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(str: 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(mask: &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, mi: 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, mi: &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(size: 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(pfn: 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(type: &memblock.memory,
252 base: bi->start, size: bi->end - bi->start)) {
253 numa_move_tail_memblk(dst: &numa_reserved_meminfo, idx: i--, src: 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(nid: bi->nid, start: high, end: bi->end,
263 mi: &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(idx: 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(idx: 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(ptr: 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(nodemask: &nodes_parsed, mi: &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, start: 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 * Sanity check to catch more bad NUMA configurations (they are amazingly
454 * common). Make sure the nodes cover all memory.
455 */
456static bool __init numa_meminfo_cover_memory(const struct numa_meminfo *mi)
457{
458 u64 numaram, e820ram;
459 int i;
460
461 numaram = 0;
462 for (i = 0; i < mi->nr_blks; i++) {
463 u64 s = mi->blk[i].start >> PAGE_SHIFT;
464 u64 e = mi->blk[i].end >> PAGE_SHIFT;
465 numaram += e - s;
466 numaram -= __absent_pages_in_range(nid: mi->blk[i].nid, start_pfn: s, end_pfn: e);
467 if ((s64)numaram < 0)
468 numaram = 0;
469 }
470
471 e820ram = max_pfn - absent_pages_in_range(start_pfn: 0, end_pfn: max_pfn);
472
473 /* We seem to lose 3 pages somewhere. Allow 1M of slack. */
474 if ((s64)(e820ram - numaram) >= (1 << (20 - PAGE_SHIFT))) {
475 printk(KERN_ERR "NUMA: nodes only cover %LuMB of your %LuMB e820 RAM. Not used.\n",
476 (numaram << PAGE_SHIFT) >> 20,
477 (e820ram << PAGE_SHIFT) >> 20);
478 return false;
479 }
480 return true;
481}
482
483/*
484 * Mark all currently memblock-reserved physical memory (which covers the
485 * kernel's own memory ranges) as hot-unswappable.
486 */
487static void __init numa_clear_kernel_node_hotplug(void)
488{
489 nodemask_t reserved_nodemask = NODE_MASK_NONE;
490 struct memblock_region *mb_region;
491 int i;
492
493 /*
494 * We have to do some preprocessing of memblock regions, to
495 * make them suitable for reservation.
496 *
497 * At this time, all memory regions reserved by memblock are
498 * used by the kernel, but those regions are not split up
499 * along node boundaries yet, and don't necessarily have their
500 * node ID set yet either.
501 *
502 * So iterate over all memory known to the x86 architecture,
503 * and use those ranges to set the nid in memblock.reserved.
504 * This will split up the memblock regions along node
505 * boundaries and will set the node IDs as well.
506 */
507 for (i = 0; i < numa_meminfo.nr_blks; i++) {
508 struct numa_memblk *mb = numa_meminfo.blk + i;
509 int ret;
510
511 ret = memblock_set_node(base: mb->start, size: mb->end - mb->start, type: &memblock.reserved, nid: mb->nid);
512 WARN_ON_ONCE(ret);
513 }
514
515 /*
516 * Now go over all reserved memblock regions, to construct a
517 * node mask of all kernel reserved memory areas.
518 *
519 * [ Note, when booting with mem=nn[kMG] or in a kdump kernel,
520 * numa_meminfo might not include all memblock.reserved
521 * memory ranges, because quirks such as trim_snb_memory()
522 * reserve specific pages for Sandy Bridge graphics. ]
523 */
524 for_each_reserved_mem_region(mb_region) {
525 int nid = memblock_get_region_node(r: mb_region);
526
527 if (nid != MAX_NUMNODES)
528 node_set(nid, reserved_nodemask);
529 }
530
531 /*
532 * Finally, clear the MEMBLOCK_HOTPLUG flag for all memory
533 * belonging to the reserved node mask.
