arch/powerpc/mm/numa.c

   1 /*
   2  * pSeries NUMA support
   3  *
   4  * Copyright (C) 2002 Anton Blanchard <[email protected]>, 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/bootmem.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/memblock.h>
  23 #include <linux/of.h>
  24 #include <linux/pfn.h>
  25 #include <linux/cpuset.h>
  26 #include <linux/node.h>
  27 #include <linux/stop_machine.h>
  28 #include <linux/proc_fs.h>
  29 #include <linux/seq_file.h>
  30 #include <linux/uaccess.h>
  31 #include <linux/slab.h>
  32 #include <asm/cputhreads.h>
  33 #include <asm/sparsemem.h>
  34 #include <asm/prom.h>
  35 #include <asm/smp.h>
  36 #include <asm/cputhreads.h>
  37 #include <asm/topology.h>
  38 #include <asm/firmware.h>
  39 #include <asm/paca.h>
  40 #include <asm/hvcall.h>
  41 #include <asm/setup.h>
  42 #include <asm/vdso.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 (node = 0; node < nr_node_ids; 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 %d 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 update_numa_cpu_lookup_table(unsigned int cpu, int node)
 146 {
 147         numa_cpu_lookup_table[cpu] = node;
 148 }
 149
 150 static void map_cpu_to_node(int cpu, int node)
 151 {
 152         update_numa_cpu_lookup_table(cpu, node);
 153
 154         dbg("adding cpu %d to node %d\n", cpu, node);
 155
 156         if (!(cpumask_test_cpu(cpu, node_to_cpumask_map[node])))
 157                 cpumask_set_cpu(cpu, node_to_cpumask_map[node]);
 158 }
 159
 160 #if defined(CONFIG_HOTPLUG_CPU) || defined(CONFIG_PPC_SPLPAR)
 161 static void unmap_cpu_from_node(unsigned long cpu)
 162 {
 163         int node = numa_cpu_lookup_table[cpu];
 164
 165         dbg("removing cpu %lu from node %d\n", cpu, node);
 166
 167         if (cpumask_test_cpu(cpu, node_to_cpumask_map[node])) {
 168                 cpumask_clear_cpu(cpu, node_to_cpumask_map[node]);
 169         } else {
 170                 printk(KERN_ERR "WARNING: cpu %lu not found in node %d\n",
 171                        cpu, node);
 172         }
 173 }
 174 #endif /* CONFIG_HOTPLUG_CPU || CONFIG_PPC_SPLPAR */
 175
 176 /* must hold reference to node during call */
 177 static const __be32 *of_get_associativity(struct device_node *dev)
 178 {
 179         return of_get_property(dev, "ibm,associativity", NULL);
 180 }
 181
 182 /*
 183  * Returns the property linux,drconf-usable-memory if
 184  * it exists (the property exists only in kexec/kdump kernels,
 185  * added by kexec-tools)
 186  */
 187 static const __be32 *of_get_usable_memory(struct device_node *memory)
 188 {
 189         const __be32 *prop;
 190         u32 len;
 191         prop = of_get_property(memory, "linux,drconf-usable-memory", &len);
 192         if (!prop || len < sizeof(unsigned int))
 193                 return NULL;
 194         return prop;
 195 }
 196
 197 int __node_distance(int a, int b)
 198 {
 199         int i;
 200         int distance = LOCAL_DISTANCE;
 201
 202         if (!form1_affinity)
 203                 return ((a == b) ? LOCAL_DISTANCE : REMOTE_DISTANCE);
 204
 205         for (i = 0; i < distance_ref_points_depth; i++) {
 206                 if (distance_lookup_table[a][i] == distance_lookup_table[b][i])
 207                         break;
 208
 209                 /* Double the distance for each NUMA level */
 210                 distance *= 2;
 211         }
 212
 213         return distance;
 214 }
 215 EXPORT_SYMBOL(__node_distance);
 216
 217 static void initialize_distance_lookup_table(int nid,
 218                 const __be32 *associativity)
 219 {
 220         int i;
 221
 222         if (!form1_affinity)
 223                 return;
 224
 225         for (i = 0; i < distance_ref_points_depth; i++) {
 226                 const __be32 *entry;
 227
 228                 entry = &associativity[be32_to_cpu(distance_ref_points[i])];
 229                 distance_lookup_table[nid][i] = of_read_number(entry, 1);
 230         }
 231 }
 232
 233 /* Returns nid in the range [0..MAX_NUMNODES-1], or -1 if no useful numa
 234  * info is found.
 235  */
 236 static int associativity_to_nid(const __be32 *associativity)
 237 {
 238         int nid = -1;
 239
 240         if (min_common_depth == -1)
 241                 goto out;
 242
 243         if (of_read_number(associativity, 1) >= min_common_depth)
 244                 nid = of_read_number(&associativity[min_common_depth], 1);
 245
 246         /* POWER4 LPAR uses 0xffff as invalid node */
 247         if (nid == 0xffff || nid >= MAX_NUMNODES)
 248                 nid = -1;
 249
 250         if (nid > 0 &&
 251             of_read_number(associativity, 1) >= distance_ref_points_depth)
 252                 initialize_distance_lookup_table(nid, associativity);
 253
 254 out:
 255         return nid;
 256 }
 257
 258 /* Returns the nid associated with the given device tree node,
 259  * or -1 if not found.
 260  */
 261 static int of_node_to_nid_single(struct device_node *device)
 262 {
 263         int nid = -1;
 264         const __be32 *tmp;
 265
 266         tmp = of_get_associativity(device);
 267         if (tmp)
 268                 nid = associativity_to_nid(tmp);
 269         return nid;
 270 }
 271
 272 /* Walk the device tree upwards, looking for an associativity id */
 273 int of_node_to_nid(struct device_node *device)
 274 {
 275         struct device_node *tmp;
 276         int nid = -1;
 277
 278         of_node_get(device);
 279         while (device) {
 280                 nid = of_node_to_nid_single(device);
 281                 if (nid != -1)
 282                         break;
 283
 284                 tmp = device;
 285                 device = of_get_parent(tmp);
 286                 of_node_put(tmp);
 287         }
 288         of_node_put(device);
 289
 290         return nid;
 291 }
 292 EXPORT_SYMBOL_GPL(of_node_to_nid);
 293
 294 static int __init find_min_common_depth(void)
 295 {
 296         int depth;
 297         struct device_node *root;
 298
 299         if (firmware_has_feature(FW_FEATURE_OPAL))
 300                 root = of_find_node_by_path("/ibm,opal");
 301         else
 302                 root = of_find_node_by_path("/rtas");
 303         if (!root)
 304                 root = of_find_node_by_path("/");
 305
 306         /*
 307          * This property is a set of 32-bit integers, each representing
 308          * an index into the ibm,associativity nodes.
 309          *
 310          * With form 0 affinity the first integer is for an SMP configuration
 311          * (should be all 0's) and the second is for a normal NUMA
 312          * configuration. We have only one level of NUMA.
