Commit | Line | Data |
---|---|---|
caab277b | 1 | // SPDX-License-Identifier: GPL-2.0-only |
1a2db300 GK |
2 | /* |
3 | * NUMA support, based on the x86 implementation. | |
4 | * | |
5 | * Copyright (C) 2015 Cavium Inc. | |
6 | * Author: Ganapatrao Kulkarni <gkulkarni@cavium.com> | |
1a2db300 GK |
7 | */ |
8 | ||
f11c7bac KW |
9 | #define pr_fmt(fmt) "NUMA: " fmt |
10 | ||
d8b47fca | 11 | #include <linux/acpi.h> |
1a2db300 GK |
12 | #include <linux/memblock.h> |
13 | #include <linux/module.h> | |
14 | #include <linux/of.h> | |
15 | ||
bfe6c8a8 | 16 | #include <asm/acpi.h> |
7af3a0a9 | 17 | #include <asm/sections.h> |
bfe6c8a8 | 18 | |
1a2db300 GK |
19 | struct pglist_data *node_data[MAX_NUMNODES] __read_mostly; |
20 | EXPORT_SYMBOL(node_data); | |
21 | nodemask_t numa_nodes_parsed __initdata; | |
22 | static int cpu_to_node_map[NR_CPUS] = { [0 ... NR_CPUS-1] = NUMA_NO_NODE }; | |
23 | ||
24 | static int numa_distance_cnt; | |
25 | static u8 *numa_distance; | |
aec03f89 | 26 | bool numa_off; |
1a2db300 GK |
27 | |
28 | static __init int numa_parse_early_param(char *opt) | |
29 | { | |
30 | if (!opt) | |
31 | return -EINVAL; | |
b3e089cd | 32 | if (str_has_prefix(opt, "off")) |
34c33370 | 33 | numa_off = true; |
f11c7bac | 34 | |
1a2db300 GK |
35 | return 0; |
36 | } | |
37 | early_param("numa", numa_parse_early_param); | |
38 | ||
39 | cpumask_var_t node_to_cpumask_map[MAX_NUMNODES]; | |
40 | EXPORT_SYMBOL(node_to_cpumask_map); | |
41 | ||
42 | #ifdef CONFIG_DEBUG_PER_CPU_MAPS | |
43 | ||
44 | /* | |
45 | * Returns a pointer to the bitmask of CPUs on Node 'node'. | |
46 | */ | |
47 | const struct cpumask *cpumask_of_node(int node) | |
48 | { | |
49 | if (WARN_ON(node >= nr_node_ids)) | |
50 | return cpu_none_mask; | |
51 | ||
52 | if (WARN_ON(node_to_cpumask_map[node] == NULL)) | |
53 | return cpu_online_mask; | |
54 | ||
55 | return node_to_cpumask_map[node]; | |
56 | } | |
57 | EXPORT_SYMBOL(cpumask_of_node); | |
58 | ||
59 | #endif | |
60 | ||
97fd6016 | 61 | static void numa_update_cpu(unsigned int cpu, bool remove) |
1a2db300 | 62 | { |
97fd6016 SH |
63 | int nid = cpu_to_node(cpu); |
64 | ||
65 | if (nid == NUMA_NO_NODE) | |
66 | return; | |
67 | ||
68 | if (remove) | |
69 | cpumask_clear_cpu(cpu, node_to_cpumask_map[nid]); | |
70 | else | |
1a2db300 GK |
71 | cpumask_set_cpu(cpu, node_to_cpumask_map[nid]); |
72 | } | |
73 | ||
97fd6016 | 74 | void numa_add_cpu(unsigned int cpu) |
1a2db300 | 75 | { |
97fd6016 SH |
76 | numa_update_cpu(cpu, false); |
77 | } | |
1a2db300 | 78 | |
97fd6016 SH |
79 | void numa_remove_cpu(unsigned int cpu) |
80 | { | |
81 | numa_update_cpu(cpu, true); | |
82 | } | |
83 | ||
84 | void numa_clear_node(unsigned int cpu) | |
85 | { | |
86 | numa_remove_cpu(cpu); | |
1a2db300 GK |
87 | set_cpu_numa_node(cpu, NUMA_NO_NODE); |
88 | } | |
89 | ||
90 | /* | |
91 | * Allocate node_to_cpumask_map based on number of available nodes | |
92 | * Requires node_possible_map to be valid. | |
93 | * | |
94 | * Note: cpumask_of_node() is not valid until after this is done. | |
95 | * (Use CONFIG_DEBUG_PER_CPU_MAPS to check this.) | |
96 | */ | |
97 | static void __init setup_node_to_cpumask_map(void) | |
98 | { | |
1a2db300 GK |
99 | int node; |