534 *
535 * Note that this will include memory regions that reside
536 * on nodes that contain kernel memory - entire nodes
537 * become hot-unpluggable:
538 */
539 for (i = 0; i < numa_meminfo.nr_blks; i++) {
540 struct numa_memblk *mb = numa_meminfo.blk + i;
541
542 if (!node_isset(mb->nid, reserved_nodemask))
543 continue;
544
545 memblock_clear_hotplug(base: mb->start, size: mb->end - mb->start);
546 }
547}
548
549static int __init numa_register_memblks(struct numa_meminfo *mi)
550{
551 int i, nid;
552
553 /* Account for nodes with cpus and no memory */
554 node_possible_map = numa_nodes_parsed;
555 numa_nodemask_from_meminfo(nodemask: &node_possible_map, mi);
556 if (WARN_ON(nodes_empty(node_possible_map)))
557 return -EINVAL;
558
559 for (i = 0; i < mi->nr_blks; i++) {
560 struct numa_memblk *mb = &mi->blk[i];
561 memblock_set_node(base: mb->start, size: mb->end - mb->start,
562 type: &memblock.memory, nid: mb->nid);
563 }
564
565 /*
566 * At very early time, the kernel have to use some memory such as
567 * loading the kernel image. We cannot prevent this anyway. So any
568 * node the kernel resides in should be un-hotpluggable.
569 *
570 * And when we come here, alloc node data won't fail.
571 */
572 numa_clear_kernel_node_hotplug();
573
574 /*
575 * If sections array is gonna be used for pfn -> nid mapping, check
576 * whether its granularity is fine enough.
577 */
578 if (IS_ENABLED(NODE_NOT_IN_PAGE_FLAGS)) {
579 unsigned long pfn_align = node_map_pfn_alignment();
580
581 if (pfn_align && pfn_align < PAGES_PER_SECTION) {
582 pr_warn("Node alignment %LuMB < min %LuMB, rejecting NUMA config\n",
583 PFN_PHYS(pfn_align) >> 20,
584 PFN_PHYS(PAGES_PER_SECTION) >> 20);
585 return -EINVAL;
586 }
587 }
588 if (!numa_meminfo_cover_memory(mi))
589 return -EINVAL;
590
591 /* Finally register nodes. */
592 for_each_node_mask(nid, node_possible_map) {
593 u64 start = PFN_PHYS(max_pfn);
594 u64 end = 0;
595
596 for (i = 0; i < mi->nr_blks; i++) {
597 if (nid != mi->blk[i].nid)
598 continue;
599 start = min(mi->blk[i].start, start);
600 end = max(mi->blk[i].end, end);
601 }
602
603 if (start >= end)
604 continue;
605
606 alloc_node_data(nid);
607 }
608
609 /* Dump memblock with node info and return. */
610 memblock_dump_all();
611 return 0;
612}
613
614/*
615 * There are unfortunately some poorly designed mainboards around that
616 * only connect memory to a single CPU. This breaks the 1:1 cpu->node
617 * mapping. To avoid this fill in the mapping for all possible CPUs,
618 * as the number of CPUs is not known yet. We round robin the existing
619 * nodes.
620 */
621static void __init numa_init_array(void)
622{
623 int rr, i;
624
625 rr = first_node(node_online_map);
626 for (i = 0; i < nr_cpu_ids; i++) {
627 if (early_cpu_to_node(cpu: i) != NUMA_NO_NODE)
628 continue;
629 numa_set_node(cpu: i, node: rr);
630 rr = next_node_in(rr, node_online_map);
631 }
632}
633
634static int __init numa_init(int (*init_func)(void))
635{
636 int i;
637 int ret;
638
639 for (i = 0; i < MAX_LOCAL_APIC; i++)
640 set_apicid_to_node(apicid: i, NUMA_NO_NODE);
641
642 nodes_clear(numa_nodes_parsed);
643 nodes_clear(node_possible_map);
644 nodes_clear(node_online_map);
645 memset(&numa_meminfo, 0, sizeof(numa_meminfo));
646 WARN_ON(memblock_set_node(0, ULLONG_MAX, &memblock.memory,
647 MAX_NUMNODES));
648 WARN_ON(memblock_set_node(0, ULLONG_MAX, &memblock.reserved,
649 MAX_NUMNODES));
650 /* In case that parsing SRAT failed. */
651 WARN_ON(memblock_clear_hotplug(0, ULLONG_MAX));
652 numa_reset_distance();
653
654 ret = init_func();
655 if (ret < 0)
656 return ret;
657
658 /*
659 * We reset memblock back to the top-down direction
660 * here because if we configured ACPI_NUMA, we have
661 * parsed SRAT in init_func(). It is ok to have the
662 * reset here even if we did't configure ACPI_NUMA
663 * or acpi numa init fails and fallbacks to dummy
664 * numa init.
