arch/powerpc/mm/numa.c

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