1// SPDX-License-Identifier: GPL-2.0-only
  2/*
  3 * NUMA support, based on the x86 implementation.
  4 *
  5 * Copyright (C) 2015 Cavium Inc.
  6 * Author: Ganapatrao Kulkarni <gkulkarni@cavium.com>
  7 */
  8
  9#define pr_fmt(fmt) "NUMA: " fmt
 10
 11#include <linux/acpi.h>
 12#include <linux/memblock.h>
 13#include <linux/module.h>
 14#include <linux/of.h>
 15
 16#include <asm/sections.h>
 17
 18struct pglist_data *node_data[MAX_NUMNODES] __read_mostly;
 19EXPORT_SYMBOL(node_data);
 20nodemask_t numa_nodes_parsed __initdata;
 21static int cpu_to_node_map[NR_CPUS] = { [0 ... NR_CPUS-1] = NUMA_NO_NODE };
 22
 23static int numa_distance_cnt;
 24static u8 *numa_distance;
 25bool numa_off;
 26
 27static __init int numa_parse_early_param(char *opt)
 28{
 29	if (!opt)
 30		return -EINVAL;
 31	if (str_has_prefix(opt, "off"))
 32		numa_off = true;
 33
 34	return 0;
 35}
 36early_param("numa", numa_parse_early_param);
 37
 38cpumask_var_t node_to_cpumask_map[MAX_NUMNODES];
 39EXPORT_SYMBOL(node_to_cpumask_map);
 40
 41#ifdef CONFIG_DEBUG_PER_CPU_MAPS
 42
 43/*
 44 * Returns a pointer to the bitmask of CPUs on Node 'node'.
 45 */
 46const struct cpumask *cpumask_of_node(int node)
 47{
 48
 49	if (node == NUMA_NO_NODE)
 50		return cpu_all_mask;
 51
 52	if (WARN_ON(node < 0 || node >= nr_node_ids))
 53		return cpu_none_mask;
 54
 55	if (WARN_ON(node_to_cpumask_map[node] == NULL))
 56		return cpu_online_mask;
 57
 58	return node_to_cpumask_map[node];
 59}
 60EXPORT_SYMBOL(cpumask_of_node);
 61
 62#endif
 63
 64static void numa_update_cpu(unsigned int cpu, bool remove)
 65{
 66	int nid = cpu_to_node(cpu);
 67
 68	if (nid == NUMA_NO_NODE)
 69		return;
 70
 71	if (remove)
 72		cpumask_clear_cpu(cpu, node_to_cpumask_map[nid]);
 73	else
 74		cpumask_set_cpu(cpu, node_to_cpumask_map[nid]);
 75}
 76
 77void numa_add_cpu(unsigned int cpu)
 78{
 79	numa_update_cpu(cpu, false);
 80}
 81
 82void numa_remove_cpu(unsigned int cpu)
 83{
 84	numa_update_cpu(cpu, true);
 85}
 86
 87void numa_clear_node(unsigned int cpu)
 88{
 89	numa_remove_cpu(cpu);
 90	set_cpu_numa_node(cpu, NUMA_NO_NODE);
 91}
 92
 93/*
 94 * Allocate node_to_cpumask_map based on number of available nodes
 95 * Requires node_possible_map to be valid.
 96 *
 97 * Note: cpumask_of_node() is not valid until after this is done.
 98 * (Use CONFIG_DEBUG_PER_CPU_MAPS to check this.)
