Linux Audio

Check our new training course

Loading...
v6.8
   1// SPDX-License-Identifier: GPL-2.0-or-later
   2/*
   3 * pSeries NUMA support
   4 *
   5 * Copyright (C) 2002 Anton Blanchard <anton@au.ibm.com>, IBM
 
 
 
 
 
   6 */
   7#define pr_fmt(fmt) "numa: " fmt
   8
   9#include <linux/threads.h>
  10#include <linux/memblock.h>
  11#include <linux/init.h>
  12#include <linux/mm.h>
  13#include <linux/mmzone.h>
  14#include <linux/export.h>
  15#include <linux/nodemask.h>
  16#include <linux/cpu.h>
  17#include <linux/notifier.h>
 
  18#include <linux/of.h>
  19#include <linux/of_address.h>
  20#include <linux/pfn.h>
  21#include <linux/cpuset.h>
  22#include <linux/node.h>
  23#include <linux/stop_machine.h>
  24#include <linux/proc_fs.h>
  25#include <linux/seq_file.h>
  26#include <linux/uaccess.h>
  27#include <linux/slab.h>
  28#include <asm/cputhreads.h>
  29#include <asm/sparsemem.h>
 
 
  30#include <asm/smp.h>
  31#include <asm/topology.h>
  32#include <asm/firmware.h>
  33#include <asm/paca.h>
  34#include <asm/hvcall.h>
  35#include <asm/setup.h>
  36#include <asm/vdso.h>
  37#include <asm/vphn.h>
  38#include <asm/drmem.h>
  39
  40static int numa_enabled = 1;
  41
  42static char *cmdline __initdata;
  43
 
 
 
  44int numa_cpu_lookup_table[NR_CPUS];
  45cpumask_var_t node_to_cpumask_map[MAX_NUMNODES];
  46struct pglist_data *node_data[MAX_NUMNODES];
  47
  48EXPORT_SYMBOL(numa_cpu_lookup_table);
  49EXPORT_SYMBOL(node_to_cpumask_map);
  50EXPORT_SYMBOL(node_data);
  51
  52static int primary_domain_index;
  53static int n_mem_addr_cells, n_mem_size_cells;
  54
  55#define FORM0_AFFINITY 0
  56#define FORM1_AFFINITY 1
  57#define FORM2_AFFINITY 2
  58static int affinity_form;
  59
  60#define MAX_DISTANCE_REF_POINTS 4
  61static int distance_ref_points_depth;
  62static const __be32 *distance_ref_points;
  63static int distance_lookup_table[MAX_NUMNODES][MAX_DISTANCE_REF_POINTS];
  64static int numa_distance_table[MAX_NUMNODES][MAX_NUMNODES] = {
  65	[0 ... MAX_NUMNODES - 1] = { [0 ... MAX_NUMNODES - 1] = -1 }
  66};
  67static int numa_id_index_table[MAX_NUMNODES] = { [0 ... MAX_NUMNODES - 1] = NUMA_NO_NODE };
  68
  69/*
  70 * Allocate node_to_cpumask_map based on number of available nodes
  71 * Requires node_possible_map to be valid.
  72 *
  73 * Note: cpumask_of_node() is not valid until after this is done.
  74 */
  75static void __init setup_node_to_cpumask_map(void)
  76{
  77	unsigned int node;
  78
  79	/* setup nr_node_ids if not done yet */
  80	if (nr_node_ids == MAX_NUMNODES)
  81		setup_nr_node_ids();
 
 
 
  82
  83	/* allocate the map */
  84	for_each_node(node)
  85		alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]);
  86
  87	/* cpumask_of_node() will now work */
  88	pr_debug("Node to cpumask map for %u nodes\n", nr_node_ids);
  89}
  90
  91static int __init fake_numa_create_new_node(unsigned long end_pfn,
  92						unsigned int *nid)
  93{
  94	unsigned long long mem;
  95	char *p = cmdline;
  96	static unsigned int fake_nid;
  97	static unsigned long long curr_boundary;
  98
  99	/*
 100	 * Modify node id, iff we started creating NUMA nodes
 101	 * We want to continue from where we left of the last time
 102	 */
 103	if (fake_nid)
 104		*nid = fake_nid;
 105	/*
 106	 * In case there are no more arguments to parse, the
 107	 * node_id should be the same as the last fake node id
 108	 * (we've handled this above).
 109	 */
 110	if (!p)
 111		return 0;
 112
 113	mem = memparse(p, &p);
 114	if (!mem)
 115		return 0;
 116
 117	if (mem < curr_boundary)
 118		return 0;
 119
 120	curr_boundary = mem;
 121
 122	if ((end_pfn << PAGE_SHIFT) > mem) {
 123		/*
 124		 * Skip commas and spaces
 125		 */
 126		while (*p == ',' || *p == ' ' || *p == '\t')
 127			p++;
 128
 129		cmdline = p;
 130		fake_nid++;
 131		*nid = fake_nid;
 132		pr_debug("created new fake_node with id %d\n", fake_nid);
 133		return 1;
 134	}
 135	return 0;
 136}
 137
 138static void __init reset_numa_cpu_lookup_table(void)
 
 
 
 
 
 
 
 
 
 
 
 139{
 140	unsigned int cpu;
 
 
 
 
 
 
 
 
 141
 142	for_each_possible_cpu(cpu)
 143		numa_cpu_lookup_table[cpu] = -1;
 144}
 145
 146void map_cpu_to_node(int cpu, int node)
 
 
 
 
 
 
 
 147{
 148	update_numa_cpu_lookup_table(cpu, node);
 149
 150	if (!(cpumask_test_cpu(cpu, node_to_cpumask_map[node]))) {
 151		pr_debug("adding cpu %d to node %d\n", cpu, node);
 
 
 
 
 
 
 
 
 
 
 
 152		cpumask_set_cpu(cpu, node_to_cpumask_map[node]);
 153	}
 154}
 155
 156#if defined(CONFIG_HOTPLUG_CPU) || defined(CONFIG_PPC_SPLPAR)
 157void unmap_cpu_from_node(unsigned long cpu)
 158{
 159	int node = numa_cpu_lookup_table[cpu];
 160
 
 
 161	if (cpumask_test_cpu(cpu, node_to_cpumask_map[node])) {
 162		cpumask_clear_cpu(cpu, node_to_cpumask_map[node]);
 163		pr_debug("removing cpu %lu from node %d\n", cpu, node);
 164	} else {
 165		pr_warn("Warning: cpu %lu not found in node %d\n", cpu, node);
 
 166	}
 167}
 168#endif /* CONFIG_HOTPLUG_CPU || CONFIG_PPC_SPLPAR */
 169
 170static int __associativity_to_nid(const __be32 *associativity,
 171				  int max_array_sz)
 172{
 173	int nid;
 174	/*
 175	 * primary_domain_index is 1 based array index.
 176	 */
 177	int index = primary_domain_index  - 1;
 178
 179	if (!numa_enabled || index >= max_array_sz)
 180		return NUMA_NO_NODE;
 181
 182	nid = of_read_number(&associativity[index], 1);
 183
 184	/* POWER4 LPAR uses 0xffff as invalid node */
 185	if (nid == 0xffff || nid >= nr_node_ids)
 186		nid = NUMA_NO_NODE;
 187	return nid;
 188}
 
 189/*
 190 * Returns nid in the range [0..nr_node_ids], or -1 if no useful NUMA
 191 * info is found.
 
 192 */
 193static int associativity_to_nid(const __be32 *associativity)
 194{
 195	int array_sz = of_read_number(associativity, 1);
 196
 197	/* Skip the first element in the associativity array */
 198	return __associativity_to_nid((associativity + 1), array_sz);
 199}
 200
 201static int __cpu_form2_relative_distance(__be32 *cpu1_assoc, __be32 *cpu2_assoc)
 202{
 203	int dist;
 204	int node1, node2;
 205
 206	node1 = associativity_to_nid(cpu1_assoc);
 207	node2 = associativity_to_nid(cpu2_assoc);
 208
 209	dist = numa_distance_table[node1][node2];
 210	if (dist <= LOCAL_DISTANCE)
 211		return 0;
 212	else if (dist <= REMOTE_DISTANCE)
 213		return 1;
 214	else
 215		return 2;
 216}
 217
 218static int __cpu_form1_relative_distance(__be32 *cpu1_assoc, __be32 *cpu2_assoc)
 219{
 220	int dist = 0;
 
 221
 222	int i, index;
 
 223
 224	for (i = 0; i < distance_ref_points_depth; i++) {
 225		index = be32_to_cpu(distance_ref_points[i]);
 226		if (cpu1_assoc[index] == cpu2_assoc[index])
 227			break;
 228		dist++;
 
 
 229	}
 230
 231	return dist;
 232}
 233
 234int cpu_relative_distance(__be32 *cpu1_assoc, __be32 *cpu2_assoc)
 
 235{
 236	/* We should not get called with FORM0 */
 237	VM_WARN_ON(affinity_form == FORM0_AFFINITY);
 238	if (affinity_form == FORM1_AFFINITY)
 239		return __cpu_form1_relative_distance(cpu1_assoc, cpu2_assoc);
 240	return __cpu_form2_relative_distance(cpu1_assoc, cpu2_assoc);
 241}
 242
 243/* must hold reference to node during call */
 244static const __be32 *of_get_associativity(struct device_node *dev)
 245{
 246	return of_get_property(dev, "ibm,associativity", NULL);
 
 
 
 247}
 248
 249int __node_distance(int a, int b)
 
 
 
 250{
 251	int i;
 252	int distance = LOCAL_DISTANCE;
 253
 254	if (affinity_form == FORM2_AFFINITY)
 255		return numa_distance_table[a][b];
 256	else if (affinity_form == FORM0_AFFINITY)
 257		return ((a == b) ? LOCAL_DISTANCE : REMOTE_DISTANCE);
 258
 259	for (i = 0; i < distance_ref_points_depth; i++) {
 260		if (distance_lookup_table[a][i] == distance_lookup_table[b][i])
 261			break;
 262
 263		/* Double the distance for each NUMA level */
 264		distance *= 2;
 265	}
 266
 267	return distance;
 
 
 
 
 268}
 269EXPORT_SYMBOL(__node_distance);
 270
 271/* Returns the nid associated with the given device tree node,
 272 * or -1 if not found.
 273 */
 274static int of_node_to_nid_single(struct device_node *device)
 275{
 276	int nid = NUMA_NO_NODE;
 277	const __be32 *tmp;
 278
 279	tmp = of_get_associativity(device);
 280	if (tmp)
 281		nid = associativity_to_nid(tmp);
 282	return nid;
 283}
 284
 285/* Walk the device tree upwards, looking for an associativity id */
 286int of_node_to_nid(struct device_node *device)
 287{
 288	int nid = NUMA_NO_NODE;
 
 289
 290	of_node_get(device);
 291	while (device) {
 292		nid = of_node_to_nid_single(device);
 293		if (nid != -1)
 294			break;
 295
 296		device = of_get_next_parent(device);
 
