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
19 static int cpu_to_node_map[NR_CPUS] = { [0 ... NR_CPUS-1] = NUMA_NO_NODE };
20
21 bool numa_off;
22
numa_parse_early_param(char * opt)23 static __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 }
34 early_param("numa", numa_parse_early_param);
35
36 cpumask_var_t node_to_cpumask_map[MAX_NUMNODES];
37 EXPORT_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 */
cpumask_of_node(int node)44 const 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 }
58 EXPORT_SYMBOL(cpumask_of_node);
59
60 #endif
61
62 #ifndef CONFIG_NUMA_EMU
numa_update_cpu(unsigned int cpu,bool remove)63 static 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
numa_add_cpu(unsigned int cpu)76 void numa_add_cpu(unsigned int cpu)
77 {
78 numa_update_cpu(cpu, false);
79 }
80
numa_remove_cpu(unsigned int cpu)81 void numa_remove_cpu(unsigned int cpu)
82 {
83 numa_update_cpu(cpu, true);
84 }
85 #endif
86
numa_clear_node(unsigned int cpu)87 void 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 */
setup_node_to_cpumask_map(void)100 static 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 /*
109 * This check should never be true but it makes it clear to compilers
110 * that node_to_cpumask_map is bound by nr_node_ids, avoiding false
111 * positive fortify warnings when accessing node_to_cpumask_map in the
112 * for loop below.
113 */
114 if (unlikely(nr_node_ids > MAX_NUMNODES)) {
115 pr_err("nr_node_ids (%u) is larger than MAX_NUMNODES (%u)\n",
116 nr_node_ids, MAX_NUMNODES);
117 return;
118 }
119
120 /* allocate and clear the mapping */
121 for (node = 0; node < nr_node_ids; node++) {
122 alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]);
123 cpumask_clear(node_to_cpumask_map[node]);
124 }
125
126 /* cpumask_of_node() will now work */
127 pr_debug("Node to cpumask map for %u nodes\n", nr_node_ids);
128 }
129
130 /*
131 * Set the cpu to node and mem mapping
132 */
numa_store_cpu_info(unsigned int cpu)133 void numa_store_cpu_info(unsigned int cpu)
134 {
135 set_cpu_numa_node(cpu, cpu_to_node_map[cpu]);
136 }
137
early_map_cpu_to_node(unsigned int cpu,int nid)138 void __init early_map_cpu_to_node(unsigned int cpu, int nid)
139 {
140 /* fallback to node 0 */
141 if (nid < 0 || nid >= MAX_NUMNODES || numa_off)
142 nid = 0;
143
144 cpu_to_node_map[cpu] = nid;
145
146 /*
147 * We should set the numa node of cpu0 as soon as possible, because it
148 * has already been set up online before. cpu_to_node(0) will soon be
149 * called.
150 */
151 if (!cpu)
152 set_cpu_numa_node(cpu, nid);
153 }
154
155 #ifdef CONFIG_HAVE_SETUP_PER_CPU_AREA
156 unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
157 EXPORT_SYMBOL(__per_cpu_offset);
158
early_cpu_to_node(int cpu)159 int early_cpu_to_node(int cpu)
160 {
161 return cpu_to_node_map[cpu];
162 }
163
pcpu_cpu_distance(unsigned int from,unsigned int to)164 static int __init pcpu_cpu_distance(unsigned int from, unsigned int to)
165 {
166 return node_distance(early_cpu_to_node(from), early_cpu_to_node(to));
167 }
168
setup_per_cpu_areas(void)169 void __init setup_per_cpu_areas(void)
170 {
171 unsigned long delta;
172 unsigned int cpu;
173 int rc = -EINVAL;
174
175 if (pcpu_chosen_fc != PCPU_FC_PAGE) {
176 /*
177 * Always reserve area for module percpu variables. That's
178 * what the legacy allocator did.
179 */
180 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
181 PERCPU_DYNAMIC_RESERVE, PAGE_SIZE,
182 pcpu_cpu_distance,
183 early_cpu_to_node);
184 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
185 if (rc < 0)
186 pr_warn("PERCPU: %s allocator failed (%d), falling back to page size\n",
187 pcpu_fc_names[pcpu_chosen_fc], rc);
188 #endif
189 }
190
191 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
192 if (rc < 0)
193 rc = pcpu_page_first_chunk(PERCPU_MODULE_RESERVE, early_cpu_to_node);
194 #endif
195 if (rc < 0)
196 panic("Failed to initialize percpu areas (err=%d).", rc);
197
198 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
199 for_each_possible_cpu(cpu)
200 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
201 }
202 #endif
203
204 /*
205 * Initialize NODE_DATA for a node on the local memory
206 */
setup_node_data(int nid,u64 start_pfn,u64 end_pfn)207 static void __init setup_node_data(int nid, u64 start_pfn, u64 end_pfn)
208 {
209 if (start_pfn >= end_pfn)
210 pr_info("Initmem setup node %d [<memory-less node>]\n", nid);
211
212 alloc_node_data(nid);
213
214 NODE_DATA(nid)->node_id = nid;
215 NODE_DATA(nid)->node_start_pfn = start_pfn;
216 NODE_DATA(nid)->node_spanned_pages = end_pfn - start_pfn;
217 }
218
numa_register_nodes(void)219 static int __init numa_register_nodes(void)
220 {
221 int nid;
222
223 /* Check the validity of the memblock/node mapping */
224 if (!memblock_validate_numa_coverage(0))
225 return -EINVAL;
226
227 /* Finally register nodes. */
228 for_each_node_mask(nid, numa_nodes_parsed) {
229 unsigned long start_pfn, end_pfn;
230
231 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
232 setup_node_data(nid, start_pfn, end_pfn);
233 node_set_online(nid);
234 }
235
236 return 0;
237 }
238
numa_init(int (* init_func)(void))239 static int __init numa_init(int (*init_func)(void))
240 {
241 int ret;
242
243 ret = numa_memblks_init(init_func, /* memblock_force_top_down */ false);
244 if (ret < 0)
245 goto out_free_distance;
246
247 if (nodes_empty(numa_nodes_parsed)) {
248 pr_info("No NUMA configuration found\n");
249 ret = -EINVAL;
250 goto out_free_distance;
251 }
252
253 ret = numa_register_nodes();
254 if (ret < 0)
255 goto out_free_distance;
256
257 setup_node_to_cpumask_map();
258
259 return 0;
260 out_free_distance:
261 numa_reset_distance();
262 return ret;
263 }
264
265 /**
266 * dummy_numa_init() - Fallback dummy NUMA init
267 *
268 * Used if there's no underlying NUMA architecture, NUMA initialization
269 * fails, or NUMA is disabled on the command line.
