xref: /openbmc/linux/arch/x86/mm/numa.c (revision 0c874100)
1 /* Common code for 32 and 64-bit NUMA */
2 #include <linux/acpi.h>
3 #include <linux/kernel.h>
4 #include <linux/mm.h>
5 #include <linux/string.h>
6 #include <linux/init.h>
7 #include <linux/memblock.h>
8 #include <linux/mmzone.h>
9 #include <linux/ctype.h>
10 #include <linux/nodemask.h>
11 #include <linux/sched.h>
12 #include <linux/topology.h>
13 
14 #include <asm/e820/api.h>
15 #include <asm/proto.h>
16 #include <asm/dma.h>
17 #include <asm/amd_nb.h>
18 
19 #include "numa_internal.h"
20 
21 int numa_off;
22 nodemask_t numa_nodes_parsed __initdata;
23 
24 struct pglist_data *node_data[MAX_NUMNODES] __read_mostly;
25 EXPORT_SYMBOL(node_data);
26 
27 static struct numa_meminfo numa_meminfo
28 #ifndef CONFIG_MEMORY_HOTPLUG
29 __initdata
30 #endif
31 ;
32 
33 static int numa_distance_cnt;
34 static u8 *numa_distance;
35 
36 static __init int numa_setup(char *opt)
37 {
38 	if (!opt)
39 		return -EINVAL;
40 	if (!strncmp(opt, "off", 3))
41 		numa_off = 1;
42 #ifdef CONFIG_NUMA_EMU
43 	if (!strncmp(opt, "fake=", 5))
44 		numa_emu_cmdline(opt + 5);
45 #endif
46 #ifdef CONFIG_ACPI_NUMA
47 	if (!strncmp(opt, "noacpi", 6))
48 		acpi_numa = -1;
49 #endif
50 	return 0;
51 }
52 early_param("numa", numa_setup);
53 
54 /*
55  * apicid, cpu, node mappings
56  */
57 s16 __apicid_to_node[MAX_LOCAL_APIC] = {
58 	[0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE
59 };
60 
61 int numa_cpu_node(int cpu)
62 {
63 	int apicid = early_per_cpu(x86_cpu_to_apicid, cpu);
64 
65 	if (apicid != BAD_APICID)
66 		return __apicid_to_node[apicid];
67 	return NUMA_NO_NODE;
68 }
69 
70 cpumask_var_t node_to_cpumask_map[MAX_NUMNODES];
71 EXPORT_SYMBOL(node_to_cpumask_map);
72 
73 /*
74  * Map cpu index to node index
75  */
76 DEFINE_EARLY_PER_CPU(int, x86_cpu_to_node_map, NUMA_NO_NODE);
77 EXPORT_EARLY_PER_CPU_SYMBOL(x86_cpu_to_node_map);
78 
79 void numa_set_node(int cpu, int node)
80 {
81 	int *cpu_to_node_map = early_per_cpu_ptr(x86_cpu_to_node_map);
82 
83 	/* early setting, no percpu area yet */
84 	if (cpu_to_node_map) {
85 		cpu_to_node_map[cpu] = node;
86 		return;
87 	}
88 
89 #ifdef CONFIG_DEBUG_PER_CPU_MAPS
90 	if (cpu >= nr_cpu_ids || !cpu_possible(cpu)) {
91 		printk(KERN_ERR "numa_set_node: invalid cpu# (%d)\n", cpu);
92 		dump_stack();
93 		return;
94 	}
95 #endif
96 	per_cpu(x86_cpu_to_node_map, cpu) = node;
97 
98 	set_cpu_numa_node(cpu, node);
99 }
100 
101 void numa_clear_node(int cpu)
102 {
103 	numa_set_node(cpu, NUMA_NO_NODE);
104 }
105 
106 /*
107  * Allocate node_to_cpumask_map based on number of available nodes
108  * Requires node_possible_map to be valid.
109  *
110  * Note: cpumask_of_node() is not valid until after this is done.
111  * (Use CONFIG_DEBUG_PER_CPU_MAPS to check this.)