 
 297	}
 298	of_node_put(device);
 299
 300	return nid;
 301}
 302EXPORT_SYMBOL(of_node_to_nid);
 303
 304static void __initialize_form1_numa_distance(const __be32 *associativity,
 305					     int max_array_sz)
 306{
 307	int i, nid;
 308
 309	if (affinity_form != FORM1_AFFINITY)
 310		return;
 311
 312	nid = __associativity_to_nid(associativity, max_array_sz);
 313	if (nid != NUMA_NO_NODE) {
 314		for (i = 0; i < distance_ref_points_depth; i++) {
 315			const __be32 *entry;
 316			int index = be32_to_cpu(distance_ref_points[i]) - 1;
 317
 318			/*
 319			 * broken hierarchy, return with broken distance table
 320			 */
 321			if (WARN(index >= max_array_sz, "Broken ibm,associativity property"))
 322				return;
 323
 324			entry = &associativity[index];
 325			distance_lookup_table[nid][i] = of_read_number(entry, 1);
 326		}
 327	}
 328}
 329
 330static void initialize_form1_numa_distance(const __be32 *associativity)
 331{
 332	int array_sz;
 333
 334	array_sz = of_read_number(associativity, 1);
 335	/* Skip the first element in the associativity array */
 336	__initialize_form1_numa_distance(associativity + 1, array_sz);
 337}
 338
 339/*
 340 * Used to update distance information w.r.t newly added node.
 341 */
 342void update_numa_distance(struct device_node *node)
 343{
 344	int nid;
 345
 346	if (affinity_form == FORM0_AFFINITY)
 347		return;
 348	else if (affinity_form == FORM1_AFFINITY) {
 349		const __be32 *associativity;
 350
 351		associativity = of_get_associativity(node);
 352		if (!associativity)
 353			return;
 354
 355		initialize_form1_numa_distance(associativity);
 356		return;
 357	}
 358
 359	/* FORM2 affinity  */
 360	nid = of_node_to_nid_single(node);
 361	if (nid == NUMA_NO_NODE)
 362		return;
 363
 364	/*
 365	 * With FORM2 we expect NUMA distance of all possible NUMA
 366	 * nodes to be provided during boot.
 367	 */
 368	WARN(numa_distance_table[nid][nid] == -1,
 369	     "NUMA distance details for node %d not provided\n", nid);
 370}
 371EXPORT_SYMBOL_GPL(update_numa_distance);
 372
 373/*
 374 * ibm,numa-lookup-index-table= {N, domainid1, domainid2, ..... domainidN}
 375 * ibm,numa-distance-table = { N, 1, 2, 4, 5, 1, 6, .... N elements}
 376 */
 377static void __init initialize_form2_numa_distance_lookup_table(void)
 378{
 379	int i, j;
 380	struct device_node *root;
 381	const __u8 *form2_distances;
 382	const __be32 *numa_lookup_index;
 383	int form2_distances_length;
 384	int max_numa_index, distance_index;
 385
 386	if (firmware_has_feature(FW_FEATURE_OPAL))
 387		root = of_find_node_by_path("/ibm,opal");
 388	else
 389		root = of_find_node_by_path("/rtas");
 390	if (!root)
 391		root = of_find_node_by_path("/");
 392
 393	numa_lookup_index = of_get_property(root, "ibm,numa-lookup-index-table", NULL);
 394	max_numa_index = of_read_number(&numa_lookup_index[0], 1);
 395
 396	/* first element of the array is the size and is encode-int */
 397	form2_distances = of_get_property(root, "ibm,numa-distance-table", NULL);
 398	form2_distances_length = of_read_number((const __be32 *)&form2_distances[0], 1);
 399	/* Skip the size which is encoded int */
 400	form2_distances += sizeof(__be32);
 401
 402	pr_debug("form2_distances_len = %d, numa_dist_indexes_len = %d\n",
 403		 form2_distances_length, max_numa_index);
 404
 405	for (i = 0; i < max_numa_index; i++)
 406		/* +1 skip the max_numa_index in the property */
 407		numa_id_index_table[i] = of_read_number(&numa_lookup_index[i + 1], 1);
 408
 409
 410	if (form2_distances_length != max_numa_index * max_numa_index) {
 411		WARN(1, "Wrong NUMA distance information\n");
 412		form2_distances = NULL; // don't use it
 413	}
 414	distance_index = 0;
 415	for (i = 0;  i < max_numa_index; i++) {
 416		for (j = 0; j < max_numa_index; j++) {
 417			int nodeA = numa_id_index_table[i];
 418			int nodeB = numa_id_index_table[j];
 419			int dist;
 420
 421			if (form2_distances)
 422				dist = form2_distances[distance_index++];
 423			else if (nodeA == nodeB)
 424				dist = LOCAL_DISTANCE;
 425			else
 426				dist = REMOTE_DISTANCE;
 427			numa_distance_table[nodeA][nodeB] = dist;
 428			pr_debug("dist[%d][%d]=%d ", nodeA, nodeB, dist);
 429		}
 430	}
 431
 432	of_node_put(root);
 433}
 434
 435static int __init find_primary_domain_index(void)
 436{
 437	int index;
 
 438	struct device_node *root;
 
 439
 440	/*
 441	 * Check for which form of affinity.
 442	 */
 443	if (firmware_has_feature(FW_FEATURE_OPAL)) {
 444		affinity_form = FORM1_AFFINITY;
 445	} else if (firmware_has_feature(FW_FEATURE_FORM2_AFFINITY)) {
 446		pr_debug("Using form 2 affinity\n");
 447		affinity_form = FORM2_AFFINITY;
 448	} else if (firmware_has_feature(FW_FEATURE_FORM1_AFFINITY)) {
 449		pr_debug("Using form 1 affinity\n");
 450		affinity_form = FORM1_AFFINITY;
 451	} else
 452		affinity_form = FORM0_AFFINITY;
 453
 454	if (firmware_has_feature(FW_FEATURE_OPAL))
 455		root = of_find_node_by_path("/ibm,opal");
 456	else
 457		root = of_find_node_by_path("/rtas");
 458	if (!root)
 459		root = of_find_node_by_path("/");
 460
 461	/*
 462	 * This property is a set of 32-bit integers, each representing
 463	 * an index into the ibm,associativity nodes.
 464	 *
 465	 * With form 0 affinity the first integer is for an SMP configuration
 466	 * (should be all 0's) and the second is for a normal NUMA
 467	 * configuration. We have only one level of NUMA.
 468	 *
 469	 * With form 1 affinity the first integer is the most significant
 470	 * NUMA boundary and the following are progressively less significant
 471	 * boundaries. There can be more than one level of NUMA.
 472	 */
 473	distance_ref_points = of_get_property(root,
 474					"ibm,associativity-reference-points",
 475					&distance_ref_points_depth);
 476
 477	if (!distance_ref_points) {
 478		pr_debug("ibm,associativity-reference-points not found.\n");
 479		goto err;
 480	}
 481
 482	distance_ref_points_depth /= sizeof(int);
 483	if (affinity_form == FORM0_AFFINITY) {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 484		if (distance_ref_points_depth < 2) {
 485			pr_warn("short ibm,associativity-reference-points\n");
 
 486			goto err;
 487		}
 488
 489		index = of_read_number(&distance_ref_points[1], 1);
 490	} else {
 491		/*
 492		 * Both FORM1 and FORM2 affinity find the primary domain details
 493		 * at the same offset.
 494		 */
 495		index = of_read_number(distance_ref_points, 1);
 496	}
 
 497	/*
 498	 * Warn and cap if the hardware supports more than
 499	 * MAX_DISTANCE_REF_POINTS domains.
 500	 */
 501	if (distance_ref_points_depth > MAX_DISTANCE_REF_POINTS) {
 502		pr_warn("distance array capped at %d entries\n",
 503			MAX_DISTANCE_REF_POINTS);
 504		distance_ref_points_depth = MAX_DISTANCE_REF_POINTS;
 505	}
 506
 507	of_node_put(root);
 508	return index;
 509
 510err:
 511	of_node_put(root);
 512	return -1;
 513}
 514
 515static void __init get_n_mem_cells(int *n_addr_cells, int *n_size_cells)
 516{
 517	struct device_node *memory = NULL;
 518
 519	memory = of_find_node_by_type(memory, "memory");
 520	if (!memory)
 521		panic("numa.c: No memory nodes found!");
 522
 523	*n_addr_cells = of_n_addr_cells(memory);
 524	*n_size_cells = of_n_size_cells(memory);
 525	of_node_put(memory);
 526}
 527
 528static unsigned long read_n_cells(int n, const __be32 **buf)
 529{
 530	unsigned long result = 0;
 531
 532	while (n--) {
 533		result = (result << 32) | of_read_number(*buf, 1);
 534		(*buf)++;
 535	}
 536	return result;
 537}
 538
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 539struct assoc_arrays {
 540	u32	n_arrays;
 541	u32	array_sz;
 542	const __be32 *arrays;
 543};
 544
 545/*
 546 * Retrieve and validate the list of associativity arrays for drconf
 547 * memory from the ibm,associativity-lookup-arrays property of the
 548 * device tree..
 549 *
 550 * The layout of the ibm,associativity-lookup-arrays property is a number N
 551 * indicating the number of associativity arrays, followed by a number M
 552 * indicating the size of each associativity array, followed by a list
 553 * of N associativity arrays.
 554 */
 555static int of_get_assoc_arrays(struct assoc_arrays *aa)
 
 556{
 557	struct device_node *memory;
 558	const __be32 *prop;
 559	u32 len;
 560
 561	memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
 562	if (!memory)
 563		return -1;
 564
 565	prop = of_get_property(memory, "ibm,associativity-lookup-arrays", &len);
 566	if (!prop || len < 2 * sizeof(unsigned int)) {
 567		of_node_put(memory);
 568		return -1;
 569	}
 570
 571	aa->n_arrays = of_read_number(prop++, 1);
 572	aa->array_sz = of_read_number(prop++, 1);
 573
 574	of_node_put(memory);
 575
 576	/* Now that we know the number of arrays and size of each array,
 577	 * revalidate the size of the property read in.
 578	 */
 579	if (len < (aa->n_arrays * aa->array_sz + 2) * sizeof(unsigned int))
 580		return -1;
 581
 582	aa->arrays = prop;
 583	return 0;
 584}
 585
 586static int __init get_nid_and_numa_distance(struct drmem_lmb *lmb)
 587{
 588	struct assoc_arrays aa = { .arrays = NULL };
 589	int default_nid = NUMA_NO_NODE;
 590	int nid = default_nid;
 591	int rc, index;
 592
 593	if ((primary_domain_index < 0) || !numa_enabled)
 594		return default_nid;
 595
 596	rc = of_get_assoc_arrays(&aa);
 597	if (rc)
 598		return default_nid;
 599
 600	if (primary_domain_index <= aa.array_sz &&
 601	    !(lmb->flags & DRCONF_MEM_AI_INVALID) && lmb->aa_index < aa.n_arrays) {
 602		const __be32 *associativity;
 603
 604		index = lmb->aa_index * aa.array_sz;
 605		associativity = &aa.arrays[index];
 606		nid = __associativity_to_nid(associativity, aa.array_sz);
 607		if (nid > 0 && affinity_form == FORM1_AFFINITY) {
 608			/*
 609			 * lookup array associativity entries have
 610			 * no length of the array as the first element.
 611			 */
 612			__initialize_form1_numa_distance(associativity, aa.array_sz);
 613		}
 614	}
 615	return nid;
 616}
 617
 618/*
 619 * This is like of_node_to_nid_single() for memory represented in the
 620 * ibm,dynamic-reconfiguration-memory node.
 621 */
 622int of_drconf_to_nid_single(struct drmem_lmb *lmb)
 
 623{
 624	struct assoc_arrays aa = { .arrays = NULL };
 625	int default_nid = NUMA_NO_NODE;
 626	int nid = default_nid;
 627	int rc, index;
 628
 629	if ((primary_domain_index < 0) || !numa_enabled)
 630		return default_nid;
 631
 632	rc = of_get_assoc_arrays(&aa);
 633	if (rc)
 634		return default_nid;
 635
 636	if (primary_domain_index <= aa.array_sz &&
 637	    !(lmb->flags & DRCONF_MEM_AI_INVALID) && lmb->aa_index < aa.n_arrays) {
 638		const __be32 *associativity;
 639
 640		index = lmb->aa_index * aa.array_sz;
 641		associativity = &aa.arrays[index];
 642		nid = __associativity_to_nid(associativity, aa.array_sz);
 643	}
 644	return nid;
 645}
 646
 647#ifdef CONFIG_PPC_SPLPAR
 648
 649static int __vphn_get_associativity(long lcpu, __be32 *associativity)
 650{
 651	long rc, hwid;
 
 
 652
 653	/*
 654	 * On a shared lpar, device tree will not have node associativity.
 655	 * At this time lppaca, or its __old_status field may not be
 656	 * updated. Hence kernel cannot detect if its on a shared lpar. So
 657	 * request an explicit associativity irrespective of whether the
 658	 * lpar is shared or dedicated. Use the device tree property as a
 659	 * fallback. cpu_to_phys_id is only valid between
 660	 * smp_setup_cpu_maps() and smp_setup_pacas().
 661	 */
 662	if (firmware_has_feature(FW_FEATURE_VPHN)) {
 663		if (cpu_to_phys_id)
 664			hwid = cpu_to_phys_id[lcpu];
 665		else
 666			hwid = get_hard_smp_processor_id(lcpu);
 667
 668		rc = hcall_vphn(hwid, VPHN_FLAG_VCPU, associativity);
 669		if (rc == H_SUCCESS)
 670			return 0;
 671	}
 672
 673	return -1;
 674}
 675
 676static int vphn_get_nid(long lcpu)
 677{
 678	__be32 associativity[VPHN_ASSOC_BUFSIZE] = {0};
 679
 680
 681	if (!__vphn_get_associativity(lcpu, associativity))
 682		return associativity_to_nid(associativity);
 683
 684	return NUMA_NO_NODE;
 685
 686}
 687#else
 688
 689static int __vphn_get_associativity(long lcpu, __be32 *associativity)
 690{
 691	return -1;
 692}
 693
 694static int vphn_get_nid(long unused)
 695{
 696	return NUMA_NO_NODE;
 697}
 698#endif  /* CONFIG_PPC_SPLPAR */
 699
 700/*
 701 * Figure out to which domain a cpu belongs and stick it there.
 702 * Return the id of the domain used.
 703 */
 704static int numa_setup_cpu(unsigned long lcpu)
 705{
 706	struct device_node *cpu;
 707	int fcpu = cpu_first_thread_sibling(lcpu);
 708	int nid = NUMA_NO_NODE;
 709
 710	if (!cpu_present(lcpu)) {
 711		set_cpu_numa_node(lcpu, first_online_node);
 712		return first_online_node;
 713	}
 714
 715	/*
 716	 * If a valid cpu-to-node mapping is already available, use it
 717	 * directly instead of querying the firmware, since it represents
 718	 * the most recent mapping notified to us by the platform (eg: VPHN).
 719	 * Since cpu_to_node binding remains the same for all threads in the
 720	 * core. If a valid cpu-to-node mapping is already available, for
 721	 * the first thread in the core, use it.
 722	 */
 723	nid = numa_cpu_lookup_table[fcpu];
 724	if (nid >= 0) {
 725		map_cpu_to_node(lcpu, nid);
 726		return nid;
 727	}
 728
 729	nid = vphn_get_nid(lcpu);
 730	if (nid != NUMA_NO_NODE)
 731		goto out_present;
 732
 733	cpu = of_get_cpu_node(lcpu, NULL);
 734
 735	if (!cpu) {
 736		WARN_ON(1);
 737		if (cpu_present(lcpu))
 738			goto out_present;
 739		else
 740			goto out;
 741	}
 742
 743	nid = of_node_to_nid_single(cpu);
 744	of_node_put(cpu);
 745
 746out_present:
 747	if (nid < 0 || !node_possible(nid))
 748		nid = first_online_node;
 
