Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/* Common code for 32 and 64-bit NUMA */
3#include <linux/acpi.h>
4#include <linux/kernel.h>
5#include <linux/mm.h>
6#include <linux/of.h>
7#include <linux/string.h>
8#include <linux/init.h>
9#include <linux/memblock.h>
10#include <linux/mmzone.h>
11#include <linux/ctype.h>
12#include <linux/nodemask.h>
13#include <linux/sched.h>
14#include <linux/topology.h>
15#include <linux/sort.h>
16
17#include <asm/e820/api.h>
18#include <asm/proto.h>
19#include <asm/dma.h>
20#include <asm/amd_nb.h>
21
22#include "numa_internal.h"
23
24int numa_off;
25nodemask_t numa_nodes_parsed __initdata;
26
27struct pglist_data *node_data[MAX_NUMNODES] __read_mostly;
28EXPORT_SYMBOL(node_data);
29
30static struct numa_meminfo numa_meminfo __initdata_or_meminfo;
31static struct numa_meminfo numa_reserved_meminfo __initdata_or_meminfo;
32
33static int numa_distance_cnt;
34static u8 *numa_distance;
35
36static __init int numa_setup(char *opt)
37{
38 if (!opt)
39 return -EINVAL;
40 if (!strncmp(opt, "off", 3))
41 numa_off = 1;
42 if (!strncmp(opt, "fake=", 5))
43 return numa_emu_cmdline(opt + 5);
44 if (!strncmp(opt, "noacpi", 6))
45 disable_srat();
46 if (!strncmp(opt, "nohmat", 6))
47 disable_hmat();
48 return 0;
49}
50early_param("numa", numa_setup);
51
52/*
53 * apicid, cpu, node mappings
54 */
55s16 __apicid_to_node[MAX_LOCAL_APIC] = {
56 [0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE
57};
58
59int numa_cpu_node(int cpu)
60{
61 u32 apicid = early_per_cpu(x86_cpu_to_apicid, cpu);
62
63 if (apicid != BAD_APICID)
64 return __apicid_to_node[apicid];
65 return NUMA_NO_NODE;
66}
67
68cpumask_var_t node_to_cpumask_map[MAX_NUMNODES];
69EXPORT_SYMBOL(node_to_cpumask_map);
70
71/*
72 * Map cpu index to node index
73 */
74DEFINE_EARLY_PER_CPU(int, x86_cpu_to_node_map, NUMA_NO_NODE);
75EXPORT_EARLY_PER_CPU_SYMBOL(x86_cpu_to_node_map);
76
77void numa_set_node(int cpu, int node)
78{
79 int *cpu_to_node_map = early_per_cpu_ptr(x86_cpu_to_node_map);
80
81 /* early setting, no percpu area yet */
82 if (cpu_to_node_map) {
83 cpu_to_node_map[cpu] = node;
84 return;
85 }
86
87#ifdef CONFIG_DEBUG_PER_CPU_MAPS
88 if (cpu >= nr_cpu_ids || !cpu_possible(cpu)) {
89 printk(KERN_ERR "numa_set_node: invalid cpu# (%d)\n", cpu);
90 dump_stack();
91 return;
92 }
93#endif
94 per_cpu(x86_cpu_to_node_map, cpu) = node;
95
96 set_cpu_numa_node(cpu, node);
97}
98
99void numa_clear_node(int cpu)
100{
101 numa_set_node(cpu, NUMA_NO_NODE);
102}
103
104/*
105 * Allocate node_to_cpumask_map based on number of available nodes
106 * Requires node_possible_map to be valid.
107 *
108 * Note: cpumask_of_node() is not valid until after this is done.
109 * (Use CONFIG_DEBUG_PER_CPU_MAPS to check this.)
110 */
111void __init setup_node_to_cpumask_map(void)
112{
113 unsigned int node;
114
115 /* setup nr_node_ids if not done yet */
116 if (nr_node_ids == MAX_NUMNODES)
117 setup_nr_node_ids();
118
119 /* allocate the map */
120 for (node = 0; node < nr_node_ids; node++)
121 alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]);
122
123 /* cpumask_of_node() will now work */
124 pr_debug("Node to cpumask map for %u nodes\n", nr_node_ids);
125}
126
127static int __init numa_add_memblk_to(int nid, u64 start, u64 end,
128 struct numa_meminfo *mi)
129{
130 /* ignore zero length blks */
131 if (start == end)
132 return 0;
133
134 /* whine about and ignore invalid blks */
135 if (start > end || nid < 0 || nid >= MAX_NUMNODES) {
136 pr_warn("Warning: invalid memblk node %d [mem %#010Lx-%#010Lx]\n",
137 nid, start, end - 1);
138 return 0;
139 }
140
141 if (mi->nr_blks >= NR_NODE_MEMBLKS) {
142 pr_err("too many memblk ranges\n");
143 return -EINVAL;
144 }
145
146 mi->blk[mi->nr_blks].start = start;
147 mi->blk[mi->nr_blks].end = end;
148 mi->blk[mi->nr_blks].nid = nid;
149 mi->nr_blks++;
150 return 0;
151}
152
153/**
154 * numa_remove_memblk_from - Remove one numa_memblk from a numa_meminfo
155 * @idx: Index of memblk to remove
156 * @mi: numa_meminfo to remove memblk from
157 *
158 * Remove @idx'th numa_memblk from @mi by shifting @mi->blk[] and
159 * decrementing @mi->nr_blks.
160 */
161void __init numa_remove_memblk_from(int idx, struct numa_meminfo *mi)
162{
163 mi->nr_blks--;
164 memmove(&mi->blk[idx], &mi->blk[idx + 1],
165 (mi->nr_blks - idx) * sizeof(mi->blk[0]));
166}
167
168/**
169 * numa_move_tail_memblk - Move a numa_memblk from one numa_meminfo to another
170 * @dst: numa_meminfo to append block to
171 * @idx: Index of memblk to remove
172 * @src: numa_meminfo to remove memblk from
173 */
174static void __init numa_move_tail_memblk(struct numa_meminfo *dst, int idx,
175 struct numa_meminfo *src)
176{
177 dst->blk[dst->nr_blks++] = src->blk[idx];
178 numa_remove_memblk_from(idx, src);
179}
180
181/**
182 * numa_add_memblk - Add one numa_memblk to numa_meminfo
183 * @nid: NUMA node ID of the new memblk
184 * @start: Start address of the new memblk
185 * @end: End address of the new memblk
186 *
187 * Add a new memblk to the default numa_meminfo.
188 *
189 * RETURNS:
190 * 0 on success, -errno on failure.
191 */
192int __init numa_add_memblk(int nid, u64 start, u64 end)
193{
194 return numa_add_memblk_to(nid, start, end, &numa_meminfo);
195}
196
197/* Allocate NODE_DATA for a node on the local memory */
198static void __init alloc_node_data(int nid)
199{
200 const size_t nd_size = roundup(sizeof(pg_data_t), PAGE_SIZE);
201 u64 nd_pa;
202 void *nd;
203 int tnid;
204
205 /*
206 * Allocate node data. Try node-local memory and then any node.
207 * Never allocate in DMA zone.
208 */
209 nd_pa = memblock_phys_alloc_try_nid(nd_size, SMP_CACHE_BYTES, nid);
210 if (!nd_pa) {
211 pr_err("Cannot find %zu bytes in any node (initial node: %d)\n",
212 nd_size, nid);
213 return;
214 }
215 nd = __va(nd_pa);
216
217 /* report and initialize */
218 printk(KERN_INFO "NODE_DATA(%d) allocated [mem %#010Lx-%#010Lx]\n", nid,
219 nd_pa, nd_pa + nd_size - 1);
220 tnid = early_pfn_to_nid(nd_pa >> PAGE_SHIFT);
221 if (tnid != nid)
222 printk(KERN_INFO " NODE_DATA(%d) on node %d\n", nid, tnid);
223
224 node_data[nid] = nd;
225 memset(NODE_DATA(nid), 0, sizeof(pg_data_t));
226
227 node_set_online(nid);
228}
229
230/**
231 * numa_cleanup_meminfo - Cleanup a numa_meminfo
232 * @mi: numa_meminfo to clean up
233 *
234 * Sanitize @mi by merging and removing unnecessary memblks. Also check for
235 * conflicts and clear unused memblks.
