xref: /openbmc/linux/arch/x86/mm/numa.c (revision e5bd61e8)
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 
22 int numa_off;
23 nodemask_t numa_nodes_parsed __initdata;
24 
25 struct pglist_data *node_data[MAX_NUMNODES] __read_mostly;
26 EXPORT_SYMBOL(node_data);
27 
28 static struct numa_meminfo numa_meminfo __initdata_or_meminfo;
29 static struct numa_meminfo numa_reserved_meminfo __initdata_or_meminfo;
30 
31 static int numa_distance_cnt;
32 static u8 *numa_distance;
33 
34 static __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 }
48 early_param("numa", numa_setup);
49 
50 /*
51  * apicid, cpu, node mappings
52  */
53 s16 __apicid_to_node[MAX_LOCAL_APIC] = {
54 	[0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE
55 };
56 
57 int 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 
66 cpumask_var_t node_to_cpumask_map[MAX_NUMNODES];
67 EXPORT_SYMBOL(node_to_cpumask_map);
68 
69 /*
70  * Map cpu index to node index
71  */
72 DEFINE_EARLY_PER_CPU(int, x86_cpu_to_node_map, NUMA_NO_NODE);
73 EXPORT_EARLY_PER_CPU_SYMBOL(x86_cpu_to_node_map);
74 
75 void 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 
97 void 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  */
109 void __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 
125 static 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  */
159 void __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  */
172 static 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  */
190 int __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 */
196 static 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  */
238 int __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 		bi->end = min(bi->end, high);
258 
259 		/* and there's no empty block */
260 		if (bi->start >= bi->end)
261 			numa_remove_memblk_from(i--, mi);
262 	}
263 
264 	/* merge neighboring / overlapping entries */
265 	for (i = 0; i < mi->nr_blks; i++) {
266 		struct numa_memblk *bi = &mi->blk[i];
267 
268 		for (j = i + 1; j < mi->nr_blks; j++) {
269 			struct numa_memblk *bj = &mi->blk[j];
270 			u64 start, end;
271 
272 			/*
273 			 * See whether there are overlapping blocks.  Whine
274 			 * about but allow overlaps of the same nid.  They
275 			 * will be merged below.
276 			 */
277 			if (bi->end > bj->start && bi->start < bj->end) {
278 				if (bi->nid != bj->nid) {
279 					pr_err("node %d [mem %#010Lx-%#010Lx] overlaps with node %d [mem %#010Lx-%#010Lx]\n",
280 					       bi->nid, bi->start, bi->end - 1,
281 					       bj->nid, bj->start, bj->end - 1);
282 					return -EINVAL;
283 				}
284 				pr_warn("Warning: node %d [mem %#010Lx-%#010Lx] overlaps with itself [mem %#010Lx-%#010Lx]\n",
285 					bi->nid, bi->start, bi->end - 1,
286 					bj->start, bj->end - 1);
287 			}
288 
289 			/*
290 			 * Join together blocks on the same node, holes
291 			 * between which don't overlap with memory on other
292 			 * nodes.
293 			 */
294 			if (bi->nid != bj->nid)
295 				continue;
296 			start = min(bi->start, bj->start);
297 			end = max(bi->end, bj->end);
298 			for (k = 0; k < mi->nr_blks; k++) {
299 				struct numa_memblk *bk = &mi->blk[k];
300 
301 				if (bi->nid == bk->nid)
302 					continue;
303 				if (start < bk->end && end > bk->start)
304 					break;
305 			}
306 			if (k < mi->nr_blks)
307 				continue;
308 			printk(KERN_INFO "NUMA: Node %d [mem %#010Lx-%#010Lx] + [mem %#010Lx-%#010Lx] -> [mem %#010Lx-%#010Lx]\n",
309 			       bi->nid, bi->start, bi->end - 1, bj->start,
310 			       bj->end - 1, start, end - 1);
311 			bi->start = start;
312 			bi->end = end;
313 			numa_remove_memblk_from(j--, mi);
314 		}
315 	}
316 
317 	/* clear unused ones */
318 	for (i = mi->nr_blks; i < ARRAY_SIZE(mi->blk); i++) {
319 		mi->blk[i].start = mi->blk[i].end = 0;
320 		mi->blk[i].nid = NUMA_NO_NODE;
321 	}
322 
323 	return 0;
324 }
325 
326 /*
327  * Set nodes, which have memory in @mi, in *@nodemask.