100 | ||
101 | /* setup nr_node_ids if not done yet */ | |
102 | if (nr_node_ids == MAX_NUMNODES) | |
103 | setup_nr_node_ids(); | |
104 | ||
105 | /* allocate and clear the mapping */ | |
106 | for (node = 0; node < nr_node_ids; node++) { | |
107 | alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]); | |
108 | cpumask_clear(node_to_cpumask_map[node]); | |
109 | } | |
110 | ||
1a2db300 | 111 | /* cpumask_of_node() will now work */ |
b9726c26 | 112 | pr_debug("Node to cpumask map for %u nodes\n", nr_node_ids); |
1a2db300 GK |
113 | } |
114 | ||
115 | /* | |
b1ce45e8 | 116 | * Set the cpu to node and mem mapping |
1a2db300 GK |
117 | */ |
118 | void numa_store_cpu_info(unsigned int cpu) | |
119 | { | |
97fd6016 | 120 | set_cpu_numa_node(cpu, cpu_to_node_map[cpu]); |
1a2db300 GK |
121 | } |
122 | ||
123 | void __init early_map_cpu_to_node(unsigned int cpu, int nid) | |
124 | { | |
125 | /* fallback to node 0 */ | |
7ba5f605 | 126 | if (nid < 0 || nid >= MAX_NUMNODES || numa_off) |
1a2db300 GK |
127 | nid = 0; |
128 | ||
129 | cpu_to_node_map[cpu] = nid; | |
7ba5f605 ZL |
130 | |
131 | /* | |
132 | * We should set the numa node of cpu0 as soon as possible, because it | |
133 | * has already been set up online before. cpu_to_node(0) will soon be | |
134 | * called. | |
135 | */ | |
136 | if (!cpu) | |
137 | set_cpu_numa_node(cpu, nid); | |
1a2db300 GK |
138 | } |
139 | ||
7af3a0a9 ZL |
140 | #ifdef CONFIG_HAVE_SETUP_PER_CPU_AREA |
141 | unsigned long __per_cpu_offset[NR_CPUS] __read_mostly; | |
142 | EXPORT_SYMBOL(__per_cpu_offset); | |
143 | ||
144 | static int __init early_cpu_to_node(int cpu) | |
145 | { | |
146 | return cpu_to_node_map[cpu]; | |
147 | } | |
148 | ||
149 | static int __init pcpu_cpu_distance(unsigned int from, unsigned int to) | |
150 | { | |
26984c3b | 151 | return node_distance(early_cpu_to_node(from), early_cpu_to_node(to)); |
7af3a0a9 ZL |
152 | } |
153 | ||
154 | static void * __init pcpu_fc_alloc(unsigned int cpu, size_t size, | |
155 | size_t align) | |
156 | { | |
157 | int nid = early_cpu_to_node(cpu); | |
158 | ||
eb31d559 | 159 | return memblock_alloc_try_nid(size, align, |
7af3a0a9 ZL |
160 | __pa(MAX_DMA_ADDRESS), MEMBLOCK_ALLOC_ACCESSIBLE, nid); |
161 | } | |
162 | ||
163 | static void __init pcpu_fc_free(void *ptr, size_t size) | |
164 | { | |
165 | memblock_free_early(__pa(ptr), size); | |
166 | } | |
167 | ||
168 | void __init setup_per_cpu_areas(void) | |
169 | { | |
170 | unsigned long delta; | |
171 | unsigned int cpu; | |
172 | int rc; | |
173 | ||
174 | /* | |
175 | * Always reserve area for module percpu variables. That's | |
176 | * what the legacy allocator did. | |
177 | */ | |
178 | rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE, | |
179 | PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, | |
180 | pcpu_cpu_distance, | |
181 | pcpu_fc_alloc, pcpu_fc_free); | |
182 | if (rc < 0) | |
183 | panic("Failed to initialize percpu areas."); | |
184 | ||
185 | delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start; | |
186 | for_each_possible_cpu(cpu) | |
187 | __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu]; | |
188 | } | |
189 | #endif | |
190 | ||
1a2db300 | 191 | /** |
