1
2#include <linux/kernel.h>
3#include <linux/mm.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>
14
15#include <asm/e820.h>
16#include <asm/proto.h>
17#include <asm/dma.h>
18#include <asm/acpi.h>
19#include <asm/amd_nb.h>
20
21#include "numa_internal.h"
22
23int __initdata numa_off;
24nodemask_t numa_nodes_parsed __initdata;
25
26struct pglist_data *node_data[MAX_NUMNODES] __read_mostly;
27EXPORT_SYMBOL(node_data);
28
29static struct numa_meminfo numa_meminfo
30#ifndef CONFIG_MEMORY_HOTPLUG
31__initdata
32#endif
33;
34
35static int numa_distance_cnt;
36static u8 *numa_distance;
37
38static __init int numa_setup(char *opt)
39{
40 if (!opt)
41 return -EINVAL;
42 if (!strncmp(opt, "off", 3))
43 numa_off = 1;
44#ifdef CONFIG_NUMA_EMU
45 if (!strncmp(opt, "fake=", 5))
46 numa_emu_cmdline(opt + 5);
47#endif
48#ifdef CONFIG_ACPI_NUMA
49 if (!strncmp(opt, "noacpi", 6))
50 acpi_numa = -1;
51#endif
52 return 0;
53}
54early_param("numa", numa_setup);
55
56
57
58
59s16 __apicid_to_node[MAX_LOCAL_APIC] = {
60 [0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE
61};
62
63int numa_cpu_node(int cpu)
64{
65 int apicid = early_per_cpu(x86_cpu_to_apicid, cpu);
66
67 if (apicid != BAD_APICID)
68 return __apicid_to_node[apicid];
69 return NUMA_NO_NODE;
70}
71
72cpumask_var_t node_to_cpumask_map[MAX_NUMNODES];
73EXPORT_SYMBOL(node_to_cpumask_map);
74
75
76
77
78DEFINE_EARLY_PER_CPU(int, x86_cpu_to_node_map, NUMA_NO_NODE);
79EXPORT_EARLY_PER_CPU_SYMBOL(x86_cpu_to_node_map);
80
81void numa_set_node(int cpu, int node)
82{
83 int *cpu_to_node_map = early_per_cpu_ptr(x86_cpu_to_node_map);
84
85
86 if (cpu_to_node_map) {
87 cpu_to_node_map[cpu] = node;
88 return;
89 }
90
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);
94 dump_stack();
95 return;
96 }
97#endif
98 per_cpu(x86_cpu_to_node_map, cpu) = node;
99
100 set_cpu_numa_node(cpu, node);
101}
102
103void numa_clear_node(int cpu)
104{
105 numa_set_node(cpu, NUMA_NO_NODE);
106}
107
108
109
110
111
112
113
114
115void __init setup_node_to_cpumask_map(void)
116{
117 unsigned int node;
118
119
120 if (nr_node_ids == MAX_NUMNODES)
121 setup_nr_node_ids();
122
123
124 for (node = 0; node < nr_node_ids; node++)
125 alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]);
126
127
128 pr_debug("Node to cpumask map for %d nodes\n", nr_node_ids);
129}
130
131static int __init numa_add_memblk_to(int nid, u64 start, u64 end,
132 struct numa_meminfo *mi)
133{
134
135 if (start == end)
136 return 0;
137
138
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);
142 return 0;
143 }
144
145 if (mi->nr_blks >= NR_NODE_MEMBLKS) {
146 pr_err("NUMA: too many memblk ranges\n");
147 return -EINVAL;
148 }
149
150 mi->blk[mi->nr_blks].start = start;
151 mi->blk[mi->nr_blks].end = end;
152 mi->blk[mi->nr_blks].nid = nid;
153 mi->nr_blks++;
154 return 0;
155}
156
157
158
159
160
161
162
163
164
165void __init numa_remove_memblk_from(int idx, struct numa_meminfo *mi)
166{
167 mi->nr_blks--;
168 memmove(&mi->blk[idx], &mi->blk[idx + 1],
169 (mi->nr_blks - idx) * sizeof(mi->blk[0]));
170}
171
172
173
174
175
176
177
178
179
180
181
182
183int __init numa_add_memblk(int nid, u64 start, u64 end)
184{
185 return numa_add_memblk_to(nid, start, end, &numa_meminfo);
186}
187
188
189static void __init setup_node_data(int nid, u64 start, u64 end)
190{