328  */
329 static void __init numa_nodemask_from_meminfo(nodemask_t *nodemask,
330 					      const struct numa_meminfo *mi)
331 {
332 	int i;
333 
334 	for (i = 0; i < ARRAY_SIZE(mi->blk); i++)
335 		if (mi->blk[i].start != mi->blk[i].end &&
336 		    mi->blk[i].nid != NUMA_NO_NODE)
337 			node_set(mi->blk[i].nid, *nodemask);
338 }
339 
340 /**
341  * numa_reset_distance - Reset NUMA distance table
342  *
343  * The current table is freed.  The next numa_set_distance() call will
344  * create a new one.
345  */
346 void __init numa_reset_distance(void)
347 {
348 	size_t size = numa_distance_cnt * numa_distance_cnt * sizeof(numa_distance[0]);
349 
350 	/* numa_distance could be 1LU marking allocation failure, test cnt */
351 	if (numa_distance_cnt)
352 		memblock_free(__pa(numa_distance), size);
353 	numa_distance_cnt = 0;
354 	numa_distance = NULL;	/* enable table creation */
355 }
356 
357 static int __init numa_alloc_distance(void)
358 {
359 	nodemask_t nodes_parsed;
360 	size_t size;
361 	int i, j, cnt = 0;
362 	u64 phys;
363 
364 	/* size the new table and allocate it */
365 	nodes_parsed = numa_nodes_parsed;
366 	numa_nodemask_from_meminfo(&nodes_parsed, &numa_meminfo);
367 
368 	for_each_node_mask(i, nodes_parsed)
369 		cnt = i;
370 	cnt++;
371 	size = cnt * cnt * sizeof(numa_distance[0]);
372 
373 	phys = memblock_find_in_range(0, PFN_PHYS(max_pfn_mapped),
374 				      size, PAGE_SIZE);
375 	if (!phys) {
376 		pr_warn("Warning: can't allocate distance table!\n");
377 		/* don't retry until explicitly reset */
378 		numa_distance = (void *)1LU;
379 		return -ENOMEM;
380 	}
381 	memblock_reserve(phys, size);
382 
383 	numa_distance = __va(phys);
384 	numa_distance_cnt = cnt;
385 
386 	/* fill with the default distances */
387 	for (i = 0; i < cnt; i++)
388 		for (j = 0; j < cnt; j++)
389 			numa_distance[i * cnt + j] = i == j ?
390 				LOCAL_DISTANCE : REMOTE_DISTANCE;
391 	printk(KERN_DEBUG "NUMA: Initialized distance table, cnt=%d\n", cnt);
392 
393 	return 0;
394 }
395 
396 /**
397  * numa_set_distance - Set NUMA distance from one NUMA to another
398  * @from: the 'from' node to set distance
399  * @to: the 'to'  node to set distance
400  * @distance: NUMA distance
401  *
402  * Set the distance from node @from to @to to @distance.  If distance table
403  * doesn't exist, one which is large enough to accommodate all the currently
404  * known nodes will be created.
405  *
406  * If such table cannot be allocated, a warning is printed and further
407  * calls are ignored until the distance table is reset with
408  * numa_reset_distance().
409  *
410  * If @from or @to is higher than the highest known node or lower than zero
411  * at the time of table creation or @distance doesn't make sense, the call
412  * is ignored.
413  * This is to allow simplification of specific NUMA config implementations.
