2 * NUMA support, based on the x86 implementation.
4 * Copyright (C) 2015 Cavium Inc.
5 * Author: Ganapatrao Kulkarni <gkulkarni@cavium.com>
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.
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.
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/>.
20 #define pr_fmt(fmt) "NUMA: " fmt
22 #include <linux/acpi.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/module.h>
29 #include <asm/sections.h>
31 struct pglist_data *node_data[MAX_NUMNODES] __read_mostly;
32 EXPORT_SYMBOL(node_data);
33 nodemask_t numa_nodes_parsed __initdata;
34 static int cpu_to_node_map[NR_CPUS] = { [0 ... NR_CPUS-1] = NUMA_NO_NODE };
36 static int numa_distance_cnt;
37 static u8 *numa_distance;
40 static __init int numa_parse_early_param(char *opt)
44 if (!strncmp(opt, "off", 3))
49 early_param("numa", numa_parse_early_param);
51 cpumask_var_t node_to_cpumask_map[MAX_NUMNODES];
52 EXPORT_SYMBOL(node_to_cpumask_map);
54 #ifdef CONFIG_DEBUG_PER_CPU_MAPS
57 * Returns a pointer to the bitmask of CPUs on Node 'node'.
59 const struct cpumask *cpumask_of_node(int node)
61 if (WARN_ON(node >= nr_node_ids))
64 if (WARN_ON(node_to_cpumask_map[node] == NULL))
65 return cpu_online_mask;
67 return node_to_cpumask_map[node];
69 EXPORT_SYMBOL(cpumask_of_node);
73 static void map_cpu_to_node(unsigned int cpu, int nid)
75 set_cpu_numa_node(cpu, nid);
77 cpumask_set_cpu(cpu, node_to_cpumask_map[nid]);
80 void numa_clear_node(unsigned int cpu)
82 int nid = cpu_to_node(cpu);
85 cpumask_clear_cpu(cpu, node_to_cpumask_map[nid]);
86 set_cpu_numa_node(cpu, NUMA_NO_NODE);
90 * Allocate node_to_cpumask_map based on number of available nodes
91 * Requires node_possible_map to be valid.
93 * Note: cpumask_of_node() is not valid until after this is done.
94 * (Use CONFIG_DEBUG_PER_CPU_MAPS to check this.)
96 static void __init setup_node_to_cpumask_map(void)
101 /* setup nr_node_ids if not done yet */
102 if (nr_node_ids == MAX_NUMNODES)
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]);
111 for_each_possible_cpu(cpu)
112 set_cpu_numa_node(cpu, NUMA_NO_NODE);
114 /* cpumask_of_node() will now work */
115 pr_debug("Node to cpumask map for %d nodes\n", nr_node_ids);
119 * Set the cpu to node and mem mapping
121 void numa_store_cpu_info(unsigned int cpu)
123 map_cpu_to_node(cpu, numa_off ? 0 : cpu_to_node_map[cpu]);
126 void __init early_map_cpu_to_node(unsigned int cpu, int nid)
128 /* fallback to node 0 */
129 if (nid < 0 || nid >= MAX_NUMNODES)
132 cpu_to_node_map[cpu] = nid;
135 #ifdef CONFIG_HAVE_SETUP_PER_CPU_AREA
136 unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
137 EXPORT_SYMBOL(__per_cpu_offset);
139 static int __init early_cpu_to_node(int cpu)
141 return cpu_to_node_map[cpu];
144 static int __init pcpu_cpu_distance(unsigned int from, unsigned int to)
146 return node_distance(from, to);
149 static void * __init pcpu_fc_alloc(unsigned int cpu, size_t size,
152 int nid = early_cpu_to_node(cpu);
154 return memblock_virt_alloc_try_nid(size, align,
155 __pa(MAX_DMA_ADDRESS), MEMBLOCK_ALLOC_ACCESSIBLE, nid);
158 static void __init pcpu_fc_free(void *ptr, size_t size)
160 memblock_free_early(__pa(ptr), size);
163 void __init setup_per_cpu_areas(void)
170 * Always reserve area for module percpu variables. That's
171 * what the legacy allocator did.
173 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
174 PERCPU_DYNAMIC_RESERVE, PAGE_SIZE,
176 pcpu_fc_alloc, pcpu_fc_free);
178 panic("Failed to initialize percpu areas.");
180 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
181 for_each_possible_cpu(cpu)
182 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
187 * numa_add_memblk - Set node id to memblk
188 * @nid: NUMA node ID of the new memblk
189 * @start: Start address of the new memblk
190 * @end: End address of the new memblk
193 * 0 on success, -errno on failure.
195 int __init numa_add_memblk(int nid, u64 start, u64 end)
199 ret = memblock_set_node(start, (end - start), &memblock.memory, nid);
201 pr_err("memblock [0x%llx - 0x%llx] failed to add on node %d\n",
202 start, (end - 1), nid);
206 node_set(nid, numa_nodes_parsed);
207 pr_info("Adding memblock [0x%llx - 0x%llx] on node %d\n",
208 start, (end - 1), nid);
213 * Initialize NODE_DATA for a node on the local memory
215 static void __init setup_node_data(int nid, u64 start_pfn, u64 end_pfn)
217 const size_t nd_size = roundup(sizeof(pg_data_t), SMP_CACHE_BYTES);
222 pr_info("Initmem setup node %d [mem %#010Lx-%#010Lx]\n",
223 nid, start_pfn << PAGE_SHIFT, (end_pfn << PAGE_SHIFT) - 1);
225 nd_pa = memblock_alloc_try_nid(nd_size, SMP_CACHE_BYTES, nid);
228 /* report and initialize */
229 pr_info("NODE_DATA [mem %#010Lx-%#010Lx]\n",
230 nd_pa, nd_pa + nd_size - 1);
231 tnid = early_pfn_to_nid(nd_pa >> PAGE_SHIFT);
233 pr_info("NODE_DATA(%d) on node %d\n", nid, tnid);
236 memset(NODE_DATA(nid), 0, sizeof(pg_data_t));
237 NODE_DATA(nid)->node_id = nid;
238 NODE_DATA(nid)->node_start_pfn = start_pfn;
239 NODE_DATA(nid)->node_spanned_pages = end_pfn - start_pfn;
245 * The current table is freed.
