arch/powerpc/mm/numa.c

   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/export.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 <linux/cpuset.h>
  24 #include <linux/node.h>
  25 #include <linux/stop_machine.h>
  26 #include <linux/proc_fs.h>
  27 #include <linux/seq_file.h>
  28 #include <linux/uaccess.h>
  29 #include <linux/slab.h>
  30 #include <asm/cputhreads.h>
  31 #include <asm/sparsemem.h>
  32 #include <asm/prom.h>
  33 #include <asm/smp.h>
  34 #include <asm/firmware.h>
  35 #include <asm/paca.h>
  36 #include <asm/hvcall.h>
  37 #include <asm/setup.h>
  38 #include <asm/vdso.h>
  39
  40 static int numa_enabled = 1;
  41
  42 static char *cmdline __initdata;
  43
  44 static int numa_debug;
  45 #define dbg(args...) if (numa_debug) { printk(KERN_INFO args); }
  46
  47 int numa_cpu_lookup_table[NR_CPUS];
  48 cpumask_var_t node_to_cpumask_map[MAX_NUMNODES];
  49 struct pglist_data *node_data[MAX_NUMNODES];
  50
  51 EXPORT_SYMBOL(numa_cpu_lookup_table);
  52 EXPORT_SYMBOL(node_to_cpumask_map);
  53 EXPORT_SYMBOL(node_data);
  54
  55 static int min_common_depth;
  56 static int n_mem_addr_cells, n_mem_size_cells;
  57 static int form1_affinity;
  58
  59 #define MAX_DISTANCE_REF_POINTS 4
  60 static int distance_ref_points_depth;
  61 static const __be32 *distance_ref_points;
  62 static int distance_lookup_table[MAX_NUMNODES][MAX_DISTANCE_REF_POINTS];
  63
  64 /*
  65  * Allocate node_to_cpumask_map based on number of available nodes
  66  * Requires node_possible_map to be valid.
  67  *
  68  * Note: cpumask_of_node() is not valid until after this is done.
  69  */
  70 static void __init setup_node_to_cpumask_map(void)
  71 {
  72         unsigned int node;
  73
  74         /* setup nr_node_ids if not done yet */
  75         if (nr_node_ids == MAX_NUMNODES)
  76                 setup_nr_node_ids();
  77
  78         /* allocate the map */
  79         for (node = 0; node < nr_node_ids; node++)
  80                 alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]);
  81
  82         /* cpumask_of_node() will now work */
  83         dbg("Node to cpumask map for %d nodes\n", nr_node_ids);
  84 }
  85
  86 static int __init fake_numa_create_new_node(unsigned long end_pfn,
  87                                                 unsigned int *nid)
  88 {
  89         unsigned long long mem;
  90         char *p = cmdline;
  91         static unsigned int fake_nid;
  92         static unsigned long long curr_boundary;
  93
  94         /*
  95          * Modify node id, iff we started creating NUMA nodes
  96          * We want to continue from where we left of the last time
  97          */
  98         if (fake_nid)
  99                 *nid = fake_nid;
 100         /*
 101          * In case there are no more arguments to parse, the
 102          * node_id should be the same as the last fake node id
 103          * (we've handled this above).
 104          */
 105         if (!p)
 106                 return 0;
 107
 108         mem = memparse(p, &p);
 109         if (!mem)
 110                 return 0;
 111
 112         if (mem < curr_boundary)
 113                 return 0;
 114
 115         curr_boundary = mem;
 116
 117         if ((end_pfn << PAGE_SHIFT) > mem) {
 118                 /*
 119                  * Skip commas and spaces
 120                  */
 121                 while (*p == ',' || *p == ' ' || *p == '\t')
 122                         p++;
 123
 124                 cmdline = p;
 125                 fake_nid++;
 126                 *nid = fake_nid;
 127                 dbg("created new fake_node with id %d\n", fake_nid);
 128                 return 1;
 129         }
 130         return 0;
 131 }
 132
 133 /*
 134  * get_node_active_region - Return active region containing pfn
 135  * Active range returned is empty if none found.
 136  * @pfn: The page to return the region for
 137  * @node_ar: Returned set to the active region containing @pfn
 138  */
 139 static void __init get_node_active_region(unsigned long pfn,
 140                                           struct node_active_region *node_ar)
 141 {
 142         unsigned long start_pfn, end_pfn;
 143         int i, nid;
 144
 145         for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
 146                 if (pfn >= start_pfn && pfn < end_pfn) {
 147                         node_ar->nid = nid;
 148                         node_ar->start_pfn = start_pfn;
 149                         node_ar->end_pfn = end_pfn;
 150                         break;
 151                 }
 152         }
 153 }
 154
 155 static void map_cpu_to_node(int cpu, int node)
 156 {
 157         numa_cpu_lookup_table[cpu] = node;
 158
 159         dbg("adding cpu %d to node %d\n", cpu, node);
 160
 161         if (!(cpumask_test_cpu(cpu, node_to_cpumask_map[node])))
 162                 cpumask_set_cpu(cpu, node_to_cpumask_map[node]);
 163 }
 164
 165 #if defined(CONFIG_HOTPLUG_CPU) || defined(CONFIG_PPC_SPLPAR)
 166 static void unmap_cpu_from_node(unsigned long cpu)
 167 {
 168         int node = numa_cpu_lookup_table[cpu];
 169
 170         dbg("removing cpu %lu from node %d\n", cpu, node);
 171
 172         if (cpumask_test_cpu(cpu, node_to_cpumask_map[node])) {
 173                 cpumask_clear_cpu(cpu, node_to_cpumask_map[node]);
 174         } else {
 175                 printk(KERN_ERR "WARNING: cpu %lu not found in node %d\n",
 176                        cpu, node);
 177         }
 178 }
 179 #endif /* CONFIG_HOTPLUG_CPU || CONFIG_PPC_SPLPAR */
 180
 181 /* must hold reference to node during call */
 182 static const __be32 *of_get_associativity(struct device_node *dev)
 183 {
 184         return of_get_property(dev, "ibm,associativity", NULL);
 185 }
 186
 187 /*
 188  * Returns the property linux,drconf-usable-memory if
 189  * it exists (the property exists only in kexec/kdump kernels,
 190  * added by kexec-tools)
 191  */
 192 static const __be32 *of_get_usable_memory(struct device_node *memory)
 193 {
 194         const __be32 *prop;
 195         u32 len;
 196         prop = of_get_property(memory, "linux,drconf-usable-memory", &len);
 197         if (!prop || len < sizeof(unsigned int))
 198                 return NULL;
 199         return prop;
 200 }
 201
 202 int __node_distance(int a, int b)
 203 {
 204         int i;
 205         int distance = LOCAL_DISTANCE;
 206
 207         if (!form1_affinity)
 208                 return ((a == b) ? LOCAL_DISTANCE : REMOTE_DISTANCE);
 209
 210         for (i = 0; i < distance_ref_points_depth; i++) {
 211                 if (distance_lookup_table[a][i] == distance_lookup_table[b][i])
 212                         break;
 213
 214                 /* Double the distance for each NUMA level */
 215                 distance *= 2;
 216         }
 217
 218         return distance;
 219 }
 220
 221 static void initialize_distance_lookup_table(int nid,
 222                 const __be32 *associativity)
 223 {
 224         int i;
 225
 226         if (!form1_affinity)
 227                 return;
 228
 229         for (i = 0; i < distance_ref_points_depth; i++) {
 230                 const __be32 *entry;
 231
 232                 entry = &associativity[be32_to_cpu(distance_ref_points[i])];
 233                 distance_lookup_table[nid][i] = of_read_number(entry, 1);
 234         }
 235 }
 236
 237 /* Returns nid in the range [0..MAX_NUMNODES-1], or -1 if no useful numa
 238  * info is found.
