arch/powerpc/mm/numa.c

   1 /*
   2  * pSeries NUMA support
   3  *
   4  * Copyright (C) 2002 Anton Blanchard <anton@au.ibm.com>, IBM
   5  *
   6  * This program is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU General Public License
   8  * as published by the Free Software Foundation; either version
   9  * 2 of the License, or (at your option) any later version.
  10  */
  11 #include <linux/threads.h>
  12 #include <linux/bootmem.h>
  13 #include <linux/init.h>
  14 #include <linux/mm.h>
  15 #include <linux/mmzone.h>
  16 #include <linux/export.h>
  17 #include <linux/nodemask.h>
  18 #include <linux/cpu.h>
  19 #include <linux/notifier.h>
  20 #include <linux/memblock.h>
  21 #include <linux/of.h>
  22 #include <linux/pfn.h>
  23 #include <linux/cpuset.h>
  24 #include <linux/node.h>
  25 #include <linux/stop_machine.h>
  26 #include <linux/proc_fs.h>
  27 #include <linux/seq_file.h>
  28 #include <linux/uaccess.h>
  29 #include <linux/slab.h>
  30 #include <asm/cputhreads.h>
  31 #include <asm/sparsemem.h>
  32 #include <asm/prom.h>
  33 #include <asm/smp.h>
  34 #include <asm/cputhreads.h>
  35 #include <asm/topology.h>
  36 #include <asm/firmware.h>
  37 #include <asm/paca.h>
  38 #include <asm/hvcall.h>
  39 #include <asm/setup.h>
  40 #include <asm/vdso.h>
  41
  42 static int numa_enabled = 1;
  43
  44 static char *cmdline __initdata;
  45
  46 static int numa_debug;
  47 #define dbg(args...) if (numa_debug) { printk(KERN_INFO args); }
  48
  49 int numa_cpu_lookup_table[NR_CPUS];
  50 cpumask_var_t node_to_cpumask_map[MAX_NUMNODES];
  51 struct pglist_data *node_data[MAX_NUMNODES];
  52
  53 EXPORT_SYMBOL(numa_cpu_lookup_table);
  54 EXPORT_SYMBOL(node_to_cpumask_map);
  55 EXPORT_SYMBOL(node_data);
  56
  57 static int min_common_depth;
  58 static int n_mem_addr_cells, n_mem_size_cells;
  59 static int form1_affinity;
  60
  61 #define MAX_DISTANCE_REF_POINTS 4
  62 static int distance_ref_points_depth;
  63 static const unsigned int *distance_ref_points;
  64 static int distance_lookup_table[MAX_NUMNODES][MAX_DISTANCE_REF_POINTS];
  65
  66 /*
  67  * Allocate node_to_cpumask_map based on number of available nodes
  68  * Requires node_possible_map to be valid.
  69  *
  70  * Note: cpumask_of_node() is not valid until after this is done.
  71  */
  72 static void __init setup_node_to_cpumask_map(void)
  73 {
  74         unsigned int node;
  75
  76         /* setup nr_node_ids if not done yet */
  77         if (nr_node_ids == MAX_NUMNODES)
  78                 setup_nr_node_ids();
  79
  80         /* allocate the map */
  81         for (node = 0; node < nr_node_ids; node++)
  82                 alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]);
  83
  84         /* cpumask_of_node() will now work */
  85         dbg("Node to cpumask map for %d nodes\n", nr_node_ids);
  86 }
  87
  88 static int __init fake_numa_create_new_node(unsigned long end_pfn,
  89                                                 unsigned int *nid)
  90 {
  91         unsigned long long mem;
  92         char *p = cmdline;
  93         static unsigned int fake_nid;
  94         static unsigned long long curr_boundary;
  95
  96         /*
  97          * Modify node id, iff we started creating NUMA nodes
  98          * We want to continue from where we left of the last time
  99          */
 100         if (fake_nid)
 101                 *nid = fake_nid;
 102         /*
 103          * In case there are no more arguments to parse, the
 104          * node_id should be the same as the last fake node id
 105          * (we've handled this above).
 106          */
 107         if (!p)
 108                 return 0;
 109
 110         mem = memparse(p, &p);
 111         if (!mem)
 112                 return 0;
 113
 114         if (mem < curr_boundary)
 115                 return 0;
 116
 117         curr_boundary = mem;
 118
 119         if ((end_pfn << PAGE_SHIFT) > mem) {
 120                 /*
 121                  * Skip commas and spaces
 122                  */
 123                 while (*p == ',' || *p == ' ' || *p == '\t')
 124                         p++;
 125
 126                 cmdline = p;
 127                 fake_nid++;
 128                 *nid = fake_nid;
 129                 dbg("created new fake_node with id %d\n", fake_nid);
 130                 return 1;
 131         }
 132         return 0;
 133 }
 134
 135 /*
 136  * get_node_active_region - Return active region containing pfn
 137  * Active range returned is empty if none found.
 138  * @pfn: The page to return the region for
 139  * @node_ar: Returned set to the active region containing @pfn
 140  */
 141 static void __init get_node_active_region(unsigned long pfn,
 142                                           struct node_active_region *node_ar)
 143 {
 144         unsigned long start_pfn, end_pfn;
 145         int i, nid;
 146
 147         for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
 148                 if (pfn >= start_pfn && pfn < end_pfn) {
 149                         node_ar->nid = nid;
 150                         node_ar->start_pfn = start_pfn;
 151                         node_ar->end_pfn = end_pfn;
 152                         break;
 153                 }
 154         }
 155 }
 156
 157 static void reset_numa_cpu_lookup_table(void)
 158 {
 159         unsigned int cpu;
 160
 161         for_each_possible_cpu(cpu)
 162                 numa_cpu_lookup_table[cpu] = -1;
 163 }
 164
 165 static void update_numa_cpu_lookup_table(unsigned int cpu, int node)
 166 {
 167         numa_cpu_lookup_table[cpu] = node;
 168 }
 169
 170 static void map_cpu_to_node(int cpu, int node)
 171 {
 172         update_numa_cpu_lookup_table(cpu, node);
 173
 174         dbg("adding cpu %d to node %d\n", cpu, node);
 175
 176         if (!(cpumask_test_cpu(cpu, node_to_cpumask_map[node])))
 177                 cpumask_set_cpu(cpu, node_to_cpumask_map[node]);
 178 }
 179
 180 #if defined(CONFIG_HOTPLUG_CPU) || defined(CONFIG_PPC_SPLPAR)
 181 static void unmap_cpu_from_node(unsigned long cpu)
 182 {
 183         int node = numa_cpu_lookup_table[cpu];
 184
 185         dbg("removing cpu %lu from node %d\n", cpu, node);
 186
 187         if (cpumask_test_cpu(cpu, node_to_cpumask_map[node])) {
 188                 cpumask_clear_cpu(cpu, node_to_cpumask_map[node]);
 189         } else {
 190                 printk(KERN_ERR "WARNING: cpu %lu not found in node %d\n",
 191                        cpu, node);
 192         }
 193 }
 194 #endif /* CONFIG_HOTPLUG_CPU || CONFIG_PPC_SPLPAR */
 195
 196 /* must hold reference to node during call */
 197 static const int *of_get_associativity(struct device_node *dev)
 198 {
 199         return of_get_property(dev, "ibm,associativity", NULL);
 200 }
 201
 202 /*
 203  * Returns the property linux,drconf-usable-memory if
 204  * it exists (the property exists only in kexec/kdump kernels,
 205  * added by kexec-tools)
 206  */
 207 static const u32 *of_get_usable_memory(struct device_node *memory)
 208 {
 209         const u32 *prop;
 210         u32 len;
 211         prop = of_get_property(memory, "linux,drconf-usable-memory", &len);
 212         if (!prop || len < sizeof(unsigned int))
 213                 return 0;
 214         return prop;
 215 }
 216
 217 int __node_distance(int a, int b)
 218 {
 219         int i;
 220         int distance = LOCAL_DISTANCE;
 221
 222         if (!form1_affinity)
 223                 return ((a == b) ? LOCAL_DISTANCE : REMOTE_DISTANCE);
 224
 225         for (i = 0; i < distance_ref_points_depth; i++) {
 226                 if (distance_lookup_table[a][i] == distance_lookup_table[b][i])
 227                         break;
 228
 229                 /* Double the distance for each NUMA level */
 230                 distance *= 2;
 231         }
 232
 233         return distance;
 234 }
 235
 236 static void initialize_distance_lookup_table(int nid,
 237                 const unsigned int *associativity)
 238 {
 239         int i;
 240
 241         if (!form1_affinity)
 242                 return;
 243
 244         for (i = 0; i < distance_ref_points_depth; i++) {
 245                 distance_lookup_table[nid][i] =
 246                         associativity[distance_ref_points[i]];
 247         }
 248 }
 249
 250 /* Returns nid in the range [0..MAX_NUMNODES-1], or -1 if no useful numa
 251  * info is found.