236 *
237 * RETURNS:
238 * 0 on success, -errno on failure.
239 */
240int __init numa_cleanup_meminfo(struct numa_meminfo *mi)
241{
242 const u64 low = 0;
243 const u64 high = PFN_PHYS(max_pfn);
244 int i, j, k;
245
246 /* first, trim all entries */
247 for (i = 0; i < mi->nr_blks; i++) {
248 struct numa_memblk *bi = &mi->blk[i];
249
250 /* move / save reserved memory ranges */
251 if (!memblock_overlaps_region(&memblock.memory,
252 bi->start, bi->end - bi->start)) {
253 numa_move_tail_memblk(&numa_reserved_meminfo, i--, mi);
254 continue;
255 }
256
257 /* make sure all non-reserved blocks are inside the limits */
258 bi->start = max(bi->start, low);
259
260 /* preserve info for non-RAM areas above 'max_pfn': */
261 if (bi->end > high) {
262 numa_add_memblk_to(bi->nid, high, bi->end,
263 &numa_reserved_meminfo);
264 bi->end = high;
265 }
266
267 /* and there's no empty block */
268 if (bi->start >= bi->end)
269 numa_remove_memblk_from(i--, mi);
270 }
271
272 /* merge neighboring / overlapping entries */
273 for (i = 0; i < mi->nr_blks; i++) {
274 struct numa_memblk *bi = &mi->blk[i];
275
276 for (j = i + 1; j < mi->nr_blks; j++) {
277 struct numa_memblk *bj = &mi->blk[j];
278 u64 start, end;
279
280 /*
281 * See whether there are overlapping blocks. Whine
282 * about but allow overlaps of the same nid. They
283 * will be merged below.
284 */
285 if (bi->end > bj->start && bi->start < bj->end) {
286 if (bi->nid != bj->nid) {
287 pr_err("node %d [mem %#010Lx-%#010Lx] overlaps with node %d [mem %#010Lx-%#010Lx]\n",
288 bi->nid, bi->start, bi->end - 1,
289 bj->nid, bj->start, bj->end - 1);
290 return -EINVAL;
291 }
292 pr_warn("Warning: node %d [mem %#010Lx-%#010Lx] overlaps with itself [mem %#010Lx-%#010Lx]\n",
293 bi->nid, bi->start, bi->end - 1,
294 bj->start, bj->end - 1);
295 }
296
297 /*
298 * Join together blocks on the same node, holes
299 * between which don't overlap with memory on other
300 * nodes.
301 */
302 if (bi->nid != bj->nid)
303 continue;
304 start = min(bi->start, bj->start);
305 end = max(bi->end, bj->end);
306 for (k = 0; k < mi->nr_blks; k++) {
307 struct numa_memblk *bk = &mi->blk[k];
308
309 if (bi->nid == bk->nid)
310 continue;
311 if (start < bk->end && end > bk->start)
312 break;
313 }
314 if (k < mi->nr_blks)
315 continue;
316 printk(KERN_INFO "NUMA: Node %d [mem %#010Lx-%#010Lx] + [mem %#010Lx-%#010Lx] -> [mem %#010Lx-%#010Lx]\n",
317 bi->nid, bi->start, bi->end - 1, bj->start,
318 bj->end - 1, start, end - 1);
319 bi->start = start;
320 bi->end = end;
321 numa_remove_memblk_from(j--, mi);
322 }
323 }
324
325 /* clear unused ones */
326 for (i = mi->nr_blks; i < ARRAY_SIZE(mi->blk); i++) {
327 mi->blk[i].start = mi->blk[i].end = 0;
328 mi->blk[i].nid = NUMA_NO_NODE;
329 }
330
331 return 0;
332}
333
334/*
335 * Set nodes, which have memory in @mi, in *@nodemask.
336 */
337static void __init numa_nodemask_from_meminfo(nodemask_t *nodemask,
338 const struct numa_meminfo *mi)
339{
340 int i;
341
342 for (i = 0; i < ARRAY_SIZE(mi->blk); i++)
343 if (mi->blk[i].start != mi->blk[i].end &&
344 mi->blk[i].nid != NUMA_NO_NODE)
345 node_set(mi->blk[i].nid, *nodemask);
346}
347
348/**
349 * numa_reset_distance - Reset NUMA distance table
350 *
351 * The current table is freed. The next numa_set_distance() call will
352 * create a new one.
353 */
354void __init numa_reset_distance(void)
355{
356 size_t size = numa_distance_cnt * numa_distance_cnt * sizeof(numa_distance[0]);
357
358 /* numa_distance could be 1LU marking allocation failure, test cnt */
359 if (numa_distance_cnt)
360 memblock_free(numa_distance, size);
361 numa_distance_cnt = 0;
362 numa_distance = NULL; /* enable table creation */
363}
364
365static int __init numa_alloc_distance(void)
366{
367 nodemask_t nodes_parsed;
368 size_t size;
369 int i, j, cnt = 0;
370 u64 phys;
371
372 /* size the new table and allocate it */
373 nodes_parsed = numa_nodes_parsed;
374 numa_nodemask_from_meminfo(&nodes_parsed, &numa_meminfo);
375
376 for_each_node_mask(i, nodes_parsed)
377 cnt = i;
378 cnt++;
379 size = cnt * cnt * sizeof(numa_distance[0]);
380
381 phys = memblock_phys_alloc_range(size, PAGE_SIZE, 0,
382 PFN_PHYS(max_pfn_mapped));
383 if (!phys) {
384 pr_warn("Warning: can't allocate distance table!\n");
385 /* don't retry until explicitly reset */
386 numa_distance = (void *)1LU;
387 return -ENOMEM;
388 }
389
390 numa_distance = __va(phys);
391 numa_distance_cnt = cnt;
392
393 /* fill with the default distances */
394 for (i = 0; i < cnt; i++)
395 for (j = 0; j < cnt; j++)
396 numa_distance[i * cnt + j] = i == j ?
397 LOCAL_DISTANCE : REMOTE_DISTANCE;
398 printk(KERN_DEBUG "NUMA: Initialized distance table, cnt=%d\n", cnt);
399
400 return 0;
401}
402
403/**
404 * numa_set_distance - Set NUMA distance from one NUMA to another
405 * @from: the 'from' node to set distance
406 * @to: the 'to' node to set distance
407 * @distance: NUMA distance
408 *
409 * Set the distance from node @from to @to to @distance. If distance table
410 * doesn't exist, one which is large enough to accommodate all the currently
411 * known nodes will be created.
412 *
413 * If such table cannot be allocated, a warning is printed and further
414 * calls are ignored until the distance table is reset with
415 * numa_reset_distance().
416 *
417 * If @from or @to is higher than the highest known node or lower than zero
418 * at the time of table creation or @distance doesn't make sense, the call
419 * is ignored.
420 * This is to allow simplification of specific NUMA config implementations.
421 */
422void __init numa_set_distance(int from, int to, int distance)
423{
424 if (!numa_distance && numa_alloc_distance() < 0)
425 return;
426
427 if (from >= numa_distance_cnt || to >= numa_distance_cnt ||
428 from < 0 || to < 0) {
429 pr_warn_once("Warning: node ids are out of bound, from=%d to=%d distance=%d\n",
430 from, to, distance);
431 return;
432 }
433
434 if ((u8)distance != distance ||
435 (from == to && distance != LOCAL_DISTANCE)) {
436 pr_warn_once("Warning: invalid distance parameter, from=%d to=%d distance=%d\n",
437 from, to, distance);
438 return;
439 }
440
441 numa_distance[from * numa_distance_cnt + to] = distance;
442}
443
444int __node_distance(int from, int to)
445{
446 if (from >= numa_distance_cnt || to >= numa_distance_cnt)
447 return from == to ? LOCAL_DISTANCE : REMOTE_DISTANCE;
448 return numa_distance[from * numa_distance_cnt + to];
449}
450EXPORT_SYMBOL(__node_distance);
451
452/*
453 * Mark all currently memblock-reserved physical memory (which covers the
454 * kernel's own memory ranges) as hot-unswappable.
455 */
456static void __init numa_clear_kernel_node_hotplug(void)
457{
458 nodemask_t reserved_nodemask = NODE_MASK_NONE;
459 struct memblock_region *mb_region;
460 int i;
461
462 /*
463 * We have to do some preprocessing of memblock regions, to
464 * make them suitable for reservation.
