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#include <linux/numa_memblks.h>
 16
 17#include <asm/sections.h>
 18
 19static int cpu_to_node_map[NR_CPUS] = { [0 ... NR_CPUS-1] = NUMA_NO_NODE };
 20
 21bool numa_off;
 22
 23static __init int numa_parse_early_param(char *opt)
 24{
 25	if (!opt)
 26		return -EINVAL;
 27	if (str_has_prefix(opt, "off"))
 28		numa_off = true;
 29	if (!strncmp(opt, "fake=", 5))
 30		return numa_emu_cmdline(opt + 5);
 31
 32	return 0;
 33}
 34early_param("numa", numa_parse_early_param);
 35
 36cpumask_var_t node_to_cpumask_map[MAX_NUMNODES];
 37EXPORT_SYMBOL(node_to_cpumask_map);
 38
 39#ifdef CONFIG_DEBUG_PER_CPU_MAPS
 40
 41/*
 42 * Returns a pointer to the bitmask of CPUs on Node 'node'.
 43 */
 44const struct cpumask *cpumask_of_node(int node)
 45{
 46
 47	if (node == NUMA_NO_NODE)
 48		return cpu_all_mask;
 49
 50	if (WARN_ON(node < 0 || node >= nr_node_ids))
 51		return cpu_none_mask;
 52
 53	if (WARN_ON(node_to_cpumask_map[node] == NULL))
 54		return cpu_online_mask;
 55
 56	return node_to_cpumask_map[node];
 57}
 58EXPORT_SYMBOL(cpumask_of_node);
 59
 60#endif
 61
 62#ifndef CONFIG_NUMA_EMU
 63static void numa_update_cpu(unsigned int cpu, bool remove)
 64{
 65	int nid = cpu_to_node(cpu);
 66
 67	if (nid == NUMA_NO_NODE)
 68		return;
 69
 70	if (remove)
 71		cpumask_clear_cpu(cpu, node_to_cpumask_map[nid]);
 72	else
 73		cpumask_set_cpu(cpu, node_to_cpumask_map[nid]);
 74}
 75
 76void numa_add_cpu(unsigned int cpu)
 77{
 78	numa_update_cpu(cpu, false);
 79}
 80
 81void numa_remove_cpu(unsigned int cpu)
 82{
 83	numa_update_cpu(cpu, true);
 84}
 85#endif
 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
147int 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
157void __init setup_per_cpu_areas(void)
158{
159	unsigned long delta;
160	unsigned int cpu;
161	int rc = -EINVAL;
162
163	if (pcpu_chosen_fc != PCPU_FC_PAGE) {
164		/*
165		 * Always reserve area for module percpu variables.  That's
166		 * what the legacy allocator did.
167		 */
168		rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
169					    PERCPU_DYNAMIC_RESERVE, PAGE_SIZE,
170					    pcpu_cpu_distance,
171					    early_cpu_to_node);
172#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
173		if (rc < 0)
174			pr_warn("PERCPU: %s allocator failed (%d), falling back to page size\n",
175				   pcpu_fc_names[pcpu_chosen_fc], rc);
176#endif
177	}
178
179#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
180	if (rc < 0)
181		rc = pcpu_page_first_chunk(PERCPU_MODULE_RESERVE, early_cpu_to_node);
182#endif
183	if (rc < 0)
184		panic("Failed to initialize percpu areas (err=%d).", rc);
185
186	delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
187	for_each_possible_cpu(cpu)
188		__per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
189}
190#endif
191
192/*
193 * Initialize NODE_DATA for a node on the local memory
194 */
195static void __init setup_node_data(int nid, u64 start_pfn, u64 end_pfn)
196{
197	if (start_pfn >= end_pfn)
198		pr_info("Initmem setup node %d [<memory-less node>]\n", nid);
199
200	alloc_node_data(nid);
201
202	NODE_DATA(nid)->node_id = nid;
203	NODE_DATA(nid)->node_start_pfn = start_pfn;
204	NODE_DATA(nid)->node_spanned_pages = end_pfn - start_pfn;
205}
206
207static int __init numa_register_nodes(void)
208{
209	int nid;
210
211	/* Check the validity of the memblock/node mapping */
212	if (!memblock_validate_numa_coverage(0))
213		return -EINVAL;
214
215	/* Finally register nodes. */
216	for_each_node_mask(nid, numa_nodes_parsed) {
217		unsigned long start_pfn, end_pfn;
218
219		get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
220		setup_node_data(nid, start_pfn, end_pfn);
221		node_set_online(nid);
222	}
223
224	/* Setup online nodes to actual nodes*/
225	node_possible_map = numa_nodes_parsed;
226
227	return 0;
228}
229
230static int __init numa_init(int (*init_func)(void))
231{
232	int ret;
233
234	nodes_clear(numa_nodes_parsed);
235	nodes_clear(node_possible_map);
236	nodes_clear(node_online_map);
237
238	ret = numa_memblks_init(init_func, /* memblock_force_top_down */ false);
239	if (ret < 0)
240		goto out_free_distance;
241
242	if (nodes_empty(numa_nodes_parsed)) {
243		pr_info("No NUMA configuration found\n");
244		ret = -EINVAL;
245		goto out_free_distance;
246	}
247
248	ret = numa_register_nodes();
249	if (ret < 0)
250		goto out_free_distance;
251
252	setup_node_to_cpumask_map();
253
254	return 0;
255out_free_distance:
256	numa_reset_distance();
257	return ret;
258}
259
260/**
261 * dummy_numa_init() - Fallback dummy NUMA init
262 *
263 * Used if there's no underlying NUMA architecture, NUMA initialization
264 * fails, or NUMA is disabled on the command line.
