1// SPDX-License-Identifier: GPL-2.0-or-later
2
3#include <linux/array_size.h>
4#include <linux/sort.h>
5#include <linux/printk.h>
6#include <linux/memblock.h>
7#include <linux/numa.h>
8#include <linux/numa_memblks.h>
9
10int numa_distance_cnt;
11static u8 *numa_distance;
12
13nodemask_t numa_nodes_parsed __initdata;
14
15static struct numa_meminfo numa_meminfo __initdata_or_meminfo;
16static struct numa_meminfo numa_reserved_meminfo __initdata_or_meminfo;
17
18/*
19 * Set nodes, which have memory in @mi, in *@nodemask.
20 */
21static void __init numa_nodemask_from_meminfo(nodemask_t *nodemask,
22 const struct numa_meminfo *mi)
23{
24 int i;
25
26 for (i = 0; i < ARRAY_SIZE(mi->blk); i++)
27 if (mi->blk[i].start != mi->blk[i].end &&
28 mi->blk[i].nid != NUMA_NO_NODE)
29 node_set(mi->blk[i].nid, *nodemask);
30}
31
32/**
33 * numa_reset_distance - Reset NUMA distance table
34 *
35 * The current table is freed. The next numa_set_distance() call will
36 * create a new one.
37 */
38void __init numa_reset_distance(void)
39{
40 size_t size = numa_distance_cnt * numa_distance_cnt * sizeof(numa_distance[0]);
41
42 /* numa_distance could be 1LU marking allocation failure, test cnt */
43 if (numa_distance_cnt)
44 memblock_free(ptr: numa_distance, size);
45 numa_distance_cnt = 0;
46 numa_distance = NULL; /* enable table creation */
47}
48
49static int __init numa_alloc_distance(void)
50{
51 nodemask_t nodes_parsed;
52 size_t size;
53 int i, j, cnt = 0;
54
55 /* size the new table and allocate it */
56 nodes_parsed = numa_nodes_parsed;
57 numa_nodemask_from_meminfo(nodemask: &nodes_parsed, mi: &numa_meminfo);
58
59 for_each_node_mask(i, nodes_parsed)
60 cnt = i;
61 cnt++;
62 size = cnt * cnt * sizeof(numa_distance[0]);
63
64 numa_distance = memblock_alloc(size, PAGE_SIZE);
65 if (!numa_distance) {
66 pr_warn("Warning: can't allocate distance table!\n");
67 /* don't retry until explicitly reset */
68 numa_distance = (void *)1LU;
69 return -ENOMEM;
70 }
71
72 numa_distance_cnt = cnt;
73
74 /* fill with the default distances */
75 for (i = 0; i < cnt; i++)
76 for (j = 0; j < cnt; j++)
77 numa_distance[i * cnt + j] = i == j ?
78 LOCAL_DISTANCE : REMOTE_DISTANCE;
79 printk(KERN_DEBUG "NUMA: Initialized distance table, cnt=%d\n", cnt);
80
81 return 0;
82}
83
84/**
85 * numa_set_distance - Set NUMA distance from one NUMA to another
86 * @from: the 'from' node to set distance
87 * @to: the 'to' node to set distance
88 * @distance: NUMA distance
89 *
90 * Set the distance from node @from to @to to @distance. If distance table
91 * doesn't exist, one which is large enough to accommodate all the currently
92 * known nodes will be created.
93 *
94 * If such table cannot be allocated, a warning is printed and further
95 * calls are ignored until the distance table is reset with
96 * numa_reset_distance().
97 *
98 * If @from or @to is higher than the highest known node or lower than zero
99 * at the time of table creation or @distance doesn't make sense, the call
100 * is ignored.
101 * This is to allow simplification of specific NUMA config implementations.
