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 */
v5.4
   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/pfn.h>
  20#include <linux/cpuset.h>
  21#include <linux/node.h>
  22#include <linux/stop_machine.h>
  23#include <linux/proc_fs.h>
  24#include <linux/seq_file.h>
  25#include <linux/uaccess.h>
  26#include <linux/slab.h>
  27#include <asm/cputhreads.h>
  28#include <asm/sparsemem.h>
  29#include <asm/prom.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/drmem.h>
  38
  39static int numa_enabled = 1;
  40
  41static char *cmdline __initdata;
  42
  43static int numa_debug;
  44#define dbg(args...) if (numa_debug) { printk(KERN_INFO args); }
  45
  46int numa_cpu_lookup_table[NR_CPUS];
  47cpumask_var_t node_to_cpumask_map[MAX_NUMNODES];
  48struct pglist_data *node_data[MAX_NUMNODES];
  49
  50EXPORT_SYMBOL(numa_cpu_lookup_table);
  51EXPORT_SYMBOL(node_to_cpumask_map);
  52EXPORT_SYMBOL(node_data);
  53
  54static int min_common_depth;
  55static int n_mem_addr_cells, n_mem_size_cells;
  56static int form1_affinity;
 
 
 
 
  57
  58#define MAX_DISTANCE_REF_POINTS 4
  59static int distance_ref_points_depth;
  60static const __be32 *distance_ref_points;
  61static int distance_lookup_table[MAX_NUMNODES][MAX_DISTANCE_REF_POINTS];
 
 
 
 
  62
  63/*
  64 * Allocate node_to_cpumask_map based on number of available nodes
  65 * Requires node_possible_map to be valid.
  66 *
  67 * Note: cpumask_of_node() is not valid until after this is done.
  68 */
  69static void __init setup_node_to_cpumask_map(void)
  70{
  71	unsigned int node;
  72
  73	/* setup nr_node_ids if not done yet */
  74	if (nr_node_ids == MAX_NUMNODES)
  75		setup_nr_node_ids();
  76
  77	/* allocate the map */
  78	for_each_node(node)
  79		alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]);
  80
  81	/* cpumask_of_node() will now work */
  82	dbg("Node to cpumask map for %u nodes\n", nr_node_ids);
  83}
  84
  85static int __init fake_numa_create_new_node(unsigned long end_pfn,
  86						unsigned int *nid)
  87{
  88	unsigned long long mem;
  89	char *p = cmdline;
  90	static unsigned int fake_nid;
  91	static unsigned long long curr_boundary;
  92
  93	/*
  94	 * Modify node id, iff we started creating NUMA nodes
  95	 * We want to continue from where we left of the last time
  96	 */
  97	if (fake_nid)
  98		*nid = fake_nid;
  99	/*
 100	 * In case there are no more arguments to parse, the
 101	 * node_id should be the same as the last fake node id
 102	 * (we've handled this above).
 103	 */
 104	if (!p)
 105		return 0;
 106
 107	mem = memparse(p, &p);
 108	if (!mem)
 109		return 0;
 110
 111	if (mem < curr_boundary)
 112		return 0;
 113
 114	curr_boundary = mem;
 115
 116	if ((end_pfn << PAGE_SHIFT) > mem) {
 117		/*
 118		 * Skip commas and spaces
 119		 */
 120		while (*p == ',' || *p == ' ' || *p == '\t')
 121			p++;
 122
 123		cmdline = p;
 124		fake_nid++;
 125		*nid = fake_nid;
 126		dbg("created new fake_node with id %d\n", fake_nid);
 127		return 1;
 128	}
 129	return 0;
 130}
 131
 132static void reset_numa_cpu_lookup_table(void)
 133{
 134	unsigned int cpu;
 135
 136	for_each_possible_cpu(cpu)
 137		numa_cpu_lookup_table[cpu] = -1;
 138}
 139
 140static void map_cpu_to_node(int cpu, int node)
 141{
 142	update_numa_cpu_lookup_table(cpu, node);
 143
 144	dbg("adding cpu %d to node %d\n", cpu, node);
 145
 146	if (!(cpumask_test_cpu(cpu, node_to_cpumask_map[node])))
 147		cpumask_set_cpu(cpu, node_to_cpumask_map[node]);
 
 148}
 149
 150#if defined(CONFIG_HOTPLUG_CPU) || defined(CONFIG_PPC_SPLPAR)
 151static void unmap_cpu_from_node(unsigned long cpu)
 152{
 153	int node = numa_cpu_lookup_table[cpu];
 154
 155	dbg("removing cpu %lu from node %d\n", cpu, node);
 156
 157	if (cpumask_test_cpu(cpu, node_to_cpumask_map[node])) {
 158		cpumask_clear_cpu(cpu, node_to_cpumask_map[node]);
 
 159	} else {
 160		printk(KERN_ERR "WARNING: cpu %lu not found in node %d\n",
 161		       cpu, node);
 162	}
 163}
 164#endif /* CONFIG_HOTPLUG_CPU || CONFIG_PPC_SPLPAR */
 165
 166int cpu_distance(__be32 *cpu1_assoc, __be32 *cpu2_assoc)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 167{
 168	int dist = 0;
 169
 170	int i, index;
 171
 172	for (i = 0; i < distance_ref_points_depth; i++) {
 173		index = be32_to_cpu(distance_ref_points[i]);
 174		if (cpu1_assoc[index] == cpu2_assoc[index])
 175			break;
 176		dist++;
 177	}
 178
 179	return dist;
 180}
 181
 
 
 
 
 
 
 
 
 
 182/* must hold reference to node during call */
 183static const __be32 *of_get_associativity(struct device_node *dev)
 184{
 185	return of_get_property(dev, "ibm,associativity", NULL);
 186}
 187
 188int __node_distance(int a, int b)
 189{
 190	int i;
 191	int distance = LOCAL_DISTANCE;
 192
 193	if (!form1_affinity)
 
