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 #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_each_node(node)
  84                 alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]);
  85
  86         /* cpumask_of_node() will now work */
  87         dbg("Node to cpumask map for %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]) - 1];
 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                 /*
 253                  * Skip the length field and send start of associativity array
 254                  */
 255                 initialize_distance_lookup_table(nid, associativity + 1);
 256         }
 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         int nid = -1;
 280
 281         of_node_get(device);
 282         while (device) {
 283                 nid = of_node_to_nid_single(device);
 284                 if (nid != -1)
 285                         break;
 286
 287                 device = of_get_next_parent(device);
 288         }
 289         of_node_put(device);
 290
 291         return nid;
 292 }
 293 EXPORT_SYMBOL_GPL(of_node_to_nid);
 294
 295 static int __init find_min_common_depth(void)
 296 {
 297         int depth;
 298         struct device_node *root;
 299
 300         if (firmware_has_feature(FW_FEATURE_OPAL))
 301                 root = of_find_node_by_path("/ibm,opal");
 302         else
 303                 root = of_find_node_by_path("/rtas");
 304         if (!root)
 305                 root = of_find_node_by_path("/");
 306
 307         /*
 308          * This property is a set of 32-bit integers, each representing
 309          * an index into the ibm,associativity nodes.
 310          *
 311          * With form 0 affinity the first integer is for an SMP configuration
 312          * (should be all 0's) and the second is for a normal NUMA
 313          * configuration. We have only one level of NUMA.
 314          *
 315          * With form 1 affinity the first integer is the most significant
 316          * NUMA boundary and the following are progressively less significant
 317          * boundaries. There can be more than one level of NUMA.
 318          */
 319         distance_ref_points = of_get_property(root,
 320                                         "ibm,associativity-reference-points",
 321                                         &distance_ref_points_depth);
 322
 323         if (!distance_ref_points) {
 324                 dbg("NUMA: ibm,associativity-reference-points not found.\n");
 325                 goto err;
 326         }
 327
 328         distance_ref_points_depth /= sizeof(int);
 329
 330         if (firmware_has_feature(FW_FEATURE_OPAL) ||
 331             firmware_has_feature(FW_FEATURE_TYPE1_AFFINITY)) {
 332                 dbg("Using form 1 affinity\n");
 333                 form1_affinity = 1;
 334         }
 335
 336         if (form1_affinity) {
 337                 depth = of_read_number(distance_ref_points, 1);
 338         } else {
 339                 if (distance_ref_points_depth < 2) {
 340                         printk(KERN_WARNING "NUMA: "
 341                                 "short ibm,associativity-reference-points\n");
 342                         goto err;
 343                 }
 344
 345                 depth = of_read_number(&distance_ref_points[1], 1);
 346         }
 347
 348         /*
 349          * Warn and cap if the hardware supports more than
 350          * MAX_DISTANCE_REF_POINTS domains.
 351          */
 352         if (distance_ref_points_depth > MAX_DISTANCE_REF_POINTS) {
 353                 printk(KERN_WARNING "NUMA: distance array capped at "
 354                         "%d entries\n", MAX_DISTANCE_REF_POINTS);
 355                 distance_ref_points_depth = MAX_DISTANCE_REF_POINTS;
 356         }
 357
 358         of_node_put(root);
 359         return depth;
 360
 361 err:
 362         of_node_put(root);
 363         return -1;
 364 }
 365
 366 static void __init get_n_mem_cells(int *n_addr_cells, int *n_size_cells)
 367 {
 368         struct device_node *memory = NULL;
 369
 370         memory = of_find_node_by_type(memory, "memory");
 371         if (!memory)
 372                 panic("numa.c: No memory nodes found!");
 373
 374         *n_addr_cells = of_n_addr_cells(memory);
 375         *n_size_cells = of_n_size_cells(memory);
 376         of_node_put(memory);
 377 }
 378
 379 static unsigned long read_n_cells(int n, const __be32 **buf)
 380 {
 381         unsigned long result = 0;
 382
 383         while (n--) {
 384                 result = (result << 32) | of_read_number(*buf, 1);
 385                 (*buf)++;
 386         }
 387         return result;
 388 }
 389
 390 /*
 391  * Read the next memblock list entry from the ibm,dynamic-memory property
 392  * and return the information in the provided of_drconf_cell structure.
 393  */
 394 static void read_drconf_cell(struct of_drconf_cell *drmem, const __be32 **cellp)
 395 {
 396         const __be32 *cp;
 397
 398         drmem->base_addr = read_n_cells(n_mem_addr_cells, cellp);
 399
 400         cp = *cellp;
 401         drmem->drc_index = of_read_number(cp, 1);
 402         drmem->reserved = of_read_number(&cp[1], 1);
 403         drmem->aa_index = of_read_number(&cp[2], 1);
 404         drmem->flags = of_read_number(&cp[3], 1);
 405
 406         *cellp = cp + 4;
 407 }
 408
 409 /*
 410  * Retrieve and validate the ibm,dynamic-memory property of the device tree.
 411  *
 412  * The layout of the ibm,dynamic-memory property is a number N of memblock
 413  * list entries followed by N memblock list entries.  Each memblock list entry
 414  * contains information as laid out in the of_drconf_cell struct above.
 415  */
 416 static int of_get_drconf_memory(struct device_node *memory, const __be32 **dm)
 417 {
 418         const __be32 *prop;
 419         u32 len, entries;
 420
 421         prop = of_get_property(memory, "ibm,dynamic-memory", &len);
 422         if (!prop || len < sizeof(unsigned int))
 423                 return 0;
 424
 425         entries = of_read_number(prop++, 1);
 426
 427         /* Now that we know the number of entries, revalidate the size
 428          * of the property read in to ensure we have everything
 429          */
 430         if (len < (entries * (n_mem_addr_cells + 4) + 1) * sizeof(unsigned int))
 431                 return 0;
 432
 433         *dm = prop;
 434         return entries;
 435 }
 436
 437 /*
 438  * Retrieve and validate the ibm,lmb-size property for drconf memory
 439  * from the device tree.
