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 | |
18 | struct pglist_data *node_data[MAX_NUMNODES] __read_mostly; |
19 | EXPORT_SYMBOL(node_data); |
20 | nodemask_t numa_nodes_parsed __initdata; |
21 | static int cpu_to_node_map[NR_CPUS] = { [0 ... NR_CPUS-1] = NUMA_NO_NODE }; |
22 | |
23 | static int numa_distance_cnt; |
24 | static u8 *numa_distance; |
25 | bool numa_off; |
26 | |
27 | static __init int numa_parse_early_param(char *opt) |
28 | { |
29 | if (!opt) |
30 | return -EINVAL; |
31 | if (str_has_prefix(str: opt, prefix: "off" )) |
32 | numa_off = true; |
33 | |
34 | return 0; |
35 | } |
36 | early_param("numa" , numa_parse_early_param); |
37 | |
38 | cpumask_var_t node_to_cpumask_map[MAX_NUMNODES]; |
39 | EXPORT_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 | */ |
46 | const 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 | } |
60 | EXPORT_SYMBOL(cpumask_of_node); |
61 | |
62 | #endif |
63 | |
64 | static 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, dstp: node_to_cpumask_map[nid]); |
73 | else |
74 | cpumask_set_cpu(cpu, dstp: node_to_cpumask_map[nid]); |
75 | } |
76 | |
77 | void numa_add_cpu(unsigned int cpu) |
78 | { |
79 | numa_update_cpu(cpu, remove: false); |
80 | } |
81 | |
82 | void numa_remove_cpu(unsigned int cpu) |
83 | { |
84 | numa_update_cpu(cpu, remove: true); |
85 | } |
86 | |
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 | */ |
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 | /* allocate and clear the mapping */ |
109 | for (node = 0; node < nr_node_ids; node++) { |
110 | alloc_bootmem_cpumask_var(mask: &node_to_cpumask_map[node]); |
111 | cpumask_clear(dstp: 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 | */ |
121 | void numa_store_cpu_info(unsigned int cpu) |
122 | { |
123 | set_cpu_numa_node(cpu, node: cpu_to_node_map[cpu]); |
124 | } |
125 | |
126 | void __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, node: nid); |
141 | } |
142 | |
143 | #ifdef CONFIG_HAVE_SETUP_PER_CPU_AREA |
144 | unsigned long __per_cpu_offset[NR_CPUS] __read_mostly; |
145 | EXPORT_SYMBOL(__per_cpu_offset); |
146 | |
147 | static int __init early_cpu_to_node(int cpu) |
148 | { |
149 | return cpu_to_node_map[cpu]; |
150 | } |
151 | |
152 | static 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 | |
157 | void __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 | cpu_distance_fn: pcpu_cpu_distance, |
171 | cpu_to_nd_fn: 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, cpu_to_nd_fn: early_cpu_to_node); |
182 | #endif |
183 | if (rc < 0) |
184 | panic(fmt: "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 | * numa_add_memblk() - Set node id to memblk |
194 | * @nid: NUMA node ID of the new memblk |
195 | * @start: Start address of the new memblk |
196 | * @end: End address of the new memblk |
197 | * |
198 | * RETURNS: |
199 | * 0 on success, -errno on failure. |
200 | */ |
201 | int __init numa_add_memblk(int nid, u64 start, u64 end) |
202 | { |
203 | int ret; |
204 | |
205 | ret = memblock_set_node(base: start, size: (end - start), type: &memblock.memory, nid); |
206 | if (ret < 0) { |
207 | pr_err("memblock [0x%llx - 0x%llx] failed to add on node %d\n" , |
208 | start, (end - 1), nid); |
209 | return ret; |
210 | } |
211 | |
212 | node_set(nid, numa_nodes_parsed); |
213 | return ret; |
214 | } |
215 | |
216 | /* |
217 | * Initialize NODE_DATA for a node on the local memory |
218 | */ |
219 | static void __init setup_node_data(int nid, u64 start_pfn, u64 end_pfn) |
220 | { |
221 | const size_t nd_size = roundup(sizeof(pg_data_t), SMP_CACHE_BYTES); |
222 | u64 nd_pa; |
223 | void *nd; |
224 | int tnid; |
225 | |
226 | if (start_pfn >= end_pfn) |
227 | pr_info("Initmem setup node %d [<memory-less node>]\n" , nid); |
228 | |
