drivers/base/arch_numa.c

0001 // SPDX-License-Identifier: GPL-2.0-only
0002 /*
0003  * NUMA support, based on the x86 implementation.
0004  *
0005  * Copyright (C) 2015 Cavium Inc.
0006  * Author: Ganapatrao Kulkarni <gkulkarni@cavium.com>
0007  */
0008
0009 #define pr_fmt(fmt) "NUMA: " fmt
0010
0011 #include <linux/acpi.h>
0012 #include <linux/memblock.h>
0013 #include <linux/module.h>
0014 #include <linux/of.h>
0015
0016 #include <asm/sections.h>
0017
0018 struct pglist_data *node_data[MAX_NUMNODES] __read_mostly;
0019 EXPORT_SYMBOL(node_data);
0020 nodemask_t numa_nodes_parsed __initdata;
0021 static int cpu_to_node_map[NR_CPUS] = { [0 ... NR_CPUS-1] = NUMA_NO_NODE };
0022
0023 static int numa_distance_cnt;
0024 static u8 *numa_distance;
0025 bool numa_off;
0026
0027 static __init int numa_parse_early_param(char *opt)
0028 {
0029     if (!opt)
0030         return -EINVAL;
0031     if (str_has_prefix(opt, "off"))
0032         numa_off = true;
0033
0034     return 0;
0035 }
0036 early_param("numa", numa_parse_early_param);
0037
0038 cpumask_var_t node_to_cpumask_map[MAX_NUMNODES];
0039 EXPORT_SYMBOL(node_to_cpumask_map);
0040
0041 #ifdef CONFIG_DEBUG_PER_CPU_MAPS
0042
0043 /*
0044  * Returns a pointer to the bitmask of CPUs on Node 'node'.
0045  */
0046 const struct cpumask *cpumask_of_node(int node)
0047 {
0048
0049     if (node == NUMA_NO_NODE)
0050         return cpu_all_mask;
0051
0052     if (WARN_ON(node < 0 || node >= nr_node_ids))
0053         return cpu_none_mask;
0054
0055     if (WARN_ON(node_to_cpumask_map[node] == NULL))
0056         return cpu_online_mask;
0057
0058     return node_to_cpumask_map[node];
0059 }
0060 EXPORT_SYMBOL(cpumask_of_node);
0061
0062 #endif
0063
0064 static void numa_update_cpu(unsigned int cpu, bool remove)
0065 {
0066     int nid = cpu_to_node(cpu);
0067
0068     if (nid == NUMA_NO_NODE)
0069         return;
0070
0071     if (remove)
0072         cpumask_clear_cpu(cpu, node_to_cpumask_map[nid]);
0073     else
0074         cpumask_set_cpu(cpu, node_to_cpumask_map[nid]);
0075 }
0076
0077 void numa_add_cpu(unsigned int cpu)
0078 {
0079     numa_update_cpu(cpu, false);
0080 }
0081
0082 void numa_remove_cpu(unsigned int cpu)
0083 {
0084     numa_update_cpu(cpu, true);
0085 }
0086
0087 void numa_clear_node(unsigned int cpu)
0088 {
0089     numa_remove_cpu(cpu);
0090     set_cpu_numa_node(cpu, NUMA_NO_NODE);
0091 }
0092
0093 /*
0094  * Allocate node_to_cpumask_map based on number of available nodes
0095  * Requires node_possible_map to be valid.
0096  *
0097  * Note: cpumask_of_node() is not valid until after this is done.
0098  * (Use CONFIG_DEBUG_PER_CPU_MAPS to check this.)