 252  */
 253 static int associativity_to_nid(const unsigned int *associativity)
 254 {
 255         int nid = -1;
 256
 257         if (min_common_depth == -1)
 258                 goto out;
 259
 260         if (associativity[0] >= min_common_depth)
 261                 nid = associativity[min_common_depth];
 262
 263         /* POWER4 LPAR uses 0xffff as invalid node */
 264         if (nid == 0xffff || nid >= MAX_NUMNODES)
 265                 nid = -1;
 266
 267         if (nid > 0 && associativity[0] >= distance_ref_points_depth)
 268                 initialize_distance_lookup_table(nid, associativity);
 269
 270 out:
 271         return nid;
 272 }
 273
 274 /* Returns the nid associated with the given device tree node,
 275  * or -1 if not found.
 276  */
 277 static int of_node_to_nid_single(struct device_node *device)
 278 {
 279         int nid = -1;
 280         const unsigned int *tmp;
 281
 282         tmp = of_get_associativity(device);
 283         if (tmp)
 284                 nid = associativity_to_nid(tmp);
 285         return nid;
 286 }
 287
 288 /* Walk the device tree upwards, looking for an associativity id */
 289 int of_node_to_nid(struct device_node *device)
 290 {
 291         struct device_node *tmp;
 292         int nid = -1;
 293
 294         of_node_get(device);
 295         while (device) {
 296                 nid = of_node_to_nid_single(device);
 297                 if (nid != -1)
 298                         break;
 299
 300                 tmp = device;
 301                 device = of_get_parent(tmp);
 302                 of_node_put(tmp);
 303         }
 304         of_node_put(device);
 305
 306         return nid;
 307 }
 308 EXPORT_SYMBOL_GPL(of_node_to_nid);
 309
 310 static int __init find_min_common_depth(void)
 311 {
 312         int depth;
 313         struct device_node *root;
 314
 315         if (firmware_has_feature(FW_FEATURE_OPAL))
 316                 root = of_find_node_by_path("/ibm,opal");
 317         else
 318                 root = of_find_node_by_path("/rtas");
 319         if (!root)
 320                 root = of_find_node_by_path("/");
 321
 322         /*
 323          * This property is a set of 32-bit integers, each representing
 324          * an index into the ibm,associativity nodes.
 325          *
 326          * With form 0 affinity the first integer is for an SMP configuration
 327          * (should be all 0's) and the second is for a normal NUMA
 328          * configuration. We have only one level of NUMA.
 329          *
 330          * With form 1 affinity the first integer is the most significant
 331          * NUMA boundary and the following are progressively less significant
 332          * boundaries. There can be more than one level of NUMA.
 333          */
 334         distance_ref_points = of_get_property(root,
 335                                         "ibm,associativity-reference-points",
 336                                         &distance_ref_points_depth);
 337
 338         if (!distance_ref_points) {
 339                 dbg("NUMA: ibm,associativity-reference-points not found.\n");
 340                 goto err;
 341         }
 342
 343         distance_ref_points_depth /= sizeof(int);
 344
 345         if (firmware_has_feature(FW_FEATURE_OPAL) ||
 346             firmware_has_feature(FW_FEATURE_TYPE1_AFFINITY)) {
 347                 dbg("Using form 1 affinity\n");
 348                 form1_affinity = 1;
 349         }
 350
 351         if (form1_affinity) {
 352                 depth = distance_ref_points[0];
 353         } else {
 354                 if (distance_ref_points_depth < 2) {
 355                         printk(KERN_WARNING "NUMA: "
 356                                 "short ibm,associativity-reference-points\n");
 357                         goto err;
 358                 }
 359
 360                 depth = distance_ref_points[1];
 361         }
 362
 363         /*
 364          * Warn and cap if the hardware supports more than
 365          * MAX_DISTANCE_REF_POINTS domains.
 366          */
 367         if (distance_ref_points_depth > MAX_DISTANCE_REF_POINTS) {
 368                 printk(KERN_WARNING "NUMA: distance array capped at "
 369                         "%d entries\n", MAX_DISTANCE_REF_POINTS);
 370                 distance_ref_points_depth = MAX_DISTANCE_REF_POINTS;
 371         }
 372
 373         of_node_put(root);
 374         return depth;
 375
 376 err:
 377         of_node_put(root);
 378         return -1;
 379 }
 380
 381 static void __init get_n_mem_cells(int *n_addr_cells, int *n_size_cells)
 382 {
 383         struct device_node *memory = NULL;
 384
 385         memory = of_find_node_by_type(memory, "memory");
 386         if (!memory)
 387                 panic("numa.c: No memory nodes found!");
 388
 389         *n_addr_cells = of_n_addr_cells(memory);
 390         *n_size_cells = of_n_size_cells(memory);
 391         of_node_put(memory);
 392 }
 393
 394 static unsigned long read_n_cells(int n, const unsigned int **buf)
 395 {
 396         unsigned long result = 0;
 397
 398         while (n--) {
 399                 result = (result << 32) | **buf;
 400                 (*buf)++;
 401         }
 402         return result;
 403 }
 404
 405 /*
 406  * Read the next memblock list entry from the ibm,dynamic-memory property
 407  * and return the information in the provided of_drconf_cell structure.
 408  */
 409 static void read_drconf_cell(struct of_drconf_cell *drmem, const u32 **cellp)
 410 {
 411         const u32 *cp;
 412
 413         drmem->base_addr = read_n_cells(n_mem_addr_cells, cellp);
 414
 415         cp = *cellp;
 416         drmem->drc_index = cp[0];
 417         drmem->reserved = cp[1];
 418         drmem->aa_index = cp[2];
 419         drmem->flags = cp[3];
 420
 421         *cellp = cp + 4;
 422 }
 423
 424 /*
 425  * Retrieve and validate the ibm,dynamic-memory property of the device tree.
 426  *
 427  * The layout of the ibm,dynamic-memory property is a number N of memblock
 428  * list entries followed by N memblock list entries.  Each memblock list entry
 429  * contains information as laid out in the of_drconf_cell struct above.
 430  */
 431 static int of_get_drconf_memory(struct device_node *memory, const u32 **dm)
 432 {
 433         const u32 *prop;
 434         u32 len, entries;
 435
 436         prop = of_get_property(memory, "ibm,dynamic-memory", &len);
 437         if (!prop || len < sizeof(unsigned int))
 438                 return 0;
 439
 440         entries = *prop++;
 441
 442         /* Now that we know the number of entries, revalidate the size
 443          * of the property read in to ensure we have everything
 444          */
 445         if (len < (entries * (n_mem_addr_cells + 4) + 1) * sizeof(unsigned int))
 446                 return 0;
 447
 448         *dm = prop;
 449         return entries;
 450 }
 451
 452 /*
 453  * Retrieve and validate the ibm,lmb-size property for drconf memory
 454  * from the device tree.
