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 bootmem allocator for a node */
193 static void __init setup_node_bootmem(int nid, u64 start, u64 end)
195 const u64 nd_low = PFN_PHYS(MAX_DMA_PFN);
196 const u64 nd_high = PFN_PHYS(max_pfn_mapped);
197 const size_t nd_size = roundup(sizeof(pg_data_t), PAGE_SIZE);
198 bool remapped = false;
204 * Don't confuse VM with a node that doesn't have the
205 * minimum amount of memory:
207 if (end && (end - start) < NODE_MIN_SIZE)
210 /* initialize remap allocator before aligning to ZONE_ALIGN */
211 init_alloc_remap(nid, start, end);
213 start = roundup(start, ZONE_ALIGN);
215 printk(KERN_INFO "Initmem setup node %d %016Lx-%016Lx\n",
219 * Allocate node data. Try remap allocator first, node-local
220 * memory and then any node. Never allocate in DMA zone.
222 nd = alloc_remap(nid, nd_size);
227 nd_pa = memblock_x86_find_in_range_node(nid, nd_low, nd_high,
228 nd_size, SMP_CACHE_BYTES);
229 if (nd_pa == MEMBLOCK_ERROR)
230 nd_pa = memblock_find_in_range(nd_low, nd_high,
231 nd_size, SMP_CACHE_BYTES);
232 if (nd_pa == MEMBLOCK_ERROR) {
233 pr_err("Cannot find %zu bytes in node %d\n",
237 memblock_x86_reserve_range(nd_pa, nd_pa + nd_size, "NODE_DATA");
241 /* report and initialize */
242 printk(KERN_INFO " NODE_DATA [%016Lx - %016Lx]%s\n",
243 nd_pa, nd_pa + nd_size - 1, remapped ? " (remapped)" : "");
244 tnid = early_pfn_to_nid(nd_pa >> PAGE_SHIFT);
245 if (!remapped && tnid != nid)
246 printk(KERN_INFO " NODE_DATA(%d) on node %d\n", nid, tnid);
249 memset(NODE_DATA(nid), 0, sizeof(pg_data_t));
250 NODE_DATA(nid)->node_id = nid;
251 NODE_DATA(nid)->node_start_pfn = start >> PAGE_SHIFT;
252 NODE_DATA(nid)->node_spanned_pages = (end - start) >> PAGE_SHIFT;
254 node_set_online(nid);
258 * numa_cleanup_meminfo - Cleanup a numa_meminfo
259 * @mi: numa_meminfo to clean up
261 * Sanitize @mi by merging and removing unncessary memblks. Also check for
262 * conflicts and clear unused memblks.
265 * 0 on success, -errno on failure.
267 int __init numa_cleanup_meminfo(struct numa_meminfo *mi)
270 const u64 high = PFN_PHYS(max_pfn);
273 for (i = 0; i < mi->nr_blks; i++) {
274 struct numa_memblk *bi = &mi->blk[i];
276 /* make sure all blocks are inside the limits */
277 bi->start = max(bi->start, low);
278 bi->end = min(bi->end, high);
280 /* and there's no empty block */
281 if (bi->start >= bi->end) {
282 numa_remove_memblk_from(i--, mi);
286 for (j = i + 1; j < mi->nr_blks; j++) {
287 struct numa_memblk *bj = &mi->blk[j];
291 * See whether there are overlapping blocks. Whine
292 * about but allow overlaps of the same nid. They
293 * will be merged below.
295 if (bi->end > bj->start && bi->start < bj->end) {
296 if (bi->nid != bj->nid) {
297 pr_err("NUMA: node %d (%Lx-%Lx) overlaps with node %d (%Lx-%Lx)\n",
298 bi->nid, bi->start, bi->end,
299 bj->nid, bj->start, bj->end);
302 pr_warning("NUMA: Warning: node %d (%Lx-%Lx) overlaps with itself (%Lx-%Lx)\n",
303 bi->nid, bi->start, bi->end,
308 * Join together blocks on the same node, holes
309 * between which don't overlap with memory on other
312 if (bi->nid != bj->nid)
314 start = max(min(bi->start, bj->start), low);
315 end = min(max(bi->end, bj->end), high);
316 for (k = 0; k < mi->nr_blks; k++) {
317 struct numa_memblk *bk = &mi->blk[k];
319 if (bi->nid == bk->nid)
321 if (start < bk->end && end > bk->start)
326 printk(KERN_INFO "NUMA: Node %d [%Lx,%Lx) + [%Lx,%Lx) -> [%Lx,%Lx)\n",
327 bi->nid, bi->start, bi->end, bj->start, bj->end,
331 numa_remove_memblk_from(j--, mi);
335 for (i = mi->nr_blks; i < ARRAY_SIZE(mi->blk); i++) {
336 mi->blk[i].start = mi->blk[i].end = 0;
337 mi->blk[i].nid = NUMA_NO_NODE;
344 * Set nodes, which have memory in @mi, in *@nodemask.