102 */
103void __init numa_set_distance(int from, int to, int distance)
104{
105 if (!numa_distance && numa_alloc_distance() < 0)
106 return;
107
108 if (from >= numa_distance_cnt || to >= numa_distance_cnt ||
109 from < 0 || to < 0) {
110 pr_warn_once("Warning: node ids are out of bound, from=%d to=%d distance=%d\n",
111 from, to, distance);
112 return;
113 }
114
115 if ((u8)distance != distance ||
116 (from == to && distance != LOCAL_DISTANCE)) {
117 pr_warn_once("Warning: invalid distance parameter, from=%d to=%d distance=%d\n",
118 from, to, distance);
119 return;
120 }
121
122 numa_distance[from * numa_distance_cnt + to] = distance;
123}
124
125int __node_distance(int from, int to)
126{
127 if (from >= numa_distance_cnt || to >= numa_distance_cnt)
128 return from == to ? LOCAL_DISTANCE : REMOTE_DISTANCE;
129 return numa_distance[from * numa_distance_cnt + to];
130}
131EXPORT_SYMBOL(__node_distance);
132
133static int __init numa_add_memblk_to(int nid, u64 start, u64 end,
134 struct numa_meminfo *mi)
135{
136 /* ignore zero length blks */
137 if (start == end)
138 return 0;
139
140 /* whine about and ignore invalid blks */
141 if (start > end || nid < 0 || nid >= MAX_NUMNODES) {
142 pr_warn("Warning: invalid memblk node %d [mem %#010Lx-%#010Lx]\n",
143 nid, start, end - 1);
144 return 0;
145 }
146
147 if (mi->nr_blks >= NR_NODE_MEMBLKS) {
148 pr_err("too many memblk ranges\n");
149 return -EINVAL;
150 }
151
152 mi->blk[mi->nr_blks].start = start;
153 mi->blk[mi->nr_blks].end = end;
154 mi->blk[mi->nr_blks].nid = nid;
155 mi->nr_blks++;
156 return 0;
157}
158
159/**
160 * numa_remove_memblk_from - Remove one numa_memblk from a numa_meminfo
161 * @idx: Index of memblk to remove
162 * @mi: numa_meminfo to remove memblk from
163 *
164 * Remove @idx'th numa_memblk from @mi by shifting @mi->blk[] and
165 * decrementing @mi->nr_blks.
166 */
167void __init numa_remove_memblk_from(int idx, struct numa_meminfo *mi)
168{
169 mi->nr_blks--;
170 memmove(&mi->blk[idx], &mi->blk[idx + 1],
171 (mi->nr_blks - idx) * sizeof(mi->blk[0]));
172}
173
174/**
175 * numa_move_tail_memblk - Move a numa_memblk from one numa_meminfo to another
176 * @dst: numa_meminfo to append block to
177 * @idx: Index of memblk to remove
178 * @src: numa_meminfo to remove memblk from
179 */
180static void __init numa_move_tail_memblk(struct numa_meminfo *dst, int idx,
181 struct numa_meminfo *src)
182{
183 dst->blk[dst->nr_blks++] = src->blk[idx];
184 numa_remove_memblk_from(idx, mi: src);
185}
186
187/**
188 * numa_add_memblk - Add one numa_memblk to numa_meminfo
189 * @nid: NUMA node ID of the new memblk
190 * @start: Start address of the new memblk
191 * @end: End address of the new memblk
192 *
193 * Add a new memblk to the default numa_meminfo.
194 *
195 * RETURNS:
196 * 0 on success, -errno on failure.
197 */
198int __init numa_add_memblk(int nid, u64 start, u64 end)
199{
200 return numa_add_memblk_to(nid, start, end, mi: &numa_meminfo);
201}
202
203/**
204 * numa_add_reserved_memblk - Add one numa_memblk to numa_reserved_meminfo
205 * @nid: NUMA node ID of the new memblk
206 * @start: Start address of the new memblk
207 * @end: End address of the new memblk
208 *
209 * Add a new memblk to the numa_reserved_meminfo.
210 *
211 * Usage Case: numa_cleanup_meminfo() reconciles all numa_memblk instances
212 * against memblock_type information and moves any that intersect reserved
213 * ranges to numa_reserved_meminfo. However, when that information is known
214 * ahead of time, we use numa_add_reserved_memblk() to add the numa_memblk
215 * to numa_reserved_meminfo directly.
216 *
217 * RETURNS:
218 * 0 on success, -errno on failure.
219 */
220int __init numa_add_reserved_memblk(int nid, u64 start, u64 end)
221{
222 return numa_add_memblk_to(nid, start, end, mi: &numa_reserved_meminfo);
223}
224
225/**
226 * numa_cleanup_meminfo - Cleanup a numa_meminfo
227 * @mi: numa_meminfo to clean up
228 *
229 * Sanitize @mi by merging and removing unnecessary memblks. Also check for
230 * conflicts and clear unused memblks.
231 *
232 * RETURNS:
233 * 0 on success, -errno on failure.