0099  */
0100 static void __init setup_node_to_cpumask_map(void)
0101 {
0102     int node;
0103
0104     /* setup nr_node_ids if not done yet */
0105     if (nr_node_ids == MAX_NUMNODES)
0106         setup_nr_node_ids();
0107
0108     /* allocate and clear the mapping */
0109     for (node = 0; node < nr_node_ids; node++) {
0110         alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]);
0111         cpumask_clear(node_to_cpumask_map[node]);
0112     }
0113
0114     /* cpumask_of_node() will now work */
0115     pr_debug("Node to cpumask map for %u nodes\n", nr_node_ids);
0116 }
0117
0118 /*
0119  * Set the cpu to node and mem mapping
0120  */
0121 void numa_store_cpu_info(unsigned int cpu)
0122 {
0123     set_cpu_numa_node(cpu, cpu_to_node_map[cpu]);
0124 }
0125
0126 void __init early_map_cpu_to_node(unsigned int cpu, int nid)
0127 {
0128     /* fallback to node 0 */
0129     if (nid < 0 || nid >= MAX_NUMNODES || numa_off)
0130         nid = 0;
0131
0132     cpu_to_node_map[cpu] = nid;
0133
0134     /*
0135      * We should set the numa node of cpu0 as soon as possible, because it
0136      * has already been set up online before. cpu_to_node(0) will soon be
0137      * called.
0138      */
0139     if (!cpu)
0140         set_cpu_numa_node(cpu, nid);
0141 }
0142
0143 #ifdef CONFIG_HAVE_SETUP_PER_CPU_AREA
0144 unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
0145 EXPORT_SYMBOL(__per_cpu_offset);
0146
0147 static int __init early_cpu_to_node(int cpu)
0148 {
0149     return cpu_to_node_map[cpu];
0150 }
0151
0152 static int __init pcpu_cpu_distance(unsigned int from, unsigned int to)
0153 {
0154     return node_distance(early_cpu_to_node(from), early_cpu_to_node(to));
0155 }
0156
0157 void __init setup_per_cpu_areas(void)
0158 {
0159     unsigned long delta;
0160     unsigned int cpu;
0161     int rc = -EINVAL;
0162
0163     if (pcpu_chosen_fc != PCPU_FC_PAGE) {
0164         /*
0165          * Always reserve area for module percpu variables.  That's
0166          * what the legacy allocator did.
0167          */
0168         rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
0169                         PERCPU_DYNAMIC_RESERVE, PAGE_SIZE,
0170                         pcpu_cpu_distance,
0171                         early_cpu_to_node);
0172 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
0173         if (rc < 0)
0174             pr_warn("PERCPU: %s allocator failed (%d), falling back to page size\n",
0175                    pcpu_fc_names[pcpu_chosen_fc], rc);
0176 #endif
0177     }
0178
0179 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
0180     if (rc < 0)
0181         rc = pcpu_page_first_chunk(PERCPU_MODULE_RESERVE, early_cpu_to_node);
0182 #endif
0183     if (rc < 0)
0184         panic("Failed to initialize percpu areas (err=%d).", rc);
0185
0186     delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
0187     for_each_possible_cpu(cpu)
0188         __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
0189 }
0190 #endif
0191
0192 /**
0193  * numa_add_memblk() - Set node id to memblk
0194  * @nid: NUMA node ID of the new memblk
0195  * @start: Start address of the new memblk
0196  * @end:  End address of the new memblk
0197  *
0198  * RETURNS:
0199  * 0 on success, -errno on failure.
0200  */
0201 int __init numa_add_memblk(int nid, u64 start, u64 end)
0202 {
0203     int ret;
0204
0205     ret = memblock_set_node(start, (end - start), &memblock.memory, nid);
0206     if (ret < 0) {
0207         pr_err("memblock [0x%llx - 0x%llx] failed to add on node %d\n",
0208             start, (end - 1), nid);
0209         return ret;
0210     }
0211
0212     node_set(nid, numa_nodes_parsed);
0213     return ret;
0214 }
0215
0216 /*
0217  * Initialize NODE_DATA for a node on the local memory
0218  */
0219 static void __init setup_node_data(int nid, u64 start_pfn, u64 end_pfn)
0220 {
0221     const size_t nd_size = roundup(sizeof(pg_data_t), SMP_CACHE_BYTES);
0222     u64 nd_pa;
0223     void *nd;
0224     int tnid;
0225
0226     if (start_pfn >= end_pfn)
0227         pr_info("Initmem setup node %d [<memory-less node>]\n", nid);
0228
0229     nd_pa = memblock_phys_alloc_try_nid(nd_size, SMP_CACHE_BYTES, nid);
0230     if (!nd_pa)
0231         panic("Cannot allocate %zu bytes for node %d data\n",
0232               nd_size, nid);
0233
0234     nd = __va(nd_pa);
0235
0236     /* report and initialize */
0237     pr_info("NODE_DATA [mem %#010Lx-%#010Lx]\n",
0238         nd_pa, nd_pa + nd_size - 1);
0239     tnid = early_pfn_to_nid(nd_pa >> PAGE_SHIFT);
0240     if (tnid != nid)
0241         pr_info("NODE_DATA(%d) on node %d\n", nid, tnid);
0242
0243     node_data[nid] = nd;
0244     memset(NODE_DATA(nid), 0, sizeof(pg_data_t));
0245     NODE_DATA(nid)->node_id = nid;
0246     NODE_DATA(nid)->node_start_pfn = start_pfn;
0247     NODE_DATA(nid)->node_spanned_pages = end_pfn - start_pfn;
0248 }
0249
0250 /*
0251  * numa_free_distance
0252  *
0253  * The current table is freed.