 
 194		return ((a == b) ? LOCAL_DISTANCE : REMOTE_DISTANCE);
 195
 196	for (i = 0; i < distance_ref_points_depth; i++) {
 197		if (distance_lookup_table[a][i] == distance_lookup_table[b][i])
 198			break;
 199
 200		/* Double the distance for each NUMA level */
 201		distance *= 2;
 202	}
 203
 204	return distance;
 205}
 206EXPORT_SYMBOL(__node_distance);
 207
 208static void initialize_distance_lookup_table(int nid,
 209		const __be32 *associativity)
 210{
 211	int i;
 212
 213	if (!form1_affinity)
 214		return;
 215
 216	for (i = 0; i < distance_ref_points_depth; i++) {
 217		const __be32 *entry;
 218
 219		entry = &associativity[be32_to_cpu(distance_ref_points[i]) - 1];
 220		distance_lookup_table[nid][i] = of_read_number(entry, 1);
 221	}
 222}
 223
 224/* Returns nid in the range [0..MAX_NUMNODES-1], or -1 if no useful numa
 225 * info is found.
 226 */
 227static int associativity_to_nid(const __be32 *associativity)
 228{
 229	int nid = NUMA_NO_NODE;
 230
 231	if (!numa_enabled)
 232		goto out;
 233
 234	if (of_read_number(associativity, 1) >= min_common_depth)
 235		nid = of_read_number(&associativity[min_common_depth], 1);
 236
 237	/* POWER4 LPAR uses 0xffff as invalid node */
 238	if (nid == 0xffff || nid >= MAX_NUMNODES)
 239		nid = NUMA_NO_NODE;
 240
 241	if (nid > 0 &&
 242		of_read_number(associativity, 1) >= distance_ref_points_depth) {
 243		/*
 244		 * Skip the length field and send start of associativity array
 245		 */
 246		initialize_distance_lookup_table(nid, associativity + 1);
 247	}
 248
 249out:
 250	return nid;
 251}
 252
 253/* Returns the nid associated with the given device tree node,
 254 * or -1 if not found.
 255 */
 256static int of_node_to_nid_single(struct device_node *device)
 257{
 258	int nid = NUMA_NO_NODE;
 259	const __be32 *tmp;
 260
 261	tmp = of_get_associativity(device);
 262	if (tmp)
 263		nid = associativity_to_nid(tmp);
 264	return nid;
 265}
 266
 267/* Walk the device tree upwards, looking for an associativity id */
 268int of_node_to_nid(struct device_node *device)
 269{
 270	int nid = NUMA_NO_NODE;
 271
 272	of_node_get(device);
 273	while (device) {
 274		nid = of_node_to_nid_single(device);
 275		if (nid != -1)
 276			break;
 277
 278		device = of_get_next_parent(device);
 279	}
 280	of_node_put(device);
 281
 282	return nid;
 283}
 284EXPORT_SYMBOL(of_node_to_nid);
 285
 286static int __init find_min_common_depth(void)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 287{
 288	int depth;
 289	struct device_node *root;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 290
 291	if (firmware_has_feature(FW_FEATURE_OPAL))
 292		root = of_find_node_by_path("/ibm,opal");
 293	else
 294		root = of_find_node_by_path("/rtas");
 295	if (!root)
 296		root = of_find_node_by_path("/");
 297
 298	/*
 299	 * This property is a set of 32-bit integers, each representing
 300	 * an index into the ibm,associativity nodes.
 301	 *
 302	 * With form 0 affinity the first integer is for an SMP configuration
 303	 * (should be all 0's) and the second is for a normal NUMA
 304	 * configuration. We have only one level of NUMA.
 305	 *
 306	 * With form 1 affinity the first integer is the most significant
 307	 * NUMA boundary and the following are progressively less significant
 308	 * boundaries. There can be more than one level of NUMA.
 309	 */
 310	distance_ref_points = of_get_property(root,
 311					"ibm,associativity-reference-points",
 312					&distance_ref_points_depth);
 313
 314	if (!distance_ref_points) {
 315		dbg("NUMA: ibm,associativity-reference-points not found.\n");
 316		goto err;
 317	}
 318
 319	distance_ref_points_depth /= sizeof(int);
 320
 321	if (firmware_has_feature(FW_FEATURE_OPAL) ||
 322	    firmware_has_feature(FW_FEATURE_TYPE1_AFFINITY)) {
 323		dbg("Using form 1 affinity\n");
 324		form1_affinity = 1;
 325	}
 326
 327	if (form1_affinity) {
 328		depth = of_read_number(distance_ref_points, 1);
 329	} else {
 330		if (distance_ref_points_depth < 2) {
 331			printk(KERN_WARNING "NUMA: "
 332				"short ibm,associativity-reference-points\n");
 333			goto err;
 334		}
 335
 336		depth = of_read_number(&distance_ref_points[1], 1);
 
 
 
 
 
 
 337	}
 338
 339	/*
 340	 * Warn and cap if the hardware supports more than
 341	 * MAX_DISTANCE_REF_POINTS domains.
 342	 */
 343	if (distance_ref_points_depth > MAX_DISTANCE_REF_POINTS) {
 344		printk(KERN_WARNING "NUMA: distance array capped at "
 345			"%d entries\n", MAX_DISTANCE_REF_POINTS);
 346		distance_ref_points_depth = MAX_DISTANCE_REF_POINTS;
 347	}
 348
 349	of_node_put(root);
 350	return depth;
 351
 352err:
 353	of_node_put(root);
 354	return -1;
 355}
 356
 357static void __init get_n_mem_cells(int *n_addr_cells, int *n_size_cells)
 358{
 359	struct device_node *memory = NULL;
 360
 361	memory = of_find_node_by_type(memory, "memory");
 362	if (!memory)
 363		panic("numa.c: No memory nodes found!");
 364
 365	*n_addr_cells = of_n_addr_cells(memory);
 366	*n_size_cells = of_n_size_cells(memory);
 367	of_node_put(memory);
 368}
 369
 370static unsigned long read_n_cells(int n, const __be32 **buf)
 371{
 372	unsigned long result = 0;
 373
 374	while (n--) {
 375		result = (result << 32) | of_read_number(*buf, 1);
 376		(*buf)++;
 377	}
 378	return result;
 379}
 380
 381struct assoc_arrays {
 382	u32	n_arrays;
 383	u32	array_sz;
 384	const __be32 *arrays;
 385};
 386
 387/*
 388 * Retrieve and validate the list of associativity arrays for drconf
 389 * memory from the ibm,associativity-lookup-arrays property of the
 390 * device tree..
 391 *
 392 * The layout of the ibm,associativity-lookup-arrays property is a number N
 393 * indicating the number of associativity arrays, followed by a number M
 394 * indicating the size of each associativity array, followed by a list
 395 * of N associativity arrays.
 396 */
 397static int of_get_assoc_arrays(struct assoc_arrays *aa)
 398{
 399	struct device_node *memory;
 400	const __be32 *prop;
 401	u32 len;
 402
 403	memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
 404	if (!memory)
 405		return -1;
 406
 407	prop = of_get_property(memory, "ibm,associativity-lookup-arrays", &len);
 408	if (!prop || len < 2 * sizeof(unsigned int)) {
 409		of_node_put(memory);
 410		return -1;
 411	}
 412
 413	aa->n_arrays = of_read_number(prop++, 1);
 414	aa->array_sz = of_read_number(prop++, 1);
 415
 416	of_node_put(memory);
 417
 418	/* Now that we know the number of arrays and size of each array,
 419	 * revalidate the size of the property read in.
 420	 */
 421	if (len < (aa->n_arrays * aa->array_sz + 2) * sizeof(unsigned int))
 422		return -1;
 423
 424	aa->arrays = prop;
 425	return 0;
 426}
 427
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 428/*
 429 * This is like of_node_to_nid_single() for memory represented in the
 430 * ibm,dynamic-reconfiguration-memory node.
 431 */
 432static int of_drconf_to_nid_single(struct drmem_lmb *lmb)
 433{
 434	struct assoc_arrays aa = { .arrays = NULL };
 435	int default_nid = NUMA_NO_NODE;
 436	int nid = default_nid;
 437	int rc, index;
 438
 439	if ((min_common_depth < 0) || !numa_enabled)
 440		return default_nid;
 441
 442	rc = of_get_assoc_arrays(&aa);
 443	if (rc)
 444		return default_nid;
 445
 446	if (min_common_depth <= aa.array_sz &&
 447	    !(lmb->flags & DRCONF_MEM_AI_INVALID) && lmb->aa_index < aa.n_arrays) {
 448		index = lmb->aa_index * aa.array_sz + min_common_depth - 1;
 449		nid = of_read_number(&aa.arrays[index], 1);
 450
 451		if (nid == 0xffff || nid >= MAX_NUMNODES)
 452			nid = default_nid;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 453
 454		if (nid > 0) {
 455			index = lmb->aa_index * aa.array_sz;
 456			initialize_distance_lookup_table(nid,
 457							&aa.arrays[index]);
 458		}
 459	}
 460
 461	return nid;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 462}
 