 239  */
 240 static int associativity_to_nid(const __be32 *associativity)
 241 {
 242         int nid = -1;
 243
 244         if (min_common_depth == -1)
 245                 goto out;
 246
 247         if (of_read_number(associativity, 1) >= min_common_depth)
 248                 nid = of_read_number(&associativity[min_common_depth], 1);
 249
 250         /* POWER4 LPAR uses 0xffff as invalid node */
 251         if (nid == 0xffff || nid >= MAX_NUMNODES)
 252                 nid = -1;
 253
 254         if (nid > 0 &&
 255             of_read_number(associativity, 1) >= distance_ref_points_depth)
 256                 initialize_distance_lookup_table(nid, associativity);
 257
 258 out:
 259         return nid;
 260 }
 261
 262 /* Returns the nid associated with the given device tree node,
 263  * or -1 if not found.
 264  */
 265 static int of_node_to_nid_single(struct device_node *device)
 266 {
 267         int nid = -1;
 268         const __be32 *tmp;
 269
 270         tmp = of_get_associativity(device);
 271         if (tmp)
 272                 nid = associativity_to_nid(tmp);
 273         return nid;
 274 }
 275
 276 /* Walk the device tree upwards, looking for an associativity id */
 277 int of_node_to_nid(struct device_node *device)
 278 {
 279         struct device_node *tmp;
 280         int nid = -1;
 281
 282         of_node_get(device);
 283         while (device) {
 284                 nid = of_node_to_nid_single(device);
 285                 if (nid != -1)
 286                         break;
 287
 288                 tmp = device;
 289                 device = of_get_parent(tmp);
 290                 of_node_put(tmp);
 291         }
 292         of_node_put(device);
 293
 294         return nid;
 295 }
 296 EXPORT_SYMBOL_GPL(of_node_to_nid);
 297
 298 static int __init find_min_common_depth(void)
 299 {
 300         int depth;
 301         struct device_node *root;
 302
 303         if (firmware_has_feature(FW_FEATURE_OPAL))
 304                 root = of_find_node_by_path("/ibm,opal");
 305         else
 306                 root = of_find_node_by_path("/rtas");
 307         if (!root)
 308                 root = of_find_node_by_path("/");
 309
 310         /*
 311          * This property is a set of 32-bit integers, each representing
 312          * an index into the ibm,associativity nodes.
 313          *
 314          * With form 0 affinity the first integer is for an SMP configuration
 315          * (should be all 0's) and the second is for a normal NUMA
 316          * configuration. We have only one level of NUMA.
 317          *
 318          * With form 1 affinity the first integer is the most significant
 319          * NUMA boundary and the following are progressively less significant
 320          * boundaries. There can be more than one level of NUMA.
 321          */
 322         distance_ref_points = of_get_property(root,
 323                                         "ibm,associativity-reference-points",
 324                                         &distance_ref_points_depth);
 325
 326         if (!distance_ref_points) {
 327                 dbg("NUMA: ibm,associativity-reference-points not found.\n");
 328                 goto err;
 329         }
 330
 331         distance_ref_points_depth /= sizeof(int);
 332
 333         if (firmware_has_feature(FW_FEATURE_OPAL) ||
 334             firmware_has_feature(FW_FEATURE_TYPE1_AFFINITY)) {
 335                 dbg("Using form 1 affinity\n");
 336                 form1_affinity = 1;
 337         }
 338
 339         if (form1_affinity) {
 340                 depth = of_read_number(distance_ref_points, 1);
 341         } else {
 342                 if (distance_ref_points_depth < 2) {
 343                         printk(KERN_WARNING "NUMA: "
 344                                 "short ibm,associativity-reference-points\n");
 345                         goto err;
 346                 }
 347
 348                 depth = of_read_number(&distance_ref_points[1], 1);
 349         }
 350
 351         /*
 352          * Warn and cap if the hardware supports more than
 353          * MAX_DISTANCE_REF_POINTS domains.
 354          */
 355         if (distance_ref_points_depth > MAX_DISTANCE_REF_POINTS) {
 356                 printk(KERN_WARNING "NUMA: distance array capped at "
 357                         "%d entries\n", MAX_DISTANCE_REF_POINTS);
 358                 distance_ref_points_depth = MAX_DISTANCE_REF_POINTS;
 359         }
 360
 361         of_node_put(root);
 362         return depth;
 363
 364 err:
 365         of_node_put(root);
 366         return -1;
 367 }
 368
 369 static void __init get_n_mem_cells(int *n_addr_cells, int *n_size_cells)
 370 {
 371         struct device_node *memory = NULL;
 372
 373         memory = of_find_node_by_type(memory, "memory");
 374         if (!memory)
 375                 panic("numa.c: No memory nodes found!");
 376
 377         *n_addr_cells = of_n_addr_cells(memory);
 378         *n_size_cells = of_n_size_cells(memory);
 379         of_node_put(memory);
 380 }
 381
 382 static unsigned long read_n_cells(int n, const __be32 **buf)
 383 {
 384         unsigned long result = 0;
 385
 386         while (n--) {
 387                 result = (result << 32) | of_read_number(*buf, 1);
 388                 (*buf)++;
 389         }
 390         return result;
 391 }
 392
 393 /*
 394  * Read the next memblock list entry from the ibm,dynamic-memory property
 395  * and return the information in the provided of_drconf_cell structure.
 396  */
 397 static void read_drconf_cell(struct of_drconf_cell *drmem, const __be32 **cellp)
 398 {
 399         const __be32 *cp;
 400
 401         drmem->base_addr = read_n_cells(n_mem_addr_cells, cellp);
 402
 403         cp = *cellp;
 404         drmem->drc_index = of_read_number(cp, 1);
 405         drmem->reserved = of_read_number(&cp[1], 1);
 406         drmem->aa_index = of_read_number(&cp[2], 1);
 407         drmem->flags = of_read_number(&cp[3], 1);
 408
 409         *cellp = cp + 4;
 410 }
 411
 412 /*
 413  * Retrieve and validate the ibm,dynamic-memory property of the device tree.
 414  *
 415  * The layout of the ibm,dynamic-memory property is a number N of memblock
 416  * list entries followed by N memblock list entries.  Each memblock list entry
 417  * contains information as laid out in the of_drconf_cell struct above.
 418  */
 419 static int of_get_drconf_memory(struct device_node *memory, const __be32 **dm)
 420 {
 421         const __be32 *prop;
 422         u32 len, entries;
 423
 424         prop = of_get_property(memory, "ibm,dynamic-memory", &len);
 425         if (!prop || len < sizeof(unsigned int))
 426                 return 0;
 427
 428         entries = of_read_number(prop++, 1);
 429
 430         /* Now that we know the number of entries, revalidate the size
 431          * of the property read in to ensure we have everything
 432          */
 433         if (len < (entries * (n_mem_addr_cells + 4) + 1) * sizeof(unsigned int))
 434                 return 0;
 435
 436         *dm = prop;
 437         return entries;
 438 }
 439
 440 /*
 441  * Retrieve and validate the ibm,lmb-size property for drconf memory
 442  * from the device tree.