0254  */
0255 void __init numa_free_distance(void)
0256 {
0257     size_t size;
0258
0259     if (!numa_distance)
0260         return;
0261
0262     size = numa_distance_cnt * numa_distance_cnt *
0263         sizeof(numa_distance[0]);
0264
0265     memblock_free(numa_distance, size);
0266     numa_distance_cnt = 0;
0267     numa_distance = NULL;
0268 }
0269
0270 /*
0271  * Create a new NUMA distance table.
0272  */
0273 static int __init numa_alloc_distance(void)
0274 {
0275     size_t size;
0276     int i, j;
0277
0278     size = nr_node_ids * nr_node_ids * sizeof(numa_distance[0]);
0279     numa_distance = memblock_alloc(size, PAGE_SIZE);
0280     if (WARN_ON(!numa_distance))
0281         return -ENOMEM;
0282
0283     numa_distance_cnt = nr_node_ids;
0284
0285     /* fill with the default distances */
0286     for (i = 0; i < numa_distance_cnt; i++)
0287         for (j = 0; j < numa_distance_cnt; j++)
0288             numa_distance[i * numa_distance_cnt + j] = i == j ?
0289                 LOCAL_DISTANCE : REMOTE_DISTANCE;
0290
0291     pr_debug("Initialized distance table, cnt=%d\n", numa_distance_cnt);
0292
0293     return 0;
0294 }
0295
0296 /**
0297  * numa_set_distance() - Set inter node NUMA distance from node to node.
0298  * @from: the 'from' node to set distance
0299  * @to: the 'to'  node to set distance
0300  * @distance: NUMA distance
0301  *
0302  * Set the distance from node @from to @to to @distance.
0303  * If distance table doesn't exist, a warning is printed.
0304  *
0305  * If @from or @to is higher than the highest known node or lower than zero
0306  * or @distance doesn't make sense, the call is ignored.
0307  */
0308 void __init numa_set_distance(int from, int to, int distance)
0309 {
0310     if (!numa_distance) {
0311         pr_warn_once("Warning: distance table not allocated yet\n");
0312         return;
0313     }
0314
0315     if (from >= numa_distance_cnt || to >= numa_distance_cnt ||
0316             from < 0 || to < 0) {
0317         pr_warn_once("Warning: node ids are out of bound, from=%d to=%d distance=%d\n",
0318                 from, to, distance);
0319         return;
0320     }
0321
0322     if ((u8)distance != distance ||
0323         (from == to && distance != LOCAL_DISTANCE)) {
0324         pr_warn_once("Warning: invalid distance parameter, from=%d to=%d distance=%d\n",
0325                  from, to, distance);
0326         return;
0327     }
0328
0329     numa_distance[from * numa_distance_cnt + to] = distance;
0330 }
0331
0332 /*
0333  * Return NUMA distance @from to @to
0334  */
0335 int __node_distance(int from, int to)
0336 {
0337     if (from >= numa_distance_cnt || to >= numa_distance_cnt)
0338         return from == to ? LOCAL_DISTANCE : REMOTE_DISTANCE;
0339     return numa_distance[from * numa_distance_cnt + to];
0340 }
0341 EXPORT_SYMBOL(__node_distance);
0342
0343 static int __init numa_register_nodes(void)
0344 {
0345     int nid;
0346     struct memblock_region *mblk;
0347
0348     /* Check that valid nid is set to memblks */
0349     for_each_mem_region(mblk) {
0350         int mblk_nid = memblock_get_region_node(mblk);
0351         phys_addr_t start = mblk->base;
0352         phys_addr_t end = mblk->base + mblk->size - 1;
0353
0354         if (mblk_nid == NUMA_NO_NODE || mblk_nid >= MAX_NUMNODES) {
0355             pr_warn("Warning: invalid memblk node %d [mem %pap-%pap]\n",
