numa_memblks.c

  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
 10#include <asm/numa.h>
 11
 12int numa_distance_cnt;
 13static u8 *numa_distance;
 14
 15nodemask_t numa_nodes_parsed __initdata;
 16
 17static struct numa_meminfo numa_meminfo __initdata_or_meminfo;
 18static struct numa_meminfo numa_reserved_meminfo __initdata_or_meminfo;
 19
 20/*
 21 * Set nodes, which have memory in @mi, in *@nodemask.
 22 */
 23static void __init numa_nodemask_from_meminfo(nodemask_t *nodemask,
 24					      const struct numa_meminfo *mi)
 25{
 26	int i;
 27
 28	for (i = 0; i < ARRAY_SIZE(mi->blk); i++)
 29		if (mi->blk[i].start != mi->blk[i].end &&
 30		    mi->blk[i].nid != NUMA_NO_NODE)
 31			node_set(mi->blk[i].nid, *nodemask);
 32}
 33
 34/**
 35 * numa_reset_distance - Reset NUMA distance table
 36 *
 37 * The current table is freed.  The next numa_set_distance() call will
 38 * create a new one.
 39 */
 40void __init numa_reset_distance(void)
 41{
 42	size_t size = numa_distance_cnt * numa_distance_cnt * sizeof(numa_distance[0]);
 43
 44	/* numa_distance could be 1LU marking allocation failure, test cnt */
 45	if (numa_distance_cnt)
 46		memblock_free(numa_distance, size);
 47	numa_distance_cnt = 0;
 48	numa_distance = NULL;	/* enable table creation */
 49}
 50
 51static int __init numa_alloc_distance(void)
 52{
 53	nodemask_t nodes_parsed;
 54	size_t size;
 55	int i, j, cnt = 0;
 56
 57	/* size the new table and allocate it */
 58	nodes_parsed = numa_nodes_parsed;
 59	numa_nodemask_from_meminfo(&nodes_parsed, &numa_meminfo);
 60
 61	for_each_node_mask(i, nodes_parsed)
 62		cnt = i;
 63	cnt++;
 64	size = cnt * cnt * sizeof(numa_distance[0]);
 65
 66	numa_distance = memblock_alloc(size, PAGE_SIZE);
 67	if (!numa_distance) {
 68		pr_warn("Warning: can't allocate distance table!\n");
 69		/* don't retry until explicitly reset */
 70		numa_distance = (void *)1LU;
 71		return -ENOMEM;
 72	}
 73
 74	numa_distance_cnt = cnt;
 75
 76	/* fill with the default distances */
 77	for (i = 0; i < cnt; i++)
 78		for (j = 0; j < cnt; j++)
 79			numa_distance[i * cnt + j] = i == j ?
 80				LOCAL_DISTANCE : REMOTE_DISTANCE;
 81	pr_debug("NUMA: Initialized distance table, cnt=%d\n", cnt);
 82
 83	return 0;
 84}
 85
 86/**
 87 * numa_set_distance - Set NUMA distance from one NUMA to another
 88 * @from: the 'from' node to set distance
 89 * @to: the 'to'  node to set distance
 90 * @distance: NUMA distance
 91 *
 92 * Set the distance from node @from to @to to @distance.  If distance table
 93 * doesn't exist, one which is large enough to accommodate all the currently
 94 * known nodes will be created.
 95 *
 96 * If such table cannot be allocated, a warning is printed and further
 97 * calls are ignored until the distance table is reset with
 98 * numa_reset_distance().
 99 *
100 * If @from or @to is higher than the highest known node or lower than zero
101 * at the time of table creation or @distance doesn't make sense, the call
102 * is ignored.
103 * This is to allow simplification of specific NUMA config implementations.