414  */
415 void __init numa_set_distance(int from, int to, int distance)
416 {
417 	if (!numa_distance && numa_alloc_distance() < 0)
418 		return;
419 
420 	if (from >= numa_distance_cnt || to >= numa_distance_cnt ||
421 			from < 0 || to < 0) {
422 		pr_warn_once("Warning: node ids are out of bound, from=%d to=%d distance=%d\n",
423 			     from, to, distance);
424 		return;
425 	}
426 
427 	if ((u8)distance != distance ||
428 	    (from == to && distance != LOCAL_DISTANCE)) {
429 		pr_warn_once("Warning: invalid distance parameter, from=%d to=%d distance=%d\n",
430 			     from, to, distance);
431 		return;
432 	}
433 
434 	numa_distance[from * numa_distance_cnt + to] = distance;
435 }
436 
437 int __node_distance(int from, int to)
438 {
439 	if (from >= numa_distance_cnt || to >= numa_distance_cnt)
440 		return from == to ? LOCAL_DISTANCE : REMOTE_DISTANCE;
441 	return numa_distance[from * numa_distance_cnt + to];
442 }
443 EXPORT_SYMBOL(__node_distance);
444 
445 /*
446  * Sanity check to catch more bad NUMA configurations (they are amazingly
447  * common).  Make sure the nodes cover all memory.
448  */
449 static bool __init numa_meminfo_cover_memory(const struct numa_meminfo *mi)
450 {
451 	u64 numaram, e820ram;
452 	int i;
453 
454 	numaram = 0;
455 	for (i = 0; i < mi->nr_blks; i++) {
456 		u64 s = mi->blk[i].start >> PAGE_SHIFT;
457 		u64 e = mi->blk[i].end >> PAGE_SHIFT;
458 		numaram += e - s;
459 		numaram -= __absent_pages_in_range(mi->blk[i].nid, s, e);
460 		if ((s64)numaram < 0)
461 			numaram = 0;
462 	}
463 
464 	e820ram = max_pfn - absent_pages_in_range(0, max_pfn);
465 
466 	/* We seem to lose 3 pages somewhere. Allow 1M of slack. */
467 	if ((s64)(e820ram - numaram) >= (1 << (20 - PAGE_SHIFT))) {
468 		printk(KERN_ERR "NUMA: nodes only cover %LuMB of your %LuMB e820 RAM. Not used.\n",
469 		       (numaram << PAGE_SHIFT) >> 20,
470 		       (e820ram << PAGE_SHIFT) >> 20);
471 		return false;
472 	}
473 	return true;
474 }
475 
476 /*
477  * Mark all currently memblock-reserved physical memory (which covers the
478  * kernel's own memory ranges) as hot-unswappable.
479  */
480 static void __init numa_clear_kernel_node_hotplug(void)
481 {
482 	nodemask_t reserved_nodemask = NODE_MASK_NONE;
483 	struct memblock_region *mb_region;
484 	int i;
485 
486 	/*
487 	 * We have to do some preprocessing of memblock regions, to
488 	 * make them suitable for reservation.
489 	 *
490 	 * At this time, all memory regions reserved by memblock are
491 	 * used by the kernel, but those regions are not split up
492 	 * along node boundaries yet, and don't necessarily have their
493 	 * node ID set yet either.
494 	 *
495 	 * So iterate over all memory known to the x86 architecture,
496 	 * and use those ranges to set the nid in memblock.reserved.
497 	 * This will split up the memblock regions along node
498 	 * boundaries and will set the node IDs as well.
499 	 */
500 	for (i = 0; i < numa_meminfo.nr_blks; i++) {
501 		struct numa_memblk *mb = numa_meminfo.blk + i;
502 		int ret;
503 
504 		ret = memblock_set_node(mb->start, mb->end - mb->start, &memblock.reserved, mb->nid);
505 		WARN_ON_ONCE(ret);
506 	}
507 
508 	/*
509 	 * Now go over all reserved memblock regions, to construct a
510 	 * node mask of all kernel reserved memory areas.