247 void __init numa_free_distance(void)
254 size = numa_distance_cnt * numa_distance_cnt *
255 sizeof(numa_distance[0]);
257 memblock_free(__pa(numa_distance), size);
258 numa_distance_cnt = 0;
259 numa_distance = NULL;
264 * Create a new NUMA distance table.
267 static int __init numa_alloc_distance(void)
273 size = nr_node_ids * nr_node_ids * sizeof(numa_distance[0]);
274 phys = memblock_find_in_range(0, PFN_PHYS(max_pfn),
279 memblock_reserve(phys, size);
281 numa_distance = __va(phys);
282 numa_distance_cnt = nr_node_ids;
284 /* fill with the default distances */
285 for (i = 0; i < numa_distance_cnt; i++)
286 for (j = 0; j < numa_distance_cnt; j++)
287 numa_distance[i * numa_distance_cnt + j] = i == j ?
288 LOCAL_DISTANCE : REMOTE_DISTANCE;
290 pr_debug("Initialized distance table, cnt=%d\n", numa_distance_cnt);
296 * numa_set_distance - Set inter node NUMA distance from node to node.
297 * @from: the 'from' node to set distance
298 * @to: the 'to' node to set distance
299 * @distance: NUMA distance
301 * Set the distance from node @from to @to to @distance.
302 * If distance table doesn't exist, a warning is printed.
304 * If @from or @to is higher than the highest known node or lower than zero
305 * or @distance doesn't make sense, the call is ignored.
308 void __init numa_set_distance(int from, int to, int distance)
310 if (!numa_distance) {
311 pr_warn_once("Warning: distance table not allocated yet\n");
315 if (from >= numa_distance_cnt || to >= numa_distance_cnt ||
316 from < 0 || to < 0) {
317 pr_warn_once("Warning: node ids are out of bound, from=%d to=%d distance=%d\n",
322 if ((u8)distance != distance ||
323 (from == to && distance != LOCAL_DISTANCE)) {
324 pr_warn_once("Warning: invalid distance parameter, from=%d to=%d distance=%d\n",
329 numa_distance[from * numa_distance_cnt + to] = distance;
333 * Return NUMA distance @from to @to
335 int __node_distance(int from, int to)
337 if (from >= numa_distance_cnt || to >= numa_distance_cnt)
338 return from == to ? LOCAL_DISTANCE : REMOTE_DISTANCE;
339 return numa_distance[from * numa_distance_cnt + to];
341 EXPORT_SYMBOL(__node_distance);
343 static int __init numa_register_nodes(void)
346 struct memblock_region *mblk;
348 /* Check that valid nid is set to memblks */
349 for_each_memblock(memory, mblk)
350 if (mblk->nid == NUMA_NO_NODE || mblk->nid >= MAX_NUMNODES) {
351 pr_warn("Warning: invalid memblk node %d [mem %#010Lx-%#010Lx]\n",
352 mblk->nid, mblk->base,
353 mblk->base + mblk->size - 1);
357 /* Finally register nodes. */
358 for_each_node_mask(nid, numa_nodes_parsed) {
359 unsigned long start_pfn, end_pfn;
361 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
362 setup_node_data(nid, start_pfn, end_pfn);
363 node_set_online(nid);
366 /* Setup online nodes to actual nodes*/
367 node_possible_map = numa_nodes_parsed;
372 static int __init numa_init(int (*init_func)(void))
376 nodes_clear(numa_nodes_parsed);
377 nodes_clear(node_possible_map);
378 nodes_clear(node_online_map);
379 numa_free_distance();
381 ret = numa_alloc_distance();
389 if (nodes_empty(numa_nodes_parsed)) {
390 pr_info("No NUMA configuration found\n");
394 ret = numa_register_nodes();
398 setup_node_to_cpumask_map();
400 /* init boot processor */
401 cpu_to_node_map[0] = 0;
402 map_cpu_to_node(0, 0);
408 * dummy_numa_init - Fallback dummy NUMA init
410 * Used if there's no underlying NUMA architecture, NUMA initialization
411 * fails, or NUMA is disabled on the command line.
413 * Must online at least one node (node 0) and add memory blocks that cover all
414 * allowed memory. It is unlikely that this function fails.
416 static int __init dummy_numa_init(void)
419 struct memblock_region *mblk;
422 pr_info("NUMA disabled\n"); /* Forced off on command line. */
423 pr_info("Faking a node at [mem %#018Lx-%#018Lx]\n",
424 0LLU, PFN_PHYS(max_pfn) - 1);
426 for_each_memblock(memory, mblk) {
427 ret = numa_add_memblk(0, mblk->base, mblk->base + mblk->size);
431 pr_err("NUMA init failed\n");
440 * arm64_numa_init - Initialize NUMA
442 * Try each configured NUMA initialization method until one succeeds. The
443 * last fallback is dummy single node config encomapssing whole memory.
445 void __init arm64_numa_init(void)
448 if (!acpi_disabled && !numa_init(arm64_acpi_numa_init))
450 if (acpi_disabled && !numa_init(of_numa_init))
454 numa_init(dummy_numa_init);