 
 749
 750	/*
 751	 * Update for the first thread of the core. All threads of a core
 752	 * have to be part of the same node. This not only avoids querying
 753	 * for every other thread in the core, but always avoids a case
 754	 * where virtual node associativity change causes subsequent threads
 755	 * of a core to be associated with different nid. However if first
 756	 * thread is already online, expect it to have a valid mapping.
 757	 */
 758	if (fcpu != lcpu) {
 759		WARN_ON(cpu_online(fcpu));
 760		map_cpu_to_node(fcpu, nid);
 761	}
 762
 763	map_cpu_to_node(lcpu, nid);
 764out:
 765	return nid;
 766}
 767
 768static void verify_cpu_node_mapping(int cpu, int node)
 769{
 770	int base, sibling, i;
 771
 772	/* Verify that all the threads in the core belong to the same node */
 773	base = cpu_first_thread_sibling(cpu);
 774
 775	for (i = 0; i < threads_per_core; i++) {
 776		sibling = base + i;
 777
 778		if (sibling == cpu || cpu_is_offline(sibling))
 779			continue;
 780
 781		if (cpu_to_node(sibling) != node) {
 782			WARN(1, "CPU thread siblings %d and %d don't belong"
 783				" to the same node!\n", cpu, sibling);
 784			break;
 785		}
 
 
 
 
 786	}
 787}
 788
 789/* Must run before sched domains notifier. */
 790static int ppc_numa_cpu_prepare(unsigned int cpu)
 791{
 792	int nid;
 793
 794	nid = numa_setup_cpu(cpu);
 795	verify_cpu_node_mapping(cpu, nid);
 796	return 0;
 797}
 798
 799static int ppc_numa_cpu_dead(unsigned int cpu)
 800{
 801	return 0;
 802}
 803
 804/*
 805 * Check and possibly modify a memory region to enforce the memory limit.
 806 *
 807 * Returns the size the region should have to enforce the memory limit.
 808 * This will either be the original value of size, a truncated value,
 809 * or zero. If the returned value of size is 0 the region should be
 810 * discarded as it lies wholly above the memory limit.
 811 */
 812static unsigned long __init numa_enforce_memory_limit(unsigned long start,
 813						      unsigned long size)
 814{
 815	/*
 816	 * We use memblock_end_of_DRAM() in here instead of memory_limit because
 817	 * we've already adjusted it for the limit and it takes care of
 818	 * having memory holes below the limit.  Also, in the case of
 819	 * iommu_is_off, memory_limit is not set but is implicitly enforced.
 820	 */
 821
 822	if (start + size <= memblock_end_of_DRAM())
 823		return size;
 824
 825	if (start >= memblock_end_of_DRAM())
 826		return 0;
 827
 828	return memblock_end_of_DRAM() - start;
 829}
 830
 831/*
 832 * Reads the counter for a given entry in
 833 * linux,drconf-usable-memory property
 834 */
 835static inline int __init read_usm_ranges(const __be32 **usm)
 836{
 837	/*
 838	 * For each lmb in ibm,dynamic-memory a corresponding
 839	 * entry in linux,drconf-usable-memory property contains
 840	 * a counter followed by that many (base, size) duple.
 841	 * read the counter from linux,drconf-usable-memory
 842	 */
 843	return read_n_cells(n_mem_size_cells, usm);
 844}
 845
 846/*
 847 * Extract NUMA information from the ibm,dynamic-reconfiguration-memory
 848 * node.  This assumes n_mem_{addr,size}_cells have been set.
 849 */
 850static int __init numa_setup_drmem_lmb(struct drmem_lmb *lmb,
 851					const __be32 **usm,
 852					void *data)
 853{
 854	unsigned int ranges, is_kexec_kdump = 0;
 855	unsigned long base, size, sz;
 
 856	int nid;
 
 857
 858	/*
 859	 * Skip this block if the reserved bit is set in flags (0x80)
 860	 * or if the block is not assigned to this partition (0x8)
 861	 */
 862	if ((lmb->flags & DRCONF_MEM_RESERVED)
 863	    || !(lmb->flags & DRCONF_MEM_ASSIGNED))
 864		return 0;
 865
 866	if (*usm)
 
 
 
 
 
 
 
 
 
 
 867		is_kexec_kdump = 1;
 868
 869	base = lmb->base_addr;
 870	size = drmem_lmb_size();
 871	ranges = 1;
 872
 873	if (is_kexec_kdump) {
 874		ranges = read_usm_ranges(usm);
 875		if (!ranges) /* there are no (base, size) duple */
 876			return 0;
 877	}
 878
 879	do {
 880		if (is_kexec_kdump) {
 881			base = read_n_cells(n_mem_addr_cells, usm);
 882			size = read_n_cells(n_mem_size_cells, usm);
 883		}
 884
 885		nid = get_nid_and_numa_distance(lmb);
 886		fake_numa_create_new_node(((base + size) >> PAGE_SHIFT),
 887					  &nid);
 888		node_set_online(nid);
 889		sz = numa_enforce_memory_limit(base, size);
 890		if (sz)
 891			memblock_set_node(base, sz, &memblock.memory, nid);
 892	} while (--ranges);
 893
 894	return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 895}
 896
 897static int __init parse_numa_properties(void)
 898{
 899	struct device_node *memory;
 
 900	int default_nid = 0;
 901	unsigned long i;
 902	const __be32 *associativity;
 903
 904	if (numa_enabled == 0) {
 905		pr_warn("disabled by user\n");
 906		return -1;
 907	}
 908
 909	primary_domain_index = find_primary_domain_index();
 910
 911	if (primary_domain_index < 0) {
 912		/*
 913		 * if we fail to parse primary_domain_index from device tree
 914		 * mark the numa disabled, boot with numa disabled.
 915		 */
 916		numa_enabled = false;
 917		return primary_domain_index;
 918	}
 919
 920	pr_debug("associativity depth for CPU/Memory: %d\n", primary_domain_index);
 921
 922	/*
 923	 * If it is FORM2 initialize the distance table here.
 924	 */
 925	if (affinity_form == FORM2_AFFINITY)
 926		initialize_form2_numa_distance_lookup_table();
 927
 928	/*
 929	 * Even though we connect cpus to numa domains later in SMP
 930	 * init, we need to know the node ids now. This is because
 931	 * each node to be onlined must have NODE_DATA etc backing it.
 932	 */
 933	for_each_present_cpu(i) {
 934		__be32 vphn_assoc[VPHN_ASSOC_BUFSIZE];
 935		struct device_node *cpu;
 936		int nid = NUMA_NO_NODE;
 937
 938		memset(vphn_assoc, 0, VPHN_ASSOC_BUFSIZE * sizeof(__be32));
 939
 940		if (__vphn_get_associativity(i, vphn_assoc) == 0) {
 941			nid = associativity_to_nid(vphn_assoc);
 942			initialize_form1_numa_distance(vphn_assoc);
 943		} else {
 944
 945			/*
 946			 * Don't fall back to default_nid yet -- we will plug
 947			 * cpus into nodes once the memory scan has discovered
 948			 * the topology.
 949			 */
 950			cpu = of_get_cpu_node(i, NULL);
 951			BUG_ON(!cpu);
 952
 953			associativity = of_get_associativity(cpu);
 954			if (associativity) {
 955				nid = associativity_to_nid(associativity);
 956				initialize_form1_numa_distance(associativity);
 957			}
 958			of_node_put(cpu);
 959		}
 960
 961		/* node_set_online() is an UB if 'nid' is negative */
 962		if (likely(nid >= 0))
 963			node_set_online(nid);
 
 
 
 
 
 964	}
 965
 966	get_n_mem_cells(&n_mem_addr_cells, &n_mem_size_cells);
 967
 968	for_each_node_by_type(memory, "memory") {
 969		unsigned long start;
 970		unsigned long size;
 971		int nid;
 972		int ranges;
 973		const __be32 *memcell_buf;
 974		unsigned int len;
 975
 976		memcell_buf = of_get_property(memory,
 977			"linux,usable-memory", &len);
 978		if (!memcell_buf || len <= 0)
 979			memcell_buf = of_get_property(memory, "reg", &len);
 980		if (!memcell_buf || len <= 0)
 981			continue;
 982
 983		/* ranges in cell */
 984		ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
 985new_range:
 986		/* these are order-sensitive, and modify the buffer pointer */
 987		start = read_n_cells(n_mem_addr_cells, &memcell_buf);
 988		size = read_n_cells(n_mem_size_cells, &memcell_buf);
 989
 990		/*
 991		 * Assumption: either all memory nodes or none will
 992		 * have associativity properties.  If none, then
 993		 * everything goes to default_nid.
 994		 */
 995		associativity = of_get_associativity(memory);
 996		if (associativity) {
 997			nid = associativity_to_nid(associativity);
 998			initialize_form1_numa_distance(associativity);
 999		} else
1000			nid = default_nid;
1001
1002		fake_numa_create_new_node(((start + size) >> PAGE_SHIFT), &nid);
1003		node_set_online(nid);
1004
1005		size = numa_enforce_memory_limit(start, size);
1006		if (size)
1007			memblock_set_node(start, size, &memblock.memory, nid);
 
 
 
 
 
 
1008
1009		if (--ranges)
1010			goto new_range;
1011	}
1012
1013	/*
1014	 * Now do the same thing for each MEMBLOCK listed in the
1015	 * ibm,dynamic-memory property in the
1016	 * ibm,dynamic-reconfiguration-memory node.
1017	 */
1018	memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1019	if (memory) {
1020		walk_drmem_lmbs(memory, NULL, numa_setup_drmem_lmb);
1021		of_node_put(memory);
1022	}
1023
1024	return 0;
1025}
1026
1027static void __init setup_nonnuma(void)
1028{
1029	unsigned long top_of_ram = memblock_end_of_DRAM();
1030	unsigned long total_ram = memblock_phys_mem_size();
1031	unsigned long start_pfn, end_pfn;
1032	unsigned int nid = 0;
1033	int i;
1034
1035	pr_debug("Top of RAM: 0x%lx, Total RAM: 0x%lx\n", top_of_ram, total_ram);
1036	pr_debug("Memory hole size: %ldMB\n", (top_of_ram - total_ram) >> 20);
 
 
 
 
 
 
1037
1038	for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, NULL) {
1039		fake_numa_create_new_node(end_pfn, &nid);
1040		memblock_set_node(PFN_PHYS(start_pfn),
1041				  PFN_PHYS(end_pfn - start_pfn),
1042				  &memblock.memory, nid);
1043		node_set_online(nid);
1044	}
1045}
1046
1047void __init dump_numa_cpu_topology(void)
1048{
1049	unsigned int node;
1050	unsigned int cpu, count;
1051
1052	if (!numa_enabled)
1053		return;
1054
1055	for_each_online_node(node) {
1056		pr_info("Node %d CPUs:", node);
1057
1058		count = 0;
1059		/*
1060		 * If we used a CPU iterator here we would miss printing
1061		 * the holes in the cpumap.
1062		 */
1063		for (cpu = 0; cpu < nr_cpu_ids; cpu++) {
1064			if (cpumask_test_cpu(cpu,
1065					node_to_cpumask_map[node])) {
1066				if (count == 0)
1067					pr_cont(" %u", cpu);
1068				++count;
1069			} else {
1070				if (count > 1)
1071					pr_cont("-%u", cpu - 1);
1072				count = 0;
1073			}
1074		}
1075
1076		if (count > 1)
1077			pr_cont("-%u", nr_cpu_ids - 1);
1078		pr_cont("\n");
1079	}
1080}
1081
1082/* Initialize NODE_DATA for a node on the local memory */
1083static void __init setup_node_data(int nid, u64 start_pfn, u64 end_pfn)
1084{
1085	u64 spanned_pages = end_pfn - start_pfn;
1086	const size_t nd_size = roundup(sizeof(pg_data_t), SMP_CACHE_BYTES);
1087	u64 nd_pa;
1088	void *nd;
1089	int tnid;
1090
1091	nd_pa = memblock_phys_alloc_try_nid(nd_size, SMP_CACHE_BYTES, nid);
1092	if (!nd_pa)
1093		panic("Cannot allocate %zu bytes for node %d data\n",
1094		      nd_size, nid);
1095
1096	nd = __va(nd_pa);
 