511 	 *
512 	 * [ Note, when booting with mem=nn[kMG] or in a kdump kernel,
513 	 *   numa_meminfo might not include all memblock.reserved
514 	 *   memory ranges, because quirks such as trim_snb_memory()
515 	 *   reserve specific pages for Sandy Bridge graphics. ]
516 	 */
517 	for_each_reserved_mem_region(mb_region) {
518 		int nid = memblock_get_region_node(mb_region);
519 
520 		if (nid != MAX_NUMNODES)
521 			node_set(nid, reserved_nodemask);
522 	}
523 
524 	/*
525 	 * Finally, clear the MEMBLOCK_HOTPLUG flag for all memory
526 	 * belonging to the reserved node mask.
527 	 *
528 	 * Note that this will include memory regions that reside
529 	 * on nodes that contain kernel memory - entire nodes
530 	 * become hot-unpluggable:
531 	 */
532 	for (i = 0; i < numa_meminfo.nr_blks; i++) {
533 		struct numa_memblk *mb = numa_meminfo.blk + i;
534 
535 		if (!node_isset(mb->nid, reserved_nodemask))
536 			continue;
537 
538 		memblock_clear_hotplug(mb->start, mb->end - mb->start);
539 	}
540 }
541 
542 static int __init numa_register_memblks(struct numa_meminfo *mi)
543 {
544 	int i, nid;
545 
546 	/* Account for nodes with cpus and no memory */
547 	node_possible_map = numa_nodes_parsed;
548 	numa_nodemask_from_meminfo(&node_possible_map, mi);
549 	if (WARN_ON(nodes_empty(node_possible_map)))
550 		return -EINVAL;
551 
552 	for (i = 0; i < mi->nr_blks; i++) {
553 		struct numa_memblk *mb = &mi->blk[i];
554 		memblock_set_node(mb->start, mb->end - mb->start,
555 				  &memblock.memory, mb->nid);
556 	}
557 
558 	/*
559 	 * At very early time, the kernel have to use some memory such as
560 	 * loading the kernel image. We cannot prevent this anyway. So any
561 	 * node the kernel resides in should be un-hotpluggable.
562 	 *
563 	 * And when we come here, alloc node data won't fail.
564 	 */
565 	numa_clear_kernel_node_hotplug();
566 
567 	/*
568 	 * If sections array is gonna be used for pfn -> nid mapping, check
569 	 * whether its granularity is fine enough.
570 	 */
571 	if (IS_ENABLED(NODE_NOT_IN_PAGE_FLAGS)) {
572 		unsigned long pfn_align = node_map_pfn_alignment();
573 
574 		if (pfn_align && pfn_align < PAGES_PER_SECTION) {
575 			pr_warn("Node alignment %LuMB < min %LuMB, rejecting NUMA config\n",
576 				PFN_PHYS(pfn_align) >> 20,
577 				PFN_PHYS(PAGES_PER_SECTION) >> 20);
578 			return -EINVAL;
579 		}
580 	}
581 	if (!numa_meminfo_cover_memory(mi))
582 		return -EINVAL;
583 
584 	/* Finally register nodes. */
585 	for_each_node_mask(nid, node_possible_map) {
586 		u64 start = PFN_PHYS(max_pfn);
587 		u64 end = 0;
588 
589 		for (i = 0; i < mi->nr_blks; i++) {
590 			if (nid != mi->blk[i].nid)
591 				continue;
592 			start = min(mi->blk[i].start, start);
593 			end = max(mi->blk[i].end, end);
594 		}
595 
596 		if (start >= end)
597 			continue;
598 
599 		/*
600 		 * Don't confuse VM with a node that doesn't have the
601 		 * minimum amount of memory:
602 		 */
603 		if (end && (end - start) < NODE_MIN_SIZE)
604 			continue;
605 
606 		alloc_node_data(nid);
607 	}
608 
609 	/* Dump memblock with node info and return. */
610 	memblock_dump_all();
611 	return 0;
612 }
613 
614 /*
615  * There are unfortunately some poorly designed mainboards around that
616  * only connect memory to a single CPU. This breaks the 1:1 cpu->node
617  * mapping. To avoid this fill in the mapping for all possible CPUs,
618  * as the number of CPUs is not known yet. We round robin the existing
619  * nodes.