 443  */
 444 static u64 of_get_lmb_size(struct device_node *memory)
 445 {
 446         const __be32 *prop;
 447         u32 len;
 448
 449         prop = of_get_property(memory, "ibm,lmb-size", &len);
 450         if (!prop || len < sizeof(unsigned int))
 451                 return 0;
 452
 453         return read_n_cells(n_mem_size_cells, &prop);
 454 }
 455
 456 struct assoc_arrays {
 457         u32     n_arrays;
 458         u32     array_sz;
 459         const __be32 *arrays;
 460 };
 461
 462 /*
 463  * Retrieve and validate the list of associativity arrays for drconf
 464  * memory from the ibm,associativity-lookup-arrays property of the
 465  * device tree..
 466  *
 467  * The layout of the ibm,associativity-lookup-arrays property is a number N
 468  * indicating the number of associativity arrays, followed by a number M
 469  * indicating the size of each associativity array, followed by a list
 470  * of N associativity arrays.
 471  */
 472 static int of_get_assoc_arrays(struct device_node *memory,
 473                                struct assoc_arrays *aa)
 474 {
 475         const __be32 *prop;
 476         u32 len;
 477
 478         prop = of_get_property(memory, "ibm,associativity-lookup-arrays", &len);
 479         if (!prop || len < 2 * sizeof(unsigned int))
 480                 return -1;
 481
 482         aa->n_arrays = of_read_number(prop++, 1);
 483         aa->array_sz = of_read_number(prop++, 1);
 484
 485         /* Now that we know the number of arrays and size of each array,
 486          * revalidate the size of the property read in.
 487          */
 488         if (len < (aa->n_arrays * aa->array_sz + 2) * sizeof(unsigned int))
 489                 return -1;
 490
 491         aa->arrays = prop;
 492         return 0;
 493 }
 494
 495 /*
 496  * This is like of_node_to_nid_single() for memory represented in the
 497  * ibm,dynamic-reconfiguration-memory node.
 498  */
 499 static int of_drconf_to_nid_single(struct of_drconf_cell *drmem,
 500                                    struct assoc_arrays *aa)
 501 {
 502         int default_nid = 0;
 503         int nid = default_nid;
 504         int index;
 505
 506         if (min_common_depth > 0 && min_common_depth <= aa->array_sz &&
 507             !(drmem->flags & DRCONF_MEM_AI_INVALID) &&
 508             drmem->aa_index < aa->n_arrays) {
 509                 index = drmem->aa_index * aa->array_sz + min_common_depth - 1;
 510                 nid = of_read_number(&aa->arrays[index], 1);
 511
 512                 if (nid == 0xffff || nid >= MAX_NUMNODES)
 513                         nid = default_nid;
 514         }
 515
 516         return nid;
 517 }
 518
 519 /*
 520  * Figure out to which domain a cpu belongs and stick it there.
 521  * Return the id of the domain used.
 522  */
 523 static int numa_setup_cpu(unsigned long lcpu)
 524 {
 525         int nid = 0;
 526         struct device_node *cpu = of_get_cpu_node(lcpu, NULL);
 527
 528         if (!cpu) {
 529                 WARN_ON(1);
 530                 goto out;
 531         }
 532
 533         nid = of_node_to_nid_single(cpu);
 534
 535         if (nid < 0 || !node_online(nid))
 536                 nid = first_online_node;
 537 out:
 538         map_cpu_to_node(lcpu, nid);
 539
 540         of_node_put(cpu);
 541
 542         return nid;
 543 }
 544
 545 static int cpu_numa_callback(struct notifier_block *nfb, unsigned long action,
 546                              void *hcpu)
 547 {
 548         unsigned long lcpu = (unsigned long)hcpu;
 549         int ret = NOTIFY_DONE;
 550
 551         switch (action) {
 552         case CPU_UP_PREPARE:
 553         case CPU_UP_PREPARE_FROZEN:
 554                 numa_setup_cpu(lcpu);
 555                 ret = NOTIFY_OK;
 556                 break;
 557 #ifdef CONFIG_HOTPLUG_CPU
 558         case CPU_DEAD:
 559         case CPU_DEAD_FROZEN:
 560         case CPU_UP_CANCELED:
 561         case CPU_UP_CANCELED_FROZEN:
 562                 unmap_cpu_from_node(lcpu);
 563                 break;
 564                 ret = NOTIFY_OK;
 565 #endif
 566         }
 567         return ret;
 568 }
 569
 570 /*
 571  * Check and possibly modify a memory region to enforce the memory limit.
 572  *
 573  * Returns the size the region should have to enforce the memory limit.
 574  * This will either be the original value of size, a truncated value,
 575  * or zero. If the returned value of size is 0 the region should be
 576  * discarded as it lies wholly above the memory limit.
 577  */
 578 static unsigned long __init numa_enforce_memory_limit(unsigned long start,
 579                                                       unsigned long size)
 580 {
 581         /*
 582          * We use memblock_end_of_DRAM() in here instead of memory_limit because
 583          * we've already adjusted it for the limit and it takes care of
 584          * having memory holes below the limit.  Also, in the case of
 585          * iommu_is_off, memory_limit is not set but is implicitly enforced.
 586          */
 587
 588         if (start + size <= memblock_end_of_DRAM())
 589                 return size;
 590
 591         if (start >= memblock_end_of_DRAM())
 592                 return 0;
 593
 594         return memblock_end_of_DRAM() - start;
 595 }
 596
 597 /*
 598  * Reads the counter for a given entry in
 599  * linux,drconf-usable-memory property
 600  */
 601 static inline int __init read_usm_ranges(const __be32 **usm)
 602 {
 603         /*
 604          * For each lmb in ibm,dynamic-memory a corresponding
 605          * entry in linux,drconf-usable-memory property contains
 606          * a counter followed by that many (base, size) duple.
 607          * read the counter from linux,drconf-usable-memory
 608          */
 609         return read_n_cells(n_mem_size_cells, usm);
 610 }
 611
 612 /*
 613  * Extract NUMA information from the ibm,dynamic-reconfiguration-memory
 614  * node.  This assumes n_mem_{addr,size}_cells have been set.
