Merge tag 'arm64-stable' of git://git.kernel.org/pub/scm/linux/kernel/git/cmarinas...
[linux-block.git] / arch / x86 / mm / numa.c
CommitLineData
71ee73e7 1/* Common code for 32 and 64-bit NUMA */
a4106eae
TH
2#include <linux/kernel.h>
3#include <linux/mm.h>
4#include <linux/string.h>
5#include <linux/init.h>
71ee73e7 6#include <linux/bootmem.h>
a4106eae
TH
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>
90321602 18#include <asm/acpi.h>
a4106eae
TH
19#include <asm/amd_nb.h>
20
21#include "numa_internal.h"
90321602
JB
22
23int __initdata numa_off;
e6df595b 24nodemask_t numa_nodes_parsed __initdata;
90321602 25
a4106eae
TH
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;
a4106eae 37
90321602
JB
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);
71ee73e7 55
71ee73e7 56/*
bbc9e2f4 57 * apicid, cpu, node mappings
71ee73e7 58 */
c4c60524 59s16 __apicid_to_node[MAX_LOCAL_APIC] = {
bbc9e2f4
TH
60 [0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE
61};
62
148f9bb8 63int numa_cpu_node(int cpu)
6bd26273
TH
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
c032ef60 72cpumask_var_t node_to_cpumask_map[MAX_NUMNODES];
71ee73e7
RR
73EXPORT_SYMBOL(node_to_cpumask_map);
74
645a7919
TH
75/*
76 * Map cpu index to node index
77 */
645a7919 78DEFINE_EARLY_PER_CPU(int, x86_cpu_to_node_map, NUMA_NO_NODE);
645a7919
TH
79EXPORT_EARLY_PER_CPU_SYMBOL(x86_cpu_to_node_map);
80
e13fe869 81void numa_set_node(int cpu, int node)
645a7919
TH
82{
83 int *cpu_to_node_map = early_per_cpu_ptr(x86_cpu_to_node_map);
84
85 /* early setting, no percpu area yet */
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
942670d0 100 set_cpu_numa_node(cpu, node);
645a7919
TH
101}
102
e13fe869 103void numa_clear_node(int cpu)
645a7919
TH
104{
105 numa_set_node(cpu, NUMA_NO_NODE);
106}
107
71ee73e7
RR
108/*
109 * Allocate node_to_cpumask_map based on number of available nodes
110 * Requires node_possible_map to be valid.
111 *
9512938b 112 * Note: cpumask_of_node() is not valid until after this is done.
71ee73e7
RR
113 * (Use CONFIG_DEBUG_PER_CPU_MAPS to check this.)
114 */
115void __init setup_node_to_cpumask_map(void)
116{
d2ad351e 117 unsigned int node;
71ee73e7
RR
118
119 /* setup nr_node_ids if not done yet */
d2ad351e
CS
120 if (nr_node_ids == MAX_NUMNODES)
121 setup_nr_node_ids();
71ee73e7
RR
122
123 /* allocate the map */
c032ef60
RR
124 for (node = 0; node < nr_node_ids; node++)
125 alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]);
71ee73e7 126
c032ef60
RR
127 /* cpumask_of_node() will now work */
128 pr_debug("Node to cpumask map for %d nodes\n", nr_node_ids);
71ee73e7
RR
129}
130
a4106eae
TH
131static int __init numa_add_memblk_to(int nid, u64 start, u64 end,
132 struct numa_meminfo *mi)
133{
134 /* ignore zero length blks */
135 if (start == end)
136 return 0;
137
138 /* whine about and ignore invalid blks */
139 if (start > end || nid < 0 || nid >= MAX_NUMNODES) {
365811d6
BH
140 pr_warning("NUMA: Warning: invalid memblk node %d [mem %#010Lx-%#010Lx]\n",
141 nid, start, end - 1);
a4106eae
TH
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 * 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
161 *
162 * Remove @idx'th numa_memblk from @mi by shifting @mi->blk[] and
163 * decrementing @mi->nr_blks.
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 * 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
177 *
178 * Add a new memblk to the default numa_meminfo.
179 *
180 * RETURNS:
181 * 0 on success, -errno on failure.
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
a56bca80
YL
188/* Initialize NODE_DATA for a node on the local memory */
189static void __init setup_node_data(int nid, u64 start, u64 end)
a4106eae 190{
a4106eae 191 const size_t nd_size = roundup(sizeof(pg_data_t), PAGE_SIZE);
38f3e1ca 192 u64 nd_pa;
7888e96b 193 void *nd;
a4106eae
TH
194 int tnid;
195
196 /*
197 * Don't confuse VM with a node that doesn't have the
198 * minimum amount of memory:
199 */
200 if (end && (end - start) < NODE_MIN_SIZE)
201 return;
202
203 start = roundup(start, ZONE_ALIGN);
204
365811d6
BH
205 printk(KERN_INFO "Initmem setup node %d [mem %#010Lx-%#010Lx]\n",
206 nid, start, end - 1);
a4106eae
TH
207
208 /*
07f4207a
PA
209 * Allocate node data. Try node-local memory and then any node.
