1 /* Common code for 32 and 64-bit NUMA */
2 #include <linux/kernel.h>
4 #include <linux/string.h>
5 #include <linux/init.h>
6 #include <linux/bootmem.h>
7 #include <linux/memblock.h>
8 #include <linux/mmzone.h>
9 #include <linux/ctype.h>
10 #include <linux/module.h>
11 #include <linux/nodemask.h>
12 #include <linux/sched.h>
13 #include <linux/topology.h>
16 #include <asm/proto.h>
19 #include <asm/amd_nb.h>
21 #include "numa_internal.h"
23 int __initdata numa_off;
24 nodemask_t numa_nodes_parsed __initdata;
26 struct pglist_data *node_data[MAX_NUMNODES] __read_mostly;
27 EXPORT_SYMBOL(node_data);
29 static struct numa_meminfo numa_meminfo
30 #ifndef CONFIG_MEMORY_HOTPLUG
35 static int numa_distance_cnt;
36 static u8 *numa_distance;
38 static __init int numa_setup(char *opt)
42 if (!strncmp(opt, "off", 3))
44 #ifdef CONFIG_NUMA_EMU
45 if (!strncmp(opt, "fake=", 5))
46 numa_emu_cmdline(opt + 5);
48 #ifdef CONFIG_ACPI_NUMA
49 if (!strncmp(opt, "noacpi", 6))
54 early_param("numa", numa_setup);
57 * apicid, cpu, node mappings
59 s16 __apicid_to_node[MAX_LOCAL_APIC] = {
60 [0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE
63 int numa_cpu_node(int cpu)
65 int apicid = early_per_cpu(x86_cpu_to_apicid, cpu);
67 if (apicid != BAD_APICID)
68 return __apicid_to_node[apicid];
72 cpumask_var_t node_to_cpumask_map[MAX_NUMNODES];
73 EXPORT_SYMBOL(node_to_cpumask_map);
76 * Map cpu index to node index
78 DEFINE_EARLY_PER_CPU(int, x86_cpu_to_node_map, NUMA_NO_NODE);
79 EXPORT_EARLY_PER_CPU_SYMBOL(x86_cpu_to_node_map);
81 void numa_set_node(int cpu, int node)
83 int *cpu_to_node_map = early_per_cpu_ptr(x86_cpu_to_node_map);
85 /* early setting, no percpu area yet */
86 if (cpu_to_node_map) {
87 cpu_to_node_map[cpu] = node;
91 #ifdef CONFIG_DEBUG_PER_CPU_MAPS
92 if (cpu >= nr_cpu_ids || !cpu_possible(cpu)) {
93 printk(KERN_ERR "numa_set_node: invalid cpu# (%d)\n", cpu);
98 per_cpu(x86_cpu_to_node_map, cpu) = node;
100 set_cpu_numa_node(cpu, node);
103 void numa_clear_node(int cpu)
105 numa_set_node(cpu, NUMA_NO_NODE);
109 * Allocate node_to_cpumask_map based on number of available nodes
110 * Requires node_possible_map to be valid.
112 * Note: cpumask_of_node() is not valid until after this is done.
113 * (Use CONFIG_DEBUG_PER_CPU_MAPS to check this.)
115 void __init setup_node_to_cpumask_map(void)
119 /* setup nr_node_ids if not done yet */
120 if (nr_node_ids == MAX_NUMNODES)
123 /* allocate the map */
124 for (node = 0; node < nr_node_ids; node++)
125 alloc_bootmem_cpumask_var(&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);
131 static int __init numa_add_memblk_to(int nid, u64 start, u64 end,
132 struct numa_meminfo *mi)
134 /* ignore zero length blks */
138 /* whine about and ignore invalid blks */
139 if (start > end || nid < 0 || nid >= MAX_NUMNODES) {
140 pr_warning("NUMA: Warning: invalid memblk node %d [mem %#010Lx-%#010Lx]\n",
141 nid, start, end - 1);
145 if (mi->nr_blks >= NR_NODE_MEMBLKS) {
146 pr_err("NUMA: too many memblk ranges\n");
150 mi->blk[mi->nr_blks].start = start;
151 mi->blk[mi->nr_blks].end = end;
152 mi->blk[mi->nr_blks].nid = nid;
158 * numa_remove_memblk_from - Remove one numa_memblk from a numa_meminfo
159 * @idx: Index of memblk to remove
160 * @mi: numa_meminfo to remove memblk from
162 * Remove @idx'th numa_memblk from @mi by shifting @mi->blk[] and
163 * decrementing @mi->nr_blks.