 615  */
 616 static void __init parse_drconf_memory(struct device_node *memory)
 617 {
 618         const __be32 *uninitialized_var(dm), *usm;
 619         unsigned int n, rc, ranges, is_kexec_kdump = 0;
 620         unsigned long lmb_size, base, size, sz;
 621         int nid;
 622         struct assoc_arrays aa = { .arrays = NULL };
 623
 624         n = of_get_drconf_memory(memory, &dm);
 625         if (!n)
 626                 return;
 627
 628         lmb_size = of_get_lmb_size(memory);
 629         if (!lmb_size)
 630                 return;
 631
 632         rc = of_get_assoc_arrays(memory, &aa);
 633         if (rc)
 634                 return;
 635
 636         /* check if this is a kexec/kdump kernel */
 637         usm = of_get_usable_memory(memory);
 638         if (usm != NULL)
 639                 is_kexec_kdump = 1;
 640
 641         for (; n != 0; --n) {
 642                 struct of_drconf_cell drmem;
 643
 644                 read_drconf_cell(&drmem, &dm);
 645
 646                 /* skip this block if the reserved bit is set in flags (0x80)
 647                    or if the block is not assigned to this partition (0x8) */
 648                 if ((drmem.flags & DRCONF_MEM_RESERVED)
 649                     || !(drmem.flags & DRCONF_MEM_ASSIGNED))
 650                         continue;
 651
 652                 base = drmem.base_addr;
 653                 size = lmb_size;
 654                 ranges = 1;
 655
 656                 if (is_kexec_kdump) {
 657                         ranges = read_usm_ranges(&usm);
 658                         if (!ranges) /* there are no (base, size) duple */
 659                                 continue;
 660                 }
 661                 do {
 662                         if (is_kexec_kdump) {
 663                                 base = read_n_cells(n_mem_addr_cells, &usm);
 664                                 size = read_n_cells(n_mem_size_cells, &usm);
 665                         }
 666                         nid = of_drconf_to_nid_single(&drmem, &aa);
 667                         fake_numa_create_new_node(
 668                                 ((base + size) >> PAGE_SHIFT),
 669                                            &nid);
 670                         node_set_online(nid);
 671                         sz = numa_enforce_memory_limit(base, size);
 672                         if (sz)
 673                                 memblock_set_node(base, sz,
 674                                                   &memblock.memory, nid);
 675                 } while (--ranges);
 676         }
 677 }
 678
 679 static int __init parse_numa_properties(void)
 680 {
 681         struct device_node *memory;
 682         int default_nid = 0;
 683         unsigned long i;
 684
 685         if (numa_enabled == 0) {
 686                 printk(KERN_WARNING "NUMA disabled by user\n");
 687                 return -1;
 688         }
 689
 690         min_common_depth = find_min_common_depth();
 691
 692         if (min_common_depth < 0)
 693                 return min_common_depth;
 694
 695         dbg("NUMA associativity depth for CPU/Memory: %d\n", min_common_depth);
 696
 697         /*
 698          * Even though we connect cpus to numa domains later in SMP
 699          * init, we need to know the node ids now. This is because
 700          * each node to be onlined must have NODE_DATA etc backing it.
 701          */
 702         for_each_present_cpu(i) {
 703                 struct device_node *cpu;
 704                 int nid;
 705
 706                 cpu = of_get_cpu_node(i, NULL);
 707                 BUG_ON(!cpu);
 708                 nid = of_node_to_nid_single(cpu);
 709                 of_node_put(cpu);
 710
 711                 /*
 712                  * Don't fall back to default_nid yet -- we will plug
 713                  * cpus into nodes once the memory scan has discovered
 714                  * the topology.
 715                  */
 716                 if (nid < 0)
 717                         continue;
 718                 node_set_online(nid);
 719         }
 720
 721         get_n_mem_cells(&n_mem_addr_cells, &n_mem_size_cells);
 722
 723         for_each_node_by_type(memory, "memory") {
 724                 unsigned long start;
 725                 unsigned long size;
 726                 int nid;
 727                 int ranges;
 728                 const __be32 *memcell_buf;
 729                 unsigned int len;
 730
 731                 memcell_buf = of_get_property(memory,
 732                         "linux,usable-memory", &len);
 733                 if (!memcell_buf || len <= 0)
 734                         memcell_buf = of_get_property(memory, "reg", &len);
 735                 if (!memcell_buf || len <= 0)
 736                         continue;
 737
 738                 /* ranges in cell */
 739                 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
 740 new_range:
 741                 /* these are order-sensitive, and modify the buffer pointer */
 742                 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
 743                 size = read_n_cells(n_mem_size_cells, &memcell_buf);
 744
 745                 /*
 746                  * Assumption: either all memory nodes or none will
 747                  * have associativity properties.  If none, then
 748                  * everything goes to default_nid.
 749                  */
 750                 nid = of_node_to_nid_single(memory);
 751                 if (nid < 0)
 752                         nid = default_nid;
 753
 754                 fake_numa_create_new_node(((start + size) >> PAGE_SHIFT), &nid);
 755                 node_set_online(nid);
 756
 757                 if (!(size = numa_enforce_memory_limit(start, size))) {
 758                         if (--ranges)
 759                                 goto new_range;
 760                         else
 761                                 continue;
 762                 }
 763
 764                 memblock_set_node(start, size, &memblock.memory, nid);
 765
 766                 if (--ranges)
 767                         goto new_range;
 768         }
 769
 770         /*
 771          * Now do the same thing for each MEMBLOCK listed in the
 772          * ibm,dynamic-memory property in the
 773          * ibm,dynamic-reconfiguration-memory node.
 774          */
 775         memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
 776         if (memory)
 777                 parse_drconf_memory(memory);
 778
 779         return 0;
 780 }
 781
 782 static void __init setup_nonnuma(void)
 783 {
 784         unsigned long top_of_ram = memblock_end_of_DRAM();
 785         unsigned long total_ram = memblock_phys_mem_size();
 786         unsigned long start_pfn, end_pfn;
 787         unsigned int nid = 0;
 788         struct memblock_region *reg;
 789
 790         printk(KERN_DEBUG "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
 791                top_of_ram, total_ram);
 792         printk(KERN_DEBUG "Memory hole size: %ldMB\n",
 793                (top_of_ram - total_ram) >> 20);
 794
 795         for_each_memblock(memory, reg) {
 796                 start_pfn = memblock_region_memory_base_pfn(reg);
 797                 end_pfn = memblock_region_memory_end_pfn(reg);
 798
 799                 fake_numa_create_new_node(end_pfn, &nid);
 800                 memblock_set_node(PFN_PHYS(start_pfn),
 801                                   PFN_PHYS(end_pfn - start_pfn),
 802                                   &memblock.memory, nid);
 803                 node_set_online(nid);
 804         }
 805 }
 806
 807 void __init dump_numa_cpu_topology(void)
 808 {
 809         unsigned int node;
 810         unsigned int cpu, count;
 811
 812         if (min_common_depth == -1 || !numa_enabled)
 813                 return;
 814
 815         for_each_online_node(node) {
 816                 printk(KERN_DEBUG "Node %d CPUs:", node);
 817
 818                 count = 0;
 819                 /*
 820                  * If we used a CPU iterator here we would miss printing
 821                  * the holes in the cpumap.
 822                  */
 823                 for (cpu = 0; cpu < nr_cpu_ids; cpu++) {
 824                         if (cpumask_test_cpu(cpu,
 825                                         node_to_cpumask_map[node])) {
 826                                 if (count == 0)
 827                                         printk(" %u", cpu);
 828                                 ++count;
 829                         } else {
 830                                 if (count > 1)
 831                                         printk("-%u", cpu - 1);
 832                                 count = 0;
 833                         }
 834                 }
 835
 836                 if (count > 1)
 837                         printk("-%u", nr_cpu_ids - 1);
 838                 printk("\n");
 839         }
 840 }
 841
 842 static void __init dump_numa_memory_topology(void)
 843 {
 844         unsigned int node;
 845         unsigned int count;
 846
 847         if (min_common_depth == -1 || !numa_enabled)
 848                 return;
 849
 850         for_each_online_node(node) {
 851                 unsigned long i;
 852
 853                 printk(KERN_DEBUG "Node %d Memory:", node);
 854
 855                 count = 0;
 856
 857                 for (i = 0; i < memblock_end_of_DRAM();
 858                      i += (1 << SECTION_SIZE_BITS)) {
 859                         if (early_pfn_to_nid(i >> PAGE_SHIFT) == node) {
 860                                 if (count == 0)
 861                                         printk(" 0x%lx", i);
 862                                 ++count;
 863                         } else {
 864                                 if (count > 0)
 865                                         printk("-0x%lx", i);
 866                                 count = 0;
 867                         }
 868                 }
 869
 870                 if (count > 0)
 871                         printk("-0x%lx", i);
 872                 printk("\n");
 873         }
 874 }
 875
 876 /*
 877  * Allocate some memory, satisfying the memblock or bootmem allocator where
 878  * required. nid is the preferred node and end is the physical address of
 879  * the highest address in the node.