210 * Never allocate in DMA zone.
a4106eae 211 */
20e6926d 212 nd_pa = memblock_alloc_nid(nd_size, SMP_CACHE_BYTES, nid);
07f4207a 213 if (!nd_pa) {
20e6926d
YL
214 pr_err("Cannot find %zu bytes in node %d\n",
215 nd_size, nid);
07f4207a 216 return;
a4106eae 217 }
07f4207a 218 nd = __va(nd_pa);
a4106eae
TH
219
220 /* report and initialize */
07f4207a
PA
221 printk(KERN_INFO " NODE_DATA [mem %#010Lx-%#010Lx]\n",
222 nd_pa, nd_pa + nd_size - 1);
a4106eae 223 tnid = early_pfn_to_nid(nd_pa >> PAGE_SHIFT);
07f4207a 224 if (tnid != nid)
a4106eae
TH
225 printk(KERN_INFO " NODE_DATA(%d) on node %d\n", nid, tnid);
226
7888e96b 227 node_data[nid] = nd;
a4106eae
TH
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 * numa_cleanup_meminfo - Cleanup a numa_meminfo
238 * @mi: numa_meminfo to clean up
239 *
240 * Sanitize @mi by merging and removing unncessary memblks. Also check for
241 * conflicts and clear unused memblks.
242 *
243 * RETURNS:
244 * 0 on success, -errno on failure.
245 */
246int __init numa_cleanup_meminfo(struct numa_meminfo *mi)
247{
248 const u64 low = 0;
38f3e1ca 249 const u64 high = PFN_PHYS(max_pfn);
a4106eae
TH
250 int i, j, k;
251
e5a10c1b 252 /* first, trim all entries */
a4106eae
TH
253 for (i = 0; i < mi->nr_blks; i++) {
254 struct numa_memblk *bi = &mi->blk[i];
255
256 /* make sure all blocks are inside the limits */
257 bi->start = max(bi->start, low);
258 bi->end = min(bi->end, high);
259
260 /* and there's no empty block */
e5a10c1b 261 if (bi->start >= bi->end)
a4106eae 262 numa_remove_memblk_from(i--, mi);
e5a10c1b
YL
263 }
264
265 /* merge neighboring / overlapping entries */
266 for (i = 0; i < mi->nr_blks; i++) {
267 struct numa_memblk *bi = &mi->blk[i];
a4106eae
TH
268
269 for (j = i + 1; j < mi->nr_blks; j++) {
270 struct numa_memblk *bj = &mi->blk[j];
38f3e1ca 271 u64 start, end;
a4106eae
TH
272
273 /*
274 * See whether there are overlapping blocks. Whine
275 * about but allow overlaps of the same nid. They
276 * will be merged below.
277 */
278 if (bi->end > bj->start && bi->start < bj->end) {
279 if (bi->nid != bj->nid) {
365811d6
BH
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);
a4106eae
TH
283 return -EINVAL;
284 }
365811d6
BH
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);
a4106eae
TH
288 }
289
290 /*
291 * Join together blocks on the same node, holes
292 * between which don't overlap with memory on other
293 * nodes.
294 */
295 if (bi->nid != bj->nid)
296 continue;
e5a10c1b
YL
297 start = min(bi->start, bj->start);
298 end = max(bi->end, bj->end);
a4106eae
TH
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;
365811d6
BH
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);
a4106eae
TH
312 bi->start = start;
313 bi->end = end;
314 numa_remove_memblk_from(j--, mi);
315 }
316 }
317
e5a10c1b 318 /* clear unused ones */
a4106eae
TH
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 * Set nodes, which have memory in @mi, in *@nodemask.
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 * numa_reset_distance - Reset NUMA distance table
343 *
344 * The current table is freed. The next numa_set_distance() call will
345 * create a new one.