234 */
235int __init numa_cleanup_meminfo(struct numa_meminfo *mi)
236{
237 const u64 low = memblock_start_of_DRAM();
238 const u64 high = memblock_end_of_DRAM();
239 int i, j, k;
240
241 /* first, trim all entries */
242 for (i = 0; i < mi->nr_blks; i++) {
243 struct numa_memblk *bi = &mi->blk[i];
244
245 /* move / save reserved memory ranges */
246 if (!memblock_overlaps_region(type: &memblock.memory,
247 base: bi->start, size: bi->end - bi->start)) {
248 numa_move_tail_memblk(dst: &numa_reserved_meminfo, idx: i--, src: mi);
249 continue;
250 }
251
252 /* make sure all non-reserved blocks are inside the limits */
253 bi->start = max(bi->start, low);
254
255 /* preserve info for non-RAM areas above 'max_pfn': */
256 if (bi->end > high) {
257 numa_add_memblk_to(nid: bi->nid, start: high, end: bi->end,
258 mi: &numa_reserved_meminfo);
259 bi->end = high;
260 }
261
262 /* and there's no empty block */
263 if (bi->start >= bi->end)
264 numa_remove_memblk_from(idx: i--, mi);
265 }
266
267 /* merge neighboring / overlapping entries */
268 for (i = 0; i < mi->nr_blks; i++) {
269 struct numa_memblk *bi = &mi->blk[i];
270
271 for (j = i + 1; j < mi->nr_blks; j++) {
272 struct numa_memblk *bj = &mi->blk[j];
273 u64 start, end;
274
275 /*
276 * See whether there are overlapping blocks. Whine
277 * about but allow overlaps of the same nid. They
278 * will be merged below.
279 */
280 if (bi->end > bj->start && bi->start < bj->end) {
281 if (bi->nid != bj->nid) {
282 pr_err("node %d [mem %#010Lx-%#010Lx] overlaps with node %d [mem %#010Lx-%#010Lx]\n",
283 bi->nid, bi->start, bi->end - 1,
284 bj->nid, bj->start, bj->end - 1);
285 return -EINVAL;
286 }
287 pr_warn("Warning: node %d [mem %#010Lx-%#010Lx] overlaps with itself [mem %#010Lx-%#010Lx]\n",
288 bi->nid, bi->start, bi->end - 1,
289 bj->start, bj->end - 1);
290 }
291
292 /*
293 * Join together blocks on the same node, holes
294 * between which don't overlap with memory on other
295 * nodes.
296 */
297 if (bi->nid != bj->nid)
298 continue;
299 start = min(bi->start, bj->start);
300 end = max(bi->end, bj->end);
301 for (k = 0; k < mi->nr_blks; k++) {
302 struct numa_memblk *bk = &mi->blk[k];
303
304 if (bi->nid == bk->nid)
305 continue;
306 if (start < bk->end && end > bk->start)
307 break;
308 }
309 if (k < mi->nr_blks)
310 continue;
311 pr_info("NUMA: Node %d [mem %#010Lx-%#010Lx] + [mem %#010Lx-%#010Lx] -> [mem %#010Lx-%#010Lx]\n",
312 bi->nid, bi->start, bi->end - 1, bj->start,
313 bj->end - 1, start, end - 1);
314 bi->start = start;
315 bi->end = end;
316 numa_remove_memblk_from(idx: j--, mi);
317 }
318 }
319
320 /* clear unused ones */
321 for (i = mi->nr_blks; i < ARRAY_SIZE(mi->blk); i++) {
322 mi->blk[i].start = mi->blk[i].end = 0;
323 mi->blk[i].nid = NUMA_NO_NODE;
324 }
325
326 return 0;
327}
328
329/*
330 * Mark all currently memblock-reserved physical memory (which covers the
331 * kernel's own memory ranges) as hot-unswappable.
332 */
333static void __init numa_clear_kernel_node_hotplug(void)
334{
335 nodemask_t reserved_nodemask = NODE_MASK_NONE;
336 struct memblock_region *mb_region;
337 int i;
338
339 /*
340 * We have to do some preprocessing of memblock regions, to
341 * make them suitable for reservation.
342 *
343 * At this time, all memory regions reserved by memblock are
344 * used by the kernel, but those regions are not split up
345 * along node boundaries yet, and don't necessarily have their
346 * node ID set yet either.
347 *
348 * So iterate over all parsed memory blocks and use those ranges to
349 * set the nid in memblock.reserved. This will split up the
350 * memblock regions along node boundaries and will set the node IDs
351 * as well.
