arch/arm64/mm/numa.c

   1 /*
   2  * NUMA support, based on the x86 implementation.
   3  *
   4  * Copyright (C) 2015 Cavium Inc.
   5  * Author: Ganapatrao Kulkarni <gkulkarni@cavium.com>
   6  *
   7  * This program is free software; you can redistribute it and/or modify
   8  * it under the terms of the GNU General Public License version 2 as
   9  * published by the Free Software Foundation.
  10  *
  11  * This program is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  14  * GNU General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * along with this program.  If not, see <http://www.gnu.org/licenses/>.
  18  */
  19
  20 #define pr_fmt(fmt) "NUMA: " fmt
  21
  22 #include <linux/acpi.h>
  23 #include <linux/memblock.h>
  24 #include <linux/module.h>
  25 #include <linux/of.h>
  26
  27 #include <asm/acpi.h>
  28 #include <asm/sections.h>
  29
  30 struct pglist_data *node_data[MAX_NUMNODES] __read_mostly;
  31 EXPORT_SYMBOL(node_data);
  32 nodemask_t numa_nodes_parsed __initdata;
  33 static int cpu_to_node_map[NR_CPUS] = { [0 ... NR_CPUS-1] = NUMA_NO_NODE };
  34
  35 static int numa_distance_cnt;
  36 static u8 *numa_distance;
  37 bool numa_off;
  38
  39 static __init int numa_parse_early_param(char *opt)
  40 {
  41         if (!opt)
  42                 return -EINVAL;
  43         if (!strncmp(opt, "off", 3))
  44                 numa_off = true;
  45
  46         return 0;
  47 }
  48 early_param("numa", numa_parse_early_param);
  49
  50 cpumask_var_t node_to_cpumask_map[MAX_NUMNODES];
  51 EXPORT_SYMBOL(node_to_cpumask_map);
  52
  53 #ifdef CONFIG_DEBUG_PER_CPU_MAPS
  54
  55 /*
  56  * Returns a pointer to the bitmask of CPUs on Node 'node'.
  57  */
  58 const struct cpumask *cpumask_of_node(int node)
  59 {
  60         if (WARN_ON(node >= nr_node_ids))
  61                 return cpu_none_mask;
  62
  63         if (WARN_ON(node_to_cpumask_map[node] == NULL))
  64                 return cpu_online_mask;
  65
  66         return node_to_cpumask_map[node];
  67 }
  68 EXPORT_SYMBOL(cpumask_of_node);
  69
  70 #endif
  71
  72 static void numa_update_cpu(unsigned int cpu, bool remove)
  73 {
  74         int nid = cpu_to_node(cpu);
  75
  76         if (nid == NUMA_NO_NODE)
  77                 return;
  78
  79         if (remove)
  80                 cpumask_clear_cpu(cpu, node_to_cpumask_map[nid]);
  81         else
  82                 cpumask_set_cpu(cpu, node_to_cpumask_map[nid]);
  83 }
  84
  85 void numa_add_cpu(unsigned int cpu)
  86 {
  87         numa_update_cpu(cpu, false);
  88 }
  89
  90 void numa_remove_cpu(unsigned int cpu)
  91 {
  92         numa_update_cpu(cpu, true);
  93 }
  94
  95 void numa_clear_node(unsigned int cpu)
  96 {
  97         numa_remove_cpu(cpu);
  98         set_cpu_numa_node(cpu, NUMA_NO_NODE);
  99 }
 100
 101 /*
 102  * Allocate node_to_cpumask_map based on number of available nodes
 103  * Requires node_possible_map to be valid.
 104  *
 105  * Note: cpumask_of_node() is not valid until after this is done.
 106  * (Use CONFIG_DEBUG_PER_CPU_MAPS to check this.)