 463
 464/*
 465 * Figure out to which domain a cpu belongs and stick it there.
 466 * Return the id of the domain used.
 467 */
 468static int numa_setup_cpu(unsigned long lcpu)
 469{
 
 
 470	int nid = NUMA_NO_NODE;
 471	struct device_node *cpu;
 
 
 
 
 472
 473	/*
 474	 * If a valid cpu-to-node mapping is already available, use it
 475	 * directly instead of querying the firmware, since it represents
 476	 * the most recent mapping notified to us by the platform (eg: VPHN).
 
 
 
 477	 */
 478	if ((nid = numa_cpu_lookup_table[lcpu]) >= 0) {
 
 479		map_cpu_to_node(lcpu, nid);
 480		return nid;
 481	}
 482
 
 
 
 
 483	cpu = of_get_cpu_node(lcpu, NULL);
 484
 485	if (!cpu) {
 486		WARN_ON(1);
 487		if (cpu_present(lcpu))
 488			goto out_present;
 489		else
 490			goto out;
 491	}
 492
 493	nid = of_node_to_nid_single(cpu);
 
 494
 495out_present:
 496	if (nid < 0 || !node_possible(nid))
 497		nid = first_online_node;
 498
 
 
 
 
 
 
 
 
 
 
 
 
 
 499	map_cpu_to_node(lcpu, nid);
 500	of_node_put(cpu);
 501out:
 502	return nid;
 503}
 504
 505static void verify_cpu_node_mapping(int cpu, int node)
 506{
 507	int base, sibling, i;
 508
 509	/* Verify that all the threads in the core belong to the same node */
 510	base = cpu_first_thread_sibling(cpu);
 511
 512	for (i = 0; i < threads_per_core; i++) {
 513		sibling = base + i;
 514
 515		if (sibling == cpu || cpu_is_offline(sibling))
 516			continue;
 517
 518		if (cpu_to_node(sibling) != node) {
 519			WARN(1, "CPU thread siblings %d and %d don't belong"
 520				" to the same node!\n", cpu, sibling);
 521			break;
 522		}
 523	}
 524}
 525
 526/* Must run before sched domains notifier. */
 527static int ppc_numa_cpu_prepare(unsigned int cpu)
 528{
 529	int nid;
 530
 531	nid = numa_setup_cpu(cpu);
 532	verify_cpu_node_mapping(cpu, nid);
 533	return 0;
 534}
 535
 536static int ppc_numa_cpu_dead(unsigned int cpu)
 537{
 538#ifdef CONFIG_HOTPLUG_CPU
 539	unmap_cpu_from_node(cpu);
 540#endif
 541	return 0;
 542}
 543
 544/*
 545 * Check and possibly modify a memory region to enforce the memory limit.
 546 *
 547 * Returns the size the region should have to enforce the memory limit.
 548 * This will either be the original value of size, a truncated value,
 549 * or zero. If the returned value of size is 0 the region should be
 550 * discarded as it lies wholly above the memory limit.
 551 */
 552static unsigned long __init numa_enforce_memory_limit(unsigned long start,
 553						      unsigned long size)
 554{
 555	/*
 556	 * We use memblock_end_of_DRAM() in here instead of memory_limit because
 557	 * we've already adjusted it for the limit and it takes care of
 558	 * having memory holes below the limit.  Also, in the case of
 559	 * iommu_is_off, memory_limit is not set but is implicitly enforced.
 560	 */
 561
 562	if (start + size <= memblock_end_of_DRAM())
 563		return size;
 564
 565	if (start >= memblock_end_of_DRAM())
 566		return 0;
 567
 568	return memblock_end_of_DRAM() - start;
 569}
 570
 571/*
 572 * Reads the counter for a given entry in
 573 * linux,drconf-usable-memory property
 574 */
 575static inline int __init read_usm_ranges(const __be32 **usm)
 576{
 577	/*
 578	 * For each lmb in ibm,dynamic-memory a corresponding
 579	 * entry in linux,drconf-usable-memory property contains
 580	 * a counter followed by that many (base, size) duple.
 581	 * read the counter from linux,drconf-usable-memory
 582	 */
 583	return read_n_cells(n_mem_size_cells, usm);
 584}
 585
 586/*
 587 * Extract NUMA information from the ibm,dynamic-reconfiguration-memory
 588 * node.  This assumes n_mem_{addr,size}_cells have been set.
 589 */
 590static void __init numa_setup_drmem_lmb(struct drmem_lmb *lmb,
 591					const __be32 **usm)
 
 592{
 593	unsigned int ranges, is_kexec_kdump = 0;
 594	unsigned long base, size, sz;
 595	int nid;
 596
 597	/*
 598	 * Skip this block if the reserved bit is set in flags (0x80)
 599	 * or if the block is not assigned to this partition (0x8)
 600	 */
 601	if ((lmb->flags & DRCONF_MEM_RESERVED)
 602	    || !(lmb->flags & DRCONF_MEM_ASSIGNED))
 603		return;
 604
 605	if (*usm)
 606		is_kexec_kdump = 1;
 607
 608	base = lmb->base_addr;
 609	size = drmem_lmb_size();
 610	ranges = 1;
 611
 612	if (is_kexec_kdump) {
 613		ranges = read_usm_ranges(usm);
 614		if (!ranges) /* there are no (base, size) duple */
 615			return;
 616	}
 617
 618	do {
 619		if (is_kexec_kdump) {
 620			base = read_n_cells(n_mem_addr_cells, usm);
 621			size = read_n_cells(n_mem_size_cells, usm);
 622		}
 623
 624		nid = of_drconf_to_nid_single(lmb);
 625		fake_numa_create_new_node(((base + size) >> PAGE_SHIFT),
 626					  &nid);
 627		node_set_online(nid);
 628		sz = numa_enforce_memory_limit(base, size);
 629		if (sz)
 630			memblock_set_node(base, sz, &memblock.memory, nid);
 631	} while (--ranges);
 
