xref: /linux/mm/numa_memblks.c (revision c532de5a67a70f8533d495f8f2aaa9a0491c3ad0)
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 static int numa_distance_cnt;
11 static u8 *numa_distance;
12 
13 nodemask_t numa_nodes_parsed __initdata;
14 
15 static struct numa_meminfo numa_meminfo __initdata_or_meminfo;
16 static struct numa_meminfo numa_reserved_meminfo __initdata_or_meminfo;
17 
18 /*
19  * Set nodes, which have memory in @mi, in *@nodemask.
20  */
21 static 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  */
38 void __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(numa_distance, size);
45 	numa_distance_cnt = 0;
46 	numa_distance = NULL;	/* enable table creation */
47 }
48 
49 static 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(&nodes_parsed, &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  */
103 void __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 
125 int __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 }
131 EXPORT_SYMBOL(__node_distance);
132 
133 static 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  */
167 void __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  */
180 static 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, 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  */
198 int __init numa_add_memblk(int nid, u64 start, u64 end)
199 {
200 	return numa_add_memblk_to(nid, start, end, &numa_meminfo);
201 }
202 
203 /**
204  * numa_cleanup_meminfo - Cleanup a numa_meminfo
205  * @mi: numa_meminfo to clean up
206  *
207  * Sanitize @mi by merging and removing unnecessary memblks.  Also check for
208  * conflicts and clear unused memblks.
209  *
210  * RETURNS:
211  * 0 on success, -errno on failure.
212  */
213 int __init numa_cleanup_meminfo(struct numa_meminfo *mi)
214 {
215 	const u64 low = memblock_start_of_DRAM();
216 	const u64 high = memblock_end_of_DRAM();
217 	int i, j, k;
218 
219 	/* first, trim all entries */
220 	for (i = 0; i < mi->nr_blks; i++) {
221 		struct numa_memblk *bi = &mi->blk[i];
222 
223 		/* move / save reserved memory ranges */
224 		if (!memblock_overlaps_region(&memblock.memory,
225 					bi->start, bi->end - bi->start)) {
226 			numa_move_tail_memblk(&numa_reserved_meminfo, i--, mi);
227 			continue;
228 		}
229 
230 		/* make sure all non-reserved blocks are inside the limits */
231 		bi->start = max(bi->start, low);
232 
233 		/* preserve info for non-RAM areas above 'max_pfn': */
234 		if (bi->end > high) {
235 			numa_add_memblk_to(bi->nid, high, bi->end,
236 					   &numa_reserved_meminfo);
237 			bi->end = high;
238 		}
239 
240 		/* and there's no empty block */
241 		if (bi->start >= bi->end)
242 			numa_remove_memblk_from(i--, mi);
243 	}
244 
245 	/* merge neighboring / overlapping entries */
246 	for (i = 0; i < mi->nr_blks; i++) {
247 		struct numa_memblk *bi = &mi->blk[i];
248 
249 		for (j = i + 1; j < mi->nr_blks; j++) {
250 			struct numa_memblk *bj = &mi->blk[j];
251 			u64 start, end;
252 
253 			/*
254 			 * See whether there are overlapping blocks.  Whine
255 			 * about but allow overlaps of the same nid.  They
256 			 * will be merged below.
257 			 */
258 			if (bi->end > bj->start && bi->start < bj->end) {
259 				if (bi->nid != bj->nid) {
260 					pr_err("node %d [mem %#010Lx-%#010Lx] overlaps with node %d [mem %#010Lx-%#010Lx]\n",
261 					       bi->nid, bi->start, bi->end - 1,
262 					       bj->nid, bj->start, bj->end - 1);
263 					return -EINVAL;
264 				}
265 				pr_warn("Warning: node %d [mem %#010Lx-%#010Lx] overlaps with itself [mem %#010Lx-%#010Lx]\n",
266 					bi->nid, bi->start, bi->end - 1,
267 					bj->start, bj->end - 1);
268 			}
269 
270 			/*
271 			 * Join together blocks on the same node, holes
272 			 * between which don't overlap with memory on other
273 			 * nodes.