 455  */
 456 static u64 of_get_lmb_size(struct device_node *memory)
 457 {
 458         const u32 *prop;
 459         u32 len;
 460
 461         prop = of_get_property(memory, "ibm,lmb-size", &len);
 462         if (!prop || len < sizeof(unsigned int))
 463                 return 0;
 464
 465         return read_n_cells(n_mem_size_cells, &prop);
 466 }
 467
 468 struct assoc_arrays {
 469         u32     n_arrays;
 470         u32     array_sz;
 471         const u32 *arrays;
 472 };
 473
 474 /*
 475  * Retrieve and validate the list of associativity arrays for drconf
 476  * memory from the ibm,associativity-lookup-arrays property of the
 477  * device tree..
 478  *
 479  * The layout of the ibm,associativity-lookup-arrays property is a number N
 480  * indicating the number of associativity arrays, followed by a number M
 481  * indicating the size of each associativity array, followed by a list
 482  * of N associativity arrays.
 483  */
 484 static int of_get_assoc_arrays(struct device_node *memory,
 485                                struct assoc_arrays *aa)
 486 {
 487         const u32 *prop;
 488         u32 len;
 489
 490         prop = of_get_property(memory, "ibm,associativity-lookup-arrays", &len);
 491         if (!prop || len < 2 * sizeof(unsigned int))
 492                 return -1;
 493
 494         aa->n_arrays = *prop++;
 495         aa->array_sz = *prop++;
 496
 497         /* Now that we know the number of arrays and size of each array,
 498          * revalidate the size of the property read in.
 499          */
 500         if (len < (aa->n_arrays * aa->array_sz + 2) * sizeof(unsigned int))
 501                 return -1;
 502
 503         aa->arrays = prop;
 504         return 0;
 505 }
 506
 507 /*
 508  * This is like of_node_to_nid_single() for memory represented in the
 509  * ibm,dynamic-reconfiguration-memory node.
 510  */
 511 static int of_drconf_to_nid_single(struct of_drconf_cell *drmem,
 512                                    struct assoc_arrays *aa)
 513 {
 514         int default_nid = 0;
 515         int nid = default_nid;
 516         int index;
 517
 518         if (min_common_depth > 0 && min_common_depth <= aa->array_sz &&
 519             !(drmem->flags & DRCONF_MEM_AI_INVALID) &&
 520             drmem->aa_index < aa->n_arrays) {
 521                 index = drmem->aa_index * aa->array_sz + min_common_depth - 1;
 522                 nid = aa->arrays[index];
 523
 524                 if (nid == 0xffff || nid >= MAX_NUMNODES)
 525                         nid = default_nid;
 526         }
 527
 528         return nid;
 529 }
 530
 531 /*
 532  * Figure out to which domain a cpu belongs and stick it there.
 533  * Return the id of the domain used.
 534  */
 535 static int __cpuinit numa_setup_cpu(unsigned long lcpu)
 536 {
 537         int nid;
 538         struct device_node *cpu;
 539
 540         /*
 541          * If a valid cpu-to-node mapping is already available, use it
 542          * directly instead of querying the firmware, since it represents
 543          * the most recent mapping notified to us by the platform (eg: VPHN).
 544          */
 545         if ((nid = numa_cpu_lookup_table[lcpu]) >= 0) {
 546                 map_cpu_to_node(lcpu, nid);
 547                 return nid;
 548         }
 549
 550         cpu = of_get_cpu_node(lcpu, NULL);
 551
 552         if (!cpu) {
 553                 WARN_ON(1);
 554                 nid = 0;
 555                 goto out;
 556         }
 557
 558         nid = of_node_to_nid_single(cpu);
 559
 560         if (nid < 0 || !node_online(nid))
 561                 nid = first_online_node;
 562 out:
 563         map_cpu_to_node(lcpu, nid);
 564
 565         of_node_put(cpu);
 566
 567         return nid;
 568 }
 569
 570 static int __cpuinit cpu_numa_callback(struct notifier_block *nfb,
 571                              unsigned long action,
 572                              void *hcpu)
 573 {
 574         unsigned long lcpu = (unsigned long)hcpu;
 575         int ret = NOTIFY_DONE;
 576
 577         switch (action) {
 578         case CPU_UP_PREPARE:
 579         case CPU_UP_PREPARE_FROZEN:
 580                 numa_setup_cpu(lcpu);
 581                 ret = NOTIFY_OK;
 582                 break;
 583 #ifdef CONFIG_HOTPLUG_CPU
 584         case CPU_DEAD:
 585         case CPU_DEAD_FROZEN:
 586         case CPU_UP_CANCELED:
 587         case CPU_UP_CANCELED_FROZEN:
 588                 unmap_cpu_from_node(lcpu);
 589                 break;
 590                 ret = NOTIFY_OK;
 591 #endif
 592         }
 593         return ret;
 594 }
 595
 596 /*
 597  * Check and possibly modify a memory region to enforce the memory limit.
 598  *
 599  * Returns the size the region should have to enforce the memory limit.
 600  * This will either be the original value of size, a truncated value,
 601  * or zero. If the returned value of size is 0 the region should be
 602  * discarded as it lies wholly above the memory limit.
 603  */
 604 static unsigned long __init numa_enforce_memory_limit(unsigned long start,
 605                                                       unsigned long size)
 606 {
 607         /*
 608          * We use memblock_end_of_DRAM() in here instead of memory_limit because
 609          * we've already adjusted it for the limit and it takes care of
 610          * having memory holes below the limit.  Also, in the case of
 611          * iommu_is_off, memory_limit is not set but is implicitly enforced.
 612          */
 613
 614         if (start + size <= memblock_end_of_DRAM())
 615                 return size;
 616
 617         if (start >= memblock_end_of_DRAM())
 618                 return 0;
 619
 620         return memblock_end_of_DRAM() - start;
 621 }
 622
 623 /*
 624  * Reads the counter for a given entry in
 625  * linux,drconf-usable-memory property
 626  */
 627 static inline int __init read_usm_ranges(const u32 **usm)
 628 {
 629         /*
 630          * For each lmb in ibm,dynamic-memory a corresponding
 631          * entry in linux,drconf-usable-memory property contains
 632          * a counter followed by that many (base, size) duple.
 633          * read the counter from linux,drconf-usable-memory
 634          */
 635         return read_n_cells(n_mem_size_cells, usm);
 636 }
 637
 638 /*
 639  * Extract NUMA information from the ibm,dynamic-reconfiguration-memory
 640  * node.  This assumes n_mem_{addr,size}_cells have been set.
