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