229 | nd_pa = memblock_phys_alloc_try_nid(size: nd_size, SMP_CACHE_BYTES, nid); |
230 | if (!nd_pa) |
231 | panic(fmt: "Cannot allocate %zu bytes for node %d data\n" , |
232 | nd_size, nid); |
233 | |
234 | nd = __va(nd_pa); |
235 | |
236 | /* report and initialize */ |
237 | pr_info("NODE_DATA [mem %#010Lx-%#010Lx]\n" , |
238 | nd_pa, nd_pa + nd_size - 1); |
239 | tnid = early_pfn_to_nid(pfn: nd_pa >> PAGE_SHIFT); |
240 | if (tnid != nid) |
241 | pr_info("NODE_DATA(%d) on node %d\n" , nid, tnid); |
242 | |
243 | node_data[nid] = nd; |
244 | memset(NODE_DATA(nid), 0, sizeof(pg_data_t)); |
245 | NODE_DATA(nid)->node_id = nid; |
246 | NODE_DATA(nid)->node_start_pfn = start_pfn; |
247 | NODE_DATA(nid)->node_spanned_pages = end_pfn - start_pfn; |
248 | } |
249 | |
250 | /* |
251 | * numa_free_distance |
252 | * |
253 | * The current table is freed. |
254 | */ |
255 | void __init numa_free_distance(void) |
256 | { |
257 | size_t size; |
258 | |
259 | if (!numa_distance) |
260 | return; |
261 | |
262 | size = numa_distance_cnt * numa_distance_cnt * |
263 | sizeof(numa_distance[0]); |
264 | |
265 | memblock_free(ptr: numa_distance, size); |
266 | numa_distance_cnt = 0; |
267 | numa_distance = NULL; |
268 | } |
269 | |
270 | /* |
271 | * Create a new NUMA distance table. |
272 | */ |
273 | static int __init numa_alloc_distance(void) |
274 | { |
275 | size_t size; |
276 | int i, j; |
277 | |
278 | size = nr_node_ids * nr_node_ids * sizeof(numa_distance[0]); |
279 | numa_distance = memblock_alloc(size, PAGE_SIZE); |
280 | if (WARN_ON(!numa_distance)) |
281 | return -ENOMEM; |
282 | |
283 | numa_distance_cnt = nr_node_ids; |
284 | |
285 | /* fill with the default distances */ |
286 | for (i = 0; i < numa_distance_cnt; i++) |
287 | for (j = 0; j < numa_distance_cnt; j++) |
288 | numa_distance[i * numa_distance_cnt + j] = i == j ? |
289 | LOCAL_DISTANCE : REMOTE_DISTANCE; |
290 | |
291 | pr_debug("Initialized distance table, cnt=%d\n" , numa_distance_cnt); |
292 | |
293 | return 0; |
294 | } |
295 | |
296 | /** |
297 | * numa_set_distance() - Set inter node NUMA distance from node to node. |
298 | * @from: the 'from' node to set distance |
299 | * @to: the 'to' node to set distance |
300 | * @distance: NUMA distance |
301 | * |
302 | * Set the distance from node @from to @to to @distance. |
303 | * If distance table doesn't exist, a warning is printed. |
304 | * |
305 | * If @from or @to is higher than the highest known node or lower than zero |
306 | * or @distance doesn't make sense, the call is ignored. |
307 | */ |
308 | void __init numa_set_distance(int from, int to, int distance) |
309 | { |
310 | if (!numa_distance) { |
311 | pr_warn_once("Warning: distance table not allocated yet\n" ); |
312 | return; |
313 | } |
314 | |
315 | if (from >= numa_distance_cnt || to >= numa_distance_cnt || |
316 | from < 0 || to < 0) { |
317 | pr_warn_once("Warning: node ids are out of bound, from=%d to=%d distance=%d\n" , |
318 | from, to, distance); |
319 | return; |
320 | } |
321 | |
322 | if ((u8)distance != distance || |
323 | (from == to && distance != LOCAL_DISTANCE)) { |
324 | pr_warn_once("Warning: invalid distance parameter, from=%d to=%d distance=%d\n" , |
325 | from, to, distance); |
326 | return; |
327 | } |
328 | |
329 | numa_distance[from * numa_distance_cnt + to] = distance; |
330 | } |
331 | |
332 | /* |
333 | * Return NUMA distance @from to @to |
334 | */ |
335 | int __node_distance(int from, int to) |
336 | { |
337 | if (from >= numa_distance_cnt || to >= numa_distance_cnt) |
338 | return from == to ? LOCAL_DISTANCE : REMOTE_DISTANCE; |
339 | return numa_distance[from * numa_distance_cnt + to]; |
340 | } |
341 | EXPORT_SYMBOL(__node_distance); |
342 | |
343 | static int __init numa_register_nodes(void) |
344 | { |
345 | int nid; |
346 | struct memblock_region *mblk; |
347 | |