265 *
266 * Must online at least one node (node 0) and add memory blocks that cover all
267 * allowed memory. It is unlikely that this function fails.
268 *
269 * Return: 0 on success, -errno on failure.
270 */
271static int __init dummy_numa_init(void)
272{
273	phys_addr_t start = memblock_start_of_DRAM();
274	phys_addr_t end = memblock_end_of_DRAM() - 1;
275	int ret;
276
277	if (numa_off)
278		pr_info("NUMA disabled\n"); /* Forced off on command line. */
279	pr_info("Faking a node at [mem %pap-%pap]\n", &start, &end);
280
281	ret = numa_add_memblk(0, start, end + 1);
282	if (ret) {
283		pr_err("NUMA init failed\n");
284		return ret;
285	}
286	node_set(0, numa_nodes_parsed);
287
288	numa_off = true;
289	return 0;
290}
291
292#ifdef CONFIG_ACPI_NUMA
293static int __init arch_acpi_numa_init(void)
294{
295	int ret;
296
297	ret = acpi_numa_init();
298	if (ret) {
299		pr_debug("Failed to initialise from firmware\n");
300		return ret;
301	}
302
303	return srat_disabled() ? -EINVAL : 0;
304}
305#else
306static int __init arch_acpi_numa_init(void)
307{
308	return -EOPNOTSUPP;
309}
310#endif
311
312/**
313 * arch_numa_init() - Initialize NUMA
314 *
315 * Try each configured NUMA initialization method until one succeeds. The
316 * last fallback is dummy single node config encompassing whole memory.
317 */
318void __init arch_numa_init(void)
319{
320	if (!numa_off) {
321		if (!acpi_disabled && !numa_init(arch_acpi_numa_init))
322			return;
323		if (acpi_disabled && !numa_init(of_numa_init))
324			return;
325	}
326
327	numa_init(dummy_numa_init);
328}
329
330#ifdef CONFIG_NUMA_EMU
331void __init numa_emu_update_cpu_to_node(int *emu_nid_to_phys,
332					unsigned int nr_emu_nids)
333{
334	int i, j;
335
336	/*
337	 * Transform cpu_to_node_map table to use emulated nids by
338	 * reverse-mapping phys_nid.  The maps should always exist but fall
339	 * back to zero just in case.
340	 */
341	for (i = 0; i < ARRAY_SIZE(cpu_to_node_map); i++) {
342		if (cpu_to_node_map[i] == NUMA_NO_NODE)
343			continue;
344		for (j = 0; j < nr_emu_nids; j++)
345			if (cpu_to_node_map[i] == emu_nid_to_phys[j])
346				break;
347		cpu_to_node_map[i] = j < nr_emu_nids ? j : 0;
348	}
349}
350
351u64 __init numa_emu_dma_end(void)
352{
353	return memblock_start_of_DRAM() + SZ_4G;
354}
355
356void debug_cpumask_set_cpu(unsigned int cpu, int node, bool enable)
357{
358	struct cpumask *mask;
359
360	if (node == NUMA_NO_NODE)
361		return;
362
363	mask = node_to_cpumask_map[node];
364	if (!cpumask_available(mask)) {
365		pr_err("node_to_cpumask_map[%i] NULL\n", node);
366		dump_stack();
367		return;
368	}
369
370	if (enable)
371		cpumask_set_cpu(cpu, mask);
372	else
373		cpumask_clear_cpu(cpu, mask);
374
375	pr_debug("%s cpu %d node %d: mask now %*pbl\n",
376		 enable ? "numa_add_cpu" : "numa_remove_cpu",
377		 cpu, node, cpumask_pr_args(mask));
378}
379#endif /* CONFIG_NUMA_EMU */