1097
1098	/* report and initialize */
1099	pr_info("  NODE_DATA [mem %#010Lx-%#010Lx]\n",
1100		nd_pa, nd_pa + nd_size - 1);
1101	tnid = early_pfn_to_nid(nd_pa >> PAGE_SHIFT);
1102	if (tnid != nid)
1103		pr_info("    NODE_DATA(%d) on node %d\n", nid, tnid);
1104
1105	node_data[nid] = nd;
1106	memset(NODE_DATA(nid), 0, sizeof(pg_data_t));
1107	NODE_DATA(nid)->node_id = nid;
1108	NODE_DATA(nid)->node_start_pfn = start_pfn;
1109	NODE_DATA(nid)->node_spanned_pages = spanned_pages;
1110}
1111
1112static void __init find_possible_nodes(void)
1113{
1114	struct device_node *rtas;
1115	const __be32 *domains = NULL;
1116	int prop_length, max_nodes;
1117	u32 i;
 
 
 
 
 
 
1118
1119	if (!numa_enabled)
1120		return;
 
 
 
1121
1122	rtas = of_find_node_by_path("/rtas");
1123	if (!rtas)
1124		return;
 
 
 
 
 
 
 
 
 
 
 
1125
1126	/*
1127	 * ibm,current-associativity-domains is a fairly recent property. If
1128	 * it doesn't exist, then fallback on ibm,max-associativity-domains.
1129	 * Current denotes what the platform can support compared to max
1130	 * which denotes what the Hypervisor can support.
1131	 *
1132	 * If the LPAR is migratable, new nodes might be activated after a LPM,
1133	 * so we should consider the max number in that case.
1134	 */
1135	if (!of_get_property(of_root, "ibm,migratable-partition", NULL))
1136		domains = of_get_property(rtas,
1137					  "ibm,current-associativity-domains",
1138					  &prop_length);
1139	if (!domains) {
1140		domains = of_get_property(rtas, "ibm,max-associativity-domains",
1141					&prop_length);
1142		if (!domains)
1143			goto out;
1144	}
1145
1146	max_nodes = of_read_number(&domains[primary_domain_index], 1);
1147	pr_info("Partition configured for %d NUMA nodes.\n", max_nodes);
1148
1149	for (i = 0; i < max_nodes; i++) {
1150		if (!node_possible(i))
1151			node_set(i, node_possible_map);
1152	}
1153
1154	prop_length /= sizeof(int);
1155	if (prop_length > primary_domain_index + 2)
1156		coregroup_enabled = 1;
1157
1158out:
1159	of_node_put(rtas);
1160}
1161
1162void __init mem_topology_setup(void)
1163{
1164	int cpu;
1165
1166	max_low_pfn = max_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
1167	min_low_pfn = MEMORY_START >> PAGE_SHIFT;
1168
1169	/*
1170	 * Linux/mm assumes node 0 to be online at boot. However this is not
1171	 * true on PowerPC, where node 0 is similar to any other node, it
1172	 * could be cpuless, memoryless node. So force node 0 to be offline
1173	 * for now. This will prevent cpuless, memoryless node 0 showing up
1174	 * unnecessarily as online. If a node has cpus or memory that need
1175	 * to be online, then node will anyway be marked online.
 
 
 
 
1176	 */
1177	node_set_offline(0);
 
 
 
1178
1179	if (parse_numa_properties())
1180		setup_nonnuma();
1181
1182	/*
1183	 * Modify the set of possible NUMA nodes to reflect information
1184	 * available about the set of online nodes, and the set of nodes
1185	 * that we expect to make use of for this platform's affinity
1186	 * calculations.
1187	 */
1188	nodes_and(node_possible_map, node_possible_map, node_online_map);
1189
1190	find_possible_nodes();
 
 
 
1191
1192	setup_node_to_cpumask_map();
 
 
 
1193
1194	reset_numa_cpu_lookup_table();
 
 
 
 
 
 
 
1195
1196	for_each_possible_cpu(cpu) {
1197		/*
1198		 * Powerpc with CONFIG_NUMA always used to have a node 0,
1199		 * even if it was memoryless or cpuless. For all cpus that
1200		 * are possible but not present, cpu_to_node() would point
1201		 * to node 0. To remove a cpuless, memoryless dummy node,
1202		 * powerpc need to make sure all possible but not present
1203		 * cpu_to_node are set to a proper node.
1204		 */
1205		numa_setup_cpu(cpu);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1206	}
1207}
1208
1209void __init initmem_init(void)
 
1210{
1211	int nid;
1212
1213	memblock_dump_all();
 
 
 
 
 
 
 
1214
1215	for_each_online_node(nid) {
1216		unsigned long start_pfn, end_pfn;
 
 
1217
1218		get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
1219		setup_node_data(nid, start_pfn, end_pfn);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1220	}
1221
1222	sparse_init();
1223
1224	/*
1225	 * We need the numa_cpu_lookup_table to be accurate for all CPUs,
1226	 * even before we online them, so that we can use cpu_to_{node,mem}
1227	 * early in boot, cf. smp_prepare_cpus().
1228	 * _nocalls() + manual invocation is used because cpuhp is not yet
1229	 * initialized for the boot CPU.
1230	 */
1231	cpuhp_setup_state_nocalls(CPUHP_POWER_NUMA_PREPARE, "powerpc/numa:prepare",
1232				  ppc_numa_cpu_prepare, ppc_numa_cpu_dead);
 
 
 
 
 
 
 
 
 
 
 
1233}
1234
1235static int __init early_numa(char *p)
1236{
1237	if (!p)
1238		return 0;
1239
1240	if (strstr(p, "off"))
1241		numa_enabled = 0;
1242
 
 
 
1243	p = strstr(p, "fake=");
1244	if (p)
1245		cmdline = p + strlen("fake=");
1246
1247	return 0;
1248}
1249early_param("numa", early_numa);
1250
1251#ifdef CONFIG_MEMORY_HOTPLUG
1252/*
1253 * Find the node associated with a hot added memory section for
1254 * memory represented in the device tree by the property
1255 * ibm,dynamic-reconfiguration-memory/ibm,dynamic-memory.
1256 */
1257static int hot_add_drconf_scn_to_nid(unsigned long scn_addr)
 
1258{
1259	struct drmem_lmb *lmb;
 
1260	unsigned long lmb_size;
1261	int nid = NUMA_NO_NODE;
 
1262
1263	lmb_size = drmem_lmb_size();
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1264
1265	for_each_drmem_lmb(lmb) {
1266		/* skip this block if it is reserved or not assigned to
1267		 * this partition */
1268		if ((lmb->flags & DRCONF_MEM_RESERVED)
1269		    || !(lmb->flags & DRCONF_MEM_ASSIGNED))
1270			continue;
1271
1272		if ((scn_addr < lmb->base_addr)
1273		    || (scn_addr >= (lmb->base_addr + lmb_size)))
1274			continue;
1275
1276		nid = of_drconf_to_nid_single(lmb);
1277		break;
1278	}
1279
1280	return nid;
1281}
1282
1283/*
1284 * Find the node associated with a hot added memory section for memory
1285 * represented in the device tree as a node (i.e. memory@XXXX) for
1286 * each memblock.
1287 */
1288static int hot_add_node_scn_to_nid(unsigned long scn_addr)
1289{
1290	struct device_node *memory;
1291	int nid = NUMA_NO_NODE;
1292
1293	for_each_node_by_type(memory, "memory") {
1294		int i = 0;
 
 
 
1295
1296		while (1) {
1297			struct resource res;
 
1298
1299			if (of_address_to_resource(memory, i++, &res))
1300				break;
1301
1302			if ((scn_addr < res.start) || (scn_addr > res.end))
 
 
 
 
1303				continue;
1304
1305			nid = of_node_to_nid_single(memory);
1306			break;
1307		}
1308
 
1309		if (nid >= 0)
1310			break;
1311	}
1312
1313	of_node_put(memory);
1314
1315	return nid;
1316}
1317
1318/*
1319 * Find the node associated with a hot added memory section.  Section
1320 * corresponds to a SPARSEMEM section, not an MEMBLOCK.  It is assumed that
1321 * sections are fully contained within a single MEMBLOCK.
1322 */
1323int hot_add_scn_to_nid(unsigned long scn_addr)
1324{
1325	struct device_node *memory = NULL;
1326	int nid;
1327
1328	if (!numa_enabled)
1329		return first_online_node;
1330
1331	memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1332	if (memory) {
1333		nid = hot_add_drconf_scn_to_nid(scn_addr);
1334		of_node_put(memory);
1335	} else {
1336		nid = hot_add_node_scn_to_nid(scn_addr);
1337	}
1338
1339	if (nid < 0 || !node_possible(nid))
1340		nid = first_online_node;
1341
 
 
 
 
 
 
 
 
 
 
 
1342	return nid;
1343}
1344
1345static u64 hot_add_drconf_memory_max(void)
1346{
1347	struct device_node *memory = NULL;
1348	struct device_node *dn = NULL;
1349	const __be64 *lrdr = NULL;
1350
1351	dn = of_find_node_by_path("/rtas");
1352	if (dn) {
1353		lrdr = of_get_property(dn, "ibm,lrdr-capacity", NULL);
1354		of_node_put(dn);
1355		if (lrdr)
1356			return be64_to_cpup(lrdr);
1357	}
1358
1359	memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1360	if (memory) {
1361		of_node_put(memory);
1362		return drmem_lmb_memory_max();
1363	}
1364	return 0;
1365}
1366
1367/*
1368 * memory_hotplug_max - return max address of memory that may be added
1369 *
1370 * This is currently only used on systems that support drconfig memory
1371 * hotplug.
1372 */
1373u64 memory_hotplug_max(void)
1374{
1375        return max(hot_add_drconf_memory_max(), memblock_end_of_DRAM());
1376}
1377#endif /* CONFIG_MEMORY_HOTPLUG */
1378
1379/* Virtual Processor Home Node (VPHN) support */
1380#ifdef CONFIG_PPC_SPLPAR
1381static int topology_inited;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1382
1383/*
1384 * Retrieve the new associativity information for a virtual processor's
1385 * home node.
1386 */
 
 
 
 
 
 
 
 
 
 
 
 
 
1387static long vphn_get_associativity(unsigned long cpu,
1388					__be32 *associativity)
1389{
1390	long rc;
1391
1392	rc = hcall_vphn(get_hard_smp_processor_id(cpu),
1393				VPHN_FLAG_VCPU, associativity);
1394
1395	switch (rc) {
1396	case H_SUCCESS:
1397		pr_debug("VPHN hcall succeeded. Reset polling...\n");
1398		goto out;
1399
1400	case H_FUNCTION:
1401		pr_err_ratelimited("VPHN unsupported. Disabling polling...\n");
 
 
1402		break;
1403	case H_HARDWARE:
1404		pr_err_ratelimited("hcall_vphn() experienced a hardware fault "
 
1405			"preventing VPHN. Disabling polling...\n");
1406		break;
1407	case H_PARAMETER:
1408		pr_err_ratelimited("hcall_vphn() was passed an invalid parameter. "
1409			"Disabling polling...\n");
1410		break;
1411	default:
1412		pr_err_ratelimited("hcall_vphn() returned %ld. Disabling polling...\n"
1413			, rc);
1414		break;
1415	}
1416out:
1417	return rc;
1418}
1419
1420void find_and_update_cpu_nid(int cpu)
 
 
 
 
1421{
1422	__be32 associativity[VPHN_ASSOC_BUFSIZE] = {0};
1423	int new_nid;
 
1424
1425	/* Use associativity from first thread for all siblings */
1426	if (vphn_get_associativity(cpu, associativity))
1427		return;
1428
1429	/* Do not have previous associativity, so find it now. */
1430	new_nid = associativity_to_nid(associativity);
1431
1432	if (new_nid < 0 || !node_possible(new_nid))
1433		new_nid = first_online_node;
1434	else
1435		// Associate node <-> cpu, so cpu_up() calls
1436		// try_online_node() on the right node.
1437		set_cpu_numa_node(cpu, new_nid);
 
 
 
 
 
 
 
 
 
 
1438
1439	pr_debug("%s:%d cpu %d nid %d\n", __func__, __LINE__, cpu, new_nid);
1440}
1441
1442int cpu_to_coregroup_id(int cpu)
1443{
1444	__be32 associativity[VPHN_ASSOC_BUFSIZE] = {0};
1445	int index;
1446
1447	if (cpu < 0 || cpu > nr_cpu_ids)
1448		return -1;
1449
1450	if (!coregroup_enabled)
1451		goto out;
 
 
1452
1453	if (!firmware_has_feature(FW_FEATURE_VPHN))
1454		goto out;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1455
1456	if (vphn_get_associativity(cpu, associativity))
1457		goto out;
 
 
 
 
1458
1459	index = of_read_number(associativity, 1);
1460	if (index > primary_domain_index + 1)
1461		return of_read_number(&associativity[index - 1], 1);
 