112  */
113 void __init setup_node_to_cpumask_map(void)
114 {
115 	unsigned int node;
116 
117 	/* setup nr_node_ids if not done yet */
118 	if (nr_node_ids == MAX_NUMNODES)
119 		setup_nr_node_ids();
120 
121 	/* allocate the map */
122 	for (node = 0; node < nr_node_ids; node++)
123 		alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]);
124 
125 	/* cpumask_of_node() will now work */
126 	pr_debug("Node to cpumask map for %d nodes\n", nr_node_ids);
127 }
128 
129 static int __init numa_add_memblk_to(int nid, u64 start, u64 end,
130 				     struct numa_meminfo *mi)
131 {
132 	/* ignore zero length blks */
133 	if (start == end)
134 		return 0;
135 
136 	/* whine about and ignore invalid blks */
137 	if (start > end || nid < 0 || nid >= MAX_NUMNODES) {
138 		pr_warn("Warning: invalid memblk node %d [mem %#010Lx-%#010Lx]\n",
139 			nid, start, end - 1);
140 		return 0;
141 	}
142 
143 	if (mi->nr_blks >= NR_NODE_MEMBLKS) {
144 		pr_err("too many memblk ranges\n");
145 		return -EINVAL;
146 	}
147 
148 	mi->blk[mi->nr_blks].start = start;
149 	mi->blk[mi->nr_blks].end = end;
150 	mi->blk[mi->nr_blks].nid = nid;
151 	mi->nr_blks++;
152 	return 0;
153 }
154 
155 /**
156  * numa_remove_memblk_from - Remove one numa_memblk from a numa_meminfo
157  * @idx: Index of memblk to remove
158  * @mi: numa_meminfo to remove memblk from
159  *
160  * Remove @idx'th numa_memblk from @mi by shifting @mi->blk[] and
161  * decrementing @mi->nr_blks.
162  */
163 void __init numa_remove_memblk_from(int idx, struct numa_meminfo *mi)
164 {
165 	mi->nr_blks--;
166 	memmove(&mi->blk[idx], &mi->blk[idx + 1],
167 		(mi->nr_blks - idx) * sizeof(mi->blk[0]));
168 }
169 
170 /**
171  * numa_add_memblk - Add one numa_memblk to numa_meminfo
172  * @nid: NUMA node ID of the new memblk
173  * @start: Start address of the new memblk
174  * @end: End address of the new memblk
175  *
176  * Add a new memblk to the default numa_meminfo.
177  *
178  * RETURNS:
179  * 0 on success, -errno on failure.
180  */
181 int __init numa_add_memblk(int nid, u64 start, u64 end)
182 {
183 	return numa_add_memblk_to(nid, start, end, &numa_meminfo);
184 }
185 
186 /* Allocate NODE_DATA for a node on the local memory */
187 static void __init alloc_node_data(int nid)
188 {
189 	const size_t nd_size = roundup(sizeof(pg_data_t), PAGE_SIZE);
190 	u64 nd_pa;
191 	void *nd;
192 	int tnid;
193 
194 	/*
195 	 * Allocate node data.  Try node-local memory and then any node.
196 	 * Never allocate in DMA zone.
197 	 */
198 	nd_pa = memblock_phys_alloc_nid(nd_size, SMP_CACHE_BYTES, nid);
199 	if (!nd_pa) {
200 		nd_pa = __memblock_alloc_base(nd_size, SMP_CACHE_BYTES,
201 					      MEMBLOCK_ALLOC_ACCESSIBLE);
202 		if (!nd_pa) {
203 			pr_err("Cannot find %zu bytes in any node (initial node: %d)\n",
204 			       nd_size, nid);
205 			return;
206 		}
207 	}
208 	nd = __va(nd_pa);
209 
210 	/* report and initialize */
211 	printk(KERN_INFO "NODE_DATA(%d) allocated [mem %#010Lx-%#010Lx]\n", nid,
212 	       nd_pa, nd_pa + nd_size - 1);
213 	tnid = early_pfn_to_nid(nd_pa >> PAGE_SHIFT);
214 	if (tnid != nid)
215 		printk(KERN_INFO "    NODE_DATA(%d) on node %d\n", nid, tnid);
216 
217 	node_data[nid] = nd;
218 	memset(NODE_DATA(nid), 0, sizeof(pg_data_t));
219 
220 	node_set_online(nid);
221 }
222 
223 /**
224  * numa_cleanup_meminfo - Cleanup a numa_meminfo
225  * @mi: numa_meminfo to clean up
226  *
227  * Sanitize @mi by merging and removing unnecessary memblks.  Also check for
228  * conflicts and clear unused memblks.