 107  */
 108 static void __init setup_node_to_cpumask_map(void)
 109 {
 110         int node;
 111
 112         /* setup nr_node_ids if not done yet */
 113         if (nr_node_ids == MAX_NUMNODES)
 114                 setup_nr_node_ids();
 115
 116         /* allocate and clear the mapping */
 117         for (node = 0; node < nr_node_ids; node++) {
 118                 alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]);
 119                 cpumask_clear(node_to_cpumask_map[node]);
 120         }
 121
 122         /* cpumask_of_node() will now work */
 123         pr_debug("Node to cpumask map for %u nodes\n", nr_node_ids);
 124 }
 125
 126 /*
 127  *  Set the cpu to node and mem mapping
 128  */
 129 void numa_store_cpu_info(unsigned int cpu)
 130 {
 131         set_cpu_numa_node(cpu, cpu_to_node_map[cpu]);
 132 }
 133
 134 void __init early_map_cpu_to_node(unsigned int cpu, int nid)
 135 {
 136         /* fallback to node 0 */
 137         if (nid < 0 || nid >= MAX_NUMNODES || numa_off)
 138                 nid = 0;
 139
 140         cpu_to_node_map[cpu] = nid;
 141
 142         /*
 143          * We should set the numa node of cpu0 as soon as possible, because it
 144          * has already been set up online before. cpu_to_node(0) will soon be
 145          * called.
 146          */
 147         if (!cpu)
 148                 set_cpu_numa_node(cpu, nid);
 149 }
 150
 151 #ifdef CONFIG_HAVE_SETUP_PER_CPU_AREA
 152 unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
 153 EXPORT_SYMBOL(__per_cpu_offset);
 154
 155 static int __init early_cpu_to_node(int cpu)
 156 {
 157         return cpu_to_node_map[cpu];
 158 }
 159
 160 static int __init pcpu_cpu_distance(unsigned int from, unsigned int to)
 161 {
 162         return node_distance(early_cpu_to_node(from), early_cpu_to_node(to));
 163 }
 164
 165 static void * __init pcpu_fc_alloc(unsigned int cpu, size_t size,
 166                                        size_t align)
 167 {
 168         int nid = early_cpu_to_node(cpu);
 169
 170         return  memblock_alloc_try_nid(size, align,
 171                         __pa(MAX_DMA_ADDRESS), MEMBLOCK_ALLOC_ACCESSIBLE, nid);
 172 }
 173
 174 static void __init pcpu_fc_free(void *ptr, size_t size)
 175 {
 176         memblock_free_early(__pa(ptr), size);
 177 }
 178
 179 void __init setup_per_cpu_areas(void)
 180 {
 181         unsigned long delta;
 182         unsigned int cpu;
 183         int rc;
 184
 185         /*
 186          * Always reserve area for module percpu variables.  That's
 187          * what the legacy allocator did.
 188          */
 189         rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
 190                                     PERCPU_DYNAMIC_RESERVE, PAGE_SIZE,
 191                                     pcpu_cpu_distance,
 192                                     pcpu_fc_alloc, pcpu_fc_free);
 193         if (rc < 0)
 194                 panic("Failed to initialize percpu areas.");
 195
 196         delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
 197         for_each_possible_cpu(cpu)
 198                 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
 199 }
 200 #endif
 201
 202 /**
 203  * numa_add_memblk - Set node id to memblk
 204  * @nid: NUMA node ID of the new memblk
 205  * @start: Start address of the new memblk
 206  * @end:  End address of the new memblk
 207  *
 208  * RETURNS:
 209  * 0 on success, -errno on failure.
 210  */
 211 int __init numa_add_memblk(int nid, u64 start, u64 end)
 212 {
 213         int ret;
 214
 215         ret = memblock_set_node(start, (end - start), &memblock.memory, nid);
 216         if (ret < 0) {
 217                 pr_err("memblock [0x%llx - 0x%llx] failed to add on node %d\n",
 218                         start, (end - 1), nid);
 219                 return ret;
 220         }
 221
 222         node_set(nid, numa_nodes_parsed);
 223         return ret;
 224 }
 225
 226 /**
 227  * Initialize NODE_DATA for a node on the local memory
 228  */
 229 static void __init setup_node_data(int nid, u64 start_pfn, u64 end_pfn)
 230 {
 231         const size_t nd_size = roundup(sizeof(pg_data_t), SMP_CACHE_BYTES);
 232         u64 nd_pa;
 233         void *nd;
 234         int tnid;
 235
 236         if (start_pfn >= end_pfn)
 237                 pr_info("Initmem setup node %d [<memory-less node>]\n", nid);
 238
 239         nd_pa = memblock_phys_alloc_try_nid(nd_size, SMP_CACHE_BYTES, nid);
 240         nd = __va(nd_pa);
 241
 242         /* report and initialize */
 243         pr_info("NODE_DATA [mem %#010Lx-%#010Lx]\n",
 244                 nd_pa, nd_pa + nd_size - 1);
 245         tnid = early_pfn_to_nid(nd_pa >> PAGE_SHIFT);
 246         if (tnid != nid)
 247                 pr_info("NODE_DATA(%d) on node %d\n", nid, tnid);
 248
 249         node_data[nid] = nd;
 250         memset(NODE_DATA(nid), 0, sizeof(pg_data_t));
 251         NODE_DATA(nid)->node_id = nid;
 252         NODE_DATA(nid)->node_start_pfn = start_pfn;
 253         NODE_DATA(nid)->node_spanned_pages = end_pfn - start_pfn;
 254 }
 255
 256 /**
 257  * numa_free_distance
 258  *
 259  * The current table is freed.