 641  */
 642 static void __init parse_drconf_memory(struct device_node *memory)
 643 {
 644         const u32 *uninitialized_var(dm), *usm;
 645         unsigned int n, rc, ranges, is_kexec_kdump = 0;
 646         unsigned long lmb_size, base, size, sz;
 647         int nid;
 648         struct assoc_arrays aa = { .arrays = NULL };
 649
 650         n = of_get_drconf_memory(memory, &dm);
 651         if (!n)
 652                 return;
 653
 654         lmb_size = of_get_lmb_size(memory);
 655         if (!lmb_size)
 656                 return;
 657
 658         rc = of_get_assoc_arrays(memory, &aa);
 659         if (rc)
 660                 return;
 661
 662         /* check if this is a kexec/kdump kernel */
 663         usm = of_get_usable_memory(memory);
 664         if (usm != NULL)
 665                 is_kexec_kdump = 1;
 666
 667         for (; n != 0; --n) {
 668                 struct of_drconf_cell drmem;
 669
 670                 read_drconf_cell(&drmem, &dm);
 671
 672                 /* skip this block if the reserved bit is set in flags (0x80)
 673                    or if the block is not assigned to this partition (0x8) */
 674                 if ((drmem.flags & DRCONF_MEM_RESERVED)
 675                     || !(drmem.flags & DRCONF_MEM_ASSIGNED))
 676                         continue;
 677
 678                 base = drmem.base_addr;
 679                 size = lmb_size;
 680                 ranges = 1;
 681
 682                 if (is_kexec_kdump) {
 683                         ranges = read_usm_ranges(&usm);
 684                         if (!ranges) /* there are no (base, size) duple */
 685                                 continue;
 686                 }
 687                 do {
 688                         if (is_kexec_kdump) {
 689                                 base = read_n_cells(n_mem_addr_cells, &usm);
 690                                 size = read_n_cells(n_mem_size_cells, &usm);
 691                         }
 692                         nid = of_drconf_to_nid_single(&drmem, &aa);
 693                         fake_numa_create_new_node(
 694                                 ((base + size) >> PAGE_SHIFT),
 695                                            &nid);
 696                         node_set_online(nid);
 697                         sz = numa_enforce_memory_limit(base, size);
 698                         if (sz)
 699                                 memblock_set_node(base, sz, nid);
 700                 } while (--ranges);
 701         }
 702 }
 703
 704 static int __init parse_numa_properties(void)
 705 {
 706         struct device_node *memory;
 707         int default_nid = 0;
 708         unsigned long i;
 709
 710         if (numa_enabled == 0) {
 711                 printk(KERN_WARNING "NUMA disabled by user\n");
 712                 return -1;
 713         }
 714
 715         min_common_depth = find_min_common_depth();
 716
 717         if (min_common_depth < 0)
 718                 return min_common_depth;
 719
 720         dbg("NUMA associativity depth for CPU/Memory: %d\n", min_common_depth);
 721
 722         /*
 723          * Even though we connect cpus to numa domains later in SMP
 724          * init, we need to know the node ids now. This is because
 725          * each node to be onlined must have NODE_DATA etc backing it.
 726          */
 727         for_each_present_cpu(i) {
 728                 struct device_node *cpu;
 729                 int nid;
 730
 731                 cpu = of_get_cpu_node(i, NULL);
 732                 BUG_ON(!cpu);
 733                 nid = of_node_to_nid_single(cpu);
 734                 of_node_put(cpu);
 735
 736                 /*
 737                  * Don't fall back to default_nid yet -- we will plug
 738                  * cpus into nodes once the memory scan has discovered
 739                  * the topology.
 740                  */
 741                 if (nid < 0)
 742                         continue;
 743                 node_set_online(nid);
 744         }
 745
 746         get_n_mem_cells(&n_mem_addr_cells, &n_mem_size_cells);
 747
 748         for_each_node_by_type(memory, "memory") {
 749                 unsigned long start;
 750                 unsigned long size;
 751                 int nid;
 752                 int ranges;
 753                 const unsigned int *memcell_buf;
 754                 unsigned int len;
 755
 756                 memcell_buf = of_get_property(memory,
 757                         "linux,usable-memory", &len);
 758                 if (!memcell_buf || len <= 0)
 759                         memcell_buf = of_get_property(memory, "reg", &len);
 760                 if (!memcell_buf || len <= 0)
 761                         continue;
 762
 763                 /* ranges in cell */
 764                 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
 765 new_range:
 766                 /* these are order-sensitive, and modify the buffer pointer */
 767                 start = read_n_cells(n_mem_addr_cells, &memcell_buf);
 768                 size = read_n_cells(n_mem_size_cells, &memcell_buf);
 769
 770                 /*
 771                  * Assumption: either all memory nodes or none will
 772                  * have associativity properties.  If none, then
 773                  * everything goes to default_nid.
 774                  */
 775                 nid = of_node_to_nid_single(memory);
 776                 if (nid < 0)
 777                         nid = default_nid;
 778
 779                 fake_numa_create_new_node(((start + size) >> PAGE_SHIFT), &nid);
 780                 node_set_online(nid);
 781
 782                 if (!(size = numa_enforce_memory_limit(start, size))) {
 783                         if (--ranges)
 784                                 goto new_range;
 785                         else
 786                                 continue;
 787                 }
 788
 789                 memblock_set_node(start, size, nid);
 790
 791                 if (--ranges)
 792                         goto new_range;
 793         }
 794
 795         /*
 796          * Now do the same thing for each MEMBLOCK listed in the
 797          * ibm,dynamic-memory property in the
 798          * ibm,dynamic-reconfiguration-memory node.
 799          */
 800         memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
 801         if (memory)
 802                 parse_drconf_memory(memory);
 803
 804         return 0;
 805 }
 806
 807 static void __init setup_nonnuma(void)
 808 {
 809         unsigned long top_of_ram = memblock_end_of_DRAM();
 810         unsigned long total_ram = memblock_phys_mem_size();
 811         unsigned long start_pfn, end_pfn;
 812         unsigned int nid = 0;
 813         struct memblock_region *reg;
 814
 815         printk(KERN_DEBUG "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
 816                top_of_ram, total_ram);
 817         printk(KERN_DEBUG "Memory hole size: %ldMB\n",
 818                (top_of_ram - total_ram) >> 20);
 819
 820         for_each_memblock(memory, reg) {
 821                 start_pfn = memblock_region_memory_base_pfn(reg);
 822                 end_pfn = memblock_region_memory_end_pfn(reg);
 823
 824                 fake_numa_create_new_node(end_pfn, &nid);
 825                 memblock_set_node(PFN_PHYS(start_pfn),
 826                                   PFN_PHYS(end_pfn - start_pfn), nid);
 827                 node_set_online(nid);
 828         }
 829 }
 830
 831 void __init dump_numa_cpu_topology(void)
 832 {
 833         unsigned int node;
 834         unsigned int cpu, count;
 835
 836         if (min_common_depth == -1 || !numa_enabled)
 837                 return;
 838
 839         for_each_online_node(node) {
 840                 printk(KERN_DEBUG "Node %d CPUs:", node);
 841
 842                 count = 0;
 843                 /*
 844                  * If we used a CPU iterator here we would miss printing
 845                  * the holes in the cpumap.
 846                  */
 847                 for (cpu = 0; cpu < nr_cpu_ids; cpu++) {
 848                         if (cpumask_test_cpu(cpu,
 849                                         node_to_cpumask_map[node])) {
 850                                 if (count == 0)
 851                                         printk(" %u", cpu);
 852                                 ++count;
 853                         } else {
 854                                 if (count > 1)
 855                                         printk("-%u", cpu - 1);
 856                                 count = 0;
 857                         }
 858                 }
 859
 860                 if (count > 1)
 861                         printk("-%u", nr_cpu_ids - 1);
 862                 printk("\n");
 863         }
 864 }
 865
 866 static void __init dump_numa_memory_topology(void)
 867 {
 868         unsigned int node;
 869         unsigned int count;
 870
 871         if (min_common_depth == -1 || !numa_enabled)
 872                 return;
 873
 874         for_each_online_node(node) {
 875                 unsigned long i;
 876
 877                 printk(KERN_DEBUG "Node %d Memory:", node);
 878
 879                 count = 0;
 880
 881                 for (i = 0; i < memblock_end_of_DRAM();
 882                      i += (1 << SECTION_SIZE_BITS)) {
 883                         if (early_pfn_to_nid(i >> PAGE_SHIFT) == node) {
 884                                 if (count == 0)
 885                                         printk(" 0x%lx", i);
 886                                 ++count;
 887                         } else {
 888                                 if (count > 0)
 889                                         printk("-0x%lx", i);
 890                                 count = 0;
 891                         }
 892                 }
 893
 894                 if (count > 0)
 895                         printk("-0x%lx", i);
 896                 printk("\n");
 897         }
 898 }
 899
 900 /*
 901  * Allocate some memory, satisfying the memblock or bootmem allocator where
 902  * required. nid is the preferred node and end is the physical address of
 903  * the highest address in the node.
