1 // SPDX-License-Identifier: GPL-2.0-only
3 * NUMA support, based on the x86 implementation.
5 * Copyright (C) 2015 Cavium Inc.
6 * Author: Ganapatrao Kulkarni <gkulkarni@cavium.com>
9 #define pr_fmt(fmt) "NUMA: " fmt
11 #include <linux/acpi.h>
12 #include <linux/memblock.h>
13 #include <linux/module.h>
16 #include <asm/sections.h>
17 #include <asm/pgalloc.h>
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 };
24 static int numa_distance_cnt;
25 static u8 *numa_distance;
28 static __init int numa_parse_early_param(char *opt)
32 if (str_has_prefix(opt, "off"))
37 early_param("numa", numa_parse_early_param);
39 cpumask_var_t node_to_cpumask_map[MAX_NUMNODES];
40 EXPORT_SYMBOL(node_to_cpumask_map);
42 #ifdef CONFIG_DEBUG_PER_CPU_MAPS
45 * Returns a pointer to the bitmask of CPUs on Node 'node'.
47 const struct cpumask *cpumask_of_node(int node)
50 if (node == NUMA_NO_NODE)
53 if (WARN_ON(node < 0 || node >= nr_node_ids))
56 if (WARN_ON(node_to_cpumask_map[node] == NULL))
57 return cpu_online_mask;
59 return node_to_cpumask_map[node];
61 EXPORT_SYMBOL(cpumask_of_node);
65 static void numa_update_cpu(unsigned int cpu, bool remove)
67 int nid = cpu_to_node(cpu);
69 if (nid == NUMA_NO_NODE)
73 cpumask_clear_cpu(cpu, node_to_cpumask_map[nid]);
75 cpumask_set_cpu(cpu, node_to_cpumask_map[nid]);
78 void numa_add_cpu(unsigned int cpu)
80 numa_update_cpu(cpu, false);
83 void numa_remove_cpu(unsigned int cpu)
85 numa_update_cpu(cpu, true);
88 void numa_clear_node(unsigned int cpu)
91 set_cpu_numa_node(cpu, NUMA_NO_NODE);
95 * Allocate node_to_cpumask_map based on number of available nodes
96 * Requires node_possible_map to be valid.
98 * Note: cpumask_of_node() is not valid until after this is done.
99 * (Use CONFIG_DEBUG_PER_CPU_MAPS to check this.)
101 static void __init setup_node_to_cpumask_map(void)
105 /* setup nr_node_ids if not done yet */
106 if (nr_node_ids == MAX_NUMNODES)
109 /* allocate and clear the mapping */
110 for (node = 0; node < nr_node_ids; node++) {
111 alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]);
112 cpumask_clear(node_to_cpumask_map[node]);
115 /* cpumask_of_node() will now work */
116 pr_debug("Node to cpumask map for %u nodes\n", nr_node_ids);
120 * Set the cpu to node and mem mapping
122 void numa_store_cpu_info(unsigned int cpu)
124 set_cpu_numa_node(cpu, cpu_to_node_map[cpu]);
127 void __init early_map_cpu_to_node(unsigned int cpu, int nid)
129 /* fallback to node 0 */
130 if (nid < 0 || nid >= MAX_NUMNODES || numa_off)
133 cpu_to_node_map[cpu] = nid;
136 * We should set the numa node of cpu0 as soon as possible, because it
137 * has already been set up online before. cpu_to_node(0) will soon be
141 set_cpu_numa_node(cpu, nid);
144 #ifdef CONFIG_HAVE_SETUP_PER_CPU_AREA
145 unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
146 EXPORT_SYMBOL(__per_cpu_offset);
148 static int __init early_cpu_to_node(int cpu)
150 return cpu_to_node_map[cpu];
153 static int __init pcpu_cpu_distance(unsigned int from, unsigned int to)
155 return node_distance(early_cpu_to_node(from), early_cpu_to_node(to));
158 static void * __init pcpu_fc_alloc(unsigned int cpu, size_t size,
161 int nid = early_cpu_to_node(cpu);
163 return memblock_alloc_try_nid(size, align,
164 __pa(MAX_DMA_ADDRESS), MEMBLOCK_ALLOC_ACCESSIBLE, nid);
167 static void __init pcpu_fc_free(void *ptr, size_t size)
169 memblock_free(ptr, size);
172 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
173 static void __init pcpu_populate_pte(unsigned long addr)
175 pgd_t *pgd = pgd_offset_k(addr);
180 p4d = p4d_offset(pgd, addr);
181 if (p4d_none(*p4d)) {
184 new = memblock_alloc(PAGE_SIZE, PAGE_SIZE);
187 p4d_populate(&init_mm, p4d, new);
190 pud = pud_offset(p4d, addr);
191 if (pud_none(*pud)) {
194 new = memblock_alloc(PAGE_SIZE, PAGE_SIZE);
197 pud_populate(&init_mm, pud, new);
200 pmd = pmd_offset(pud, addr);
201 if (!pmd_present(*pmd)) {
204 new = memblock_alloc(PAGE_SIZE, PAGE_SIZE);
207 pmd_populate_kernel(&init_mm, pmd, new);
213 panic("%s: Failed to allocate %lu bytes align=%lx from=%lx\n",
214 __func__, PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
218 void __init setup_per_cpu_areas(void)
224 if (pcpu_chosen_fc != PCPU_FC_PAGE) {
226 * Always reserve area for module percpu variables. That's
227 * what the legacy allocator did.