465 *
466 * At this time, all memory regions reserved by memblock are
467 * used by the kernel, but those regions are not split up
468 * along node boundaries yet, and don't necessarily have their
469 * node ID set yet either.
470 *
471 * So iterate over all memory known to the x86 architecture,
472 * and use those ranges to set the nid in memblock.reserved.
473 * This will split up the memblock regions along node
474 * boundaries and will set the node IDs as well.
475 */
476 for (i = 0; i < numa_meminfo.nr_blks; i++) {
477 struct numa_memblk *mb = numa_meminfo.blk + i;
478 int ret;
479
480 ret = memblock_set_node(mb->start, mb->end - mb->start, &memblock.reserved, mb->nid);
481 WARN_ON_ONCE(ret);
482 }
483
484 /*
485 * Now go over all reserved memblock regions, to construct a
486 * node mask of all kernel reserved memory areas.
487 *
488 * [ Note, when booting with mem=nn[kMG] or in a kdump kernel,
489 * numa_meminfo might not include all memblock.reserved
490 * memory ranges, because quirks such as trim_snb_memory()
491 * reserve specific pages for Sandy Bridge graphics. ]
492 */
493 for_each_reserved_mem_region(mb_region) {
494 int nid = memblock_get_region_node(mb_region);
495
496 if (nid != MAX_NUMNODES)
497 node_set(nid, reserved_nodemask);
498 }
499
500 /*
501 * Finally, clear the MEMBLOCK_HOTPLUG flag for all memory
502 * belonging to the reserved node mask.
503 *
504 * Note that this will include memory regions that reside
505 * on nodes that contain kernel memory - entire nodes
506 * become hot-unpluggable:
507 */
508 for (i = 0; i < numa_meminfo.nr_blks; i++) {
509 struct numa_memblk *mb = numa_meminfo.blk + i;
510
511 if (!node_isset(mb->nid, reserved_nodemask))
512 continue;
513
514 memblock_clear_hotplug(mb->start, mb->end - mb->start);
515 }
516}
517
518static int __init numa_register_memblks(struct numa_meminfo *mi)
519{
520 int i, nid;
521
522 /* Account for nodes with cpus and no memory */
523 node_possible_map = numa_nodes_parsed;
524 numa_nodemask_from_meminfo(&node_possible_map, mi);
525 if (WARN_ON(nodes_empty(node_possible_map)))
526 return -EINVAL;
527
528 for (i = 0; i < mi->nr_blks; i++) {
529 struct numa_memblk *mb = &mi->blk[i];
530 memblock_set_node(mb->start, mb->end - mb->start,
531 &memblock.memory, mb->nid);
532 }
533
534 /*
535 * At very early time, the kernel have to use some memory such as
536 * loading the kernel image. We cannot prevent this anyway. So any
537 * node the kernel resides in should be un-hotpluggable.
538 *
539 * And when we come here, alloc node data won't fail.
540 */
541 numa_clear_kernel_node_hotplug();
542
543 /*
544 * If sections array is gonna be used for pfn -> nid mapping, check
545 * whether its granularity is fine enough.
546 */
547 if (IS_ENABLED(NODE_NOT_IN_PAGE_FLAGS)) {
548 unsigned long pfn_align = node_map_pfn_alignment();
549
550 if (pfn_align && pfn_align < PAGES_PER_SECTION) {
551 pr_warn("Node alignment %LuMB < min %LuMB, rejecting NUMA config\n",
552 PFN_PHYS(pfn_align) >> 20,
553 PFN_PHYS(PAGES_PER_SECTION) >> 20);
554 return -EINVAL;
555 }
556 }
557
558 if (!memblock_validate_numa_coverage(SZ_1M))
559 return -EINVAL;
560
561 /* Finally register nodes. */
562 for_each_node_mask(nid, node_possible_map) {
563 u64 start = PFN_PHYS(max_pfn);
564 u64 end = 0;
565
566 for (i = 0; i < mi->nr_blks; i++) {
567 if (nid != mi->blk[i].nid)
568 continue;
569 start = min(mi->blk[i].start, start);
570 end = max(mi->blk[i].end, end);
571 }
572
573 if (start >= end)
574 continue;
575
576 alloc_node_data(nid);
577 }
578
579 /* Dump memblock with node info and return. */
580 memblock_dump_all();
581 return 0;
582}
583
584/*
585 * There are unfortunately some poorly designed mainboards around that
586 * only connect memory to a single CPU. This breaks the 1:1 cpu->node
587 * mapping. To avoid this fill in the mapping for all possible CPUs,
588 * as the number of CPUs is not known yet. We round robin the existing
589 * nodes.
590 */
591static void __init numa_init_array(void)
592{
593 int rr, i;
594
595 rr = first_node(node_online_map);
596 for (i = 0; i < nr_cpu_ids; i++) {
597 if (early_cpu_to_node(i) != NUMA_NO_NODE)
598 continue;
599 numa_set_node(i, rr);
600 rr = next_node_in(rr, node_online_map);
601 }
602}
603
604static int __init numa_init(int (*init_func)(void))
605{
606 int i;
607 int ret;
608
609 for (i = 0; i < MAX_LOCAL_APIC; i++)
610 set_apicid_to_node(i, NUMA_NO_NODE);
611
612 nodes_clear(numa_nodes_parsed);
613 nodes_clear(node_possible_map);
614 nodes_clear(node_online_map);
615 memset(&numa_meminfo, 0, sizeof(numa_meminfo));
616 WARN_ON(memblock_set_node(0, ULLONG_MAX, &memblock.memory,
617 MAX_NUMNODES));
618 WARN_ON(memblock_set_node(0, ULLONG_MAX, &memblock.reserved,
619 MAX_NUMNODES));
620 /* In case that parsing SRAT failed. */
621 WARN_ON(memblock_clear_hotplug(0, ULLONG_MAX));
622 numa_reset_distance();
623
624 ret = init_func();
625 if (ret < 0)
626 return ret;
627
628 /*
629 * We reset memblock back to the top-down direction
630 * here because if we configured ACPI_NUMA, we have
631 * parsed SRAT in init_func(). It is ok to have the
632 * reset here even if we did't configure ACPI_NUMA
633 * or acpi numa init fails and fallbacks to dummy
634 * numa init.
635 */
636 memblock_set_bottom_up(false);
637
638 ret = numa_cleanup_meminfo(&numa_meminfo);
639 if (ret < 0)
640 return ret;
641
642 numa_emulation(&numa_meminfo, numa_distance_cnt);
643
644 ret = numa_register_memblks(&numa_meminfo);
645 if (ret < 0)
646 return ret;
647
648 for (i = 0; i < nr_cpu_ids; i++) {
649 int nid = early_cpu_to_node(i);
650
651 if (nid == NUMA_NO_NODE)
652 continue;
653 if (!node_online(nid))
654 numa_clear_node(i);
655 }
656 numa_init_array();
657
658 return 0;
659}
660
661/**
662 * dummy_numa_init - Fallback dummy NUMA init
663 *
664 * Used if there's no underlying NUMA architecture, NUMA initialization
665 * fails, or NUMA is disabled on the command line.
666 *
667 * Must online at least one node and add memory blocks that cover all
668 * allowed memory. This function must not fail.
669 */
670static int __init dummy_numa_init(void)
671{
672 printk(KERN_INFO "%s\n",
673 numa_off ? "NUMA turned off" : "No NUMA configuration found");
674 printk(KERN_INFO "Faking a node at [mem %#018Lx-%#018Lx]\n",
675 0LLU, PFN_PHYS(max_pfn) - 1);
676
677 node_set(0, numa_nodes_parsed);
678 numa_add_memblk(0, 0, PFN_PHYS(max_pfn));
679
680 return 0;
681}
682
683/**
684 * x86_numa_init - Initialize NUMA
685 *
686 * Try each configured NUMA initialization method until one succeeds. The
687 * last fallback is dummy single node config encompassing whole memory and
688 * never fails.
689 */
690void __init x86_numa_init(void)
691{
692 if (!numa_off) {
693#ifdef CONFIG_ACPI_NUMA
694 if (!numa_init(x86_acpi_numa_init))
695 return;
696#endif
697#ifdef CONFIG_AMD_NUMA
698 if (!numa_init(amd_numa_init))
699 return;
700#endif
701 if (acpi_disabled && !numa_init(of_numa_init))
702 return;
703 }
704
705 numa_init(dummy_numa_init);
706}
707
708
709/*
710 * A node may exist which has one or more Generic Initiators but no CPUs and no
711 * memory.