 904  *
 905  * Returns the virtual address of the memory.
 906  */
 907 static void __init *careful_zallocation(int nid, unsigned long size,
 908                                        unsigned long align,
 909                                        unsigned long end_pfn)
 910 {
 911         void *ret;
 912         int new_nid;
 913         unsigned long ret_paddr;
 914
 915         ret_paddr = __memblock_alloc_base(size, align, end_pfn << PAGE_SHIFT);
 916
 917         /* retry over all memory */
 918         if (!ret_paddr)
 919                 ret_paddr = __memblock_alloc_base(size, align, memblock_end_of_DRAM());
 920
 921         if (!ret_paddr)
 922                 panic("numa.c: cannot allocate %lu bytes for node %d",
 923                       size, nid);
 924
 925         ret = __va(ret_paddr);
 926
 927         /*
 928          * We initialize the nodes in numeric order: 0, 1, 2...
 929          * and hand over control from the MEMBLOCK allocator to the
 930          * bootmem allocator.  If this function is called for
 931          * node 5, then we know that all nodes <5 are using the
 932          * bootmem allocator instead of the MEMBLOCK allocator.
 933          *
 934          * So, check the nid from which this allocation came
 935          * and double check to see if we need to use bootmem
 936          * instead of the MEMBLOCK.  We don't free the MEMBLOCK memory
 937          * since it would be useless.
 938          */
 939         new_nid = early_pfn_to_nid(ret_paddr >> PAGE_SHIFT);
 940         if (new_nid < nid) {
 941                 ret = __alloc_bootmem_node(NODE_DATA(new_nid),
 942                                 size, align, 0);
 943
 944                 dbg("alloc_bootmem %p %lx\n", ret, size);
 945         }
 946
 947         memset(ret, 0, size);
 948         return ret;
 949 }
 950
 951 static struct notifier_block __cpuinitdata ppc64_numa_nb = {
 952         .notifier_call = cpu_numa_callback,
 953         .priority = 1 /* Must run before sched domains notifier. */
 954 };
 955
 956 static void __init mark_reserved_regions_for_nid(int nid)
 957 {
 958         struct pglist_data *node = NODE_DATA(nid);
 959         struct memblock_region *reg;
 960
 961         for_each_memblock(reserved, reg) {
 962                 unsigned long physbase = reg->base;
 963                 unsigned long size = reg->size;
 964                 unsigned long start_pfn = physbase >> PAGE_SHIFT;
 965                 unsigned long end_pfn = PFN_UP(physbase + size);
 966                 struct node_active_region node_ar;
 967                 unsigned long node_end_pfn = node->node_start_pfn +
 968                                              node->node_spanned_pages;
 969
 970                 /*
 971                  * Check to make sure that this memblock.reserved area is
 972                  * within the bounds of the node that we care about.
 973                  * Checking the nid of the start and end points is not
 974                  * sufficient because the reserved area could span the
 975                  * entire node.
 976                  */
 977                 if (end_pfn <= node->node_start_pfn ||
 978                     start_pfn >= node_end_pfn)
 979                         continue;
 980
 981                 get_node_active_region(start_pfn, &node_ar);
 982                 while (start_pfn < end_pfn &&
 983                         node_ar.start_pfn < node_ar.end_pfn) {
 984                         unsigned long reserve_size = size;
 985                         /*
 986                          * if reserved region extends past active region
 987                          * then trim size to active region
 988                          */
 989                         if (end_pfn > node_ar.end_pfn)
 990                                 reserve_size = (node_ar.end_pfn << PAGE_SHIFT)
 991                                         - physbase;
 992                         /*
 993                          * Only worry about *this* node, others may not
 994                          * yet have valid NODE_DATA().
 995                          */
 996                         if (node_ar.nid == nid) {
 997                                 dbg("reserve_bootmem %lx %lx nid=%d\n",
 998                                         physbase, reserve_size, node_ar.nid);
 999                                 reserve_bootmem_node(NODE_DATA(node_ar.nid),
1000                                                 physbase, reserve_size,
1001                                                 BOOTMEM_DEFAULT);
1002                         }
1003                         /*
1004                          * if reserved region is contained in the active region
1005                          * then done.
1006                          */
1007                         if (end_pfn <= node_ar.end_pfn)
1008                                 break;
1009
1010                         /*
1011                          * reserved region extends past the active region
1012                          *   get next active region that contains this
1013                          *   reserved region
1014                          */
1015                         start_pfn = node_ar.end_pfn;
1016                         physbase = start_pfn << PAGE_SHIFT;
1017                         size = size - reserve_size;
1018                         get_node_active_region(start_pfn, &node_ar);
1019                 }
1020         }
1021 }
1022
1023
1024 void __init do_init_bootmem(void)
1025 {
1026         int nid;
1027
1028         min_low_pfn = 0;
1029         max_low_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
1030         max_pfn = max_low_pfn;
1031
1032         if (parse_numa_properties())
1033                 setup_nonnuma();
1034         else
1035                 dump_numa_memory_topology();
1036
1037         for_each_online_node(nid) {
1038                 unsigned long start_pfn, end_pfn;
1039                 void *bootmem_vaddr;
1040                 unsigned long bootmap_pages;
1041
1042                 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
1043
1044                 /*
1045                  * Allocate the node structure node local if possible
1046                  *
1047                  * Be careful moving this around, as it relies on all
1048                  * previous nodes' bootmem to be initialized and have
1049                  * all reserved areas marked.
1050                  */
1051                 NODE_DATA(nid) = careful_zallocation(nid,
1052                                         sizeof(struct pglist_data),
1053                                         SMP_CACHE_BYTES, end_pfn);
1054
1055                 dbg("node %d\n", nid);
1056                 dbg("NODE_DATA() = %p\n", NODE_DATA(nid));
1057
1058                 NODE_DATA(nid)->bdata = &bootmem_node_data[nid];
1059                 NODE_DATA(nid)->node_start_pfn = start_pfn;
1060                 NODE_DATA(nid)->node_spanned_pages = end_pfn - start_pfn;
1061
1062                 if (NODE_DATA(nid)->node_spanned_pages == 0)
1063                         continue;
1064
1065                 dbg("start_paddr = %lx\n", start_pfn << PAGE_SHIFT);
1066                 dbg("end_paddr = %lx\n", end_pfn << PAGE_SHIFT);
1067
1068                 bootmap_pages = bootmem_bootmap_pages(end_pfn - start_pfn);
1069                 bootmem_vaddr = careful_zallocation(nid,
1070                                         bootmap_pages << PAGE_SHIFT,
1071                                         PAGE_SIZE, end_pfn);
1072
1073                 dbg("bootmap_vaddr = %p\n", bootmem_vaddr);
1074
1075                 init_bootmem_node(NODE_DATA(nid),
1076                                   __pa(bootmem_vaddr) >> PAGE_SHIFT,
1077                                   start_pfn, end_pfn);
1078
1079                 free_bootmem_with_active_regions(nid, end_pfn);
1080                 /*
1081                  * Be very careful about moving this around.  Future
1082                  * calls to careful_zallocation() depend on this getting
1083                  * done correctly.
1084                  */
1085                 mark_reserved_regions_for_nid(nid);
1086                 sparse_memory_present_with_active_regions(nid);
1087         }
1088
1089         init_bootmem_done = 1;
1090
1091         /*
1092          * Now bootmem is initialised we can create the node to cpumask
1093          * lookup tables and setup the cpu callback to populate them.