229  *
230  * RETURNS:
231  * 0 on success, -errno on failure.
232  */
233 int __init numa_cleanup_meminfo(struct numa_meminfo *mi)
234 {
235 	const u64 low = 0;
236 	const u64 high = PFN_PHYS(max_pfn);
237 	int i, j, k;
238 
239 	/* first, trim all entries */
240 	for (i = 0; i < mi->nr_blks; i++) {
241 		struct numa_memblk *bi = &mi->blk[i];
242 
243 		/* make sure all blocks are inside the limits */
244 		bi->start = max(bi->start, low);
245 		bi->end = min(bi->end, high);
246 
247 		/* and there's no empty or non-exist block */
248 		if (bi->start >= bi->end ||
249 		    !memblock_overlaps_region(&memblock.memory,
250 			bi->start, bi->end - bi->start))
251 			numa_remove_memblk_from(i--, mi);
252 	}
253 
254 	/* merge neighboring / overlapping entries */
255 	for (i = 0; i < mi->nr_blks; i++) {
256 		struct numa_memblk *bi = &mi->blk[i];
257 
258 		for (j = i + 1; j < mi->nr_blks; j++) {
259 			struct numa_memblk *bj = &mi->blk[j];
260 			u64 start, end;
261 
262 			/*
263 			 * See whether there are overlapping blocks.  Whine
264 			 * about but allow overlaps of the same nid.  They
265 			 * will be merged below.
266 			 */
267 			if (bi->end > bj->start && bi->start < bj->end) {
268 				if (bi->nid != bj->nid) {
269 					pr_err("node %d [mem %#010Lx-%#010Lx] overlaps with node %d [mem %#010Lx-%#010Lx]\n",
270 					       bi->nid, bi->start, bi->end - 1,
271 					       bj->nid, bj->start, bj->end - 1);
272 					return -EINVAL;
273 				}
274 				pr_warn("Warning: node %d [mem %#010Lx-%#010Lx] overlaps with itself [mem %#010Lx-%#010Lx]\n",
275 					bi->nid, bi->start, bi->end - 1,
276 					bj->start, bj->end - 1);
277 			}
278 
279 			/*
280 			 * Join together blocks on the same node, holes
281 			 * between which don't overlap with memory on other
282 			 * nodes.
283 			 */
284 			if (bi->nid != bj->nid)
285 				continue;
286 			start = min(bi->start, bj->start);
287 			end = max(bi->end, bj->end);
288 			for (k = 0; k < mi->nr_blks; k++) {
289 				struct numa_memblk *bk = &mi->blk[k];
290 
291 				if (bi->nid == bk->nid)
292 					continue;
293 				if (start < bk->end && end > bk->start)
294 					break;
295 			}
296 			if (k < mi->nr_blks)
297 				continue;
298 			printk(KERN_INFO "NUMA: Node %d [mem %#010Lx-%#010Lx] + [mem %#010Lx-%#010Lx] -> [mem %#010Lx-%#010Lx]\n",
299 			       bi->nid, bi->start, bi->end - 1, bj->start,
300 			       bj->end - 1, start, end - 1);
301 			bi->start = start;
302 			bi->end = end;
303 			numa_remove_memblk_from(j--, mi);
304 		}
305 	}
306 
307 	/* clear unused ones */
308 	for (i = mi->nr_blks; i < ARRAY_SIZE(mi->blk); i++) {
309 		mi->blk[i].start = mi->blk[i].end = 0;
310 		mi->blk[i].nid = NUMA_NO_NODE;
311 	}
312 
313 	return 0;
314 }
315 
316 /*
317  * Set nodes, which have memory in @mi, in *@nodemask.