104 */
105void __init numa_set_distance(int from, int to, int distance)
106{
107	if (!numa_distance && numa_alloc_distance() < 0)
108		return;
109
110	if (from >= numa_distance_cnt || to >= numa_distance_cnt ||
111			from < 0 || to < 0) {
112		pr_warn_once("Warning: node ids are out of bound, from=%d to=%d distance=%d\n",
113			     from, to, distance);
114		return;
115	}
116
117	if ((u8)distance != distance ||
118	    (from == to && distance != LOCAL_DISTANCE)) {
119		pr_warn_once("Warning: invalid distance parameter, from=%d to=%d distance=%d\n",
120			     from, to, distance);
121		return;
122	}
123
124	numa_distance[from * numa_distance_cnt + to] = distance;
125}
126
127int __node_distance(int from, int to)
128{
129	if (from >= numa_distance_cnt || to >= numa_distance_cnt)
130		return from == to ? LOCAL_DISTANCE : REMOTE_DISTANCE;
131	return numa_distance[from * numa_distance_cnt + to];
132}
133EXPORT_SYMBOL(__node_distance);
134
135static int __init numa_add_memblk_to(int nid, u64 start, u64 end,
136				     struct numa_meminfo *mi)
137{
138	/* ignore zero length blks */
139	if (start == end)
140		return 0;
141
142	/* whine about and ignore invalid blks */
143	if (start > end || nid < 0 || nid >= MAX_NUMNODES) {
144		pr_warn("Warning: invalid memblk node %d [mem %#010Lx-%#010Lx]\n",
145			nid, start, end - 1);
146		return 0;
147	}
148
149	if (mi->nr_blks >= NR_NODE_MEMBLKS) {
150		pr_err("too many memblk ranges\n");
151		return -EINVAL;
152	}
153
154	mi->blk[mi->nr_blks].start = start;
155	mi->blk[mi->nr_blks].end = end;
156	mi->blk[mi->nr_blks].nid = nid;
157	mi->nr_blks++;
158	return 0;
159}
160
161/**
162 * numa_remove_memblk_from - Remove one numa_memblk from a numa_meminfo
163 * @idx: Index of memblk to remove
164 * @mi: numa_meminfo to remove memblk from
165 *
166 * Remove @idx'th numa_memblk from @mi by shifting @mi->blk[] and
167 * decrementing @mi->nr_blks.
168 */
169void __init numa_remove_memblk_from(int idx, struct numa_meminfo *mi)
170{
171	mi->nr_blks--;
172	memmove(&mi->blk[idx], &mi->blk[idx + 1],
173		(mi->nr_blks - idx) * sizeof(mi->blk[0]));
174}
175
176/**
177 * numa_move_tail_memblk - Move a numa_memblk from one numa_meminfo to another
178 * @dst: numa_meminfo to append block to
179 * @idx: Index of memblk to remove
180 * @src: numa_meminfo to remove memblk from
181 */
182static void __init numa_move_tail_memblk(struct numa_meminfo *dst, int idx,
183					 struct numa_meminfo *src)
184{
185	dst->blk[dst->nr_blks++] = src->blk[idx];
186	numa_remove_memblk_from(idx, src);
187}
188
189/**
190 * numa_add_memblk - Add one numa_memblk to numa_meminfo
191 * @nid: NUMA node ID of the new memblk
192 * @start: Start address of the new memblk
193 * @end: End address of the new memblk
194 *
195 * Add a new memblk to the default numa_meminfo.
196 *
197 * RETURNS:
198 * 0 on success, -errno on failure.
199 */
200int __init numa_add_memblk(int nid, u64 start, u64 end)
201{
202	return numa_add_memblk_to(nid, start, end, &numa_meminfo);
203}
204
205/**
206 * numa_add_reserved_memblk - Add one numa_memblk to numa_reserved_meminfo
207 * @nid: NUMA node ID of the new memblk
208 * @start: Start address of the new memblk
209 * @end: End address of the new memblk
210 *
211 * Add a new memblk to the numa_reserved_meminfo.
212 *
213 * Usage Case: numa_cleanup_meminfo() reconciles all numa_memblk instances
214 * against memblock_type information and moves any that intersect reserved
215 * ranges to numa_reserved_meminfo. However, when that information is known
216 * ahead of time, we use numa_add_reserved_memblk() to add the numa_memblk
217 * to numa_reserved_meminfo directly.
218 *
219 * RETURNS:
220 * 0 on success, -errno on failure.
221 */
222int __init numa_add_reserved_memblk(int nid, u64 start, u64 end)
223{
224	return numa_add_memblk_to(nid, start, end, &numa_reserved_meminfo);
225}
226
227/**
228 * numa_cleanup_meminfo - Cleanup a numa_meminfo
229 * @mi: numa_meminfo to clean up
230 *
231 * Sanitize @mi by merging and removing unnecessary memblks.  Also check for
232 * conflicts and clear unused memblks.
233 *
234 * RETURNS:
235 * 0 on success, -errno on failure.