 440  */
 441 static u64 of_get_lmb_size(struct device_node *memory)
 442 {
 443         const __be32 *prop;
 444         u32 len;
 445
 446         prop = of_get_property(memory, "ibm,lmb-size", &len);
 447         if (!prop || len < sizeof(unsigned int))
 448                 return 0;
 449
 450         return read_n_cells(n_mem_size_cells, &prop);
 451 }
 452
 453 struct assoc_arrays {
 454         u32     n_arrays;
 455         u32     array_sz;
 456         const __be32 *arrays;
 457 };
 458
 459 /*
 460  * Retrieve and validate the list of associativity arrays for drconf
 461  * memory from the ibm,associativity-lookup-arrays property of the
 462  * device tree..
 463  *
 464  * The layout of the ibm,associativity-lookup-arrays property is a number N
 465  * indicating the number of associativity arrays, followed by a number M
 466  * indicating the size of each associativity array, followed by a list
 467  * of N associativity arrays.
 468  */
 469 static int of_get_assoc_arrays(struct device_node *memory,
 470                                struct assoc_arrays *aa)
 471 {
 472         const __be32 *prop;
 473         u32 len;
 474
 475         prop = of_get_property(memory, "ibm,associativity-lookup-arrays", &len);
 476         if (!prop || len < 2 * sizeof(unsigned int))
 477                 return -1;
 478
 479         aa->n_arrays = of_read_number(prop++, 1);
 480         aa->array_sz = of_read_number(prop++, 1);
 481
 482         /* Now that we know the number of arrays and size of each array,
 483          * revalidate the size of the property read in.
 484          */
 485         if (len < (aa->n_arrays * aa->array_sz + 2) * sizeof(unsigned int))
 486                 return -1;
 487
 488         aa->arrays = prop;
 489         return 0;
 490 }
 491
 492 /*
 493  * This is like of_node_to_nid_single() for memory represented in the
 494  * ibm,dynamic-reconfiguration-memory node.
 495  */
 496 static int of_drconf_to_nid_single(struct of_drconf_cell *drmem,
 497                                    struct assoc_arrays *aa)
 498 {
 499         int default_nid = 0;
 500         int nid = default_nid;
 501         int index;
 502
 503         if (min_common_depth > 0 && min_common_depth <= aa->array_sz &&
 504             !(drmem->flags & DRCONF_MEM_AI_INVALID) &&
 505             drmem->aa_index < aa->n_arrays) {
 506                 index = drmem->aa_index * aa->array_sz + min_common_depth - 1;
 507                 nid = of_read_number(&aa->arrays[index], 1);
 508
 509                 if (nid == 0xffff || nid >= MAX_NUMNODES)
 510                         nid = default_nid;
 511
 512                 if (nid > 0) {
 513                         index = drmem->aa_index * aa->array_sz;
 514                         initialize_distance_lookup_table(nid,
 515                                                         &aa->arrays[index]);
 516                 }
 517         }
 518
 519         return nid;
 520 }
 521
 522 /*
 523  * Figure out to which domain a cpu belongs and stick it there.
 524  * Return the id of the domain used.
 525  */
 526 static int numa_setup_cpu(unsigned long lcpu)
 527 {
 528         int nid = -1;
 529         struct device_node *cpu;
 530
 531         /*
 532          * If a valid cpu-to-node mapping is already available, use it
 533          * directly instead of querying the firmware, since it represents
 534          * the most recent mapping notified to us by the platform (eg: VPHN).
 535          */
 536         if ((nid = numa_cpu_lookup_table[lcpu]) >= 0) {
 537                 map_cpu_to_node(lcpu, nid);
 538                 return nid;
 539         }
 540
 541         cpu = of_get_cpu_node(lcpu, NULL);
 542
 543         if (!cpu) {
 544                 WARN_ON(1);
 545                 if (cpu_present(lcpu))
 546                         goto out_present;
 547                 else
 548                         goto out;
 549         }
 550
 551         nid = of_node_to_nid_single(cpu);
 552
 553 out_present:
 554         if (nid < 0 || !node_possible(nid))
 555                 nid = first_online_node;
 556
 557         map_cpu_to_node(lcpu, nid);
 558         of_node_put(cpu);
 559 out:
 560         return nid;
 561 }
 562
 563 static void verify_cpu_node_mapping(int cpu, int node)
 564 {
 565         int base, sibling, i;
 566
 567         /* Verify that all the threads in the core belong to the same node */
 568         base = cpu_first_thread_sibling(cpu);
 569
 570         for (i = 0; i < threads_per_core; i++) {
 571                 sibling = base + i;
 572
 573                 if (sibling == cpu || cpu_is_offline(sibling))
 574                         continue;
 575
 576                 if (cpu_to_node(sibling) != node) {
 577                         WARN(1, "CPU thread siblings %d and %d don't belong"
 578                                 " to the same node!\n", cpu, sibling);
 579                         break;
 580                 }
 581         }
 582 }
 583
 584 static int cpu_numa_callback(struct notifier_block *nfb, unsigned long action,
 585                              void *hcpu)
 586 {
 587         unsigned long lcpu = (unsigned long)hcpu;
 588         int ret = NOTIFY_DONE, nid;
 589
 590         switch (action) {
 591         case CPU_UP_PREPARE:
 592         case CPU_UP_PREPARE_FROZEN:
 593                 nid = numa_setup_cpu(lcpu);
 594                 verify_cpu_node_mapping((int)lcpu, nid);
 595                 ret = NOTIFY_OK;
 596                 break;
 597 #ifdef CONFIG_HOTPLUG_CPU
 598         case CPU_DEAD:
 599         case CPU_DEAD_FROZEN:
 600         case CPU_UP_CANCELED:
 601         case CPU_UP_CANCELED_FROZEN:
 602                 unmap_cpu_from_node(lcpu);
 603                 ret = NOTIFY_OK;
 604                 break;
 605 #endif
 606         }
 607         return ret;
 608 }
 609
 610 /*
 611  * Check and possibly modify a memory region to enforce the memory limit.
 612  *
 613  * Returns the size the region should have to enforce the memory limit.
 614  * This will either be the original value of size, a truncated value,
 615  * or zero. If the returned value of size is 0 the region should be
 616  * discarded as it lies wholly above the memory limit.
 617  */
 618 static unsigned long __init numa_enforce_memory_limit(unsigned long start,
 619                                                       unsigned long size)
 620 {
 621         /*
 622          * We use memblock_end_of_DRAM() in here instead of memory_limit because
 623          * we've already adjusted it for the limit and it takes care of
 624          * having memory holes below the limit.  Also, in the case of
 625          * iommu_is_off, memory_limit is not set but is implicitly enforced.