 880  *
 881  * Returns the virtual address of the memory.
 882  */
 883 static void __init *careful_zallocation(int nid, unsigned long size,
 884                                        unsigned long align,
 885                                        unsigned long end_pfn)
 886 {
 887         void *ret;
 888         int new_nid;
 889         unsigned long ret_paddr;
 890
 891         ret_paddr = __memblock_alloc_base(size, align, end_pfn << PAGE_SHIFT);
 892
 893         /* retry over all memory */
 894         if (!ret_paddr)
 895                 ret_paddr = __memblock_alloc_base(size, align, memblock_end_of_DRAM());
 896
 897         if (!ret_paddr)
 898                 panic("numa.c: cannot allocate %lu bytes for node %d",
 899                       size, nid);
 900
 901         ret = __va(ret_paddr);
 902
 903         /*
 904          * We initialize the nodes in numeric order: 0, 1, 2...
 905          * and hand over control from the MEMBLOCK allocator to the
 906          * bootmem allocator.  If this function is called for
 907          * node 5, then we know that all nodes <5 are using the
 908          * bootmem allocator instead of the MEMBLOCK allocator.
 909          *
 910          * So, check the nid from which this allocation came
 911          * and double check to see if we need to use bootmem
 912          * instead of the MEMBLOCK.  We don't free the MEMBLOCK memory
 913          * since it would be useless.
 914          */
 915         new_nid = early_pfn_to_nid(ret_paddr >> PAGE_SHIFT);
 916         if (new_nid < nid) {
 917                 ret = __alloc_bootmem_node(NODE_DATA(new_nid),
 918                                 size, align, 0);
 919
 920                 dbg("alloc_bootmem %p %lx\n", ret, size);
 921         }
 922
 923         memset(ret, 0, size);
 924         return ret;
 925 }
 926
 927 static struct notifier_block ppc64_numa_nb = {
 928         .notifier_call = cpu_numa_callback,
 929         .priority = 1 /* Must run before sched domains notifier. */
 930 };
 931
 932 static void __init mark_reserved_regions_for_nid(int nid)
 933 {
 934         struct pglist_data *node = NODE_DATA(nid);
 935         struct memblock_region *reg;
 936
 937         for_each_memblock(reserved, reg) {
 938                 unsigned long physbase = reg->base;
 939                 unsigned long size = reg->size;
 940                 unsigned long start_pfn = physbase >> PAGE_SHIFT;
 941                 unsigned long end_pfn = PFN_UP(physbase + size);
 942                 struct node_active_region node_ar;
 943                 unsigned long node_end_pfn = pgdat_end_pfn(node);
 944
 945                 /*
 946                  * Check to make sure that this memblock.reserved area is
 947                  * within the bounds of the node that we care about.
 948                  * Checking the nid of the start and end points is not
 949                  * sufficient because the reserved area could span the
 950                  * entire node.
 951                  */
 952                 if (end_pfn <= node->node_start_pfn ||
 953                     start_pfn >= node_end_pfn)
 954                         continue;
 955
 956                 get_node_active_region(start_pfn, &node_ar);
 957                 while (start_pfn < end_pfn &&
 958                         node_ar.start_pfn < node_ar.end_pfn) {
 959                         unsigned long reserve_size = size;
 960                         /*
 961                          * if reserved region extends past active region
 962                          * then trim size to active region
 963                          */
 964                         if (end_pfn > node_ar.end_pfn)
 965                                 reserve_size = (node_ar.end_pfn << PAGE_SHIFT)
 966                                         - physbase;
 967                         /*
 968                          * Only worry about *this* node, others may not
 969                          * yet have valid NODE_DATA().
 970                          */
 971                         if (node_ar.nid == nid) {
 972                                 dbg("reserve_bootmem %lx %lx nid=%d\n",
 973                                         physbase, reserve_size, node_ar.nid);
 974                                 reserve_bootmem_node(NODE_DATA(node_ar.nid),
 975                                                 physbase, reserve_size,
 976                                                 BOOTMEM_DEFAULT);
 977                         }
 978                         /*
 979                          * if reserved region is contained in the active region
 980                          * then done.
 981                          */
 982                         if (end_pfn <= node_ar.end_pfn)
 983                                 break;
 984
 985                         /*
 986                          * reserved region extends past the active region
 987                          *   get next active region that contains this
 988                          *   reserved region
 989                          */
 990                         start_pfn = node_ar.end_pfn;
 991                         physbase = start_pfn << PAGE_SHIFT;
 992                         size = size - reserve_size;
 993                         get_node_active_region(start_pfn, &node_ar);
 994                 }
 995         }
 996 }
 997
 998
 999 void __init do_init_bootmem(void)
1000 {
1001         int nid;
1002
1003         min_low_pfn = 0;
1004         max_low_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
1005         max_pfn = max_low_pfn;
1006
1007         if (parse_numa_properties())
1008                 setup_nonnuma();
1009         else
1010                 dump_numa_memory_topology();
1011
1012         for_each_online_node(nid) {
1013                 unsigned long start_pfn, end_pfn;
1014                 void *bootmem_vaddr;
1015                 unsigned long bootmap_pages;
1016
1017                 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
1018
1019                 /*
1020                  * Allocate the node structure node local if possible
1021                  *
1022                  * Be careful moving this around, as it relies on all
1023                  * previous nodes' bootmem to be initialized and have
1024                  * all reserved areas marked.
1025                  */
1026                 NODE_DATA(nid) = careful_zallocation(nid,
1027                                         sizeof(struct pglist_data),
1028                                         SMP_CACHE_BYTES, end_pfn);
1029
1030                 dbg("node %d\n", nid);
1031                 dbg("NODE_DATA() = %p\n", NODE_DATA(nid));
1032
1033                 NODE_DATA(nid)->bdata = &bootmem_node_data[nid];
1034                 NODE_DATA(nid)->node_start_pfn = start_pfn;
1035                 NODE_DATA(nid)->node_spanned_pages = end_pfn - start_pfn;
1036
1037                 if (NODE_DATA(nid)->node_spanned_pages == 0)
1038                         continue;
1039
1040                 dbg("start_paddr = %lx\n", start_pfn << PAGE_SHIFT);
1041                 dbg("end_paddr = %lx\n", end_pfn << PAGE_SHIFT);
1042
1043                 bootmap_pages = bootmem_bootmap_pages(end_pfn - start_pfn);
1044                 bootmem_vaddr = careful_zallocation(nid,
1045                                         bootmap_pages << PAGE_SHIFT,
1046                                         PAGE_SIZE, end_pfn);
1047
1048                 dbg("bootmap_vaddr = %p\n", bootmem_vaddr);
1049
1050                 init_bootmem_node(NODE_DATA(nid),
1051                                   __pa(bootmem_vaddr) >> PAGE_SHIFT,
1052                                   start_pfn, end_pfn);
1053
1054                 free_bootmem_with_active_regions(nid, end_pfn);
1055                 /*
1056                  * Be very careful about moving this around.  Future
1057                  * calls to careful_zallocation() depend on this getting
1058                  * done correctly.
1059                  */
1060                 mark_reserved_regions_for_nid(nid);
1061                 sparse_memory_present_with_active_regions(nid);
1062         }
1063
1064         init_bootmem_done = 1;
1065
1066         /*
1067          * Now bootmem is initialised we can create the node to cpumask
1068          * lookup tables and setup the cpu callback to populate them.
