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] __cpuinitdata = {
60 [0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE
63 int __cpuinit 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 __cpuinit 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 if (node != NUMA_NO_NODE)
101 set_cpu_numa_node(cpu, node);
104 void __cpuinit numa_clear_node(int cpu)
106 numa_set_node(cpu, NUMA_NO_NODE);
110 * Allocate node_to_cpumask_map based on number of available nodes
111 * Requires node_possible_map to be valid.
113 * Note: node_to_cpumask() is not valid until after this is done.
114 * (Use CONFIG_DEBUG_PER_CPU_MAPS to check this.)
116 void __init setup_node_to_cpumask_map(void)
118 unsigned int node, num = 0;
120 /* setup nr_node_ids if not done yet */
121 if (nr_node_ids == MAX_NUMNODES) {
122 for_each_node_mask(node, node_possible_map)
124 nr_node_ids = num + 1;
127 /* allocate the map */
128 for (node = 0; node < nr_node_ids; node++)
129 alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]);
131 /* cpumask_of_node() will now work */
132 pr_debug("Node to cpumask map for %d nodes\n", nr_node_ids);
135 static int __init numa_add_memblk_to(int nid, u64 start, u64 end,
136 struct numa_meminfo *mi)
138 /* ignore zero length blks */
142 /* whine about and ignore invalid blks */
143 if (start > end || nid < 0 || nid >= MAX_NUMNODES) {
144 pr_warning("NUMA: Warning: invalid memblk node %d (%Lx-%Lx)\n",
149 if (mi->nr_blks >= NR_NODE_MEMBLKS) {
150 pr_err("NUMA: too many memblk ranges\n");
154 mi->blk[mi->nr_blks].start = start;
155 mi->blk[mi->nr_blks].end = end;
156 mi->blk[mi->nr_blks].nid = nid;
162 * numa_remove_memblk_from - Remove one numa_memblk from a numa_meminfo
163 * @idx: Index of memblk to remove
164 * @mi: numa_meminfo to remove memblk from
166 * Remove @idx'th numa_memblk from @mi by shifting @mi->blk[] and
167 * decrementing @mi->nr_blks.
169 void __init numa_remove_memblk_from(int idx, struct numa_meminfo *mi)
172 memmove(&mi->blk[idx], &mi->blk[idx + 1],
173 (mi->nr_blks - idx) * sizeof(mi->blk[0]));
177 * numa_add_memblk - Add one numa_memblk to numa_meminfo
178 * @nid: NUMA node ID of the new memblk
179 * @start: Start address of the new memblk
180 * @end: End address of the new memblk
182 * Add a new memblk to the default numa_meminfo.
185 * 0 on success, -errno on failure.
187 int __init numa_add_memblk(int nid, u64 start, u64 end)
189 return numa_add_memblk_to(nid, start, end, &numa_meminfo);
192 /* Initialize NODE_DATA for a node on the local memory */
193 static void __init setup_node_data(int nid, u64 start, u64 end)
195 const size_t nd_size = roundup(sizeof(pg_data_t), PAGE_SIZE);
196 bool remapped = false;
202 * Don't confuse VM with a node that doesn't have the
203 * minimum amount of memory:
205 if (end && (end - start) < NODE_MIN_SIZE)
208 /* initialize remap allocator before aligning to ZONE_ALIGN */
209 init_alloc_remap(nid, start, end);
211 start = roundup(start, ZONE_ALIGN);
213 printk(KERN_INFO "Initmem setup node %d %016Lx-%016Lx\n",
217 * Allocate node data. Try remap allocator first, node-local
218 * memory and then any node. Never allocate in DMA zone.
220 nd = alloc_remap(nid, nd_size);
225 nd_pa = memblock_alloc_nid(nd_size, SMP_CACHE_BYTES, nid);
227 pr_err("Cannot find %zu bytes in node %d\n",
234 /* report and initialize */
235 printk(KERN_INFO " NODE_DATA [%016Lx - %016Lx]%s\n",
236 nd_pa, nd_pa + nd_size - 1, remapped ? " (remapped)" : "");
237 tnid = early_pfn_to_nid(nd_pa >> PAGE_SHIFT);
238 if (!remapped && tnid != nid)
239 printk(KERN_INFO " NODE_DATA(%d) on node %d\n", nid, tnid);
242 memset(NODE_DATA(nid), 0, sizeof(pg_data_t));
243 NODE_DATA(nid)->node_id = nid;
244 NODE_DATA(nid)->node_start_pfn = start >> PAGE_SHIFT;
245 NODE_DATA(nid)->node_spanned_pages = (end - start) >> PAGE_SHIFT;
247 node_set_online(nid);
251 * numa_cleanup_meminfo - Cleanup a numa_meminfo
252 * @mi: numa_meminfo to clean up
254 * Sanitize @mi by merging and removing unncessary memblks. Also check for
255 * conflicts and clear unused memblks.