 99 */
100static void __init setup_node_to_cpumask_map(void)
101{
102	int node;
103
104	/* setup nr_node_ids if not done yet */
105	if (nr_node_ids == MAX_NUMNODES)
106		setup_nr_node_ids();
107
108	/* allocate and clear the mapping */
109	for (node = 0; node < nr_node_ids; node++) {
110		alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]);
111		cpumask_clear(node_to_cpumask_map[node]);
112	}
113
114	/* cpumask_of_node() will now work */
115	pr_debug("Node to cpumask map for %u nodes\n", nr_node_ids);
116}
117
118/*
119 * Set the cpu to node and mem mapping
120 */
121void numa_store_cpu_info(unsigned int cpu)
122{
123	set_cpu_numa_node(cpu, cpu_to_node_map[cpu]);
124}
125
126void __init early_map_cpu_to_node(unsigned int cpu, int nid)
127{
128	/* fallback to node 0 */
129	if (nid < 0 || nid >= MAX_NUMNODES || numa_off)
130		nid = 0;
131
132	cpu_to_node_map[cpu] = nid;
133
134	/*
135	 * We should set the numa node of cpu0 as soon as possible, because it
136	 * has already been set up online before. cpu_to_node(0) will soon be
137	 * called.
138	 */
139	if (!cpu)
140		set_cpu_numa_node(cpu, nid);
141}
142
143#ifdef CONFIG_HAVE_SETUP_PER_CPU_AREA
144unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
145EXPORT_SYMBOL(__per_cpu_offset);
146
147static int __init early_cpu_to_node(int cpu)
148{
149	return cpu_to_node_map[cpu];
150}
151
152static int __init pcpu_cpu_distance(unsigned int from, unsigned int to)
153{
154	return node_distance(early_cpu_to_node(from), early_cpu_to_node(to));
155}
156
157static void * __init pcpu_fc_alloc(unsigned int cpu, size_t size,
158				       size_t align)
159{
160	int nid = early_cpu_to_node(cpu);
161
162	return  memblock_alloc_try_nid(size, align,
163			__pa(MAX_DMA_ADDRESS), MEMBLOCK_ALLOC_ACCESSIBLE, nid);
164}
165
166static void __init pcpu_fc_free(void *ptr, size_t size)
167{
168	memblock_free_early(__pa(ptr), size);
169}
170
171void __init setup_per_cpu_areas(void)
172{
173	unsigned long delta;
174	unsigned int cpu;
175	int rc;
176
177	/*
178	 * Always reserve area for module percpu variables.  That's
179	 * what the legacy allocator did.
180	 */
181	rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
182				    PERCPU_DYNAMIC_RESERVE, PAGE_SIZE,
183				    pcpu_cpu_distance,
184				    pcpu_fc_alloc, pcpu_fc_free);
185	if (rc < 0)
186		panic("Failed to initialize percpu areas.");
187
188	delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
189	for_each_possible_cpu(cpu)
190		__per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
191}
192#endif
193
194/**
195 * numa_add_memblk() - Set node id to memblk
196 * @nid: NUMA node ID of the new memblk
197 * @start: Start address of the new memblk
198 * @end:  End address of the new memblk
199 *
200 * RETURNS:
201 * 0 on success, -errno on failure.
202 */
203int __init numa_add_memblk(int nid, u64 start, u64 end)
204{
205	int ret;
206
207	ret = memblock_set_node(start, (end - start), &memblock.memory, nid);
208	if (ret < 0) {
209		pr_err("memblock [0x%llx - 0x%llx] failed to add on node %d\n",
210			start, (end - 1), nid);
211		return ret;
212	}
213
214	node_set(nid, numa_nodes_parsed);
215	return ret;
216}
217
218/*
219 * Initialize NODE_DATA for a node on the local memory
220 */
221static void __init setup_node_data(int nid, u64 start_pfn, u64 end_pfn)
222{
223	const size_t nd_size = roundup(sizeof(pg_data_t), SMP_CACHE_BYTES);
224	u64 nd_pa;
225	void *nd;
226	int tnid;
227
228	if (start_pfn >= end_pfn)
229		pr_info("Initmem setup node %d [<memory-less node>]\n", nid);
230
231	nd_pa = memblock_phys_alloc_try_nid(nd_size, SMP_CACHE_BYTES, nid);
232	if (!nd_pa)
233		panic("Cannot allocate %zu bytes for node %d data\n",
234		      nd_size, nid);
235
236	nd = __va(nd_pa);
237
238	/* report and initialize */
239	pr_info("NODE_DATA [mem %#010Lx-%#010Lx]\n",
240		nd_pa, nd_pa + nd_size - 1);
241	tnid = early_pfn_to_nid(nd_pa >> PAGE_SHIFT);
242	if (tnid != nid)
243		pr_info("NODE_DATA(%d) on node %d\n", nid, tnid);
244
245	node_data[nid] = nd;
246	memset(NODE_DATA(nid), 0, sizeof(pg_data_t));
247	NODE_DATA(nid)->node_id = nid;
248	NODE_DATA(nid)->node_start_pfn = start_pfn;
249	NODE_DATA(nid)->node_spanned_pages = end_pfn - start_pfn;
250}
251
252/*
253 * numa_free_distance
254 *
255 * The current table is freed.