b1ce45e8 | 192 | * numa_add_memblk() - Set node id to memblk |
1a2db300 GK |
193 | * @nid: NUMA node ID of the new memblk |
194 | * @start: Start address of the new memblk | |
8ccbbdaa | 195 | * @end: End address of the new memblk |
1a2db300 GK |
196 | * |
197 | * RETURNS: | |
198 | * 0 on success, -errno on failure. | |
199 | */ | |
8ccbbdaa | 200 | int __init numa_add_memblk(int nid, u64 start, u64 end) |
1a2db300 GK |
201 | { |
202 | int ret; | |
203 | ||
8ccbbdaa | 204 | ret = memblock_set_node(start, (end - start), &memblock.memory, nid); |
1a2db300 | 205 | if (ret < 0) { |
f11c7bac | 206 | pr_err("memblock [0x%llx - 0x%llx] failed to add on node %d\n", |
8ccbbdaa | 207 | start, (end - 1), nid); |
1a2db300 GK |
208 | return ret; |
209 | } | |
210 | ||
211 | node_set(nid, numa_nodes_parsed); | |
1a2db300 GK |
212 | return ret; |
213 | } | |
214 | ||
b1ce45e8 | 215 | /* |
1a2db300 GK |
216 | * Initialize NODE_DATA for a node on the local memory |
217 | */ | |
218 | static void __init setup_node_data(int nid, u64 start_pfn, u64 end_pfn) | |
219 | { | |
220 | const size_t nd_size = roundup(sizeof(pg_data_t), SMP_CACHE_BYTES); | |
221 | u64 nd_pa; | |
222 | void *nd; | |
223 | int tnid; | |
224 | ||
ece4b206 | 225 | if (start_pfn >= end_pfn) |
3f7a09f4 | 226 | pr_info("Initmem setup node %d [<memory-less node>]\n", nid); |
1a2db300 | 227 | |
9a8dd708 | 228 | nd_pa = memblock_phys_alloc_try_nid(nd_size, SMP_CACHE_BYTES, nid); |
33755574 MR |
229 | if (!nd_pa) |
230 | panic("Cannot allocate %zu bytes for node %d data\n", | |
231 | nd_size, nid); | |
232 | ||
1a2db300 GK |
233 | nd = __va(nd_pa); |
234 | ||
235 | /* report and initialize */ | |
f11c7bac | 236 | pr_info("NODE_DATA [mem %#010Lx-%#010Lx]\n", |
1a2db300 GK |
237 | nd_pa, nd_pa + nd_size - 1); |
238 | tnid = early_pfn_to_nid(nd_pa >> PAGE_SHIFT); | |
239 | if (tnid != nid) | |
f11c7bac | 240 | pr_info("NODE_DATA(%d) on node %d\n", nid, tnid); |
1a2db300 GK |
241 | |
242 | node_data[nid] = nd; | |
243 | memset(NODE_DATA(nid), 0, sizeof(pg_data_t)); | |
244 | NODE_DATA(nid)->node_id = nid; | |
245 | NODE_DATA(nid)->node_start_pfn = start_pfn; | |
246 | NODE_DATA(nid)->node_spanned_pages = end_pfn - start_pfn; | |
247 | } | |
248 | ||
b1ce45e8 | 249 | /* |
1a2db300 GK |
250 | * numa_free_distance |
251 | * | |
252 | * The current table is freed. | |
253 | */ | |
254 | void __init numa_free_distance(void) | |
255 | { | |
256 | size_t size; | |
257 | ||
258 | if (!numa_distance) | |
259 | return; | |
260 | ||
261 | size = numa_distance_cnt * numa_distance_cnt * | |
262 | sizeof(numa_distance[0]); | |
263 | ||
264 | memblock_free(__pa(numa_distance), size); | |
265 | numa_distance_cnt = 0; | |
266 | numa_distance = NULL; | |
267 | } | |
268 | ||
b1ce45e8 | 269 | /* |
1a2db300 | 270 | * Create a new NUMA distance table. |
1a2db300 GK |
271 | */ |
272 | static int __init numa_alloc_distance(void) | |
273 | { | |
274 | size_t size; | |
275 | u64 phys; | |
276 | int i, j; | |
277 | ||
278 | size = nr_node_ids * nr_node_ids * sizeof(numa_distance[0]); | |
279 | phys = memblock_find_in_range(0, PFN_PHYS(max_pfn), | |
280 | size, PAGE_SIZE); | |
281 | if (WARN_ON(!phys)) | |
282 | return -ENOMEM; | |
283 | ||
284 | memblock_reserve(phys, size); | |