1069          */
1070         setup_node_to_cpumask_map();
1071
1072         register_cpu_notifier(&ppc64_numa_nb);
1073         cpu_numa_callback(&ppc64_numa_nb, CPU_UP_PREPARE,
1074                           (void *)(unsigned long)boot_cpuid);
1075 }
1076
1077 void __init paging_init(void)
1078 {
1079         unsigned long max_zone_pfns[MAX_NR_ZONES];
1080         memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
1081         max_zone_pfns[ZONE_DMA] = memblock_end_of_DRAM() >> PAGE_SHIFT;
1082         free_area_init_nodes(max_zone_pfns);
1083 }
1084
1085 static int __init early_numa(char *p)
1086 {
1087         if (!p)
1088                 return 0;
1089
1090         if (strstr(p, "off"))
1091                 numa_enabled = 0;
1092
1093         if (strstr(p, "debug"))
1094                 numa_debug = 1;
1095
1096         p = strstr(p, "fake=");
1097         if (p)
1098                 cmdline = p + strlen("fake=");
1099
1100         return 0;
1101 }
1102 early_param("numa", early_numa);
1103
1104 #ifdef CONFIG_MEMORY_HOTPLUG
1105 /*
1106  * Find the node associated with a hot added memory section for
1107  * memory represented in the device tree by the property
1108  * ibm,dynamic-reconfiguration-memory/ibm,dynamic-memory.
1109  */
1110 static int hot_add_drconf_scn_to_nid(struct device_node *memory,
1111                                      unsigned long scn_addr)
1112 {
1113         const __be32 *dm;
1114         unsigned int drconf_cell_cnt, rc;
1115         unsigned long lmb_size;
1116         struct assoc_arrays aa;
1117         int nid = -1;
1118
1119         drconf_cell_cnt = of_get_drconf_memory(memory, &dm);
1120         if (!drconf_cell_cnt)
1121                 return -1;
1122
1123         lmb_size = of_get_lmb_size(memory);
1124         if (!lmb_size)
1125                 return -1;
1126
1127         rc = of_get_assoc_arrays(memory, &aa);
1128         if (rc)
1129                 return -1;
1130
1131         for (; drconf_cell_cnt != 0; --drconf_cell_cnt) {
1132                 struct of_drconf_cell drmem;
1133
1134                 read_drconf_cell(&drmem, &dm);
1135
1136                 /* skip this block if it is reserved or not assigned to
1137                  * this partition */
1138                 if ((drmem.flags & DRCONF_MEM_RESERVED)
1139                     || !(drmem.flags & DRCONF_MEM_ASSIGNED))
1140                         continue;
1141
1142                 if ((scn_addr < drmem.base_addr)
1143                     || (scn_addr >= (drmem.base_addr + lmb_size)))
1144                         continue;
1145
1146                 nid = of_drconf_to_nid_single(&drmem, &aa);
1147                 break;
1148         }
1149
1150         return nid;
1151 }
1152
1153 /*
1154  * Find the node associated with a hot added memory section for memory
1155  * represented in the device tree as a node (i.e. memory@XXXX) for
1156  * each memblock.
1157  */
1158 static int hot_add_node_scn_to_nid(unsigned long scn_addr)
1159 {
1160         struct device_node *memory;
1161         int nid = -1;
1162
1163         for_each_node_by_type(memory, "memory") {
1164                 unsigned long start, size;
1165                 int ranges;
1166                 const __be32 *memcell_buf;
1167                 unsigned int len;
1168
1169                 memcell_buf = of_get_property(memory, "reg", &len);
1170                 if (!memcell_buf || len <= 0)
1171                         continue;
1172
1173                 /* ranges in cell */
1174                 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
1175
1176                 while (ranges--) {
1177                         start = read_n_cells(n_mem_addr_cells, &memcell_buf);
1178                         size = read_n_cells(n_mem_size_cells, &memcell_buf);
1179
1180                         if ((scn_addr < start) || (scn_addr >= (start + size)))
1181                                 continue;
1182
1183                         nid = of_node_to_nid_single(memory);
1184                         break;
1185                 }
1186
1187                 if (nid >= 0)
1188                         break;
1189         }
1190
1191         of_node_put(memory);
1192
1193         return nid;
1194 }
1195
1196 /*
1197  * Find the node associated with a hot added memory section.  Section
1198  * corresponds to a SPARSEMEM section, not an MEMBLOCK.  It is assumed that
1199  * sections are fully contained within a single MEMBLOCK.
1200  */
1201 int hot_add_scn_to_nid(unsigned long scn_addr)
1202 {
1203         struct device_node *memory = NULL;
1204         int nid, found = 0;
1205
1206         if (!numa_enabled || (min_common_depth < 0))
1207                 return first_online_node;
1208
1209         memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1210         if (memory) {
1211                 nid = hot_add_drconf_scn_to_nid(memory, scn_addr);
1212                 of_node_put(memory);
1213         } else {
1214                 nid = hot_add_node_scn_to_nid(scn_addr);
1215         }
1216
1217         if (nid < 0 || !node_online(nid))
1218                 nid = first_online_node;
1219
1220         if (NODE_DATA(nid)->node_spanned_pages)
1221                 return nid;
1222
1223         for_each_online_node(nid) {
1224                 if (NODE_DATA(nid)->node_spanned_pages) {
1225                         found = 1;
1226                         break;
1227                 }
1228         }
1229
1230         BUG_ON(!found);
1231         return nid;
1232 }
1233
1234 static u64 hot_add_drconf_memory_max(void)
1235 {
1236         struct device_node *memory = NULL;
1237         unsigned int drconf_cell_cnt = 0;
1238         u64 lmb_size = 0;
1239         const __be32 *dm = NULL;
1240
1241         memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1242         if (memory) {
1243                 drconf_cell_cnt = of_get_drconf_memory(memory, &dm);
1244                 lmb_size = of_get_lmb_size(memory);
1245                 of_node_put(memory);
1246         }
1247         return lmb_size * drconf_cell_cnt;
1248 }
1249
1250 /*
1251  * memory_hotplug_max - return max address of memory that may be added
1252  *
1253  * This is currently only used on systems that support drconfig memory
1254  * hotplug.
1255  */
1256 u64 memory_hotplug_max(void)
1257 {
1258         return max(hot_add_drconf_memory_max(), memblock_end_of_DRAM());
1259 }
1260 #endif /* CONFIG_MEMORY_HOTPLUG */
1261
1262 /* Virtual Processor Home Node (VPHN) support */
1263 #ifdef CONFIG_PPC_SPLPAR
1264 struct topology_update_data {
1265         struct topology_update_data *next;
1266         unsigned int cpu;
1267         int old_nid;
1268         int new_nid;
1269 };
1270
1271 static u8 vphn_cpu_change_counts[NR_CPUS][MAX_DISTANCE_REF_POINTS];
1272 static cpumask_t cpu_associativity_changes_mask;
1273 static int vphn_enabled;
1274 static int prrn_enabled;
1275 static void reset_topology_timer(void);
1276
1277 /*
1278  * Store the current values of the associativity change counters in the
1279  * hypervisor.