 
 632}
 633
 634static int __init parse_numa_properties(void)
 635{
 636	struct device_node *memory;
 637	int default_nid = 0;
 638	unsigned long i;
 
 639
 640	if (numa_enabled == 0) {
 641		printk(KERN_WARNING "NUMA disabled by user\n");
 642		return -1;
 643	}
 644
 645	min_common_depth = find_min_common_depth();
 646
 647	if (min_common_depth < 0) {
 648		/*
 649		 * if we fail to parse min_common_depth from device tree
 650		 * mark the numa disabled, boot with numa disabled.
 651		 */
 652		numa_enabled = false;
 653		return min_common_depth;
 654	}
 655
 656	dbg("NUMA associativity depth for CPU/Memory: %d\n", min_common_depth);
 
 
 
 
 
 
 657
 658	/*
 659	 * Even though we connect cpus to numa domains later in SMP
 660	 * init, we need to know the node ids now. This is because
 661	 * each node to be onlined must have NODE_DATA etc backing it.
 662	 */
 663	for_each_present_cpu(i) {
 
 664		struct device_node *cpu;
 665		int nid;
 
 
 666
 667		cpu = of_get_cpu_node(i, NULL);
 668		BUG_ON(!cpu);
 669		nid = of_node_to_nid_single(cpu);
 670		of_node_put(cpu);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 671
 672		/*
 673		 * Don't fall back to default_nid yet -- we will plug
 674		 * cpus into nodes once the memory scan has discovered
 675		 * the topology.
 676		 */
 677		if (nid < 0)
 678			continue;
 679		node_set_online(nid);
 680	}
 681
 682	get_n_mem_cells(&n_mem_addr_cells, &n_mem_size_cells);
 683
 684	for_each_node_by_type(memory, "memory") {
 685		unsigned long start;
 686		unsigned long size;
 687		int nid;
 688		int ranges;
 689		const __be32 *memcell_buf;
 690		unsigned int len;
 691
 692		memcell_buf = of_get_property(memory,
 693			"linux,usable-memory", &len);
 694		if (!memcell_buf || len <= 0)
 695			memcell_buf = of_get_property(memory, "reg", &len);
 696		if (!memcell_buf || len <= 0)
 697			continue;
 698
 699		/* ranges in cell */
 700		ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
 701new_range:
 702		/* these are order-sensitive, and modify the buffer pointer */
 703		start = read_n_cells(n_mem_addr_cells, &memcell_buf);
 704		size = read_n_cells(n_mem_size_cells, &memcell_buf);
 705
 706		/*
 707		 * Assumption: either all memory nodes or none will
 708		 * have associativity properties.  If none, then
 709		 * everything goes to default_nid.
 710		 */
 711		nid = of_node_to_nid_single(memory);
 712		if (nid < 0)
 
 
 
 713			nid = default_nid;
 714
 715		fake_numa_create_new_node(((start + size) >> PAGE_SHIFT), &nid);
 716		node_set_online(nid);
 717
 718		size = numa_enforce_memory_limit(start, size);
 719		if (size)
 720			memblock_set_node(start, size, &memblock.memory, nid);
 721
 722		if (--ranges)
 723			goto new_range;
 724	}
 725
 726	/*
 727	 * Now do the same thing for each MEMBLOCK listed in the
 728	 * ibm,dynamic-memory property in the
 729	 * ibm,dynamic-reconfiguration-memory node.
 730	 */
 731	memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
 732	if (memory) {
 733		walk_drmem_lmbs(memory, numa_setup_drmem_lmb);
 734		of_node_put(memory);
 735	}
 736
 737	return 0;
 738}
 739
 740static void __init setup_nonnuma(void)
 741{
 742	unsigned long top_of_ram = memblock_end_of_DRAM();
 743	unsigned long total_ram = memblock_phys_mem_size();
 744	unsigned long start_pfn, end_pfn;
 745	unsigned int nid = 0;
 746	struct memblock_region *reg;
 747
 748	printk(KERN_DEBUG "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
 749	       top_of_ram, total_ram);
 750	printk(KERN_DEBUG "Memory hole size: %ldMB\n",
 751	       (top_of_ram - total_ram) >> 20);
 752
 753	for_each_memblock(memory, reg) {
 754		start_pfn = memblock_region_memory_base_pfn(reg);
 755		end_pfn = memblock_region_memory_end_pfn(reg);
 756
 
 757		fake_numa_create_new_node(end_pfn, &nid);
 758		memblock_set_node(PFN_PHYS(start_pfn),
 759				  PFN_PHYS(end_pfn - start_pfn),
 760				  &memblock.memory, nid);
 761		node_set_online(nid);
 762	}
 763}
 764
 765void __init dump_numa_cpu_topology(void)
 766{
 767	unsigned int node;
 768	unsigned int cpu, count;
 769
 770	if (!numa_enabled)
 771		return;
 772
 773	for_each_online_node(node) {
 774		pr_info("Node %d CPUs:", node);
 775
 776		count = 0;
 777		/*
 778		 * If we used a CPU iterator here we would miss printing
 779		 * the holes in the cpumap.
 780		 */
 781		for (cpu = 0; cpu < nr_cpu_ids; cpu++) {
 782			if (cpumask_test_cpu(cpu,
 783					node_to_cpumask_map[node])) {
 784				if (count == 0)
 785					pr_cont(" %u", cpu);
 786				++count;
 787			} else {
 788				if (count > 1)
 789					pr_cont("-%u", cpu - 1);
 790				count = 0;
 791			}
 792		}
 793
 794		if (count > 1)
 795			pr_cont("-%u", nr_cpu_ids - 1);
 796		pr_cont("\n");
 797	}
 798}
 799
 800/* Initialize NODE_DATA for a node on the local memory */
 801static void __init setup_node_data(int nid, u64 start_pfn, u64 end_pfn)
 802{
 803	u64 spanned_pages = end_pfn - start_pfn;
 804	const size_t nd_size = roundup(sizeof(pg_data_t), SMP_CACHE_BYTES);
 805	u64 nd_pa;
 806	void *nd;
 807	int tnid;
 808
 809	nd_pa = memblock_phys_alloc_try_nid(nd_size, SMP_CACHE_BYTES, nid);
 810	if (!nd_pa)
 811		panic("Cannot allocate %zu bytes for node %d data\n",
 812		      nd_size, nid);
 813
 814	nd = __va(nd_pa);
 815
 816	/* report and initialize */
 817	pr_info("  NODE_DATA [mem %#010Lx-%#010Lx]\n",
 818		nd_pa, nd_pa + nd_size - 1);
 819	tnid = early_pfn_to_nid(nd_pa >> PAGE_SHIFT);
 820	if (tnid != nid)
 821		pr_info("    NODE_DATA(%d) on node %d\n", nid, tnid);
 822
 823	node_data[nid] = nd;
 824	memset(NODE_DATA(nid), 0, sizeof(pg_data_t));
 825	NODE_DATA(nid)->node_id = nid;
 826	NODE_DATA(nid)->node_start_pfn = start_pfn;
 827	NODE_DATA(nid)->node_spanned_pages = spanned_pages;
 828}
 829
 830static void __init find_possible_nodes(void)
 831{
 832	struct device_node *rtas;
 833	u32 numnodes, i;
 