236 */
237int __init numa_cleanup_meminfo(struct numa_meminfo *mi)
238{
239	const u64 low = memblock_start_of_DRAM();
240	const u64 high = memblock_end_of_DRAM();
241	int i, j, k;
242
243	/* first, trim all entries */
244	for (i = 0; i < mi->nr_blks; i++) {
245		struct numa_memblk *bi = &mi->blk[i];
246
247		/* move / save reserved memory ranges */
248		if (!memblock_overlaps_region(&memblock.memory,
249					bi->start, bi->end - bi->start)) {
250			numa_move_tail_memblk(&numa_reserved_meminfo, i--, mi);
251			continue;
252		}
253
254		/* make sure all non-reserved blocks are inside the limits */
255		bi->start = max(bi->start, low);
256
257		/* preserve info for non-RAM areas above 'max_pfn': */
258		if (bi->end > high) {
259			numa_add_memblk_to(bi->nid, high, bi->end,
260					   &numa_reserved_meminfo);
261			bi->end = high;
262		}
263
264		/* and there's no empty block */
265		if (bi->start >= bi->end)
266			numa_remove_memblk_from(i--, mi);
267	}
268
269	/* merge neighboring / overlapping entries */
270	for (i = 0; i < mi->nr_blks; i++) {
271		struct numa_memblk *bi = &mi->blk[i];
272
273		for (j = i + 1; j < mi->nr_blks; j++) {
274			struct numa_memblk *bj = &mi->blk[j];
275			u64 start, end;
276
277			/*
278			 * See whether there are overlapping blocks.  Whine
279			 * about but allow overlaps of the same nid.  They
280			 * will be merged below.
281			 */
282			if (bi->end > bj->start && bi->start < bj->end) {
283				if (bi->nid != bj->nid) {
284					pr_err("node %d [mem %#010Lx-%#010Lx] overlaps with node %d [mem %#010Lx-%#010Lx]\n",
285					       bi->nid, bi->start, bi->end - 1,
286					       bj->nid, bj->start, bj->end - 1);
287					return -EINVAL;
288				}
289				pr_warn("Warning: node %d [mem %#010Lx-%#010Lx] overlaps with itself [mem %#010Lx-%#010Lx]\n",
290					bi->nid, bi->start, bi->end - 1,
291					bj->start, bj->end - 1);
292			}
293
294			/*
295			 * Join together blocks on the same node, holes
296			 * between which don't overlap with memory on other
297			 * nodes.
298			 */
299			if (bi->nid != bj->nid)
300				continue;
301			start = min(bi->start, bj->start);
302			end = max(bi->end, bj->end);
303			for (k = 0; k < mi->nr_blks; k++) {
304				struct numa_memblk *bk = &mi->blk[k];
305
306				if (bi->nid == bk->nid)
307					continue;
308				if (start < bk->end && end > bk->start)
309					break;
310			}
311			if (k < mi->nr_blks)
312				continue;
313			pr_info("NUMA: Node %d [mem %#010Lx-%#010Lx] + [mem %#010Lx-%#010Lx] -> [mem %#010Lx-%#010Lx]\n",
314			       bi->nid, bi->start, bi->end - 1, bj->start,
315			       bj->end - 1, start, end - 1);
316			bi->start = start;
317			bi->end = end;
318			numa_remove_memblk_from(j--, mi);
319		}
320	}
321
322	/* clear unused ones */
323	for (i = mi->nr_blks; i < ARRAY_SIZE(mi->blk); i++) {
324		mi->blk[i].start = mi->blk[i].end = 0;
325		mi->blk[i].nid = NUMA_NO_NODE;
326	}
327
328	return 0;
329}
330
331/*
332 * Mark all currently memblock-reserved physical memory (which covers the
333 * kernel's own memory ranges) as hot-unswappable.
334 */
335static void __init numa_clear_kernel_node_hotplug(void)
336{
337	nodemask_t reserved_nodemask = NODE_MASK_NONE;
338	struct memblock_region *mb_region;
339	int i;
340
341	/*
342	 * We have to do some preprocessing of memblock regions, to
343	 * make them suitable for reservation.
344	 *
345	 * At this time, all memory regions reserved by memblock are
346	 * used by the kernel, but those regions are not split up
347	 * along node boundaries yet, and don't necessarily have their
348	 * node ID set yet either.
349	 *
350	 * So iterate over all parsed memory blocks and use those ranges to
351	 * set the nid in memblock.reserved.  This will split up the
352	 * memblock regions along node boundaries and will set the node IDs
353	 * as well.