274 			 */
275 			if (bi->nid != bj->nid)
276 				continue;
277 			start = min(bi->start, bj->start);
278 			end = max(bi->end, bj->end);
279 			for (k = 0; k < mi->nr_blks; k++) {
280 				struct numa_memblk *bk = &mi->blk[k];
281 
282 				if (bi->nid == bk->nid)
283 					continue;
284 				if (start < bk->end && end > bk->start)
285 					break;
286 			}
287 			if (k < mi->nr_blks)
288 				continue;
289 			pr_info("NUMA: Node %d [mem %#010Lx-%#010Lx] + [mem %#010Lx-%#010Lx] -> [mem %#010Lx-%#010Lx]\n",
290 			       bi->nid, bi->start, bi->end - 1, bj->start,
291 			       bj->end - 1, start, end - 1);
292 			bi->start = start;
293 			bi->end = end;
294 			numa_remove_memblk_from(j--, mi);
295 		}
296 	}
297 
298 	/* clear unused ones */
299 	for (i = mi->nr_blks; i < ARRAY_SIZE(mi->blk); i++) {
300 		mi->blk[i].start = mi->blk[i].end = 0;
301 		mi->blk[i].nid = NUMA_NO_NODE;
302 	}
303 
304 	return 0;
305 }
306 
307 /*
308  * Mark all currently memblock-reserved physical memory (which covers the
309  * kernel's own memory ranges) as hot-unswappable.
310  */
311 static void __init numa_clear_kernel_node_hotplug(void)
312 {
313 	nodemask_t reserved_nodemask = NODE_MASK_NONE;
314 	struct memblock_region *mb_region;
315 	int i;
316 
317 	/*
318 	 * We have to do some preprocessing of memblock regions, to
319 	 * make them suitable for reservation.
320 	 *
321 	 * At this time, all memory regions reserved by memblock are
322 	 * used by the kernel, but those regions are not split up
323 	 * along node boundaries yet, and don't necessarily have their
324 	 * node ID set yet either.
325 	 *
326 	 * So iterate over all parsed memory blocks and use those ranges to
327 	 * set the nid in memblock.reserved.  This will split up the
328 	 * memblock regions along node boundaries and will set the node IDs
329 	 * as well.
330 	 */
331 	for (i = 0; i < numa_meminfo.nr_blks; i++) {
332 		struct numa_memblk *mb = numa_meminfo.blk + i;
333 		int ret;
334 
335 		ret = memblock_set_node(mb->start, mb->end - mb->start,
336 					&memblock.reserved, mb->nid);
337 		WARN_ON_ONCE(ret);
338 	}
339 
340 	/*
341 	 * Now go over all reserved memblock regions, to construct a
342 	 * node mask of all kernel reserved memory areas.
343 	 *
344 	 * [ Note, when booting with mem=nn[kMG] or in a kdump kernel,
345 	 *   numa_meminfo might not include all memblock.reserved
346 	 *   memory ranges, because quirks such as trim_snb_memory()
347 	 *   reserve specific pages for Sandy Bridge graphics. ]
348 	 */
349 	for_each_reserved_mem_region(mb_region) {
350 		int nid = memblock_get_region_node(mb_region);
351 
352 		if (numa_valid_node(nid))
353 			node_set(nid, reserved_nodemask);
354 	}
355 
356 	/*
357 	 * Finally, clear the MEMBLOCK_HOTPLUG flag for all memory
358 	 * belonging to the reserved node mask.
