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 alloc_node_data(int nid)
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 nd_pa = memblock_alloc_nid(nd_size, SMP_CACHE_BYTES, nid);
201 if (!nd_pa) {
202 nd_pa = __memblock_alloc_base(nd_size, SMP_CACHE_BYTES,
203 MEMBLOCK_ALLOC_ACCESSIBLE);
204 if (!nd_pa) {
205 pr_err("Cannot find %zu bytes in node %d\n",
206 nd_size, nid);
207 return;
208 }
209 }
210 nd = __va(nd_pa);
211
212
213 printk(KERN_INFO "NODE_DATA(%d) allocated [mem %#010Lx-%#010Lx]\n", nid,
214 nd_pa, nd_pa + nd_size - 1);
215 tnid = early_pfn_to_nid(nd_pa >> PAGE_SHIFT);
216 if (tnid != nid)
217 printk(KERN_INFO " NODE_DATA(%d) on node %d\n", nid, tnid);
218
219 node_data[nid] = nd;
220 memset(NODE_DATA(nid), 0, sizeof(pg_data_t));
221
222 node_set_online(nid);
223}
224
225
226
227
228
229
230
231
232
233
234
235int __init numa_cleanup_meminfo(struct numa_meminfo *mi)
236{
237 const u64 low = 0;
238 const u64 high = PFN_PHYS(max_pfn);
239 int i, j, k;
240
241
242 for (i = 0; i < mi->nr_blks; i++) {
243 struct numa_memblk *bi = &mi->blk[i];
244
245
246 bi->start = max(bi->start, low);
247 bi->end = min(bi->end, high);
248
249
250 if (bi->start >= bi->end)
251 numa_remove_memblk_from(i--, mi);
252 }
253
254
255 for (i = 0; i < mi->nr_blks; i++) {
256 struct numa_memblk *bi = &mi->blk[i];
257
258 for (j = i + 1; j < mi->nr_blks; j++) {
259 struct numa_memblk *bj = &mi->blk[j];
260 u64 start, end;
261
262
263
264
265
266
267 if (bi->end > bj->start && bi->start < bj->end) {
268 if (bi->nid != bj->nid) {
269 pr_err("NUMA: node %d [mem %#010Lx-%#010Lx] overlaps with node %d [mem %#010Lx-%#010Lx]\n",
270 bi->nid, bi->start, bi->end - 1,
271 bj->nid, bj->start, bj->end - 1);
272 return -EINVAL;
273 }
274 pr_warning("NUMA: Warning: node %d [mem %#010Lx-%#010Lx] overlaps with itself [mem %#010Lx-%#010Lx]\n",
275 bi->nid, bi->start, bi->end - 1,
276 bj->start, bj->end - 1);
277 }
278
279
280
281
282
283
284 if (bi->nid != bj->nid)
285 continue;
286 start = min(bi->start, bj->start);
287 end = max(bi->end, bj->end);
288 for (k = 0; k < mi->nr_blks; k++) {
289 struct numa_memblk *bk = &mi->blk[k];
290
291 if (bi->nid == bk->nid)
292 continue;
293 if (start < bk->end && end > bk->start)
294 break;
295 }
296 if (k < mi->nr_blks)
297 continue;
298 printk(KERN_INFO "NUMA: Node %d [mem %#010Lx-%#010Lx] + [mem %#010Lx-%#010Lx] -> [mem %#010Lx-%#010Lx]\n",
299 bi->nid, bi->start, bi->end - 1, bj->start,
300 bj->end - 1, start, end - 1);
301 bi->start = start;
302 bi->end = end;
303 numa_remove_memblk_from(j--, mi);
304 }
305 }
306
307
308 for (i = mi->nr_blks; i < ARRAY_SIZE(mi->blk); i++) {
309 mi->blk[i].start = mi->blk[i].end = 0;
310 mi->blk[i].nid = NUMA_NO_NODE;
311 }
312
313 return 0;
314}
315
316
317
318
319static void __init numa_nodemask_from_meminfo(nodemask_t *nodemask,
320 const struct numa_meminfo *mi)
321{
322 int i;
323
324 for (i = 0; i < ARRAY_SIZE(mi->blk); i++)
325 if (mi->blk[i].start != mi->blk[i].end &&
326 mi->blk[i].nid != NUMA_NO_NODE)
327 node_set(mi->blk[i].nid, *nodemask);
328}
329
330
331
332
333
334
335
336void __init numa_reset_distance(void)
337{
338 size_t size = numa_distance_cnt * numa_distance_cnt * sizeof(numa_distance[0]);
339
340
341 if (numa_distance_cnt)
342 memblock_free(__pa(numa_distance), size);
343 numa_distance_cnt = 0;
344 numa_distance = NULL;
345}
346
347static int __init numa_alloc_distance(void)
348{
349 nodemask_t nodes_parsed;
350 size_t size;
351 int i, j, cnt = 0;
352 u64 phys;
353
354
355 nodes_parsed = numa_nodes_parsed;
356 numa_nodemask_from_meminfo(&nodes_parsed, &numa_meminfo);
357
358 for_each_node_mask(i, nodes_parsed)
359 cnt = i;
360 cnt++;
361 size = cnt * cnt * sizeof(numa_distance[0]);
362
363 phys = memblock_find_in_range(0, PFN_PHYS(max_pfn_mapped),
364 size, PAGE_SIZE);
365 if (!phys) {
366 pr_warning("NUMA: Warning: can't allocate distance table!\n");
367
368 numa_distance = (void *)1LU;
369 return -ENOMEM;
370 }
371 memblock_reserve(phys, size);
372
373 numa_distance = __va(phys);
374 numa_distance_cnt = cnt;
375
376
377 for (i = 0; i < cnt; i++)
378 for (j = 0; j < cnt; j++)
379 numa_distance[i * cnt + j] = i == j ?