191 const size_t nd_size = roundup(sizeof(pg_data_t), PAGE_SIZE);
192 u64 nd_pa;
193 void *nd;
194 int tnid;
195
196
197
198
199
200 if (end && (end - start) < NODE_MIN_SIZE)
201 return;
202
203 start = roundup(start, ZONE_ALIGN);
204
205 printk(KERN_INFO "Initmem setup node %d [mem %#010Lx-%#010Lx]\n",
206 nid, start, end - 1);
207
208
209
210
211
212 nd_pa = memblock_alloc_nid(nd_size, SMP_CACHE_BYTES, nid);
213 if (!nd_pa) {
214 pr_err("Cannot find %zu bytes in node %d\n",
215 nd_size, nid);
216 return;
217 }
218 nd = __va(nd_pa);
219
220
221 printk(KERN_INFO " NODE_DATA [mem %#010Lx-%#010Lx]\n",
222 nd_pa, nd_pa + nd_size - 1);
223 tnid = early_pfn_to_nid(nd_pa >> PAGE_SHIFT);
224 if (tnid != nid)
225 printk(KERN_INFO " NODE_DATA(%d) on node %d\n", nid, tnid);
226
227 node_data[nid] = nd;
228 memset(NODE_DATA(nid), 0, sizeof(pg_data_t));
229 NODE_DATA(nid)->node_id = nid;
230 NODE_DATA(nid)->node_start_pfn = start >> PAGE_SHIFT;
231 NODE_DATA(nid)->node_spanned_pages = (end - start) >> PAGE_SHIFT;
232
233 node_set_online(nid);
234}
235
236
237
238
239
240
241
242
243
244
245
246int __init numa_cleanup_meminfo(struct numa_meminfo *mi)
247{
248 const u64 low = 0;
249 const u64 high = PFN_PHYS(max_pfn);
250 int i, j, k;
251
252
253 for (i = 0; i < mi->nr_blks; i++) {
254 struct numa_memblk *bi = &mi->blk[i];
255
256
257 bi->start = max(bi->start, low);
258 bi->end = min(bi->end, high);
259
260
261 if (bi->start >= bi->end)
262 numa_remove_memblk_from(i--, mi);
263 }
264
265
266 for (i = 0; i < mi->nr_blks; i++) {
267 struct numa_memblk *bi = &mi->blk[i];
268
269 for (j = i + 1; j < mi->nr_blks; j++) {
270 struct numa_memblk *bj = &mi->blk[j];
271 u64 start, end;
272
273
274
275
276
277
278 if (bi->end > bj->start && bi->start < bj->end) {
279 if (bi->nid != bj->nid) {
280 pr_err("NUMA: node %d [mem %#010Lx-%#010Lx] overlaps with node %d [mem %#010Lx-%#010Lx]\n",
281 bi->nid, bi->start, bi->end - 1,
282 bj->nid, bj->start, bj->end - 1);
283 return -EINVAL;
284 }
285 pr_warning("NUMA: Warning: node %d [mem %#010Lx-%#010Lx] overlaps with itself [mem %#010Lx-%#010Lx]\n",
286 bi->nid, bi->start, bi->end - 1,
287 bj->start, bj->end - 1);
288 }
289
290
291
292
293
294
295 if (bi->nid != bj->nid)
296 continue;
297 start = min(bi->start, bj->start);
298 end = max(bi->end, bj->end);
299 for (k = 0; k < mi->nr_blks; k++) {
300 struct numa_memblk *bk = &mi->blk[k];
301
302 if (bi->nid == bk->nid)
303 continue;
304 if (start < bk->end && end > bk->start)
305 break;
306 }
307 if (k < mi->nr_blks)
308 continue;
309 printk(KERN_INFO "NUMA: Node %d [mem %#010Lx-%#010Lx] + [mem %#010Lx-%#010Lx] -> [mem %#010Lx-%#010Lx]\n",
310 bi->nid, bi->start, bi->end - 1, bj->start,
311 bj->end - 1, start, end - 1);
312 bi->start = start;
313 bi->end = end;
314 numa_remove_memblk_from(j--, mi);
315 }
316 }
317
318
319 for (i = mi->nr_blks; i < ARRAY_SIZE(mi->blk); i++) {
320 mi->blk[i].start = mi->blk[i].end = 0;
321 mi->blk[i].nid = NUMA_NO_NODE;
322 }
323
324 return 0;
325}
326
327
328
329
330static void __init numa_nodemask_from_meminfo(nodemask_t *nodemask,
331 const struct numa_meminfo *mi)
332{
333 int i;
334
335 for (i = 0; i < ARRAY_SIZE(mi->blk); i++)
336 if (mi->blk[i].start != mi->blk[i].end &&
337 mi->blk[i].nid != NUMA_NO_NODE)
338 node_set(mi->blk[i].nid, *nodemask);
339}
340
341
342
343
344