258 * 0 on success, -errno on failure.
260 int __init numa_cleanup_meminfo(struct numa_meminfo *mi)
263 const u64 high = PFN_PHYS(max_pfn);
266 /* first, trim all entries */
267 for (i = 0; i < mi->nr_blks; i++) {
268 struct numa_memblk *bi = &mi->blk[i];
270 /* make sure all blocks are inside the limits */
271 bi->start = max(bi->start, low);
272 bi->end = min(bi->end, high);
274 /* and there's no empty block */
275 if (bi->start >= bi->end)
276 numa_remove_memblk_from(i--, mi);
279 /* merge neighboring / overlapping entries */
280 for (i = 0; i < mi->nr_blks; i++) {
281 struct numa_memblk *bi = &mi->blk[i];
283 for (j = i + 1; j < mi->nr_blks; j++) {
284 struct numa_memblk *bj = &mi->blk[j];
288 * See whether there are overlapping blocks. Whine
289 * about but allow overlaps of the same nid. They
290 * will be merged below.
292 if (bi->end > bj->start && bi->start < bj->end) {
293 if (bi->nid != bj->nid) {
294 pr_err("NUMA: node %d (%Lx-%Lx) overlaps with node %d (%Lx-%Lx)\n",
295 bi->nid, bi->start, bi->end,
296 bj->nid, bj->start, bj->end);
299 pr_warning("NUMA: Warning: node %d (%Lx-%Lx) overlaps with itself (%Lx-%Lx)\n",
300 bi->nid, bi->start, bi->end,
305 * Join together blocks on the same node, holes
306 * between which don't overlap with memory on other
309 if (bi->nid != bj->nid)
311 start = min(bi->start, bj->start);
312 end = max(bi->end, bj->end);
313 for (k = 0; k < mi->nr_blks; k++) {
314 struct numa_memblk *bk = &mi->blk[k];
316 if (bi->nid == bk->nid)
318 if (start < bk->end && end > bk->start)
323 printk(KERN_INFO "NUMA: Node %d [%Lx,%Lx) + [%Lx,%Lx) -> [%Lx,%Lx)\n",
324 bi->nid, bi->start, bi->end, bj->start, bj->end,
328 numa_remove_memblk_from(j--, mi);
332 /* clear unused ones */
333 for (i = mi->nr_blks; i < ARRAY_SIZE(mi->blk); i++) {
334 mi->blk[i].start = mi->blk[i].end = 0;
335 mi->blk[i].nid = NUMA_NO_NODE;
342 * Set nodes, which have memory in @mi, in *@nodemask.
344 static void __init numa_nodemask_from_meminfo(nodemask_t *nodemask,
345 const struct numa_meminfo *mi)
349 for (i = 0; i < ARRAY_SIZE(mi->blk); i++)
350 if (mi->blk[i].start != mi->blk[i].end &&
351 mi->blk[i].nid != NUMA_NO_NODE)
352 node_set(mi->blk[i].nid, *nodemask);
356 * numa_reset_distance - Reset NUMA distance table
358 * The current table is freed. The next numa_set_distance() call will
361 void __init numa_reset_distance(void)
363 size_t size = numa_distance_cnt * numa_distance_cnt * sizeof(numa_distance[0]);
365 /* numa_distance could be 1LU marking allocation failure, test cnt */
366 if (numa_distance_cnt)
367 memblock_x86_free_range(__pa(numa_distance),
368 __pa(numa_distance) + size);
369 numa_distance_cnt = 0;
370 numa_distance = NULL; /* enable table creation */
373 static int __init numa_alloc_distance(void)
375 nodemask_t nodes_parsed;
380 /* size the new table and allocate it */
381 nodes_parsed = numa_nodes_parsed;
382 numa_nodemask_from_meminfo(&nodes_parsed, &numa_meminfo);
384 for_each_node_mask(i, nodes_parsed)
387 size = cnt * cnt * sizeof(numa_distance[0]);
389 phys = memblock_find_in_range(0, PFN_PHYS(max_pfn_mapped),
392 pr_warning("NUMA: Warning: can't allocate distance table!\n");
393 /* don't retry until explicitly reset */
394 numa_distance = (void *)1LU;
397 memblock_x86_reserve_range(phys, phys + size, "NUMA DIST");
399 numa_distance = __va(phys);
400 numa_distance_cnt = cnt;
402 /* fill with the default distances */
403 for (i = 0; i < cnt; i++)
404 for (j = 0; j < cnt; j++)
405 numa_distance[i * cnt + j] = i == j ?