380 LOCAL_DISTANCE : REMOTE_DISTANCE;
381 printk(KERN_DEBUG "NUMA: Initialized distance table, cnt=%d\n", cnt);
382
383 return 0;
384}
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405void __init numa_set_distance(int from, int to, int distance)
406{
407 if (!numa_distance && numa_alloc_distance() < 0)
408 return;
409
410 if (from >= numa_distance_cnt || to >= numa_distance_cnt ||
411 from < 0 || to < 0) {
412 pr_warn_once("NUMA: Warning: node ids are out of bound, from=%d to=%d distance=%d\n",
413 from, to, distance);
414 return;
415 }
416
417 if ((u8)distance != distance ||
418 (from == to && distance != LOCAL_DISTANCE)) {
419 pr_warn_once("NUMA: Warning: invalid distance parameter, from=%d to=%d distance=%d\n",
420 from, to, distance);
421 return;
422 }
423
424 numa_distance[from * numa_distance_cnt + to] = distance;
425}
426
427int __node_distance(int from, int to)
428{
429 if (from >= numa_distance_cnt || to >= numa_distance_cnt)
430 return from == to ? LOCAL_DISTANCE : REMOTE_DISTANCE;
431 return numa_distance[from * numa_distance_cnt + to];
432}
433EXPORT_SYMBOL(__node_distance);
434
435
436
437
438
439static bool __init numa_meminfo_cover_memory(const struct numa_meminfo *mi)
440{
441 u64 numaram, e820ram;
442 int i;
443
444 numaram = 0;
445 for (i = 0; i < mi->nr_blks; i++) {
446 u64 s = mi->blk[i].start >> PAGE_SHIFT;
447 u64 e = mi->blk[i].end >> PAGE_SHIFT;
448 numaram += e - s;
449 numaram -= __absent_pages_in_range(mi->blk[i].nid, s, e);
450 if ((s64)numaram < 0)
451 numaram = 0;
452 }
453
454 e820ram = max_pfn - absent_pages_in_range(0, max_pfn);
455
456
457 if ((s64)(e820ram - numaram) >= (1 << (20 - PAGE_SHIFT))) {
458 printk(KERN_ERR "NUMA: nodes only cover %LuMB of your %LuMB e820 RAM. Not used.\n",
459 (numaram << PAGE_SHIFT) >> 20,
460 (e820ram << PAGE_SHIFT) >> 20);
461 return false;
462 }
463 return true;
464}
465
466static void __init numa_clear_kernel_node_hotplug(void)
467{
468 int i, nid;
469 nodemask_t numa_kernel_nodes = NODE_MASK_NONE;
470 unsigned long start, end;
471 struct memblock_region *r;
472
473
474
475
476
477
478 for (i = 0; i < numa_meminfo.nr_blks; i++) {
479 struct numa_memblk *mb = &numa_meminfo.blk[i];
480
481 memblock_set_node(mb->start, mb->end - mb->start,
482 &memblock.reserved, mb->nid);
483 }
484
485
486 for_each_memblock(reserved, r)
487 node_set(r->nid, numa_kernel_nodes);
488
489
490 for (i = 0; i < numa_meminfo.nr_blks; i++) {
491 nid = numa_meminfo.blk[i].nid;
492 if (!node_isset(nid, numa_kernel_nodes))
493 continue;
494
495 start = numa_meminfo.blk[i].start;
496 end = numa_meminfo.blk[i].end;
497
498 memblock_clear_hotplug(start, end - start);
499 }
500}
501
502static int __init numa_register_memblks(struct numa_meminfo *mi)
503{
504 unsigned long uninitialized_var(pfn_align);
505 int i, nid;
506
507
508 node_possible_map = numa_nodes_parsed;
509 numa_nodemask_from_meminfo(&node_possible_map, mi);
510 if (WARN_ON(nodes_empty(node_possible_map)))
511 return -EINVAL;
512
513 for (i = 0; i < mi->nr_blks; i++) {
514 struct numa_memblk *mb = &mi->blk[i];
515 memblock_set_node(mb->start, mb->end - mb->start,
516 &memblock.memory, mb->nid);
517 }
518
519
520
521
522
523
524
525
526 numa_clear_kernel_node_hotplug();
527
528
529
530
531
532#ifdef NODE_NOT_IN_PAGE_FLAGS
533 pfn_align = node_map_pfn_alignment();