712 *
713 * This function must be called after init_cpu_to_node(), to ensure that any
714 * memoryless CPU nodes have already been brought online, and before the
715 * node_data[nid] is needed for zone list setup in build_all_zonelists().
716 *
717 * When this function is called, any nodes containing either memory and/or CPUs
718 * will already be online and there is no need to do anything extra, even if
719 * they also contain one or more Generic Initiators.
720 */
721void __init init_gi_nodes(void)
722{
723 int nid;
724
725 /*
726 * Exclude this node from
727 * bringup_nonboot_cpus
728 * cpu_up
729 * __try_online_node
730 * register_one_node
731 * because node_subsys is not initialized yet.
732 * TODO remove dependency on node_online
733 */
734 for_each_node_state(nid, N_GENERIC_INITIATOR)
735 if (!node_online(nid))
736 node_set_online(nid);
737}
738
739/*
740 * Setup early cpu_to_node.
741 *
742 * Populate cpu_to_node[] only if x86_cpu_to_apicid[],
743 * and apicid_to_node[] tables have valid entries for a CPU.
744 * This means we skip cpu_to_node[] initialisation for NUMA
745 * emulation and faking node case (when running a kernel compiled
746 * for NUMA on a non NUMA box), which is OK as cpu_to_node[]
747 * is already initialized in a round robin manner at numa_init_array,
748 * prior to this call, and this initialization is good enough
749 * for the fake NUMA cases.
750 *
751 * Called before the per_cpu areas are setup.
752 */
753void __init init_cpu_to_node(void)
754{
755 int cpu;
756 u32 *cpu_to_apicid = early_per_cpu_ptr(x86_cpu_to_apicid);
757
758 BUG_ON(cpu_to_apicid == NULL);
759
760 for_each_possible_cpu(cpu) {
761 int node = numa_cpu_node(cpu);
762
763 if (node == NUMA_NO_NODE)
764 continue;
765
766 /*
767 * Exclude this node from
768 * bringup_nonboot_cpus
769 * cpu_up
770 * __try_online_node
771 * register_one_node
772 * because node_subsys is not initialized yet.
773 * TODO remove dependency on node_online
774 */
775 if (!node_online(node))
776 node_set_online(node);
777
778 numa_set_node(cpu, node);
779 }
780}
781
782#ifndef CONFIG_DEBUG_PER_CPU_MAPS
783
784# ifndef CONFIG_NUMA_EMU
785void numa_add_cpu(int cpu)
786{
787 cpumask_set_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
788}
789
790void numa_remove_cpu(int cpu)
791{
792 cpumask_clear_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
793}
794# endif /* !CONFIG_NUMA_EMU */
795
796#else /* !CONFIG_DEBUG_PER_CPU_MAPS */
797
798int __cpu_to_node(int cpu)
799{
800 if (early_per_cpu_ptr(x86_cpu_to_node_map)) {
801 printk(KERN_WARNING
802 "cpu_to_node(%d): usage too early!\n", cpu);
803 dump_stack();
804 return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
805 }
806 return per_cpu(x86_cpu_to_node_map, cpu);
807}
808EXPORT_SYMBOL(__cpu_to_node);
809
810/*
811 * Same function as cpu_to_node() but used if called before the
812 * per_cpu areas are setup.
813 */
814int early_cpu_to_node(int cpu)
815{
816 if (early_per_cpu_ptr(x86_cpu_to_node_map))
817 return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
818
819 if (!cpu_possible(cpu)) {
820 printk(KERN_WARNING
821 "early_cpu_to_node(%d): no per_cpu area!\n", cpu);
822 dump_stack();
823 return NUMA_NO_NODE;
824 }
825 return per_cpu(x86_cpu_to_node_map, cpu);
826}
827
828void debug_cpumask_set_cpu(int cpu, int node, bool enable)
829{
830 struct cpumask *mask;
831
832 if (node == NUMA_NO_NODE) {
833 /* early_cpu_to_node() already emits a warning and trace */
834 return;
835 }
836 mask = node_to_cpumask_map[node];
837 if (!cpumask_available(mask)) {
838 pr_err("node_to_cpumask_map[%i] NULL\n", node);
839 dump_stack();
840 return;
841 }
842
843 if (enable)
844 cpumask_set_cpu(cpu, mask);
845 else
846 cpumask_clear_cpu(cpu, mask);
847
848 printk(KERN_DEBUG "%s cpu %d node %d: mask now %*pbl\n",
849 enable ? "numa_add_cpu" : "numa_remove_cpu",
850 cpu, node, cpumask_pr_args(mask));
851 return;
852}
853
854# ifndef CONFIG_NUMA_EMU
855static void numa_set_cpumask(int cpu, bool enable)
856{
857 debug_cpumask_set_cpu(cpu, early_cpu_to_node(cpu), enable);
858}
859
860void numa_add_cpu(int cpu)
861{
862 numa_set_cpumask(cpu, true);
863}
864
865void numa_remove_cpu(int cpu)
866{
867 numa_set_cpumask(cpu, false);
868}
869# endif /* !CONFIG_NUMA_EMU */
870
871/*
872 * Returns a pointer to the bitmask of CPUs on Node 'node'.
873 */
874const struct cpumask *cpumask_of_node(int node)
875{
876 if ((unsigned)node >= nr_node_ids) {
877 printk(KERN_WARNING
878 "cpumask_of_node(%d): (unsigned)node >= nr_node_ids(%u)\n",
879 node, nr_node_ids);
880 dump_stack();
881 return cpu_none_mask;
882 }
883 if (!cpumask_available(node_to_cpumask_map[node])) {
884 printk(KERN_WARNING
885 "cpumask_of_node(%d): no node_to_cpumask_map!\n",
886 node);
887 dump_stack();
888 return cpu_online_mask;
889 }
890 return node_to_cpumask_map[node];
891}
892EXPORT_SYMBOL(cpumask_of_node);
893
894#endif /* !CONFIG_DEBUG_PER_CPU_MAPS */
895
896#ifdef CONFIG_NUMA_KEEP_MEMINFO
897static int meminfo_to_nid(struct numa_meminfo *mi, u64 start)
898{
899 int i;
900
901 for (i = 0; i < mi->nr_blks; i++)
902 if (mi->blk[i].start <= start && mi->blk[i].end > start)
903 return mi->blk[i].nid;
904 return NUMA_NO_NODE;
905}
906
907int phys_to_target_node(phys_addr_t start)
908{
909 int nid = meminfo_to_nid(&numa_meminfo, start);
910
911 /*
912 * Prefer online nodes, but if reserved memory might be
913 * hot-added continue the search with reserved ranges.
914 */
915 if (nid != NUMA_NO_NODE)
916 return nid;
917
918 return meminfo_to_nid(&numa_reserved_meminfo, start);
919}
920EXPORT_SYMBOL_GPL(phys_to_target_node);
921
922int memory_add_physaddr_to_nid(u64 start)
923{
924 int nid = meminfo_to_nid(&numa_meminfo, start);
925
926 if (nid == NUMA_NO_NODE)
927 nid = numa_meminfo.blk[0].nid;
928 return nid;
929}
930EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
931
932static int __init cmp_memblk(const void *a, const void *b)
933{
934 const struct numa_memblk *ma = *(const struct numa_memblk **)a;
935 const struct numa_memblk *mb = *(const struct numa_memblk **)b;
936
937 return (ma->start > mb->start) - (ma->start < mb->start);
938}
939
940static struct numa_memblk *numa_memblk_list[NR_NODE_MEMBLKS] __initdata;
941
942/**
943 * numa_fill_memblks - Fill gaps in numa_meminfo memblks
944 * @start: address to begin fill
945 * @end: address to end fill
946 *
947 * Find and extend numa_meminfo memblks to cover the physical
948 * address range @start-@end
949 *
950 * RETURNS:
951 * 0 : Success
952 * NUMA_NO_MEMBLK : No memblks exist in address range @start-@end
953 */
954
955int __init numa_fill_memblks(u64 start, u64 end)
956{
957 struct numa_memblk **blk = &numa_memblk_list[0];
958 struct numa_meminfo *mi = &numa_meminfo;
959 int count = 0;
960 u64 prev_end;
961
962 /*
963 * Create a list of pointers to numa_meminfo memblks that
964 * overlap start, end. The list is used to make in-place
965 * changes that fill out the numa_meminfo memblks.