229 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
230 PERCPU_DYNAMIC_RESERVE, PAGE_SIZE,
232 pcpu_fc_alloc, pcpu_fc_free);
233 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
235 pr_warn("PERCPU: %s allocator failed (%d), falling back to page size\n",
236 pcpu_fc_names[pcpu_chosen_fc], rc);
240 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
242 rc = pcpu_page_first_chunk(PERCPU_MODULE_RESERVE,
248 panic("Failed to initialize percpu areas (err=%d).", rc);
250 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
251 for_each_possible_cpu(cpu)
252 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
257 * numa_add_memblk() - Set node id to memblk
258 * @nid: NUMA node ID of the new memblk
259 * @start: Start address of the new memblk
260 * @end: End address of the new memblk
263 * 0 on success, -errno on failure.
265 int __init numa_add_memblk(int nid, u64 start, u64 end)
269 ret = memblock_set_node(start, (end - start), &memblock.memory, nid);
271 pr_err("memblock [0x%llx - 0x%llx] failed to add on node %d\n",
272 start, (end - 1), nid);
276 node_set(nid, numa_nodes_parsed);
281 * Initialize NODE_DATA for a node on the local memory
283 static void __init setup_node_data(int nid, u64 start_pfn, u64 end_pfn)
285 const size_t nd_size = roundup(sizeof(pg_data_t), SMP_CACHE_BYTES);
290 if (start_pfn >= end_pfn)
291 pr_info("Initmem setup node %d [<memory-less node>]\n", nid);
293 nd_pa = memblock_phys_alloc_try_nid(nd_size, SMP_CACHE_BYTES, nid);
295 panic("Cannot allocate %zu bytes for node %d data\n",
300 /* report and initialize */
301 pr_info("NODE_DATA [mem %#010Lx-%#010Lx]\n",
302 nd_pa, nd_pa + nd_size - 1);
303 tnid = early_pfn_to_nid(nd_pa >> PAGE_SHIFT);
305 pr_info("NODE_DATA(%d) on node %d\n", nid, tnid);
308 memset(NODE_DATA(nid), 0, sizeof(pg_data_t));
309 NODE_DATA(nid)->node_id = nid;
310 NODE_DATA(nid)->node_start_pfn = start_pfn;
311 NODE_DATA(nid)->node_spanned_pages = end_pfn - start_pfn;
317 * The current table is freed.
319 void __init numa_free_distance(void)
326 size = numa_distance_cnt * numa_distance_cnt *
327 sizeof(numa_distance[0]);
329 memblock_free(numa_distance, size);
330 numa_distance_cnt = 0;
331 numa_distance = NULL;
335 * Create a new NUMA distance table.
337 static int __init numa_alloc_distance(void)
342 size = nr_node_ids * nr_node_ids * sizeof(numa_distance[0]);
343 numa_distance = memblock_alloc(size, PAGE_SIZE);
344 if (WARN_ON(!numa_distance))
347 numa_distance_cnt = nr_node_ids;
349 /* fill with the default distances */
350 for (i = 0; i < numa_distance_cnt; i++)
351 for (j = 0; j < numa_distance_cnt; j++)
352 numa_distance[i * numa_distance_cnt + j] = i == j ?
353 LOCAL_DISTANCE : REMOTE_DISTANCE;
355 pr_debug("Initialized distance table, cnt=%d\n", numa_distance_cnt);
361 * numa_set_distance() - Set inter node NUMA distance from node to node.