 
 834
 835	if (!numa_enabled)
 836		return;
 837
 838	rtas = of_find_node_by_path("/rtas");
 839	if (!rtas)
 840		return;
 841
 842	if (of_property_read_u32_index(rtas,
 843				"ibm,max-associativity-domains",
 844				min_common_depth, &numnodes))
 845		goto out;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 846
 847	for (i = 0; i < numnodes; i++) {
 848		if (!node_possible(i))
 849			node_set(i, node_possible_map);
 850	}
 851
 
 
 
 
 852out:
 853	of_node_put(rtas);
 854}
 855
 856void __init mem_topology_setup(void)
 857{
 858	int cpu;
 859
 
 
 
 
 
 
 
 
 
 
 
 
 
 860	if (parse_numa_properties())
 861		setup_nonnuma();
 862
 863	/*
 864	 * Modify the set of possible NUMA nodes to reflect information
 865	 * available about the set of online nodes, and the set of nodes
 866	 * that we expect to make use of for this platform's affinity
 867	 * calculations.
 868	 */
 869	nodes_and(node_possible_map, node_possible_map, node_online_map);
 870
 871	find_possible_nodes();
 872
 873	setup_node_to_cpumask_map();
 874
 875	reset_numa_cpu_lookup_table();
 876
 877	for_each_present_cpu(cpu)
 
 
 
 
 
 
 
 
 878		numa_setup_cpu(cpu);
 