256 */
257void __init numa_free_distance(void)
258{
259	size_t size;
260
261	if (!numa_distance)
262		return;
263
264	size = numa_distance_cnt * numa_distance_cnt *
265		sizeof(numa_distance[0]);
266
267	memblock_free(__pa(numa_distance), size);
268	numa_distance_cnt = 0;
269	numa_distance = NULL;
270}
271
272/*
273 * Create a new NUMA distance table.
274 */
275static int __init numa_alloc_distance(void)
276{
277	size_t size;
278	u64 phys;
279	int i, j;
280
281	size = nr_node_ids * nr_node_ids * sizeof(numa_distance[0]);
282	phys = memblock_find_in_range(0, PFN_PHYS(max_pfn),
283				      size, PAGE_SIZE);
284	if (WARN_ON(!phys))
285		return -ENOMEM;
286
287	memblock_reserve(phys, size);
288
289	numa_distance = __va(phys);
290	numa_distance_cnt = nr_node_ids;
291
292	/* fill with the default distances */
293	for (i = 0; i < numa_distance_cnt; i++)
294		for (j = 0; j < numa_distance_cnt; j++)
295			numa_distance[i * numa_distance_cnt + j] = i == j ?
296				LOCAL_DISTANCE : REMOTE_DISTANCE;
297
298	pr_debug("Initialized distance table, cnt=%d\n", numa_distance_cnt);
299
300	return 0;
301}
302
303/**
304 * numa_set_distance() - Set inter node NUMA distance from node to node.
305 * @from: the 'from' node to set distance
306 * @to: the 'to'  node to set distance
307 * @distance: NUMA distance
308 *
309 * Set the distance from node @from to @to to @distance.
310 * If distance table doesn't exist, a warning is printed.
311 *
312 * If @from or @to is higher than the highest known node or lower than zero
313 * or @distance doesn't make sense, the call is ignored.
314 */
315void __init numa_set_distance(int from, int to, int distance)
316{
317	if (!numa_distance) {
318		pr_warn_once("Warning: distance table not allocated yet\n");
319		return;
320	}
321
322	if (from >= numa_distance_cnt || to >= numa_distance_cnt ||
323			from < 0 || to < 0) {
324		pr_warn_once("Warning: node ids are out of bound, from=%d to=%d distance=%d\n",
325			    from, to, distance);
326		return;
327	}
328
329	if ((u8)distance != distance ||
330	    (from == to && distance != LOCAL_DISTANCE)) {
331		pr_warn_once("Warning: invalid distance parameter, from=%d to=%d distance=%d\n",
332			     from, to, distance);
333		return;
334	}
335
336	numa_distance[from * numa_distance_cnt + to] = distance;
337}
338
339/*
340 * Return NUMA distance @from to @to
341 */
342int __node_distance(int from, int to)
343{
344	if (from >= numa_distance_cnt || to >= numa_distance_cnt)
345		return from == to ? LOCAL_DISTANCE : REMOTE_DISTANCE;
346	return numa_distance[from * numa_distance_cnt + to];
347}
348EXPORT_SYMBOL(__node_distance);
349
350static int __init numa_register_nodes(void)
351{
352	int nid;
353	struct memblock_region *mblk;
354
355	/* Check that valid nid is set to memblks */
356	for_each_mem_region(mblk) {
357		int mblk_nid = memblock_get_region_node(mblk);
358		phys_addr_t start = mblk->base;
359		phys_addr_t end = mblk->base + mblk->size - 1;
360