 313          *
 314          * With form 1 affinity the first integer is the most significant
 315          * NUMA boundary and the following are progressively less significant
 316          * boundaries. There can be more than one level of NUMA.
 317          */
 318         distance_ref_points = of_get_property(root,
 319                                         "ibm,associativity-reference-points",
 320                                         &distance_ref_points_depth);
 321
 322         if (!distance_ref_points) {
 323                 dbg("NUMA: ibm,associativity-reference-points not found.\n");
 324                 goto err;
 325         }
 326
 327         distance_ref_points_depth /= sizeof(int);
 328
 329         if (firmware_has_feature(FW_FEATURE_OPAL) ||
 330             firmware_has_feature(FW_FEATURE_TYPE1_AFFINITY)) {
 331                 dbg("Using form 1 affinity\n");
 332                 form1_affinity = 1;
 333         }
 334
 335         if (form1_affinity) {
 336                 depth = of_read_number(distance_ref_points, 1);
 337         } else {
 338                 if (distance_ref_points_depth < 2) {
 339                         printk(KERN_WARNING "NUMA: "
 340                                 "short ibm,associativity-reference-points\n");
 341                         goto err;
 342                 }
 343
 344                 depth = of_read_number(&distance_ref_points[1], 1);
 345         }
 346
 347         /*
 348          * Warn and cap if the hardware supports more than
 349          * MAX_DISTANCE_REF_POINTS domains.
 350          */
 351         if (distance_ref_points_depth > MAX_DISTANCE_REF_POINTS) {
 352                 printk(KERN_WARNING "NUMA: distance array capped at "
 353                         "%d entries\n", MAX_DISTANCE_REF_POINTS);
 354                 distance_ref_points_depth = MAX_DISTANCE_REF_POINTS;
 355         }
 356
 357         of_node_put(root);
 358         return depth;
 359
 360 err:
 361         of_node_put(root);
 362         return -1;
 363 }
 364
 365 static void __init get_n_mem_cells(int *n_addr_cells, int *n_size_cells)
 366 {
 367         struct device_node *memory = NULL;
 368
 369         memory = of_find_node_by_type(memory, "memory");
 370         if (!memory)
 371                 panic("numa.c: No memory nodes found!");
 372
 373         *n_addr_cells = of_n_addr_cells(memory);
 374         *n_size_cells = of_n_size_cells(memory);
 375         of_node_put(memory);
 376 }
 377
 378 static unsigned long read_n_cells(int n, const __be32 **buf)
 379 {
 380         unsigned long result = 0;
 381
 382         while (n--) {
 383                 result = (result << 32) | of_read_number(*buf, 1);
 384                 (*buf)++;
 385         }
 386         return result;
 387 }
 388
 389 /*
 390  * Read the next memblock list entry from the ibm,dynamic-memory property
 391  * and return the information in the provided of_drconf_cell structure.
 392  */
 393 static void read_drconf_cell(struct of_drconf_cell *drmem, const __be32 **cellp)
 394 {
 395         const __be32 *cp;
 396
 397         drmem->base_addr = read_n_cells(n_mem_addr_cells, cellp);
 398
 399         cp = *cellp;
 400         drmem->drc_index = of_read_number(cp, 1);
 401         drmem->reserved = of_read_number(&cp[1], 1);
 402         drmem->aa_index = of_read_number(&cp[2], 1);
 403         drmem->flags = of_read_number(&cp[3], 1);
 404
 405         *cellp = cp + 4;
 406 }
 407
 408 /*
 409  * Retrieve and validate the ibm,dynamic-memory property of the device tree.
 410  *
 411  * The layout of the ibm,dynamic-memory property is a number N of memblock
 412  * list entries followed by N memblock list entries.  Each memblock list entry
 413  * contains information as laid out in the of_drconf_cell struct above.
 414  */
 415 static int of_get_drconf_memory(struct device_node *memory, const __be32 **dm)
 416 {
 417         const __be32 *prop;
 418         u32 len, entries;
 419
 420         prop = of_get_property(memory, "ibm,dynamic-memory", &len);
 421         if (!prop || len < sizeof(unsigned int))
 422                 return 0;
 423
 424         entries = of_read_number(prop++, 1);
 425
 426         /* Now that we know the number of entries, revalidate the size
 427          * of the property read in to ensure we have everything
 428          */
 429         if (len < (entries * (n_mem_addr_cells + 4) + 1) * sizeof(unsigned int))
 430                 return 0;
 431
 432         *dm = prop;
 433         return entries;
 434 }
 435
 436 /*
 437  * Retrieve and validate the ibm,lmb-size property for drconf memory
 438  * from the device tree.
 439  */
 440 static u64 of_get_lmb_size(struct device_node *memory)
 441 {
 442         const __be32 *prop;
 443         u32 len;
 444
 445         prop = of_get_property(memory, "ibm,lmb-size", &len);
 446         if (!prop || len < sizeof(unsigned int))
 447                 return 0;
 448
 449         return read_n_cells(n_mem_size_cells, &prop);
 450 }
 451
 452 struct assoc_arrays {
 453         u32     n_arrays;
 454         u32     array_sz;
 455         const __be32 *arrays;
 456 };
 457
 458 /*
 459  * Retrieve and validate the list of associativity arrays for drconf
 460  * memory from the ibm,associativity-lookup-arrays property of the
 461  * device tree..
 462  *
 463  * The layout of the ibm,associativity-lookup-arrays property is a number N
 464  * indicating the number of associativity arrays, followed by a number M
 465  * indicating the size of each associativity array, followed by a list
 466  * of N associativity arrays.
 467  */
 468 static int of_get_assoc_arrays(struct device_node *memory,
 469                                struct assoc_arrays *aa)
 470 {
 471         const __be32 *prop;
 472         u32 len;
 473
 474         prop = of_get_property(memory, "ibm,associativity-lookup-arrays", &len);
 475         if (!prop || len < 2 * sizeof(unsigned int))
 476                 return -1;
 477
 478         aa->n_arrays = of_read_number(prop++, 1);
 479         aa->array_sz = of_read_number(prop++, 1);
 480
 481         /* Now that we know the number of arrays and size of each array,
 482          * revalidate the size of the property read in.
 483          */
 484         if (len < (aa->n_arrays * aa->array_sz + 2) * sizeof(unsigned int))
 485                 return -1;
 486
 487         aa->arrays = prop;
 488         return 0;
 489 }
 490
 491 /*
 492  * This is like of_node_to_nid_single() for memory represented in the
 493  * ibm,dynamic-reconfiguration-memory node.
 494  */
 495 static int of_drconf_to_nid_single(struct of_drconf_cell *drmem,
 496                                    struct assoc_arrays *aa)
 497 {
 498         int default_nid = 0;
 499         int nid = default_nid;
 500         int index;
 501
 502         if (min_common_depth > 0 && min_common_depth <= aa->array_sz &&
 503             !(drmem->flags & DRCONF_MEM_AI_INVALID) &&
 504             drmem->aa_index < aa->n_arrays) {
 505                 index = drmem->aa_index * aa->array_sz + min_common_depth - 1;
 506                 nid = of_read_number(&aa->arrays[index], 1);
 507
 508                 if (nid == 0xffff || nid >= MAX_NUMNODES)
 509                         nid = default_nid;
 510         }
 511
 512         return nid;
 513 }
 514
 515 /*
 516  * Figure out to which domain a cpu belongs and stick it there.