346 */
347void __init numa_reset_distance(void)
348{
349 size_t size = numa_distance_cnt * numa_distance_cnt * sizeof(numa_distance[0]);
350
351 /* numa_distance could be 1LU marking allocation failure, test cnt */
352 if (numa_distance_cnt)
24aa0788 353 memblock_free(__pa(numa_distance), size);
a4106eae
TH
354 numa_distance_cnt = 0;
355 numa_distance = NULL; /* enable table creation */
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 /* size the new table and allocate it */
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
38f3e1ca 374 phys = memblock_find_in_range(0, PFN_PHYS(max_pfn_mapped),
a4106eae 375 size, PAGE_SIZE);
1f5026a7 376 if (!phys) {
a4106eae
TH
377 pr_warning("NUMA: Warning: can't allocate distance table!\n");
378 /* don't retry until explicitly reset */
379 numa_distance = (void *)1LU;
380 return -ENOMEM;
381 }
24aa0788 382 memblock_reserve(phys, size);
a4106eae
TH
383
384 numa_distance = __va(phys);
385 numa_distance_cnt = cnt;
386
387 /* fill with the default distances */
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 * numa_set_distance - Set NUMA distance from one NUMA to another
399 * @from: the 'from' node to set distance
400 * @to: the 'to' node to set distance
401 * @distance: NUMA distance
402 *
403 * Set the distance from node @from to @to to @distance. If distance table
404 * doesn't exist, one which is large enough to accommodate all the currently
405 * known nodes will be created.
406 *
407 * If such table cannot be allocated, a warning is printed and further
408 * calls are ignored until the distance table is reset with
409 * numa_reset_distance().
410 *
54eed6cb
PH
411 * If @from or @to is higher than the highest known node or lower than zero
412 * at the time of table creation or @distance doesn't make sense, the call
413 * is ignored.
a4106eae
TH
414 * This is to allow simplification of specific NUMA config implementations.
415 */
416void __init numa_set_distance(int from, int to, int distance)
417{
418 if (!numa_distance && numa_alloc_distance() < 0)
419 return;
420
54eed6cb
PH
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",
a4106eae
TH
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 * Sanity check to catch more bad NUMA configurations (they are amazingly
448 * common). Make sure the nodes cover all memory.
449 */
450static bool __init numa_meminfo_cover_memory(const struct numa_meminfo *mi)
451{
38f3e1ca 452 u64 numaram, e820ram;
a4106eae
TH
453 int i;
454
455 numaram = 0;
456 for (i = 0; i < mi->nr_blks; i++) {
38f3e1ca
TH
457 u64 s = mi->blk[i].start >> PAGE_SHIFT;
458 u64 e = mi->blk[i].end >> PAGE_SHIFT;
a4106eae
TH
459 numaram += e - s;
460 numaram -= __absent_pages_in_range(mi->blk[i].nid, s, e);
38f3e1ca 461 if ((s64)numaram < 0)
a4106eae
TH
462 numaram = 0;
463 }
464
474b881b
TH
465 e820ram = max_pfn - absent_pages_in_range(0, max_pfn);
466
a4106eae 467 /* We seem to lose 3 pages somewhere. Allow 1M of slack. */
38f3e1ca
TH
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",
a4106eae
TH
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{
1e01979c 479 unsigned long uninitialized_var(pfn_align);
a4106eae
TH
480 int i, nid;
481
482 /* Account for nodes with cpus and no memory */
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
0608f70c
TH
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 }
1e01979c
TH
492
493 /*
494 * If sections array is gonna be used for pfn -> nid mapping, check
495 * whether its granularity is fine enough.
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
a4106eae
TH
506 if (!numa_meminfo_cover_memory(mi))
507 return -EINVAL;
508
509 /* Finally register nodes. */
510 for_each_node_mask(nid, node_possible_map) {
38f3e1ca 511 u64 start = PFN_PHYS(max_pfn);
a4106eae
TH
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)
a56bca80 522 setup_node_data(nid, start, end);
a4106eae
TH
523 }
524
0608f70c
TH
525 /* Dump memblock with node info and return. */
526 memblock_dump_all();
a4106eae
TH
527 return 0;
528}
a4106eae 529
8db78cc4
TH
530/*
531 * There are unfortunately some poorly designed mainboards around that
532 * only connect memory to a single CPU. This breaks the 1:1 cpu->node
533 * mapping. To avoid this fill in the mapping for all possible CPUs,
534 * as the number of CPUs is not known yet. We round robin the existing
535 * nodes.
536 */
752d4f37 537static void __init numa_init_array(void)
8db78cc4
TH
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
a4106eae
TH
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
20e6926d 560 nodes_clear(numa_nodes_parsed);
a4106eae
TH
561 nodes_clear(node_possible_map);
562 nodes_clear(node_online_map);
20e6926d 563 memset(&numa_meminfo, 0, sizeof(numa_meminfo));
0608f70c 564 WARN_ON(memblock_set_node(0, ULLONG_MAX, MAX_NUMNODES));
a4106eae
TH
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 * dummy_numa_init - Fallback dummy NUMA init
594 *
595 * Used if there's no underlying NUMA architecture, NUMA initialization
596 * fails, or NUMA is disabled on the command line.