534 if (pfn_align && pfn_align < PAGES_PER_SECTION) {
535 printk(KERN_WARNING "Node alignment %LuMB < min %LuMB, rejecting NUMA config\n",
536 PFN_PHYS(pfn_align) >> 20,
537 PFN_PHYS(PAGES_PER_SECTION) >> 20);
538 return -EINVAL;
539 }
540#endif
541 if (!numa_meminfo_cover_memory(mi))
542 return -EINVAL;
543
544
545 for_each_node_mask(nid, node_possible_map) {
546 u64 start = PFN_PHYS(max_pfn);
547 u64 end = 0;
548
549 for (i = 0; i < mi->nr_blks; i++) {
550 if (nid != mi->blk[i].nid)
551 continue;
552 start = min(mi->blk[i].start, start);
553 end = max(mi->blk[i].end, end);
554 }
555
556 if (start >= end)
557 continue;
558
559
560
561
562
563 if (end && (end - start) < NODE_MIN_SIZE)
564 continue;
565
566 alloc_node_data(nid);
567 }
568
569
570 memblock_dump_all();
571 return 0;
572}
573
574
575
576
577
578
579
580
581static void __init numa_init_array(void)
582{
583 int rr, i;
584
585 rr = first_node(node_online_map);
586 for (i = 0; i < nr_cpu_ids; i++) {
587 if (early_cpu_to_node(i) != NUMA_NO_NODE)
588 continue;
589 numa_set_node(i, rr);
590 rr = next_node(rr, node_online_map);
591 if (rr == MAX_NUMNODES)
592 rr = first_node(node_online_map);
593 }
594}
595
596static int __init numa_init(int (*init_func)(void))
597{
598 int i;
599 int ret;
600
601 for (i = 0; i < MAX_LOCAL_APIC; i++)
602 set_apicid_to_node(i, NUMA_NO_NODE);
603
604 nodes_clear(numa_nodes_parsed);
605 nodes_clear(node_possible_map);
606 nodes_clear(node_online_map);
607 memset(&numa_meminfo, 0, sizeof(numa_meminfo));
608 WARN_ON(memblock_set_node(0, ULLONG_MAX, &memblock.memory,
609 MAX_NUMNODES));
610 WARN_ON(memblock_set_node(0, ULLONG_MAX, &memblock.reserved,
611 MAX_NUMNODES));
612
613 WARN_ON(memblock_clear_hotplug(0, ULLONG_MAX));
614 numa_reset_distance();
615
616 ret = init_func();
617 if (ret < 0)
618 return ret;
619
620
621
622
623
624
625
626
627
628 memblock_set_bottom_up(false);
629
630 ret = numa_cleanup_meminfo(&numa_meminfo);
631 if (ret < 0)
632 return ret;
633
634 numa_emulation(&numa_meminfo, numa_distance_cnt);
635
636 ret = numa_register_memblks(&numa_meminfo);
637 if (ret < 0)
638 return ret;
639
640 for (i = 0; i < nr_cpu_ids; i++) {
641 int nid = early_cpu_to_node(i);
642
643 if (nid == NUMA_NO_NODE)
644 continue;
645 if (!node_online(nid))
646 numa_clear_node(i);
647 }
648 numa_init_array();
649
650 return 0;
651}
652
653
654
655
656
657
658
659
660
661
662static int __init dummy_numa_init(void)
663{
664 printk(KERN_INFO "%s\n",
665 numa_off ? "NUMA turned off" : "No NUMA configuration found");
666 printk(KERN_INFO "Faking a node at [mem %#018Lx-%#018Lx]\n",
667 0LLU, PFN_PHYS(max_pfn) - 1);
668
669 node_set(0, numa_nodes_parsed);
670 numa_add_memblk(0, 0, PFN_PHYS(max_pfn));
671
672 return 0;
673}
674
675
676
677
678
679
680
681
682void __init x86_numa_init(void)
683{
684 if (!numa_off) {
685#ifdef CONFIG_ACPI_NUMA
686 if (!numa_init(x86_acpi_numa_init))
687 return;
688#endif
689#ifdef CONFIG_AMD_NUMA
690 if (!numa_init(amd_numa_init))
691 return;
692#endif
693 }
694
695 numa_init(dummy_numa_init);
696}
697
698static __init int find_near_online_node(int node)
699{
700 int n, val;
701 int min_val = INT_MAX;
702 int best_node = -1;
703
704 for_each_online_node(n) {
705 val = node_distance(node, n);
706
707 if (val < min_val) {
708 min_val = val;
709 best_node = n;
710 }
711 }
712