346 static void __init numa_nodemask_from_meminfo(nodemask_t *nodemask,
347 const struct numa_meminfo *mi)
351 for (i = 0; i < ARRAY_SIZE(mi->blk); i++)
352 if (mi->blk[i].start != mi->blk[i].end &&
353 mi->blk[i].nid != NUMA_NO_NODE)
354 node_set(mi->blk[i].nid, *nodemask);
358 * numa_reset_distance - Reset NUMA distance table
360 * The current table is freed. The next numa_set_distance() call will
363 void __init numa_reset_distance(void)
365 size_t size = numa_distance_cnt * numa_distance_cnt * sizeof(numa_distance[0]);
367 /* numa_distance could be 1LU marking allocation failure, test cnt */
368 if (numa_distance_cnt)
369 memblock_x86_free_range(__pa(numa_distance),
370 __pa(numa_distance) + size);
371 numa_distance_cnt = 0;
372 numa_distance = NULL; /* enable table creation */
375 static int __init numa_alloc_distance(void)
377 nodemask_t nodes_parsed;
382 /* size the new table and allocate it */
383 nodes_parsed = numa_nodes_parsed;
384 numa_nodemask_from_meminfo(&nodes_parsed, &numa_meminfo);
386 for_each_node_mask(i, nodes_parsed)
389 size = cnt * cnt * sizeof(numa_distance[0]);
391 phys = memblock_find_in_range(0, PFN_PHYS(max_pfn_mapped),
393 if (phys == MEMBLOCK_ERROR) {
394 pr_warning("NUMA: Warning: can't allocate distance table!\n");
395 /* don't retry until explicitly reset */
396 numa_distance = (void *)1LU;
399 memblock_x86_reserve_range(phys, phys + size, "NUMA DIST");
401 numa_distance = __va(phys);
402 numa_distance_cnt = cnt;
404 /* fill with the default distances */
405 for (i = 0; i < cnt; i++)
406 for (j = 0; j < cnt; j++)
407 numa_distance[i * cnt + j] = i == j ?
408 LOCAL_DISTANCE : REMOTE_DISTANCE;
409 printk(KERN_DEBUG "NUMA: Initialized distance table, cnt=%d\n", cnt);
415 * numa_set_distance - Set NUMA distance from one NUMA to another
416 * @from: the 'from' node to set distance
417 * @to: the 'to' node to set distance
418 * @distance: NUMA distance
420 * Set the distance from node @from to @to to @distance. If distance table
421 * doesn't exist, one which is large enough to accommodate all the currently
422 * known nodes will be created.
424 * If such table cannot be allocated, a warning is printed and further
425 * calls are ignored until the distance table is reset with
426 * numa_reset_distance().
428 * If @from or @to is higher than the highest known node at the time of
429 * table creation or @distance doesn't make sense, the call is ignored.
430 * This is to allow simplification of specific NUMA config implementations.
432 void __init numa_set_distance(int from, int to, int distance)
434 if (!numa_distance && numa_alloc_distance() < 0)
437 if (from >= numa_distance_cnt || to >= numa_distance_cnt) {
438 printk_once(KERN_DEBUG "NUMA: Debug: distance out of bound, from=%d to=%d distance=%d\n",
443 if ((u8)distance != distance ||
444 (from == to && distance != LOCAL_DISTANCE)) {
445 pr_warn_once("NUMA: Warning: invalid distance parameter, from=%d to=%d distance=%d\n",
450 numa_distance[from * numa_distance_cnt + to] = distance;
453 int __node_distance(int from, int to)
455 if (from >= numa_distance_cnt || to >= numa_distance_cnt)
456 return from == to ? LOCAL_DISTANCE : REMOTE_DISTANCE;
457 return numa_distance[from * numa_distance_cnt + to];
459 EXPORT_SYMBOL(__node_distance);
462 * Sanity check to catch more bad NUMA configurations (they are amazingly
463 * common). Make sure the nodes cover all memory.