359 	 *
360 	 * Note that this will include memory regions that reside
361 	 * on nodes that contain kernel memory - entire nodes
362 	 * become hot-unpluggable:
363 	 */
364 	for (i = 0; i < numa_meminfo.nr_blks; i++) {
365 		struct numa_memblk *mb = numa_meminfo.blk + i;
366 
367 		if (!node_isset(mb->nid, reserved_nodemask))
368 			continue;
369 
370 		memblock_clear_hotplug(mb->start, mb->end - mb->start);
371 	}
372 }
373 
374 static int __init numa_register_meminfo(struct numa_meminfo *mi)
375 {
376 	int i;
377 
378 	/* Account for nodes with cpus and no memory */
379 	node_possible_map = numa_nodes_parsed;
380 	numa_nodemask_from_meminfo(&node_possible_map, mi);
381 	if (WARN_ON(nodes_empty(node_possible_map)))
382 		return -EINVAL;
383 
384 	for (i = 0; i < mi->nr_blks; i++) {
385 		struct numa_memblk *mb = &mi->blk[i];
386 
387 		memblock_set_node(mb->start, mb->end - mb->start,
388 				  &memblock.memory, mb->nid);
389 	}
390 
391 	/*
392 	 * At very early time, the kernel have to use some memory such as
393 	 * loading the kernel image. We cannot prevent this anyway. So any
394 	 * node the kernel resides in should be un-hotpluggable.
395 	 *
396 	 * And when we come here, alloc node data won't fail.
397 	 */
398 	numa_clear_kernel_node_hotplug();
399 
400 	/*
401 	 * If sections array is gonna be used for pfn -> nid mapping, check
402 	 * whether its granularity is fine enough.
403 	 */
404 	if (IS_ENABLED(NODE_NOT_IN_PAGE_FLAGS)) {
405 		unsigned long pfn_align = node_map_pfn_alignment();
406 
407 		if (pfn_align && pfn_align < PAGES_PER_SECTION) {
408 			unsigned long node_align_mb = PFN_PHYS(pfn_align) >> 20;
409 
410 			unsigned long sect_align_mb = PFN_PHYS(PAGES_PER_SECTION) >> 20;
411 
412 			pr_warn("Node alignment %luMB < min %luMB, rejecting NUMA config\n",
413 				node_align_mb, sect_align_mb);
414 			return -EINVAL;
415 		}
416 	}
417 
418 	return 0;
419 }
420 
421 int __init numa_memblks_init(int (*init_func)(void),
422 			     bool memblock_force_top_down)
423 {
424 	phys_addr_t max_addr = (phys_addr_t)ULLONG_MAX;
425 	int ret;
426 
427 	nodes_clear(numa_nodes_parsed);
428 	nodes_clear(node_possible_map);
429 	nodes_clear(node_online_map);
430 	memset(&numa_meminfo, 0, sizeof(numa_meminfo));
431 	WARN_ON(memblock_set_node(0, max_addr, &memblock.memory, NUMA_NO_NODE));
432 	WARN_ON(memblock_set_node(0, max_addr, &memblock.reserved,
433 				  NUMA_NO_NODE));
434 	/* In case that parsing SRAT failed. */
435 	WARN_ON(memblock_clear_hotplug(0, max_addr));
436 	numa_reset_distance();
437 
438 	ret = init_func();
439 	if (ret < 0)
440 		return ret;
441 
442 	/*
443 	 * We reset memblock back to the top-down direction
444 	 * here because if we configured ACPI_NUMA, we have
445 	 * parsed SRAT in init_func(). It is ok to have the
446 	 * reset here even if we did't configure ACPI_NUMA
447 	 * or acpi numa init fails and fallbacks to dummy
448 	 * numa init.