1280  */
1281 static void setup_cpu_associativity_change_counters(void)
1282 {
1283         int cpu;
1284
1285         /* The VPHN feature supports a maximum of 8 reference points */
1286         BUILD_BUG_ON(MAX_DISTANCE_REF_POINTS > 8);
1287
1288         for_each_possible_cpu(cpu) {
1289                 int i;
1290                 u8 *counts = vphn_cpu_change_counts[cpu];
1291                 volatile u8 *hypervisor_counts = lppaca[cpu].vphn_assoc_counts;
1292
1293                 for (i = 0; i < distance_ref_points_depth; i++)
1294                         counts[i] = hypervisor_counts[i];
1295         }
1296 }
1297
1298 /*
1299  * The hypervisor maintains a set of 8 associativity change counters in
1300  * the VPA of each cpu that correspond to the associativity levels in the
1301  * ibm,associativity-reference-points property. When an associativity
1302  * level changes, the corresponding counter is incremented.
1303  *
1304  * Set a bit in cpu_associativity_changes_mask for each cpu whose home
1305  * node associativity levels have changed.
1306  *
1307  * Returns the number of cpus with unhandled associativity changes.
1308  */
1309 static int update_cpu_associativity_changes_mask(void)
1310 {
1311         int cpu;
1312         cpumask_t *changes = &cpu_associativity_changes_mask;
1313
1314         for_each_possible_cpu(cpu) {
1315                 int i, changed = 0;
1316                 u8 *counts = vphn_cpu_change_counts[cpu];
1317                 volatile u8 *hypervisor_counts = lppaca[cpu].vphn_assoc_counts;
1318
1319                 for (i = 0; i < distance_ref_points_depth; i++) {
1320                         if (hypervisor_counts[i] != counts[i]) {
1321                                 counts[i] = hypervisor_counts[i];
1322                                 changed = 1;
1323                         }
1324                 }
1325                 if (changed) {
1326                         cpumask_or(changes, changes, cpu_sibling_mask(cpu));
1327                         cpu = cpu_last_thread_sibling(cpu);
1328                 }
1329         }
1330
1331         return cpumask_weight(changes);
1332 }
1333
1334 /*
1335  * 6 64-bit registers unpacked into 12 32-bit associativity values. To form
1336  * the complete property we have to add the length in the first cell.
1337  */
1338 #define VPHN_ASSOC_BUFSIZE (6*sizeof(u64)/sizeof(u32) + 1)
1339
1340 /*
1341  * Convert the associativity domain numbers returned from the hypervisor
1342  * to the sequence they would appear in the ibm,associativity property.
1343  */
1344 static int vphn_unpack_associativity(const long *packed, __be32 *unpacked)
1345 {
1346         int i, nr_assoc_doms = 0;
1347         const __be16 *field = (const __be16 *) packed;
1348
1349 #define VPHN_FIELD_UNUSED       (0xffff)
1350 #define VPHN_FIELD_MSB          (0x8000)
1351 #define VPHN_FIELD_MASK         (~VPHN_FIELD_MSB)
1352
1353         for (i = 1; i < VPHN_ASSOC_BUFSIZE; i++) {
1354                 if (be16_to_cpup(field) == VPHN_FIELD_UNUSED) {
1355                         /* All significant fields processed, and remaining
1356                          * fields contain the reserved value of all 1's.
1357                          * Just store them.
1358                          */
1359                         unpacked[i] = *((__be32 *)field);
1360                         field += 2;
1361                 } else if (be16_to_cpup(field) & VPHN_FIELD_MSB) {
1362                         /* Data is in the lower 15 bits of this field */
1363                         unpacked[i] = cpu_to_be32(
1364                                 be16_to_cpup(field) & VPHN_FIELD_MASK);
1365                         field++;
1366                         nr_assoc_doms++;
1367                 } else {
1368                         /* Data is in the lower 15 bits of this field
1369                          * concatenated with the next 16 bit field
1370                          */
1371                         unpacked[i] = *((__be32 *)field);
1372                         field += 2;
1373                         nr_assoc_doms++;
1374                 }
1375         }
1376
1377         /* The first cell contains the length of the property */
1378         unpacked[0] = cpu_to_be32(nr_assoc_doms);
1379
1380         return nr_assoc_doms;
1381 }
1382
1383 /*
1384  * Retrieve the new associativity information for a virtual processor's
1385  * home node.
1386  */
1387 static long hcall_vphn(unsigned long cpu, __be32 *associativity)
1388 {
1389         long rc;
1390         long retbuf[PLPAR_HCALL9_BUFSIZE] = {0};
1391         u64 flags = 1;
1392         int hwcpu = get_hard_smp_processor_id(cpu);
1393
1394         rc = plpar_hcall9(H_HOME_NODE_ASSOCIATIVITY, retbuf, flags, hwcpu);
1395         vphn_unpack_associativity(retbuf, associativity);
1396
1397         return rc;
1398 }
1399
1400 static long vphn_get_associativity(unsigned long cpu,
1401                                         __be32 *associativity)
1402 {
1403         long rc;
1404
1405         rc = hcall_vphn(cpu, associativity);
1406
1407         switch (rc) {
1408         case H_FUNCTION:
1409                 printk(KERN_INFO
1410                         "VPHN is not supported. Disabling polling...\n");
1411                 stop_topology_update();
1412                 break;
1413         case H_HARDWARE:
1414                 printk(KERN_ERR
1415                         "hcall_vphn() experienced a hardware fault "
1416                         "preventing VPHN. Disabling polling...\n");
1417                 stop_topology_update();
1418         }
1419
1420         return rc;
1421 }
1422
1423 /*
1424  * Update the CPU maps and sysfs entries for a single CPU when its NUMA
1425  * characteristics change. This function doesn't perform any locking and is
1426  * only safe to call from stop_machine().
1427  */
1428 static int update_cpu_topology(void *data)
1429 {
1430         struct topology_update_data *update;
1431         unsigned long cpu;
1432
1433         if (!data)
1434                 return -EINVAL;
1435
1436         cpu = smp_processor_id();
1437
1438         for (update = data; update; update = update->next) {
1439                 if (cpu != update->cpu)
1440                         continue;
1441
1442                 unmap_cpu_from_node(update->cpu);
1443                 map_cpu_to_node(update->cpu, update->new_nid);
1444                 vdso_getcpu_init();
1445         }
1446
1447         return 0;
1448 }
1449
1450 /*
1451  * Update the node maps and sysfs entries for each cpu whose home node
1452  * has changed. Returns 1 when the topology has changed, and 0 otherwise.
1453  */
1454 int arch_update_cpu_topology(void)
1455 {
1456         unsigned int cpu, sibling, changed = 0;
1457         struct topology_update_data *updates, *ud;
1458         __be32 associativity[VPHN_ASSOC_BUFSIZE] = {0};
1459         cpumask_t updated_cpus;
1460         struct device *dev;
1461         int weight, new_nid, i = 0;
1462
1463         weight = cpumask_weight(&cpu_associativity_changes_mask);
1464         if (!weight)
1465                 return 0;
1466
1467         updates = kzalloc(weight * (sizeof(*updates)), GFP_KERNEL);
1468         if (!updates)
1469                 return 0;
1470
1471         cpumask_clear(&updated_cpus);
1472
1473         for_each_cpu(cpu, &cpu_associativity_changes_mask) {
1474                 /*
1475                  * If siblings aren't flagged for changes, updates list
1476                  * will be too short. Skip on this update and set for next
1477                  * update.