165 void __init numa_remove_memblk_from(int idx, struct numa_meminfo *mi)
168 memmove(&mi->blk[idx], &mi->blk[idx + 1],
169 (mi->nr_blks - idx) * sizeof(mi->blk[0]));
173 * numa_add_memblk - Add one numa_memblk to numa_meminfo
174 * @nid: NUMA node ID of the new memblk
175 * @start: Start address of the new memblk
176 * @end: End address of the new memblk
178 * Add a new memblk to the default numa_meminfo.
181 * 0 on success, -errno on failure.
183 int __init numa_add_memblk(int nid, u64 start, u64 end)
185 return numa_add_memblk_to(nid, start, end, &numa_meminfo);
188 /* Initialize NODE_DATA for a node on the local memory */
189 static void __init setup_node_data(int nid, u64 start, u64 end)
191 const size_t nd_size = roundup(sizeof(pg_data_t), PAGE_SIZE);
197 * Don't confuse VM with a node that doesn't have the
198 * minimum amount of memory:
200 if (end && (end - start) < NODE_MIN_SIZE)
203 start = roundup(start, ZONE_ALIGN);
205 printk(KERN_INFO "Initmem setup node %d [mem %#010Lx-%#010Lx]\n",
206 nid, start, end - 1);
209 * Allocate node data. Try node-local memory and then any node.
210 * Never allocate in DMA zone.
212 nd_pa = memblock_alloc_nid(nd_size, SMP_CACHE_BYTES, nid);
214 nd_pa = __memblock_alloc_base(nd_size, SMP_CACHE_BYTES,
215 MEMBLOCK_ALLOC_ACCESSIBLE);
217 pr_err("Cannot find %zu bytes in node %d\n",
224 /* report and initialize */
225 printk(KERN_INFO " NODE_DATA [mem %#010Lx-%#010Lx]\n",
226 nd_pa, nd_pa + nd_size - 1);
227 tnid = early_pfn_to_nid(nd_pa >> PAGE_SHIFT);
229 printk(KERN_INFO " NODE_DATA(%d) on node %d\n", nid, tnid);
232 memset(NODE_DATA(nid), 0, sizeof(pg_data_t));
233 NODE_DATA(nid)->node_id = nid;
234 NODE_DATA(nid)->node_start_pfn = start >> PAGE_SHIFT;
235 NODE_DATA(nid)->node_spanned_pages = (end - start) >> PAGE_SHIFT;
237 node_set_online(nid);
241 * numa_cleanup_meminfo - Cleanup a numa_meminfo
242 * @mi: numa_meminfo to clean up
244 * Sanitize @mi by merging and removing unncessary memblks. Also check for
245 * conflicts and clear unused memblks.
248 * 0 on success, -errno on failure.
250 int __init numa_cleanup_meminfo(struct numa_meminfo *mi)
253 const u64 high = PFN_PHYS(max_pfn);
256 /* first, trim all entries */
257 for (i = 0; i < mi->nr_blks; i++) {
258 struct numa_memblk *bi = &mi->blk[i];
260 /* make sure all blocks are inside the limits */
261 bi->start = max(bi->start, low);
262 bi->end = min(bi->end, high);
264 /* and there's no empty block */
265 if (bi->start >= bi->end)
266 numa_remove_memblk_from(i--, mi);
269 /* merge neighboring / overlapping entries */
270 for (i = 0; i < mi->nr_blks; i++) {
271 struct numa_memblk *bi = &mi->blk[i];
273 for (j = i + 1; j < mi->nr_blks; j++) {
274 struct numa_memblk *bj = &mi->blk[j];
278 * See whether there are overlapping blocks. Whine
279 * about but allow overlaps of the same nid. They
280 * will be merged below.