966 */
967 for (int i = 0; i < mi->nr_blks; i++) {
968 struct numa_memblk *bi = &mi->blk[i];
969
970 if (memblock_addrs_overlap(start, end - start, bi->start,
971 bi->end - bi->start)) {
972 blk[count] = &mi->blk[i];
973 count++;
974 }
975 }
976 if (!count)
977 return NUMA_NO_MEMBLK;
978
979 /* Sort the list of pointers in memblk->start order */
980 sort(&blk[0], count, sizeof(blk[0]), cmp_memblk, NULL);
981
982 /* Make sure the first/last memblks include start/end */
983 blk[0]->start = min(blk[0]->start, start);
984 blk[count - 1]->end = max(blk[count - 1]->end, end);
985
986 /*
987 * Fill any gaps by tracking the previous memblks
988 * end address and backfilling to it if needed.
989 */
990 prev_end = blk[0]->end;
991 for (int i = 1; i < count; i++) {
992 struct numa_memblk *curr = blk[i];
993
994 if (prev_end >= curr->start) {
995 if (prev_end < curr->end)
996 prev_end = curr->end;
997 } else {
998 curr->start = prev_end;
999 prev_end = curr->end;
1000 }
1001 }
1002 return 0;
1003}
1004
1005#endif
1/* Common code for 32 and 64-bit NUMA */
2#include <linux/acpi.h>
3#include <linux/kernel.h>
4#include <linux/mm.h>
5#include <linux/string.h>
6#include <linux/init.h>
7#include <linux/bootmem.h>
8#include <linux/memblock.h>
9#include <linux/mmzone.h>
10#include <linux/ctype.h>
11#include <linux/nodemask.h>
12#include <linux/sched.h>
13#include <linux/topology.h>
14
15#include <asm/e820/api.h>
16#include <asm/proto.h>
17#include <asm/dma.h>
18#include <asm/amd_nb.h>
19
20#include "numa_internal.h"
21
22int numa_off;
23nodemask_t numa_nodes_parsed __initdata;
24
25struct pglist_data *node_data[MAX_NUMNODES] __read_mostly;
26EXPORT_SYMBOL(node_data);
27
28static struct numa_meminfo numa_meminfo
29#ifndef CONFIG_MEMORY_HOTPLUG
30__initdata
31#endif
32;
33
34static int numa_distance_cnt;
35static u8 *numa_distance;
36
37static __init int numa_setup(char *opt)
38{
39 if (!opt)
40 return -EINVAL;
41 if (!strncmp(opt, "off", 3))
42 numa_off = 1;
43#ifdef CONFIG_NUMA_EMU
44 if (!strncmp(opt, "fake=", 5))
45 numa_emu_cmdline(opt + 5);
46#endif
47#ifdef CONFIG_ACPI_NUMA
48 if (!strncmp(opt, "noacpi", 6))
49 acpi_numa = -1;
50#endif
51 return 0;
52}
53early_param("numa", numa_setup);
54
55/*
56 * apicid, cpu, node mappings
57 */
58s16 __apicid_to_node[MAX_LOCAL_APIC] = {
59 [0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE
60};
61
62int numa_cpu_node(int cpu)
63{
64 int apicid = early_per_cpu(x86_cpu_to_apicid, cpu);
65
66 if (apicid != BAD_APICID)
67 return __apicid_to_node[apicid];
68 return NUMA_NO_NODE;
69}
70
71cpumask_var_t node_to_cpumask_map[MAX_NUMNODES];
72EXPORT_SYMBOL(node_to_cpumask_map);
73
74/*
75 * Map cpu index to node index
76 */
77DEFINE_EARLY_PER_CPU(int, x86_cpu_to_node_map, NUMA_NO_NODE);
78EXPORT_EARLY_PER_CPU_SYMBOL(x86_cpu_to_node_map);
79
80void numa_set_node(int cpu, int node)
81{
82 int *cpu_to_node_map = early_per_cpu_ptr(x86_cpu_to_node_map);
83
84 /* early setting, no percpu area yet */
85 if (cpu_to_node_map) {
86 cpu_to_node_map[cpu] = node;
87 return;
88 }
89
90#ifdef CONFIG_DEBUG_PER_CPU_MAPS
91 if (cpu >= nr_cpu_ids || !cpu_possible(cpu)) {
92 printk(KERN_ERR "numa_set_node: invalid cpu# (%d)\n", cpu);
93 dump_stack();
94 return;
95 }
96#endif
97 per_cpu(x86_cpu_to_node_map, cpu) = node;
98
99 set_cpu_numa_node(cpu, node);
100}
101
102void numa_clear_node(int cpu)
103{
104 numa_set_node(cpu, NUMA_NO_NODE);
105}
106
107/*
108 * Allocate node_to_cpumask_map based on number of available nodes
109 * Requires node_possible_map to be valid.
110 *
111 * Note: cpumask_of_node() is not valid until after this is done.
112 * (Use CONFIG_DEBUG_PER_CPU_MAPS to check this.)
113 */
114void __init setup_node_to_cpumask_map(void)
115{
116 unsigned int node;
117
118 /* setup nr_node_ids if not done yet */
119 if (nr_node_ids == MAX_NUMNODES)
120 setup_nr_node_ids();
121
122 /* allocate the map */
123 for (node = 0; node < nr_node_ids; node++)
124 alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]);
125
126 /* cpumask_of_node() will now work */
127 pr_debug("Node to cpumask map for %d nodes\n", nr_node_ids);
128}
129
130static int __init numa_add_memblk_to(int nid, u64 start, u64 end,
131 struct numa_meminfo *mi)
132{
133 /* ignore zero length blks */
134 if (start == end)
135 return 0;
136
137 /* whine about and ignore invalid blks */
138 if (start > end || nid < 0 || nid >= MAX_NUMNODES) {
139 pr_warning("NUMA: Warning: invalid memblk node %d [mem %#010Lx-%#010Lx]\n",
140 nid, start, end - 1);
141 return 0;
142 }
143
144 if (mi->nr_blks >= NR_NODE_MEMBLKS) {
145 pr_err("NUMA: too many memblk ranges\n");
146 return -EINVAL;
147 }
148
149 mi->blk[mi->nr_blks].start = start;
150 mi->blk[mi->nr_blks].end = end;
151 mi->blk[mi->nr_blks].nid = nid;
152 mi->nr_blks++;
153 return 0;
154}
155
156/**
157 * numa_remove_memblk_from - Remove one numa_memblk from a numa_meminfo
158 * @idx: Index of memblk to remove
159 * @mi: numa_meminfo to remove memblk from
160 *
161 * Remove @idx'th numa_memblk from @mi by shifting @mi->blk[] and
162 * decrementing @mi->nr_blks.
163 */
164void __init numa_remove_memblk_from(int idx, struct numa_meminfo *mi)
165{
166 mi->nr_blks--;
167 memmove(&mi->blk[idx], &mi->blk[idx + 1],
168 (mi->nr_blks - idx) * sizeof(mi->blk[0]));
169}
170
171/**
172 * numa_add_memblk - Add one numa_memblk to numa_meminfo
173 * @nid: NUMA node ID of the new memblk
174 * @start: Start address of the new memblk
175 * @end: End address of the new memblk
176 *
177 * Add a new memblk to the default numa_meminfo.
178 *
179 * RETURNS:
180 * 0 on success, -errno on failure.
181 */
182int __init numa_add_memblk(int nid, u64 start, u64 end)
183{
184 return numa_add_memblk_to(nid, start, end, &numa_meminfo);
185}
186
187/* Allocate NODE_DATA for a node on the local memory */
188static void __init alloc_node_data(int nid)
189{
190 const size_t nd_size = roundup(sizeof(pg_data_t), PAGE_SIZE);
191 u64 nd_pa;
192 void *nd;
193 int tnid;
194
195 /*
196 * Allocate node data. Try node-local memory and then any node.
197 * Never allocate in DMA zone.