 260  */
 261 void __init numa_free_distance(void)
 262 {
 263         size_t size;
 264
 265         if (!numa_distance)
 266                 return;
 267
 268         size = numa_distance_cnt * numa_distance_cnt *
 269                 sizeof(numa_distance[0]);
 270
 271         memblock_free(__pa(numa_distance), size);
 272         numa_distance_cnt = 0;
 273         numa_distance = NULL;
 274 }
 275
 276 /**
 277  *
 278  * Create a new NUMA distance table.
 279  *
 280  */
 281 static int __init numa_alloc_distance(void)
 282 {
 283         size_t size;
 284         u64 phys;
 285         int i, j;
 286
 287         size = nr_node_ids * nr_node_ids * sizeof(numa_distance[0]);
 288         phys = memblock_find_in_range(0, PFN_PHYS(max_pfn),
 289                                       size, PAGE_SIZE);
 290         if (WARN_ON(!phys))
 291                 return -ENOMEM;
 292
 293         memblock_reserve(phys, size);
 294
 295         numa_distance = __va(phys);
 296         numa_distance_cnt = nr_node_ids;
 297
 298         /* fill with the default distances */
 299         for (i = 0; i < numa_distance_cnt; i++)
 300                 for (j = 0; j < numa_distance_cnt; j++)
 301                         numa_distance[i * numa_distance_cnt + j] = i == j ?
 302                                 LOCAL_DISTANCE : REMOTE_DISTANCE;
 303
 304         pr_debug("Initialized distance table, cnt=%d\n", numa_distance_cnt);
 305
 306         return 0;
 307 }
 308
 309 /**
 310  * numa_set_distance - Set inter node NUMA distance from node to node.
 311  * @from: the 'from' node to set distance
 312  * @to: the 'to'  node to set distance
 313  * @distance: NUMA distance
 314  *
 315  * Set the distance from node @from to @to to @distance.
 316  * If distance table doesn't exist, a warning is printed.
 317  *
 318  * If @from or @to is higher than the highest known node or lower than zero
 319  * or @distance doesn't make sense, the call is ignored.