665 */
666 memblock_set_bottom_up(enable: false);
667
668 ret = numa_cleanup_meminfo(mi: &numa_meminfo);
669 if (ret < 0)
670 return ret;
671
672 numa_emulation(numa_meminfo: &numa_meminfo, numa_dist_cnt: numa_distance_cnt);
673
674 ret = numa_register_memblks(mi: &numa_meminfo);
675 if (ret < 0)
676 return ret;
677
678 for (i = 0; i < nr_cpu_ids; i++) {
679 int nid = early_cpu_to_node(cpu: i);
680
681 if (nid == NUMA_NO_NODE)
682 continue;
683 if (!node_online(nid))
684 numa_clear_node(cpu: i);
685 }
686 numa_init_array();
687
688 return 0;
689}
690
691/**
692 * dummy_numa_init - Fallback dummy NUMA init
693 *
694 * Used if there's no underlying NUMA architecture, NUMA initialization
695 * fails, or NUMA is disabled on the command line.
696 *
697 * Must online at least one node and add memory blocks that cover all
698 * allowed memory. This function must not fail.
699 */
700static int __init dummy_numa_init(void)
701{
702 printk(KERN_INFO "%s\n",
703 numa_off ? "NUMA turned off" : "No NUMA configuration found");
704 printk(KERN_INFO "Faking a node at [mem %#018Lx-%#018Lx]\n",
705 0LLU, PFN_PHYS(max_pfn) - 1);
706
707 node_set(0, numa_nodes_parsed);
708 numa_add_memblk(nid: 0, start: 0, PFN_PHYS(max_pfn));
709
710 return 0;
711}
712
713/**
714 * x86_numa_init - Initialize NUMA
715 *
716 * Try each configured NUMA initialization method until one succeeds. The
717 * last fallback is dummy single node config encompassing whole memory and
718 * never fails.
719 */
720void __init x86_numa_init(void)
721{
722 if (!numa_off) {
723#ifdef CONFIG_ACPI_NUMA
724 if (!numa_init(init_func: x86_acpi_numa_init))
725 return;
726#endif
727#ifdef CONFIG_AMD_NUMA
728 if (!numa_init(init_func: amd_numa_init))
729 return;
730#endif
731 if (acpi_disabled && !numa_init(init_func: of_numa_init))
732 return;
733 }
734
735 numa_init(init_func: dummy_numa_init);
736}
737
738
739/*
740 * A node may exist which has one or more Generic Initiators but no CPUs and no
741 * memory.
742 *
743 * This function must be called after init_cpu_to_node(), to ensure that any
744 * memoryless CPU nodes have already been brought online, and before the
745 * node_data[nid] is needed for zone list setup in build_all_zonelists().
746 *
747 * When this function is called, any nodes containing either memory and/or CPUs
748 * will already be online and there is no need to do anything extra, even if
749 * they also contain one or more Generic Initiators.
750 */
751void __init init_gi_nodes(void)
752{
753 int nid;
754
755 /*
756 * Exclude this node from
757 * bringup_nonboot_cpus
758 * cpu_up
759 * __try_online_node
760 * register_one_node
761 * because node_subsys is not initialized yet.
762 * TODO remove dependency on node_online
763 */
764 for_each_node_state(nid, N_GENERIC_INITIATOR)
765 if (!node_online(nid))
766 node_set_online(nid);
767}
768
769/*
770 * Setup early cpu_to_node.
771 *
772 * Populate cpu_to_node[] only if x86_cpu_to_apicid[],
773 * and apicid_to_node[] tables have valid entries for a CPU.
774 * This means we skip cpu_to_node[] initialisation for NUMA
775 * emulation and faking node case (when running a kernel compiled
776 * for NUMA on a non NUMA box), which is OK as cpu_to_node[]
777 * is already initialized in a round robin manner at numa_init_array,
778 * prior to this call, and this initialization is good enough
779 * for the fake NUMA cases.
780 *
781 * Called before the per_cpu areas are setup.