 517  * Return the id of the domain used.
 518  */
 519 static int numa_setup_cpu(unsigned long lcpu)
 520 {
 521         int nid = -1;
 522         struct device_node *cpu;
 523
 524         /*
 525          * If a valid cpu-to-node mapping is already available, use it
 526          * directly instead of querying the firmware, since it represents
 527          * the most recent mapping notified to us by the platform (eg: VPHN).
 528          */
 529         if ((nid = numa_cpu_lookup_table[lcpu]) >= 0) {
 530                 map_cpu_to_node(lcpu, nid);
 531                 return nid;
 532         }
 533
 534         cpu = of_get_cpu_node(lcpu, NULL);
 535
 536         if (!cpu) {
 537                 WARN_ON(1);
 538                 if (cpu_present(lcpu))
 539                         goto out_present;
 540                 else
 541                         goto out;
 542         }
 543
 544         nid = of_node_to_nid_single(cpu);
 545
 546 out_present:
 547         if (nid < 0 || !node_online(nid))
 548                 nid = first_online_node;
 549
 550         map_cpu_to_node(lcpu, nid);
 551         of_node_put(cpu);
 552 out:
 553         return nid;
 554 }
 555
 556 static void verify_cpu_node_mapping(int cpu, int node)
 557 {
 558         int base, sibling, i;
 559
 560         /* Verify that all the threads in the core belong to the same node */
 561         base = cpu_first_thread_sibling(cpu);
 562
 563         for (i = 0; i < threads_per_core; i++) {
 564                 sibling = base + i;
 565
 566                 if (sibling == cpu || cpu_is_offline(sibling))
 567                         continue;
 568
 569                 if (cpu_to_node(sibling) != node) {
 570                         WARN(1, "CPU thread siblings %d and %d don't belong"
 571                                 " to the same node!\n", cpu, sibling);
 572                         break;
 573                 }
 574         }
 575 }
 576
 577 static int cpu_numa_callback(struct notifier_block *nfb, unsigned long action,
 578                              void *hcpu)
 579 {
 580         unsigned long lcpu = (unsigned long)hcpu;
 581         int ret = NOTIFY_DONE, nid;
 582
 583         switch (action) {
 584         case CPU_UP_PREPARE:
 585         case CPU_UP_PREPARE_FROZEN:
 586                 nid = numa_setup_cpu(lcpu);
 587                 verify_cpu_node_mapping((int)lcpu, nid);
 588                 ret = NOTIFY_OK;
 589                 break;
 590 #ifdef CONFIG_HOTPLUG_CPU
 591         case CPU_DEAD:
 592         case CPU_DEAD_FROZEN:
 593         case CPU_UP_CANCELED:
 594         case CPU_UP_CANCELED_FROZEN:
 595                 unmap_cpu_from_node(lcpu);
 596                 ret = NOTIFY_OK;
 597                 break;
 598 #endif
 599         }
 600         return ret;
 601 }
 602
 603 /*
 604  * Check and possibly modify a memory region to enforce the memory limit.
 605  *
 606  * Returns the size the region should have to enforce the memory limit.
 607  * This will either be the original value of size, a truncated value,
 608  * or zero. If the returned value of size is 0 the region should be
 609  * discarded as it lies wholly above the memory limit.
 610  */
 611 static unsigned long __init numa_enforce_memory_limit(unsigned long start,
 612                                                       unsigned long size)
 613 {
 614         /*
 615          * We use memblock_end_of_DRAM() in here instead of memory_limit because
 616          * we've already adjusted it for the limit and it takes care of
 617          * having memory holes below the limit.  Also, in the case of
 618          * iommu_is_off, memory_limit is not set but is implicitly enforced.
 619          */
 620
 621         if (start + size <= memblock_end_of_DRAM())
 622                 return size;
 623
 624         if (start >= memblock_end_of_DRAM())
 625                 return 0;
 626
 627         return memblock_end_of_DRAM() - start;
 628 }
 629
 630 /*
 631  * Reads the counter for a given entry in
 632  * linux,drconf-usable-memory property
 633  */
 634 static inline int __init read_usm_ranges(const __be32 **usm)
 635 {
 636         /*
 637          * For each lmb in ibm,dynamic-memory a corresponding
 638          * entry in linux,drconf-usable-memory property contains
 639          * a counter followed by that many (base, size) duple.
 640          * read the counter from linux,drconf-usable-memory
 641          */
 642         return read_n_cells(n_mem_size_cells, usm);
 643 }
 644
 645 /*
 646  * Extract NUMA information from the ibm,dynamic-reconfiguration-memory
 647  * node.  This assumes n_mem_{addr,size}_cells have been set.
 648  */
 649 static void __init parse_drconf_memory(struct device_node *memory)
 650 {
 651         const __be32 *uninitialized_var(dm), *usm;
 652         unsigned int n, rc, ranges, is_kexec_kdump = 0;
 653         unsigned long lmb_size, base, size, sz;
 654         int nid;
 655         struct assoc_arrays aa = { .arrays = NULL };
 656
 657         n = of_get_drconf_memory(memory, &dm);
 658         if (!n)
 659                 return;
 660
 661         lmb_size = of_get_lmb_size(memory);
 662         if (!lmb_size)
 663                 return;
 664
 665         rc = of_get_assoc_arrays(memory, &aa);
 666         if (rc)
 667                 return;
 668
 669         /* check if this is a kexec/kdump kernel */
 670         usm = of_get_usable_memory(memory);
 671         if (usm != NULL)
 672                 is_kexec_kdump = 1;
 673
 674         for (; n != 0; --n) {
 675                 struct of_drconf_cell drmem;
 676
 677                 read_drconf_cell(&drmem, &dm);
 678
 679                 /* skip this block if the reserved bit is set in flags (0x80)
 680                    or if the block is not assigned to this partition (0x8) */
 681                 if ((drmem.flags & DRCONF_MEM_RESERVED)
 682                     || !(drmem.flags & DRCONF_MEM_ASSIGNED))
 683                         continue;
 684
 685                 base = drmem.base_addr;
 686                 size = lmb_size;
 687                 ranges = 1;
 688
 689                 if (is_kexec_kdump) {
 690                         ranges = read_usm_ranges(&usm);
 691                         if (!ranges) /* there are no (base, size) duple */
 692                                 continue;
 693                 }
 694                 do {
 695                         if (is_kexec_kdump) {
 696                                 base = read_n_cells(n_mem_addr_cells, &usm);
 697                                 size = read_n_cells(n_mem_size_cells, &usm);
 698                         }
 699                         nid = of_drconf_to_nid_single(&drmem, &aa);
 700                         fake_numa_create_new_node(
 701                                 ((base + size) >> PAGE_SHIFT),
 702                                            &nid);
 703                         node_set_online(nid);
 704                         sz = numa_enforce_memory_limit(base, size);
 705                         if (sz)
 706                                 memblock_set_node(base, sz,
 707                                                   &memblock.memory, nid);
 708                 } while (--ranges);
 709         }
 710 }
 711
 712 static int __init parse_numa_properties(void)
 713 {
 714         struct device_node *memory;
 715         int default_nid = 0;
 716         unsigned long i;
 717
 718         if (numa_enabled == 0) {
 719                 printk(KERN_WARNING "NUMA disabled by user\n");
 720                 return -1;
 721         }
 722
 723         min_common_depth = find_min_common_depth();
 724
 725         if (min_common_depth < 0)
 726                 return min_common_depth;
 727
 728         dbg("NUMA associativity depth for CPU/Memory: %d\n", min_common_depth);
 729
 730         /*
 731          * Even though we connect cpus to numa domains later in SMP
 732          * init, we need to know the node ids now. This is because
 733          * each node to be onlined must have NODE_DATA etc backing it.