 
 
 
 
 
1462
1463out:
1464	return cpu_to_core_id(cpu);
1465}
 
1466
1467static int topology_update_init(void)
 
 
 
1468{
1469	topology_inited = 1;
1470	return 0;
1471}
1472device_initcall(topology_update_init);
1473#endif /* CONFIG_PPC_SPLPAR */
v3.1
 
   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/module.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 <asm/sparsemem.h>
  26#include <asm/prom.h>
  27#include <asm/system.h>
  28#include <asm/smp.h>
 
  29#include <asm/firmware.h>
  30#include <asm/paca.h>
  31#include <asm/hvcall.h>
 
 
 
 
  32
  33static int numa_enabled = 1;
  34
  35static char *cmdline __initdata;
  36
  37static int numa_debug;
  38#define dbg(args...) if (numa_debug) { printk(KERN_INFO args); }
  39
  40int numa_cpu_lookup_table[NR_CPUS];
  41cpumask_var_t node_to_cpumask_map[MAX_NUMNODES];
  42struct pglist_data *node_data[MAX_NUMNODES];
  43
  44EXPORT_SYMBOL(numa_cpu_lookup_table);
  45EXPORT_SYMBOL(node_to_cpumask_map);
  46EXPORT_SYMBOL(node_data);
  47
  48static int min_common_depth;
  49static int n_mem_addr_cells, n_mem_size_cells;
  50static int form1_affinity;
 
 
 
 
  51
  52#define MAX_DISTANCE_REF_POINTS 4
  53static int distance_ref_points_depth;
  54static const unsigned int *distance_ref_points;
  55static int distance_lookup_table[MAX_NUMNODES][MAX_DISTANCE_REF_POINTS];
 
 
 
 
  56
  57/*
  58 * Allocate node_to_cpumask_map based on number of available nodes
  59 * Requires node_possible_map to be valid.
  60 *
  61 * Note: node_to_cpumask() is not valid until after this is done.
  62 */
  63static void __init setup_node_to_cpumask_map(void)
  64{
  65	unsigned int node, num = 0;
  66
  67	/* setup nr_node_ids if not done yet */
  68	if (nr_node_ids == MAX_NUMNODES) {
  69		for_each_node_mask(node, node_possible_map)
  70			num = node;
  71		nr_node_ids = num + 1;
  72	}
  73
  74	/* allocate the map */
  75	for (node = 0; node < nr_node_ids; node++)
  76		alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]);
  77
  78	/* cpumask_of_node() will now work */
  79	dbg("Node to cpumask map for %d nodes\n", nr_node_ids);
  80}
  81
  82static int __cpuinit fake_numa_create_new_node(unsigned long end_pfn,
  83						unsigned int *nid)
  84{
  85	unsigned long long mem;
  86	char *p = cmdline;
  87	static unsigned int fake_nid;
  88	static unsigned long long curr_boundary;
  89
  90	/*
  91	 * Modify node id, iff we started creating NUMA nodes
  92	 * We want to continue from where we left of the last time
  93	 */
  94	if (fake_nid)
  95		*nid = fake_nid;
  96	/*
  97	 * In case there are no more arguments to parse, the
  98	 * node_id should be the same as the last fake node id
  99	 * (we've handled this above).
 100	 */
 101	if (!p)
 102		return 0;
 103
 104	mem = memparse(p, &p);
 105	if (!mem)
 106		return 0;
 107
 108	if (mem < curr_boundary)
 109		return 0;
 110
 111	curr_boundary = mem;
 112
 113	if ((end_pfn << PAGE_SHIFT) > mem) {
 114		/*
 115		 * Skip commas and spaces
 116		 */
 117		while (*p == ',' || *p == ' ' || *p == '\t')
 118			p++;
 119
 120		cmdline = p;
 121		fake_nid++;
 122		*nid = fake_nid;
 123		dbg("created new fake_node with id %d\n", fake_nid);
 124		return 1;
 125	}
 126	return 0;
 127}
 128
 129/*
 130 * get_active_region_work_fn - A helper function for get_node_active_region
 131 *	Returns datax set to the start_pfn and end_pfn if they contain
 132 *	the initial value of datax->start_pfn between them
 133 * @start_pfn: start page(inclusive) of region to check
 134 * @end_pfn: end page(exclusive) of region to check
 135 * @datax: comes in with ->start_pfn set to value to search for and
 136 *	goes out with active range if it contains it
 137 * Returns 1 if search value is in range else 0
 138 */
 139static int __init get_active_region_work_fn(unsigned long start_pfn,
 140					unsigned long end_pfn, void *datax)
 141{
 142	struct node_active_region *data;
 143	data = (struct node_active_region *)datax;
 144
 145	if (start_pfn <= data->start_pfn && end_pfn > data->start_pfn) {
 146		data->start_pfn = start_pfn;
 147		data->end_pfn = end_pfn;
 148		return 1;
 149	}
 150	return 0;
 151
 
 
 152}
 153
 154/*
 155 * get_node_active_region - Return active region containing start_pfn
 156 * Active range returned is empty if none found.
 157 * @start_pfn: The page to return the region for.
 158 * @node_ar: Returned set to the active region containing start_pfn
 159 */
 160static void __init get_node_active_region(unsigned long start_pfn,
 161		       struct node_active_region *node_ar)
 162{
 163	int nid = early_pfn_to_nid(start_pfn);
 164
 165	node_ar->nid = nid;
 166	node_ar->start_pfn = start_pfn;
 167	node_ar->end_pfn = start_pfn;
 168	work_with_active_regions(nid, get_active_region_work_fn, node_ar);
 169}
 170
 171static void map_cpu_to_node(int cpu, int node)
 172{
 173	numa_cpu_lookup_table[cpu] = node;
 174
 175	dbg("adding cpu %d to node %d\n", cpu, node);
 176
 177	if (!(cpumask_test_cpu(cpu, node_to_cpumask_map[node])))
 178		cpumask_set_cpu(cpu, node_to_cpumask_map[node]);
 
 179}
 180
 181#if defined(CONFIG_HOTPLUG_CPU) || defined(CONFIG_PPC_SPLPAR)
 182static void unmap_cpu_from_node(unsigned long cpu)
 183{
 184	int node = numa_cpu_lookup_table[cpu];
 185
 186	dbg("removing cpu %lu from node %d\n", cpu, node);
 187
 188	if (cpumask_test_cpu(cpu, node_to_cpumask_map[node])) {
 189		cpumask_clear_cpu(cpu, node_to_cpumask_map[node]);
 
 190	} else {
 191		printk(KERN_ERR "WARNING: cpu %lu not found in node %d\n",
 192		       cpu, node);
 193	}
 194}
 195#endif /* CONFIG_HOTPLUG_CPU || CONFIG_PPC_SPLPAR */
 196
 197/* must hold reference to node during call */
 198static const int *of_get_associativity(struct device_node *dev)
 199{
 200	return of_get_property(dev, "ibm,associativity", NULL);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 201}
 202
 203/*
 204 * Returns the property linux,drconf-usable-memory if
 205 * it exists (the property exists only in kexec/kdump kernels,
 206 * added by kexec-tools)
 207 */
 208static const u32 *of_get_usable_memory(struct device_node *memory)
 
 
 
 
 
 
 
 
 209{
 210	const u32 *prop;
 211	u32 len;
 212	prop = of_get_property(memory, "linux,drconf-usable-memory", &len);
 213	if (!prop || len < sizeof(unsigned int))
 
 
 
 
 214		return 0;
 215	return prop;
 
 
 
 216}
 217
 218int __node_distance(int a, int b)
 219{
 220	int i;
 221	int distance = LOCAL_DISTANCE;
 222
 223	if (!form1_affinity)
 224		return distance;
 225
 226	for (i = 0; i < distance_ref_points_depth; i++) {
 227		if (distance_lookup_table[a][i] == distance_lookup_table[b][i])
 
 228			break;
 229
 230		/* Double the distance for each NUMA level */
 231		distance *= 2;
 232	}
 233
 234	return distance;
 235}
 236
 237static void initialize_distance_lookup_table(int nid,
 238		const unsigned int *associativity)
 239{
 240	int i;
 
 
 
 
 
 241
 242	if (!form1_affinity)
 243		return;
 244
 245	for (i = 0; i < distance_ref_points_depth; i++) {
 246		distance_lookup_table[nid][i] =
 247			associativity[distance_ref_points[i]];
 248	}
 249}
 250
 251/* Returns nid in the range [0..MAX_NUMNODES-1], or -1 if no useful numa
 252 * info is found.
 253 */
 254static int associativity_to_nid(const unsigned int *associativity)
 255{
 256	int nid = -1;
 
 257
 258	if (min_common_depth == -1)
 259		goto out;
 
 
 260
 261	if (associativity[0] >= min_common_depth)
 262		nid = associativity[min_common_depth];
 
 263
 264	/* POWER4 LPAR uses 0xffff as invalid node */
 265	if (nid == 0xffff || nid >= MAX_NUMNODES)
 266		nid = -1;
 267
 268	if (nid > 0 && associativity[0] >= distance_ref_points_depth)
 269		initialize_distance_lookup_table(nid, associativity);
 270
 271out:
 272	return nid;
 273}
 
 274
 275/* Returns the nid associated with the given device tree node,
 276 * or -1 if not found.
 277 */
 278static int of_node_to_nid_single(struct device_node *device)
 279{
 280	int nid = -1;
 281	const unsigned int *tmp;
 282
 283	tmp = of_get_associativity(device);
 284	if (tmp)
 285		nid = associativity_to_nid(tmp);
 286	return nid;
 287}
 288
 289/* Walk the device tree upwards, looking for an associativity id */
 290int of_node_to_nid(struct device_node *device)
 291{
 292	struct device_node *tmp;
 293	int nid = -1;
 294
 295	of_node_get(device);
 296	while (device) {
 297		nid = of_node_to_nid_single(device);
 298		if (nid != -1)
 299			break;
 300
 301	        tmp = device;
 302		device = of_get_parent(tmp);
 303		of_node_put(tmp);
 304	}
 305	of_node_put(device);
 306
 307	return nid;
 308}
 309EXPORT_SYMBOL_GPL(of_node_to_nid);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 310
 311static int __init find_min_common_depth(void)
 312{
 313	int depth;
 314	struct device_node *chosen;
 315	struct device_node *root;
 316	const char *vec5;
 317
 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#define VEC5_AFFINITY_BYTE	5
 346#define VEC5_AFFINITY		0x80
 347	chosen = of_find_node_by_path("/chosen");
 348	if (chosen) {
 349		vec5 = of_get_property(chosen, "ibm,architecture-vec-5", NULL);
 350		if (vec5 && (vec5[VEC5_AFFINITY_BYTE] & VEC5_AFFINITY)) {
 351			dbg("Using form 1 affinity\n");
 352			form1_affinity = 1;
 353		}
 354	}
 355
 356	if (form1_affinity) {
 357		depth = distance_ref_points[0];
 358	} else {
 359		if (distance_ref_points_depth < 2) {
 360			printk(KERN_WARNING "NUMA: "
 361				"short ibm,associativity-reference-points\n");
 362			goto err;
 363		}
 364
 365		depth = distance_ref_points[1];
 
 
 
 
 