198 */
199 nd_pa = memblock_alloc_nid(nd_size, SMP_CACHE_BYTES, nid);
200 if (!nd_pa) {
201 nd_pa = __memblock_alloc_base(nd_size, SMP_CACHE_BYTES,
202 MEMBLOCK_ALLOC_ACCESSIBLE);
203 if (!nd_pa) {
204 pr_err("Cannot find %zu bytes in any node (initial node: %d)\n",
205 nd_size, nid);
206 return;
207 }
208 }
209 nd = __va(nd_pa);
210
211 /* report and initialize */
212 printk(KERN_INFO "NODE_DATA(%d) allocated [mem %#010Lx-%#010Lx]\n", nid,
213 nd_pa, nd_pa + nd_size - 1);
214 tnid = early_pfn_to_nid(nd_pa >> PAGE_SHIFT);
215 if (tnid != nid)
216 printk(KERN_INFO " NODE_DATA(%d) on node %d\n", nid, tnid);
217
218 node_data[nid] = nd;
219 memset(NODE_DATA(nid), 0, sizeof(pg_data_t));
220
221 node_set_online(nid);
222}
223
224/**
225 * numa_cleanup_meminfo - Cleanup a numa_meminfo
226 * @mi: numa_meminfo to clean up
227 *
228 * Sanitize @mi by merging and removing unnecessary memblks. Also check for
229 * conflicts and clear unused memblks.
230 *
231 * RETURNS:
232 * 0 on success, -errno on failure.
233 */
234int __init numa_cleanup_meminfo(struct numa_meminfo *mi)
235{
236 const u64 low = 0;
237 const u64 high = PFN_PHYS(max_pfn);
238 int i, j, k;
239
240 /* first, trim all entries */
241 for (i = 0; i < mi->nr_blks; i++) {
242 struct numa_memblk *bi = &mi->blk[i];
243
244 /* make sure all blocks are inside the limits */
245 bi->start = max(bi->start, low);
246 bi->end = min(bi->end, high);
247
248 /* and there's no empty or non-exist block */
249 if (bi->start >= bi->end ||
250 !memblock_overlaps_region(&memblock.memory,
251 bi->start, bi->end - bi->start))
252 numa_remove_memblk_from(i--, mi);
253 }
254
255 /* merge neighboring / overlapping entries */
256 for (i = 0; i < mi->nr_blks; i++) {
257 struct numa_memblk *bi = &mi->blk[i];
258
259 for (j = i + 1; j < mi->nr_blks; j++) {
260 struct numa_memblk *bj = &mi->blk[j];
261 u64 start, end;
262
263 /*
264 * See whether there are overlapping blocks. Whine
265 * about but allow overlaps of the same nid. They
266 * will be merged below.
267 */
268 if (bi->end > bj->start && bi->start < bj->end) {
269 if (bi->nid != bj->nid) {
270 pr_err("NUMA: node %d [mem %#010Lx-%#010Lx] overlaps with node %d [mem %#010Lx-%#010Lx]\n",
271 bi->nid, bi->start, bi->end - 1,
272 bj->nid, bj->start, bj->end - 1);
273 return -EINVAL;
274 }
275 pr_warning("NUMA: Warning: node %d [mem %#010Lx-%#010Lx] overlaps with itself [mem %#010Lx-%#010Lx]\n",
276 bi->nid, bi->start, bi->end - 1,
277 bj->start, bj->end - 1);
278 }
279
280 /*
281 * Join together blocks on the same node, holes
282 * between which don't overlap with memory on other
283 * nodes.
284 */
285 if (bi->nid != bj->nid)
286 continue;
287 start = min(bi->start, bj->start);
288 end = max(bi->end, bj->end);
289 for (k = 0; k < mi->nr_blks; k++) {
290 struct numa_memblk *bk = &mi->blk[k];
291
292 if (bi->nid == bk->nid)
293 continue;
294 if (start < bk->end && end > bk->start)
295 break;
296 }
297 if (k < mi->nr_blks)
298 continue;
299 printk(KERN_INFO "NUMA: Node %d [mem %#010Lx-%#010Lx] + [mem %#010Lx-%#010Lx] -> [mem %#010Lx-%#010Lx]\n",
300 bi->nid, bi->start, bi->end - 1, bj->start,
301 bj->end - 1, start, end - 1);
302 bi->start = start;
303 bi->end = end;
304 numa_remove_memblk_from(j--, mi);
305 }
306 }
307
308 /* clear unused ones */
309 for (i = mi->nr_blks; i < ARRAY_SIZE(mi->blk); i++) {
310 mi->blk[i].start = mi->blk[i].end = 0;
311 mi->blk[i].nid = NUMA_NO_NODE;
312 }
313
314 return 0;
315}
316
317/*
318 * Set nodes, which have memory in @mi, in *@nodemask.
319 */
320static void __init numa_nodemask_from_meminfo(nodemask_t *nodemask,
321 const struct numa_meminfo *mi)
322{
323 int i;
324
325 for (i = 0; i < ARRAY_SIZE(mi->blk); i++)
326 if (mi->blk[i].start != mi->blk[i].end &&
327 mi->blk[i].nid != NUMA_NO_NODE)
328 node_set(mi->blk[i].nid, *nodemask);
329}
330
331/**
332 * numa_reset_distance - Reset NUMA distance table
333 *
334 * The current table is freed. The next numa_set_distance() call will
335 * create a new one.
336 */
337void __init numa_reset_distance(void)
338{
339 size_t size = numa_distance_cnt * numa_distance_cnt * sizeof(numa_distance[0]);
340
341 /* numa_distance could be 1LU marking allocation failure, test cnt */
342 if (numa_distance_cnt)
343 memblock_free(__pa(numa_distance), size);
344 numa_distance_cnt = 0;
345 numa_distance = NULL; /* enable table creation */
346}
347
348static int __init numa_alloc_distance(void)
349{
350 nodemask_t nodes_parsed;
351 size_t size;
352 int i, j, cnt = 0;
353 u64 phys;
354
355 /* size the new table and allocate it */
356 nodes_parsed = numa_nodes_parsed;
357 numa_nodemask_from_meminfo(&nodes_parsed, &numa_meminfo);
358
359 for_each_node_mask(i, nodes_parsed)
360 cnt = i;
361 cnt++;
362 size = cnt * cnt * sizeof(numa_distance[0]);
363
364 phys = memblock_find_in_range(0, PFN_PHYS(max_pfn_mapped),
365 size, PAGE_SIZE);
366 if (!phys) {
367 pr_warning("NUMA: Warning: can't allocate distance table!\n");
368 /* don't retry until explicitly reset */
369 numa_distance = (void *)1LU;
370 return -ENOMEM;
371 }
372 memblock_reserve(phys, size);
373
374 numa_distance = __va(phys);
375 numa_distance_cnt = cnt;
376
377 /* fill with the default distances */
378 for (i = 0; i < cnt; i++)
379 for (j = 0; j < cnt; j++)
380 numa_distance[i * cnt + j] = i == j ?
381 LOCAL_DISTANCE : REMOTE_DISTANCE;
382 printk(KERN_DEBUG "NUMA: Initialized distance table, cnt=%d\n", cnt);
383
384 return 0;
385}
386
387/**
388 * numa_set_distance - Set NUMA distance from one NUMA to another
389 * @from: the 'from' node to set distance
390 * @to: the 'to' node to set distance
391 * @distance: NUMA distance
392 *
393 * Set the distance from node @from to @to to @distance. If distance table
394 * doesn't exist, one which is large enough to accommodate all the currently
395 * known nodes will be created.
396 *
397 * If such table cannot be allocated, a warning is printed and further
398 * calls are ignored until the distance table is reset with
399 * numa_reset_distance().
400 *
401 * If @from or @to is higher than the highest known node or lower than zero
402 * at the time of table creation or @distance doesn't make sense, the call
403 * is ignored.
404 * This is to allow simplification of specific NUMA config implementations.