1094          */
1095         setup_node_to_cpumask_map();
1096
1097         reset_numa_cpu_lookup_table();
1098         register_cpu_notifier(&ppc64_numa_nb);
1099         cpu_numa_callback(&ppc64_numa_nb, CPU_UP_PREPARE,
1100                           (void *)(unsigned long)boot_cpuid);
1101 }
1102
1103 void __init paging_init(void)
1104 {
1105         unsigned long max_zone_pfns[MAX_NR_ZONES];
1106         memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
1107         max_zone_pfns[ZONE_DMA] = memblock_end_of_DRAM() >> PAGE_SHIFT;
1108         free_area_init_nodes(max_zone_pfns);
1109 }
1110
1111 static int __init early_numa(char *p)
1112 {
1113         if (!p)
1114                 return 0;
1115
1116         if (strstr(p, "off"))
1117                 numa_enabled = 0;
1118
1119         if (strstr(p, "debug"))
1120                 numa_debug = 1;
1121
1122         p = strstr(p, "fake=");
1123         if (p)
1124                 cmdline = p + strlen("fake=");
1125
1126         return 0;
1127 }
1128 early_param("numa", early_numa);
1129
1130 #ifdef CONFIG_MEMORY_HOTPLUG
1131 /*
1132  * Find the node associated with a hot added memory section for
1133  * memory represented in the device tree by the property
1134  * ibm,dynamic-reconfiguration-memory/ibm,dynamic-memory.
1135  */
1136 static int hot_add_drconf_scn_to_nid(struct device_node *memory,
1137                                      unsigned long scn_addr)
1138 {
1139         const u32 *dm;
1140         unsigned int drconf_cell_cnt, rc;
1141         unsigned long lmb_size;
1142         struct assoc_arrays aa;
1143         int nid = -1;
1144
1145         drconf_cell_cnt = of_get_drconf_memory(memory, &dm);
1146         if (!drconf_cell_cnt)
1147                 return -1;
1148
1149         lmb_size = of_get_lmb_size(memory);
1150         if (!lmb_size)
1151                 return -1;
1152
1153         rc = of_get_assoc_arrays(memory, &aa);
1154         if (rc)
1155                 return -1;
1156
1157         for (; drconf_cell_cnt != 0; --drconf_cell_cnt) {
1158                 struct of_drconf_cell drmem;
1159
1160                 read_drconf_cell(&drmem, &dm);
1161
1162                 /* skip this block if it is reserved or not assigned to
1163                  * this partition */
1164                 if ((drmem.flags & DRCONF_MEM_RESERVED)
1165                     || !(drmem.flags & DRCONF_MEM_ASSIGNED))
1166                         continue;
1167
1168                 if ((scn_addr < drmem.base_addr)
1169                     || (scn_addr >= (drmem.base_addr + lmb_size)))
1170                         continue;
1171
1172                 nid = of_drconf_to_nid_single(&drmem, &aa);
1173                 break;
1174         }
1175
1176         return nid;
1177 }
1178
1179 /*
1180  * Find the node associated with a hot added memory section for memory
1181  * represented in the device tree as a node (i.e. memory@XXXX) for
1182  * each memblock.
1183  */
1184 int hot_add_node_scn_to_nid(unsigned long scn_addr)
1185 {
1186         struct device_node *memory;
1187         int nid = -1;
1188
1189         for_each_node_by_type(memory, "memory") {
1190                 unsigned long start, size;
1191                 int ranges;
1192                 const unsigned int *memcell_buf;
1193                 unsigned int len;
1194
1195                 memcell_buf = of_get_property(memory, "reg", &len);
1196                 if (!memcell_buf || len <= 0)
1197                         continue;
1198
1199                 /* ranges in cell */
1200                 ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
1201
1202                 while (ranges--) {
1203                         start = read_n_cells(n_mem_addr_cells, &memcell_buf);
1204                         size = read_n_cells(n_mem_size_cells, &memcell_buf);
1205
1206                         if ((scn_addr < start) || (scn_addr >= (start + size)))
1207                                 continue;
1208
1209                         nid = of_node_to_nid_single(memory);
1210                         break;
1211                 }
1212
1213                 if (nid >= 0)
1214                         break;
1215         }
1216
1217         of_node_put(memory);
1218
1219         return nid;
1220 }
1221
1222 /*
1223  * Find the node associated with a hot added memory section.  Section
1224  * corresponds to a SPARSEMEM section, not an MEMBLOCK.  It is assumed that
1225  * sections are fully contained within a single MEMBLOCK.
1226  */
1227 int hot_add_scn_to_nid(unsigned long scn_addr)
1228 {
1229         struct device_node *memory = NULL;
1230         int nid, found = 0;
1231
1232         if (!numa_enabled || (min_common_depth < 0))
1233                 return first_online_node;
1234
1235         memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1236         if (memory) {
1237                 nid = hot_add_drconf_scn_to_nid(memory, scn_addr);
1238                 of_node_put(memory);
1239         } else {
1240                 nid = hot_add_node_scn_to_nid(scn_addr);
1241         }
1242
1243         if (nid < 0 || !node_online(nid))
1244                 nid = first_online_node;
1245
1246         if (NODE_DATA(nid)->node_spanned_pages)
1247                 return nid;
1248
1249         for_each_online_node(nid) {
1250                 if (NODE_DATA(nid)->node_spanned_pages) {
1251                         found = 1;
1252                         break;
1253                 }
1254         }
1255
1256         BUG_ON(!found);
1257         return nid;
1258 }
1259
1260 static u64 hot_add_drconf_memory_max(void)
1261 {
1262         struct device_node *memory = NULL;
1263         unsigned int drconf_cell_cnt = 0;
1264         u64 lmb_size = 0;
1265         const u32 *dm = 0;
1266
1267         memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1268         if (memory) {
1269                 drconf_cell_cnt = of_get_drconf_memory(memory, &dm);
1270                 lmb_size = of_get_lmb_size(memory);
1271                 of_node_put(memory);
1272         }
1273         return lmb_size * drconf_cell_cnt;
1274 }
1275
1276 /*
1277  * memory_hotplug_max - return max address of memory that may be added
1278  *
1279  * This is currently only used on systems that support drconfig memory
1280  * hotplug.
1281  */
1282 u64 memory_hotplug_max(void)
1283 {
1284         return max(hot_add_drconf_memory_max(), memblock_end_of_DRAM());
1285 }
1286 #endif /* CONFIG_MEMORY_HOTPLUG */
1287
1288 /* Virtual Processor Home Node (VPHN) support */
1289 #ifdef CONFIG_PPC_SPLPAR
1290 struct topology_update_data {
1291         struct topology_update_data *next;
1292         unsigned int cpu;
1293         int old_nid;
1294         int new_nid;
1295 };
1296
1297 static u8 vphn_cpu_change_counts[NR_CPUS][MAX_DISTANCE_REF_POINTS];
1298 static cpumask_t cpu_associativity_changes_mask;
1299 static int vphn_enabled;
1300 static int prrn_enabled;
1301 static void reset_topology_timer(void);
1302
1303 /*
1304  * Store the current values of the associativity change counters in the
1305  * hypervisor.
1306  */
1307 static void setup_cpu_associativity_change_counters(void)
1308 {
1309         int cpu;
1310
1311         /* The VPHN feature supports a maximum of 8 reference points */
1312         BUILD_BUG_ON(MAX_DISTANCE_REF_POINTS > 8);
1313
1314         for_each_possible_cpu(cpu) {
1315                 int i;
1316                 u8 *counts = vphn_cpu_change_counts[cpu];
1317                 volatile u8 *hypervisor_counts = lppaca[cpu].vphn_assoc_counts;
1318
1319                 for (i = 0; i < distance_ref_points_depth; i++)
1320                         counts[i] = hypervisor_counts[i];
1321         }
1322 }
1323
1324 /*
1325  * The hypervisor maintains a set of 8 associativity change counters in
1326  * the VPA of each cpu that correspond to the associativity levels in the
1327  * ibm,associativity-reference-points property. When an associativity
1328  * level changes, the corresponding counter is incremented.