348 | /* Check that valid nid is set to memblks */ |
349 | for_each_mem_region(mblk) { |
350 | int mblk_nid = memblock_get_region_node(r: mblk); |
351 | phys_addr_t start = mblk->base; |
352 | phys_addr_t end = mblk->base + mblk->size - 1; |
353 | |
354 | if (mblk_nid == NUMA_NO_NODE || mblk_nid >= MAX_NUMNODES) { |
355 | pr_warn("Warning: invalid memblk node %d [mem %pap-%pap]\n" , |
356 | mblk_nid, &start, &end); |
357 | return -EINVAL; |
358 | } |
359 | } |
360 | |
361 | /* Finally register nodes. */ |
362 | for_each_node_mask(nid, numa_nodes_parsed) { |
363 | unsigned long start_pfn, end_pfn; |
364 | |
365 | get_pfn_range_for_nid(nid, start_pfn: &start_pfn, end_pfn: &end_pfn); |
366 | setup_node_data(nid, start_pfn, end_pfn); |
367 | node_set_online(nid); |
368 | } |
369 | |
370 | /* Setup online nodes to actual nodes*/ |
371 | node_possible_map = numa_nodes_parsed; |
372 | |
373 | return 0; |
374 | } |
375 | |
376 | static int __init numa_init(int (*init_func)(void)) |
377 | { |
378 | int ret; |
379 | |
380 | nodes_clear(numa_nodes_parsed); |
381 | nodes_clear(node_possible_map); |
382 | nodes_clear(node_online_map); |
383 | |
384 | ret = numa_alloc_distance(); |
385 | if (ret < 0) |
386 | return ret; |
387 | |
388 | ret = init_func(); |
389 | if (ret < 0) |
390 | goto out_free_distance; |
391 | |
392 | if (nodes_empty(numa_nodes_parsed)) { |
393 | pr_info("No NUMA configuration found\n" ); |
394 | ret = -EINVAL; |
395 | goto out_free_distance; |
396 | } |
397 | |
398 | ret = numa_register_nodes(); |
399 | if (ret < 0) |
400 | goto out_free_distance; |
401 | |
402 | setup_node_to_cpumask_map(); |
403 | |
404 | return 0; |
405 | out_free_distance: |
406 | numa_free_distance(); |
407 | return ret; |
408 | } |
409 | |
410 | /** |
411 | * dummy_numa_init() - Fallback dummy NUMA init |
412 | * |
413 | * Used if there's no underlying NUMA architecture, NUMA initialization |
414 | * fails, or NUMA is disabled on the command line. |
415 | * |
416 | * Must online at least one node (node 0) and add memory blocks that cover all |
417 | * allowed memory. It is unlikely that this function fails. |
418 | * |
419 | * Return: 0 on success, -errno on failure. |
420 | */ |
421 | static int __init dummy_numa_init(void) |
422 | { |
423 | phys_addr_t start = memblock_start_of_DRAM(); |
424 | phys_addr_t end = memblock_end_of_DRAM() - 1; |
425 | int ret; |
426 | |
427 | if (numa_off) |
428 | pr_info("NUMA disabled\n" ); /* Forced off on command line. */ |
429 | pr_info("Faking a node at [mem %pap-%pap]\n" , &start, &end); |
430 | |
431 | ret = numa_add_memblk(nid: 0, start, end: end + 1); |
432 | if (ret) { |
433 | pr_err("NUMA init failed\n" ); |
434 | return ret; |
435 | } |
436 | |
437 | numa_off = true; |
438 | return 0; |
439 | } |
440 | |
441 | #ifdef CONFIG_ACPI_NUMA |
442 | static int __init arch_acpi_numa_init(void) |
443 | { |
444 | int ret; |
445 | |
446 | ret = acpi_numa_init(); |
447 | if (ret) { |
448 | pr_info("Failed to initialise from firmware\n" ); |
449 | return ret; |
450 | } |
451 | |
452 | return srat_disabled() ? -EINVAL : 0; |
453 | } |
454 | #else |
455 | static int __init arch_acpi_numa_init(void) |
456 | { |
457 | return -EOPNOTSUPP; |
458 | } |
459 | #endif |
460 | |
461 | /** |
462 | * arch_numa_init() - Initialize NUMA |
463 | * |
464 | * Try each configured NUMA initialization method until one succeeds. The |
465 | * last fallback is dummy single node config encompassing whole memory. |
466 | */ |
467 | void __init arch_numa_init(void) |
468 | { |
469 | if (!numa_off) { |
470 | if (!acpi_disabled && !numa_init(init_func: arch_acpi_numa_init)) |
471 | return; |
472 | if (acpi_disabled && !numa_init(init_func: of_numa_init)) |
473 | return; |
474 | } |
475 | |
476 | numa_init(init_func: dummy_numa_init); |
477 | } |
478 | |