406 LOCAL_DISTANCE : REMOTE_DISTANCE;
407 printk(KERN_DEBUG "NUMA: Initialized distance table, cnt=%d\n", cnt);
413 * numa_set_distance - Set NUMA distance from one NUMA to another
414 * @from: the 'from' node to set distance
415 * @to: the 'to' node to set distance
416 * @distance: NUMA distance
418 * Set the distance from node @from to @to to @distance. If distance table
419 * doesn't exist, one which is large enough to accommodate all the currently
420 * known nodes will be created.
422 * If such table cannot be allocated, a warning is printed and further
423 * calls are ignored until the distance table is reset with
424 * numa_reset_distance().
426 * If @from or @to is higher than the highest known node at the time of
427 * table creation or @distance doesn't make sense, the call is ignored.
428 * This is to allow simplification of specific NUMA config implementations.
430 void __init numa_set_distance(int from, int to, int distance)
432 if (!numa_distance && numa_alloc_distance() < 0)
435 if (from >= numa_distance_cnt || to >= numa_distance_cnt) {
436 printk_once(KERN_DEBUG "NUMA: Debug: distance out of bound, from=%d to=%d distance=%d\n",
441 if ((u8)distance != distance ||
442 (from == to && distance != LOCAL_DISTANCE)) {
443 pr_warn_once("NUMA: Warning: invalid distance parameter, from=%d to=%d distance=%d\n",
448 numa_distance[from * numa_distance_cnt + to] = distance;
451 int __node_distance(int from, int to)
453 if (from >= numa_distance_cnt || to >= numa_distance_cnt)
454 return from == to ? LOCAL_DISTANCE : REMOTE_DISTANCE;
455 return numa_distance[from * numa_distance_cnt + to];
457 EXPORT_SYMBOL(__node_distance);
460 * Sanity check to catch more bad NUMA configurations (they are amazingly
461 * common). Make sure the nodes cover all memory.
463 static bool __init numa_meminfo_cover_memory(const struct numa_meminfo *mi)
465 u64 numaram, e820ram;
469 for (i = 0; i < mi->nr_blks; i++) {
470 u64 s = mi->blk[i].start >> PAGE_SHIFT;
471 u64 e = mi->blk[i].end >> PAGE_SHIFT;
473 numaram -= __absent_pages_in_range(mi->blk[i].nid, s, e);
474 if ((s64)numaram < 0)
478 e820ram = max_pfn - (memblock_x86_hole_size(0,
479 PFN_PHYS(max_pfn)) >> PAGE_SHIFT);
480 /* We seem to lose 3 pages somewhere. Allow 1M of slack. */
481 if ((s64)(e820ram - numaram) >= (1 << (20 - PAGE_SHIFT))) {
482 printk(KERN_ERR "NUMA: nodes only cover %LuMB of your %LuMB e820 RAM. Not used.\n",
483 (numaram << PAGE_SHIFT) >> 20,
484 (e820ram << PAGE_SHIFT) >> 20);
490 static int __init numa_register_memblks(struct numa_meminfo *mi)
492 unsigned long uninitialized_var(pfn_align);
495 /* Account for nodes with cpus and no memory */
496 node_possible_map = numa_nodes_parsed;
497 numa_nodemask_from_meminfo(&node_possible_map, mi);
498 if (WARN_ON(nodes_empty(node_possible_map)))
501 for (i = 0; i < mi->nr_blks; i++) {
502 struct numa_memblk *mb = &mi->blk[i];
503 memblock_set_node(mb->start, mb->end - mb->start, mb->nid);
507 * If sections array is gonna be used for pfn -> nid mapping, check
508 * whether its granularity is fine enough.