782 */
783void __init init_cpu_to_node(void)
784{
785 int cpu;
786 u32 *cpu_to_apicid = early_per_cpu_ptr(x86_cpu_to_apicid);
787
788 BUG_ON(cpu_to_apicid == NULL);
789
790 for_each_possible_cpu(cpu) {
791 int node = numa_cpu_node(cpu);
792
793 if (node == NUMA_NO_NODE)
794 continue;
795
796 /*
797 * Exclude this node from
798 * bringup_nonboot_cpus
799 * cpu_up
800 * __try_online_node
801 * register_one_node
802 * because node_subsys is not initialized yet.
803 * TODO remove dependency on node_online
804 */
805 if (!node_online(node))
806 node_set_online(nid: node);
807
808 numa_set_node(cpu, node);
809 }
810}
811
812#ifndef CONFIG_DEBUG_PER_CPU_MAPS
813
814# ifndef CONFIG_NUMA_EMU
815void numa_add_cpu(int cpu)
816{
817 cpumask_set_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
818}
819
820void numa_remove_cpu(int cpu)
821{
822 cpumask_clear_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
823}
824# endif /* !CONFIG_NUMA_EMU */
825
826#else /* !CONFIG_DEBUG_PER_CPU_MAPS */
827
828int __cpu_to_node(int cpu)
829{
830 if (early_per_cpu_ptr(x86_cpu_to_node_map)) {
831 printk(KERN_WARNING
832 "cpu_to_node(%d): usage too early!\n", cpu);
833 dump_stack();
834 return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
835 }
836 return per_cpu(x86_cpu_to_node_map, cpu);
837}
838EXPORT_SYMBOL(__cpu_to_node);
839
840/*
841 * Same function as cpu_to_node() but used if called before the
842 * per_cpu areas are setup.
843 */
844int early_cpu_to_node(int cpu)
845{
846 if (early_per_cpu_ptr(x86_cpu_to_node_map))
847 return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
848
849 if (!cpu_possible(cpu)) {
850 printk(KERN_WARNING
851 "early_cpu_to_node(%d): no per_cpu area!\n", cpu);
852 dump_stack();
853 return NUMA_NO_NODE;
854 }
855 return per_cpu(x86_cpu_to_node_map, cpu);
856}
857
858void debug_cpumask_set_cpu(int cpu, int node, bool enable)
859{
860 struct cpumask *mask;
861
862 if (node == NUMA_NO_NODE) {
863 /* early_cpu_to_node() already emits a warning and trace */
864 return;
865 }
866 mask = node_to_cpumask_map[node];
867 if (!cpumask_available(mask)) {
868 pr_err("node_to_cpumask_map[%i] NULL\n", node);
869 dump_stack();
870 return;
871 }
872
873 if (enable)
874 cpumask_set_cpu(cpu, dstp: mask);
875 else
876 cpumask_clear_cpu(cpu, dstp: mask);
877
878 printk(KERN_DEBUG "%s cpu %d node %d: mask now %*pbl\n",
879 enable ? "numa_add_cpu" : "numa_remove_cpu",
880 cpu, node, cpumask_pr_args(mask));
881 return;
882}
883
884# ifndef CONFIG_NUMA_EMU
885static void numa_set_cpumask(int cpu, bool enable)
886{
887 debug_cpumask_set_cpu(cpu, early_cpu_to_node(cpu), enable);
888}
889
890void numa_add_cpu(int cpu)
891{
892 numa_set_cpumask(cpu, true);
893}
894
895void numa_remove_cpu(int cpu)
896{
897 numa_set_cpumask(cpu, false);
898}
899# endif /* !CONFIG_NUMA_EMU */
900
901/*
902 * Returns a pointer to the bitmask of CPUs on Node 'node'.
