)) 1215 new_nid = first_online_node; 1216 1217 if (NODE_DATA(new_nid) == NULL) { 1218 #ifdef CONFIG_MEMORY_HOTPLUG 1219 /* 1220 * Need to ensure that NODE_DATA is initialized for a node from 1221 * available memory (see memblock_alloc_try_nid). If unable to 1222 * init the node, then default to nearest node that has memory 1223 * installed. Skip onlining a node if the subsystems are not 1224 * yet initialized. 1225 */ 1226 if (!topology_inited || try_online_node(new_nid)) 1227 new_nid = first_online_node; 1228 #else 1229 /* 1230 * Default to using the nearest node that has memory installed. 1231 * Otherwise, it would be necessary to patch the kernel MM code 1232 * to deal with more memoryless-node error conditions. 1233 */ 1234 new_nid = first_online_node; 1235 #endif 1236 } 1237 1238 pr_debug("%s:%d cpu %d nid %d\n", __FUNCTION__, __LINE__, 1239 cpu, new_nid); 1240 return new_nid; 1241 } 1242 1243 /* 1244 * Update the CPU maps and sysfs entries for a single CPU when its NUMA 1245 * characteristics change. This function doesn't perform any locking and is 1246 * only safe to call from stop_machine(). 1247 */ 1248 static int update_cpu_topology(void *data) 1249 { 1250 struct topology_update_data *update; 1251 unsigned long cpu; 1252 1253 if (!data) 1254 return -EINVAL; 1255 1256 cpu = smp_processor_id(); 1257 1258 for (update = data; update; update = update->next) { 1259 int new_nid = update->new_nid; 1260 if (cpu != update->cpu) 1261 continue; 1262 1263 unmap_cpu_from_node(cpu); 1264 map_cpu_to_node(cpu, new_nid); 1265 set_cpu_numa_node(cpu, new_nid); 1266 set_cpu_numa_mem(cpu, local_memory_node(new_nid)); 1267 vdso_getcpu_init(); 1268 } 1269 1270 return 0; 1271 } 1272 1273 static int update_lookup_table(void *data) 1274 { 1275 struct topology_update_data *update; 1276 1277 if (!data) 1278 return -EINVAL; 1279 1280 /* 1281 * Upon topology update, the numa-cpu lookup table needs to be updated 1282 * for all threads in the core, including offline CPUs, to ensure that 1283 * future hotplug operations respect the cpu-to-node associativity 1284 * properly. 1285 */ 1286 for (update = data; update; update = update->next) { 1287 int nid, base, j; 1288 1289 nid = update->new_nid; 1290 base = cpu_first_thread_sibling(update->cpu); 1291 1292 for (j = 0; j < threads_per_core; j++) { 1293 update_numa_cpu_lookup_table(base + j, nid); 1294 } 1295 } 1296 1297 return 0; 1298 } 1299 1300 /* 1301 * Update the node maps and sysfs entries for each cpu whose home node 1302 * has changed. Returns 1 when the topology has changed, and 0 otherwise. 1303 * 1304 * cpus_locked says whether we already hold cpu_hotplug_lock. 1305 */ 1306 int numa_update_cpu_topology(bool cpus_locked) 1307 { 1308 unsigned int cpu, sibling, changed = 0; 1309 struct topology_update_data *updates, *ud; 1310 cpumask_t updated_cpus; 1311 struct device *dev; 1312 int weight, new_nid, i = 0; 1313 1314 if (!prrn_enabled && !vphn_enabled && topology_inited) 1315 return 0; 1316 1317 weight = cpumask_weight(&cpu_associativity_changes_mask); 1318 if (!weight) 1319 return 0; 1320 1321 updates = kcalloc(weight, sizeof(*updates), GFP_KERNEL); 1322 if (!updates) 1323 return 0; 1324 1325 cpumask_clear(&updated_cpus); 1326 1327 for_each_cpu(cpu, &cpu_associativity_changes_mask) { 1328 /* 1329 * If siblings aren't flagged for changes, updates list 1330 * will be too short. Skip on this update and set for next 1331 * update. 