345
346
347void __init numa_reset_distance(void)
348{
349 size_t size = numa_distance_cnt * numa_distance_cnt * sizeof(numa_distance[0]);
350
351
352 if (numa_distance_cnt)
353 memblock_free(__pa(numa_distance), size);
354 numa_distance_cnt = 0;
355 numa_distance = NULL;
356}
357
358static int __init numa_alloc_distance(void)
359{
360 nodemask_t nodes_parsed;
361 size_t size;
362 int i, j, cnt = 0;
363 u64 phys;
364
365
366 nodes_parsed = numa_nodes_parsed;
367 numa_nodemask_from_meminfo(&nodes_parsed, &numa_meminfo);
368
369 for_each_node_mask(i, nodes_parsed)
370 cnt = i;
371 cnt++;
372 size = cnt * cnt * sizeof(numa_distance[0]);
373
374 phys = memblock_find_in_range(0, PFN_PHYS(max_pfn_mapped),
375 size, PAGE_SIZE);
376 if (!phys) {
377 pr_warning("NUMA: Warning: can't allocate distance table!\n");
378
379 numa_distance = (void *)1LU;
380 return -ENOMEM;
381 }
382 memblock_reserve(phys, size);
383
384 numa_distance = __va(phys);
385 numa_distance_cnt = cnt;
386
387
388 for (i = 0; i < cnt; i++)
389 for (j = 0; j < cnt; j++)
390 numa_distance[i * cnt + j] = i == j ?
391 LOCAL_DISTANCE : REMOTE_DISTANCE;
392 printk(KERN_DEBUG "NUMA: Initialized distance table, cnt=%d\n", cnt);
393
394 return 0;
395}
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416void __init numa_set_distance(int from, int to, int distance)
417{
418 if (!numa_distance && numa_alloc_distance() < 0)
419 return;
420
421 if (from >= numa_distance_cnt || to >= numa_distance_cnt ||
422 from < 0 || to < 0) {
423 pr_warn_once("NUMA: Warning: node ids are out of bound, from=%d to=%d distance=%d\n",
424 from, to, distance);
425 return;
426 }
427
428 if ((u8)distance != distance ||
429 (from == to && distance != LOCAL_DISTANCE)) {
430 pr_warn_once("NUMA: Warning: invalid distance parameter, from=%d to=%d distance=%d\n",
431 from, to, distance);
432 return;
433 }
434
435 numa_distance[from * numa_distance_cnt + to] = distance;
436}
437
438int __node_distance(int from, int to)
439{
440 if (from >= numa_distance_cnt || to >= numa_distance_cnt)
441 return from == to ? LOCAL_DISTANCE : REMOTE_DISTANCE;
442 return numa_distance[from * numa_distance_cnt + to];
443}
444EXPORT_SYMBOL(__node_distance);
445
446
447
448
449
450static bool __init numa_meminfo_cover_memory(const struct numa_meminfo *mi)
451{
452 u64 numaram, e820ram;
453 int i;
454
455 numaram = 0;
456 for (i = 0; i < mi->nr_blks; i++) {
457 u64 s = mi->blk[i].start >> PAGE_SHIFT;
458 u64 e = mi->blk[i].end >> PAGE_SHIFT;
459 numaram += e - s;
460 numaram -= __absent_pages_in_range(mi->blk[i].nid, s, e);
461 if ((s64)numaram < 0)
462 numaram = 0;
463 }
464
465 e820ram = max_pfn - absent_pages_in_range(0, max_pfn);
466
467
468 if ((s64)(e820ram - numaram) >= (1 << (20 - PAGE_SHIFT))) {
469 printk(KERN_ERR "NUMA: nodes only cover %LuMB of your %LuMB e820 RAM. Not used.\n",
470 (numaram << PAGE_SHIFT) >> 20,
471 (e820ram << PAGE_SHIFT) >> 20);
472 return false;
473 }
474 return true;
475}
476
477static int __init numa_register_memblks(struct numa_meminfo *mi)
478{
479 unsigned long uninitialized_var(pfn_align);
480 int i, nid;
481
482
483 node_possible_map = numa_nodes_parsed;
484 numa_nodemask_from_meminfo(&node_possible_map, mi);
485 if (WARN_ON(nodes_empty(node_possible_map)))
486 return -EINVAL;
487
488 for (i = 0; i < mi->nr_blks; i++) {