 879}
 880
 881void __init initmem_init(void)
 882{
 883	int nid;
 884
 885	max_low_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
 886	max_pfn = max_low_pfn;
 887
 888	memblock_dump_all();
 889
 890	for_each_online_node(nid) {
 891		unsigned long start_pfn, end_pfn;
 892
 893		get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
 894		setup_node_data(nid, start_pfn, end_pfn);
 895		sparse_memory_present_with_active_regions(nid);
 896	}
 897
 898	sparse_init();
 899
 900	/*
 901	 * We need the numa_cpu_lookup_table to be accurate for all CPUs,
 902	 * even before we online them, so that we can use cpu_to_{node,mem}
 903	 * early in boot, cf. smp_prepare_cpus().
 904	 * _nocalls() + manual invocation is used because cpuhp is not yet
 905	 * initialized for the boot CPU.
 906	 */
 907	cpuhp_setup_state_nocalls(CPUHP_POWER_NUMA_PREPARE, "powerpc/numa:prepare",
 908				  ppc_numa_cpu_prepare, ppc_numa_cpu_dead);
 909}
 910
 911static int __init early_numa(char *p)
 912{
 913	if (!p)
 914		return 0;
 915
 916	if (strstr(p, "off"))
 917		numa_enabled = 0;
 918
 919	if (strstr(p, "debug"))
 920		numa_debug = 1;
 921
 922	p = strstr(p, "fake=");
 923	if (p)
 924		cmdline = p + strlen("fake=");
 925
 926	return 0;
 927}
 928early_param("numa", early_numa);
 929
 930/*
 931 * The platform can inform us through one of several mechanisms
 932 * (post-migration device tree updates, PRRN or VPHN) that the NUMA
 933 * assignment of a resource has changed. This controls whether we act
 934 * on that. Disabled by default.
 935 */
 936static bool topology_updates_enabled;
 937
 938static int __init early_topology_updates(char *p)
 939{
 940	if (!p)
 941		return 0;
 942
 943	if (!strcmp(p, "on")) {
 944		pr_warn("Caution: enabling topology updates\n");
 945		topology_updates_enabled = true;
 946	}
 947
 948	return 0;
 949}
 950early_param("topology_updates", early_topology_updates);
 951
 952#ifdef CONFIG_MEMORY_HOTPLUG
 953/*
 954 * Find the node associated with a hot added memory section for
 955 * memory represented in the device tree by the property
 956 * ibm,dynamic-reconfiguration-memory/ibm,dynamic-memory.
 957 */
 958static int hot_add_drconf_scn_to_nid(unsigned long scn_addr)
 959{
 960	struct drmem_lmb *lmb;
 961	unsigned long lmb_size;
 962	int nid = NUMA_NO_NODE;
 963
 964	lmb_size = drmem_lmb_size();
 965
 966	for_each_drmem_lmb(lmb) {
 967		/* skip this block if it is reserved or not assigned to
 968		 * this partition */
 969		if ((lmb->flags & DRCONF_MEM_RESERVED)
 970		    || !(lmb->flags & DRCONF_MEM_ASSIGNED))
 971			continue;
 972
 973		if ((scn_addr < lmb->base_addr)
 974		    || (scn_addr >= (lmb->base_addr + lmb_size)))
 975			continue;
 976
 977		nid = of_drconf_to_nid_single(lmb);
 978		break;
 979	}
 980
 981	return nid;
 982}
 983
 984/*
 985 * Find the node associated with a hot added memory section for memory
 986 * represented in the device tree as a node (i.e. memory@XXXX) for
 987 * each memblock.
 988 */
 989static int hot_add_node_scn_to_nid(unsigned long scn_addr)
 990{
 991	struct device_node *memory;
 992	int nid = NUMA_NO_NODE;
 993
 994	for_each_node_by_type(memory, "memory") {
 995		unsigned long start, size;
 996		int ranges;
 997		const __be32 *memcell_buf;
 998		unsigned int len;
 999
1000		memcell_buf = of_get_property(memory, "reg", &len);
1001		if (!memcell_buf || len <= 0)
1002			continue;
1003
1004		/* ranges in cell */
1005		ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
1006
1007		while (ranges--) {
1008			start = read_n_cells(n_mem_addr_cells, &memcell_buf);
1009			size = read_n_cells(n_mem_size_cells, &memcell_buf);
1010
1011			if ((scn_addr < start) || (scn_addr >= (start + size)))
1012				continue;
1013
1014			nid = of_node_to_nid_single(memory);
1015			break;
1016		}
1017
1018		if (nid >= 0)
1019			break;
1020	}
1021
1022	of_node_put(memory);
1023
1024	return nid;
1025}
1026
1027/*
1028 * Find the node associated with a hot added memory section.  Section
1029 * corresponds to a SPARSEMEM section, not an MEMBLOCK.  It is assumed that
1030 * sections are fully contained within a single MEMBLOCK.
1031 */
1032int hot_add_scn_to_nid(unsigned long scn_addr)
1033{
1034	struct device_node *memory = NULL;
1035	int nid;
1036
1037	if (!numa_enabled)
1038		return first_online_node;
1039
1040	memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1041	if (memory) {
1042		nid = hot_add_drconf_scn_to_nid(scn_addr);
1043		of_node_put(memory);
1044	} else {
1045		nid = hot_add_node_scn_to_nid(scn_addr);
1046	}
1047
1048	if (nid < 0 || !node_possible(nid))
1049		nid = first_online_node;
1050
1051	return nid;
1052}
1053
1054static u64 hot_add_drconf_memory_max(void)
1055{
1056	struct device_node *memory = NULL;
1057	struct device_node *dn = NULL;
1058	const __be64 *lrdr = NULL;
1059
1060	dn = of_find_node_by_path("/rtas");
1061	if (dn) {
1062		lrdr = of_get_property(dn, "ibm,lrdr-capacity", NULL);
1063		of_node_put(dn);
1064		if (lrdr)
1065			return be64_to_cpup(lrdr);
1066	}
1067
1068	memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
1069	if (memory) {
1070		of_node_put(memory);
1071		return drmem_lmb_memory_max();
1072	}
1073	return 0;
1074}
1075
1076/*
1077 * memory_hotplug_max - return max address of memory that may be added
1078 *
1079 * This is currently only used on systems that support drconfig memory
1080 * hotplug.
1081 */
1082u64 memory_hotplug_max(void)
1083{
1084        return max(hot_add_drconf_memory_max(), memblock_end_of_DRAM());
1085}
1086#endif /* CONFIG_MEMORY_HOTPLUG */
1087
1088/* Virtual Processor Home Node (VPHN) support */
1089#ifdef CONFIG_PPC_SPLPAR
1090struct topology_update_data {
1091	struct topology_update_data *next;
1092	unsigned int cpu;
1093	int old_nid;
1094	int new_nid;
1095};
1096
1097#define TOPOLOGY_DEF_TIMER_SECS	60
1098
1099static u8 vphn_cpu_change_counts[NR_CPUS][MAX_DISTANCE_REF_POINTS];
1100static cpumask_t cpu_associativity_changes_mask;
1101static int vphn_enabled;
1102static int prrn_enabled;
1103static void reset_topology_timer(void);
1104static int topology_timer_secs = 1;
1105static int topology_inited;
1106
1107/*
1108 * Change polling interval for associativity changes.
1109 */
1110int timed_topology_update(int nsecs)
1111{
1112	if (vphn_enabled) {
1113		if (nsecs > 0)
1114			topology_timer_secs = nsecs;
1115		else
1116			topology_timer_secs = TOPOLOGY_DEF_TIMER_SECS;
1117
1118		reset_topology_timer();
1119	}
1120
1121	return 0;
1122}
1123
1124/*
1125 * Store the current values of the associativity change counters in the
1126 * hypervisor.
1127 */
1128static void setup_cpu_associativity_change_counters(void)
1129{
1130	int cpu;
1131
1132	/* The VPHN feature supports a maximum of 8 reference points */
1133	BUILD_BUG_ON(MAX_DISTANCE_REF_POINTS > 8);
1134
1135	for_each_possible_cpu(cpu) {
1136		int i;
1137		u8 *counts = vphn_cpu_change_counts[cpu];
1138		volatile u8 *hypervisor_counts = lppaca_of(cpu).vphn_assoc_counts;
1139
1140		for (i = 0; i < distance_ref_points_depth; i++)
1141			counts[i] = hypervisor_counts[i];
1142	}
1143}
1144
1145/*
1146 * The hypervisor maintains a set of 8 associativity change counters in
1147 * the VPA of each cpu that correspond to the associativity levels in the
1148 * ibm,associativity-reference-points property. When an associativity
1149 * level changes, the corresponding counter is incremented.
1150 *
1151 * Set a bit in cpu_associativity_changes_mask for each cpu whose home
1152 * node associativity levels have changed.
1153 *
1154 * Returns the number of cpus with unhandled associativity changes.
1155 */
1156static int update_cpu_associativity_changes_mask(void)
1157{
1158	int cpu;
1159	cpumask_t *changes = &cpu_associativity_changes_mask;
1160
1161	for_each_possible_cpu(cpu) {
1162		int i, changed = 0;
1163		u8 *counts = vphn_cpu_change_counts[cpu];
1164		volatile u8 *hypervisor_counts = lppaca_of(cpu).vphn_assoc_counts;
1165
1166		for (i = 0; i < distance_ref_points_depth; i++) {
1167			if (hypervisor_counts[i] != counts[i]) {
1168				counts[i] = hypervisor_counts[i];
1169				changed = 1;
1170			}
1171		}
1172		if (changed) {
1173			cpumask_or(changes, changes, cpu_sibling_mask(cpu));
1174			cpu = cpu_last_thread_sibling(cpu);
1175		}
1176	}
1177
1178	return cpumask_weight(changes);
1179}
1180
1181/*
1182 * Retrieve the new associativity information for a virtual processor's
1183 * home node.
1184 */
1185static long vphn_get_associativity(unsigned long cpu,
1186					__be32 *associativity)
1187{
1188	long rc;
1189
1190	rc = hcall_vphn(get_hard_smp_processor_id(cpu),
1191				VPHN_FLAG_VCPU, associativity);
1192
1193	switch (rc) {
 
 
 
 
1194	case H_FUNCTION:
1195		printk_once(KERN_INFO
1196			"VPHN is not supported. Disabling polling...\n");
1197		stop_topology_update();
1198		break;
1199	case H_HARDWARE:
1200		printk(KERN_ERR
1201			"hcall_vphn() experienced a hardware fault "
1202			"preventing VPHN. Disabling polling...\n");
1203		stop_topology_update();
1204		break;
1205	case H_SUCCESS:
1206		dbg("VPHN hcall succeeded. Reset polling...\n");
1207		timed_topology_update(0);
 
 
 
 
1208		break;
1209	}
1210
1211	return rc;
1212}
1213
1214int find_and_online_cpu_nid(int cpu)
1215{
1216	__be32 associativity[VPHN_ASSOC_BUFSIZE] = {0};
1217	int new_nid;
1218
1219	/* Use associativity from first thread for all siblings */
1220	if (vphn_get_associativity(cpu, associativity))
1221		return cpu_to_node(cpu);
1222
 
1223	new_nid = associativity_to_nid(associativity);
 