282 if (bi->end > bj->start && bi->start < bj->end) {
283 if (bi->nid != bj->nid) {
284 pr_err("NUMA: node %d [mem %#010Lx-%#010Lx] overlaps with node %d [mem %#010Lx-%#010Lx]\n",
285 bi->nid, bi->start, bi->end - 1,
286 bj->nid, bj->start, bj->end - 1);
289 pr_warning("NUMA: Warning: node %d [mem %#010Lx-%#010Lx] overlaps with itself [mem %#010Lx-%#010Lx]\n",
290 bi->nid, bi->start, bi->end - 1,
291 bj->start, bj->end - 1);
295 * Join together blocks on the same node, holes
296 * between which don't overlap with memory on other
299 if (bi->nid != bj->nid)
301 start = min(bi->start, bj->start);
302 end = max(bi->end, bj->end);
303 for (k = 0; k < mi->nr_blks; k++) {
304 struct numa_memblk *bk = &mi->blk[k];
306 if (bi->nid == bk->nid)
308 if (start < bk->end && end > bk->start)
313 printk(KERN_INFO "NUMA: Node %d [mem %#010Lx-%#010Lx] + [mem %#010Lx-%#010Lx] -> [mem %#010Lx-%#010Lx]\n",
314 bi->nid, bi->start, bi->end - 1, bj->start,
315 bj->end - 1, start, end - 1);
318 numa_remove_memblk_from(j--, mi);
322 /* clear unused ones */
323 for (i = mi->nr_blks; i < ARRAY_SIZE(mi->blk); i++) {
324 mi->blk[i].start = mi->blk[i].end = 0;
325 mi->blk[i].nid = NUMA_NO_NODE;
332 * Set nodes, which have memory in @mi, in *@nodemask.
334 static void __init numa_nodemask_from_meminfo(nodemask_t *nodemask,
335 const struct numa_meminfo *mi)
339 for (i = 0; i < ARRAY_SIZE(mi->blk); i++)
340 if (mi->blk[i].start != mi->blk[i].end &&
341 mi->blk[i].nid != NUMA_NO_NODE)
342 node_set(mi->blk[i].nid, *nodemask);
346 * numa_reset_distance - Reset NUMA distance table
348 * The current table is freed. The next numa_set_distance() call will
351 void __init numa_reset_distance(void)
353 size_t size = numa_distance_cnt * numa_distance_cnt * sizeof(numa_distance[0]);
355 /* numa_distance could be 1LU marking allocation failure, test cnt */
356 if (numa_distance_cnt)
357 memblock_free(__pa(numa_distance), size);
358 numa_distance_cnt = 0;
359 numa_distance = NULL; /* enable table creation */
362 static int __init numa_alloc_distance(void)
364 nodemask_t nodes_parsed;
369 /* size the new table and allocate it */
370 nodes_parsed = numa_nodes_parsed;
371 numa_nodemask_from_meminfo(&nodes_parsed, &numa_meminfo);
373 for_each_node_mask(i, nodes_parsed)
376 size = cnt * cnt * sizeof(numa_distance[0]);
378 phys = memblock_find_in_range(0, PFN_PHYS(max_pfn_mapped),
381 pr_warning("NUMA: Warning: can't allocate distance table!\n");
382 /* don't retry until explicitly reset */
383 numa_distance = (void *)1LU;
386 memblock_reserve(phys, size);
388 numa_distance = __va(phys);
389 numa_distance_cnt = cnt;
391 /* fill with the default distances */
392 for (i = 0; i < cnt; i++)
393 for (j = 0; j < cnt; j++)
394 numa_distance[i * cnt + j] = i == j ?
395 LOCAL_DISTANCE : REMOTE_DISTANCE;
396 printk(KERN_DEBUG "NUMA: Initialized distance table, cnt=%d\n", cnt);
402 * numa_set_distance - Set NUMA distance from one NUMA to another
403 * @from: the 'from' node to set distance
404 * @to: the 'to' node to set distance
405 * @distance: NUMA distance
407 * Set the distance from node @from to @to to @distance. If distance table
408 * doesn't exist, one which is large enough to accommodate all the currently
409 * known nodes will be created.