597 *
598 * Must online at least one node and add memory blocks that cover all
599 * allowed memory. This function must not fail.
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");
365811d6
BH
605 printk(KERN_INFO "Faking a node at [mem %#018Lx-%#018Lx]\n",
606 0LLU, PFN_PHYS(max_pfn) - 1);
a4106eae
TH
607
608 node_set(0, numa_nodes_parsed);
38f3e1ca 609 numa_add_memblk(0, 0, PFN_PHYS(max_pfn));
a4106eae
TH
610
611 return 0;
612}
613
614/**
615 * x86_numa_init - Initialize NUMA
616 *
617 * Try each configured NUMA initialization method until one succeeds. The
618 * last fallback is dummy single node config encomapssing whole memory and
619 * never fails.
620 */
621void __init x86_numa_init(void)
622{
623 if (!numa_off) {
bd6709a9
TH
624#ifdef CONFIG_X86_NUMAQ
625 if (!numa_init(numaq_numa_init))
626 return;
627#endif
a4106eae
TH
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}
a4106eae 640
8db78cc4
TH
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 * Setup early cpu_to_node.
661 *
662 * Populate cpu_to_node[] only if x86_cpu_to_apicid[],
663 * and apicid_to_node[] tables have valid entries for a CPU.
664 * This means we skip cpu_to_node[] initialisation for NUMA
665 * emulation and faking node case (when running a kernel compiled
666 * for NUMA on a non NUMA box), which is OK as cpu_to_node[]
667 * is already initialized in a round robin manner at numa_init_array,
668 * prior to this call, and this initialization is good enough
669 * for the fake NUMA cases.
670 *
671 * Called before the per_cpu areas are setup.
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
de2d9445
TH
691#ifndef CONFIG_DEBUG_PER_CPU_MAPS
692
693# ifndef CONFIG_NUMA_EMU
148f9bb8 694void numa_add_cpu(int cpu)
de2d9445
TH
695{
696 cpumask_set_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
697}
698
148f9bb8 699void numa_remove_cpu(int cpu)
de2d9445
TH
700{
701 cpumask_clear_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
702}
703# endif /* !CONFIG_NUMA_EMU */
704
705#else /* !CONFIG_DEBUG_PER_CPU_MAPS */
645a7919
TH
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 * Same function as cpu_to_node() but used if called before the
721 * per_cpu areas are setup.
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
7a6c6547 737void debug_cpumask_set_cpu(int cpu, int node, bool enable)
de2d9445 738{
de2d9445
TH
739 struct cpumask *mask;
740 char buf[64];
741
14392fd3
DR
742 if (node == NUMA_NO_NODE) {
743 /* early_cpu_to_node() already emits a warning and trace */
7a6c6547 744 return;
14392fd3 745 }
de2d9445
TH
746 mask = node_to_cpumask_map[node];
747 if (!mask) {
748 pr_err("node_to_cpumask_map[%i] NULL\n", node);
749 dump_stack();
7a6c6547 750 return;
de2d9445
TH
751 }
752
7a6c6547
DR
753 if (enable)
754 cpumask_set_cpu(cpu, mask);
755 else
756 cpumask_clear_cpu(cpu, mask);
757
de2d9445
TH
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);
7a6c6547 762 return;
de2d9445
TH
763}
764
765# ifndef CONFIG_NUMA_EMU
148f9bb8 766static void numa_set_cpumask(int cpu, bool enable)
de2d9445 767{
7a6c6547 768 debug_cpumask_set_cpu(cpu, early_cpu_to_node(cpu), enable);
de2d9445
TH
769}
770
148f9bb8 771void numa_add_cpu(int cpu)
de2d9445 772{
7a6c6547 773 numa_set_cpumask(cpu, true);
de2d9445
TH
774}
775
148f9bb8 776void numa_remove_cpu(int cpu)
de2d9445 777{
7a6c6547 778 numa_set_cpumask(cpu, false);
de2d9445
TH
779}
780# endif /* !CONFIG_NUMA_EMU */
781
71ee73e7
RR
782/*
783 * Returns a pointer to the bitmask of CPUs on Node 'node'.
784 */
73e907de 785const struct cpumask *cpumask_of_node(int node)
71ee73e7 786{
71ee73e7
RR
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 }
c032ef60
RR
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 }
0b966252 801 return node_to_cpumask_map[node];
71ee73e7
RR
802}
803EXPORT_SYMBOL(cpumask_of_node);
645a7919 804
de2d9445 805#endif /* !CONFIG_DEBUG_PER_CPU_MAPS */
a4106eae 806
bd6709a9 807#ifdef CONFIG_MEMORY_HOTPLUG
a4106eae
TH
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