318  */
319 static void __init numa_nodemask_from_meminfo(nodemask_t *nodemask,
320 					      const struct numa_meminfo *mi)
321 {
322 	int i;
323 
324 	for (i = 0; i < ARRAY_SIZE(mi->blk); i++)
325 		if (mi->blk[i].start != mi->blk[i].end &&
326 		    mi->blk[i].nid != NUMA_NO_NODE)
327 			node_set(mi->blk[i].nid, *nodemask);
328 }
329 
330 /**
331  * numa_reset_distance - Reset NUMA distance table
332  *
333  * The current table is freed.  The next numa_set_distance() call will
334  * create a new one.
335  */
336 void __init numa_reset_distance(void)
337 {
338 	size_t size = numa_distance_cnt * numa_distance_cnt * sizeof(numa_distance[0]);
339 
340 	/* numa_distance could be 1LU marking allocation failure, test cnt */
341 	if (numa_distance_cnt)
342 		memblock_free(__pa(numa_distance), size);
343 	numa_distance_cnt = 0;
344 	numa_distance = NULL;	/* enable table creation */
345 }
346 
347 static int __init numa_alloc_distance(void)
348 {
349 	nodemask_t nodes_parsed;
350 	size_t size;
351 	int i, j, cnt = 0;
352 	u64 phys;
353 
354 	/* size the new table and allocate it */
355 	nodes_parsed = numa_nodes_parsed;
356 	numa_nodemask_from_meminfo(&nodes_parsed, &numa_meminfo);
357 
358 	for_each_node_mask(i, nodes_parsed)
359 		cnt = i;
360 	cnt++;
361 	size = cnt * cnt * sizeof(numa_distance[0]);
362 
363 	phys = memblock_find_in_range(0, PFN_PHYS(max_pfn_mapped),
364 				      size, PAGE_SIZE);
365 	if (!phys) {
366 		pr_warn("Warning: can't allocate distance table!\n");
367 		/* don't retry until explicitly reset */
368 		numa_distance = (void *)1LU;
369 		return -ENOMEM;
370 	}
371 	memblock_reserve(phys, size);
372 
373 	numa_distance = __va(phys);
374 	numa_distance_cnt = cnt;
375 
376 	/* fill with the default distances */
377 	for (i = 0; i < cnt; i++)
378 		for (j = 0; j < cnt; j++)
379 			numa_distance[i * cnt + j] = i == j ?
380 				LOCAL_DISTANCE : REMOTE_DISTANCE;
381 	printk(KERN_DEBUG "NUMA: Initialized distance table, cnt=%d\n", cnt);
382 
383 	return 0;
384 }
385 
386 /**
387  * numa_set_distance - Set NUMA distance from one NUMA to another
388  * @from: the 'from' node to set distance
389  * @to: the 'to'  node to set distance
390  * @distance: NUMA distance
391  *
392  * Set the distance from node @from to @to to @distance.  If distance table
393  * doesn't exist, one which is large enough to accommodate all the currently
394  * known nodes will be created.
395  *
396  * If such table cannot be allocated, a warning is printed and further
397  * calls are ignored until the distance table is reset with
398  * numa_reset_distance().
399  *
400  * If @from or @to is higher than the highest known node or lower than zero
401  * at the time of table creation or @distance doesn't make sense, the call
402  * is ignored.
403  * This is to allow simplification of specific NUMA config implementations.
404  */
405 void __init numa_set_distance(int from, int to, int distance)
406 {
407 	if (!numa_distance && numa_alloc_distance() < 0)
408 		return;
409 
410 	if (from >= numa_distance_cnt || to >= numa_distance_cnt ||
411 			from < 0 || to < 0) {
412 		pr_warn_once("Warning: node ids are out of bound, from=%d to=%d distance=%d\n",
413 			     from, to, distance);
414 		return;
415 	}
416 
417 	if ((u8)distance != distance ||
418 	    (from == to && distance != LOCAL_DISTANCE)) {
419 		pr_warn_once("Warning: invalid distance parameter, from=%d to=%d distance=%d\n",
420 			     from, to, distance);
421 		return;
422 	}
423 
424 	numa_distance[from * numa_distance_cnt + to] = distance;
425 }
426 
427 int __node_distance(int from, int to)
428 {
429 	if (from >= numa_distance_cnt || to >= numa_distance_cnt)
430 		return from == to ? LOCAL_DISTANCE : REMOTE_DISTANCE;
431 	return numa_distance[from * numa_distance_cnt + to];
432 }
433 EXPORT_SYMBOL(__node_distance);
434 
435 /*
436  * Sanity check to catch more bad NUMA configurations (they are amazingly
437  * common).  Make sure the nodes cover all memory.
