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)
62 if (node == NUMA_NO_NODE)
65 if (WARN_ON(node < 0 || node >= nr_node_ids))
68 if (WARN_ON(node_to_cpumask_map[node] == NULL))
69 return cpu_online_mask;
71 return node_to_cpumask_map[node];
73 EXPORT_SYMBOL(cpumask_of_node);
77 static void numa_update_cpu(unsigned int cpu, bool remove)
79 int nid = cpu_to_node(cpu);
81 if (nid == NUMA_NO_NODE)
85 cpumask_clear_cpu(cpu, node_to_cpumask_map[nid]);
87 cpumask_set_cpu(cpu, node_to_cpumask_map[nid]);
90 void numa_add_cpu(unsigned int cpu)
92 numa_update_cpu(cpu, false);
95 void numa_remove_cpu(unsigned int cpu)
97 numa_update_cpu(cpu, true);
100 void numa_clear_node(unsigned int cpu)
102 numa_remove_cpu(cpu);
103 set_cpu_numa_node(cpu, NUMA_NO_NODE);
107 * Allocate node_to_cpumask_map based on number of available nodes
108 * Requires node_possible_map to be valid.
110 * Note: cpumask_of_node() is not valid until after this is done.
111 * (Use CONFIG_DEBUG_PER_CPU_MAPS to check this.)
113 static void __init setup_node_to_cpumask_map(void)
117 /* setup nr_node_ids if not done yet */
118 if (nr_node_ids == MAX_NUMNODES)
121 /* allocate and clear the mapping */
122 for (node = 0; node < nr_node_ids; node++) {
123 alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]);
124 cpumask_clear(node_to_cpumask_map[node]);
127 /* cpumask_of_node() will now work */
128 pr_debug("Node to cpumask map for %d nodes\n", nr_node_ids);
132 * Set the cpu to node and mem mapping
134 void numa_store_cpu_info(unsigned int cpu)
136 set_cpu_numa_node(cpu, cpu_to_node_map[cpu]);
139 void __init early_map_cpu_to_node(unsigned int cpu, int nid)
141 /* fallback to node 0 */
142 if (nid < 0 || nid >= MAX_NUMNODES || numa_off)
145 cpu_to_node_map[cpu] = nid;
148 * We should set the numa node of cpu0 as soon as possible, because it
149 * has already been set up online before. cpu_to_node(0) will soon be
153 set_cpu_numa_node(cpu, nid);
156 #ifdef CONFIG_HAVE_SETUP_PER_CPU_AREA
157 unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
158 EXPORT_SYMBOL(__per_cpu_offset);
160 static int __init early_cpu_to_node(int cpu)
162 return cpu_to_node_map[cpu];
165 static int __init pcpu_cpu_distance(unsigned int from, unsigned int to)
167 return node_distance(early_cpu_to_node(from), early_cpu_to_node(to));
170 static void * __init pcpu_fc_alloc(unsigned int cpu, size_t size,
173 int nid = early_cpu_to_node(cpu);
175 return memblock_virt_alloc_try_nid(size, align,
176 __pa(MAX_DMA_ADDRESS), MEMBLOCK_ALLOC_ACCESSIBLE, nid);
179 static void __init pcpu_fc_free(void *ptr, size_t size)
181 memblock_free_early(__pa(ptr), size);
184 void __init setup_per_cpu_areas(void)
191 * Always reserve area for module percpu variables. That's
192 * what the legacy allocator did.
194 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
195 PERCPU_DYNAMIC_RESERVE, PAGE_SIZE,
197 pcpu_fc_alloc, pcpu_fc_free);
199 panic("Failed to initialize percpu areas.");
201 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
202 for_each_possible_cpu(cpu)
203 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
208 * numa_add_memblk - Set node id to memblk
209 * @nid: NUMA node ID of the new memblk
210 * @start: Start address of the new memblk
211 * @end: End address of the new memblk
214 * 0 on success, -errno on failure.
216 int __init numa_add_memblk(int nid, u64 start, u64 end)
220 ret = memblock_set_node(start, (end - start), &memblock.memory, nid);
222 pr_err("memblock [0x%llx - 0x%llx] failed to add on node %d\n",
223 start, (end - 1), nid);
227 node_set(nid, numa_nodes_parsed);
232 * Initialize NODE_DATA for a node on the local memory
234 static void __init setup_node_data(int nid, u64 start_pfn, u64 end_pfn)
236 const size_t nd_size = roundup(sizeof(pg_data_t), SMP_CACHE_BYTES);
241 if (start_pfn >= end_pfn)
242 pr_info("Initmem setup node %d [<memory-less node>]\n", nid);
244 nd_pa = memblock_alloc_try_nid(nd_size, SMP_CACHE_BYTES, nid);
247 /* report and initialize */
248 pr_info("NODE_DATA [mem %#010Lx-%#010Lx]\n",
249 nd_pa, nd_pa + nd_size - 1);
250 tnid = early_pfn_to_nid(nd_pa >> PAGE_SHIFT);
252 pr_info("NODE_DATA(%d) on node %d\n", nid, tnid);
255 memset(NODE_DATA(nid), 0, sizeof(pg_data_t));
256 NODE_DATA(nid)->node_id = nid;
257 NODE_DATA(nid)->node_start_pfn = start_pfn;
258 NODE_DATA(nid)->node_spanned_pages = end_pfn - start_pfn;
264 * The current table is freed.