 626          */
 627
 628         if (start + size <= memblock_end_of_DRAM())
 629                 return size;
 630
 631         if (start >= memblock_end_of_DRAM())
 632                 return 0;
 633
 634         return memblock_end_of_DRAM() - start;
 635 }
 636
 637 /*
 638  * Reads the counter for a given entry in
 639  * linux,drconf-usable-memory property
 640  */
 641 static inline int __init read_usm_ranges(const __be32 **usm)
 642 {
 643         /*
 644          * For each lmb in ibm,dynamic-memory a corresponding
 645          * entry in linux,drconf-usable-memory property contains
 646          * a counter followed by that many (base, size) duple.
 647          * read the counter from linux,drconf-usable-memory
 648          */
 649         return read_n_cells(n_mem_size_cells, usm);
 650 }
 651
 652 /*
 653  * Extract NUMA information from the ibm,dynamic-reconfiguration-memory
 654  * node.  This assumes n_mem_{addr,size}_cells have been set.
 655  */
 656 static void __init parse_drconf_memory(struct device_node *memory)
 657 {
 658         const __be32 *uninitialized_var(dm), *usm;
 659         unsigned int n, rc, ranges, is_kexec_kdump = 0;
 660         unsigned long lmb_size, base, size, sz;
 661         int nid;
 662         struct assoc_arrays aa = { .arrays = NULL };
 663
 664         n = of_get_drconf_memory(memory, &dm);
 665         if (!n)
 666                 return;
 667
 668         lmb_size = of_get_lmb_size(memory);
 669         if (!lmb_size)
 670                 return;
 671
 672         rc = of_get_assoc_arrays(memory, &aa);
 673         if (rc)
 674                 return;
 675
 676         /* check if this is a kexec/kdump kernel */
 677         usm = of_get_usable_memory(memory);
 678         if (usm != NULL)
 679                 is_kexec_kdump = 1;
 680
 681         for (; n != 0; --n) {
 682                 struct of_drconf_cell drmem;
 683
 684                 read_drconf_cell(&drmem, &dm);
 685
 686                 /* skip this block if the reserved bit is set in flags (0x80)
 687                    or if the block is not assigned to this partition (0x8) */
 688                 if ((drmem.flags & DRCONF_MEM_RESERVED)
 689                     || !(drmem.flags & DRCONF_MEM_ASSIGNED))
 690                         continue;
 691
 692                 base = drmem.base_addr;
 693                 size = lmb_size;
 694                 ranges = 1;
 695
 696                 if (is_kexec_kdump) {
 697                         ranges = read_usm_ranges(&usm);
 698                         if (!ranges) /* there are no (base, size) duple */
 699                                 continue;
 700                 }
 701                 do {
 702                         if (is_kexec_kdump) {
 703                                 base = read_n_cells(n_mem_addr_cells, &usm);
 704                                 size = read_n_cells(n_mem_size_cells, &usm);
 705                         }
 706                         nid = of_drconf_to_nid_single(&drmem, &aa);
 707                         fake_numa_create_new_node(
 708                                 ((base + size) >> PAGE_SHIFT),
 709                                            &nid);
 710                         node_set_online(nid);
 711                         sz = numa_enforce_memory_limit(base, size);
 712                         if (sz)
 713                                 memblock_set_node(base, sz,
 714                                                   &memblock.memory, nid);
 715                 } while (--ranges);
 716         }
 717 }
 718
 719 static int __init parse_numa_properties(void)
 720 {
 721         struct device_node *memory;
 722         int default_nid = 0;
 723         unsigned long i;
 724
 725         if (numa_enabled == 0) {
 726                 printk(KERN_WARNING "NUMA disabled by user\n");
 727                 return -1;
 728         }
 729
 730         min_common_depth = find_min_common_depth();
 731
 732         if (min_common_depth < 0)
 733                 return min_common_depth;
 734
 735         dbg("NUMA associativity depth for CPU/Memory: %d\n", min_common_depth);
 736
 737         /*
 738          * Even though we connect cpus to numa domains later in SMP
 739          * init, we need to know the node ids now. This is because
 740          * each node to be onlined must have NODE_DATA etc backing it.
 741          */
 742         for_each_present_cpu(i) {
 743                 struct device_node *cpu;
 744                 int nid;
 745
 746                 cpu = of_get_cpu_node(i, NULL);
 747                 BUG_ON(!cpu);
 748                 nid = of_node_to_nid_single(cpu);
 749                 of_node_put(cpu);
 750
 751                 /*
 752                  * Don't fall back to default_nid yet -- we will plug
 753                  * cpus into nodes once the memory scan has discovered
 754                  * the topology.
 755                  */
 756                 if (nid < 0)
 757                         continue;
 758                 node_set_online(nid);
 759         }
 760
 761         get_n_mem_cells(&n_mem_addr_cells, &n_mem_size_cells);
 762
 763         for_each_node_by_type(memory, "memory") {
 764                 unsigned long start;
 765                 unsigned long size;
 766                 int nid;
 767                 int ranges;
 768                 const __be32 *memcell_buf;
 769                 unsigned int len;
 770
 771                 memcell_buf = of_get_property(memory,
 772                         "linux,usable-memory", &len);
 773                 if (!memcell_buf || len <= 0)
 774                         memcell_buf = of_get_property(memory, "reg", &len);
 775                 if (!memcell_buf || len <= 0)
 776                         continue;
 777
 778                 /* ranges in cell */
 779                 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
 780 new_range:
 781                 /* these are order-sensitive, and modify the buffer pointer */
 782                 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
 783                 size = read_n_cells(n_mem_size_cells, &memcell_buf);
 784
 785                 /*
 786                  * Assumption: either all memory nodes or none will
 787                  * have associativity properties.  If none, then
 788                  * everything goes to default_nid.
 789                  */
 790                 nid = of_node_to_nid_single(memory);
 791                 if (nid < 0)
 792                         nid = default_nid;
 793
 794                 fake_numa_create_new_node(((start + size) >> PAGE_SHIFT), &nid);
 795                 node_set_online(nid);
 796
 797                 if (!(size = numa_enforce_memory_limit(start, size))) {
 798                         if (--ranges)
 799                                 goto new_range;
 800                         else
 801                                 continue;
 802                 }
 803
 804                 memblock_set_node(start, size, &memblock.memory, nid);
 805
 806                 if (--ranges)
 807                         goto new_range;
 808         }
 809
 810         /*
 811          * Now do the same thing for each MEMBLOCK listed in the
 812          * ibm,dynamic-memory property in the
 813          * ibm,dynamic-reconfiguration-memory node.
