arch/arm64/mm/numa.c

   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/acpi.h>
  17 #include <asm/sections.h>
  18
  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;
  26 bool numa_off;
  27
  28 static __init int numa_parse_early_param(char *opt)
  29 {
  30         if (!opt)
  31                 return -EINVAL;
  32         if (str_has_prefix(opt, "off"))
  33                 numa_off = true;
  34
  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
  50         if (node == NUMA_NO_NODE)
  51                 return cpu_all_mask;
  52
  53         if (WARN_ON(node < 0 || node >= nr_node_ids))
  54                 return cpu_none_mask;
  55
  56         if (WARN_ON(node_to_cpumask_map[node] == NULL))
  57                 return cpu_online_mask;
  58
  59         return node_to_cpumask_map[node];
  60 }
  61 EXPORT_SYMBOL(cpumask_of_node);
  62
  63 #endif
  64
  65 static void numa_update_cpu(unsigned int cpu, bool remove)
  66 {
  67         int nid = cpu_to_node(cpu);
  68
  69         if (nid == NUMA_NO_NODE)
  70                 return;
  71
  72         if (remove)
  73                 cpumask_clear_cpu(cpu, node_to_cpumask_map[nid]);
  74         else
  75                 cpumask_set_cpu(cpu, node_to_cpumask_map[nid]);
  76 }
  77
  78 void numa_add_cpu(unsigned int cpu)
  79 {
  80         numa_update_cpu(cpu, false);
  81 }
  82
  83 void numa_remove_cpu(unsigned int cpu)
  84 {
  85         numa_update_cpu(cpu, true);
  86 }
  87
  88 void numa_clear_node(unsigned int cpu)
  89 {
  90         numa_remove_cpu(cpu);
  91         set_cpu_numa_node(cpu, NUMA_NO_NODE);
  92 }
  93
  94 /*
  95  * Allocate node_to_cpumask_map based on number of available nodes
  96  * Requires node_possible_map to be valid.
  97  *
  98  * Note: cpumask_of_node() is not valid until after this is done.
  99  * (Use CONFIG_DEBUG_PER_CPU_MAPS to check this.)
 100  */
 101 static void __init setup_node_to_cpumask_map(void)
 102 {
 103         int node;
 104
 105         /* setup nr_node_ids if not done yet */
 106         if (nr_node_ids == MAX_NUMNODES)
 107                 setup_nr_node_ids();
 108
 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]);
 113         }
 114
 115         /* cpumask_of_node() will now work */
 116         pr_debug("Node to cpumask map for %u nodes\n", nr_node_ids);
 117 }
 118
 119 /*
 120  * Set the cpu to node and mem mapping
 121  */
 122 void numa_store_cpu_info(unsigned int cpu)
 123 {
 124         set_cpu_numa_node(cpu, cpu_to_node_map[cpu]);
 125 }
 126
 127 void __init early_map_cpu_to_node(unsigned int cpu, int nid)
 128 {
 129         /* fallback to node 0 */
 130         if (nid < 0 || nid >= MAX_NUMNODES || numa_off)
 131                 nid = 0;
 132
 133         cpu_to_node_map[cpu] = nid;
 134
 135         /*
 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
 138          * called.
 139          */
 140         if (!cpu)
 141                 set_cpu_numa_node(cpu, nid);
 142 }
 143
 144 #ifdef CONFIG_HAVE_SETUP_PER_CPU_AREA
 145 unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
 146 EXPORT_SYMBOL(__per_cpu_offset);
 147
 148 static int __init early_cpu_to_node(int cpu)
 149 {
 150         return cpu_to_node_map[cpu];
 151 }
 152
 153 static int __init pcpu_cpu_distance(unsigned int from, unsigned int to)
 154 {
 155         return node_distance(early_cpu_to_node(from), early_cpu_to_node(to));
 156 }
 157
 158 static void * __init pcpu_fc_alloc(unsigned int cpu, size_t size,
 159                                        size_t align)
 160 {
 161         int nid = early_cpu_to_node(cpu);
 162
 163         return  memblock_alloc_try_nid(size, align,
 164                         __pa(MAX_DMA_ADDRESS), MEMBLOCK_ALLOC_ACCESSIBLE, nid);
 165 }
 166
 167 static void __init pcpu_fc_free(void *ptr, size_t size)
 168 {
 169         memblock_free_early(__pa(ptr), size);
 170 }
 171
 172 void __init setup_per_cpu_areas(void)
 173 {
 174         unsigned long delta;
 175         unsigned int cpu;
 176         int rc;
 177
 178         /*
 179          * Always reserve area for module percpu variables.  That's
 180          * what the legacy allocator did.