465 static bool __init numa_meminfo_cover_memory(const struct numa_meminfo *mi)
467 u64 numaram, e820ram;
471 for (i = 0; i < mi->nr_blks; i++) {
472 u64 s = mi->blk[i].start >> PAGE_SHIFT;
473 u64 e = mi->blk[i].end >> PAGE_SHIFT;
475 numaram -= __absent_pages_in_range(mi->blk[i].nid, s, e);
476 if ((s64)numaram < 0)
480 e820ram = max_pfn - (memblock_x86_hole_size(0,
481 PFN_PHYS(max_pfn)) >> PAGE_SHIFT);
482 /* We seem to lose 3 pages somewhere. Allow 1M of slack. */
483 if ((s64)(e820ram - numaram) >= (1 << (20 - PAGE_SHIFT))) {
484 printk(KERN_ERR "NUMA: nodes only cover %LuMB of your %LuMB e820 RAM. Not used.\n",
485 (numaram << PAGE_SHIFT) >> 20,
486 (e820ram << PAGE_SHIFT) >> 20);
492 static int __init numa_register_memblks(struct numa_meminfo *mi)
496 /* Account for nodes with cpus and no memory */
497 node_possible_map = numa_nodes_parsed;
498 numa_nodemask_from_meminfo(&node_possible_map, mi);
499 if (WARN_ON(nodes_empty(node_possible_map)))
502 for (i = 0; i < mi->nr_blks; i++)
503 memblock_x86_register_active_regions(mi->blk[i].nid,
504 mi->blk[i].start >> PAGE_SHIFT,
505 mi->blk[i].end >> PAGE_SHIFT);
507 /* for out of order entries */
509 if (!numa_meminfo_cover_memory(mi))
512 /* Finally register nodes. */
513 for_each_node_mask(nid, node_possible_map) {
514 u64 start = PFN_PHYS(max_pfn);
517 for (i = 0; i < mi->nr_blks; i++) {
518 if (nid != mi->blk[i].nid)
520 start = min(mi->blk[i].start, start);
521 end = max(mi->blk[i].end, end);
525 setup_node_bootmem(nid, start, end);
532 * There are unfortunately some poorly designed mainboards around that
533 * only connect memory to a single CPU. This breaks the 1:1 cpu->node
534 * mapping. To avoid this fill in the mapping for all possible CPUs,
535 * as the number of CPUs is not known yet. We round robin the existing
538 void __init numa_init_array(void)
542 rr = first_node(node_online_map);
543 for (i = 0; i < nr_cpu_ids; i++) {
544 if (early_cpu_to_node(i) != NUMA_NO_NODE)
546 numa_set_node(i, rr);
547 rr = next_node(rr, node_online_map);
548 if (rr == MAX_NUMNODES)
549 rr = first_node(node_online_map);
553 static int __init numa_init(int (*init_func)(void))
558 for (i = 0; i < MAX_LOCAL_APIC; i++)
559 set_apicid_to_node(i, NUMA_NO_NODE);
561 nodes_clear(numa_nodes_parsed);
562 nodes_clear(node_possible_map);
563 nodes_clear(node_online_map);
564 memset(&numa_meminfo, 0, sizeof(numa_meminfo));
565 remove_all_active_ranges();
566 numa_reset_distance();
571 ret = numa_cleanup_meminfo(&numa_meminfo);
575 numa_emulation(&numa_meminfo, numa_distance_cnt);
577 ret = numa_register_memblks(&numa_meminfo);
581 for (i = 0; i < nr_cpu_ids; i++) {
582 int nid = early_cpu_to_node(i);
584 if (nid == NUMA_NO_NODE)
586 if (!node_online(nid))
594 * dummy_numa_init - Fallback dummy NUMA init
596 * Used if there's no underlying NUMA architecture, NUMA initialization
597 * fails, or NUMA is disabled on the command line.
599 * Must online at least one node and add memory blocks that cover all
600 * allowed memory. This function must not fail.
602 static int __init dummy_numa_init(void)
604 printk(KERN_INFO "%s\n",
605 numa_off ? "NUMA turned off" : "No NUMA configuration found");
606 printk(KERN_INFO "Faking a node at %016Lx-%016Lx\n",
607 0LLU, PFN_PHYS(max_pfn));
609 node_set(0, numa_nodes_parsed);
610 numa_add_memblk(0, 0, PFN_PHYS(max_pfn));
616 * x86_numa_init - Initialize NUMA
618 * Try each configured NUMA initialization method until one succeeds. The
619 * last fallback is dummy single node config encomapssing whole memory and
622 void __init x86_numa_init(void)
625 #ifdef CONFIG_X86_NUMAQ
626 if (!numa_init(numaq_numa_init))
629 #ifdef CONFIG_ACPI_NUMA
630 if (!numa_init(x86_acpi_numa_init))
633 #ifdef CONFIG_AMD_NUMA
634 if (!numa_init(amd_numa_init))
639 numa_init(dummy_numa_init);
642 static __init int find_near_online_node(int node)
645 int min_val = INT_MAX;
648 for_each_online_node(n) {
649 val = node_distance(node, n);
661 * Setup early cpu_to_node.