285 | ||
286 | numa_distance = __va(phys); | |
287 | numa_distance_cnt = nr_node_ids; | |
288 | ||
289 | /* fill with the default distances */ | |
290 | for (i = 0; i < numa_distance_cnt; i++) | |
291 | for (j = 0; j < numa_distance_cnt; j++) | |
292 | numa_distance[i * numa_distance_cnt + j] = i == j ? | |
293 | LOCAL_DISTANCE : REMOTE_DISTANCE; | |
294 | ||
f11c7bac | 295 | pr_debug("Initialized distance table, cnt=%d\n", numa_distance_cnt); |
1a2db300 GK |
296 | |
297 | return 0; | |
298 | } | |
299 | ||
300 | /** | |
b1ce45e8 | 301 | * numa_set_distance() - Set inter node NUMA distance from node to node. |
1a2db300 GK |
302 | * @from: the 'from' node to set distance |
303 | * @to: the 'to' node to set distance | |
304 | * @distance: NUMA distance | |
305 | * | |
306 | * Set the distance from node @from to @to to @distance. | |
307 | * If distance table doesn't exist, a warning is printed. | |
308 | * | |
309 | * If @from or @to is higher than the highest known node or lower than zero | |
310 | * or @distance doesn't make sense, the call is ignored. | |
1a2db300 GK |
311 | */ |
312 | void __init numa_set_distance(int from, int to, int distance) | |
313 | { | |
314 | if (!numa_distance) { | |
f11c7bac | 315 | pr_warn_once("Warning: distance table not allocated yet\n"); |
1a2db300 GK |
316 | return; |
317 | } | |
318 | ||
319 | if (from >= numa_distance_cnt || to >= numa_distance_cnt || | |
320 | from < 0 || to < 0) { | |
f11c7bac | 321 | pr_warn_once("Warning: node ids are out of bound, from=%d to=%d distance=%d\n", |
1a2db300 GK |
322 | from, to, distance); |
323 | return; | |
324 | } | |
325 | ||
326 | if ((u8)distance != distance || | |
327 | (from == to && distance != LOCAL_DISTANCE)) { | |
f11c7bac | 328 | pr_warn_once("Warning: invalid distance parameter, from=%d to=%d distance=%d\n", |
1a2db300 GK |
329 | from, to, distance); |
330 | return; | |
331 | } | |
332 | ||
333 | numa_distance[from * numa_distance_cnt + to] = distance; | |
334 | } | |
335 | ||
b1ce45e8 | 336 | /* |
1a2db300 GK |
337 | * Return NUMA distance @from to @to |
338 | */ | |
339 | int __node_distance(int from, int to) | |
340 | { | |
341 | if (from >= numa_distance_cnt || to >= numa_distance_cnt) | |
342 | return from == to ? LOCAL_DISTANCE : REMOTE_DISTANCE; | |
343 | return numa_distance[from * numa_distance_cnt + to]; | |
344 | } | |
345 | EXPORT_SYMBOL(__node_distance); | |
346 | ||
347 | static int __init numa_register_nodes(void) | |
348 | { | |
349 | int nid; | |
350 | struct memblock_region *mblk; | |
351 | ||
352 | /* Check that valid nid is set to memblks */ | |
d622abf7 MR |
353 | for_each_memblock(memory, mblk) { |
354 | int mblk_nid = memblock_get_region_node(mblk); | |
355 | ||
356 | if (mblk_nid == NUMA_NO_NODE || mblk_nid >= MAX_NUMNODES) { | |
f11c7bac | 357 | pr_warn("Warning: invalid memblk node %d [mem %#010Lx-%#010Lx]\n", |
d622abf7 | 358 | mblk_nid, mblk->base, |
1a2db300 GK |
359 | mblk->base + mblk->size - 1); |
360 | return -EINVAL; | |
361 | } | |
d622abf7 | 362 | } |
1a2db300 GK |
363 | |
364 | /* Finally register nodes. */ | |
365 | for_each_node_mask(nid, numa_nodes_parsed) { | |
366 | unsigned long start_pfn, end_pfn; | |
367 | ||