411 * If such table cannot be allocated, a warning is printed and further
412 * calls are ignored until the distance table is reset with
413 * numa_reset_distance().
415 * If @from or @to is higher than the highest known node or lower than zero
416 * at the time of table creation or @distance doesn't make sense, the call
418 * This is to allow simplification of specific NUMA config implementations.
420 void __init numa_set_distance(int from, int to, int distance)
422 if (!numa_distance && numa_alloc_distance() < 0)
425 if (from >= numa_distance_cnt || to >= numa_distance_cnt ||
426 from < 0 || to < 0) {
427 pr_warn_once("NUMA: Warning: node ids are out of bound, from=%d to=%d distance=%d\n",
432 if ((u8)distance != distance ||
433 (from == to && distance != LOCAL_DISTANCE)) {
434 pr_warn_once("NUMA: Warning: invalid distance parameter, from=%d to=%d distance=%d\n",
439 numa_distance[from * numa_distance_cnt + to] = distance;
442 int __node_distance(int from, int to)
444 if (from >= numa_distance_cnt || to >= numa_distance_cnt)
445 return from == to ? LOCAL_DISTANCE : REMOTE_DISTANCE;
446 return numa_distance[from * numa_distance_cnt + to];
448 EXPORT_SYMBOL(__node_distance);
451 * Sanity check to catch more bad NUMA configurations (they are amazingly
452 * common). Make sure the nodes cover all memory.
454 static bool __init numa_meminfo_cover_memory(const struct numa_meminfo *mi)
456 u64 numaram, e820ram;
460 for (i = 0; i < mi->nr_blks; i++) {
461 u64 s = mi->blk[i].start >> PAGE_SHIFT;
462 u64 e = mi->blk[i].end >> PAGE_SHIFT;
464 numaram -= __absent_pages_in_range(mi->blk[i].nid, s, e);
465 if ((s64)numaram < 0)
469 e820ram = max_pfn - absent_pages_in_range(0, max_pfn);
471 /* We seem to lose 3 pages somewhere. Allow 1M of slack. */
472 if ((s64)(e820ram - numaram) >= (1 << (20 - PAGE_SHIFT))) {
473 printk(KERN_ERR "NUMA: nodes only cover %LuMB of your %LuMB e820 RAM. Not used.\n",
474 (numaram << PAGE_SHIFT) >> 20,
475 (e820ram << PAGE_SHIFT) >> 20);
481 static int __init numa_register_memblks(struct numa_meminfo *mi)
483 unsigned long uninitialized_var(pfn_align);
486 /* Account for nodes with cpus and no memory */
487 node_possible_map = numa_nodes_parsed;
488 numa_nodemask_from_meminfo(&node_possible_map, mi);
489 if (WARN_ON(nodes_empty(node_possible_map)))
492 for (i = 0; i < mi->nr_blks; i++) {
493 struct numa_memblk *mb = &mi->blk[i];
494 memblock_set_node(mb->start, mb->end - mb->start,
495 &memblock.memory, mb->nid);
499 * If sections array is gonna be used for pfn -> nid mapping, check
500 * whether its granularity is fine enough.