1332 */ 1333 if (!cpumask_subset(cpu_sibling_mask(cpu), 1334 &cpu_associativity_changes_mask)) { 1335 pr_info("Sibling bits not set for associativity " 1336 "change, cpu%d\n", cpu); 1337 cpumask_or(&cpu_associativity_changes_mask, 1338 &cpu_associativity_changes_mask, 1339 cpu_sibling_mask(cpu)); 1340 cpu = cpu_last_thread_sibling(cpu); 1341 continue; 1342 } 1343 1344 new_nid = find_and_online_cpu_nid(cpu); 1345 1346 if (new_nid == numa_cpu_lookup_table[cpu]) { 1347 cpumask_andnot(&cpu_associativity_changes_mask, 1348 &cpu_associativity_changes_mask, 1349 cpu_sibling_mask(cpu)); 1350 dbg("Assoc chg gives same node %d for cpu%d\n", 1351 new_nid, cpu); 1352 cpu = cpu_last_thread_sibling(cpu); 1353 continue; 1354 } 1355 1356 for_each_cpu(sibling, cpu_sibling_mask(cpu)) { 1357 ud = &updates[i++]; 1358 ud->next = &updates[i]; 1359 ud->cpu = sibling; 1360 ud->new_nid = new_nid; 1361 ud->old_nid = numa_cpu_lookup_table[sibling]; 1362 cpumask_set_cpu(sibling, &updated_cpus); 1363 } 1364 cpu = cpu_last_thread_sibling(cpu); 1365 } 1366 1367 /* 1368 * Prevent processing of 'updates' from overflowing array 1369 * where last entry filled in a 'next' pointer. 1370 */ 1371 if (i) 1372 updates[i-1].next = NULL; 1373 1374 pr_debug("Topology update for the following CPUs:\n"); 1375 if (cpumask_weight(&updated_cpus)) { 1376 for (ud = &updates[0]; ud; ud = ud->next) { 1377 pr_debug("cpu %d moving from node %d " 1378 "to %d\n", ud->cpu, 1379 ud->old_nid, ud->new_nid); 1380 } 1381 } 1382 1383 /* 1384 * In cases where we have nothing to update (because the updates list 1385 * is too short or because the new topology is same as the old one), 1386 * skip invoking update_cpu_topology() via stop-machine(). This is 1387 * necessary (and not just a fast-path optimization) since stop-machine 1388 * can end up electing a random CPU to run update_cpu_topology(), and 1389 * thus trick us into setting up incorrect cpu-node mappings (since 1390 * 'updates' is kzalloc()'ed). 1391 * 1392 * And for the similar reason, we will skip all the following updating. 1393 */ 1394 if (!cpumask_weight(&updated_cpus)) 1395 goto out; 1396 1397 if (cpus_locked) 1398 stop_machine_cpuslocked(update_cpu_topology, &updates[0], 1399 &updated_cpus); 1400 else 1401 stop_machine(update_cpu_topology, &updates[0], &updated_cpus); 1402 1403 /* 1404 * Update the numa-cpu lookup table with the new mappings, even for 1405 * offline CPUs. It is best to perform this update from the stop- 1406 * machine context. 1407 */ 1408 if (cpus_locked) 1409 stop_machine_cpuslocked(update_lookup_table, &updates[0], 1410 cpumask_of(raw_smp_processor_id())); 1411 else 1412 stop_machine(update_lookup_table, &updates[0], 1413 cpumask_of(raw_smp_processor_id())); 1414 1415 for (ud = &updates[0]; ud; ud = ud->next) { 1416 unregister_cpu_under_node(ud->cpu, ud->old_nid); 1417 register_cpu_under_node(ud->cpu, ud->new_nid); 1418 1419 dev = get_cpu_device(ud->cpu); 1420 if (dev) 1421 kobject_uevent(&dev->kobj, KOBJ_CHANGE); 1422 cpumask_clear_cpu(ud->cpu, &cpu_associativity_changes_mask); 1423 changed = 1; 1424 } 1425 1426 out: 1427 kfree(updates); 1428 return changed; 1429 } 1430 1431 int arch_update_cpu_topology(void) 1432 { 1433 return numa_update_cpu_topology(true); 1434 } 1435 1436 static void topology_work_fn(struct work_struct *work) 1437 { 1438 rebuild_sched_domains(); 1439 } 1440 static DECLARE_WORK(topology_work, topology_work_fn); 1441 1442 static void topology_schedule_update(void) 1443 { 1444 schedule_work(&topology_work); 1445 } 1446 1447 static void topology_timer_fn(struct timer_list *unused) 1448 { 1449 if (prrn_enabled && cpumask_weight(&cpu_associativity_changes_mask)) 1450 topology_schedule_update(); 1451 else if (vphn_enabled) { 1452 if (update_cpu_associativity_changes_mask() > 0) 1453 topology_schedule_update(); 1454 reset_topology_timer(); 1455 } 1456 } 1457 static struct timer_list topology_timer; 1458 1459 static void reset_topology_timer(void) 1460 { 1461 if (vphn_enabled) 1462 mod_timer(&topology_timer, jiffies + topology_timer_secs * HZ); 1463 } 1464 1465 #ifdef CONFIG_SMP 1466 1467 static int dt_update_callback(struct