 814          */
 815         memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
 816         if (memory)
 817                 parse_drconf_memory(memory);
 818
 819         return 0;
 820 }
 821
 822 static void __init setup_nonnuma(void)
 823 {
 824         unsigned long top_of_ram = memblock_end_of_DRAM();
 825         unsigned long total_ram = memblock_phys_mem_size();
 826         unsigned long start_pfn, end_pfn;
 827         unsigned int nid = 0;
 828         struct memblock_region *reg;
 829
 830         printk(KERN_DEBUG "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
 831                top_of_ram, total_ram);
 832         printk(KERN_DEBUG "Memory hole size: %ldMB\n",
 833                (top_of_ram - total_ram) >> 20);
 834
 835         for_each_memblock(memory, reg) {
 836                 start_pfn = memblock_region_memory_base_pfn(reg);
 837                 end_pfn = memblock_region_memory_end_pfn(reg);
 838
 839                 fake_numa_create_new_node(end_pfn, &nid);
 840                 memblock_set_node(PFN_PHYS(start_pfn),
 841                                   PFN_PHYS(end_pfn - start_pfn),
 842                                   &memblock.memory, nid);
 843                 node_set_online(nid);
 844         }
 845 }
 846
 847 void __init dump_numa_cpu_topology(void)
 848 {
 849         unsigned int node;
 850         unsigned int cpu, count;
 851
 852         if (min_common_depth == -1 || !numa_enabled)
 853                 return;
 854
 855         for_each_online_node(node) {
 856                 printk(KERN_DEBUG "Node %d CPUs:", node);
 857
 858                 count = 0;
 859                 /*
 860                  * If we used a CPU iterator here we would miss printing
 861                  * the holes in the cpumap.
 862                  */
 863                 for (cpu = 0; cpu < nr_cpu_ids; cpu++) {
 864                         if (cpumask_test_cpu(cpu,
 865                                         node_to_cpumask_map[node])) {
 866                                 if (count == 0)
 867                                         printk(" %u", cpu);
 868                                 ++count;
 869                         } else {
 870                                 if (count > 1)
 871                                         printk("-%u", cpu - 1);
 872                                 count = 0;
 873                         }
 874                 }
 875
 876                 if (count > 1)
 877                         printk("-%u", nr_cpu_ids - 1);
 878                 printk("\n");
 879         }
 880 }
 881
 882 static void __init dump_numa_memory_topology(void)
 883 {
 884         unsigned int node;
 885         unsigned int count;
 886
 887         if (min_common_depth == -1 || !numa_enabled)
 888                 return;
 889
 890         for_each_online_node(node) {
 891                 unsigned long i;
 892
 893                 printk(KERN_DEBUG "Node %d Memory:", node);
 894
 895                 count = 0;
 896
 897                 for (i = 0; i < memblock_end_of_DRAM();
 898                      i += (1 << SECTION_SIZE_BITS)) {
 899                         if (early_pfn_to_nid(i >> PAGE_SHIFT) == node) {
 900                                 if (count == 0)
 901                                         printk(" 0x%lx", i);
 902                                 ++count;
 903                         } else {
 904                                 if (count > 0)
 905                                         printk("-0x%lx", i);
 906                                 count = 0;
 907                         }
 908                 }
 909
 910                 if (count > 0)
 911                         printk("-0x%lx", i);
 912                 printk("\n");
 913         }
 914 }
 915
 916 static struct notifier_block ppc64_numa_nb = {
 917         .notifier_call = cpu_numa_callback,
 918         .priority = 1 /* Must run before sched domains notifier. */
 919 };
 920
 921 /* Initialize NODE_DATA for a node on the local memory */
 922 static void __init setup_node_data(int nid, u64 start_pfn, u64 end_pfn)
 923 {
 924         u64 spanned_pages = end_pfn - start_pfn;
 925         const size_t nd_size = roundup(sizeof(pg_data_t), SMP_CACHE_BYTES);
 926         u64 nd_pa;
 927         void *nd;
 928         int tnid;
 929
 930         if (spanned_pages)
 931                 pr_info("Initmem setup node %d [mem %#010Lx-%#010Lx]\n",
 932                         nid, start_pfn << PAGE_SHIFT,
 933                         (end_pfn << PAGE_SHIFT) - 1);
 934         else
 935                 pr_info("Initmem setup node %d\n", nid);
 936
 937         nd_pa = memblock_alloc_try_nid(nd_size, SMP_CACHE_BYTES, nid);
 938         nd = __va(nd_pa);
 939
 940         /* report and initialize */
 941         pr_info("  NODE_DATA [mem %#010Lx-%#010Lx]\n",
 942                 nd_pa, nd_pa + nd_size - 1);
 943         tnid = early_pfn_to_nid(nd_pa >> PAGE_SHIFT);
 944         if (tnid != nid)
 945                 pr_info("    NODE_DATA(%d) on node %d\n", nid, tnid);
 946
 947         node_data[nid] = nd;
 948         memset(NODE_DATA(nid), 0, sizeof(pg_data_t));
 949         NODE_DATA(nid)->node_id = nid;
 950         NODE_DATA(nid)->node_start_pfn = start_pfn;
 951         NODE_DATA(nid)->node_spanned_pages = spanned_pages;
 952 }
 953
 954 static void __init find_possible_nodes(void)
 955 {
 956         struct device_node *rtas;
 957         u32 numnodes, i;
 958
 959         if (min_common_depth <= 0)
 960                 return;
 961
 962         rtas = of_find_node_by_path("/rtas");
 963         if (!rtas)
 964                 return;
 965
 966         if (of_property_read_u32_index(rtas,
 967                                 "ibm,max-associativity-domains",
 968                                 min_common_depth, &numnodes))
 969                 goto out;
 970
 971         for (i = 0; i < numnodes; i++) {
 972                 if (!node_possible(i))
 973                         node_set(i, node_possible_map);
 974         }
 975
 976 out:
 977         of_node_put(rtas);
 978 }
 979
 980 void __init initmem_init(void)
 981 {
 982         int nid, cpu;
 983
 984         max_low_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
 985         max_pfn = max_low_pfn;
 986
 987         if (parse_numa_properties())
 988                 setup_nonnuma();
 989         else
 990                 dump_numa_memory_topology();
 991
 992         memblock_dump_all();
 993
 994         /*
 995          * Modify the set of possible NUMA nodes to reflect information
 996          * available about the set of online nodes, and the set of nodes
 997          * that we expect to make use of for this platform's affinity
 998          * calculations.