0356                 mblk_nid, &start, &end);
0357             return -EINVAL;
0358         }
0359     }
0360
0361     /* Finally register nodes. */
0362     for_each_node_mask(nid, numa_nodes_parsed) {
0363         unsigned long start_pfn, end_pfn;
0364
0365         get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
0366         setup_node_data(nid, start_pfn, end_pfn);
0367         node_set_online(nid);
0368     }
0369
0370     /* Setup online nodes to actual nodes*/
0371     node_possible_map = numa_nodes_parsed;
0372
0373     return 0;
0374 }
0375
0376 static int __init numa_init(int (*init_func)(void))
0377 {
0378     int ret;
0379
0380     nodes_clear(numa_nodes_parsed);
0381     nodes_clear(node_possible_map);
0382     nodes_clear(node_online_map);
0383
0384     ret = numa_alloc_distance();
0385     if (ret < 0)
0386         return ret;
0387
0388     ret = init_func();
0389     if (ret < 0)
0390         goto out_free_distance;
0391
0392     if (nodes_empty(numa_nodes_parsed)) {
0393         pr_info("No NUMA configuration found\n");
0394         ret = -EINVAL;
0395         goto out_free_distance;
0396     }
0397
0398     ret = numa_register_nodes();
0399     if (ret < 0)
0400         goto out_free_distance;
0401
0402     setup_node_to_cpumask_map();
0403
0404     return 0;
0405 out_free_distance:
0406     numa_free_distance();
0407     return ret;
0408 }
0409
0410 /**
0411  * dummy_numa_init() - Fallback dummy NUMA init
0412  *
0413  * Used if there's no underlying NUMA architecture, NUMA initialization
0414  * fails, or NUMA is disabled on the command line.
0415  *
0416  * Must online at least one node (node 0) and add memory blocks that cover all
0417  * allowed memory. It is unlikely that this function fails.
0418  *
0419  * Return: 0 on success, -errno on failure.
0420  */
0421 static int __init dummy_numa_init(void)
0422 {
0423     phys_addr_t start = memblock_start_of_DRAM();
0424     phys_addr_t end = memblock_end_of_DRAM() - 1;
0425     int ret;
0426
0427     if (numa_off)
0428         pr_info("NUMA disabled\n"); /* Forced off on command line. */
0429     pr_info("Faking a node at [mem %pap-%pap]\n", &start, &end);
0430
0431     ret = numa_add_memblk(0, start, end + 1);
0432     if (ret) {
0433         pr_err("NUMA init failed\n");
0434         return ret;
0435     }
0436
0437     numa_off = true;
0438     return 0;
0439 }
0440
0441 #ifdef CONFIG_ACPI_NUMA
0442 static int __init arch_acpi_numa_init(void)
0443 {
0444     int ret;
0445
0446     ret = acpi_numa_init();
0447     if (ret) {
0448         pr_info("Failed to initialise from firmware\n");
0449         return ret;
0450     }
0451
0452     return srat_disabled() ? -EINVAL : 0;
0453 }
0454 #else
0455 static int __init arch_acpi_numa_init(void)
0456 {
0457     return -EOPNOTSUPP;
0458 }
0459 #endif
0460
0461 /**
0462  * arch_numa_init() - Initialize NUMA
0463  *
0464  * Try each configured NUMA initialization method until one succeeds. The
0465  * last fallback is dummy single node config encompassing whole memory.
0466  */
0467 void __init arch_numa_init(void)
0468 {
0469     if (!numa_off) {
0470         if (!acpi_disabled && !numa_init(arch_acpi_numa_init))
0471             return;
0472         if (acpi_disabled && !numa_init(of_numa_init))
0473             return;
0474     }
0475
0476     numa_init(dummy_numa_init);
0477 }