361		if (mblk_nid == NUMA_NO_NODE || mblk_nid >= MAX_NUMNODES) {
362			pr_warn("Warning: invalid memblk node %d [mem %pap-%pap]\n",
363				mblk_nid, &start, &end);
364			return -EINVAL;
365		}
366	}
367
368	/* Finally register nodes. */
369	for_each_node_mask(nid, numa_nodes_parsed) {
370		unsigned long start_pfn, end_pfn;
371
372		get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
373		setup_node_data(nid, start_pfn, end_pfn);
374		node_set_online(nid);
375	}
376
377	/* Setup online nodes to actual nodes*/
378	node_possible_map = numa_nodes_parsed;
379
380	return 0;
381}
382
383static int __init numa_init(int (*init_func)(void))
384{
385	int ret;
386
387	nodes_clear(numa_nodes_parsed);
388	nodes_clear(node_possible_map);
389	nodes_clear(node_online_map);
390
391	ret = numa_alloc_distance();
392	if (ret < 0)
393		return ret;
394
395	ret = init_func();
396	if (ret < 0)
397		goto out_free_distance;
398
399	if (nodes_empty(numa_nodes_parsed)) {
400		pr_info("No NUMA configuration found\n");
401		ret = -EINVAL;
402		goto out_free_distance;
403	}
404
405	ret = numa_register_nodes();
406	if (ret < 0)
407		goto out_free_distance;
408
409	setup_node_to_cpumask_map();
410
411	return 0;
412out_free_distance:
413	numa_free_distance();
414	return ret;
415}
416
417/**
418 * dummy_numa_init() - Fallback dummy NUMA init
419 *
420 * Used if there's no underlying NUMA architecture, NUMA initialization
421 * fails, or NUMA is disabled on the command line.
422 *
423 * Must online at least one node (node 0) and add memory blocks that cover all
424 * allowed memory. It is unlikely that this function fails.
425 *
426 * Return: 0 on success, -errno on failure.
427 */
428static int __init dummy_numa_init(void)
429{
430	phys_addr_t start = memblock_start_of_DRAM();
431	phys_addr_t end = memblock_end_of_DRAM() - 1;
432	int ret;
433
434	if (numa_off)
435		pr_info("NUMA disabled\n"); /* Forced off on command line. */
436	pr_info("Faking a node at [mem %pap-%pap]\n", &start, &end);
437
438	ret = numa_add_memblk(0, start, end + 1);
439	if (ret) {
440		pr_err("NUMA init failed\n");
441		return ret;
442	}
443
444	numa_off = true;
445	return 0;
446}
447
448#ifdef CONFIG_ACPI_NUMA
449static int __init arch_acpi_numa_init(void)
450{
451	int ret;
452
453	ret = acpi_numa_init();
454	if (ret) {
455		pr_info("Failed to initialise from firmware\n");
456		return ret;
457	}
458
459	return srat_disabled() ? -EINVAL : 0;
460}
461#else
462static int __init arch_acpi_numa_init(void)
463{
464	return -EOPNOTSUPP;
465}
466#endif
467
468/**
469 * arch_numa_init() - Initialize NUMA
470 *
471 * Try each configured NUMA initialization method until one succeeds. The
472 * last fallback is dummy single node config encompassing whole memory.
473 */
474void __init arch_numa_init(void)
475{
476	if (!numa_off) {
477		if (!acpi_disabled && !numa_init(arch_acpi_numa_init))
478			return;
479		if (acpi_disabled && !numa_init(of_numa_init))
480			return;
481	}
482
483	numa_init(dummy_numa_init);
484}