 999          */
1000         nodes_and(node_possible_map, node_possible_map, node_online_map);
1001
1002         find_possible_nodes();
1003
1004         for_each_online_node(nid) {
1005                 unsigned long start_pfn, end_pfn;
1006
1007                 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
1008                 setup_node_data(nid, start_pfn, end_pfn);
1009                 sparse_memory_present_with_active_regions(nid);
1010         }
1011
1012         sparse_init();
1013
1014         setup_node_to_cpumask_map();
1015
1016         reset_numa_cpu_lookup_table();
1017         register_cpu_notifier(&ppc64_numa_nb);
1018         /*
1019          * We need the numa_cpu_lookup_table to be accurate for all CPUs,
1020          * even before we online them, so that we can use cpu_to_{node,mem}
1021          * early in boot, cf. smp_prepare_cpus().
1022          */
1023         for_each_present_cpu(cpu) {
1024                 numa_setup_cpu((unsigned long)cpu);
1025         }
1026 }
1027
1028 static int __init early_numa(char *p)
1029 {
1030         if (!p)
1031                 return 0;
1032
1033         if (strstr(p, "off"))
1034                 numa_enabled = 0;
1035
1036         if (strstr(p, "debug"))
1037                 numa_debug = 1;
1038
1039         p = strstr(p, "fake=");
1040         if (p)
1041                 cmdline = p + strlen("fake=");
1042
1043         return 0;
1044 }
1045 early_param("numa", early_numa);
1046
1047 static bool topology_updates_enabled = true;
1048
1049 static int __init early_topology_updates(char *p)
1050 {
1051         if (!p)
1052                 return 0;
1053
1054         if (!strcmp(p, "off")) {
1055                 pr_info("Disabling topology updates\n");
1056                 topology_updates_enabled = false;
1057         }
1058
1059         return 0;
1060 }
1061 early_param("topology_updates", early_topology_updates);
1062
1063 #ifdef CONFIG_MEMORY_HOTPLUG
1064 /*
1065  * Find the node associated with a hot added memory section for
1066  * memory represented in the device tree by the property
1067  * ibm,dynamic-reconfiguration-memory/ibm,dynamic-memory.
1068  */
1069 static int hot_add_drconf_scn_to_nid(struct device_node *memory,
1070                                      unsigned long scn_addr)
1071 {
1072         const __be32 *dm;
1073         unsigned int drconf_cell_cnt, rc;
1074         unsigned long lmb_size;
1075         struct assoc_arrays aa;
1076         int nid = -1;
1077
1078         drconf_cell_cnt = of_get_drconf_memory(memory, &dm);
1079         if (!drconf_cell_cnt)
1080                 return -1;
1081
1082         lmb_size = of_get_lmb_size(memory);
1083         if (!lmb_size)
1084                 return -1;
1085
1086         rc = of_get_assoc_arrays(memory, &aa);
1087         if (rc)
1088                 return -1;
1089
1090         for (; drconf_cell_cnt != 0; --drconf_cell_cnt) {
1091                 struct of_drconf_cell drmem;
1092
1093                 read_drconf_cell(&drmem, &dm);
1094
1095                 /* skip this block if it is reserved or not assigned to
1096                  * this partition */
1097                 if ((drmem.flags & DRCONF_MEM_RESERVED)
1098                     || !(drmem.flags & DRCONF_MEM_ASSIGNED))
1099                         continue;
1100
1101                 if ((scn_addr < drmem.base_addr)
1102                     || (scn_addr >= (drmem.base_addr + lmb_size)))
1103                         continue;
1104
1105                 nid = of_drconf_to_nid_single(&drmem, &aa);
1106                 break;
1107         }
1108
1109         return nid;
1110 }
1111
1112 /*
1113  * Find the node associated with a hot added memory section for memory
1114  * represented in the device tree as a node (i.e. memory@XXXX) for
1115  * each memblock.
1116  */
1117 static int hot_add_node_scn_to_nid(unsigned long scn_addr)
1118 {
1119         struct device_node *memory;
1120         int nid = -1;
1121
1122         for_each_node_by_type(memory, "memory") {
1123                 unsigned long start, size;
1124                 int ranges;
1125                 const __be32 *memcell_buf;
1126                 unsigned int len;
1127
1128                 memcell_buf = of_get_property(memory, "reg", &len);
1129                 if (!memcell_buf || len <= 0)
1130                         continue;
1131
1132                 /* ranges in cell */
1133                 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
1134
1135                 while (ranges--) {
1136                         start = read_n_cells(n_mem_addr_cells, &memcell_buf);
1137                         size = read_n_cells(n_mem_size_cells, &memcell_buf);
1138
1139                         if ((scn_addr < start) || (scn_addr >= (start + size)))
1140                                 continue;
1141
1142                         nid = of_node_to_nid_single(memory);
1143                         break;
1144                 }
1145
1146                 if (nid >= 0)
1147                         break;
1148         }
1149
1150         of_node_put(memory);
1151
1152         return nid;
1153 }
1154
1155 /*
1156  * Find the node associated with a hot added memory section.  Section
1157  * corresponds to a SPARSEMEM section, not an MEMBLOCK.  It is assumed that
1158  * sections are fully contained within a single MEMBLOCK.