489 struct numa_memblk *mb = &mi->blk[i];
490 memblock_set_node(mb->start, mb->end - mb->start, mb->nid);
491 }
492
493
494
495
496
497#ifdef NODE_NOT_IN_PAGE_FLAGS
498 pfn_align = node_map_pfn_alignment();
499 if (pfn_align && pfn_align < PAGES_PER_SECTION) {
500 printk(KERN_WARNING "Node alignment %LuMB < min %LuMB, rejecting NUMA config\n",
501 PFN_PHYS(pfn_align) >> 20,
502 PFN_PHYS(PAGES_PER_SECTION) >> 20);
503 return -EINVAL;
504 }
505#endif
506 if (!numa_meminfo_cover_memory(mi))
507 return -EINVAL;
508
509
510 for_each_node_mask(nid, node_possible_map) {
511 u64 start = PFN_PHYS(max_pfn);
512 u64 end = 0;
513
514 for (i = 0; i < mi->nr_blks; i++) {
515 if (nid != mi->blk[i].nid)
516 continue;
517 start = min(mi->blk[i].start, start);
518 end = max(mi->blk[i].end, end);
519 }
520
521 if (start < end)
522 setup_node_data(nid, start, end);
523 }
524
525
526 memblock_dump_all();
527 return 0;
528}
529
530
531
532
533
534
535
536
537static void __init numa_init_array(void)
538{
539 int rr, i;
540
541 rr = first_node(node_online_map);
542 for (i = 0; i < nr_cpu_ids; i++) {
543 if (early_cpu_to_node(i) != NUMA_NO_NODE)
544 continue;
545 numa_set_node(i, rr);
546 rr = next_node(rr, node_online_map);
547 if (rr == MAX_NUMNODES)
548 rr = first_node(node_online_map);
549 }
550}
551
552static int __init numa_init(int (*init_func)(void))
553{
554 int i;
555 int ret;
556
557 for (i = 0; i < MAX_LOCAL_APIC; i++)
558 set_apicid_to_node(i, NUMA_NO_NODE);
559
560 nodes_clear(numa_nodes_parsed);
561 nodes_clear(node_possible_map);
562 nodes_clear(node_online_map);
563 memset(&numa_meminfo, 0, sizeof(numa_meminfo));
564 WARN_ON(memblock_set_node(0, ULLONG_MAX, MAX_NUMNODES));
565 numa_reset_distance();
566
567 ret = init_func();
568 if (ret < 0)
569 return ret;
570 ret = numa_cleanup_meminfo(&numa_meminfo);
571 if (ret < 0)
572 return ret;
573
574 numa_emulation(&numa_meminfo, numa_distance_cnt);
575
576 ret = numa_register_memblks(&numa_meminfo);
577 if (ret < 0)
578 return ret;
579
580 for (i = 0; i < nr_cpu_ids; i++) {
581 int nid = early_cpu_to_node(i);
582
583 if (nid == NUMA_NO_NODE)
584 continue;
585 if (!node_online(nid))
586 numa_clear_node(i);
587 }
588 numa_init_array();
589 return 0;
590}
591
592
593
594
595
596
597
598
599
600
601static int __init dummy_numa_init(void)
602{
603 printk(KERN_INFO "%s\n",
604 numa_off ? "NUMA turned off" : "No NUMA configuration found");
605 printk(KERN_INFO "Faking a node at [mem %#018Lx-%#018Lx]\n",
606 0LLU, PFN_PHYS(max_pfn) - 1);
607
608 node_set(0, numa_nodes_parsed);
609 numa_add_memblk(0, 0, PFN_PHYS(max_pfn));
610
611 return 0;
612}
613
614
615
616
617
618
619
620
621void __init x86_numa_init(void)
622{
623 if (!numa_off) {
624#ifdef CONFIG_X86_NUMAQ
625 if (!numa_init(numaq_numa_init))
626 return;
627#endif
628#ifdef CONFIG_ACPI_NUMA
629 if (!numa_init(x86_acpi_numa_init))
630 return;
631#endif
632#ifdef CONFIG_AMD_NUMA
633 if (!numa_init(amd_numa_init))
634 return;
635#endif
636 }
637
638 numa_init(dummy_numa_init);
639}
640
641static __init int find_near_online_node(int node)
642{
643 int n, val;
644 int min_val = INT_MAX;
645 int best_node = -1;
646
647 for_each_online_node(n) {
648 val = node_distance(node, n);
649
650 if (val < min_val) {
651 min_val = val;
652 best_node = n;
653 }
654 }
655
656 return best_node;
657}