438  */
439 static bool __init numa_meminfo_cover_memory(const struct numa_meminfo *mi)
440 {
441 	u64 numaram, e820ram;
442 	int i;
443 
444 	numaram = 0;
445 	for (i = 0; i < mi->nr_blks; i++) {
446 		u64 s = mi->blk[i].start >> PAGE_SHIFT;
447 		u64 e = mi->blk[i].end >> PAGE_SHIFT;
448 		numaram += e - s;
449 		numaram -= __absent_pages_in_range(mi->blk[i].nid, s, e);
450 		if ((s64)numaram < 0)
451 			numaram = 0;
452 	}
453 
454 	e820ram = max_pfn - absent_pages_in_range(0, max_pfn);
455 
456 	/* We seem to lose 3 pages somewhere. Allow 1M of slack. */
457 	if ((s64)(e820ram - numaram) >= (1 << (20 - PAGE_SHIFT))) {
458 		printk(KERN_ERR "NUMA: nodes only cover %LuMB of your %LuMB e820 RAM. Not used.\n",
459 		       (numaram << PAGE_SHIFT) >> 20,
460 		       (e820ram << PAGE_SHIFT) >> 20);
461 		return false;
462 	}
463 	return true;
464 }
465 
466 /*
467  * Mark all currently memblock-reserved physical memory (which covers the
468  * kernel's own memory ranges) as hot-unswappable.
469  */
470 static void __init numa_clear_kernel_node_hotplug(void)
471 {
472 	nodemask_t reserved_nodemask = NODE_MASK_NONE;
473 	struct memblock_region *mb_region;
474 	int i;
475 
476 	/*
477 	 * We have to do some preprocessing of memblock regions, to
478 	 * make them suitable for reservation.
479 	 *
480 	 * At this time, all memory regions reserved by memblock are
481 	 * used by the kernel, but those regions are not split up
482 	 * along node boundaries yet, and don't necessarily have their
483 	 * node ID set yet either.
484 	 *
485 	 * So iterate over all memory known to the x86 architecture,
486 	 * and use those ranges to set the nid in memblock.reserved.
487 	 * This will split up the memblock regions along node
488 	 * boundaries and will set the node IDs as well.
489 	 */
490 	for (i = 0; i < numa_meminfo.nr_blks; i++) {
491 		struct numa_memblk *mb = numa_meminfo.blk + i;
492 		int ret;
493 
494 		ret = memblock_set_node(mb->start, mb->end - mb->start, &memblock.reserved, mb->nid);
495 		WARN_ON_ONCE(ret);
496 	}
497 
498 	/*
499 	 * Now go over all reserved memblock regions, to construct a
500 	 * node mask of all kernel reserved memory areas.
501 	 *
502 	 * [ Note, when booting with mem=nn[kMG] or in a kdump kernel,
503 	 *   numa_meminfo might not include all memblock.reserved
504 	 *   memory ranges, because quirks such as trim_snb_memory()
505 	 *   reserve specific pages for Sandy Bridge graphics. ]
506 	 */
507 	for_each_memblock(reserved, mb_region) {
508 		if (mb_region->nid != MAX_NUMNODES)
509 			node_set(mb_region->nid, reserved_nodemask);
510 	}
511 
512 	/*
513 	 * Finally, clear the MEMBLOCK_HOTPLUG flag for all memory
514 	 * belonging to the reserved node mask.