 734          */
 735         for_each_present_cpu(i) {
 736                 struct device_node *cpu;
 737                 int nid;
 738
 739                 cpu = of_get_cpu_node(i, NULL);
 740                 BUG_ON(!cpu);
 741                 nid = of_node_to_nid_single(cpu);
 742                 of_node_put(cpu);
 743
 744                 /*
 745                  * Don't fall back to default_nid yet -- we will plug
 746                  * cpus into nodes once the memory scan has discovered
 747                  * the topology.
 748                  */
 749                 if (nid < 0)
 750                         continue;
 751                 node_set_online(nid);
 752         }
 753
 754         get_n_mem_cells(&n_mem_addr_cells, &n_mem_size_cells);
 755
 756         for_each_node_by_type(memory, "memory") {
 757                 unsigned long start;
 758                 unsigned long size;
 759                 int nid;
 760                 int ranges;
 761                 const __be32 *memcell_buf;
 762                 unsigned int len;
 763
 764                 memcell_buf = of_get_property(memory,
 765                         "linux,usable-memory", &len);
 766                 if (!memcell_buf || len <= 0)
 767                         memcell_buf = of_get_property(memory, "reg", &len);
 768                 if (!memcell_buf || len <= 0)
 769                         continue;
 770
 771                 /* ranges in cell */
 772                 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
 773 new_range:
 774                 /* these are order-sensitive, and modify the buffer pointer */
 775                 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
 776                 size = read_n_cells(n_mem_size_cells, &memcell_buf);
 777
 778                 /*
 779                  * Assumption: either all memory nodes or none will
 780                  * have associativity properties.  If none, then
 781                  * everything goes to default_nid.
 782                  */
 783                 nid = of_node_to_nid_single(memory);
 784                 if (nid < 0)
 785                         nid = default_nid;
 786
 787                 fake_numa_create_new_node(((start + size) >> PAGE_SHIFT), &nid);
 788                 node_set_online(nid);
 789
 790                 if (!(size = numa_enforce_memory_limit(start, size))) {
 791                         if (--ranges)
 792                                 goto new_range;
 793                         else
 794                                 continue;
 795                 }
 796
 797                 memblock_set_node(start, size, &memblock.memory, nid);
 798
 799                 if (--ranges)
 800                         goto new_range;
 801         }
 802
 803         /*
 804          * Now do the same thing for each MEMBLOCK listed in the
 805          * ibm,dynamic-memory property in the
 806          * ibm,dynamic-reconfiguration-memory node.
 807          */
 808         memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
 809         if (memory)
 810                 parse_drconf_memory(memory);
 811
 812         return 0;
 813 }
 814
 815 static void __init setup_nonnuma(void)
 816 {
 817         unsigned long top_of_ram = memblock_end_of_DRAM();
 818         unsigned long total_ram = memblock_phys_mem_size();
 819         unsigned long start_pfn, end_pfn;
 820         unsigned int nid = 0;
 821         struct memblock_region *reg;
 822
 823         printk(KERN_DEBUG "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
 824                top_of_ram, total_ram);
 825         printk(KERN_DEBUG "Memory hole size: %ldMB\n",
 826                (top_of_ram - total_ram) >> 20);
 827
 828         for_each_memblock(memory, reg) {
 829                 start_pfn = memblock_region_memory_base_pfn(reg);
 830                 end_pfn = memblock_region_memory_end_pfn(reg);
 831
 832                 fake_numa_create_new_node(end_pfn, &nid);
 833                 memblock_set_node(PFN_PHYS(start_pfn),
 834                                   PFN_PHYS(end_pfn - start_pfn),
 835                                   &memblock.memory, nid);
 836                 node_set_online(nid);
 837         }
 838 }
 839
 840 void __init dump_numa_cpu_topology(void)
 841 {
 842         unsigned int node;
 843         unsigned int cpu, count;
 844
 845         if (min_common_depth == -1 || !numa_enabled)
 846                 return;
 847
 848         for_each_online_node(node) {
 849                 printk(KERN_DEBUG "Node %d CPUs:", node);
 850
 851                 count = 0;
 852                 /*
 853                  * If we used a CPU iterator here we would miss printing
 854                  * the holes in the cpumap.