 320  *
 321  */
 322 void __init numa_set_distance(int from, int to, int distance)
 323 {
 324         if (!numa_distance) {
 325                 pr_warn_once("Warning: distance table not allocated yet\n");
 326                 return;
 327         }
 328
 329         if (from >= numa_distance_cnt || to >= numa_distance_cnt ||
 330                         from < 0 || to < 0) {
 331                 pr_warn_once("Warning: node ids are out of bound, from=%d to=%d distance=%d\n",
 332                             from, to, distance);
 333                 return;
 334         }
 335
 336         if ((u8)distance != distance ||
 337             (from == to && distance != LOCAL_DISTANCE)) {
 338                 pr_warn_once("Warning: invalid distance parameter, from=%d to=%d distance=%d\n",
 339                              from, to, distance);
 340                 return;
 341         }
 342
 343         numa_distance[from * numa_distance_cnt + to] = distance;
 344 }
 345
 346 /**
 347  * Return NUMA distance @from to @to
 348  */
 349 int __node_distance(int from, int to)
 350 {
 351         if (from >= numa_distance_cnt || to >= numa_distance_cnt)
 352                 return from == to ? LOCAL_DISTANCE : REMOTE_DISTANCE;
 353         return numa_distance[from * numa_distance_cnt + to];
 354 }
 355 EXPORT_SYMBOL(__node_distance);
 356
 357 static int __init numa_register_nodes(void)
 358 {
 359         int nid;
 360         struct memblock_region *mblk;
 361
 362         /* Check that valid nid is set to memblks */
 363         for_each_memblock(memory, mblk)
 364                 if (mblk->nid == NUMA_NO_NODE || mblk->nid >= MAX_NUMNODES) {
 365                         pr_warn("Warning: invalid memblk node %d [mem %#010Lx-%#010Lx]\n",
 366                                 mblk->nid, mblk->base,
 367                                 mblk->base + mblk->size - 1);
 368                         return -EINVAL;
 369                 }
 370
 371         /* Finally register nodes. */
 372         for_each_node_mask(nid, numa_nodes_parsed) {
 373                 unsigned long start_pfn, end_pfn;
 374
 375                 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
 376                 setup_node_data(nid, start_pfn, end_pfn);
 377                 node_set_online(nid);
 378         }
 379
 380         /* Setup online nodes to actual nodes*/
 381         node_possible_map = numa_nodes_parsed;
 382
 383         return 0;
 384 }
 385
 386 static int __init numa_init(int (*init_func)(void))
 387 {
 388         int ret;
 389
 390         nodes_clear(numa_nodes_parsed);
 391         nodes_clear(node_possible_map);
 392         nodes_clear(node_online_map);
 393
 394         ret = numa_alloc_distance();
 395         if (ret < 0)
 396                 return ret;
 397
 398         ret = init_func();
 399         if (ret < 0)
 400                 goto out_free_distance;
 401
 402         if (nodes_empty(numa_nodes_parsed)) {
 403                 pr_info("No NUMA configuration found\n");
 404                 ret = -EINVAL;
 405                 goto out_free_distance;
 406         }
 407
 408         ret = numa_register_nodes();
 409         if (ret < 0)
 410                 goto out_free_distance;
 411
 412         setup_node_to_cpumask_map();
 413
 414         return 0;
 415 out_free_distance:
 416         numa_free_distance();
 417         return ret;
 418 }
 419
 420 /**
 421  * dummy_numa_init - Fallback dummy NUMA init
 422  *
 423  * Used if there's no underlying NUMA architecture, NUMA initialization
 424  * fails, or NUMA is disabled on the command line.
 425  *
 426  * Must online at least one node (node 0) and add memory blocks that cover all
 427  * allowed memory. It is unlikely that this function fails.
 428  */
 429 static int __init dummy_numa_init(void)
 430 {
 431         int ret;
 432         struct memblock_region *mblk;
 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",
 437                 memblock_start_of_DRAM(), memblock_end_of_DRAM() - 1);
 438
 439         for_each_memblock(memory, mblk) {
 440                 ret = numa_add_memblk(0, mblk->base, mblk->base + mblk->size);
 441                 if (!ret)
 442                         continue;
 443
 444                 pr_err("NUMA init failed\n");
 445                 return ret;
 446         }
 447
 448         numa_off = true;
 449         return 0;
 450 }
 451
 452 /**
 453  * arm64_numa_init - Initialize NUMA
 454  *
 455  * Try each configured NUMA initialization method until one succeeds.  The
 456  * last fallback is dummy single node config encomapssing whole memory.
 457  */
 458 void __init arm64_numa_init(void)
 459 {
 460         if (!numa_off) {
 461                 if (!acpi_disabled && !numa_init(arm64_acpi_numa_init))
 462                         return;
 463                 if (acpi_disabled && !numa_init(of_numa_init))
 464                         return;
 465         }
 466
 467         numa_init(dummy_numa_init);
 468 }
 469
 470 /*
 471  * We hope that we will be hotplugging memory on nodes we already know about,
 472  * such that acpi_get_node() succeeds and we never fall back to this...
 473  */
 474 int memory_add_physaddr_to_nid(u64 addr)
 475 {
 476         pr_warn("Unknown node for memory at 0x%llx, assuming node 0\n", addr);
 477         return 0;
 478 }