620  */
621 static void __init numa_init_array(void)
622 {
623 	int rr, i;
624 
625 	rr = first_node(node_online_map);
626 	for (i = 0; i < nr_cpu_ids; i++) {
627 		if (early_cpu_to_node(i) != NUMA_NO_NODE)
628 			continue;
629 		numa_set_node(i, rr);
630 		rr = next_node_in(rr, node_online_map);
631 	}
632 }
633 
634 static int __init numa_init(int (*init_func)(void))
635 {
636 	int i;
637 	int ret;
638 
639 	for (i = 0; i < MAX_LOCAL_APIC; i++)
640 		set_apicid_to_node(i, NUMA_NO_NODE);
641 
642 	nodes_clear(numa_nodes_parsed);
643 	nodes_clear(node_possible_map);
644 	nodes_clear(node_online_map);
645 	memset(&numa_meminfo, 0, sizeof(numa_meminfo));
646 	WARN_ON(memblock_set_node(0, ULLONG_MAX, &memblock.memory,
647 				  MAX_NUMNODES));
648 	WARN_ON(memblock_set_node(0, ULLONG_MAX, &memblock.reserved,
649 				  MAX_NUMNODES));
650 	/* In case that parsing SRAT failed. */
651 	WARN_ON(memblock_clear_hotplug(0, ULLONG_MAX));
652 	numa_reset_distance();
653 
654 	ret = init_func();
655 	if (ret < 0)
656 		return ret;
657 
658 	/*
659 	 * We reset memblock back to the top-down direction
660 	 * here because if we configured ACPI_NUMA, we have
661 	 * parsed SRAT in init_func(). It is ok to have the
662 	 * reset here even if we did't configure ACPI_NUMA
663 	 * or acpi numa init fails and fallbacks to dummy
664 	 * numa init.
665 	 */
666 	memblock_set_bottom_up(false);
667 
668 	ret = numa_cleanup_meminfo(&numa_meminfo);
669 	if (ret < 0)
670 		return ret;
671 
672 	numa_emulation(&numa_meminfo, numa_distance_cnt);
673 
674 	ret = numa_register_memblks(&numa_meminfo);
675 	if (ret < 0)
676 		return ret;
677 
678 	for (i = 0; i < nr_cpu_ids; i++) {
679 		int nid = early_cpu_to_node(i);
680 
681 		if (nid == NUMA_NO_NODE)
682 			continue;
683 		if (!node_online(nid))
684 			numa_clear_node(i);
685 	}
686 	numa_init_array();
687 
688 	return 0;
689 }
690 
691 /**
692  * dummy_numa_init - Fallback dummy NUMA init
693  *
694  * Used if there's no underlying NUMA architecture, NUMA initialization
695  * fails, or NUMA is disabled on the command line.
696  *
697  * Must online at least one node and add memory blocks that cover all
698  * allowed memory.  This function must not fail.
699  */
700 static int __init dummy_numa_init(void)
701 {
702 	printk(KERN_INFO "%s\n",
703 	       numa_off ? "NUMA turned off" : "No NUMA configuration found");
704 	printk(KERN_INFO "Faking a node at [mem %#018Lx-%#018Lx]\n",
705 	       0LLU, PFN_PHYS(max_pfn) - 1);
706 
707 	node_set(0, numa_nodes_parsed);
708 	numa_add_memblk(0, 0, PFN_PHYS(max_pfn));
709 
710 	return 0;
711 }
712 
713 /**
714  * x86_numa_init - Initialize NUMA
715  *
716  * Try each configured NUMA initialization method until one succeeds.  The
717  * last fallback is dummy single node config encompassing whole memory and
718  * never fails.