903 */
904const struct cpumask *cpumask_of_node(int node)
905{
906 if ((unsigned)node >= nr_node_ids) {
907 printk(KERN_WARNING
908 "cpumask_of_node(%d): (unsigned)node >= nr_node_ids(%u)\n",
909 node, nr_node_ids);
910 dump_stack();
911 return cpu_none_mask;
912 }
913 if (!cpumask_available(mask: node_to_cpumask_map[node])) {
914 printk(KERN_WARNING
915 "cpumask_of_node(%d): no node_to_cpumask_map!\n",
916 node);
917 dump_stack();
918 return cpu_online_mask;
919 }
920 return node_to_cpumask_map[node];
921}
922EXPORT_SYMBOL(cpumask_of_node);
923
924#endif /* !CONFIG_DEBUG_PER_CPU_MAPS */
925
926#ifdef CONFIG_NUMA_KEEP_MEMINFO
927static int meminfo_to_nid(struct numa_meminfo *mi, u64 start)
928{
929 int i;
930
931 for (i = 0; i < mi->nr_blks; i++)
932 if (mi->blk[i].start <= start && mi->blk[i].end > start)
933 return mi->blk[i].nid;
934 return NUMA_NO_NODE;
935}
936
937int phys_to_target_node(phys_addr_t start)
938{
939 int nid = meminfo_to_nid(mi: &numa_meminfo, start);
940
941 /*
942 * Prefer online nodes, but if reserved memory might be
943 * hot-added continue the search with reserved ranges.
944 */
945 if (nid != NUMA_NO_NODE)
946 return nid;
947
948 return meminfo_to_nid(mi: &numa_reserved_meminfo, start);
949}
950EXPORT_SYMBOL_GPL(phys_to_target_node);
951
952int memory_add_physaddr_to_nid(u64 start)
953{
954 int nid = meminfo_to_nid(mi: &numa_meminfo, start);
955
956 if (nid == NUMA_NO_NODE)
957 nid = numa_meminfo.blk[0].nid;
958 return nid;
959}
960EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
961
962static int __init cmp_memblk(const void *a, const void *b)
963{
964 const struct numa_memblk *ma = *(const struct numa_memblk **)a;
965 const struct numa_memblk *mb = *(const struct numa_memblk **)b;
966
967 return ma->start - mb->start;
968}
969
970static struct numa_memblk *numa_memblk_list[NR_NODE_MEMBLKS] __initdata;
971
972/**
973 * numa_fill_memblks - Fill gaps in numa_meminfo memblks
974 * @start: address to begin fill
975 * @end: address to end fill
976 *
977 * Find and extend numa_meminfo memblks to cover the @start-@end
978 * physical address range, such that the first memblk includes
979 * @start, the last memblk includes @end, and any gaps in between
980 * are filled.
981 *
982 * RETURNS:
983 * 0 : Success
984 * NUMA_NO_MEMBLK : No memblk exists in @start-@end range
985 */
986
987int __init numa_fill_memblks(u64 start, u64 end)
988{
989 struct numa_memblk **blk = &numa_memblk_list[0];
990 struct numa_meminfo *mi = &numa_meminfo;
991 int count = 0;
992 u64 prev_end;
993
994 /*
995 * Create a list of pointers to numa_meminfo memblks that
996 * overlap start, end. Exclude (start == bi->end) since
997 * end addresses in both a CFMWS range and a memblk range
998 * are exclusive.
999 *
1000 * This list of pointers is used to make in-place changes
1001 * that fill out the numa_meminfo memblks.
1002 */
1003 for (int i = 0; i < mi->nr_blks; i++) {
1004 struct numa_memblk *bi = &mi->blk[i];
1005
1006 if (start < bi->end && end >= bi->start) {
1007 blk[count] = &mi->blk[i];
1008 count++;
1009 }
1010 }
1011 if (!count)
1012 return NUMA_NO_MEMBLK;
1013
1014 /* Sort the list of pointers in memblk->start order */
1015 sort(base: &blk[0], num: count, size: sizeof(blk[0]), cmp_func: cmp_memblk, NULL);
1016
1017 /* Make sure the first/last memblks include start/end */
1018 blk[0]->start = min(blk[0]->start, start);
1019 blk[count - 1]->end = max(blk[count - 1]->end, end);
1020
1021 /*
1022 * Fill any gaps by tracking the previous memblks
1023 * end address and backfilling to it if needed.
1024 */
1025 prev_end = blk[0]->end;
1026 for (int i = 1; i < count; i++) {
1027 struct numa_memblk *curr = blk[i];
1028
1029 if (prev_end >= curr->start) {
1030 if (prev_end < curr->end)
1031 prev_end = curr->end;
1032 } else {
1033 curr->start = prev_end;
1034 prev_end = curr->end;
1035 }
1036 }
1037 return 0;
1038}
1039
1040#endif
1041

source code of linux/arch/x86/mm/numa.c