352 */
353 for (i = 0; i < numa_meminfo.nr_blks; i++) {
354 struct numa_memblk *mb = numa_meminfo.blk + i;
355 int ret;
356
357 ret = memblock_set_node(base: mb->start, size: mb->end - mb->start,
358 type: &memblock.reserved, nid: mb->nid);
359 WARN_ON_ONCE(ret);
360 }
361
362 /*
363 * Now go over all reserved memblock regions, to construct a
364 * node mask of all kernel reserved memory areas.
365 *
366 * [ Note, when booting with mem=nn[kMG] or in a kdump kernel,
367 * numa_meminfo might not include all memblock.reserved
368 * memory ranges, because quirks such as trim_snb_memory()
369 * reserve specific pages for Sandy Bridge graphics. ]
370 */
371 for_each_reserved_mem_region(mb_region) {
372 int nid = memblock_get_region_node(r: mb_region);
373
374 if (numa_valid_node(nid))
375 node_set(nid, reserved_nodemask);
376 }
377
378 /*
379 * Finally, clear the MEMBLOCK_HOTPLUG flag for all memory
380 * belonging to the reserved node mask.
381 *
382 * Note that this will include memory regions that reside
383 * on nodes that contain kernel memory - entire nodes
384 * become hot-unpluggable:
385 */
386 for (i = 0; i < numa_meminfo.nr_blks; i++) {
387 struct numa_memblk *mb = numa_meminfo.blk + i;
388
389 if (!node_isset(mb->nid, reserved_nodemask))
390 continue;
391
392 memblock_clear_hotplug(base: mb->start, size: mb->end - mb->start);
393 }
394}
395
396static int __init numa_register_meminfo(struct numa_meminfo *mi)
397{
398 int i;
399
400 /* Account for nodes with cpus and no memory */
401 node_possible_map = numa_nodes_parsed;
402 numa_nodemask_from_meminfo(nodemask: &node_possible_map, mi);
403 if (WARN_ON(nodes_empty(node_possible_map)))
404 return -EINVAL;
405
406 for (i = 0; i < mi->nr_blks; i++) {
407 struct numa_memblk *mb = &mi->blk[i];
408
409 memblock_set_node(base: mb->start, size: mb->end - mb->start,
410 type: &memblock.memory, nid: mb->nid);
411 }
412
413 /*
414 * At very early time, the kernel have to use some memory such as
415 * loading the kernel image. We cannot prevent this anyway. So any
416 * node the kernel resides in should be un-hotpluggable.
417 *
418 * And when we come here, alloc node data won't fail.
419 */
420 numa_clear_kernel_node_hotplug();
421
422 /*
423 * If sections array is gonna be used for pfn -> nid mapping, check
424 * whether its granularity is fine enough.
425 */
426 if (IS_ENABLED(NODE_NOT_IN_PAGE_FLAGS)) {
427 unsigned long pfn_align = node_map_pfn_alignment();
428
429 if (pfn_align && pfn_align < PAGES_PER_SECTION) {
430 unsigned long node_align_mb = PFN_PHYS(pfn_align) >> 20;
431
432 unsigned long sect_align_mb = PFN_PHYS(PAGES_PER_SECTION) >> 20;
433
434 pr_warn("Node alignment %luMB < min %luMB, rejecting NUMA config\n",
435 node_align_mb, sect_align_mb);
436 return -EINVAL;
437 }
438 }
439
440 return 0;
441}
442
443int __init numa_memblks_init(int (*init_func)(void),
444 bool memblock_force_top_down)
445{
446 phys_addr_t max_addr = (phys_addr_t)ULLONG_MAX;
447 int ret;
448
449 nodes_clear(numa_nodes_parsed);
450 nodes_clear(node_possible_map);
451 nodes_clear(node_online_map);
452 memset(&numa_meminfo, 0, sizeof(numa_meminfo));
453 WARN_ON(memblock_set_node(0, max_addr, &memblock.memory, NUMA_NO_NODE));
454 WARN_ON(memblock_set_node(0, max_addr, &memblock.reserved,
455 NUMA_NO_NODE));
456 /* In case that parsing SRAT failed. */
457 WARN_ON(memblock_clear_hotplug(0, max_addr));
458 numa_reset_distance();
459
460 ret = init_func();
461 if (ret < 0)
462 return ret;
463
464 /*
465 * We reset memblock back to the top-down direction
466 * here because if we configured ACPI_NUMA, we have
467 * parsed SRAT in init_func(). It is ok to have the
468 * reset here even if we did't configure ACPI_NUMA
469 * or acpi numa init fails and fallbacks to dummy
470 * numa init.