502 #ifdef NODE_NOT_IN_PAGE_FLAGS
503 pfn_align = node_map_pfn_alignment();
504 if (pfn_align && pfn_align < PAGES_PER_SECTION) {
505 printk(KERN_WARNING "Node alignment %LuMB < min %LuMB, rejecting NUMA config\n",
506 PFN_PHYS(pfn_align) >> 20,
507 PFN_PHYS(PAGES_PER_SECTION) >> 20);
511 if (!numa_meminfo_cover_memory(mi))
514 /* Finally register nodes. */
515 for_each_node_mask(nid, node_possible_map) {
516 u64 start = PFN_PHYS(max_pfn);
519 for (i = 0; i < mi->nr_blks; i++) {
520 if (nid != mi->blk[i].nid)
522 start = min(mi->blk[i].start, start);
523 end = max(mi->blk[i].end, end);
527 setup_node_data(nid, start, end);
530 /* Dump memblock with node info and return. */
536 * There are unfortunately some poorly designed mainboards around that
537 * only connect memory to a single CPU. This breaks the 1:1 cpu->node
538 * mapping. To avoid this fill in the mapping for all possible CPUs,
539 * as the number of CPUs is not known yet. We round robin the existing
542 static void __init numa_init_array(void)
546 rr = first_node(node_online_map);
547 for (i = 0; i < nr_cpu_ids; i++) {
548 if (early_cpu_to_node(i) != NUMA_NO_NODE)
550 numa_set_node(i, rr);
551 rr = next_node(rr, node_online_map);
552 if (rr == MAX_NUMNODES)
553 rr = first_node(node_online_map);
557 static int __init numa_init(int (*init_func)(void))
562 for (i = 0; i < MAX_LOCAL_APIC; i++)
563 set_apicid_to_node(i, NUMA_NO_NODE);
565 nodes_clear(numa_nodes_parsed);
566 nodes_clear(node_possible_map);
567 nodes_clear(node_online_map);
568 memset(&numa_meminfo, 0, sizeof(numa_meminfo));
569 WARN_ON(memblock_set_node(0, ULLONG_MAX, &memblock.memory,
571 numa_reset_distance();
578 * We reset memblock back to the top-down direction
579 * here because if we configured ACPI_NUMA, we have
580 * parsed SRAT in init_func(). It is ok to have the
581 * reset here even if we did't configure ACPI_NUMA
582 * or acpi numa init fails and fallbacks to dummy
585 memblock_set_bottom_up(false);
587 ret = numa_cleanup_meminfo(&numa_meminfo);
591 numa_emulation(&numa_meminfo, numa_distance_cnt);
593 ret = numa_register_memblks(&numa_meminfo);
597 for (i = 0; i < nr_cpu_ids; i++) {
598 int nid = early_cpu_to_node(i);
600 if (nid == NUMA_NO_NODE)
602 if (!node_online(nid))
610 * dummy_numa_init - Fallback dummy NUMA init
612 * Used if there's no underlying NUMA architecture, NUMA initialization
613 * fails, or NUMA is disabled on the command line.
615 * Must online at least one node and add memory blocks that cover all
616 * allowed memory. This function must not fail.
618 static int __init dummy_numa_init(void)
620 printk(KERN_INFO "%s\n",
621 numa_off ? "NUMA turned off" : "No NUMA configuration found");
622 printk(KERN_INFO "Faking a node at [mem %#018Lx-%#018Lx]\n",
623 0LLU, PFN_PHYS(max_pfn) - 1);
625 node_set(0, numa_nodes_parsed);
626 numa_add_memblk(0, 0, PFN_PHYS(max_pfn));
632 * x86_numa_init - Initialize NUMA
634 * Try each configured NUMA initialization method until one succeeds. The
635 * last fallback is dummy single node config encomapssing whole memory and
638 void __init x86_numa_init(void)
641 #ifdef CONFIG_X86_NUMAQ
642 if (!numa_init(numaq_numa_init))
645 #ifdef CONFIG_ACPI_NUMA
646 if (!numa_init(x86_acpi_numa_init))
649 #ifdef CONFIG_AMD_NUMA
650 if (!numa_init(amd_numa_init))
655 numa_init(dummy_numa_init);
658 static __init int find_near_online_node(int node)
661 int min_val = INT_MAX;
664 for_each_online_node(n) {
665 val = node_distance(node, n);
677 * Setup early cpu_to_node.
679 * Populate cpu_to_node[] only if x86_cpu_to_apicid[],
680 * and apicid_to_node[] tables have valid entries for a CPU.
681 * This means we skip cpu_to_node[] initialisation for NUMA
682 * emulation and faking node case (when running a kernel compiled
683 * for NUMA on a non NUMA box), which is OK as cpu_to_node[]
684 * is already initialized in a round robin manner at numa_init_array,
685 * prior to this call, and this initialization is good enough
686 * for the fake NUMA cases.