449 	 */
450 	if (memblock_force_top_down)
451 		memblock_set_bottom_up(false);
452 
453 	ret = numa_cleanup_meminfo(&numa_meminfo);
454 	if (ret < 0)
455 		return ret;
456 
457 	numa_emulation(&numa_meminfo, numa_distance_cnt);
458 
459 	return numa_register_meminfo(&numa_meminfo);
460 }
461 
462 static int __init cmp_memblk(const void *a, const void *b)
463 {
464 	const struct numa_memblk *ma = *(const struct numa_memblk **)a;
465 	const struct numa_memblk *mb = *(const struct numa_memblk **)b;
466 
467 	return (ma->start > mb->start) - (ma->start < mb->start);
468 }
469 
470 static struct numa_memblk *numa_memblk_list[NR_NODE_MEMBLKS] __initdata;
471 
472 /**
473  * numa_fill_memblks - Fill gaps in numa_meminfo memblks
474  * @start: address to begin fill
475  * @end: address to end fill
476  *
477  * Find and extend numa_meminfo memblks to cover the physical
478  * address range @start-@end
479  *
480  * RETURNS:
481  * 0		  : Success
482  * NUMA_NO_MEMBLK : No memblks exist in address range @start-@end
483  */
484 
485 int __init numa_fill_memblks(u64 start, u64 end)
486 {
487 	struct numa_memblk **blk = &numa_memblk_list[0];
488 	struct numa_meminfo *mi = &numa_meminfo;
489 	int count = 0;
490 	u64 prev_end;
491 
492 	/*
493 	 * Create a list of pointers to numa_meminfo memblks that
494 	 * overlap start, end. The list is used to make in-place
495 	 * changes that fill out the numa_meminfo memblks.
496 	 */
497 	for (int i = 0; i < mi->nr_blks; i++) {
498 		struct numa_memblk *bi = &mi->blk[i];
499 
500 		if (memblock_addrs_overlap(start, end - start, bi->start,
501 					   bi->end - bi->start)) {
502 			blk[count] = &mi->blk[i];
503 			count++;
504 		}
505 	}
506 	if (!count)
507 		return NUMA_NO_MEMBLK;
508 
509 	/* Sort the list of pointers in memblk->start order */
510 	sort(&blk[0], count, sizeof(blk[0]), cmp_memblk, NULL);
511 
512 	/* Make sure the first/last memblks include start/end */
513 	blk[0]->start = min(blk[0]->start, start);
514 	blk[count - 1]->end = max(blk[count - 1]->end, end);
515 
516 	/*
517 	 * Fill any gaps by tracking the previous memblks
518 	 * end address and backfilling to it if needed.
519 	 */
520 	prev_end = blk[0]->end;
521 	for (int i = 1; i < count; i++) {
522 		struct numa_memblk *curr = blk[i];
523 
524 		if (prev_end >= curr->start) {
525 			if (prev_end < curr->end)
526 				prev_end = curr->end;
527 		} else {
528 			curr->start = prev_end;
529 			prev_end = curr->end;
530 		}
531 	}
532 	return 0;
533 }
534 
535 #ifdef CONFIG_NUMA_KEEP_MEMINFO
536 static int meminfo_to_nid(struct numa_meminfo *mi, u64 start)
537 {
538 	int i;
539 
540 	for (i = 0; i < mi->nr_blks; i++)
541 		if (mi->blk[i].start <= start && mi->blk[i].end > start)
542 			return mi->blk[i].nid;
543 	return NUMA_NO_NODE;
544 }
545 
546 int phys_to_target_node(u64 start)
547 {
548 	int nid = meminfo_to_nid(&numa_meminfo, start);
549 
550 	/*
551 	 * Prefer online nodes, but if reserved memory might be
552 	 * hot-added continue the search with reserved ranges.
553 	 */
554 	if (nid != NUMA_NO_NODE)
555 		return nid;
556 
557 	return meminfo_to_nid(&numa_reserved_meminfo, start);
558 }
559 EXPORT_SYMBOL_GPL(phys_to_target_node);
560 
561 int memory_add_physaddr_to_nid(u64 start)
562 {
563 	int nid = meminfo_to_nid(&numa_meminfo, start);
564 
565 	if (nid == NUMA_NO_NODE)
566 		nid = numa_meminfo.blk[0].nid;
567 	return nid;
568 }
569 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
570 
571 #endif /* CONFIG_NUMA_KEEP_MEMINFO */
572