 855                  */
 856                 for (cpu = 0; cpu < nr_cpu_ids; cpu++) {
 857                         if (cpumask_test_cpu(cpu,
 858                                         node_to_cpumask_map[node])) {
 859                                 if (count == 0)
 860                                         printk(" %u", cpu);
 861                                 ++count;
 862                         } else {
 863                                 if (count > 1)
 864                                         printk("-%u", cpu - 1);
 865                                 count = 0;
 866                         }
 867                 }
 868
 869                 if (count > 1)
 870                         printk("-%u", nr_cpu_ids - 1);
 871                 printk("\n");
 872         }
 873 }
 874
 875 static void __init dump_numa_memory_topology(void)
 876 {
 877         unsigned int node;
 878         unsigned int count;
 879
 880         if (min_common_depth == -1 || !numa_enabled)
 881                 return;
 882
 883         for_each_online_node(node) {
 884                 unsigned long i;
 885
 886                 printk(KERN_DEBUG "Node %d Memory:", node);
 887
 888                 count = 0;
 889
 890                 for (i = 0; i < memblock_end_of_DRAM();
 891                      i += (1 << SECTION_SIZE_BITS)) {
 892                         if (early_pfn_to_nid(i >> PAGE_SHIFT) == node) {
 893                                 if (count == 0)
 894                                         printk(" 0x%lx", i);
 895                                 ++count;
 896                         } else {
 897                                 if (count > 0)
 898                                         printk("-0x%lx", i);
 899                                 count = 0;
 900                         }
 901                 }
 902
 903                 if (count > 0)
 904                         printk("-0x%lx", i);
 905                 printk("\n");
 906         }
 907 }
 908
 909 static struct notifier_block ppc64_numa_nb = {
 910         .notifier_call = cpu_numa_callback,
 911         .priority = 1 /* Must run before sched domains notifier. */
 912 };
 913
 914 /* Initialize NODE_DATA for a node on the local memory */
 915 static void __init setup_node_data(int nid, u64 start_pfn, u64 end_pfn)
 916 {
 917         u64 spanned_pages = end_pfn - start_pfn;
 918         const size_t nd_size = roundup(sizeof(pg_data_t), SMP_CACHE_BYTES);
 919         u64 nd_pa;
 920         void *nd;
 921         int tnid;
 922
 923         if (spanned_pages)
 924                 pr_info("Initmem setup node %d [mem %#010Lx-%#010Lx]\n",
 925                         nid, start_pfn << PAGE_SHIFT,
 926                         (end_pfn << PAGE_SHIFT) - 1);
 927         else
 928                 pr_info("Initmem setup node %d\n", nid);
 929
 930         nd_pa = memblock_alloc_try_nid(nd_size, SMP_CACHE_BYTES, nid);
 931         nd = __va(nd_pa);
 932
 933         /* report and initialize */
 934         pr_info("  NODE_DATA [mem %#010Lx-%#010Lx]\n",
 935                 nd_pa, nd_pa + nd_size - 1);
 936         tnid = early_pfn_to_nid(nd_pa >> PAGE_SHIFT);
 937         if (tnid != nid)
 938                 pr_info("    NODE_DATA(%d) on node %d\n", nid, tnid);
 939
 940         node_data[nid] = nd;
 941         memset(NODE_DATA(nid), 0, sizeof(pg_data_t));
 942         NODE_DATA(nid)->node_id = nid;
 943         NODE_DATA(nid)->node_start_pfn = start_pfn;
 944         NODE_DATA(nid)->node_spanned_pages = spanned_pages;
 945 }
 946
 947 void __init initmem_init(void)
 948 {
 949         int nid, cpu;
 950
 951         max_low_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
 952         max_pfn = max_low_pfn;
 953
 954         if (parse_numa_properties())
 955                 setup_nonnuma();
 956         else
 957                 dump_numa_memory_topology();
 958
 959         memblock_dump_all();
 960
 961         for_each_online_node(nid) {
 962                 unsigned long start_pfn, end_pfn;
 963
 964                 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
 965                 setup_node_data(nid, start_pfn, end_pfn);
 966                 sparse_memory_present_with_active_regions(nid);
 967         }
 968
 969         sparse_init();
 970
 971         setup_node_to_cpumask_map();
 972
 973         reset_numa_cpu_lookup_table();
 974         register_cpu_notifier(&ppc64_numa_nb);
 975         /*
 976          * We need the numa_cpu_lookup_table to be accurate for all CPUs,
 977          * even before we online them, so that we can use cpu_to_{node,mem}
 978          * early in boot, cf. smp_prepare_cpus().
 979          */
 980         for_each_present_cpu(cpu) {
 981                 numa_setup_cpu((unsigned long)cpu);
 982         }
 983 }
 984
 985 static int __init early_numa(char *p)
 986 {
 987         if (!p)
 988                 return 0;
 989
 990         if (strstr(p, "off"))
 991                 numa_enabled = 0;
 992
 993         if (strstr(p, "debug"))
 994                 numa_debug = 1;
 995
 996         p = strstr(p, "fake=");
 997         if (p)
 998                 cmdline = p + strlen("fake=");
 999
1000         return 0;
1001 }
1002 early_param("numa", early_numa);
1003
1004 static bool topology_updates_enabled = true;
1005
1006 static int __init early_topology_updates(char *p)
1007 {
1008         if (!p)
1009                 return 0;
1010
1011         if (!strcmp(p, "off")) {
1012                 pr_info("Disabling topology updates\n");
1013                 topology_updates_enabled = false;
1014         }
1015
1016         return 0;
1017 }
1018 early_param("topology_updates", early_topology_updates);
1019
1020 #ifdef CONFIG_MEMORY_HOTPLUG
1021 /*
1022  * Find the node associated with a hot added memory section for
1023  * memory represented in the device tree by the property
1024  * ibm,dynamic-reconfiguration-memory/ibm,dynamic-memory.
1025  */
1026 static int hot_add_drconf_scn_to_nid(struct device_node *memory,
1027                                      unsigned long scn_addr)
1028 {
1029         const __be32 *dm;
1030         unsigned int drconf_cell_cnt, rc;
1031         unsigned long lmb_size;
1032         struct assoc_arrays aa;
1033         int nid = -1;
1034
1035         drconf_cell_cnt = of_get_drconf_memory(memory, &dm);
1036         if (!drconf_cell_cnt)
1037                 return -1;
1038
1039         lmb_size = of_get_lmb_size(memory);
1040         if (!lmb_size)
1041                 return -1;
1042
1043         rc = of_get_assoc_arrays(memory, &aa);
1044         if (rc)
1045                 return -1;
1046
1047         for (; drconf_cell_cnt != 0; --drconf_cell_cnt) {
1048                 struct of_drconf_cell drmem;
1049
1050                 read_drconf_cell(&drmem, &dm);
1051
1052                 /* skip this block if it is reserved or not assigned to
1053                  * this partition */
1054                 if ((drmem.flags & DRCONF_MEM_RESERVED)
1055                     || !(drmem.flags & DRCONF_MEM_ASSIGNED))
1056                         continue;
1057
1058                 if ((scn_addr < drmem.base_addr)
1059                     || (scn_addr >= (drmem.base_addr + lmb_size)))
1060                         continue;
1061
1062                 nid = of_drconf_to_nid_single(&drmem, &aa);
1063                 break;
1064         }
1065
1066         return nid;
1067 }
1068
1069 /*
1070  * Find the node associated with a hot added memory section for memory
1071  * represented in the device tree as a node (i.e. memory@XXXX) for
1072  * each memblock.
1073  */
1074 static int hot_add_node_scn_to_nid(unsigned long scn_addr)
1075 {
1076         struct device_node *memory;
1077         int nid = -1;
1078
1079         for_each_node_by_type(memory, "memory") {
1080                 unsigned long start, size;
1081                 int ranges;
1082                 const __be32 *memcell_buf;
1083                 unsigned int len;
1084
1085                 memcell_buf = of_get_property(memory, "reg", &len);
1086                 if (!memcell_buf || len <= 0)
1087                         continue;
1088
1089                 /* ranges in cell */
1090                 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
1091
1092                 while (ranges--) {
1093                         start = read_n_cells(n_mem_addr_cells, &memcell_buf);
1094                         size = read_n_cells(n_mem_size_cells, &memcell_buf);
1095
1096                         if ((scn_addr < start) || (scn_addr >= (start + size)))
1097                                 continue;
1098
1099                         nid = of_node_to_nid_single(memory);
1100                         break;
1101                 }
1102
1103                 if (nid >= 0)
1104                         break;
1105         }
1106
1107         of_node_put(memory);
1108
1109         return nid;
1110 }
1111
1112 /*
1113  * Find the node associated with a hot added memory section.  Section
1114  * corresponds to a SPARSEMEM section, not an MEMBLOCK.  It is assumed that
1115  * sections are fully contained within a single MEMBLOCK.