 
 366	}
 367
 368	/*
 369	 * Warn and cap if the hardware supports more than
 370	 * MAX_DISTANCE_REF_POINTS domains.
 371	 */
 372	if (distance_ref_points_depth > MAX_DISTANCE_REF_POINTS) {
 373		printk(KERN_WARNING "NUMA: distance array capped at "
 374			"%d entries\n", MAX_DISTANCE_REF_POINTS);
 375		distance_ref_points_depth = MAX_DISTANCE_REF_POINTS;
 376	}
 377
 378	of_node_put(root);
 379	return depth;
 380
 381err:
 382	of_node_put(root);
 383	return -1;
 384}
 385
 386static void __init get_n_mem_cells(int *n_addr_cells, int *n_size_cells)
 387{
 388	struct device_node *memory = NULL;
 389
 390	memory = of_find_node_by_type(memory, "memory");
 391	if (!memory)
 392		panic("numa.c: No memory nodes found!");
 393
 394	*n_addr_cells = of_n_addr_cells(memory);
 395	*n_size_cells = of_n_size_cells(memory);
 396	of_node_put(memory);
 397}
 398
 399static unsigned long __devinit read_n_cells(int n, const unsigned int **buf)
 400{
 401	unsigned long result = 0;
 402
 403	while (n--) {
 404		result = (result << 32) | **buf;
 405		(*buf)++;
 406	}
 407	return result;
 408}
 409
 410struct of_drconf_cell {
 411	u64	base_addr;
 412	u32	drc_index;
 413	u32	reserved;
 414	u32	aa_index;
 415	u32	flags;
 416};
 417
 418#define DRCONF_MEM_ASSIGNED	0x00000008
 419#define DRCONF_MEM_AI_INVALID	0x00000040
 420#define DRCONF_MEM_RESERVED	0x00000080
 421
 422/*
 423 * Read the next memblock list entry from the ibm,dynamic-memory property
 424 * and return the information in the provided of_drconf_cell structure.
 425 */
 426static void read_drconf_cell(struct of_drconf_cell *drmem, const u32 **cellp)
 427{
 428	const u32 *cp;
 429
 430	drmem->base_addr = read_n_cells(n_mem_addr_cells, cellp);
 431
 432	cp = *cellp;
 433	drmem->drc_index = cp[0];
 434	drmem->reserved = cp[1];
 435	drmem->aa_index = cp[2];
 436	drmem->flags = cp[3];
 437
 438	*cellp = cp + 4;
 439}
 440
 441/*
 442 * Retrieve and validate the ibm,dynamic-memory property of the device tree.
 443 *
 444 * The layout of the ibm,dynamic-memory property is a number N of memblock
 445 * list entries followed by N memblock list entries.  Each memblock list entry
 446 * contains information as laid out in the of_drconf_cell struct above.
 447 */
 448static int of_get_drconf_memory(struct device_node *memory, const u32 **dm)
 449{
 450	const u32 *prop;
 451	u32 len, entries;
 452
 453	prop = of_get_property(memory, "ibm,dynamic-memory", &len);
 454	if (!prop || len < sizeof(unsigned int))
 455		return 0;
 456
 457	entries = *prop++;
 458
 459	/* Now that we know the number of entries, revalidate the size
 460	 * of the property read in to ensure we have everything
 461	 */
 462	if (len < (entries * (n_mem_addr_cells + 4) + 1) * sizeof(unsigned int))
 463		return 0;
 464
 465	*dm = prop;
 466	return entries;
 467}
 468
 469/*
 470 * Retrieve and validate the ibm,lmb-size property for drconf memory
 471 * from the device tree.
 472 */
 473static u64 of_get_lmb_size(struct device_node *memory)
 474{
 475	const u32 *prop;
 476	u32 len;
 477
 478	prop = of_get_property(memory, "ibm,lmb-size", &len);
 479	if (!prop || len < sizeof(unsigned int))
 480		return 0;
 481
 482	return read_n_cells(n_mem_size_cells, &prop);
 483}
 484
 485struct assoc_arrays {
 486	u32	n_arrays;
 487	u32	array_sz;
 488	const u32 *arrays;
 489};
 490
 491/*
 492 * Retrieve and validate the list of associativity arrays for drconf
 493 * memory from the ibm,associativity-lookup-arrays property of the
 494 * device tree..
 495 *
 496 * The layout of the ibm,associativity-lookup-arrays property is a number N
 497 * indicating the number of associativity arrays, followed by a number M
 498 * indicating the size of each associativity array, followed by a list
 499 * of N associativity arrays.
 500 */
 501static int of_get_assoc_arrays(struct device_node *memory,
 502			       struct assoc_arrays *aa)
 503{
 504	const u32 *prop;
 
 505	u32 len;
 506
 
 
 
 
 507	prop = of_get_property(memory, "ibm,associativity-lookup-arrays", &len);
 508	if (!prop || len < 2 * sizeof(unsigned int))
 
 509		return -1;
 
 510
 511	aa->n_arrays = *prop++;
 512	aa->array_sz = *prop++;
 513
 514	/* Now that we know the number of arrrays and size of each array,
 
 
 515	 * revalidate the size of the property read in.
 516	 */
 517	if (len < (aa->n_arrays * aa->array_sz + 2) * sizeof(unsigned int))
 518		return -1;
 519
 520	aa->arrays = prop;
 521	return 0;
 522}
 523
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 524/*
 525 * This is like of_node_to_nid_single() for memory represented in the
 526 * ibm,dynamic-reconfiguration-memory node.
 527 */
 528static int of_drconf_to_nid_single(struct of_drconf_cell *drmem,
 529				   struct assoc_arrays *aa)
 530{
 531	int default_nid = 0;
 
 532	int nid = default_nid;
 533	int index;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 534
 535	if (min_common_depth > 0 && min_common_depth <= aa->array_sz &&
 536	    !(drmem->flags & DRCONF_MEM_AI_INVALID) &&
 537	    drmem->aa_index < aa->n_arrays) {
 538		index = drmem->aa_index * aa->array_sz + min_common_depth - 1;
 539		nid = aa->arrays[index];
 540
 541		if (nid == 0xffff || nid >= MAX_NUMNODES)
 542			nid = default_nid;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 543	}
 544
 545	return nid;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 546}
 
 547
 548/*
 549 * Figure out to which domain a cpu belongs and stick it there.
 550 * Return the id of the domain used.
 551 */
 552static int __cpuinit numa_setup_cpu(unsigned long lcpu)
 553{
 554	int nid = 0;
 555	struct device_node *cpu = of_get_cpu_node(lcpu, NULL);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 556
 557	if (!cpu) {
 558		WARN_ON(1);
 559		goto out;
 
 
 
 560	}
 561
 562	nid = of_node_to_nid_single(cpu);
 
 563
 564	if (nid < 0 || !node_online(nid))
 
 565		nid = first_online_node;
 566out:
 567	map_cpu_to_node(lcpu, nid);
 568
 569	of_node_put(cpu);
 
 
 
 
 
 
 
 
 
 
 
 570
 
 
 571	return nid;
 572}
 573
 574static int __cpuinit cpu_numa_callback(struct notifier_block *nfb,
 575			     unsigned long action,
 576			     void *hcpu)
 577{
 578	unsigned long lcpu = (unsigned long)hcpu;
 579	int ret = NOTIFY_DONE;
 580
 581	switch (action) {
 582	case CPU_UP_PREPARE:
 583	case CPU_UP_PREPARE_FROZEN:
 584		numa_setup_cpu(lcpu);
 585		ret = NOTIFY_OK;
 586		break;
 587#ifdef CONFIG_HOTPLUG_CPU
 588	case CPU_DEAD:
 589	case CPU_DEAD_FROZEN:
 590	case CPU_UP_CANCELED:
 591	case CPU_UP_CANCELED_FROZEN:
 592		unmap_cpu_from_node(lcpu);
 593		break;
 594		ret = NOTIFY_OK;
 595#endif
 596	}
 597	return ret;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 598}
 599
 600/*
 601 * Check and possibly modify a memory region to enforce the memory limit.
 602 *
 603 * Returns the size the region should have to enforce the memory limit.
 604 * This will either be the original value of size, a truncated value,
 605 * or zero. If the returned value of size is 0 the region should be
 606 * discarded as it lies wholly above the memory limit.
 607 */
 608static unsigned long __init numa_enforce_memory_limit(unsigned long start,
 609						      unsigned long size)
 610{
 611	/*
 612	 * We use memblock_end_of_DRAM() in here instead of memory_limit because
 613	 * we've already adjusted it for the limit and it takes care of
 614	 * having memory holes below the limit.  Also, in the case of
 615	 * iommu_is_off, memory_limit is not set but is implicitly enforced.
 616	 */
 617
 618	if (start + size <= memblock_end_of_DRAM())
 619		return size;
 620
 621	if (start >= memblock_end_of_DRAM())
 622		return 0;
 623
 624	return memblock_end_of_DRAM() - start;
 625}
 626
 627/*
 628 * Reads the counter for a given entry in
 629 * linux,drconf-usable-memory property
 630 */
 631static inline int __init read_usm_ranges(const u32 **usm)
 632{
 633	/*
 634	 * For each lmb in ibm,dynamic-memory a corresponding
 635	 * entry in linux,drconf-usable-memory property contains
 636	 * a counter followed by that many (base, size) duple.
 637	 * read the counter from linux,drconf-usable-memory
 638	 */
 639	return read_n_cells(n_mem_size_cells, usm);
 640}
 641
 642/*
 643 * Extract NUMA information from the ibm,dynamic-reconfiguration-memory
 644 * node.  This assumes n_mem_{addr,size}_cells have been set.
 645 */
 646static void __init parse_drconf_memory(struct device_node *memory)
 
 
 647{
 648	const u32 *dm, *usm;
 649	unsigned int n, rc, ranges, is_kexec_kdump = 0;
 650	unsigned long lmb_size, base, size, sz;
 651	int nid;
 652	struct assoc_arrays aa;
 653
 654	n = of_get_drconf_memory(memory, &dm);
 655	if (!n)
 656		return;
 
 
 
 
 657
 658	lmb_size = of_get_lmb_size(memory);
 659	if (!lmb_size)
 660		return;
 661
 662	rc = of_get_assoc_arrays(memory, &aa);
 663	if (rc)
 664		return;
 665
 666	/* check if this is a kexec/kdump kernel */
 667	usm = of_get_usable_memory(memory);
 668	if (usm != NULL)
 669		is_kexec_kdump = 1;
 670
 671	for (; n != 0; --n) {
 672		struct of_drconf_cell drmem;
 
 673
 674		read_drconf_cell(&drmem, &dm);
 
 
 
 
 675
 676		/* skip this block if the reserved bit is set in flags (0x80)
 677		   or if the block is not assigned to this partition (0x8) */
 678		if ((drmem.flags & DRCONF_MEM_RESERVED)
 679		    || !(drmem.flags & DRCONF_MEM_ASSIGNED))
 680			continue;
 681
 682		base = drmem.base_addr;
 683		size = lmb_size;
 684		ranges = 1;
 
 
 
 
 
 685
 686		if (is_kexec_kdump) {
 687			ranges = read_usm_ranges(&usm);
 688			if (!ranges) /* there are no (base, size) duple */
 689				continue;
 690		}
 691		do {
 692			if (is_kexec_kdump) {
 693				base = read_n_cells(n_mem_addr_cells, &usm);
 694				size = read_n_cells(n_mem_size_cells, &usm);
 695			}
 696			nid = of_drconf_to_nid_single(&drmem, &aa);
 697			fake_numa_create_new_node(
 698				((base + size) >> PAGE_SHIFT),
 699					   &nid);
 700			node_set_online(nid);
 701			sz = numa_enforce_memory_limit(base, size);
 702			if (sz)
 703				add_active_range(nid, base >> PAGE_SHIFT,
 704						 (base >> PAGE_SHIFT)
 705						 + (sz >> PAGE_SHIFT));
 706		} while (--ranges);
 707	}
 708}
 709
 710static int __init parse_numa_properties(void)
 711{
 712	struct device_node *cpu = NULL;
 713	struct device_node *memory = NULL;
 714	int default_nid = 0;
 715	unsigned long i;
 
 716
 717	if (numa_enabled == 0) {
 718		printk(KERN_WARNING "NUMA disabled by user\n");
 719		return -1;
 720	}
 721
 722	min_common_depth = find_min_common_depth();
 723
 724	if (min_common_depth < 0)
 725		return min_common_depth;
 
 
 
 
 
 
 
 
 726
 727	dbg("NUMA associativity depth for CPU/Memory: %d\n", min_common_depth);
 
 
 
 
 728
 729	/*
 730	 * Even though we connect cpus to numa domains later in SMP
 731	 * init, we need to know the node ids now. This is because
 732	 * each node to be onlined must have NODE_DATA etc backing it.
 733	 */
 734	for_each_present_cpu(i) {
 735		int nid;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 736
 737		cpu = of_get_cpu_node(i, NULL);
 738		BUG_ON(!cpu);
 739		nid = of_node_to_nid_single(cpu);
 740		of_node_put(cpu);
 
 
 
 741
 742		/*
 743		 * Don't fall back to default_nid yet -- we will plug
 744		 * cpus into nodes once the memory scan has discovered
 745		 * the topology.
 746		 */
 747		if (nid < 0)
 748			continue;
 749		node_set_online(nid);
 750	}
 751
 752	get_n_mem_cells(&n_mem_addr_cells, &n_mem_size_cells);
 753	memory = NULL;
 754	while ((memory = of_find_node_by_type(memory, "memory")) != NULL) {
 755		unsigned long start;
 756		unsigned long size;
 757		int nid;
 758		int ranges;
 759		const unsigned int *memcell_buf;
 760		unsigned int len;
 761
 762		memcell_buf = of_get_property(memory,
 763			"linux,usable-memory", &len);
 764		if (!memcell_buf || len <= 0)
 765			memcell_buf = of_get_property(memory, "reg", &len);
 766		if (!memcell_buf || len <= 0)
 767			continue;
 768
 769		/* ranges in cell */
 770		ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
 771new_range:
 772		/* these are order-sensitive, and modify the buffer pointer */
 773		start = read_n_cells(n_mem_addr_cells, &memcell_buf);
 774		size = read_n_cells(n_mem_size_cells, &memcell_buf);
 775
 776		/*
 777		 * Assumption: either all memory nodes or none will
 778		 * have associativity properties.  If none, then
 779		 * everything goes to default_nid.
 780		 */
 781		nid = of_node_to_nid_single(memory);
 782		if (nid < 0)
 
 
 
 783			nid = default_nid;
 784
 785		fake_numa_create_new_node(((start + size) >> PAGE_SHIFT), &nid);
 786		node_set_online(nid);
 787
 788		if (!(size = numa_enforce_memory_limit(start, size))) {
 789			if (--ranges)
 790				goto new_range;
 791			else
 792				continue;
 793		}
 794
 795		add_active_range(nid, start >> PAGE_SHIFT,
 796				(start >> PAGE_SHIFT) + (size >> PAGE_SHIFT));
 797
 798		if (--ranges)
 799			goto new_range;
 800	}
 801
 802	/*
 803	 * Now do the same thing for each MEMBLOCK listed in the ibm,dynamic-memory
 804	 * property in the ibm,dynamic-reconfiguration-memory node.
 