1224	if (new_nid < 0 || !node_possible(new_nid))
1225		new_nid = first_online_node;
1226
1227	if (NODE_DATA(new_nid) == NULL) {
1228#ifdef CONFIG_MEMORY_HOTPLUG
1229		/*
1230		 * Need to ensure that NODE_DATA is initialized for a node from
1231		 * available memory (see memblock_alloc_try_nid). If unable to
1232		 * init the node, then default to nearest node that has memory
1233		 * installed. Skip onlining a node if the subsystems are not
1234		 * yet initialized.
1235		 */
1236		if (!topology_inited || try_online_node(new_nid))
1237			new_nid = first_online_node;
1238#else
1239		/*
1240		 * Default to using the nearest node that has memory installed.
1241		 * Otherwise, it would be necessary to patch the kernel MM code
1242		 * to deal with more memoryless-node error conditions.
1243		 */
1244		new_nid = first_online_node;
1245#endif
1246	}
1247
1248	pr_debug("%s:%d cpu %d nid %d\n", __FUNCTION__, __LINE__,
1249		cpu, new_nid);
1250	return new_nid;
1251}
1252
1253/*
1254 * Update the CPU maps and sysfs entries for a single CPU when its NUMA
1255 * characteristics change. This function doesn't perform any locking and is
1256 * only safe to call from stop_machine().
1257 */
1258static int update_cpu_topology(void *data)
1259{
1260	struct topology_update_data *update;
1261	unsigned long cpu;
1262
1263	if (!data)
1264		return -EINVAL;
1265
1266	cpu = smp_processor_id();
1267
1268	for (update = data; update; update = update->next) {
1269		int new_nid = update->new_nid;
1270		if (cpu != update->cpu)
1271			continue;
1272
1273		unmap_cpu_from_node(cpu);
1274		map_cpu_to_node(cpu, new_nid);
1275		set_cpu_numa_node(cpu, new_nid);
1276		set_cpu_numa_mem(cpu, local_memory_node(new_nid));
1277		vdso_getcpu_init();
1278	}
1279
1280	return 0;
1281}
1282
1283static int update_lookup_table(void *data)
1284{
1285	struct topology_update_data *update;
 