362 * @from: the 'from' node to set distance
363 * @to: the 'to' node to set distance
364 * @distance: NUMA distance
366 * Set the distance from node @from to @to to @distance.
367 * If distance table doesn't exist, a warning is printed.
369 * If @from or @to is higher than the highest known node or lower than zero
370 * or @distance doesn't make sense, the call is ignored.
372 void __init numa_set_distance(int from, int to, int distance)
374 if (!numa_distance) {
375 pr_warn_once("Warning: distance table not allocated yet\n");
379 if (from >= numa_distance_cnt || to >= numa_distance_cnt ||
380 from < 0 || to < 0) {
381 pr_warn_once("Warning: node ids are out of bound, from=%d to=%d distance=%d\n",
386 if ((u8)distance != distance ||
387 (from == to && distance != LOCAL_DISTANCE)) {
388 pr_warn_once("Warning: invalid distance parameter, from=%d to=%d distance=%d\n",
393 numa_distance[from * numa_distance_cnt + to] = distance;
397 * Return NUMA distance @from to @to
399 int __node_distance(int from, int to)
401 if (from >= numa_distance_cnt || to >= numa_distance_cnt)
402 return from == to ? LOCAL_DISTANCE : REMOTE_DISTANCE;
403 return numa_distance[from * numa_distance_cnt + to];
405 EXPORT_SYMBOL(__node_distance);
407 static int __init numa_register_nodes(void)
410 struct memblock_region *mblk;
412 /* Check that valid nid is set to memblks */
413 for_each_mem_region(mblk) {
414 int mblk_nid = memblock_get_region_node(mblk);
415 phys_addr_t start = mblk->base;
416 phys_addr_t end = mblk->base + mblk->size - 1;
418 if (mblk_nid == NUMA_NO_NODE || mblk_nid >= MAX_NUMNODES) {
419 pr_warn("Warning: invalid memblk node %d [mem %pap-%pap]\n",
420 mblk_nid, &start, &end);
425 /* Finally register nodes. */
426 for_each_node_mask(nid, numa_nodes_parsed) {
427 unsigned long start_pfn, end_pfn;
429 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
430 setup_node_data(nid, start_pfn, end_pfn);
431 node_set_online(nid);
434 /* Setup online nodes to actual nodes*/
435 node_possible_map = numa_nodes_parsed;
440 static int __init numa_init(int (*init_func)(void))
444 nodes_clear(numa_nodes_parsed);
445 nodes_clear(node_possible_map);
446 nodes_clear(node_online_map);
448 ret = numa_alloc_distance();
454 goto out_free_distance;
456 if (nodes_empty(numa_nodes_parsed)) {
457 pr_info("No NUMA configuration found\n");
459 goto out_free_distance;
462 ret = numa_register_nodes();
464 goto out_free_distance;
466 setup_node_to_cpumask_map();
470 numa_free_distance();
475 * dummy_numa_init() - Fallback dummy NUMA init
477 * Used if there's no underlying NUMA architecture, NUMA initialization
478 * fails, or NUMA is disabled on the command line.
480 * Must online at least one node (node 0) and add memory blocks that cover all
481 * allowed memory. It is unlikely that this function fails.
483 * Return: 0 on success, -errno on failure.
485 static int __init dummy_numa_init(void)
487 phys_addr_t start = memblock_start_of_DRAM();
488 phys_addr_t end = memblock_end_of_DRAM() - 1;
492 pr_info("NUMA disabled\n"); /* Forced off on command line. */
493 pr_info("Faking a node at [mem %pap-%pap]\n", &start, &end);
495 ret = numa_add_memblk(0, start, end + 1);
497 pr_err("NUMA init failed\n");
505 #ifdef CONFIG_ACPI_NUMA
506 static int __init arch_acpi_numa_init(void)
510 ret = acpi_numa_init();
512 pr_info("Failed to initialise from firmware\n");
516 return srat_disabled() ? -EINVAL : 0;
519 static int __init arch_acpi_numa_init(void)
526 * arch_numa_init() - Initialize NUMA
528 * Try each configured NUMA initialization method until one succeeds. The
529 * last fallback is dummy single node config encompassing whole memory.
531 void __init arch_numa_init(void)
534 if (!acpi_disabled && !numa_init(arch_acpi_numa_init))
536 if (acpi_disabled && !numa_init(of_numa_init))
540 numa_init(dummy_numa_init);