1329  *
1330  * Set a bit in cpu_associativity_changes_mask for each cpu whose home
1331  * node associativity levels have changed.
1332  *
1333  * Returns the number of cpus with unhandled associativity changes.
1334  */
1335 static int update_cpu_associativity_changes_mask(void)
1336 {
1337         int cpu;
1338         cpumask_t *changes = &cpu_associativity_changes_mask;
1339
1340         for_each_possible_cpu(cpu) {
1341                 int i, changed = 0;
1342                 u8 *counts = vphn_cpu_change_counts[cpu];
1343                 volatile u8 *hypervisor_counts = lppaca[cpu].vphn_assoc_counts;
1344
1345                 for (i = 0; i < distance_ref_points_depth; i++) {
1346                         if (hypervisor_counts[i] != counts[i]) {
1347                                 counts[i] = hypervisor_counts[i];
1348                                 changed = 1;
1349                         }
1350                 }
1351                 if (changed) {
1352                         cpumask_or(changes, changes, cpu_sibling_mask(cpu));
1353                         cpu = cpu_last_thread_sibling(cpu);
1354                 }
1355         }
1356
1357         return cpumask_weight(changes);
1358 }
1359
1360 /*
1361  * 6 64-bit registers unpacked into 12 32-bit associativity values. To form
1362  * the complete property we have to add the length in the first cell.
1363  */
1364 #define VPHN_ASSOC_BUFSIZE (6*sizeof(u64)/sizeof(u32) + 1)
1365
1366 /*
1367  * Convert the associativity domain numbers returned from the hypervisor
1368  * to the sequence they would appear in the ibm,associativity property.
1369  */
1370 static int vphn_unpack_associativity(const long *packed, unsigned int *unpacked)
1371 {
1372         int i, nr_assoc_doms = 0;
1373         const u16 *field = (const u16*) packed;
1374
1375 #define VPHN_FIELD_UNUSED       (0xffff)
1376 #define VPHN_FIELD_MSB          (0x8000)
1377 #define VPHN_FIELD_MASK         (~VPHN_FIELD_MSB)
1378
1379         for (i = 1; i < VPHN_ASSOC_BUFSIZE; i++) {
1380                 if (*field == VPHN_FIELD_UNUSED) {
1381                         /* All significant fields processed, and remaining
1382                          * fields contain the reserved value of all 1's.
1383                          * Just store them.
1384                          */
1385                         unpacked[i] = *((u32*)field);
1386                         field += 2;
1387                 } else if (*field & VPHN_FIELD_MSB) {
1388                         /* Data is in the lower 15 bits of this field */
1389                         unpacked[i] = *field & VPHN_FIELD_MASK;
1390                         field++;
1391                         nr_assoc_doms++;
1392                 } else {
1393                         /* Data is in the lower 15 bits of this field
1394                          * concatenated with the next 16 bit field
1395                          */
1396                         unpacked[i] = *((u32*)field);
1397                         field += 2;
1398                         nr_assoc_doms++;
1399                 }
1400         }
1401
1402         /* The first cell contains the length of the property */
1403         unpacked[0] = nr_assoc_doms;
1404
1405         return nr_assoc_doms;
1406 }
1407
1408 /*
1409  * Retrieve the new associativity information for a virtual processor's
1410  * home node.
1411  */
1412 static long hcall_vphn(unsigned long cpu, unsigned int *associativity)
1413 {
1414         long rc;
1415         long retbuf[PLPAR_HCALL9_BUFSIZE] = {0};
1416         u64 flags = 1;
1417         int hwcpu = get_hard_smp_processor_id(cpu);
1418
1419         rc = plpar_hcall9(H_HOME_NODE_ASSOCIATIVITY, retbuf, flags, hwcpu);
1420         vphn_unpack_associativity(retbuf, associativity);
1421
1422         return rc;
1423 }
1424
1425 static long vphn_get_associativity(unsigned long cpu,
1426                                         unsigned int *associativity)
1427 {
1428         long rc;
1429
1430         rc = hcall_vphn(cpu, associativity);
1431
1432         switch (rc) {
1433         case H_FUNCTION:
1434                 printk(KERN_INFO
1435                         "VPHN is not supported. Disabling polling...\n");
1436                 stop_topology_update();
1437                 break;
1438         case H_HARDWARE:
1439                 printk(KERN_ERR
1440                         "hcall_vphn() experienced a hardware fault "
1441                         "preventing VPHN. Disabling polling...\n");
1442                 stop_topology_update();
1443         }
1444
1445         return rc;
1446 }
1447
1448 /*
1449  * Update the CPU maps and sysfs entries for a single CPU when its NUMA
1450  * characteristics change. This function doesn't perform any locking and is
1451  * only safe to call from stop_machine().
1452  */
1453 static int update_cpu_topology(void *data)
1454 {
1455         struct topology_update_data *update;
1456         unsigned long cpu;
1457
1458         if (!data)
1459                 return -EINVAL;
1460
1461         cpu = smp_processor_id();
1462
1463         for (update = data; update; update = update->next) {
1464                 if (cpu != update->cpu)
1465                         continue;
1466
1467                 unmap_cpu_from_node(update->cpu);
1468                 map_cpu_to_node(update->cpu, update->new_nid);
1469                 vdso_getcpu_init();
1470         }
1471
1472         return 0;
1473 }
1474
1475 static int update_lookup_table(void *data)
1476 {
1477         struct topology_update_data *update;
1478
1479         if (!data)
1480                 return -EINVAL;
1481
1482         /*
1483          * Upon topology update, the numa-cpu lookup table needs to be updated
1484          * for all threads in the core, including offline CPUs, to ensure that
1485          * future hotplug operations respect the cpu-to-node associativity
1486          * properly.
1487          */
1488         for (update = data; update; update = update->next) {
1489                 int nid, base, j;
1490
1491                 nid = update->new_nid;
1492                 base = cpu_first_thread_sibling(update->cpu);
1493
1494                 for (j = 0; j < threads_per_core; j++) {
1495                         update_numa_cpu_lookup_table(base + j, nid);
1496                 }
1497         }
1498
1499         return 0;
1500 }
1501
1502 /*
1503  * Update the node maps and sysfs entries for each cpu whose home node
1504  * has changed. Returns 1 when the topology has changed, and 0 otherwise.
1505  */
1506 int arch_update_cpu_topology(void)
1507 {
1508         unsigned int cpu, sibling, changed = 0;
1509         struct topology_update_data *updates, *ud;
1510         unsigned int associativity[VPHN_ASSOC_BUFSIZE] = {0};
1511         cpumask_t updated_cpus;
1512         struct device *dev;
1513         int weight, new_nid, i = 0;
1514
1515         weight = cpumask_weight(&cpu_associativity_changes_mask);
1516         if (!weight)
1517                 return 0;
1518
1519         updates = kzalloc(weight * (sizeof(*updates)), GFP_KERNEL);
1520         if (!updates)
1521                 return 0;
1522
1523         cpumask_clear(&updated_cpus);
1524
1525         for_each_cpu(cpu, &cpu_associativity_changes_mask) {
1526                 /*
1527                  * If siblings aren't flagged for changes, updates list
1528                  * will be too short. Skip on this update and set for next
1529                  * update.