405 */
406void __init numa_set_distance(int from, int to, int distance)
407{
408 if (!numa_distance && numa_alloc_distance() < 0)
409 return;
410
411 if (from >= numa_distance_cnt || to >= numa_distance_cnt ||
412 from < 0 || to < 0) {
413 pr_warn_once("NUMA: Warning: node ids are out of bound, from=%d to=%d distance=%d\n",
414 from, to, distance);
415 return;
416 }
417
418 if ((u8)distance != distance ||
419 (from == to && distance != LOCAL_DISTANCE)) {
420 pr_warn_once("NUMA: Warning: invalid distance parameter, from=%d to=%d distance=%d\n",
421 from, to, distance);
422 return;
423 }
424
425 numa_distance[from * numa_distance_cnt + to] = distance;
426}
427
428int __node_distance(int from, int to)
429{
430 if (from >= numa_distance_cnt || to >= numa_distance_cnt)
431 return from == to ? LOCAL_DISTANCE : REMOTE_DISTANCE;
432 return numa_distance[from * numa_distance_cnt + to];
433}
434EXPORT_SYMBOL(__node_distance);
435
436/*
437 * Sanity check to catch more bad NUMA configurations (they are amazingly
438 * common). Make sure the nodes cover all memory.
439 */
440static bool __init numa_meminfo_cover_memory(const struct numa_meminfo *mi)
441{
442 u64 numaram, e820ram;
443 int i;
444
445 numaram = 0;
446 for (i = 0; i < mi->nr_blks; i++) {
447 u64 s = mi->blk[i].start >> PAGE_SHIFT;
448 u64 e = mi->blk[i].end >> PAGE_SHIFT;
449 numaram += e - s;
450 numaram -= __absent_pages_in_range(mi->blk[i].nid, s, e);
451 if ((s64)numaram < 0)
452 numaram = 0;
453 }
454
455 e820ram = max_pfn - absent_pages_in_range(0, max_pfn);
456
457 /* We seem to lose 3 pages somewhere. Allow 1M of slack. */
458 if ((s64)(e820ram - numaram) >= (1 << (20 - PAGE_SHIFT))) {
459 printk(KERN_ERR "NUMA: nodes only cover %LuMB of your %LuMB e820 RAM. Not used.\n",
460 (numaram << PAGE_SHIFT) >> 20,
461 (e820ram << PAGE_SHIFT) >> 20);
462 return false;
463 }
464 return true;
465}
466
467/*
468 * Mark all currently memblock-reserved physical memory (which covers the
469 * kernel's own memory ranges) as hot-unswappable.
470 */
471static void __init numa_clear_kernel_node_hotplug(void)
472{
473 nodemask_t reserved_nodemask = NODE_MASK_NONE;
474 struct memblock_region *mb_region;
475 int i;
476
477 /*
478 * We have to do some preprocessing of memblock regions, to
479 * make them suitable for reservation.
480 *
481 * At this time, all memory regions reserved by memblock are
482 * used by the kernel, but those regions are not split up
483 * along node boundaries yet, and don't necessarily have their
484 * node ID set yet either.
485 *
486 * So iterate over all memory known to the x86 architecture,
487 * and use those ranges to set the nid in memblock.reserved.
488 * This will split up the memblock regions along node
489 * boundaries and will set the node IDs as well.
490 */
491 for (i = 0; i < numa_meminfo.nr_blks; i++) {
492 struct numa_memblk *mb = numa_meminfo.blk + i;
493 int ret;
494
495 ret = memblock_set_node(mb->start, mb->end - mb->start, &memblock.reserved, mb->nid);
496 WARN_ON_ONCE(ret);
497 }
498
499 /*
500 * Now go over all reserved memblock regions, to construct a
501 * node mask of all kernel reserved memory areas.
502 *
503 * [ Note, when booting with mem=nn[kMG] or in a kdump kernel,
504 * numa_meminfo might not include all memblock.reserved
505 * memory ranges, because quirks such as trim_snb_memory()
506 * reserve specific pages for Sandy Bridge graphics. ]
507 */
508 for_each_memblock(reserved, mb_region) {
509 if (mb_region->nid != MAX_NUMNODES)
510 node_set(mb_region->nid, reserved_nodemask);
511 }
512
513 /*
514 * Finally, clear the MEMBLOCK_HOTPLUG flag for all memory
515 * belonging to the reserved node mask.
516 *
517 * Note that this will include memory regions that reside
518 * on nodes that contain kernel memory - entire nodes
519 * become hot-unpluggable:
520 */
521 for (i = 0; i < numa_meminfo.nr_blks; i++) {
522 struct numa_memblk *mb = numa_meminfo.blk + i;
523
524 if (!node_isset(mb->nid, reserved_nodemask))
525 continue;
526
527 memblock_clear_hotplug(mb->start, mb->end - mb->start);
528 }
529}
530
531static int __init numa_register_memblks(struct numa_meminfo *mi)
532{
533 unsigned long uninitialized_var(pfn_align);
534 int i, nid;
535
536 /* Account for nodes with cpus and no memory */
537 node_possible_map = numa_nodes_parsed;
538 numa_nodemask_from_meminfo(&node_possible_map, mi);
539 if (WARN_ON(nodes_empty(node_possible_map)))
540 return -EINVAL;
541
542 for (i = 0; i < mi->nr_blks; i++) {
543 struct numa_memblk *mb = &mi->blk[i];
544 memblock_set_node(mb->start, mb->end - mb->start,
545 &memblock.memory, mb->nid);
546 }
547
548 /*
549 * At very early time, the kernel have to use some memory such as
550 * loading the kernel image. We cannot prevent this anyway. So any
551 * node the kernel resides in should be un-hotpluggable.
552 *
553 * And when we come here, alloc node data won't fail.
554 */
555 numa_clear_kernel_node_hotplug();
556
557 /*
558 * If sections array is gonna be used for pfn -> nid mapping, check
559 * whether its granularity is fine enough.
560 */
561#ifdef NODE_NOT_IN_PAGE_FLAGS
562 pfn_align = node_map_pfn_alignment();
563 if (pfn_align && pfn_align < PAGES_PER_SECTION) {
564 printk(KERN_WARNING "Node alignment %LuMB < min %LuMB, rejecting NUMA config\n",
565 PFN_PHYS(pfn_align) >> 20,
566 PFN_PHYS(PAGES_PER_SECTION) >> 20);
567 return -EINVAL;
568 }
569#endif
570 if (!numa_meminfo_cover_memory(mi))
571 return -EINVAL;
572
573 /* Finally register nodes. */
574 for_each_node_mask(nid, node_possible_map) {
575 u64 start = PFN_PHYS(max_pfn);
576 u64 end = 0;
577
578 for (i = 0; i < mi->nr_blks; i++) {
579 if (nid != mi->blk[i].nid)
580 continue;
581 start = min(mi->blk[i].start, start);
582 end = max(mi->blk[i].end, end);
583 }
584
585 if (start >= end)
586 continue;
587
588 /*
589 * Don't confuse VM with a node that doesn't have the
590 * minimum amount of memory:
591 */
592 if (end && (end - start) < NODE_MIN_SIZE)
593 continue;
594
595 alloc_node_data(nid);
596 }
597
598 /* Dump memblock with node info and return. */
599 memblock_dump_all();
600 return 0;
601}
602
603/*
604 * There are unfortunately some poorly designed mainboards around that
605 * only connect memory to a single CPU. This breaks the 1:1 cpu->node
606 * mapping. To avoid this fill in the mapping for all possible CPUs,
607 * as the number of CPUs is not known yet. We round robin the existing
608 * nodes.
609 */
610static void __init numa_init_array(void)
611{
612 int rr, i;
613
614 rr = first_node(node_online_map);
615 for (i = 0; i < nr_cpu_ids; i++) {
616 if (early_cpu_to_node(i) != NUMA_NO_NODE)
617 continue;
618 numa_set_node(i, rr);
619 rr = next_node_in(rr, node_online_map);
620 }
621}
622
623static int __init numa_init(int (*init_func)(void))
624{
625 int i;
626 int ret;
627
628 for (i = 0; i < MAX_LOCAL_APIC; i++)
629 set_apicid_to_node(i, NUMA_NO_NODE);
630
631 nodes_clear(numa_nodes_parsed);
632 nodes_clear(node_possible_map);
633 nodes_clear(node_online_map);
634 memset(&numa_meminfo, 0, sizeof(numa_meminfo));
635 WARN_ON(memblock_set_node(0, ULLONG_MAX, &memblock.memory,
636 MAX_NUMNODES));
637 WARN_ON(memblock_set_node(0, ULLONG_MAX, &memblock.reserved,
638 MAX_NUMNODES));
639 /* In case that parsing SRAT failed. */
640 WARN_ON(memblock_clear_hotplug(0, ULLONG_MAX));
641 numa_reset_distance();
642
643 ret = init_func();
644 if (ret < 0)
645 return ret;
646
647 /*
648 * We reset memblock back to the top-down direction
649 * here because if we configured ACPI_NUMA, we have
650 * parsed SRAT in init_func(). It is ok to have the
651 * reset here even if we did't configure ACPI_NUMA
652 * or acpi numa init fails and fallbacks to dummy
653 * numa init.