1116  */
1117 int hot_add_scn_to_nid(unsigned long scn_addr)
1118 {
1119         struct device_node *memory = NULL;
1120         int nid, found = 0;
1121
1122         if (!numa_enabled || (min_common_depth < 0))
1123                 return first_online_node;
1124
1125         memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1126         if (memory) {
1127                 nid = hot_add_drconf_scn_to_nid(memory, scn_addr);
1128                 of_node_put(memory);
1129         } else {
1130                 nid = hot_add_node_scn_to_nid(scn_addr);
1131         }
1132
1133         if (nid < 0 || !node_online(nid))
1134                 nid = first_online_node;
1135
1136         if (NODE_DATA(nid)->node_spanned_pages)
1137                 return nid;
1138
1139         for_each_online_node(nid) {
1140                 if (NODE_DATA(nid)->node_spanned_pages) {
1141                         found = 1;
1142                         break;
1143                 }
1144         }
1145
1146         BUG_ON(!found);
1147         return nid;
1148 }
1149
1150 static u64 hot_add_drconf_memory_max(void)
1151 {
1152         struct device_node *memory = NULL;
1153         unsigned int drconf_cell_cnt = 0;
1154         u64 lmb_size = 0;
1155         const __be32 *dm = NULL;
1156
1157         memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1158         if (memory) {
1159                 drconf_cell_cnt = of_get_drconf_memory(memory, &dm);
1160                 lmb_size = of_get_lmb_size(memory);
1161                 of_node_put(memory);
1162         }
1163         return lmb_size * drconf_cell_cnt;
1164 }
1165
1166 /*
1167  * memory_hotplug_max - return max address of memory that may be added
1168  *
1169  * This is currently only used on systems that support drconfig memory
1170  * hotplug.
1171  */
1172 u64 memory_hotplug_max(void)
1173 {
1174         return max(hot_add_drconf_memory_max(), memblock_end_of_DRAM());
1175 }
1176 #endif /* CONFIG_MEMORY_HOTPLUG */
1177
1178 /* Virtual Processor Home Node (VPHN) support */
1179 #ifdef CONFIG_PPC_SPLPAR
1180 struct topology_update_data {
1181         struct topology_update_data *next;
1182         unsigned int cpu;
1183         int old_nid;
1184         int new_nid;
1185 };
1186
1187 static u8 vphn_cpu_change_counts[NR_CPUS][MAX_DISTANCE_REF_POINTS];
1188 static cpumask_t cpu_associativity_changes_mask;
1189 static int vphn_enabled;
1190 static int prrn_enabled;
1191 static void reset_topology_timer(void);
1192
1193 /*
1194  * Store the current values of the associativity change counters in the
1195  * hypervisor.
1196  */
1197 static void setup_cpu_associativity_change_counters(void)
1198 {
1199         int cpu;
1200
1201         /* The VPHN feature supports a maximum of 8 reference points */
1202         BUILD_BUG_ON(MAX_DISTANCE_REF_POINTS > 8);
1203
1204         for_each_possible_cpu(cpu) {
1205                 int i;
1206                 u8 *counts = vphn_cpu_change_counts[cpu];
1207                 volatile u8 *hypervisor_counts = lppaca[cpu].vphn_assoc_counts;
1208
1209                 for (i = 0; i < distance_ref_points_depth; i++)
1210                         counts[i] = hypervisor_counts[i];
1211         }
1212 }
1213
1214 /*
1215  * The hypervisor maintains a set of 8 associativity change counters in
1216  * the VPA of each cpu that correspond to the associativity levels in the
1217  * ibm,associativity-reference-points property. When an associativity
1218  * level changes, the corresponding counter is incremented.
1219  *
1220  * Set a bit in cpu_associativity_changes_mask for each cpu whose home
1221  * node associativity levels have changed.
1222  *
1223  * Returns the number of cpus with unhandled associativity changes.
1224  */
1225 static int update_cpu_associativity_changes_mask(void)
1226 {
1227         int cpu;
1228         cpumask_t *changes = &cpu_associativity_changes_mask;
1229
1230         for_each_possible_cpu(cpu) {
1231                 int i, changed = 0;
1232                 u8 *counts = vphn_cpu_change_counts[cpu];
1233                 volatile u8 *hypervisor_counts = lppaca[cpu].vphn_assoc_counts;
1234
1235                 for (i = 0; i < distance_ref_points_depth; i++) {
1236                         if (hypervisor_counts[i] != counts[i]) {
1237                                 counts[i] = hypervisor_counts[i];
1238                                 changed = 1;
1239                         }
1240                 }
1241                 if (changed) {
1242                         cpumask_or(changes, changes, cpu_sibling_mask(cpu));
1243                         cpu = cpu_last_thread_sibling(cpu);
1244                 }
1245         }
1246
1247         return cpumask_weight(changes);
1248 }
1249
1250 /*
1251  * 6 64-bit registers unpacked into 12 32-bit associativity values. To form
1252  * the complete property we have to add the length in the first cell.
1253  */
1254 #define VPHN_ASSOC_BUFSIZE (6*sizeof(u64)/sizeof(u32) + 1)
1255
1256 /*
1257  * Convert the associativity domain numbers returned from the hypervisor
1258  * to the sequence they would appear in the ibm,associativity property.
1259  */
1260 static int vphn_unpack_associativity(const long *packed, __be32 *unpacked)
1261 {
1262         int i, nr_assoc_doms = 0;
1263         const __be16 *field = (const __be16 *) packed;
1264
1265 #define VPHN_FIELD_UNUSED       (0xffff)
1266 #define VPHN_FIELD_MSB          (0x8000)
1267 #define VPHN_FIELD_MASK         (~VPHN_FIELD_MSB)
1268
1269         for (i = 1; i < VPHN_ASSOC_BUFSIZE; i++) {
1270                 if (be16_to_cpup(field) == VPHN_FIELD_UNUSED) {
1271                         /* All significant fields processed, and remaining
1272                          * fields contain the reserved value of all 1's.
1273                          * Just store them.
1274                          */
1275                         unpacked[i] = *((__be32 *)field);
1276                         field += 2;
1277                 } else if (be16_to_cpup(field) & VPHN_FIELD_MSB) {
1278                         /* Data is in the lower 15 bits of this field */
1279                         unpacked[i] = cpu_to_be32(
1280                                 be16_to_cpup(field) & VPHN_FIELD_MASK);
1281                         field++;
1282                         nr_assoc_doms++;
1283                 } else {
1284                         /* Data is in the lower 15 bits of this field
1285                          * concatenated with the next 16 bit field
1286                          */
1287                         unpacked[i] = *((__be32 *)field);
1288                         field += 2;
1289                         nr_assoc_doms++;
1290                 }
1291         }
1292
1293         /* The first cell contains the length of the property */
1294         unpacked[0] = cpu_to_be32(nr_assoc_doms);
1295
1296         return nr_assoc_doms;
1297 }
1298
1299 /*
1300  * Retrieve the new associativity information for a virtual processor's
1301  * home node.