471 */
472 if (memblock_force_top_down)
473 memblock_set_bottom_up(enable: false);
474
475 ret = numa_cleanup_meminfo(mi: &numa_meminfo);
476 if (ret < 0)
477 return ret;
478
479 numa_emulation(numa_meminfo: &numa_meminfo, numa_dist_cnt: numa_distance_cnt);
480
481 return numa_register_meminfo(mi: &numa_meminfo);
482}
483
484static int __init cmp_memblk(const void *a, const void *b)
485{
486 const struct numa_memblk *ma = *(const struct numa_memblk **)a;
487 const struct numa_memblk *mb = *(const struct numa_memblk **)b;
488
489 return (ma->start > mb->start) - (ma->start < mb->start);
490}
491
492static struct numa_memblk *numa_memblk_list[NR_NODE_MEMBLKS] __initdata;
493
494/**
495 * numa_fill_memblks - Fill gaps in numa_meminfo memblks
496 * @start: address to begin fill
497 * @end: address to end fill
498 *
499 * Find and extend numa_meminfo memblks to cover the physical
500 * address range @start-@end
501 *
502 * RETURNS:
503 * 0 : Success
504 * NUMA_NO_MEMBLK : No memblks exist in address range @start-@end
505 */
506
507int __init numa_fill_memblks(u64 start, u64 end)
508{
509 struct numa_memblk **blk = &numa_memblk_list[0];
510 struct numa_meminfo *mi = &numa_meminfo;
511 int count = 0;
512 u64 prev_end;
513
514 /*
515 * Create a list of pointers to numa_meminfo memblks that
516 * overlap start, end. The list is used to make in-place
517 * changes that fill out the numa_meminfo memblks.
518 */
519 for (int i = 0; i < mi->nr_blks; i++) {
520 struct numa_memblk *bi = &mi->blk[i];
521
522 if (memblock_addrs_overlap(base1: start, size1: end - start, base2: bi->start,
523 size2: bi->end - bi->start)) {
524 blk[count] = &mi->blk[i];
525 count++;
526 }
527 }
528 if (!count)
529 return NUMA_NO_MEMBLK;
530
531 /* Sort the list of pointers in memblk->start order */
532 sort(base: &blk[0], num: count, size: sizeof(blk[0]), cmp_func: cmp_memblk, NULL);
533
534 /* Make sure the first/last memblks include start/end */
535 blk[0]->start = min(blk[0]->start, start);
536 blk[count - 1]->end = max(blk[count - 1]->end, end);
537
538 /*
539 * Fill any gaps by tracking the previous memblks
540 * end address and backfilling to it if needed.
541 */
542 prev_end = blk[0]->end;
543 for (int i = 1; i < count; i++) {
544 struct numa_memblk *curr = blk[i];
545
546 if (prev_end >= curr->start) {
547 if (prev_end < curr->end)
548 prev_end = curr->end;
549 } else {
550 curr->start = prev_end;
551 prev_end = curr->end;
552 }
553 }
554 return 0;
555}
556
557#ifdef CONFIG_NUMA_KEEP_MEMINFO
558static int meminfo_to_nid(struct numa_meminfo *mi, u64 start)
559{
560 int i;
561
562 for (i = 0; i < mi->nr_blks; i++)
563 if (mi->blk[i].start <= start && mi->blk[i].end > start)
564 return mi->blk[i].nid;
565 return NUMA_NO_NODE;
566}
567
568int phys_to_target_node(u64 start)
569{
570 int nid = meminfo_to_nid(mi: &numa_meminfo, start);
571
572 /*
573 * Prefer online nodes, but if reserved memory might be
574 * hot-added continue the search with reserved ranges.
575 */
576 if (nid != NUMA_NO_NODE)
577 return nid;
578
579 return meminfo_to_nid(mi: &numa_reserved_meminfo, start);
580}
581EXPORT_SYMBOL_GPL(phys_to_target_node);
582
583int memory_add_physaddr_to_nid(u64 start)
584{
585 int nid = meminfo_to_nid(mi: &numa_meminfo, start);
586
587 if (nid == NUMA_NO_NODE)
588 nid = numa_meminfo.blk[0].nid;
589 return nid;
590}
591EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
592
593#endif /* CONFIG_NUMA_KEEP_MEMINFO */
594

source code of linux/mm/numa_memblks.c