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673void __init init_cpu_to_node(void)
674{
675 int cpu;
676 u16 *cpu_to_apicid = early_per_cpu_ptr(x86_cpu_to_apicid);
677
678 BUG_ON(cpu_to_apicid == NULL);
679
680 for_each_possible_cpu(cpu) {
681 int node = numa_cpu_node(cpu);
682
683 if (node == NUMA_NO_NODE)
684 continue;
685 if (!node_online(node))
686 node = find_near_online_node(node);
687 numa_set_node(cpu, node);
688 }
689}
690
691#ifndef CONFIG_DEBUG_PER_CPU_MAPS
692
693# ifndef CONFIG_NUMA_EMU
694void numa_add_cpu(int cpu)
695{
696 cpumask_set_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
697}
698
699void numa_remove_cpu(int cpu)
700{
701 cpumask_clear_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
702}
703# endif
704
705#else
706
707int __cpu_to_node(int cpu)
708{
709 if (early_per_cpu_ptr(x86_cpu_to_node_map)) {
710 printk(KERN_WARNING
711 "cpu_to_node(%d): usage too early!\n", cpu);
712 dump_stack();
713 return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
714 }
715 return per_cpu(x86_cpu_to_node_map, cpu);
716}
717EXPORT_SYMBOL(__cpu_to_node);
718
719
720
721
722
723int early_cpu_to_node(int cpu)
724{
725 if (early_per_cpu_ptr(x86_cpu_to_node_map))
726 return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
727
728 if (!cpu_possible(cpu)) {
729 printk(KERN_WARNING
730 "early_cpu_to_node(%d): no per_cpu area!\n", cpu);
731 dump_stack();
732 return NUMA_NO_NODE;
733 }
734 return per_cpu(x86_cpu_to_node_map, cpu);
735}
736
737void debug_cpumask_set_cpu(int cpu, int node, bool enable)
738{
739 struct cpumask *mask;
740 char buf[64];
741
742 if (node == NUMA_NO_NODE) {
743
744 return;
745 }
746 mask = node_to_cpumask_map[node];
747 if (!mask) {
748 pr_err("node_to_cpumask_map[%i] NULL\n", node);
749 dump_stack();
750 return;
751 }
752
753 if (enable)
754 cpumask_set_cpu(cpu, mask);
755 else
756 cpumask_clear_cpu(cpu, mask);
757
758 cpulist_scnprintf(buf, sizeof(buf), mask);
759 printk(KERN_DEBUG "%s cpu %d node %d: mask now %s\n",
760 enable ? "numa_add_cpu" : "numa_remove_cpu",
761 cpu, node, buf);
762 return;
763}
764
765# ifndef CONFIG_NUMA_EMU
766static void numa_set_cpumask(int cpu, bool enable)
767{
768 debug_cpumask_set_cpu(cpu, early_cpu_to_node(cpu), enable);
769}
770
771void numa_add_cpu(int cpu)
772{
773 numa_set_cpumask(cpu, true);
774}
775
776void numa_remove_cpu(int cpu)
777{
778 numa_set_cpumask(cpu, false);
779}
780# endif
781
782
783
784
785const struct cpumask *cpumask_of_node(int node)
786{
787 if (node >= nr_node_ids) {
788 printk(KERN_WARNING
789 "cpumask_of_node(%d): node > nr_node_ids(%d)\n",
790 node, nr_node_ids);
791 dump_stack();
792 return cpu_none_mask;
793 }
794 if (node_to_cpumask_map[node] == NULL) {
795 printk(KERN_WARNING
796 "cpumask_of_node(%d): no node_to_cpumask_map!\n",
797 node);
798 dump_stack();
799 return cpu_online_mask;
800 }
801 return node_to_cpumask_map[node];
802}
803EXPORT_SYMBOL(cpumask_of_node);
804
805#endif
806
807#ifdef CONFIG_MEMORY_HOTPLUG
808int memory_add_physaddr_to_nid(u64 start)
809{
810 struct numa_meminfo *mi = &numa_meminfo;
811 int nid = mi->blk[0].nid;
812 int i;
813
814 for (i = 0; i < mi->nr_blks; i++)
815 if (mi->blk[i].start <= start && mi->blk[i].end > start)
816 nid = mi->blk[i].nid;
817 return nid;
818}
819EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
820#endif
821