515 	 *
516 	 * Note that this will include memory regions that reside
517 	 * on nodes that contain kernel memory - entire nodes
518 	 * become hot-unpluggable:
519 	 */
520 	for (i = 0; i < numa_meminfo.nr_blks; i++) {
521 		struct numa_memblk *mb = numa_meminfo.blk + i;
522 
523 		if (!node_isset(mb->nid, reserved_nodemask))
524 			continue;
525 
526 		memblock_clear_hotplug(mb->start, mb->end - mb->start);
527 	}
528 }
529 
530 static int __init numa_register_memblks(struct numa_meminfo *mi)
531 {
532 	unsigned long uninitialized_var(pfn_align);
533 	int i, nid;
534 
535 	/* Account for nodes with cpus and no memory */
536 	node_possible_map = numa_nodes_parsed;
537 	numa_nodemask_from_meminfo(&node_possible_map, mi);
538 	if (WARN_ON(nodes_empty(node_possible_map)))
539 		return -EINVAL;
540 
541 	for (i = 0; i < mi->nr_blks; i++) {
542 		struct numa_memblk *mb = &mi->blk[i];
543 		memblock_set_node(mb->start, mb->end - mb->start,
544 				  &memblock.memory, mb->nid);
545 	}
546 
547 	/*
548 	 * At very early time, the kernel have to use some memory such as
549 	 * loading the kernel image. We cannot prevent this anyway. So any
550 	 * node the kernel resides in should be un-hotpluggable.
551 	 *
552 	 * And when we come here, alloc node data won't fail.
553 	 */
554 	numa_clear_kernel_node_hotplug();
555 
556 	/*
557 	 * If sections array is gonna be used for pfn -> nid mapping, check
558 	 * whether its granularity is fine enough.
559 	 */
560 #ifdef NODE_NOT_IN_PAGE_FLAGS
561 	pfn_align = node_map_pfn_alignment();
562 	if (pfn_align && pfn_align < PAGES_PER_SECTION) {
563 		printk(KERN_WARNING "Node alignment %LuMB < min %LuMB, rejecting NUMA config\n",
564 		       PFN_PHYS(pfn_align) >> 20,
565 		       PFN_PHYS(PAGES_PER_SECTION) >> 20);
566 		return -EINVAL;
567 	}
568 #endif
569 	if (!numa_meminfo_cover_memory(mi))
570 		return -EINVAL;
571 
572 	/* Finally register nodes. */
573 	for_each_node_mask(nid, node_possible_map) {
574 		u64 start = PFN_PHYS(max_pfn);
575 		u64 end = 0;
576 
577 		for (i = 0; i < mi->nr_blks; i++) {
578 			if (nid != mi->blk[i].nid)
579 				continue;
580 			start = min(mi->blk[i].start, start);
581 			end = max(mi->blk[i].end, end);
582 		}
583 
584 		if (start >= end)
585 			continue;
586 
587 		/*
588 		 * Don't confuse VM with a node that doesn't have the
589 		 * minimum amount of memory:
590 		 */
591 		if (end && (end - start) < NODE_MIN_SIZE)
592 			continue;
593 
594 		alloc_node_data(nid);
595 	}
596 
597 	/* Dump memblock with node info and return. */
598 	memblock_dump_all();
599 	return 0;
600 }
601 
602 /*
603  * There are unfortunately some poorly designed mainboards around that
604  * only connect memory to a single CPU. This breaks the 1:1 cpu->node
605  * mapping. To avoid this fill in the mapping for all possible CPUs,
606  * as the number of CPUs is not known yet. We round robin the existing
607  * nodes.
608  */
609 static void __init numa_init_array(void)
610 {
611 	int rr, i;
612 
613 	rr = first_node(node_online_map);
614 	for (i = 0; i < nr_cpu_ids; i++) {
615 		if (early_cpu_to_node(i) != NUMA_NO_NODE)
616 			continue;
617 		numa_set_node(i, rr);
618 		rr = next_node_in(rr, node_online_map);
619 	}
620 }
621 
622 static int __init numa_init(int (*init_func)(void))
623 {
624 	int i;
625 	int ret;
626 
627 	for (i = 0; i < MAX_LOCAL_APIC; i++)
628 		set_apicid_to_node(i, NUMA_NO_NODE);
629 
630 	nodes_clear(numa_nodes_parsed);
631 	nodes_clear(node_possible_map);
632 	nodes_clear(node_online_map);
633 	memset(&numa_meminfo, 0, sizeof(numa_meminfo));
634 	WARN_ON(memblock_set_node(0, ULLONG_MAX, &memblock.memory,
635 				  MAX_NUMNODES));
636 	WARN_ON(memblock_set_node(0, ULLONG_MAX, &memblock.reserved,
637 				  MAX_NUMNODES));
638 	/* In case that parsing SRAT failed. */
639 	WARN_ON(memblock_clear_hotplug(0, ULLONG_MAX));
640 	numa_reset_distance();
641 
642 	ret = init_func();
643 	if (ret < 0)
644 		return ret;
645 
646 	/*
647 	 * We reset memblock back to the top-down direction
648 	 * here because if we configured ACPI_NUMA, we have
649 	 * parsed SRAT in init_func(). It is ok to have the
650 	 * reset here even if we did't configure ACPI_NUMA
651 	 * or acpi numa init fails and fallbacks to dummy
652 	 * numa init.