510 #ifdef NODE_NOT_IN_PAGE_FLAGS
511 pfn_align = node_map_pfn_alignment();
512 if (pfn_align && pfn_align < PAGES_PER_SECTION) {
513 printk(KERN_WARNING "Node alignment %LuMB < min %LuMB, rejecting NUMA config\n",
514 PFN_PHYS(pfn_align) >> 20,
515 PFN_PHYS(PAGES_PER_SECTION) >> 20);
519 if (!numa_meminfo_cover_memory(mi))
522 /* Finally register nodes. */
523 for_each_node_mask(nid, node_possible_map) {
524 u64 start = PFN_PHYS(max_pfn);
527 for (i = 0; i < mi->nr_blks; i++) {
528 if (nid != mi->blk[i].nid)
530 start = min(mi->blk[i].start, start);
531 end = max(mi->blk[i].end, end);
535 setup_node_data(nid, start, end);
538 /* Dump memblock with node info and return. */
544 * There are unfortunately some poorly designed mainboards around that
545 * only connect memory to a single CPU. This breaks the 1:1 cpu->node
546 * mapping. To avoid this fill in the mapping for all possible CPUs,
547 * as the number of CPUs is not known yet. We round robin the existing
550 static void __init numa_init_array(void)
554 rr = first_node(node_online_map);
555 for (i = 0; i < nr_cpu_ids; i++) {
556 if (early_cpu_to_node(i) != NUMA_NO_NODE)
558 numa_set_node(i, rr);
559 rr = next_node(rr, node_online_map);
560 if (rr == MAX_NUMNODES)
561 rr = first_node(node_online_map);
565 static int __init numa_init(int (*init_func)(void))
570 for (i = 0; i < MAX_LOCAL_APIC; i++)
571 set_apicid_to_node(i, NUMA_NO_NODE);
573 nodes_clear(numa_nodes_parsed);
574 nodes_clear(node_possible_map);
575 nodes_clear(node_online_map);
576 memset(&numa_meminfo, 0, sizeof(numa_meminfo));
577 WARN_ON(memblock_set_node(0, ULLONG_MAX, MAX_NUMNODES));
578 numa_reset_distance();
583 ret = numa_cleanup_meminfo(&numa_meminfo);
587 numa_emulation(&numa_meminfo, numa_distance_cnt);
589 ret = numa_register_memblks(&numa_meminfo);
593 for (i = 0; i < nr_cpu_ids; i++) {
594 int nid = early_cpu_to_node(i);
596 if (nid == NUMA_NO_NODE)
598 if (!node_online(nid))
606 * dummy_numa_init - Fallback dummy NUMA init
608 * Used if there's no underlying NUMA architecture, NUMA initialization
609 * fails, or NUMA is disabled on the command line.
611 * Must online at least one node and add memory blocks that cover all
612 * allowed memory. This function must not fail.
614 static int __init dummy_numa_init(void)
616 printk(KERN_INFO "%s\n",
617 numa_off ? "NUMA turned off" : "No NUMA configuration found");
618 printk(KERN_INFO "Faking a node at %016Lx-%016Lx\n",
619 0LLU, PFN_PHYS(max_pfn));
621 node_set(0, numa_nodes_parsed);
622 numa_add_memblk(0, 0, PFN_PHYS(max_pfn));
628 * x86_numa_init - Initialize NUMA
630 * Try each configured NUMA initialization method until one succeeds. The
631 * last fallback is dummy single node config encomapssing whole memory and
634 void __init x86_numa_init(void)
637 #ifdef CONFIG_X86_NUMAQ
638 if (!numa_init(numaq_numa_init))
641 #ifdef CONFIG_ACPI_NUMA
642 if (!numa_init(x86_acpi_numa_init))
645 #ifdef CONFIG_AMD_NUMA
646 if (!numa_init(amd_numa_init))
651 numa_init(dummy_numa_init);
654 static __init int find_near_online_node(int node)
657 int min_val = INT_MAX;
660 for_each_online_node(n) {
661 val = node_distance(node, n);
673 * Setup early cpu_to_node.
675 * Populate cpu_to_node[] only if x86_cpu_to_apicid[],
676 * and apicid_to_node[] tables have valid entries for a CPU.