266 void __init numa_free_distance(void)
273 size = numa_distance_cnt * numa_distance_cnt *
274 sizeof(numa_distance[0]);
276 memblock_free(__pa(numa_distance), size);
277 numa_distance_cnt = 0;
278 numa_distance = NULL;
283 * Create a new NUMA distance table.
286 static int __init numa_alloc_distance(void)
292 size = nr_node_ids * nr_node_ids * sizeof(numa_distance[0]);
293 phys = memblock_find_in_range(0, PFN_PHYS(max_pfn),
298 memblock_reserve(phys, size);
300 numa_distance = __va(phys);
301 numa_distance_cnt = nr_node_ids;
303 /* fill with the default distances */
304 for (i = 0; i < numa_distance_cnt; i++)
305 for (j = 0; j < numa_distance_cnt; j++)
306 numa_distance[i * numa_distance_cnt + j] = i == j ?
307 LOCAL_DISTANCE : REMOTE_DISTANCE;
309 pr_debug("Initialized distance table, cnt=%d\n", numa_distance_cnt);
315 * numa_set_distance - Set inter node NUMA distance from node to node.
316 * @from: the 'from' node to set distance
317 * @to: the 'to' node to set distance
318 * @distance: NUMA distance
320 * Set the distance from node @from to @to to @distance.
321 * If distance table doesn't exist, a warning is printed.
323 * If @from or @to is higher than the highest known node or lower than zero
324 * or @distance doesn't make sense, the call is ignored.
327 void __init numa_set_distance(int from, int to, int distance)
329 if (!numa_distance) {
330 pr_warn_once("Warning: distance table not allocated yet\n");
334 if (from >= numa_distance_cnt || to >= numa_distance_cnt ||
335 from < 0 || to < 0) {
336 pr_warn_once("Warning: node ids are out of bound, from=%d to=%d distance=%d\n",
341 if ((u8)distance != distance ||
342 (from == to && distance != LOCAL_DISTANCE)) {
343 pr_warn_once("Warning: invalid distance parameter, from=%d to=%d distance=%d\n",
348 numa_distance[from * numa_distance_cnt + to] = distance;
352 * Return NUMA distance @from to @to
354 int __node_distance(int from, int to)
356 if (from >= numa_distance_cnt || to >= numa_distance_cnt)
357 return from == to ? LOCAL_DISTANCE : REMOTE_DISTANCE;
358 return numa_distance[from * numa_distance_cnt + to];
360 EXPORT_SYMBOL(__node_distance);
362 static int __init numa_register_nodes(void)
365 struct memblock_region *mblk;
367 /* Check that valid nid is set to memblks */
368 for_each_memblock(memory, mblk)
369 if (mblk->nid == NUMA_NO_NODE || mblk->nid >= MAX_NUMNODES) {
370 pr_warn("Warning: invalid memblk node %d [mem %#010Lx-%#010Lx]\n",
371 mblk->nid, mblk->base,
372 mblk->base + mblk->size - 1);
376 /* Finally register nodes. */
377 for_each_node_mask(nid, numa_nodes_parsed) {
378 unsigned long start_pfn, end_pfn;
380 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
381 setup_node_data(nid, start_pfn, end_pfn);
382 node_set_online(nid);
385 /* Setup online nodes to actual nodes*/
386 node_possible_map = numa_nodes_parsed;
391 static int __init numa_init(int (*init_func)(void))
395 nodes_clear(numa_nodes_parsed);
396 nodes_clear(node_possible_map);
397 nodes_clear(node_online_map);
398 numa_free_distance();
400 ret = numa_alloc_distance();
408 if (nodes_empty(numa_nodes_parsed)) {
409 pr_info("No NUMA configuration found\n");
413 ret = numa_register_nodes();
417 setup_node_to_cpumask_map();
423 * dummy_numa_init - Fallback dummy NUMA init
425 * Used if there's no underlying NUMA architecture, NUMA initialization
426 * fails, or NUMA is disabled on the command line.
428 * Must online at least one node (node 0) and add memory blocks that cover all
429 * allowed memory. It is unlikely that this function fails.
431 static int __init dummy_numa_init(void)
434 struct memblock_region *mblk;
437 pr_info("NUMA disabled\n"); /* Forced off on command line. */
438 pr_info("Faking a node at [mem %#018Lx-%#018Lx]\n",
439 memblock_start_of_DRAM(), memblock_end_of_DRAM() - 1);
441 for_each_memblock(memory, mblk) {
442 ret = numa_add_memblk(0, mblk->base, mblk->base + mblk->size);
446 pr_err("NUMA init failed\n");
455 * arm64_numa_init - Initialize NUMA
457 * Try each configured NUMA initialization method until one succeeds. The
458 * last fallback is dummy single node config encomapssing whole memory.
460 void __init arm64_numa_init(void)
463 if (!acpi_disabled && !numa_init(arm64_acpi_numa_init))
465 if (acpi_disabled && !numa_init(of_numa_init))
469 numa_init(dummy_numa_init);