653 	 */
654 	memblock_set_bottom_up(false);
655 
656 	ret = numa_cleanup_meminfo(&numa_meminfo);
657 	if (ret < 0)
658 		return ret;
659 
660 	numa_emulation(&numa_meminfo, numa_distance_cnt);
661 
662 	ret = numa_register_memblks(&numa_meminfo);
663 	if (ret < 0)
664 		return ret;
665 
666 	for (i = 0; i < nr_cpu_ids; i++) {
667 		int nid = early_cpu_to_node(i);
668 
669 		if (nid == NUMA_NO_NODE)
670 			continue;
671 		if (!node_online(nid))
672 			numa_clear_node(i);
673 	}
674 	numa_init_array();
675 
676 	return 0;
677 }
678 
679 /**
680  * dummy_numa_init - Fallback dummy NUMA init
681  *
682  * Used if there's no underlying NUMA architecture, NUMA initialization
683  * fails, or NUMA is disabled on the command line.
684  *
685  * Must online at least one node and add memory blocks that cover all
686  * allowed memory.  This function must not fail.
687  */
688 static int __init dummy_numa_init(void)
689 {
690 	printk(KERN_INFO "%s\n",
691 	       numa_off ? "NUMA turned off" : "No NUMA configuration found");
692 	printk(KERN_INFO "Faking a node at [mem %#018Lx-%#018Lx]\n",
693 	       0LLU, PFN_PHYS(max_pfn) - 1);
694 
695 	node_set(0, numa_nodes_parsed);
696 	numa_add_memblk(0, 0, PFN_PHYS(max_pfn));
697 
698 	return 0;
699 }
700 
701 /**
702  * x86_numa_init - Initialize NUMA
703  *
704  * Try each configured NUMA initialization method until one succeeds.  The
705  * last fallback is dummy single node config encomapssing whole memory and
706  * never fails.
707  */
708 void __init x86_numa_init(void)
709 {
710 	if (!numa_off) {
711 #ifdef CONFIG_ACPI_NUMA
712 		if (!numa_init(x86_acpi_numa_init))
713 			return;
714 #endif
715 #ifdef CONFIG_AMD_NUMA
716 		if (!numa_init(amd_numa_init))
717 			return;
718 #endif
719 	}
720 
721 	numa_init(dummy_numa_init);
722 }
723 
724 static void __init init_memory_less_node(int nid)
725 {
726 	unsigned long zones_size[MAX_NR_ZONES] = {0};
727 	unsigned long zholes_size[MAX_NR_ZONES] = {0};
728 
729 	/* Allocate and initialize node data. Memory-less node is now online.*/
730 	alloc_node_data(nid);
731 	free_area_init_node(nid, zones_size, 0, zholes_size);
732 
733 	/*
734 	 * All zonelists will be built later in start_kernel() after per cpu
735 	 * areas are initialized.