1159  */
1160 int hot_add_scn_to_nid(unsigned long scn_addr)
1161 {
1162         struct device_node *memory = NULL;
1163         int nid, found = 0;
1164
1165         if (!numa_enabled || (min_common_depth < 0))
1166                 return first_online_node;
1167
1168         memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1169         if (memory) {
1170                 nid = hot_add_drconf_scn_to_nid(memory, scn_addr);
1171                 of_node_put(memory);
1172         } else {
1173                 nid = hot_add_node_scn_to_nid(scn_addr);
1174         }
1175
1176         if (nid < 0 || !node_online(nid))
1177                 nid = first_online_node;
1178
1179         if (NODE_DATA(nid)->node_spanned_pages)
1180                 return nid;
1181
1182         for_each_online_node(nid) {
1183                 if (NODE_DATA(nid)->node_spanned_pages) {
1184                         found = 1;
1185                         break;
1186                 }
1187         }
1188
1189         BUG_ON(!found);
1190         return nid;
1191 }
1192
1193 static u64 hot_add_drconf_memory_max(void)
1194 {
1195         struct device_node *memory = NULL;
1196         unsigned int drconf_cell_cnt = 0;
1197         u64 lmb_size = 0;
1198         const __be32 *dm = NULL;
1199
1200         memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1201         if (memory) {
1202                 drconf_cell_cnt = of_get_drconf_memory(memory, &dm);
1203                 lmb_size = of_get_lmb_size(memory);
1204                 of_node_put(memory);
1205         }
1206         return lmb_size * drconf_cell_cnt;
1207 }
1208
1209 /*
1210  * memory_hotplug_max - return max address of memory that may be added
1211  *
1212  * This is currently only used on systems that support drconfig memory
1213  * hotplug.
1214  */
1215 u64 memory_hotplug_max(void)
1216 {
1217         return max(hot_add_drconf_memory_max(), memblock_end_of_DRAM());
1218 }
1219 #endif /* CONFIG_MEMORY_HOTPLUG */
1220
1221 /* Virtual Processor Home Node (VPHN) support */
1222 #ifdef CONFIG_PPC_SPLPAR
1223
1224 #include "vphn.h"
1225
1226 struct topology_update_data {
1227         struct topology_update_data *next;
1228         unsigned int cpu;
1229         int old_nid;
1230         int new_nid;
1231 };
1232
1233 static u8 vphn_cpu_change_counts[NR_CPUS][MAX_DISTANCE_REF_POINTS];
1234 static cpumask_t cpu_associativity_changes_mask;
1235 static int vphn_enabled;
1236 static int prrn_enabled;
1237 static void reset_topology_timer(void);
1238
1239 /*
1240  * Store the current values of the associativity change counters in the
1241  * hypervisor.
1242  */
1243 static void setup_cpu_associativity_change_counters(void)
1244 {
1245         int cpu;
1246
1247         /* The VPHN feature supports a maximum of 8 reference points */
1248         BUILD_BUG_ON(MAX_DISTANCE_REF_POINTS > 8);
1249
1250         for_each_possible_cpu(cpu) {
1251                 int i;
1252                 u8 *counts = vphn_cpu_change_counts[cpu];
1253                 volatile u8 *hypervisor_counts = lppaca[cpu].vphn_assoc_counts;
1254
1255                 for (i = 0; i < distance_ref_points_depth; i++)
1256                         counts[i] = hypervisor_counts[i];
1257         }
1258 }
1259
1260 /*
1261  * The hypervisor maintains a set of 8 associativity change counters in
1262  * the VPA of each cpu that correspond to the associativity levels in the
1263  * ibm,associativity-reference-points property. When an associativity
1264  * level changes, the corresponding counter is incremented.
1265  *
1266  * Set a bit in cpu_associativity_changes_mask for each cpu whose home
1267  * node associativity levels have changed.
1268  *
1269  * Returns the number of cpus with unhandled associativity changes.
1270  */
1271 static int update_cpu_associativity_changes_mask(void)
1272 {
1273         int cpu;
1274         cpumask_t *changes = &cpu_associativity_changes_mask;
1275
1276         for_each_possible_cpu(cpu) {
1277                 int i, changed = 0;
1278                 u8 *counts = vphn_cpu_change_counts[cpu];
1279                 volatile u8 *hypervisor_counts = lppaca[cpu].vphn_assoc_counts;
1280
1281                 for (i = 0; i < distance_ref_points_depth; i++) {
1282                         if (hypervisor_counts[i] != counts[i]) {
1283                                 counts[i] = hypervisor_counts[i];
1284                                 changed = 1;
1285                         }
1286                 }
1287                 if (changed) {
1288                         cpumask_or(changes, changes, cpu_sibling_mask(cpu));
1289                         cpu = cpu_last_thread_sibling(cpu);
1290                 }
1291         }
1292
1293         return cpumask_weight(changes);
1294 }
1295
1296 /*
1297  * Retrieve the new associativity information for a virtual processor's
1298  * home node.
1299  */
1300 static long hcall_vphn(unsigned long cpu, __be32 *associativity)
1301 {
1302         long rc;
1303         long retbuf[PLPAR_HCALL9_BUFSIZE] = {0};
1304         u64 flags = 1;
1305         int hwcpu = get_hard_smp_processor_id(cpu);
1306
1307         rc = plpar_hcall9(H_HOME_NODE_ASSOCIATIVITY, retbuf, flags, hwcpu);
1308         vphn_unpack_associativity(retbuf, associativity);
1309
1310         return rc;
1311 }
1312
1313 static long vphn_get_associativity(unsigned long cpu,
1314                                         __be32 *associativity)
1315 {
1316         long rc;
1317
1318         rc = hcall_vphn(cpu, associativity);
1319
1320         switch (rc) {
1321         case H_FUNCTION:
1322                 printk_once(KERN_INFO
1323                         "VPHN is not supported. Disabling polling...\n");
1324                 stop_topology_update();
1325                 break;
1326         case H_HARDWARE:
1327                 printk(KERN_ERR
1328                         "hcall_vphn() experienced a hardware fault "
1329                         "preventing VPHN. Disabling polling...\n");
1330                 stop_topology_update();
1331         }
1332
1333         return rc;
1334 }
1335
1336 static inline int find_and_online_cpu_nid(int cpu)
1337 {
1338         __be32 associativity[VPHN_ASSOC_BUFSIZE] = {0};
1339         int new_nid;
1340
1341         /* Use associativity from first thread for all siblings */
1342         vphn_get_associativity(cpu, associativity);
1343         new_nid = associativity_to_nid(associativity);
1344         if (new_nid < 0 || !node_possible(new_nid))
1345                 new_nid = first_online_node;
1346
1347         if (NODE_DATA(new_nid) == NULL) {
1348 #ifdef CONFIG_MEMORY_HOTPLUG
1349                 /*
1350                  * Need to ensure that NODE_DATA is initialized for a node from
1351                  * available memory (see memblock_alloc_try_nid). If unable to
1352                  * init the node, then default to nearest node that has memory
1353                  * installed.