354	 */
355	for (i = 0; i < numa_meminfo.nr_blks; i++) {
356		struct numa_memblk *mb = numa_meminfo.blk + i;
357		int ret;
358
359		ret = memblock_set_node(mb->start, mb->end - mb->start,
360					&memblock.reserved, mb->nid);
361		WARN_ON_ONCE(ret);
362	}
363
364	/*
365	 * Now go over all reserved memblock regions, to construct a
366	 * node mask of all kernel reserved memory areas.
367	 *
368	 * [ Note, when booting with mem=nn[kMG] or in a kdump kernel,
369	 *   numa_meminfo might not include all memblock.reserved
370	 *   memory ranges, because quirks such as trim_snb_memory()
371	 *   reserve specific pages for Sandy Bridge graphics. ]
372	 */
373	for_each_reserved_mem_region(mb_region) {
374		int nid = memblock_get_region_node(mb_region);
375
376		if (numa_valid_node(nid))
377			node_set(nid, reserved_nodemask);
378	}
379
380	/*
381	 * Finally, clear the MEMBLOCK_HOTPLUG flag for all memory
382	 * belonging to the reserved node mask.
383	 *
384	 * Note that this will include memory regions that reside
385	 * on nodes that contain kernel memory - entire nodes
386	 * become hot-unpluggable:
387	 */
388	for (i = 0; i < numa_meminfo.nr_blks; i++) {
389		struct numa_memblk *mb = numa_meminfo.blk + i;
390
391		if (!node_isset(mb->nid, reserved_nodemask))
392			continue;
393
394		memblock_clear_hotplug(mb->start, mb->end - mb->start);
395	}
396}
397
398static int __init numa_register_meminfo(struct numa_meminfo *mi)
399{
400	int i;
401
402	/* Account for nodes with cpus and no memory */
403	node_possible_map = numa_nodes_parsed;
404	numa_nodemask_from_meminfo(&node_possible_map, mi);
405	if (WARN_ON(nodes_empty(node_possible_map)))
406		return -EINVAL;
407
408	for (i = 0; i < mi->nr_blks; i++) {
409		struct numa_memblk *mb = &mi->blk[i];
410
411		memblock_set_node(mb->start, mb->end - mb->start,
412				  &memblock.memory, mb->nid);
413	}
414
415	/*
416	 * At very early time, the kernel have to use some memory such as
417	 * loading the kernel image. We cannot prevent this anyway. So any
418	 * node the kernel resides in should be un-hotpluggable.
419	 *
420	 * And when we come here, alloc node data won't fail.
421	 */
422	numa_clear_kernel_node_hotplug();
423
424	/*
425	 * If sections array is gonna be used for pfn -> nid mapping, check
426	 * whether its granularity is fine enough.
427	 */
428	if (IS_ENABLED(NODE_NOT_IN_PAGE_FLAGS)) {
429		unsigned long pfn_align = node_map_pfn_alignment();
430
431		if (pfn_align && pfn_align < PAGES_PER_SECTION) {
432			unsigned long node_align_mb = PFN_PHYS(pfn_align) / SZ_1M;
433
434			unsigned long sect_align_mb = PFN_PHYS(PAGES_PER_SECTION) / SZ_1M;
435
436			pr_warn("Node alignment %luMB < min %luMB, rejecting NUMA config\n",
437				node_align_mb, sect_align_mb);
438			return -EINVAL;
439		}
440	}
441
442	return 0;
443}
444
445int __init numa_memblks_init(int (*init_func)(void),
446			     bool memblock_force_top_down)
447{
448	phys_addr_t max_addr = (phys_addr_t)ULLONG_MAX;
449	int ret;
450
451	nodes_clear(numa_nodes_parsed);
452	nodes_clear(node_possible_map);
453	nodes_clear(node_online_map);
454	memset(&numa_meminfo, 0, sizeof(numa_meminfo));
455	WARN_ON(memblock_set_node(0, max_addr, &memblock.memory, NUMA_NO_NODE));
456	WARN_ON(memblock_set_node(0, max_addr, &memblock.reserved,
457				  NUMA_NO_NODE));
458	/* In case that parsing SRAT failed. */
459	WARN_ON(memblock_clear_hotplug(0, max_addr));
460	numa_reset_distance();
461
462	ret = init_func();
463	if (ret < 0)
464		return ret;
465
466	/*
467	 * We reset memblock back to the top-down direction
468	 * here because if we configured ACPI_NUMA, we have
469	 * parsed SRAT in init_func(). It is ok to have the
470	 * reset here even if we did't configure ACPI_NUMA
471	 * or acpi numa init fails and fallbacks to dummy
472	 * numa init.