688 * Called before the per_cpu areas are setup.
690 void __init init_cpu_to_node(void)
693 u16 *cpu_to_apicid = early_per_cpu_ptr(x86_cpu_to_apicid);
695 BUG_ON(cpu_to_apicid == NULL);
697 for_each_possible_cpu(cpu) {
698 int node = numa_cpu_node(cpu);
700 if (node == NUMA_NO_NODE)
702 if (!node_online(node))
703 node = find_near_online_node(node);
704 numa_set_node(cpu, node);
708 #ifndef CONFIG_DEBUG_PER_CPU_MAPS
710 # ifndef CONFIG_NUMA_EMU
711 void numa_add_cpu(int cpu)
713 cpumask_set_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
716 void numa_remove_cpu(int cpu)
718 cpumask_clear_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
720 # endif /* !CONFIG_NUMA_EMU */
722 #else /* !CONFIG_DEBUG_PER_CPU_MAPS */
724 int __cpu_to_node(int cpu)
726 if (early_per_cpu_ptr(x86_cpu_to_node_map)) {
728 "cpu_to_node(%d): usage too early!\n", cpu);
730 return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
732 return per_cpu(x86_cpu_to_node_map, cpu);
734 EXPORT_SYMBOL(__cpu_to_node);
737 * Same function as cpu_to_node() but used if called before the
738 * per_cpu areas are setup.
740 int early_cpu_to_node(int cpu)
742 if (early_per_cpu_ptr(x86_cpu_to_node_map))
743 return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
745 if (!cpu_possible(cpu)) {
747 "early_cpu_to_node(%d): no per_cpu area!\n", cpu);
751 return per_cpu(x86_cpu_to_node_map, cpu);
754 void debug_cpumask_set_cpu(int cpu, int node, bool enable)
756 struct cpumask *mask;
759 if (node == NUMA_NO_NODE) {
760 /* early_cpu_to_node() already emits a warning and trace */
763 mask = node_to_cpumask_map[node];
765 pr_err("node_to_cpumask_map[%i] NULL\n", node);
771 cpumask_set_cpu(cpu, mask);
773 cpumask_clear_cpu(cpu, mask);
775 cpulist_scnprintf(buf, sizeof(buf), mask);
776 printk(KERN_DEBUG "%s cpu %d node %d: mask now %s\n",
777 enable ? "numa_add_cpu" : "numa_remove_cpu",
782 # ifndef CONFIG_NUMA_EMU
783 static void numa_set_cpumask(int cpu, bool enable)
785 debug_cpumask_set_cpu(cpu, early_cpu_to_node(cpu), enable);
788 void numa_add_cpu(int cpu)
790 numa_set_cpumask(cpu, true);
793 void numa_remove_cpu(int cpu)
795 numa_set_cpumask(cpu, false);
797 # endif /* !CONFIG_NUMA_EMU */
800 * Returns a pointer to the bitmask of CPUs on Node 'node'.
802 const struct cpumask *cpumask_of_node(int node)
804 if (node >= nr_node_ids) {
806 "cpumask_of_node(%d): node > nr_node_ids(%d)\n",
809 return cpu_none_mask;
811 if (node_to_cpumask_map[node] == NULL) {
813 "cpumask_of_node(%d): no node_to_cpumask_map!\n",
816 return cpu_online_mask;
818 return node_to_cpumask_map[node];
820 EXPORT_SYMBOL(cpumask_of_node);
822 #endif /* !CONFIG_DEBUG_PER_CPU_MAPS */
824 #ifdef CONFIG_MEMORY_HOTPLUG
825 int memory_add_physaddr_to_nid(u64 start)
827 struct numa_meminfo *mi = &numa_meminfo;
828 int nid = mi->blk[0].nid;
831 for (i = 0; i < mi->nr_blks; i++)
832 if (mi->blk[i].start <= start && mi->blk[i].end > start)
833 nid = mi->blk[i].nid;
836 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);