654 */
655 memblock_set_bottom_up(false);
656
657 ret = numa_cleanup_meminfo(&numa_meminfo);
658 if (ret < 0)
659 return ret;
660
661 numa_emulation(&numa_meminfo, numa_distance_cnt);
662
663 ret = numa_register_memblks(&numa_meminfo);
664 if (ret < 0)
665 return ret;
666
667 for (i = 0; i < nr_cpu_ids; i++) {
668 int nid = early_cpu_to_node(i);
669
670 if (nid == NUMA_NO_NODE)
671 continue;
672 if (!node_online(nid))
673 numa_clear_node(i);
674 }
675 numa_init_array();
676
677 return 0;
678}
679
680/**
681 * dummy_numa_init - Fallback dummy NUMA init
682 *
683 * Used if there's no underlying NUMA architecture, NUMA initialization
684 * fails, or NUMA is disabled on the command line.
685 *
686 * Must online at least one node and add memory blocks that cover all
687 * allowed memory. This function must not fail.
688 */
689static int __init dummy_numa_init(void)
690{
691 printk(KERN_INFO "%s\n",
692 numa_off ? "NUMA turned off" : "No NUMA configuration found");
693 printk(KERN_INFO "Faking a node at [mem %#018Lx-%#018Lx]\n",
694 0LLU, PFN_PHYS(max_pfn) - 1);
695
696 node_set(0, numa_nodes_parsed);
697 numa_add_memblk(0, 0, PFN_PHYS(max_pfn));
698
699 return 0;
700}
701
702/**
703 * x86_numa_init - Initialize NUMA
704 *
705 * Try each configured NUMA initialization method until one succeeds. The
706 * last fallback is dummy single node config encomapssing whole memory and
707 * never fails.
708 */
709void __init x86_numa_init(void)
710{
711 if (!numa_off) {
712#ifdef CONFIG_ACPI_NUMA
713 if (!numa_init(x86_acpi_numa_init))
714 return;
715#endif
716#ifdef CONFIG_AMD_NUMA
717 if (!numa_init(amd_numa_init))
718 return;
719#endif
720 }
721
722 numa_init(dummy_numa_init);
723}
724
725static void __init init_memory_less_node(int nid)
726{
727 unsigned long zones_size[MAX_NR_ZONES] = {0};
728 unsigned long zholes_size[MAX_NR_ZONES] = {0};
729
730 /* Allocate and initialize node data. Memory-less node is now online.*/
731 alloc_node_data(nid);
732 free_area_init_node(nid, zones_size, 0, zholes_size);
733
734 /*
735 * All zonelists will be built later in start_kernel() after per cpu
736 * areas are initialized.
737 */
738}
739
740/*
741 * Setup early cpu_to_node.
742 *
743 * Populate cpu_to_node[] only if x86_cpu_to_apicid[],
744 * and apicid_to_node[] tables have valid entries for a CPU.
745 * This means we skip cpu_to_node[] initialisation for NUMA
746 * emulation and faking node case (when running a kernel compiled
747 * for NUMA on a non NUMA box), which is OK as cpu_to_node[]
748 * is already initialized in a round robin manner at numa_init_array,
749 * prior to this call, and this initialization is good enough
750 * for the fake NUMA cases.
751 *
752 * Called before the per_cpu areas are setup.
753 */
754void __init init_cpu_to_node(void)
755{
756 int cpu;
757 u16 *cpu_to_apicid = early_per_cpu_ptr(x86_cpu_to_apicid);
758
759 BUG_ON(cpu_to_apicid == NULL);
760
761 for_each_possible_cpu(cpu) {
762 int node = numa_cpu_node(cpu);
763
764 if (node == NUMA_NO_NODE)
765 continue;
766
767 if (!node_online(node))
768 init_memory_less_node(node);
769
770 numa_set_node(cpu, node);
771 }
772}
773
774#ifndef CONFIG_DEBUG_PER_CPU_MAPS
775
776# ifndef CONFIG_NUMA_EMU
777void numa_add_cpu(int cpu)
778{
779 cpumask_set_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
780}
781
782void numa_remove_cpu(int cpu)
783{
784 cpumask_clear_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
785}
786# endif /* !CONFIG_NUMA_EMU */
787
788#else /* !CONFIG_DEBUG_PER_CPU_MAPS */
789
790int __cpu_to_node(int cpu)
791{
792 if (early_per_cpu_ptr(x86_cpu_to_node_map)) {
793 printk(KERN_WARNING
794 "cpu_to_node(%d): usage too early!\n", cpu);
795 dump_stack();
796 return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
797 }
798 return per_cpu(x86_cpu_to_node_map, cpu);
799}
800EXPORT_SYMBOL(__cpu_to_node);
801
802/*
803 * Same function as cpu_to_node() but used if called before the
804 * per_cpu areas are setup.
805 */
806int early_cpu_to_node(int cpu)
807{
808 if (early_per_cpu_ptr(x86_cpu_to_node_map))
809 return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
810
811 if (!cpu_possible(cpu)) {
812 printk(KERN_WARNING
813 "early_cpu_to_node(%d): no per_cpu area!\n", cpu);
814 dump_stack();
815 return NUMA_NO_NODE;
816 }
817 return per_cpu(x86_cpu_to_node_map, cpu);
818}
819
820void debug_cpumask_set_cpu(int cpu, int node, bool enable)
821{
822 struct cpumask *mask;
823
824 if (node == NUMA_NO_NODE) {
825 /* early_cpu_to_node() already emits a warning and trace */
826 return;
827 }
828 mask = node_to_cpumask_map[node];
829 if (!mask) {
830 pr_err("node_to_cpumask_map[%i] NULL\n", node);
831 dump_stack();
832 return;
833 }
834
835 if (enable)
836 cpumask_set_cpu(cpu, mask);
837 else
838 cpumask_clear_cpu(cpu, mask);
839
840 printk(KERN_DEBUG "%s cpu %d node %d: mask now %*pbl\n",
841 enable ? "numa_add_cpu" : "numa_remove_cpu",
842 cpu, node, cpumask_pr_args(mask));
843 return;
844}
845
846# ifndef CONFIG_NUMA_EMU
847static void numa_set_cpumask(int cpu, bool enable)
848{
849 debug_cpumask_set_cpu(cpu, early_cpu_to_node(cpu), enable);
850}
851
852void numa_add_cpu(int cpu)
853{
854 numa_set_cpumask(cpu, true);
855}
856
857void numa_remove_cpu(int cpu)
858{
859 numa_set_cpumask(cpu, false);
860}
861# endif /* !CONFIG_NUMA_EMU */
862
863/*
864 * Returns a pointer to the bitmask of CPUs on Node 'node'.
865 */
866const struct cpumask *cpumask_of_node(int node)
867{
868 if (node >= nr_node_ids) {
869 printk(KERN_WARNING
870 "cpumask_of_node(%d): node > nr_node_ids(%d)\n",
871 node, nr_node_ids);
872 dump_stack();
873 return cpu_none_mask;
874 }
875 if (node_to_cpumask_map[node] == NULL) {
876 printk(KERN_WARNING
877 "cpumask_of_node(%d): no node_to_cpumask_map!\n",
878 node);
879 dump_stack();
880 return cpu_online_mask;
881 }
882 return node_to_cpumask_map[node];
883}
884EXPORT_SYMBOL(cpumask_of_node);
885
886#endif /* !CONFIG_DEBUG_PER_CPU_MAPS */
887
888#ifdef CONFIG_MEMORY_HOTPLUG
889int memory_add_physaddr_to_nid(u64 start)
890{
891 struct numa_meminfo *mi = &numa_meminfo;
892 int nid = mi->blk[0].nid;
893 int i;
894
895 for (i = 0; i < mi->nr_blks; i++)
896 if (mi->blk[i].start <= start && mi->blk[i].end > start)
897 nid = mi->blk[i].nid;
898 return nid;
899}
900EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
901#endif