 181          */
 182         rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
 183                                     PERCPU_DYNAMIC_RESERVE, PAGE_SIZE,
 184                                     pcpu_cpu_distance,
 185                                     pcpu_fc_alloc, pcpu_fc_free);
 186         if (rc < 0)
 187                 panic("Failed to initialize percpu areas.");
 188
 189         delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
 190         for_each_possible_cpu(cpu)
 191                 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
 192 }
 193 #endif
 194
 195 /**
 196  * numa_add_memblk() - Set node id to memblk
 197  * @nid: NUMA node ID of the new memblk
 198  * @start: Start address of the new memblk
 199  * @end:  End address of the new memblk
 200  *
 201  * RETURNS:
 202  * 0 on success, -errno on failure.
 203  */
 204 int __init numa_add_memblk(int nid, u64 start, u64 end)
 205 {
 206         int ret;
 207
 208         ret = memblock_set_node(start, (end - start), &memblock.memory, nid);
 209         if (ret < 0) {
 210                 pr_err("memblock [0x%llx - 0x%llx] failed to add on node %d\n",
 211                         start, (end - 1), nid);
 212                 return ret;
 213         }
 214
 215         node_set(nid, numa_nodes_parsed);
 216         return ret;
 217 }
 218
 219 /*
 220  * Initialize NODE_DATA for a node on the local memory
 221  */
 222 static void __init setup_node_data(int nid, u64 start_pfn, u64 end_pfn)
 223 {
 224         const size_t nd_size = roundup(sizeof(pg_data_t), SMP_CACHE_BYTES);
 225         u64 nd_pa;
 226         void *nd;
 227         int tnid;
 228
 229         if (start_pfn >= end_pfn)
 230                 pr_info("Initmem setup node %d [<memory-less node>]\n", nid);
 231
 232         nd_pa = memblock_phys_alloc_try_nid(nd_size, SMP_CACHE_BYTES, nid);
 233         if (!nd_pa)
 234                 panic("Cannot allocate %zu bytes for node %d data\n",
 235                       nd_size, nid);
 236
 237         nd = __va(nd_pa);
 238
 239         /* report and initialize */
 240         pr_info("NODE_DATA [mem %#010Lx-%#010Lx]\n",
 241                 nd_pa, nd_pa + nd_size - 1);
 242         tnid = early_pfn_to_nid(nd_pa >> PAGE_SHIFT);
 243         if (tnid != nid)
 244                 pr_info("NODE_DATA(%d) on node %d\n", nid, tnid);
 245
 246         node_data[nid] = nd;
 247         memset(NODE_DATA(nid), 0, sizeof(pg_data_t));
 248         NODE_DATA(nid)->node_id = nid;
 249         NODE_DATA(nid)->node_start_pfn = start_pfn;
 250         NODE_DATA(nid)->node_spanned_pages = end_pfn - start_pfn;
 251 }
 252
 253 /*
 254  * numa_free_distance
 255  *
 256  * The current table is freed.
 257  */
 258 void __init numa_free_distance(void)
 259 {
 260         size_t size;
 261
 262         if (!numa_distance)
 263                 return;
 264
 265         size = numa_distance_cnt * numa_distance_cnt *
 266                 sizeof(numa_distance[0]);
 267
 268         memblock_free(__pa(numa_distance), size);
 269         numa_distance_cnt = 0;
 270         numa_distance = NULL;
 271 }
 272
 273 /*
 274  * Create a new NUMA distance table.
 275  */
 276 static int __init numa_alloc_distance(void)
 277 {
 278         size_t size;
 279         u64 phys;
 280         int i, j;
 281
 282         size = nr_node_ids * nr_node_ids * sizeof(numa_distance[0]);
 283         phys = memblock_find_in_range(0, PFN_PHYS(max_pfn),
 284                                       size, PAGE_SIZE);
 285         if (WARN_ON(!phys))
 286                 return -ENOMEM;
 287
 288         memblock_reserve(phys, size);
 289
 290         numa_distance = __va(phys);
 291         numa_distance_cnt = nr_node_ids;
 292
 293         /* fill with the default distances */
 294         for (i = 0; i < numa_distance_cnt; i++)
 295                 for (j = 0; j < numa_distance_cnt; j++)
 296                         numa_distance[i * numa_distance_cnt + j] = i == j ?
 297                                 LOCAL_DISTANCE : REMOTE_DISTANCE;
 298
 299         pr_debug("Initialized distance table, cnt=%d\n", numa_distance_cnt);
 300
 301         return 0;
 302 }
 303
 304 /**
 305  * numa_set_distance() - Set inter node NUMA distance from node to node.
 306  * @from: the 'from' node to set distance
 307  * @to: the 'to'  node to set distance
 308  * @distance: NUMA distance
 309  *
 310  * Set the distance from node @from to @to to @distance.
 311  * If distance table doesn't exist, a warning is printed.
 312  *
 313  * If @from or @to is higher than the highest known node or lower than zero
 314  * or @distance doesn't make sense, the call is ignored.