713 return best_node;
714}
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730void __init init_cpu_to_node(void)
731{
732 int cpu;
733 u16 *cpu_to_apicid = early_per_cpu_ptr(x86_cpu_to_apicid);
734
735 BUG_ON(cpu_to_apicid == NULL);
736
737 for_each_possible_cpu(cpu) {
738 int node = numa_cpu_node(cpu);
739
740 if (node == NUMA_NO_NODE)
741 continue;
742 if (!node_online(node))
743 node = find_near_online_node(node);
744 numa_set_node(cpu, node);
745 }
746}
747
748#ifndef CONFIG_DEBUG_PER_CPU_MAPS
749
750# ifndef CONFIG_NUMA_EMU
751void numa_add_cpu(int cpu)
752{
753 cpumask_set_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
754}
755
756void numa_remove_cpu(int cpu)
757{
758 cpumask_clear_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
759}
760# endif
761
762#else
763
764int __cpu_to_node(int cpu)
765{
766 if (early_per_cpu_ptr(x86_cpu_to_node_map)) {
767 printk(KERN_WARNING
768 "cpu_to_node(%d): usage too early!\n", cpu);
769 dump_stack();
770 return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
771 }
772 return per_cpu(x86_cpu_to_node_map, cpu);
773}
774EXPORT_SYMBOL(__cpu_to_node);
775
776
777
778
779
780int early_cpu_to_node(int cpu)
781{
782 if (early_per_cpu_ptr(x86_cpu_to_node_map))
783 return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
784
785 if (!cpu_possible(cpu)) {
786 printk(KERN_WARNING
787 "early_cpu_to_node(%d): no per_cpu area!\n", cpu);
788 dump_stack();
789 return NUMA_NO_NODE;
790 }
791 return per_cpu(x86_cpu_to_node_map, cpu);
792}
793
794void debug_cpumask_set_cpu(int cpu, int node, bool enable)
795{
796 struct cpumask *mask;
797
798 if (node == NUMA_NO_NODE) {
799
800 return;
801 }
802 mask = node_to_cpumask_map[node];
803 if (!mask) {
804 pr_err("node_to_cpumask_map[%i] NULL\n", node);
805 dump_stack();
806 return;
807 }
808
809 if (enable)
810 cpumask_set_cpu(cpu, mask);
811 else
812 cpumask_clear_cpu(cpu, mask);
813
814 printk(KERN_DEBUG "%s cpu %d node %d: mask now %*pbl\n",
815 enable ? "numa_add_cpu" : "numa_remove_cpu",
816 cpu, node, cpumask_pr_args(mask));
817 return;
818}
819
820# ifndef CONFIG_NUMA_EMU
821static void numa_set_cpumask(int cpu, bool enable)
822{
823 debug_cpumask_set_cpu(cpu, early_cpu_to_node(cpu), enable);
824}
825
826void numa_add_cpu(int cpu)
827{
828 numa_set_cpumask(cpu, true);
829}
830
831void numa_remove_cpu(int cpu)
832{
833 numa_set_cpumask(cpu, false);
834}
835# endif
836
837
838
839
840const struct cpumask *cpumask_of_node(int node)
841{
842 if (node >= nr_node_ids) {
843 printk(KERN_WARNING
844 "cpumask_of_node(%d): node > nr_node_ids(%d)\n",
845 node, nr_node_ids);
846 dump_stack();
847 return cpu_none_mask;
848 }
849 if (node_to_cpumask_map[node] == NULL) {
850 printk(KERN_WARNING
851 "cpumask_of_node(%d): no node_to_cpumask_map!\n",
852 node);
853 dump_stack();
854 return cpu_online_mask;
855 }
856 return node_to_cpumask_map[node];
857}
858EXPORT_SYMBOL(cpumask_of_node);
859
860#endif
861
862#ifdef CONFIG_MEMORY_HOTPLUG
863int memory_add_physaddr_to_nid(u64 start)
864{
865 struct numa_meminfo *mi = &numa_meminfo;
866 int nid = mi->blk[0].nid;
867 int i;
868
869 for (i = 0; i < mi->nr_blks; i++)
870 if (mi->blk[i].start <= start && mi->blk[i].end > start)
871 nid = mi->blk[i].nid;
872 return nid;
873}
874EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
875#endif
876