1286
1287	if (!data)
1288		return -EINVAL;
1289
1290	/*
1291	 * Upon topology update, the numa-cpu lookup table needs to be updated
1292	 * for all threads in the core, including offline CPUs, to ensure that
1293	 * future hotplug operations respect the cpu-to-node associativity
1294	 * properly.
1295	 */
1296	for (update = data; update; update = update->next) {
1297		int nid, base, j;
1298
1299		nid = update->new_nid;
1300		base = cpu_first_thread_sibling(update->cpu);
1301
1302		for (j = 0; j < threads_per_core; j++) {
1303			update_numa_cpu_lookup_table(base + j, nid);
1304		}
1305	}
1306
1307	return 0;
1308}
1309
1310/*
1311 * Update the node maps and sysfs entries for each cpu whose home node
1312 * has changed. Returns 1 when the topology has changed, and 0 otherwise.
1313 *
1314 * cpus_locked says whether we already hold cpu_hotplug_lock.
1315 */
1316int numa_update_cpu_topology(bool cpus_locked)
1317{
1318	unsigned int cpu, sibling, changed = 0;
1319	struct topology_update_data *updates, *ud;
1320	cpumask_t updated_cpus;
1321	struct device *dev;
1322	int weight, new_nid, i = 0;
1323
1324	if (!prrn_enabled && !vphn_enabled && topology_inited)
1325		return 0;
1326
1327	weight = cpumask_weight(&cpu_associativity_changes_mask);
1328	if (!weight)
1329		return 0;
1330
1331	updates = kcalloc(weight, sizeof(*updates), GFP_KERNEL);
1332	if (!updates)
1333		return 0;
1334
1335	cpumask_clear(&updated_cpus);
1336
1337	for_each_cpu(cpu, &cpu_associativity_changes_mask) {
1338		/*
1339		 * If siblings aren't flagged for changes, updates list
1340		 * will be too short. Skip on this update and set for next
1341		 * update.
1342		 */
1343		if (!cpumask_subset(cpu_sibling_mask(cpu),
1344					&cpu_associativity_changes_mask)) {
1345			pr_info("Sibling bits not set for associativity "
1346					"change, cpu%d\n", cpu);
1347			cpumask_or(&cpu_associativity_changes_mask,
1348					&cpu_associativity_changes_mask,
1349					cpu_sibling_mask(cpu));
1350			cpu = cpu_last_thread_sibling(cpu);
1351			continue;
1352		}
1353
1354		new_nid = find_and_online_cpu_nid(cpu);
1355
1356		if (new_nid == numa_cpu_lookup_table[cpu]) {
1357			cpumask_andnot(&cpu_associativity_changes_mask,
1358					&cpu_associativity_changes_mask,
1359					cpu_sibling_mask(cpu));
1360			dbg("Assoc chg gives same node %d for cpu%d\n",
1361					new_nid, cpu);
1362			cpu = cpu_last_thread_sibling(cpu);
1363			continue;
1364		}
1365
1366		for_each_cpu(sibling, cpu_sibling_mask(cpu)) {
1367			ud = &updates[i++];
1368			ud->next = &updates[i];
1369			ud->cpu = sibling;
1370			ud->new_nid = new_nid;
1371			ud->old_nid = numa_cpu_lookup_table[sibling];
1372			cpumask_set_cpu(sibling, &updated_cpus);
1373		}
1374		cpu = cpu_last_thread_sibling(cpu);
1375	}
1376
1377	/*
1378	 * Prevent processing of 'updates' from overflowing array
1379	 * where last entry filled in a 'next' pointer.
1380	 */
1381	if (i)
1382		updates[i-1].next = NULL;
1383
1384	pr_debug("Topology update for the following CPUs:\n");
1385	if (cpumask_weight(&updated_cpus)) {
1386		for (ud = &updates[0]; ud; ud = ud->next) {
1387			pr_debug("cpu %d moving from node %d "
1388					  "to %d\n", ud->cpu,
1389					  ud->old_nid, ud->new_nid);
1390		}
1391	}
1392
1393	/*
1394	 * In cases where we have nothing to update (because the updates list
1395	 * is too short or because the new topology is same as the old one),
1396	 * skip invoking update_cpu_topology() via stop-machine(). This is
1397	 * necessary (and not just a fast-path optimization) since stop-machine
1398	 * can end up electing a random CPU to run update_cpu_topology(), and
1399	 * thus trick us into setting up incorrect cpu-node mappings (since
1400	 * 'updates' is kzalloc()'ed).
1401	 *
1402	 * And for the similar reason, we will skip all the following updating.
1403	 */
1404	if (!cpumask_weight(&updated_cpus))
1405		goto out;
1406
1407	if (cpus_locked)
1408		stop_machine_cpuslocked(update_cpu_topology, &updates[0],
1409					&updated_cpus);
1410	else
1411		stop_machine(update_cpu_topology, &updates[0], &updated_cpus);
1412
1413	/*
1414	 * Update the numa-cpu lookup table with the new mappings, even for
1415	 * offline CPUs. It is best to perform this update from the stop-
1416	 * machine context.
1417	 */
1418	if (cpus_locked)
1419		stop_machine_cpuslocked(update_lookup_table, &updates[0],
1420					cpumask_of(raw_smp_processor_id()));
1421	else
1422		stop_machine(update_lookup_table, &updates[0],
1423			     cpumask_of(raw_smp_processor_id()));
1424
1425	for (ud = &updates[0]; ud; ud = ud->next) {
1426		unregister_cpu_under_node(ud->cpu, ud->old_nid);
1427		register_cpu_under_node(ud->cpu, ud->new_nid);
1428
1429		dev = get_cpu_device(ud->cpu);
1430		if (dev)
1431			kobject_uevent(&dev->kobj, KOBJ_CHANGE);
1432		cpumask_clear_cpu(ud->cpu, &cpu_associativity_changes_mask);
1433		changed = 1;
1434	}
1435
1436out:
1437	kfree(updates);
1438	return changed;
1439}
1440
1441int arch_update_cpu_topology(void)
1442{
1443	return numa_update_cpu_topology(true);
1444}
1445
1446static void topology_work_fn(struct work_struct *work)
1447{
1448	rebuild_sched_domains();
1449}
1450static DECLARE_WORK(topology_work, topology_work_fn);
1451
1452static void topology_schedule_update(void)
1453{
1454	schedule_work(&topology_work);
1455}
1456
1457static void topology_timer_fn(struct timer_list *unused)
1458{
1459	if (prrn_enabled && cpumask_weight(&cpu_associativity_changes_mask))
1460		topology_schedule_update();
1461	else if (vphn_enabled) {
1462		if (update_cpu_associativity_changes_mask() > 0)
1463			topology_schedule_update();
1464		reset_topology_timer();
1465	}
1466}
1467static struct timer_list topology_timer;
1468
1469static void reset_topology_timer(void)
1470{
1471	if (vphn_enabled)
1472		mod_timer(&topology_timer, jiffies + topology_timer_secs * HZ);
1473}
1474
1475#ifdef CONFIG_SMP
1476
1477static int dt_update_callback(struct notifier_block *nb,
1478				unsigned long action, void *data)
1479{
1480	struct of_reconfig_data *update = data;
1481	int rc = NOTIFY_DONE;
1482
1483	switch (action) {
1484	case OF_RECONFIG_UPDATE_PROPERTY:
1485		if (of_node_is_type(update->dn, "cpu") &&
1486		    !of_prop_cmp(update->prop->name, "ibm,associativity")) {
1487			u32 core_id;
1488			of_property_read_u32(update->dn, "reg", &core_id);
1489			rc = dlpar_cpu_readd(core_id);
1490			rc = NOTIFY_OK;
1491		}
1492		break;
1493	}
1494
1495	return rc;
1496}
1497
1498static struct notifier_block dt_update_nb = {
1499	.notifier_call = dt_update_callback,
1500};
1501
1502#endif
1503
1504/*
1505 * Start polling for associativity changes.
1506 */
1507int start_topology_update(void)
1508{
1509	int rc = 0;
1510
1511	if (!topology_updates_enabled)
1512		return 0;
1513
1514	if (firmware_has_feature(FW_FEATURE_PRRN)) {
1515		if (!prrn_enabled) {
1516			prrn_enabled = 1;
1517#ifdef CONFIG_SMP
1518			rc = of_reconfig_notifier_register(&dt_update_nb);
1519#endif
1520		}
1521	}
1522	if (firmware_has_feature(FW_FEATURE_VPHN) &&
1523		   lppaca_shared_proc(get_lppaca())) {
1524		if (!vphn_enabled) {
1525			vphn_enabled = 1;
1526			setup_cpu_associativity_change_counters();
1527			timer_setup(&topology_timer, topology_timer_fn,
1528				    TIMER_DEFERRABLE);
1529			reset_topology_timer();
1530		}
1531	}
1532
1533	pr_info("Starting topology update%s%s\n",
1534		(prrn_enabled ? " prrn_enabled" : ""),
1535		(vphn_enabled ? " vphn_enabled" : ""));
1536
1537	return rc;
1538}
1539
1540/*
1541 * Disable polling for VPHN associativity changes.
1542 */
1543int stop_topology_update(void)
1544{
1545	int rc = 0;
1546
1547	if (!topology_updates_enabled)
1548		return 0;
1549
1550	if (prrn_enabled) {
1551		prrn_enabled = 0;
1552#ifdef CONFIG_SMP
1553		rc = of_reconfig_notifier_unregister(&dt_update_nb);
1554#endif
1555	}
1556	if (vphn_enabled) {
1557		vphn_enabled = 0;
1558		rc = del_timer_sync(&topology_timer);
1559	}
1560
1561	pr_info("Stopping topology update\n");
1562
1563	return rc;
1564}
1565
1566int prrn_is_enabled(void)
1567{
1568	return prrn_enabled;
1569}
1570
1571void __init shared_proc_topology_init(void)
1572{
1573	if (lppaca_shared_proc(get_lppaca())) {
1574		bitmap_fill(cpumask_bits(&cpu_associativity_changes_mask),
1575			    nr_cpumask_bits);
1576		numa_update_cpu_topology(false);
1577	}
1578}
1579
1580static int topology_read(struct seq_file *file, void *v)
1581{
1582	if (vphn_enabled || prrn_enabled)
1583		seq_puts(file, "on\n");
1584	else
1585		seq_puts(file, "off\n");
1586
1587	return 0;
1588}
1589
1590static int topology_open(struct inode *inode, struct file *file)
1591{
1592	return single_open(file, topology_read, NULL);
1593}
1594
1595static ssize_t topology_write(struct file *file, const char __user *buf,
1596			      size_t count, loff_t *off)
1597{
1598	char kbuf[4]; /* "on" or "off" plus null. */
1599	int read_len;
1600
1601	read_len = count < 3 ? count : 3;
1602	if (copy_from_user(kbuf, buf, read_len))
1603		return -EINVAL;
1604
1605	kbuf[read_len] = '\0';
1606
1607	if (!strncmp(kbuf, "on", 2)) {
1608		topology_updates_enabled = true;
1609		start_topology_update();
1610	} else if (!strncmp(kbuf, "off", 3)) {
1611		stop_topology_update();
1612		topology_updates_enabled = false;
1613	} else
1614		return -EINVAL;
1615
1616	return count;
1617}
1618
1619static const struct file_operations topology_ops = {
1620	.read = seq_read,
1621	.write = topology_write,
1622	.open = topology_open,
1623	.release = single_release
1624};
1625
1626static int topology_update_init(void)
1627{
1628	start_topology_update();
1629
1630	if (vphn_enabled)
1631		topology_schedule_update();
1632
1633	if (!proc_create("powerpc/topology_updates", 0644, NULL, &topology_ops))
1634		return -ENOMEM;
1635
1636	topology_inited = 1;
1637	return 0;
1638}
1639device_initcall(topology_update_init);
1640#endif /* CONFIG_PPC_SPLPAR */