736 	 */
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  */
753 void __init init_cpu_to_node(void)
754 {
755 	int cpu;
756 	u16 *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 		if (!node_online(node))
767 			init_memory_less_node(node);
768 
769 		numa_set_node(cpu, node);
770 	}
771 }
772 
773 #ifndef CONFIG_DEBUG_PER_CPU_MAPS
774 
775 # ifndef CONFIG_NUMA_EMU
776 void numa_add_cpu(int cpu)
777 {
778 	cpumask_set_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
779 }
780 
781 void numa_remove_cpu(int cpu)
782 {
783 	cpumask_clear_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
784 }
785 # endif	/* !CONFIG_NUMA_EMU */
786 
787 #else	/* !CONFIG_DEBUG_PER_CPU_MAPS */
788 
789 int __cpu_to_node(int cpu)
790 {
791 	if (early_per_cpu_ptr(x86_cpu_to_node_map)) {
792 		printk(KERN_WARNING
793 			"cpu_to_node(%d): usage too early!\n", cpu);
794 		dump_stack();
795 		return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
796 	}
797 	return per_cpu(x86_cpu_to_node_map, cpu);
798 }
799 EXPORT_SYMBOL(__cpu_to_node);
800 
801 /*
802  * Same function as cpu_to_node() but used if called before the
803  * per_cpu areas are setup.
804  */
805 int early_cpu_to_node(int cpu)
806 {
807 	if (early_per_cpu_ptr(x86_cpu_to_node_map))
808 		return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
809 
810 	if (!cpu_possible(cpu)) {
811 		printk(KERN_WARNING
812 			"early_cpu_to_node(%d): no per_cpu area!\n", cpu);
813 		dump_stack();
814 		return NUMA_NO_NODE;
815 	}
816 	return per_cpu(x86_cpu_to_node_map, cpu);
817 }
818 
819 void debug_cpumask_set_cpu(int cpu, int node, bool enable)
820 {
821 	struct cpumask *mask;
822 
823 	if (node == NUMA_NO_NODE) {
824 		/* early_cpu_to_node() already emits a warning and trace */
825 		return;
826 	}
827 	mask = node_to_cpumask_map[node];
828 	if (!mask) {
829 		pr_err("node_to_cpumask_map[%i] NULL\n", node);
830 		dump_stack();
831 		return;
832 	}
833 
834 	if (enable)
835 		cpumask_set_cpu(cpu, mask);
836 	else
837 		cpumask_clear_cpu(cpu, mask);
838 
839 	printk(KERN_DEBUG "%s cpu %d node %d: mask now %*pbl\n",
840 		enable ? "numa_add_cpu" : "numa_remove_cpu",
841 		cpu, node, cpumask_pr_args(mask));
842 	return;
843 }
844 
845 # ifndef CONFIG_NUMA_EMU
846 static void numa_set_cpumask(int cpu, bool enable)
847 {
848 	debug_cpumask_set_cpu(cpu, early_cpu_to_node(cpu), enable);
849 }
850 
851 void numa_add_cpu(int cpu)
852 {
853 	numa_set_cpumask(cpu, true);
854 }
855 
856 void numa_remove_cpu(int cpu)
857 {
858 	numa_set_cpumask(cpu, false);
859 }
860 # endif	/* !CONFIG_NUMA_EMU */
861 
862 /*
863  * Returns a pointer to the bitmask of CPUs on Node 'node'.
864  */
865 const struct cpumask *cpumask_of_node(int node)
866 {
867 	if (node >= nr_node_ids) {
868 		printk(KERN_WARNING
869 			"cpumask_of_node(%d): node > nr_node_ids(%d)\n",
870 			node, nr_node_ids);
871 		dump_stack();
872 		return cpu_none_mask;
873 	}
874 	if (node_to_cpumask_map[node] == NULL) {
875 		printk(KERN_WARNING
876 			"cpumask_of_node(%d): no node_to_cpumask_map!\n",
877 			node);
878 		dump_stack();
879 		return cpu_online_mask;
880 	}
881 	return node_to_cpumask_map[node];
882 }
883 EXPORT_SYMBOL(cpumask_of_node);
884 
885 #endif	/* !CONFIG_DEBUG_PER_CPU_MAPS */
886 
887 #ifdef CONFIG_MEMORY_HOTPLUG
888 int memory_add_physaddr_to_nid(u64 start)
889 {
890 	struct numa_meminfo *mi = &numa_meminfo;
891 	int nid = mi->blk[0].nid;
892 	int i;
893 
894 	for (i = 0; i < mi->nr_blks; i++)
895 		if (mi->blk[i].start <= start && mi->blk[i].end > start)
896 			nid = mi->blk[i].nid;
897 	return nid;
898 }
899 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
900 #endif
901