719  */
720 void __init x86_numa_init(void)
721 {
722 	if (!numa_off) {
723 #ifdef CONFIG_ACPI_NUMA
724 		if (!numa_init(x86_acpi_numa_init))
725 			return;
726 #endif
727 #ifdef CONFIG_AMD_NUMA
728 		if (!numa_init(amd_numa_init))
729 			return;
730 #endif
731 	}
732 
733 	numa_init(dummy_numa_init);
734 }
735 
736 static void __init init_memory_less_node(int nid)
737 {
738 	/* Allocate and initialize node data. Memory-less node is now online.*/
739 	alloc_node_data(nid);
740 	free_area_init_memoryless_node(nid);
741 
742 	/*
743 	 * All zonelists will be built later in start_kernel() after per cpu
744 	 * areas are initialized.
745 	 */
746 }
747 
748 /*
749  * A node may exist which has one or more Generic Initiators but no CPUs and no
750  * memory.
751  *
752  * This function must be called after init_cpu_to_node(), to ensure that any
753  * memoryless CPU nodes have already been brought online, and before the
754  * node_data[nid] is needed for zone list setup in build_all_zonelists().
755  *
756  * When this function is called, any nodes containing either memory and/or CPUs
757  * will already be online and there is no need to do anything extra, even if
758  * they also contain one or more Generic Initiators.
759  */
760 void __init init_gi_nodes(void)
761 {
762 	int nid;
763 
764 	for_each_node_state(nid, N_GENERIC_INITIATOR)
765 		if (!node_online(nid))
766 			init_memory_less_node(nid);
767 }
768 
769 /*
770  * Setup early cpu_to_node.
771  *
772  * Populate cpu_to_node[] only if x86_cpu_to_apicid[],
773  * and apicid_to_node[] tables have valid entries for a CPU.
774  * This means we skip cpu_to_node[] initialisation for NUMA
775  * emulation and faking node case (when running a kernel compiled
776  * for NUMA on a non NUMA box), which is OK as cpu_to_node[]
777  * is already initialized in a round robin manner at numa_init_array,
778  * prior to this call, and this initialization is good enough
779  * for the fake NUMA cases.
780  *
781  * Called before the per_cpu areas are setup.
782  */
783 void __init init_cpu_to_node(void)
784 {
785 	int cpu;
786 	u16 *cpu_to_apicid = early_per_cpu_ptr(x86_cpu_to_apicid);
787 
788 	BUG_ON(cpu_to_apicid == NULL);
789 
790 	for_each_possible_cpu(cpu) {
791 		int node = numa_cpu_node(cpu);
792 
793 		if (node == NUMA_NO_NODE)
794 			continue;
795 
796 		if (!node_online(node))
797 			init_memory_less_node(node);
798 
799 		numa_set_node(cpu, node);
800 	}
801 }
802 
803 #ifndef CONFIG_DEBUG_PER_CPU_MAPS
804 
805 # ifndef CONFIG_NUMA_EMU
806 void numa_add_cpu(int cpu)
807 {
808 	cpumask_set_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
809 }
810 
811 void numa_remove_cpu(int cpu)
812 {
813 	cpumask_clear_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
814 }
815 # endif	/* !CONFIG_NUMA_EMU */
816 
817 #else	/* !CONFIG_DEBUG_PER_CPU_MAPS */
818 
819 int __cpu_to_node(int cpu)
820 {
821 	if (early_per_cpu_ptr(x86_cpu_to_node_map)) {
822 		printk(KERN_WARNING
823 			"cpu_to_node(%d): usage too early!\n", cpu);
824 		dump_stack();
825 		return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
826 	}
827 	return per_cpu(x86_cpu_to_node_map, cpu);
828 }
829 EXPORT_SYMBOL(__cpu_to_node);
830 
831 /*
832  * Same function as cpu_to_node() but used if called before the
833  * per_cpu areas are setup.