1354                  */
1355                 if (try_online_node(new_nid))
1356                         new_nid = first_online_node;
1357 #else
1358                 /*
1359                  * Default to using the nearest node that has memory installed.
1360                  * Otherwise, it would be necessary to patch the kernel MM code
1361                  * to deal with more memoryless-node error conditions.
1362                  */
1363                 new_nid = first_online_node;
1364 #endif
1365         }
1366
1367         return new_nid;
1368 }
1369
1370 /*
1371  * Update the CPU maps and sysfs entries for a single CPU when its NUMA
1372  * characteristics change. This function doesn't perform any locking and is
1373  * only safe to call from stop_machine().
1374  */
1375 static int update_cpu_topology(void *data)
1376 {
1377         struct topology_update_data *update;
1378         unsigned long cpu;
1379
1380         if (!data)
1381                 return -EINVAL;
1382
1383         cpu = smp_processor_id();
1384
1385         for (update = data; update; update = update->next) {
1386                 int new_nid = update->new_nid;
1387                 if (cpu != update->cpu)
1388                         continue;
1389
1390                 unmap_cpu_from_node(cpu);
1391                 map_cpu_to_node(cpu, new_nid);
1392                 set_cpu_numa_node(cpu, new_nid);
1393                 set_cpu_numa_mem(cpu, local_memory_node(new_nid));
1394                 vdso_getcpu_init();
1395         }
1396
1397         return 0;
1398 }
1399
1400 static int update_lookup_table(void *data)
1401 {
1402         struct topology_update_data *update;
1403
1404         if (!data)
1405                 return -EINVAL;
1406
1407         /*
1408          * Upon topology update, the numa-cpu lookup table needs to be updated
1409          * for all threads in the core, including offline CPUs, to ensure that
1410          * future hotplug operations respect the cpu-to-node associativity
1411          * properly.
1412          */
1413         for (update = data; update; update = update->next) {
1414                 int nid, base, j;
1415
1416                 nid = update->new_nid;
1417                 base = cpu_first_thread_sibling(update->cpu);
1418
1419                 for (j = 0; j < threads_per_core; j++) {
1420                         update_numa_cpu_lookup_table(base + j, nid);
1421                 }
1422         }
1423
1424         return 0;
1425 }
1426
1427 /*
1428  * Update the node maps and sysfs entries for each cpu whose home node
1429  * has changed. Returns 1 when the topology has changed, and 0 otherwise.
1430  */
1431 int arch_update_cpu_topology(void)
1432 {
1433         unsigned int cpu, sibling, changed = 0;
1434         struct topology_update_data *updates, *ud;
1435         cpumask_t updated_cpus;
1436         struct device *dev;
1437         int weight, new_nid, i = 0;
1438
1439         if (!prrn_enabled && !vphn_enabled)
1440                 return 0;
1441
1442         weight = cpumask_weight(&cpu_associativity_changes_mask);
1443         if (!weight)
1444                 return 0;
1445
1446         updates = kzalloc(weight * (sizeof(*updates)), GFP_KERNEL);
1447         if (!updates)
1448                 return 0;
1449
1450         cpumask_clear(&updated_cpus);
1451
1452         for_each_cpu(cpu, &cpu_associativity_changes_mask) {
1453                 /*
1454                  * If siblings aren't flagged for changes, updates list
1455                  * will be too short. Skip on this update and set for next
1456                  * update.
1457                  */
1458                 if (!cpumask_subset(cpu_sibling_mask(cpu),
1459                                         &cpu_associativity_changes_mask)) {
1460                         pr_info("Sibling bits not set for associativity "
1461                                         "change, cpu%d\n", cpu);
1462                         cpumask_or(&cpu_associativity_changes_mask,
1463                                         &cpu_associativity_changes_mask,
1464                                         cpu_sibling_mask(cpu));
1465                         cpu = cpu_last_thread_sibling(cpu);
1466                         continue;
1467                 }
1468
1469                 new_nid = find_and_online_cpu_nid(cpu);
1470
1471                 if (new_nid == numa_cpu_lookup_table[cpu]) {
1472                         cpumask_andnot(&cpu_associativity_changes_mask,
1473                                         &cpu_associativity_changes_mask,
1474                                         cpu_sibling_mask(cpu));
1475                         cpu = cpu_last_thread_sibling(cpu);
1476                         continue;
1477                 }
1478
1479                 for_each_cpu(sibling, cpu_sibling_mask(cpu)) {
1480                         ud = &updates[i++];
1481                         ud->cpu = sibling;
1482                         ud->new_nid = new_nid;
1483                         ud->old_nid = numa_cpu_lookup_table[sibling];
1484                         cpumask_set_cpu(sibling, &updated_cpus);
1485                         if (i < weight)
1486                                 ud->next = &updates[i];
1487                 }
1488                 cpu = cpu_last_thread_sibling(cpu);
1489         }
1490
1491         pr_debug("Topology update for the following CPUs:\n");
1492         if (cpumask_weight(&updated_cpus)) {
1493                 for (ud = &updates[0]; ud; ud = ud->next) {
1494                         pr_debug("cpu %d moving from node %d "
1495                                           "to %d\n", ud->cpu,
1496                                           ud->old_nid, ud->new_nid);
1497                 }
1498         }
1499
1500         /*
1501          * In cases where we have nothing to update (because the updates list
1502          * is too short or because the new topology is same as the old one),
1503          * skip invoking update_cpu_topology() via stop-machine(). This is
1504          * necessary (and not just a fast-path optimization) since stop-machine
1505          * can end up electing a random CPU to run update_cpu_topology(), and
1506          * thus trick us into setting up incorrect cpu-node mappings (since
1507          * 'updates' is kzalloc()'ed).
1508          *
1509          * And for the similar reason, we will skip all the following updating.
1510          */
1511         if (!cpumask_weight(&updated_cpus))
1512                 goto out;
1513
1514         stop_machine(update_cpu_topology, &updates[0], &updated_cpus);
1515
1516         /*
1517          * Update the numa-cpu lookup table with the new mappings, even for
1518          * offline CPUs. It is best to perform this update from the stop-
1519          * machine context.