473	 */
474	if (memblock_force_top_down)
475		memblock_set_bottom_up(false);
476
477	ret = numa_cleanup_meminfo(&numa_meminfo);
478	if (ret < 0)
479		return ret;
480
481	numa_emulation(&numa_meminfo, numa_distance_cnt);
482
483	return numa_register_meminfo(&numa_meminfo);
484}
485
486static int __init cmp_memblk(const void *a, const void *b)
487{
488	const struct numa_memblk *ma = *(const struct numa_memblk **)a;
489	const struct numa_memblk *mb = *(const struct numa_memblk **)b;
490
491	return (ma->start > mb->start) - (ma->start < mb->start);
492}
493
494static struct numa_memblk *numa_memblk_list[NR_NODE_MEMBLKS] __initdata;
495
496/**
497 * numa_fill_memblks - Fill gaps in numa_meminfo memblks
498 * @start: address to begin fill
499 * @end: address to end fill
500 *
501 * Find and extend numa_meminfo memblks to cover the physical
502 * address range @start-@end
503 *
504 * RETURNS:
505 * 0		  : Success
506 * NUMA_NO_MEMBLK : No memblks exist in address range @start-@end
507 */
508
509int __init numa_fill_memblks(u64 start, u64 end)
510{
511	struct numa_memblk **blk = &numa_memblk_list[0];
512	struct numa_meminfo *mi = &numa_meminfo;
513	int count = 0;
514	u64 prev_end;
515
516	/*
517	 * Create a list of pointers to numa_meminfo memblks that
518	 * overlap start, end. The list is used to make in-place
519	 * changes that fill out the numa_meminfo memblks.
520	 */
521	for (int i = 0; i < mi->nr_blks; i++) {
522		struct numa_memblk *bi = &mi->blk[i];
523
524		if (memblock_addrs_overlap(start, end - start, bi->start,
525					   bi->end - bi->start)) {
526			blk[count] = &mi->blk[i];
527			count++;
528		}
529	}
530	if (!count)
531		return NUMA_NO_MEMBLK;
532
533	/* Sort the list of pointers in memblk->start order */
534	sort(&blk[0], count, sizeof(blk[0]), cmp_memblk, NULL);
535
536	/* Make sure the first/last memblks include start/end */
537	blk[0]->start = min(blk[0]->start, start);
538	blk[count - 1]->end = max(blk[count - 1]->end, end);
539
540	/*
541	 * Fill any gaps by tracking the previous memblks
542	 * end address and backfilling to it if needed.
543	 */
544	prev_end = blk[0]->end;
545	for (int i = 1; i < count; i++) {
546		struct numa_memblk *curr = blk[i];
547
548		if (prev_end >= curr->start) {
549			if (prev_end < curr->end)
550				prev_end = curr->end;
551		} else {
552			curr->start = prev_end;
553			prev_end = curr->end;
554		}
555	}
556	return 0;
557}
558
559#ifdef CONFIG_NUMA_KEEP_MEMINFO
560static int meminfo_to_nid(struct numa_meminfo *mi, u64 start)
561{
562	int i;
563
564	for (i = 0; i < mi->nr_blks; i++)
565		if (mi->blk[i].start <= start && mi->blk[i].end > start)
566			return mi->blk[i].nid;
567	return NUMA_NO_NODE;
568}
569
570int phys_to_target_node(u64 start)
571{
572	int nid = meminfo_to_nid(&numa_meminfo, start);
573	int reserved_nid = meminfo_to_nid(&numa_reserved_meminfo, start);
574
575	/*
576	 * Prefer online nodes unless the address is also described
577	 * by reserved ranges, in which case use the reserved nid.
578	 */
579	if (nid != NUMA_NO_NODE && reserved_nid == NUMA_NO_NODE)
580		return nid;
581
582	return reserved_nid;
583}
584EXPORT_SYMBOL_GPL(phys_to_target_node);
585
586int memory_add_physaddr_to_nid(u64 start)
587{
588	int nid = meminfo_to_nid(&numa_meminfo, start);
589
590	if (nid == NUMA_NO_NODE)
591		nid = numa_meminfo.blk[0].nid;
592	return nid;
593}
594EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
595
596#endif /* CONFIG_NUMA_KEEP_MEMINFO */