834  */
835 int early_cpu_to_node(int cpu)
836 {
837 	if (early_per_cpu_ptr(x86_cpu_to_node_map))
838 		return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
839 
840 	if (!cpu_possible(cpu)) {
841 		printk(KERN_WARNING
842 			"early_cpu_to_node(%d): no per_cpu area!\n", cpu);
843 		dump_stack();
844 		return NUMA_NO_NODE;
845 	}
846 	return per_cpu(x86_cpu_to_node_map, cpu);
847 }
848 
849 void debug_cpumask_set_cpu(int cpu, int node, bool enable)
850 {
851 	struct cpumask *mask;
852 
853 	if (node == NUMA_NO_NODE) {
854 		/* early_cpu_to_node() already emits a warning and trace */
855 		return;
856 	}
857 	mask = node_to_cpumask_map[node];
858 	if (!mask) {
859 		pr_err("node_to_cpumask_map[%i] NULL\n", node);
860 		dump_stack();
861 		return;
862 	}
863 
864 	if (enable)
865 		cpumask_set_cpu(cpu, mask);
866 	else
867 		cpumask_clear_cpu(cpu, mask);
868 
869 	printk(KERN_DEBUG "%s cpu %d node %d: mask now %*pbl\n",
870 		enable ? "numa_add_cpu" : "numa_remove_cpu",
871 		cpu, node, cpumask_pr_args(mask));
872 	return;
873 }
874 
875 # ifndef CONFIG_NUMA_EMU
876 static void numa_set_cpumask(int cpu, bool enable)
877 {
878 	debug_cpumask_set_cpu(cpu, early_cpu_to_node(cpu), enable);
879 }
880 
881 void numa_add_cpu(int cpu)
882 {
883 	numa_set_cpumask(cpu, true);
884 }
885 
886 void numa_remove_cpu(int cpu)
887 {
888 	numa_set_cpumask(cpu, false);
889 }
890 # endif	/* !CONFIG_NUMA_EMU */
891 
892 /*
893  * Returns a pointer to the bitmask of CPUs on Node 'node'.
894  */
895 const struct cpumask *cpumask_of_node(int node)
896 {
897 	if ((unsigned)node >= nr_node_ids) {
898 		printk(KERN_WARNING
899 			"cpumask_of_node(%d): (unsigned)node >= nr_node_ids(%u)\n",
900 			node, nr_node_ids);
901 		dump_stack();
902 		return cpu_none_mask;
903 	}
904 	if (node_to_cpumask_map[node] == NULL) {
905 		printk(KERN_WARNING
906 			"cpumask_of_node(%d): no node_to_cpumask_map!\n",
907 			node);
908 		dump_stack();
909 		return cpu_online_mask;
910 	}
911 	return node_to_cpumask_map[node];
912 }
913 EXPORT_SYMBOL(cpumask_of_node);
914 
915 #endif	/* !CONFIG_DEBUG_PER_CPU_MAPS */
916 
917 #ifdef CONFIG_NUMA_KEEP_MEMINFO
918 static int meminfo_to_nid(struct numa_meminfo *mi, u64 start)
919 {
920 	int i;
921 
922 	for (i = 0; i < mi->nr_blks; i++)
923 		if (mi->blk[i].start <= start && mi->blk[i].end > start)
924 			return mi->blk[i].nid;
925 	return NUMA_NO_NODE;
926 }
927 
928 int phys_to_target_node(phys_addr_t start)
929 {
930 	int nid = meminfo_to_nid(&numa_meminfo, start);
931 
932 	/*
933 	 * Prefer online nodes, but if reserved memory might be
934 	 * hot-added continue the search with reserved ranges.
935 	 */
936 	if (nid != NUMA_NO_NODE)
937 		return nid;
938 
939 	return meminfo_to_nid(&numa_reserved_meminfo, start);
940 }
941 EXPORT_SYMBOL_GPL(phys_to_target_node);
942 
943 int memory_add_physaddr_to_nid(u64 start)
944 {
945 	int nid = meminfo_to_nid(&numa_meminfo, start);
946 
947 	if (nid == NUMA_NO_NODE)
948 		nid = numa_meminfo.blk[0].nid;
949 	return nid;
950 }
951 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
952 #endif
953