1530                  */
1531                 if (!cpumask_subset(cpu_sibling_mask(cpu),
1532                                         &cpu_associativity_changes_mask)) {
1533                         pr_info("Sibling bits not set for associativity "
1534                                         "change, cpu%d\n", cpu);
1535                         cpumask_or(&cpu_associativity_changes_mask,
1536                                         &cpu_associativity_changes_mask,
1537                                         cpu_sibling_mask(cpu));
1538                         cpu = cpu_last_thread_sibling(cpu);
1539                         continue;
1540                 }
1541
1542                 /* Use associativity from first thread for all siblings */
1543                 vphn_get_associativity(cpu, associativity);
1544                 new_nid = associativity_to_nid(associativity);
1545                 if (new_nid < 0 || !node_online(new_nid))
1546                         new_nid = first_online_node;
1547
1548                 if (new_nid == numa_cpu_lookup_table[cpu]) {
1549                         cpumask_andnot(&cpu_associativity_changes_mask,
1550                                         &cpu_associativity_changes_mask,
1551                                         cpu_sibling_mask(cpu));
1552                         cpu = cpu_last_thread_sibling(cpu);
1553                         continue;
1554                 }
1555
1556                 for_each_cpu(sibling, cpu_sibling_mask(cpu)) {
1557                         ud = &updates[i++];
1558                         ud->cpu = sibling;
1559                         ud->new_nid = new_nid;
1560                         ud->old_nid = numa_cpu_lookup_table[sibling];
1561                         cpumask_set_cpu(sibling, &updated_cpus);
1562                         if (i < weight)
1563                                 ud->next = &updates[i];
1564                 }
1565                 cpu = cpu_last_thread_sibling(cpu);
1566         }
1567
1568         stop_machine(update_cpu_topology, &updates[0], &updated_cpus);
1569
1570         /*
1571          * Update the numa-cpu lookup table with the new mappings, even for
1572          * offline CPUs. It is best to perform this update from the stop-
1573          * machine context.
1574          */
1575         stop_machine(update_lookup_table, &updates[0],
1576                                         cpumask_of(raw_smp_processor_id()));
1577
1578         for (ud = &updates[0]; ud; ud = ud->next) {
1579                 unregister_cpu_under_node(ud->cpu, ud->old_nid);
1580                 register_cpu_under_node(ud->cpu, ud->new_nid);
1581
1582                 dev = get_cpu_device(ud->cpu);
1583                 if (dev)
1584                         kobject_uevent(&dev->kobj, KOBJ_CHANGE);
1585                 cpumask_clear_cpu(ud->cpu, &cpu_associativity_changes_mask);
1586                 changed = 1;
1587         }
1588
1589         kfree(updates);
1590         return changed;
1591 }
1592
1593 static void topology_work_fn(struct work_struct *work)
1594 {
1595         rebuild_sched_domains();
1596 }
1597 static DECLARE_WORK(topology_work, topology_work_fn);
1598
1599 void topology_schedule_update(void)
1600 {
1601         schedule_work(&topology_work);
1602 }
1603
1604 static void topology_timer_fn(unsigned long ignored)
1605 {
1606         if (prrn_enabled && cpumask_weight(&cpu_associativity_changes_mask))
1607                 topology_schedule_update();
1608         else if (vphn_enabled) {
1609                 if (update_cpu_associativity_changes_mask() > 0)
1610                         topology_schedule_update();
1611                 reset_topology_timer();
1612         }
1613 }
1614 static struct timer_list topology_timer =
1615         TIMER_INITIALIZER(topology_timer_fn, 0, 0);
1616
1617 static void reset_topology_timer(void)
1618 {
1619         topology_timer.data = 0;
1620         topology_timer.expires = jiffies + 60 * HZ;
1621         mod_timer(&topology_timer, topology_timer.expires);
1622 }
1623
1624 #ifdef CONFIG_SMP
1625
1626 static void stage_topology_update(int core_id)
1627 {
1628         cpumask_or(&cpu_associativity_changes_mask,
1629                 &cpu_associativity_changes_mask, cpu_sibling_mask(core_id));
1630         reset_topology_timer();
1631 }
1632
1633 static int dt_update_callback(struct notifier_block *nb,
1634                                 unsigned long action, void *data)
1635 {
1636         struct of_prop_reconfig *update;
1637         int rc = NOTIFY_DONE;
1638
1639         switch (action) {
1640         case OF_RECONFIG_UPDATE_PROPERTY:
1641                 update = (struct of_prop_reconfig *)data;
1642                 if (!of_prop_cmp(update->dn->type, "cpu") &&
1643                     !of_prop_cmp(update->prop->name, "ibm,associativity")) {
1644                         u32 core_id;
1645                         of_property_read_u32(update->dn, "reg", &core_id);
1646                         stage_topology_update(core_id);
1647                         rc = NOTIFY_OK;
1648                 }
1649                 break;
1650         }
1651
1652         return rc;
1653 }
1654
1655 static struct notifier_block dt_update_nb = {
1656         .notifier_call = dt_update_callback,
1657 };
1658
1659 #endif
1660
1661 /*
1662  * Start polling for associativity changes.
1663  */
1664 int start_topology_update(void)
1665 {
1666         int rc = 0;
1667
1668         if (firmware_has_feature(FW_FEATURE_PRRN)) {
1669                 if (!prrn_enabled) {
1670                         prrn_enabled = 1;
1671                         vphn_enabled = 0;
1672 #ifdef CONFIG_SMP
1673                         rc = of_reconfig_notifier_register(&dt_update_nb);
1674 #endif
1675                 }
1676         } else if (firmware_has_feature(FW_FEATURE_VPHN) &&
1677                    get_lppaca()->shared_proc) {
1678                 if (!vphn_enabled) {
1679                         prrn_enabled = 0;
1680                         vphn_enabled = 1;
1681                         setup_cpu_associativity_change_counters();
1682                         init_timer_deferrable(&topology_timer);
1683                         reset_topology_timer();
1684                 }
1685         }
1686
1687         return rc;
1688 }
1689
1690 /*
1691  * Disable polling for VPHN associativity changes.
1692  */
1693 int stop_topology_update(void)
1694 {
1695         int rc = 0;
1696
1697         if (prrn_enabled) {
1698                 prrn_enabled = 0;
1699 #ifdef CONFIG_SMP
1700                 rc = of_reconfig_notifier_unregister(&dt_update_nb);
1701 #endif
1702         } else if (vphn_enabled) {
1703                 vphn_enabled = 0;
1704                 rc = del_timer_sync(&topology_timer);
1705         }
1706
1707         return rc;
1708 }
1709
1710 int prrn_is_enabled(void)
1711 {
1712         return prrn_enabled;
1713 }
1714
1715 static int topology_read(struct seq_file *file, void *v)
1716 {
1717         if (vphn_enabled || prrn_enabled)
1718                 seq_puts(file, "on\n");
1719         else
1720                 seq_puts(file, "off\n");
1721
1722         return 0;
1723 }
1724
1725 static int topology_open(struct inode *inode, struct file *file)
1726 {
1727         return single_open(file, topology_read, NULL);
1728 }
1729
1730 static ssize_t topology_write(struct file *file, const char __user *buf,
1731                               size_t count, loff_t *off)
1732 {
1733         char kbuf[4]; /* "on" or "off" plus null. */
1734         int read_len;
1735
1736         read_len = count < 3 ? count : 3;
1737         if (copy_from_user(kbuf, buf, read_len))
1738                 return -EINVAL;
1739
1740         kbuf[read_len] = '\0';
1741
1742         if (!strncmp(kbuf, "on", 2))
1743                 start_topology_update();
1744         else if (!strncmp(kbuf, "off", 3))
1745                 stop_topology_update();
1746         else
1747                 return -EINVAL;
1748
1749         return count;
1750 }
1751
1752 static const struct file_operations topology_ops = {
1753         .read = seq_read,
1754         .write = topology_write,
1755         .open = topology_open,
1756         .release = single_release
1757 };
1758
1759 static int topology_update_init(void)
1760 {
1761         start_topology_update();
1762         proc_create("powerpc/topology_updates", 644, NULL, &topology_ops);
1763
1764         return 0;
1765 }
1766 device_initcall(topology_update_init);
1767 #endif /* CONFIG_PPC_SPLPAR */