1302  */
1303 static long hcall_vphn(unsigned long cpu, __be32 *associativity)
1304 {
1305         long rc;
1306         long retbuf[PLPAR_HCALL9_BUFSIZE] = {0};
1307         u64 flags = 1;
1308         int hwcpu = get_hard_smp_processor_id(cpu);
1309         int i;
1310
1311         rc = plpar_hcall9(H_HOME_NODE_ASSOCIATIVITY, retbuf, flags, hwcpu);
1312         for (i = 0; i < 6; i++)
1313                 retbuf[i] = cpu_to_be64(retbuf[i]);
1314         vphn_unpack_associativity(retbuf, associativity);
1315
1316         return rc;
1317 }
1318
1319 static long vphn_get_associativity(unsigned long cpu,
1320                                         __be32 *associativity)
1321 {
1322         long rc;
1323
1324         rc = hcall_vphn(cpu, associativity);
1325
1326         switch (rc) {
1327         case H_FUNCTION:
1328                 printk(KERN_INFO
1329                         "VPHN is not supported. Disabling polling...\n");
1330                 stop_topology_update();
1331                 break;
1332         case H_HARDWARE:
1333                 printk(KERN_ERR
1334                         "hcall_vphn() experienced a hardware fault "
1335                         "preventing VPHN. Disabling polling...\n");
1336                 stop_topology_update();
1337         }
1338
1339         return rc;
1340 }
1341
1342 /*
1343  * Update the CPU maps and sysfs entries for a single CPU when its NUMA
1344  * characteristics change. This function doesn't perform any locking and is
1345  * only safe to call from stop_machine().
1346  */
1347 static int update_cpu_topology(void *data)
1348 {
1349         struct topology_update_data *update;
1350         unsigned long cpu;
1351
1352         if (!data)
1353                 return -EINVAL;
1354
1355         cpu = smp_processor_id();
1356
1357         for (update = data; update; update = update->next) {
1358                 int new_nid = update->new_nid;
1359                 if (cpu != update->cpu)
1360                         continue;
1361
1362                 unmap_cpu_from_node(cpu);
1363                 map_cpu_to_node(cpu, new_nid);
1364                 set_cpu_numa_node(cpu, new_nid);
1365                 set_cpu_numa_mem(cpu, local_memory_node(new_nid));
1366                 vdso_getcpu_init();
1367         }
1368
1369         return 0;
1370 }
1371
1372 static int update_lookup_table(void *data)
1373 {
1374         struct topology_update_data *update;
1375
1376         if (!data)
1377                 return -EINVAL;
1378
1379         /*
1380          * Upon topology update, the numa-cpu lookup table needs to be updated
1381          * for all threads in the core, including offline CPUs, to ensure that
1382          * future hotplug operations respect the cpu-to-node associativity
1383          * properly.
1384          */
1385         for (update = data; update; update = update->next) {
1386                 int nid, base, j;
1387
1388                 nid = update->new_nid;
1389                 base = cpu_first_thread_sibling(update->cpu);
1390
1391                 for (j = 0; j < threads_per_core; j++) {
1392                         update_numa_cpu_lookup_table(base + j, nid);
1393                 }
1394         }
1395
1396         return 0;
1397 }
1398
1399 /*
1400  * Update the node maps and sysfs entries for each cpu whose home node
1401  * has changed. Returns 1 when the topology has changed, and 0 otherwise.
1402  */
1403 int arch_update_cpu_topology(void)
1404 {
1405         unsigned int cpu, sibling, changed = 0;
1406         struct topology_update_data *updates, *ud;
1407         __be32 associativity[VPHN_ASSOC_BUFSIZE] = {0};
1408         cpumask_t updated_cpus;
1409         struct device *dev;
1410         int weight, new_nid, i = 0;
1411
1412         if (!prrn_enabled && !vphn_enabled)
1413                 return 0;
1414
1415         weight = cpumask_weight(&cpu_associativity_changes_mask);
1416         if (!weight)
1417                 return 0;
1418
1419         updates = kzalloc(weight * (sizeof(*updates)), GFP_KERNEL);
1420         if (!updates)
1421                 return 0;
1422
1423         cpumask_clear(&updated_cpus);
1424
1425         for_each_cpu(cpu, &cpu_associativity_changes_mask) {
1426                 /*
1427                  * If siblings aren't flagged for changes, updates list
1428                  * will be too short. Skip on this update and set for next
1429                  * update.
1430                  */
1431                 if (!cpumask_subset(cpu_sibling_mask(cpu),
1432                                         &cpu_associativity_changes_mask)) {
1433                         pr_info("Sibling bits not set for associativity "
1434                                         "change, cpu%d\n", cpu);
1435                         cpumask_or(&cpu_associativity_changes_mask,
1436                                         &cpu_associativity_changes_mask,
1437                                         cpu_sibling_mask(cpu));
1438                         cpu = cpu_last_thread_sibling(cpu);
1439                         continue;
1440                 }
1441
1442                 /* Use associativity from first thread for all siblings */
1443                 vphn_get_associativity(cpu, associativity);
1444                 new_nid = associativity_to_nid(associativity);
1445                 if (new_nid < 0 || !node_online(new_nid))
1446                         new_nid = first_online_node;
1447
1448                 if (new_nid == numa_cpu_lookup_table[cpu]) {
1449                         cpumask_andnot(&cpu_associativity_changes_mask,
1450                                         &cpu_associativity_changes_mask,
1451                                         cpu_sibling_mask(cpu));
1452                         cpu = cpu_last_thread_sibling(cpu);
1453                         continue;
1454                 }
1455
1456                 for_each_cpu(sibling, cpu_sibling_mask(cpu)) {
1457                         ud = &updates[i++];
1458                         ud->cpu = sibling;
1459                         ud->new_nid = new_nid;
1460                         ud->old_nid = numa_cpu_lookup_table[sibling];
1461                         cpumask_set_cpu(sibling, &updated_cpus);
1462                         if (i < weight)
1463                                 ud->next = &updates[i];
1464                 }
1465                 cpu = cpu_last_thread_sibling(cpu);
1466         }
1467
1468         pr_debug("Topology update for the following CPUs:\n");
1469         if (cpumask_weight(&updated_cpus)) {
1470                 for (ud = &updates[0]; ud; ud = ud->next) {
1471                         pr_debug("cpu %d moving from node %d "
1472                                           "to %d\n", ud->cpu,
1473                                           ud->old_nid, ud->new_nid);
1474                 }
1475         }
1476
1477         /*
1478          * In cases where we have nothing to update (because the updates list
1479          * is too short or because the new topology is same as the old one),
1480          * skip invoking update_cpu_topology() via stop-machine(). This is
1481          * necessary (and not just a fast-path optimization) since stop-machine
1482          * can end up electing a random CPU to run update_cpu_topology(), and
1483          * thus trick us into setting up incorrect cpu-node mappings (since
1484          * 'updates' is kzalloc()'ed).