 805	 */
 806	memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
 807	if (memory)
 808		parse_drconf_memory(memory);
 
 
 809
 810	return 0;
 811}
 812
 813static void __init setup_nonnuma(void)
 814{
 815	unsigned long top_of_ram = memblock_end_of_DRAM();
 816	unsigned long total_ram = memblock_phys_mem_size();
 817	unsigned long start_pfn, end_pfn;
 818	unsigned int nid = 0;
 819	struct memblock_region *reg;
 820
 821	printk(KERN_DEBUG "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
 822	       top_of_ram, total_ram);
 823	printk(KERN_DEBUG "Memory hole size: %ldMB\n",
 824	       (top_of_ram - total_ram) >> 20);
 825
 826	for_each_memblock(memory, reg) {
 827		start_pfn = memblock_region_memory_base_pfn(reg);
 828		end_pfn = memblock_region_memory_end_pfn(reg);
 829
 
 830		fake_numa_create_new_node(end_pfn, &nid);
 831		add_active_range(nid, start_pfn, end_pfn);
 
 
 832		node_set_online(nid);
 833	}
 834}
 835
 836void __init dump_numa_cpu_topology(void)
 837{
 838	unsigned int node;
 839	unsigned int cpu, count;
 840
 841	if (min_common_depth == -1 || !numa_enabled)
 842		return;
 843
 844	for_each_online_node(node) {
 845		printk(KERN_DEBUG "Node %d CPUs:", node);
 846
 847		count = 0;
 848		/*
 849		 * If we used a CPU iterator here we would miss printing
 850		 * the holes in the cpumap.
 851		 */
 852		for (cpu = 0; cpu < nr_cpu_ids; cpu++) {
 853			if (cpumask_test_cpu(cpu,
 854					node_to_cpumask_map[node])) {
 855				if (count == 0)
 856					printk(" %u", cpu);
 857				++count;
 858			} else {
 859				if (count > 1)
 860					printk("-%u", cpu - 1);
 861				count = 0;
 862			}
 863		}
 864
 865		if (count > 1)
 866			printk("-%u", nr_cpu_ids - 1);
 867		printk("\n");
 868	}
 869}
 870
 871static void __init dump_numa_memory_topology(void)
 
 872{
 873	unsigned int node;
 874	unsigned int count;
 
 
 
 875
 876	if (min_common_depth == -1 || !numa_enabled)
 877		return;
 
 
 878
 879	for_each_online_node(node) {
 880		unsigned long i;
 881
 882		printk(KERN_DEBUG "Node %d Memory:", node);
 
 
 
 
 
 883
 884		count = 0;
 
 
 
 
 
 885
 886		for (i = 0; i < memblock_end_of_DRAM();
 887		     i += (1 << SECTION_SIZE_BITS)) {
 888			if (early_pfn_to_nid(i >> PAGE_SHIFT) == node) {
 889				if (count == 0)
 890					printk(" 0x%lx", i);
 891				++count;
 892			} else {
 893				if (count > 0)
 894					printk("-0x%lx", i);
 895				count = 0;
 896			}
 897		}
 898
 899		if (count > 0)
 900			printk("-0x%lx", i);
 901		printk("\n");
 902	}
 903}
 904
 905/*
 906 * Allocate some memory, satisfying the memblock or bootmem allocator where
 907 * required. nid is the preferred node and end is the physical address of
 908 * the highest address in the node.
 909 *
 910 * Returns the virtual address of the memory.
 911 */
 912static void __init *careful_zallocation(int nid, unsigned long size,
 913				       unsigned long align,
 914				       unsigned long end_pfn)
 915{
 916	void *ret;
 917	int new_nid;
 918	unsigned long ret_paddr;
 919
 920	ret_paddr = __memblock_alloc_base(size, align, end_pfn << PAGE_SHIFT);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 921
 922	/* retry over all memory */
 923	if (!ret_paddr)
 924		ret_paddr = __memblock_alloc_base(size, align, memblock_end_of_DRAM());
 925
 926	if (!ret_paddr)
 927		panic("numa.c: cannot allocate %lu bytes for node %d",
 928		      size, nid);
 929
 930	ret = __va(ret_paddr);
 
 931
 932	/*
 933	 * We initialize the nodes in numeric order: 0, 1, 2...
 934	 * and hand over control from the MEMBLOCK allocator to the
 935	 * bootmem allocator.  If this function is called for
 936	 * node 5, then we know that all nodes <5 are using the
 937	 * bootmem allocator instead of the MEMBLOCK allocator.
 938	 *
 939	 * So, check the nid from which this allocation came
 940	 * and double check to see if we need to use bootmem
 941	 * instead of the MEMBLOCK.  We don't free the MEMBLOCK memory
 942	 * since it would be useless.
 943	 */
 944	new_nid = early_pfn_to_nid(ret_paddr >> PAGE_SHIFT);
 945	if (new_nid < nid) {
 946		ret = __alloc_bootmem_node(NODE_DATA(new_nid),
 947				size, align, 0);
 948
 949		dbg("alloc_bootmem %p %lx\n", ret, size);
 950	}
 951
 952	memset(ret, 0, size);
 953	return ret;
 954}
 
 
 
 
 955
 956static struct notifier_block __cpuinitdata ppc64_numa_nb = {
 957	.notifier_call = cpu_numa_callback,
 958	.priority = 1 /* Must run before sched domains notifier. */
 959};
 960
 961static void mark_reserved_regions_for_nid(int nid)
 962{
 963	struct pglist_data *node = NODE_DATA(nid);
 964	struct memblock_region *reg;
 965
 966	for_each_memblock(reserved, reg) {
 967		unsigned long physbase = reg->base;
 968		unsigned long size = reg->size;
 969		unsigned long start_pfn = physbase >> PAGE_SHIFT;
 970		unsigned long end_pfn = PFN_UP(physbase + size);
 971		struct node_active_region node_ar;
 972		unsigned long node_end_pfn = node->node_start_pfn +
 973					     node->node_spanned_pages;
 974
 
 975		/*
 976		 * Check to make sure that this memblock.reserved area is
 977		 * within the bounds of the node that we care about.
 978		 * Checking the nid of the start and end points is not
 979		 * sufficient because the reserved area could span the
 980		 * entire node.
 
 981		 */
 982		if (end_pfn <= node->node_start_pfn ||
 983		    start_pfn >= node_end_pfn)
 984			continue;
 985
 986		get_node_active_region(start_pfn, &node_ar);
 987		while (start_pfn < end_pfn &&
 988			node_ar.start_pfn < node_ar.end_pfn) {
 989			unsigned long reserve_size = size;
 990			/*
 991			 * if reserved region extends past active region
 992			 * then trim size to active region
 993			 */
 994			if (end_pfn > node_ar.end_pfn)
 995				reserve_size = (node_ar.end_pfn << PAGE_SHIFT)
 996					- physbase;
 997			/*
 998			 * Only worry about *this* node, others may not
 999			 * yet have valid NODE_DATA().
1000			 */
1001			if (node_ar.nid == nid) {
1002				dbg("reserve_bootmem %lx %lx nid=%d\n",
1003					physbase, reserve_size, node_ar.nid);
1004				reserve_bootmem_node(NODE_DATA(node_ar.nid),
1005						physbase, reserve_size,
1006						BOOTMEM_DEFAULT);
1007			}
1008			/*
1009			 * if reserved region is contained in the active region
1010			 * then done.
1011			 */
1012			if (end_pfn <= node_ar.end_pfn)
1013				break;
1014
1015			/*
1016			 * reserved region extends past the active region
1017			 *   get next active region that contains this
1018			 *   reserved region
1019			 */
1020			start_pfn = node_ar.end_pfn;
1021			physbase = start_pfn << PAGE_SHIFT;
1022			size = size - reserve_size;
1023			get_node_active_region(start_pfn, &node_ar);
1024		}
1025	}
1026}
1027
1028
1029void __init do_init_bootmem(void)
1030{
1031	int nid;
1032
1033	min_low_pfn = 0;
1034	max_low_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
1035	max_pfn = max_low_pfn;
1036
1037	if (parse_numa_properties())
1038		setup_nonnuma();
1039	else
1040		dump_numa_memory_topology();
1041
1042	for_each_online_node(nid) {
1043		unsigned long start_pfn, end_pfn;
1044		void *bootmem_vaddr;
1045		unsigned long bootmap_pages;
1046
1047		get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
1048
1049		/*
1050		 * Allocate the node structure node local if possible
1051		 *
1052		 * Be careful moving this around, as it relies on all
1053		 * previous nodes' bootmem to be initialized and have
1054		 * all reserved areas marked.
1055		 */
1056		NODE_DATA(nid) = careful_zallocation(nid,
1057					sizeof(struct pglist_data),
1058					SMP_CACHE_BYTES, end_pfn);
1059
1060  		dbg("node %d\n", nid);
1061		dbg("NODE_DATA() = %p\n", NODE_DATA(nid));
1062
1063		NODE_DATA(nid)->bdata = &bootmem_node_data[nid];
1064		NODE_DATA(nid)->node_start_pfn = start_pfn;
1065		NODE_DATA(nid)->node_spanned_pages = end_pfn - start_pfn;
1066
1067		if (NODE_DATA(nid)->node_spanned_pages == 0)
1068  			continue;
1069
1070  		dbg("start_paddr = %lx\n", start_pfn << PAGE_SHIFT);
1071  		dbg("end_paddr = %lx\n", end_pfn << PAGE_SHIFT);
1072
1073		bootmap_pages = bootmem_bootmap_pages(end_pfn - start_pfn);
1074		bootmem_vaddr = careful_zallocation(nid,
1075					bootmap_pages << PAGE_SHIFT,
1076					PAGE_SIZE, end_pfn);
1077
1078		dbg("bootmap_vaddr = %p\n", bootmem_vaddr);
1079
1080		init_bootmem_node(NODE_DATA(nid),
1081				  __pa(bootmem_vaddr) >> PAGE_SHIFT,
1082				  start_pfn, end_pfn);
1083
1084		free_bootmem_with_active_regions(nid, end_pfn);
1085		/*
1086		 * Be very careful about moving this around.  Future
1087		 * calls to careful_zallocation() depend on this getting
1088		 * done correctly.
1089		 */
1090		mark_reserved_regions_for_nid(nid);
1091		sparse_memory_present_with_active_regions(nid);
1092	}
1093
1094	init_bootmem_done = 1;
1095
1096	/*
1097	 * Now bootmem is initialised we can create the node to cpumask
1098	 * lookup tables and setup the cpu callback to populate them.
 
 
 
1099	 */
1100	setup_node_to_cpumask_map();
1101
1102	register_cpu_notifier(&ppc64_numa_nb);
1103	cpu_numa_callback(&ppc64_numa_nb, CPU_UP_PREPARE,
1104			  (void *)(unsigned long)boot_cpuid);
1105}
1106
1107void __init paging_init(void)
1108{
1109	unsigned long max_zone_pfns[MAX_NR_ZONES];
1110	memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
1111	max_zone_pfns[ZONE_DMA] = memblock_end_of_DRAM() >> PAGE_SHIFT;
1112	free_area_init_nodes(max_zone_pfns);
1113}
1114
1115static int __init early_numa(char *p)
1116{
1117	if (!p)
1118		return 0;
1119
1120	if (strstr(p, "off"))
1121		numa_enabled = 0;
1122
1123	if (strstr(p, "debug"))
1124		numa_debug = 1;
1125
1126	p = strstr(p, "fake=");
1127	if (p)
1128		cmdline = p + strlen("fake=");
1129
1130	return 0;
1131}
1132early_param("numa", early_numa);
1133
1134#ifdef CONFIG_MEMORY_HOTPLUG
1135/*
1136 * Find the node associated with a hot added memory section for
1137 * memory represented in the device tree by the property
1138 * ibm,dynamic-reconfiguration-memory/ibm,dynamic-memory.
1139 */
1140static int hot_add_drconf_scn_to_nid(struct device_node *memory,
1141				     unsigned long scn_addr)
1142{
1143	const u32 *dm;
1144	unsigned int drconf_cell_cnt, rc;
1145	unsigned long lmb_size;
1146	struct assoc_arrays aa;
1147	int nid = -1;
1148
1149	drconf_cell_cnt = of_get_drconf_memory(memory, &dm);
1150	if (!drconf_cell_cnt)
1151		return -1;
1152
1153	lmb_size = of_get_lmb_size(memory);
1154	if (!lmb_size)
1155		return -1;
1156
1157	rc = of_get_assoc_arrays(memory, &aa);
1158	if (rc)
1159		return -1;
1160
1161	for (; drconf_cell_cnt != 0; --drconf_cell_cnt) {
1162		struct of_drconf_cell drmem;
1163
1164		read_drconf_cell(&drmem, &dm);
1165
 