 315  */
 316 void __init numa_set_distance(int from, int to, int distance)
 317 {
 318         if (!numa_distance) {
 319                 pr_warn_once("Warning: distance table not allocated yet\n");
 320                 return;
 321         }
 322
 323         if (from >= numa_distance_cnt || to >= numa_distance_cnt ||
 324                         from < 0 || to < 0) {
 325                 pr_warn_once("Warning: node ids are out of bound, from=%d to=%d distance=%d\n",
 326                             from, to, distance);
 327                 return;
 328         }
 329
 330         if ((u8)distance != distance ||
 331             (from == to && distance != LOCAL_DISTANCE)) {
 332                 pr_warn_once("Warning: invalid distance parameter, from=%d to=%d distance=%d\n",
 333                              from, to, distance);
 334                 return;
 335         }
 336
 337         numa_distance[from * numa_distance_cnt + to] = distance;
 338 }
 339
 340 /*
 341  * Return NUMA distance @from to @to
 342  */
 343 int __node_distance(int from, int to)
 344 {
 345         if (from >= numa_distance_cnt || to >= numa_distance_cnt)
 346                 return from == to ? LOCAL_DISTANCE : REMOTE_DISTANCE;
 347         return numa_distance[from * numa_distance_cnt + to];
 348 }
 349 EXPORT_SYMBOL(__node_distance);
 350
 351 static int __init numa_register_nodes(void)
 352 {
 353         int nid;
 354         struct memblock_region *mblk;
 355
 356         /* Check that valid nid is set to memblks */
 357         for_each_mem_region(mblk) {
 358                 int mblk_nid = memblock_get_region_node(mblk);
 359
 360                 if (mblk_nid == NUMA_NO_NODE || mblk_nid >= MAX_NUMNODES) {
 361                         pr_warn("Warning: invalid memblk node %d [mem %#010Lx-%#010Lx]\n",
 362                                 mblk_nid, mblk->base,
 363                                 mblk->base + mblk->size - 1);
 364                         return -EINVAL;
 365                 }
 366         }
 367
 368         /* Finally register nodes. */
 369         for_each_node_mask(nid, numa_nodes_parsed) {
 370                 unsigned long start_pfn, end_pfn;
 371
 372                 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
 373                 setup_node_data(nid, start_pfn, end_pfn);
 374                 node_set_online(nid);
 375         }
 376
 377         /* Setup online nodes to actual nodes*/
 378         node_possible_map = numa_nodes_parsed;
 379
 380         return 0;
 381 }
 382
 383 static int __init numa_init(int (*init_func)(void))
 384 {
 385         int ret;
 386
 387         nodes_clear(numa_nodes_parsed);
 388         nodes_clear(node_possible_map);
 389         nodes_clear(node_online_map);
 390
 391         ret = numa_alloc_distance();
 392         if (ret < 0)
 393                 return ret;
 394
 395         ret = init_func();
 396         if (ret < 0)
 397                 goto out_free_distance;
 398
 399         if (nodes_empty(numa_nodes_parsed)) {
 400                 pr_info("No NUMA configuration found\n");
 401                 ret = -EINVAL;
 402                 goto out_free_distance;
 403         }
 404
 405         ret = numa_register_nodes();
 406         if (ret < 0)
 407                 goto out_free_distance;
 408
 409         setup_node_to_cpumask_map();
 410
 411         return 0;
 412 out_free_distance:
 413         numa_free_distance();
 414         return ret;
 415 }
 416
 417 /**
 418  * dummy_numa_init() - Fallback dummy NUMA init
 419  *
 420  * Used if there's no underlying NUMA architecture, NUMA initialization
 421  * fails, or NUMA is disabled on the command line.
 422  *
 423  * Must online at least one node (node 0) and add memory blocks that cover all
 424  * allowed memory. It is unlikely that this function fails.
 425  *
 426  * Return: 0 on success, -errno on failure.
 427  */
 428 static int __init dummy_numa_init(void)
 429 {
 430         phys_addr_t start = memblock_start_of_DRAM();
 431         phys_addr_t end = memblock_end_of_DRAM();
 432         int ret;
 433
 434         if (numa_off)
 435                 pr_info("NUMA disabled\n"); /* Forced off on command line. */
 436         pr_info("Faking a node at [mem %#018Lx-%#018Lx]\n", start, end - 1);
 437
 438         ret = numa_add_memblk(0, start, end);
 439         if (ret) {
 440                 pr_err("NUMA init failed\n");
 441                 return ret;
 442         }
 443
 444         numa_off = true;
 445         return 0;
 446 }
 447
 448 /**
 449  * arm64_numa_init() - Initialize NUMA
 450  *
 451  * Try each configured NUMA initialization method until one succeeds. The
 452  * last fallback is dummy single node config encompassing whole memory.
 453  */
 454 void __init arm64_numa_init(void)
 455 {
 456         if (!numa_off) {
 457                 if (!acpi_disabled && !numa_init(arm64_acpi_numa_init))
 458                         return;
 459                 if (acpi_disabled && !numa_init(of_numa_init))
 460                         return;
 461         }
 462
 463         numa_init(dummy_numa_init);
 464 }