270 *
271 * Must online at least one node (node 0) and add memory blocks that cover all
272 * allowed memory. It is unlikely that this function fails.
273 *
274 * Return: 0 on success, -errno on failure.
275 */
dummy_numa_init(void)276 static int __init dummy_numa_init(void)
277 {
278 phys_addr_t start = memblock_start_of_DRAM();
279 phys_addr_t end = memblock_end_of_DRAM() - 1;
280 int ret;
281
282 if (numa_off)
283 pr_info("NUMA disabled\n"); /* Forced off on command line. */
284 pr_info("Faking a node at [mem %pap-%pap]\n", &start, &end);
285
286 ret = numa_add_memblk(0, start, end + 1);
287 if (ret) {
288 pr_err("NUMA init failed\n");
289 return ret;
290 }
291
292 numa_off = true;
293 return 0;
294 }
295
296 #ifdef CONFIG_ACPI_NUMA
arch_acpi_numa_init(void)297 static int __init arch_acpi_numa_init(void)
298 {
299 int ret;
300
301 ret = acpi_numa_init();
302 if (ret) {
303 pr_debug("Failed to initialise from firmware\n");
304 return ret;
305 }
306
307 return srat_disabled() ? -EINVAL : 0;
308 }
309 #else
arch_acpi_numa_init(void)310 static int __init arch_acpi_numa_init(void)
311 {
312 return -EOPNOTSUPP;
313 }
314 #endif
315
316 /**
317 * arch_numa_init() - Initialize NUMA
318 *
319 * Try each configured NUMA initialization method until one succeeds. The
320 * last fallback is dummy single node config encompassing whole memory.
321 */
arch_numa_init(void)322 void __init arch_numa_init(void)
323 {
324 if (!numa_off) {
325 if (!acpi_disabled && !numa_init(arch_acpi_numa_init))
326 return;
327 if (acpi_disabled && !numa_init(of_numa_init))
328 return;
329 }
330
331 numa_init(dummy_numa_init);
332 }
333
334 #ifdef CONFIG_NUMA_EMU
numa_emu_update_cpu_to_node(int * emu_nid_to_phys,unsigned int nr_emu_nids)335 void __init numa_emu_update_cpu_to_node(int *emu_nid_to_phys,
336 unsigned int nr_emu_nids)
337 {
338 int i, j;
339
340 /*
341 * Transform cpu_to_node_map table to use emulated nids by
342 * reverse-mapping phys_nid. The maps should always exist but fall
343 * back to zero just in case.
344 */
345 for (i = 0; i < ARRAY_SIZE(cpu_to_node_map); i++) {
346 if (cpu_to_node_map[i] == NUMA_NO_NODE)
347 continue;
348 for (j = 0; j < nr_emu_nids; j++)
349 if (cpu_to_node_map[i] == emu_nid_to_phys[j])
350 break;
351 cpu_to_node_map[i] = j < nr_emu_nids ? j : 0;
352 }
353 }
354
numa_emu_dma_end(void)355 u64 __init numa_emu_dma_end(void)
356 {
357 return memblock_start_of_DRAM() + SZ_4G;
358 }
359
debug_cpumask_set_cpu(unsigned int cpu,int node,bool enable)360 void debug_cpumask_set_cpu(unsigned int cpu, int node, bool enable)
361 {
362 struct cpumask *mask;
363
364 if (node == NUMA_NO_NODE)
365 return;
366
367 mask = node_to_cpumask_map[node];
368 if (!cpumask_available(mask)) {
369 pr_err("node_to_cpumask_map[%i] NULL\n", node);
370 dump_stack();
371 return;
372 }
373
374 if (enable)
375 cpumask_set_cpu(cpu, mask);
376 else
377 cpumask_clear_cpu(cpu, mask);
378
379 pr_debug("%s cpu %d node %d: mask now %*pbl\n",
380 enable ? "numa_add_cpu" : "numa_remove_cpu",
381 cpu, node, cpumask_pr_args(mask));
382 }
383 #endif /* CONFIG_NUMA_EMU */
384