1478                  */
1479                 if (!cpumask_subset(cpu_sibling_mask(cpu),
1480                                         &cpu_associativity_changes_mask)) {
1481                         pr_info("Sibling bits not set for associativity "
1482                                         "change, cpu%d\n", cpu);
1483                         cpumask_or(&cpu_associativity_changes_mask,
1484                                         &cpu_associativity_changes_mask,
1485                                         cpu_sibling_mask(cpu));
1486                         cpu = cpu_last_thread_sibling(cpu);
1487                         continue;
1488                 }
1489
1490                 /* Use associativity from first thread for all siblings */
1491                 vphn_get_associativity(cpu, associativity);
1492                 new_nid = associativity_to_nid(associativity);
1493                 if (new_nid < 0 || !node_online(new_nid))
1494                         new_nid = first_online_node;
1495
1496                 if (new_nid == numa_cpu_lookup_table[cpu]) {
1497                         cpumask_andnot(&cpu_associativity_changes_mask,
1498                                         &cpu_associativity_changes_mask,
1499                                         cpu_sibling_mask(cpu));
1500                         cpu = cpu_last_thread_sibling(cpu);
1501                         continue;
1502                 }
1503
1504                 for_each_cpu(sibling, cpu_sibling_mask(cpu)) {
1505                         ud = &updates[i++];
1506                         ud->cpu = sibling;
1507                         ud->new_nid = new_nid;
1508                         ud->old_nid = numa_cpu_lookup_table[sibling];
1509                         cpumask_set_cpu(sibling, &updated_cpus);
1510                         if (i < weight)
1511                                 ud->next = &updates[i];
1512                 }
1513                 cpu = cpu_last_thread_sibling(cpu);
1514         }
1515
1516         stop_machine(update_cpu_topology, &updates[0], &updated_cpus);
1517
1518         for (ud = &updates[0]; ud; ud = ud->next) {
1519                 unregister_cpu_under_node(ud->cpu, ud->old_nid);
1520                 register_cpu_under_node(ud->cpu, ud->new_nid);
1521
1522                 dev = get_cpu_device(ud->cpu);
1523                 if (dev)
1524                         kobject_uevent(&dev->kobj, KOBJ_CHANGE);
1525                 cpumask_clear_cpu(ud->cpu, &cpu_associativity_changes_mask);
1526                 changed = 1;
1527         }
1528
1529         kfree(updates);
1530         return changed;
1531 }
1532
1533 static void topology_work_fn(struct work_struct *work)
1534 {
1535         rebuild_sched_domains();
1536 }
1537 static DECLARE_WORK(topology_work, topology_work_fn);
1538
1539 static void topology_schedule_update(void)
1540 {
1541         schedule_work(&topology_work);
1542 }
1543
1544 static void topology_timer_fn(unsigned long ignored)
1545 {
1546         if (prrn_enabled && cpumask_weight(&cpu_associativity_changes_mask))
1547                 topology_schedule_update();
1548         else if (vphn_enabled) {
1549                 if (update_cpu_associativity_changes_mask() > 0)
1550                         topology_schedule_update();
1551                 reset_topology_timer();
1552         }
1553 }
1554 static struct timer_list topology_timer =
1555         TIMER_INITIALIZER(topology_timer_fn, 0, 0);
1556
1557 static void reset_topology_timer(void)
1558 {
1559         topology_timer.data = 0;
1560         topology_timer.expires = jiffies + 60 * HZ;
1561         mod_timer(&topology_timer, topology_timer.expires);
1562 }
1563
1564 #ifdef CONFIG_SMP
1565
1566 static void stage_topology_update(int core_id)
1567 {
1568         cpumask_or(&cpu_associativity_changes_mask,
1569                 &cpu_associativity_changes_mask, cpu_sibling_mask(core_id));
1570         reset_topology_timer();
1571 }
1572
1573 static int dt_update_callback(struct notifier_block *nb,
1574                                 unsigned long action, void *data)
1575 {
1576         struct of_prop_reconfig *update;
1577         int rc = NOTIFY_DONE;
1578
1579         switch (action) {
1580         case OF_RECONFIG_UPDATE_PROPERTY:
1581                 update = (struct of_prop_reconfig *)data;
1582                 if (!of_prop_cmp(update->dn->type, "cpu") &&
1583                     !of_prop_cmp(update->prop->name, "ibm,associativity")) {
1584                         u32 core_id;
1585                         of_property_read_u32(update->dn, "reg", &core_id);
1586                         stage_topology_update(core_id);
1587                         rc = NOTIFY_OK;
1588                 }
1589                 break;
1590         }
1591
1592         return rc;
1593 }
1594
1595 static struct notifier_block dt_update_nb = {
1596         .notifier_call = dt_update_callback,
1597 };
1598
1599 #endif
1600
1601 /*
1602  * Start polling for associativity changes.
1603  */
1604 int start_topology_update(void)
1605 {
1606         int rc = 0;
1607
1608         if (firmware_has_feature(FW_FEATURE_PRRN)) {
1609                 if (!prrn_enabled) {
1610                         prrn_enabled = 1;
1611                         vphn_enabled = 0;
1612 #ifdef CONFIG_SMP
1613                         rc = of_reconfig_notifier_register(&dt_update_nb);
1614 #endif
1615                 }
1616         } else if (firmware_has_feature(FW_FEATURE_VPHN) &&
1617                    lppaca_shared_proc(get_lppaca())) {
1618                 if (!vphn_enabled) {
1619                         prrn_enabled = 0;
1620                         vphn_enabled = 1;
1621                         setup_cpu_associativity_change_counters();
1622                         init_timer_deferrable(&topology_timer);
1623                         reset_topology_timer();
1624                 }
1625         }
1626
1627         return rc;
1628 }
1629
1630 /*
1631  * Disable polling for VPHN associativity changes.
1632  */
1633 int stop_topology_update(void)
1634 {
1635         int rc = 0;
1636
1637         if (prrn_enabled) {
1638                 prrn_enabled = 0;
1639 #ifdef CONFIG_SMP
1640                 rc = of_reconfig_notifier_unregister(&dt_update_nb);
1641 #endif
1642         } else if (vphn_enabled) {
1643                 vphn_enabled = 0;
1644                 rc = del_timer_sync(&topology_timer);
1645         }
1646
1647         return rc;
1648 }
1649
1650 int prrn_is_enabled(void)
1651 {
1652         return prrn_enabled;
1653 }
1654
1655 static int topology_read(struct seq_file *file, void *v)
1656 {
1657         if (vphn_enabled || prrn_enabled)
1658                 seq_puts(file, "on\n");
1659         else
1660                 seq_puts(file, "off\n");
1661
1662         return 0;
1663 }
1664
1665 static int topology_open(struct inode *inode, struct file *file)
1666 {
1667         return single_open(file, topology_read, NULL);
1668 }
1669
1670 static ssize_t topology_write(struct file *file, const char __user *buf,
1671                               size_t count, loff_t *off)
1672 {
1673         char kbuf[4]; /* "on" or "off" plus null. */
1674         int read_len;
1675
1676         read_len = count < 3 ? count : 3;
1677         if (copy_from_user(kbuf, buf, read_len))
1678                 return -EINVAL;
1679
1680         kbuf[read_len] = '\0';
1681
1682         if (!strncmp(kbuf, "on", 2))
1683                 start_topology_update();
1684         else if (!strncmp(kbuf, "off", 3))
1685                 stop_topology_update();
1686         else
1687                 return -EINVAL;
1688
1689         return count;
1690 }
1691
1692 static const struct file_operations topology_ops = {
1693         .read = seq_read,
1694         .write = topology_write,
1695         .open = topology_open,
1696         .release = single_release
1697 };
1698
1699 static int topology_update_init(void)
1700 {
1701         start_topology_update();
1702         proc_create("powerpc/topology_updates", 644, NULL, &topology_ops);
1703
1704         return 0;
1705 }
1706 device_initcall(topology_update_init);
1707 #endif /* CONFIG_PPC_SPLPAR */