663 * Populate cpu_to_node[] only if x86_cpu_to_apicid[],
664 * and apicid_to_node[] tables have valid entries for a CPU.
665 * This means we skip cpu_to_node[] initialisation for NUMA
666 * emulation and faking node case (when running a kernel compiled
667 * for NUMA on a non NUMA box), which is OK as cpu_to_node[]
668 * is already initialized in a round robin manner at numa_init_array,
669 * prior to this call, and this initialization is good enough
670 * for the fake NUMA cases.
672 * Called before the per_cpu areas are setup.
674 void __init init_cpu_to_node(void)
677 u16 *cpu_to_apicid = early_per_cpu_ptr(x86_cpu_to_apicid);
679 BUG_ON(cpu_to_apicid == NULL);
681 for_each_possible_cpu(cpu) {
682 int node = numa_cpu_node(cpu);
684 if (node == NUMA_NO_NODE)
686 if (!node_online(node))
687 node = find_near_online_node(node);
688 numa_set_node(cpu, node);
692 #ifndef CONFIG_DEBUG_PER_CPU_MAPS
694 # ifndef CONFIG_NUMA_EMU
695 void __cpuinit numa_add_cpu(int cpu)
697 cpumask_set_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
700 void __cpuinit numa_remove_cpu(int cpu)
702 cpumask_clear_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
704 # endif /* !CONFIG_NUMA_EMU */
706 #else /* !CONFIG_DEBUG_PER_CPU_MAPS */
708 int __cpu_to_node(int cpu)
710 if (early_per_cpu_ptr(x86_cpu_to_node_map)) {
712 "cpu_to_node(%d): usage too early!\n", cpu);
714 return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
716 return per_cpu(x86_cpu_to_node_map, cpu);
718 EXPORT_SYMBOL(__cpu_to_node);
721 * Same function as cpu_to_node() but used if called before the
722 * per_cpu areas are setup.
724 int early_cpu_to_node(int cpu)
726 if (early_per_cpu_ptr(x86_cpu_to_node_map))
727 return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
729 if (!cpu_possible(cpu)) {
731 "early_cpu_to_node(%d): no per_cpu area!\n", cpu);
735 return per_cpu(x86_cpu_to_node_map, cpu);
738 void debug_cpumask_set_cpu(int cpu, int node, bool enable)
740 struct cpumask *mask;
743 if (node == NUMA_NO_NODE) {
744 /* early_cpu_to_node() already emits a warning and trace */
747 mask = node_to_cpumask_map[node];
749 pr_err("node_to_cpumask_map[%i] NULL\n", node);
755 cpumask_set_cpu(cpu, mask);
757 cpumask_clear_cpu(cpu, mask);
759 cpulist_scnprintf(buf, sizeof(buf), mask);
760 printk(KERN_DEBUG "%s cpu %d node %d: mask now %s\n",
761 enable ? "numa_add_cpu" : "numa_remove_cpu",
766 # ifndef CONFIG_NUMA_EMU
767 static void __cpuinit numa_set_cpumask(int cpu, bool enable)
769 debug_cpumask_set_cpu(cpu, early_cpu_to_node(cpu), enable);
772 void __cpuinit numa_add_cpu(int cpu)
774 numa_set_cpumask(cpu, true);
777 void __cpuinit numa_remove_cpu(int cpu)
779 numa_set_cpumask(cpu, false);
781 # endif /* !CONFIG_NUMA_EMU */
784 * Returns a pointer to the bitmask of CPUs on Node 'node'.
786 const struct cpumask *cpumask_of_node(int node)
788 if (node >= nr_node_ids) {
790 "cpumask_of_node(%d): node > nr_node_ids(%d)\n",
793 return cpu_none_mask;
795 if (node_to_cpumask_map[node] == NULL) {
797 "cpumask_of_node(%d): no node_to_cpumask_map!\n",
800 return cpu_online_mask;
802 return node_to_cpumask_map[node];
804 EXPORT_SYMBOL(cpumask_of_node);
806 #endif /* !CONFIG_DEBUG_PER_CPU_MAPS */
808 #ifdef CONFIG_MEMORY_HOTPLUG
809 int memory_add_physaddr_to_nid(u64 start)
811 struct numa_meminfo *mi = &numa_meminfo;
812 int nid = mi->blk[0].nid;
815 for (i = 0; i < mi->nr_blks; i++)
816 if (mi->blk[i].start <= start && mi->blk[i].end > start)
817 nid = mi->blk[i].nid;
820 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);