677 * This means we skip cpu_to_node[] initialisation for NUMA
678 * emulation and faking node case (when running a kernel compiled
679 * for NUMA on a non NUMA box), which is OK as cpu_to_node[]
680 * is already initialized in a round robin manner at numa_init_array,
681 * prior to this call, and this initialization is good enough
682 * for the fake NUMA cases.
684 * Called before the per_cpu areas are setup.
686 void __init init_cpu_to_node(void)
689 u16 *cpu_to_apicid = early_per_cpu_ptr(x86_cpu_to_apicid);
691 BUG_ON(cpu_to_apicid == NULL);
693 for_each_possible_cpu(cpu) {
694 int node = numa_cpu_node(cpu);
696 if (node == NUMA_NO_NODE)
698 if (!node_online(node))
699 node = find_near_online_node(node);
700 numa_set_node(cpu, node);
704 #ifndef CONFIG_DEBUG_PER_CPU_MAPS
706 # ifndef CONFIG_NUMA_EMU
707 void __cpuinit numa_add_cpu(int cpu)
709 cpumask_set_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
712 void __cpuinit numa_remove_cpu(int cpu)
714 cpumask_clear_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
716 # endif /* !CONFIG_NUMA_EMU */
718 #else /* !CONFIG_DEBUG_PER_CPU_MAPS */
720 int __cpu_to_node(int cpu)
722 if (early_per_cpu_ptr(x86_cpu_to_node_map)) {
724 "cpu_to_node(%d): usage too early!\n", cpu);
726 return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
728 return per_cpu(x86_cpu_to_node_map, cpu);
730 EXPORT_SYMBOL(__cpu_to_node);
733 * Same function as cpu_to_node() but used if called before the
734 * per_cpu areas are setup.
736 int early_cpu_to_node(int cpu)
738 if (early_per_cpu_ptr(x86_cpu_to_node_map))
739 return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
741 if (!cpu_possible(cpu)) {
743 "early_cpu_to_node(%d): no per_cpu area!\n", cpu);
747 return per_cpu(x86_cpu_to_node_map, cpu);
750 void debug_cpumask_set_cpu(int cpu, int node, bool enable)
752 struct cpumask *mask;
755 if (node == NUMA_NO_NODE) {
756 /* early_cpu_to_node() already emits a warning and trace */
759 mask = node_to_cpumask_map[node];
761 pr_err("node_to_cpumask_map[%i] NULL\n", node);
767 cpumask_set_cpu(cpu, mask);
769 cpumask_clear_cpu(cpu, mask);
771 cpulist_scnprintf(buf, sizeof(buf), mask);
772 printk(KERN_DEBUG "%s cpu %d node %d: mask now %s\n",
773 enable ? "numa_add_cpu" : "numa_remove_cpu",
778 # ifndef CONFIG_NUMA_EMU
779 static void __cpuinit numa_set_cpumask(int cpu, bool enable)
781 debug_cpumask_set_cpu(cpu, early_cpu_to_node(cpu), enable);
784 void __cpuinit numa_add_cpu(int cpu)
786 numa_set_cpumask(cpu, true);
789 void __cpuinit numa_remove_cpu(int cpu)
791 numa_set_cpumask(cpu, false);
793 # endif /* !CONFIG_NUMA_EMU */
796 * Returns a pointer to the bitmask of CPUs on Node 'node'.
798 const struct cpumask *cpumask_of_node(int node)
800 if (node >= nr_node_ids) {
802 "cpumask_of_node(%d): node > nr_node_ids(%d)\n",
805 return cpu_none_mask;
807 if (node_to_cpumask_map[node] == NULL) {
809 "cpumask_of_node(%d): no node_to_cpumask_map!\n",
812 return cpu_online_mask;
814 return node_to_cpumask_map[node];
816 EXPORT_SYMBOL(cpumask_of_node);
818 #endif /* !CONFIG_DEBUG_PER_CPU_MAPS */
820 #ifdef CONFIG_MEMORY_HOTPLUG
821 int memory_add_physaddr_to_nid(u64 start)
823 struct numa_meminfo *mi = &numa_meminfo;
824 int nid = mi->blk[0].nid;
827 for (i = 0; i < mi->nr_blks; i++)
828 if (mi->blk[i].start <= start && mi->blk[i].end > start)
829 nid = mi->blk[i].nid;
832 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);