1520          */
1521         stop_machine(update_lookup_table, &updates[0],
1522                                         cpumask_of(raw_smp_processor_id()));
1523
1524         for (ud = &updates[0]; ud; ud = ud->next) {
1525                 unregister_cpu_under_node(ud->cpu, ud->old_nid);
1526                 register_cpu_under_node(ud->cpu, ud->new_nid);
1527
1528                 dev = get_cpu_device(ud->cpu);
1529                 if (dev)
1530                         kobject_uevent(&dev->kobj, KOBJ_CHANGE);
1531                 cpumask_clear_cpu(ud->cpu, &cpu_associativity_changes_mask);
1532                 changed = 1;
1533         }
1534
1535 out:
1536         kfree(updates);
1537         return changed;
1538 }
1539
1540 static void topology_work_fn(struct work_struct *work)
1541 {
1542         rebuild_sched_domains();
1543 }
1544 static DECLARE_WORK(topology_work, topology_work_fn);
1545
1546 static void topology_schedule_update(void)
1547 {
1548         schedule_work(&topology_work);
1549 }
1550
1551 static void topology_timer_fn(unsigned long ignored)
1552 {
1553         if (prrn_enabled && cpumask_weight(&cpu_associativity_changes_mask))
1554                 topology_schedule_update();
1555         else if (vphn_enabled) {
1556                 if (update_cpu_associativity_changes_mask() > 0)
1557                         topology_schedule_update();
1558                 reset_topology_timer();
1559         }
1560 }
1561 static struct timer_list topology_timer =
1562         TIMER_INITIALIZER(topology_timer_fn, 0, 0);
1563
1564 static void reset_topology_timer(void)
1565 {
1566         topology_timer.data = 0;
1567         topology_timer.expires = jiffies + 60 * HZ;
1568         mod_timer(&topology_timer, topology_timer.expires);
1569 }
1570
1571 #ifdef CONFIG_SMP
1572
1573 static void stage_topology_update(int core_id)
1574 {
1575         cpumask_or(&cpu_associativity_changes_mask,
1576                 &cpu_associativity_changes_mask, cpu_sibling_mask(core_id));
1577         reset_topology_timer();
1578 }
1579
1580 static int dt_update_callback(struct notifier_block *nb,
1581                                 unsigned long action, void *data)
1582 {
1583         struct of_reconfig_data *update = data;
1584         int rc = NOTIFY_DONE;
1585
1586         switch (action) {
1587         case OF_RECONFIG_UPDATE_PROPERTY:
1588                 if (!of_prop_cmp(update->dn->type, "cpu") &&
1589                     !of_prop_cmp(update->prop->name, "ibm,associativity")) {
1590                         u32 core_id;
1591                         of_property_read_u32(update->dn, "reg", &core_id);
1592                         stage_topology_update(core_id);
1593                         rc = NOTIFY_OK;
1594                 }
1595                 break;
1596         }
1597
1598         return rc;
1599 }
1600
1601 static struct notifier_block dt_update_nb = {
1602         .notifier_call = dt_update_callback,
1603 };
1604
1605 #endif
1606
1607 /*
1608  * Start polling for associativity changes.
1609  */
1610 int start_topology_update(void)
1611 {
1612         int rc = 0;
1613
1614         if (!topology_updates_enabled)
1615                 return 0;
1616
1617         if (firmware_has_feature(FW_FEATURE_PRRN)) {
1618                 if (!prrn_enabled) {
1619                         prrn_enabled = 1;
1620                         vphn_enabled = 0;
1621 #ifdef CONFIG_SMP
1622                         rc = of_reconfig_notifier_register(&dt_update_nb);
1623 #endif
1624                 }
1625         } else if (firmware_has_feature(FW_FEATURE_VPHN) &&
1626                    lppaca_shared_proc(get_lppaca())) {
1627                 if (!vphn_enabled) {
1628                         prrn_enabled = 0;
1629                         vphn_enabled = 1;
1630                         setup_cpu_associativity_change_counters();
1631                         init_timer_deferrable(&topology_timer);
1632                         reset_topology_timer();
1633                 }
1634         }
1635
1636         return rc;
1637 }
1638
1639 /*
1640  * Disable polling for VPHN associativity changes.
1641  */
1642 int stop_topology_update(void)
1643 {
1644         int rc = 0;
1645
1646         if (!topology_updates_enabled)
1647                 return 0;
1648
1649         if (prrn_enabled) {
1650                 prrn_enabled = 0;
1651 #ifdef CONFIG_SMP
1652                 rc = of_reconfig_notifier_unregister(&dt_update_nb);
1653 #endif
1654         } else if (vphn_enabled) {
1655                 vphn_enabled = 0;
1656                 rc = del_timer_sync(&topology_timer);
1657         }
1658
1659         return rc;
1660 }
1661
1662 int prrn_is_enabled(void)
1663 {
1664         return prrn_enabled;
1665 }
1666
1667 static int topology_read(struct seq_file *file, void *v)
1668 {
1669         if (vphn_enabled || prrn_enabled)
1670                 seq_puts(file, "on\n");
1671         else
1672                 seq_puts(file, "off\n");
1673
1674         return 0;
1675 }
1676
1677 static int topology_open(struct inode *inode, struct file *file)
1678 {
1679         return single_open(file, topology_read, NULL);
1680 }
1681
1682 static ssize_t topology_write(struct file *file, const char __user *buf,
1683                               size_t count, loff_t *off)
1684 {
1685         char kbuf[4]; /* "on" or "off" plus null. */
1686         int read_len;
1687
1688         read_len = count < 3 ? count : 3;
1689         if (copy_from_user(kbuf, buf, read_len))
1690                 return -EINVAL;
1691
1692         kbuf[read_len] = '\0';
1693
1694         if (!strncmp(kbuf, "on", 2)) {
1695                 topology_updates_enabled = true;
1696                 start_topology_update();
1697         } else if (!strncmp(kbuf, "off", 3)) {
1698                 stop_topology_update();
1699                 topology_updates_enabled = false;
1700         } else
1701                 return -EINVAL;
1702
1703         return count;
1704 }
1705
1706 static const struct file_operations topology_ops = {
1707         .read = seq_read,
1708         .write = topology_write,
1709         .open = topology_open,
1710         .release = single_release
1711 };
1712
1713 static int topology_update_init(void)
1714 {
1715         start_topology_update();
1716
1717         if (!proc_create("powerpc/topology_updates", 0644, NULL, &topology_ops))
1718                 return -ENOMEM;
1719
1720         return 0;
1721 }
1722 device_initcall(topology_update_init);
1723 #endif /* CONFIG_PPC_SPLPAR */