notifier_block *nb, 1468 unsigned long action, void *data) 1469 { 1470 struct of_reconfig_data *update = data; 1471 int rc = NOTIFY_DONE; 1472 1473 switch (action) { 1474 case OF_RECONFIG_UPDATE_PROPERTY: 1475 if (of_node_is_type(update->dn, "cpu") && 1476 !of_prop_cmp(update->prop->name, "ibm,associativity")) { 1477 u32 core_id; 1478 of_property_read_u32(update->dn, "reg", &core_id); 1479 rc = dlpar_cpu_readd(core_id); 1480 rc = NOTIFY_OK; 1481 } 1482 break; 1483 } 1484 1485 return rc; 1486 } 1487 1488 static struct notifier_block dt_update_nb = { 1489 .notifier_call = dt_update_callback, 1490 }; 1491 1492 #endif 1493 1494 /* 1495 * Start polling for associativity changes. 1496 */ 1497 int start_topology_update(void) 1498 { 1499 int rc = 0; 1500 1501 if (firmware_has_feature(FW_FEATURE_PRRN)) { 1502 if (!prrn_enabled) { 1503 prrn_enabled = 1; 1504 #ifdef CONFIG_SMP 1505 rc = of_reconfig_notifier_register(&dt_update_nb); 1506 #endif 1507 } 1508 } 1509 if (firmware_has_feature(FW_FEATURE_VPHN) && 1510 lppaca_shared_proc(get_lppaca())) { 1511 if (!vphn_enabled) { 1512 vphn_enabled = 1; 1513 setup_cpu_associativity_change_counters(); 1514 timer_setup(&topology_timer, topology_timer_fn, 1515 TIMER_DEFERRABLE); 1516 reset_topology_timer(); 1517 } 1518 } 1519 1520 pr_info("Starting topology update%s%s\n", 1521 (prrn_enabled ? " prrn_enabled" : ""), 1522 (vphn_enabled ? " vphn_enabled" : "")); 1523 1524 return rc; 1525 } 1526 1527 /* 1528 * Disable polling for VPHN associativity changes. 1529 */ 1530 int stop_topology_update(void) 1531 { 1532 int rc = 0; 1533 1534 if (prrn_enabled) { 1535 prrn_enabled = 0; 1536 #ifdef CONFIG_SMP 1537 rc = of_reconfig_notifier_unregister(&dt_update_nb); 1538 #endif 1539 } 1540 if (vphn_enabled) { 1541 vphn_enabled = 0; 1542 rc = del_timer_sync(&topology_timer); 1543 } 1544 1545 pr_info("Stopping topology update\n"); 1546 1547 return rc; 1548 } 1549 1550 int prrn_is_enabled(void) 1551 { 1552 return prrn_enabled; 1553 } 1554 1555 void __init shared_proc_topology_init(void) 1556 { 1557 if (lppaca_shared_proc(get_lppaca())) { 1558 bitmap_fill(cpumask_bits(&cpu_associativity_changes_mask), 1559 nr_cpumask_bits); 1560 numa_update_cpu_topology(false); 1561 } 1562 } 1563 1564 static int topology_read(struct seq_file *file, void *v) 1565 { 1566 if (vphn_enabled || prrn_enabled) 1567 seq_puts(file, "on\n"); 1568 else 1569 seq_puts(file, "off\n"); 1570 1571 return 0; 1572 } 1573 1574 static int topology_open(struct inode *inode, struct file *file) 1575 { 1576 return single_open(file, topology_read, NULL); 1577 } 1578 1579 static ssize_t topology_write(struct file *file, const char __user *buf, 1580 size_t count, loff_t *off) 1581 { 1582 char kbuf[4]; /* "on" or "off" plus null. */ 1583 int read_len; 1584 1585 read_len = count < 3 ? count : 3; 1586 if (copy_from_user(kbuf, buf, read_len)) 1587 return -EINVAL; 1588 1589 kbuf[read_len] = '\0'; 1590 1591 if (!strncmp(kbuf, "on", 2)) 1592 start_topology_update(); 1593 else if (!strncmp(kbuf, "off", 3)) 1594 stop_topology_update(); 1595 else 1596 return -EINVAL; 1597 1598 return count; 1599 } 1600 1601 static const struct file_operations topology_ops = { 1602 .read = seq_read, 1603 .write = topology_write, 1604 .open = topology_open, 1605 .release = single_release 1606 }; 1607 1608 static int topology_update_init(void) 1609 { 1610 /* Do not poll for changes if disabled at boot */ 1611 if (topology_updates_enabled) 1612 start_topology_update(); 1613 1614 if (vphn_enabled) 1615 topology_schedule_update(); 1616 1617 if (!proc_create("powerpc/topology_updates", 0644, NULL, &topology_ops)) 1618 return -ENOMEM; 1619 1620 topology_inited = 1; 1621 return 0; 1622 } 1623 device_initcall(topology_update_init); 1624 #endif /* CONFIG_PPC_SPLPAR */ 1625