368 | get_pfn_range_for_nid(nid, &start_pfn, &end_pfn); | |
369 | setup_node_data(nid, start_pfn, end_pfn); | |
370 | node_set_online(nid); | |
371 | } | |
372 | ||
373 | /* Setup online nodes to actual nodes*/ | |
374 | node_possible_map = numa_nodes_parsed; | |
375 | ||
376 | return 0; | |
377 | } | |
378 | ||
379 | static int __init numa_init(int (*init_func)(void)) | |
380 | { | |
381 | int ret; | |
382 | ||
383 | nodes_clear(numa_nodes_parsed); | |
384 | nodes_clear(node_possible_map); | |
385 | nodes_clear(node_online_map); | |
1a2db300 GK |
386 | |
387 | ret = numa_alloc_distance(); | |
388 | if (ret < 0) | |
389 | return ret; | |
390 | ||
391 | ret = init_func(); | |
392 | if (ret < 0) | |
52338088 | 393 | goto out_free_distance; |
1a2db300 | 394 | |
794224ea ZL |
395 | if (nodes_empty(numa_nodes_parsed)) { |
396 | pr_info("No NUMA configuration found\n"); | |
52338088 AK |
397 | ret = -EINVAL; |
398 | goto out_free_distance; | |
794224ea | 399 | } |
1a2db300 GK |
400 | |
401 | ret = numa_register_nodes(); | |
402 | if (ret < 0) | |
52338088 | 403 | goto out_free_distance; |
1a2db300 GK |
404 | |
405 | setup_node_to_cpumask_map(); | |
406 | ||
1a2db300 | 407 | return 0; |
52338088 AK |
408 | out_free_distance: |
409 | numa_free_distance(); | |
410 | return ret; | |
1a2db300 GK |
411 | } |
412 | ||
413 | /** | |
b1ce45e8 | 414 | * dummy_numa_init() - Fallback dummy NUMA init |
1a2db300 GK |
415 | * |
416 | * Used if there's no underlying NUMA architecture, NUMA initialization | |
417 | * fails, or NUMA is disabled on the command line. | |
418 | * | |
419 | * Must online at least one node (node 0) and add memory blocks that cover all | |
420 | * allowed memory. It is unlikely that this function fails. | |
b1ce45e8 QC |
421 | * |
422 | * Return: 0 on success, -errno on failure. | |
1a2db300 GK |
423 | */ |
424 | static int __init dummy_numa_init(void) | |
425 | { | |
426 | int ret; | |
427 | struct memblock_region *mblk; | |
428 | ||
34c33370 DD |
429 | if (numa_off) |
430 | pr_info("NUMA disabled\n"); /* Forced off on command line. */ | |
f11c7bac | 431 | pr_info("Faking a node at [mem %#018Lx-%#018Lx]\n", |
77cfe950 | 432 | memblock_start_of_DRAM(), memblock_end_of_DRAM() - 1); |
1a2db300 GK |
433 | |
434 | for_each_memblock(memory, mblk) { | |
8ccbbdaa | 435 | ret = numa_add_memblk(0, mblk->base, mblk->base + mblk->size); |
1a2db300 GK |
436 | if (!ret) |
437 | continue; | |
438 | ||
439 | pr_err("NUMA init failed\n"); | |
440 | return ret; | |
441 | } | |
442 | ||
34c33370 | 443 | numa_off = true; |
1a2db300 GK |
444 | return 0; |
445 | } | |
446 | ||
447 | /** | |
b1ce45e8 | 448 | * arm64_numa_init() - Initialize NUMA |
1a2db300 | 449 | * |
b1ce45e8 | 450 | * Try each configured NUMA initialization method until one succeeds. The |
1a2db300 GK |
451 | * last fallback is dummy single node config encomapssing whole memory. |
452 | */ | |
453 | void __init arm64_numa_init(void) | |
454 | { | |
455 | if (!numa_off) { | |
d8b47fca HG |
456 | if (!acpi_disabled && !numa_init(arm64_acpi_numa_init)) |
457 | return; | |
458 | if (acpi_disabled && !numa_init(of_numa_init)) | |
1a2db300 GK |
459 | return; |
460 | } | |
461 | ||
462 | numa_init(dummy_numa_init); | |
463 | } |