1485          *
1486          * And for the similar reason, we will skip all the following updating.
1487          */
1488         if (!cpumask_weight(&updated_cpus))
1489                 goto out;
1490
1491         stop_machine(update_cpu_topology, &updates[0], &updated_cpus);
1492
1493         /*
1494          * Update the numa-cpu lookup table with the new mappings, even for
1495          * offline CPUs. It is best to perform this update from the stop-
1496          * machine context.
1497          */
1498         stop_machine(update_lookup_table, &updates[0],
1499                                         cpumask_of(raw_smp_processor_id()));
1500
1501         for (ud = &updates[0]; ud; ud = ud->next) {
1502                 unregister_cpu_under_node(ud->cpu, ud->old_nid);
1503                 register_cpu_under_node(ud->cpu, ud->new_nid);
1504
1505                 dev = get_cpu_device(ud->cpu);
1506                 if (dev)
1507                         kobject_uevent(&dev->kobj, KOBJ_CHANGE);
1508                 cpumask_clear_cpu(ud->cpu, &cpu_associativity_changes_mask);
1509                 changed = 1;
1510         }
1511
1512 out:
1513         kfree(updates);
1514         return changed;
1515 }
1516
1517 static void topology_work_fn(struct work_struct *work)
1518 {
1519         rebuild_sched_domains();
1520 }
1521 static DECLARE_WORK(topology_work, topology_work_fn);
1522
1523 static void topology_schedule_update(void)
1524 {
1525         schedule_work(&topology_work);
1526 }
1527
1528 static void topology_timer_fn(unsigned long ignored)
1529 {
1530         if (prrn_enabled && cpumask_weight(&cpu_associativity_changes_mask))
1531                 topology_schedule_update();
1532         else if (vphn_enabled) {
1533                 if (update_cpu_associativity_changes_mask() > 0)
1534                         topology_schedule_update();
1535                 reset_topology_timer();
1536         }
1537 }
1538 static struct timer_list topology_timer =
1539         TIMER_INITIALIZER(topology_timer_fn, 0, 0);
1540
1541 static void reset_topology_timer(void)
1542 {
1543         topology_timer.data = 0;
1544         topology_timer.expires = jiffies + 60 * HZ;
1545         mod_timer(&topology_timer, topology_timer.expires);
1546 }
1547
1548 #ifdef CONFIG_SMP
1549
1550 static void stage_topology_update(int core_id)
1551 {
1552         cpumask_or(&cpu_associativity_changes_mask,
1553                 &cpu_associativity_changes_mask, cpu_sibling_mask(core_id));
1554         reset_topology_timer();
1555 }
1556
1557 static int dt_update_callback(struct notifier_block *nb,
1558                                 unsigned long action, void *data)
1559 {
1560         struct of_reconfig_data *update = data;
1561         int rc = NOTIFY_DONE;
1562
1563         switch (action) {
1564         case OF_RECONFIG_UPDATE_PROPERTY:
1565                 if (!of_prop_cmp(update->dn->type, "cpu") &&
1566                     !of_prop_cmp(update->prop->name, "ibm,associativity")) {
1567                         u32 core_id;
1568                         of_property_read_u32(update->dn, "reg", &core_id);
1569                         stage_topology_update(core_id);
1570                         rc = NOTIFY_OK;
1571                 }
1572                 break;
1573         }
1574
1575         return rc;
1576 }
1577
1578 static struct notifier_block dt_update_nb = {
1579         .notifier_call = dt_update_callback,
1580 };
1581
1582 #endif
1583
1584 /*
1585  * Start polling for associativity changes.
1586  */
1587 int start_topology_update(void)
1588 {
1589         int rc = 0;
1590
1591         if (firmware_has_feature(FW_FEATURE_PRRN)) {
1592                 if (!prrn_enabled) {
1593                         prrn_enabled = 1;
1594                         vphn_enabled = 0;
1595 #ifdef CONFIG_SMP
1596                         rc = of_reconfig_notifier_register(&dt_update_nb);
1597 #endif
1598                 }
1599         } else if (firmware_has_feature(FW_FEATURE_VPHN) &&
1600                    lppaca_shared_proc(get_lppaca())) {
1601                 if (!vphn_enabled) {
1602                         prrn_enabled = 0;
1603                         vphn_enabled = 1;
1604                         setup_cpu_associativity_change_counters();
1605                         init_timer_deferrable(&topology_timer);
1606                         reset_topology_timer();
1607                 }
1608         }
1609
1610         return rc;
1611 }
1612
1613 /*
1614  * Disable polling for VPHN associativity changes.
1615  */
1616 int stop_topology_update(void)
1617 {
1618         int rc = 0;
1619
1620         if (prrn_enabled) {
1621                 prrn_enabled = 0;
1622 #ifdef CONFIG_SMP
1623                 rc = of_reconfig_notifier_unregister(&dt_update_nb);
1624 #endif
1625         } else if (vphn_enabled) {
1626                 vphn_enabled = 0;
1627                 rc = del_timer_sync(&topology_timer);
1628         }
1629
1630         return rc;
1631 }
1632
1633 int prrn_is_enabled(void)
1634 {
1635         return prrn_enabled;
1636 }
1637
1638 static int topology_read(struct seq_file *file, void *v)
1639 {
1640         if (vphn_enabled || prrn_enabled)
1641                 seq_puts(file, "on\n");
1642         else
1643                 seq_puts(file, "off\n");
1644
1645         return 0;
1646 }
1647
1648 static int topology_open(struct inode *inode, struct file *file)
1649 {
1650         return single_open(file, topology_read, NULL);
1651 }
1652
1653 static ssize_t topology_write(struct file *file, const char __user *buf,
1654                               size_t count, loff_t *off)
1655 {
1656         char kbuf[4]; /* "on" or "off" plus null. */
1657         int read_len;
1658
1659         read_len = count < 3 ? count : 3;
1660         if (copy_from_user(kbuf, buf, read_len))
1661                 return -EINVAL;
1662
1663         kbuf[read_len] = '\0';
1664
1665         if (!strncmp(kbuf, "on", 2))
1666                 start_topology_update();
1667         else if (!strncmp(kbuf, "off", 3))
1668                 stop_topology_update();
1669         else
1670                 return -EINVAL;
1671
1672         return count;
1673 }
1674
1675 static const struct file_operations topology_ops = {
1676         .read = seq_read,
1677         .write = topology_write,
1678         .open = topology_open,
1679         .release = single_release
1680 };
1681
1682 static int topology_update_init(void)
1683 {
1684         /* Do not poll for changes if disabled at boot */
1685         if (topology_updates_enabled)
1686                 start_topology_update();
1687
1688         if (!proc_create("powerpc/topology_updates", 0644, NULL, &topology_ops))
1689                 return -ENOMEM;
1690
1691         return 0;
1692 }
1693 device_initcall(topology_update_init);
1694 #endif /* CONFIG_PPC_SPLPAR */