1166		/* skip this block if it is reserved or not assigned to
1167		 * this partition */
1168		if ((drmem.flags & DRCONF_MEM_RESERVED)
1169		    || !(drmem.flags & DRCONF_MEM_ASSIGNED))
1170			continue;
1171
1172		if ((scn_addr < drmem.base_addr)
1173		    || (scn_addr >= (drmem.base_addr + lmb_size)))
1174			continue;
1175
1176		nid = of_drconf_to_nid_single(&drmem, &aa);
1177		break;
1178	}
1179
1180	return nid;
1181}
1182
1183/*
1184 * Find the node associated with a hot added memory section for memory
1185 * represented in the device tree as a node (i.e. memory@XXXX) for
1186 * each memblock.
1187 */
1188int hot_add_node_scn_to_nid(unsigned long scn_addr)
1189{
1190	struct device_node *memory = NULL;
1191	int nid = -1;
1192
1193	while ((memory = of_find_node_by_type(memory, "memory")) != NULL) {
1194		unsigned long start, size;
1195		int ranges;
1196		const unsigned int *memcell_buf;
1197		unsigned int len;
1198
1199		memcell_buf = of_get_property(memory, "reg", &len);
1200		if (!memcell_buf || len <= 0)
1201			continue;
1202
1203		/* ranges in cell */
1204		ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
1205
1206		while (ranges--) {
1207			start = read_n_cells(n_mem_addr_cells, &memcell_buf);
1208			size = read_n_cells(n_mem_size_cells, &memcell_buf);
1209
1210			if ((scn_addr < start) || (scn_addr >= (start + size)))
1211				continue;
1212
1213			nid = of_node_to_nid_single(memory);
1214			break;
1215		}
1216
1217		of_node_put(memory);
1218		if (nid >= 0)
1219			break;
1220	}
1221
 
 
1222	return nid;
1223}
1224
1225/*
1226 * Find the node associated with a hot added memory section.  Section
1227 * corresponds to a SPARSEMEM section, not an MEMBLOCK.  It is assumed that
1228 * sections are fully contained within a single MEMBLOCK.
1229 */
1230int hot_add_scn_to_nid(unsigned long scn_addr)
1231{
1232	struct device_node *memory = NULL;
1233	int nid, found = 0;
1234
1235	if (!numa_enabled || (min_common_depth < 0))
1236		return first_online_node;
1237
1238	memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1239	if (memory) {
1240		nid = hot_add_drconf_scn_to_nid(memory, scn_addr);
1241		of_node_put(memory);
1242	} else {
1243		nid = hot_add_node_scn_to_nid(scn_addr);
1244	}
1245
1246	if (nid < 0 || !node_online(nid))
1247		nid = first_online_node;
1248
1249	if (NODE_DATA(nid)->node_spanned_pages)
1250		return nid;
1251
1252	for_each_online_node(nid) {
1253		if (NODE_DATA(nid)->node_spanned_pages) {
1254			found = 1;
1255			break;
1256		}
1257	}
1258
1259	BUG_ON(!found);
1260	return nid;
1261}
1262
1263static u64 hot_add_drconf_memory_max(void)
1264{
1265        struct device_node *memory = NULL;
1266        unsigned int drconf_cell_cnt = 0;
1267        u64 lmb_size = 0;
1268        const u32 *dm = 0;
1269
1270        memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1271        if (memory) {
1272                drconf_cell_cnt = of_get_drconf_memory(memory, &dm);
1273                lmb_size = of_get_lmb_size(memory);
1274                of_node_put(memory);
1275        }
1276        return lmb_size * drconf_cell_cnt;
 
 
 
 
 
 
1277}
1278
1279/*
1280 * memory_hotplug_max - return max address of memory that may be added
1281 *
1282 * This is currently only used on systems that support drconfig memory
1283 * hotplug.
1284 */
1285u64 memory_hotplug_max(void)
1286{
1287        return max(hot_add_drconf_memory_max(), memblock_end_of_DRAM());
1288}
1289#endif /* CONFIG_MEMORY_HOTPLUG */
1290
1291/* Virtual Processor Home Node (VPHN) support */
1292#ifdef CONFIG_PPC_SPLPAR
1293static u8 vphn_cpu_change_counts[NR_CPUS][MAX_DISTANCE_REF_POINTS];
1294static cpumask_t cpu_associativity_changes_mask;
1295static int vphn_enabled;
1296static void set_topology_timer(void);
1297
1298/*
1299 * Store the current values of the associativity change counters in the
1300 * hypervisor.
1301 */
1302static void setup_cpu_associativity_change_counters(void)
1303{
1304	int cpu;
1305
1306	/* The VPHN feature supports a maximum of 8 reference points */
1307	BUILD_BUG_ON(MAX_DISTANCE_REF_POINTS > 8);
1308
1309	for_each_possible_cpu(cpu) {
1310		int i;
1311		u8 *counts = vphn_cpu_change_counts[cpu];
1312		volatile u8 *hypervisor_counts = lppaca[cpu].vphn_assoc_counts;
1313
1314		for (i = 0; i < distance_ref_points_depth; i++)
1315			counts[i] = hypervisor_counts[i];
1316	}
1317}
1318
1319/*
1320 * The hypervisor maintains a set of 8 associativity change counters in
1321 * the VPA of each cpu that correspond to the associativity levels in the
1322 * ibm,associativity-reference-points property. When an associativity
1323 * level changes, the corresponding counter is incremented.
1324 *
1325 * Set a bit in cpu_associativity_changes_mask for each cpu whose home
1326 * node associativity levels have changed.
1327 *
1328 * Returns the number of cpus with unhandled associativity changes.
1329 */
1330static int update_cpu_associativity_changes_mask(void)
1331{
1332	int cpu, nr_cpus = 0;
1333	cpumask_t *changes = &cpu_associativity_changes_mask;
1334
1335	cpumask_clear(changes);
1336
1337	for_each_possible_cpu(cpu) {
1338		int i, changed = 0;
1339		u8 *counts = vphn_cpu_change_counts[cpu];
1340		volatile u8 *hypervisor_counts = lppaca[cpu].vphn_assoc_counts;
1341
1342		for (i = 0; i < distance_ref_points_depth; i++) {
1343			if (hypervisor_counts[i] != counts[i]) {
1344				counts[i] = hypervisor_counts[i];
1345				changed = 1;
1346			}
1347		}
1348		if (changed) {
1349			cpumask_set_cpu(cpu, changes);
1350			nr_cpus++;
1351		}
1352	}
1353
1354	return nr_cpus;
1355}
1356
1357/*
1358 * 6 64-bit registers unpacked into 12 32-bit associativity values. To form
1359 * the complete property we have to add the length in the first cell.
1360 */
1361#define VPHN_ASSOC_BUFSIZE (6*sizeof(u64)/sizeof(u32) + 1)
1362
1363/*
1364 * Convert the associativity domain numbers returned from the hypervisor
1365 * to the sequence they would appear in the ibm,associativity property.
1366 */
1367static int vphn_unpack_associativity(const long *packed, unsigned int *unpacked)
1368{
1369	int i, nr_assoc_doms = 0;
1370	const u16 *field = (const u16*) packed;
1371
1372#define VPHN_FIELD_UNUSED	(0xffff)
1373#define VPHN_FIELD_MSB		(0x8000)
1374#define VPHN_FIELD_MASK		(~VPHN_FIELD_MSB)
1375
1376	for (i = 1; i < VPHN_ASSOC_BUFSIZE; i++) {
1377		if (*field == VPHN_FIELD_UNUSED) {
1378			/* All significant fields processed, and remaining
1379			 * fields contain the reserved value of all 1's.
1380			 * Just store them.
1381			 */
1382			unpacked[i] = *((u32*)field);
1383			field += 2;
1384		} else if (*field & VPHN_FIELD_MSB) {
1385			/* Data is in the lower 15 bits of this field */
1386			unpacked[i] = *field & VPHN_FIELD_MASK;
1387			field++;
1388			nr_assoc_doms++;
1389		} else {
1390			/* Data is in the lower 15 bits of this field
1391			 * concatenated with the next 16 bit field
1392			 */
1393			unpacked[i] = *((u32*)field);
1394			field += 2;
1395			nr_assoc_doms++;
1396		}
1397	}
1398
1399	/* The first cell contains the length of the property */
1400	unpacked[0] = nr_assoc_doms;
1401
1402	return nr_assoc_doms;
1403}
1404
1405/*
1406 * Retrieve the new associativity information for a virtual processor's
1407 * home node.
1408 */
1409static long hcall_vphn(unsigned long cpu, unsigned int *associativity)
1410{
1411	long rc;
1412	long retbuf[PLPAR_HCALL9_BUFSIZE] = {0};
1413	u64 flags = 1;
1414	int hwcpu = get_hard_smp_processor_id(cpu);
1415
1416	rc = plpar_hcall9(H_HOME_NODE_ASSOCIATIVITY, retbuf, flags, hwcpu);
1417	vphn_unpack_associativity(retbuf, associativity);
1418
1419	return rc;
1420}
1421
1422static long vphn_get_associativity(unsigned long cpu,
1423					unsigned int *associativity)
1424{
1425	long rc;
1426
1427	rc = hcall_vphn(cpu, associativity);
 
1428
1429	switch (rc) {
 
 
 
 
1430	case H_FUNCTION:
1431		printk(KERN_INFO
1432			"VPHN is not supported. Disabling polling...\n");
1433		stop_topology_update();
1434		break;
1435	case H_HARDWARE:
1436		printk(KERN_ERR
1437			"hcall_vphn() experienced a hardware fault "
1438			"preventing VPHN. Disabling polling...\n");
1439		stop_topology_update();
 
 
 
 
 
 
 
 
1440	}
1441
1442	return rc;
1443}
1444
1445/*
1446 * Update the node maps and sysfs entries for each cpu whose home node
1447 * has changed.
1448 */
1449int arch_update_cpu_topology(void)
1450{
1451	int cpu, nid, old_nid;
1452	unsigned int associativity[VPHN_ASSOC_BUFSIZE] = {0};
1453	struct sys_device *sysdev;
1454
1455	for_each_cpu(cpu,&cpu_associativity_changes_mask) {
1456		vphn_get_associativity(cpu, associativity);
1457		nid = associativity_to_nid(associativity);
1458
1459		if (nid < 0 || !node_online(nid))
1460			nid = first_online_node;
1461
1462		old_nid = numa_cpu_lookup_table[cpu];
1463
1464		/* Disable hotplug while we update the cpu
1465		 * masks and sysfs.
1466		 */
1467		get_online_cpus();
1468		unregister_cpu_under_node(cpu, old_nid);
1469		unmap_cpu_from_node(cpu);
1470		map_cpu_to_node(cpu, nid);
1471		register_cpu_under_node(cpu, nid);
1472		put_online_cpus();
1473
1474		sysdev = get_cpu_sysdev(cpu);
1475		if (sysdev)
1476			kobject_uevent(&sysdev->kobj, KOBJ_CHANGE);
1477	}
1478
1479	return 1;
1480}
1481
1482static void topology_work_fn(struct work_struct *work)
1483{
1484	rebuild_sched_domains();
1485}
1486static DECLARE_WORK(topology_work, topology_work_fn);
 
 
1487
1488void topology_schedule_update(void)
1489{
1490	schedule_work(&topology_work);
1491}
1492
1493static void topology_timer_fn(unsigned long ignored)
1494{
1495	if (!vphn_enabled)
1496		return;
1497	if (update_cpu_associativity_changes_mask() > 0)
1498		topology_schedule_update();
1499	set_topology_timer();
1500}
1501static struct timer_list topology_timer =
1502	TIMER_INITIALIZER(topology_timer_fn, 0, 0);
1503
1504static void set_topology_timer(void)
1505{
1506	topology_timer.data = 0;
1507	topology_timer.expires = jiffies + 60 * HZ;
1508	add_timer(&topology_timer);
1509}
1510
1511/*
1512 * Start polling for VPHN associativity changes.
1513 */
1514int start_topology_update(void)
1515{
1516	int rc = 0;
1517
1518	/* Disabled until races with load balancing are fixed */
1519	if (0 && firmware_has_feature(FW_FEATURE_VPHN) &&
1520	    get_lppaca()->shared_proc) {
1521		vphn_enabled = 1;
1522		setup_cpu_associativity_change_counters();
1523		init_timer_deferrable(&topology_timer);
1524		set_topology_timer();
1525		rc = 1;
1526	}
1527
1528	return rc;
 
1529}
1530__initcall(start_topology_update);
1531
1532/*
1533 * Disable polling for VPHN associativity changes.
1534 */
1535int stop_topology_update(void)
1536{
1537	vphn_enabled = 0;
1538	return del_timer_sync(&topology_timer);
1539}
 
1540#endif /* CONFIG_PPC_SPLPAR */