xref: /openbmc/linux/arch/x86/mm/numa.c (revision 239480ab)
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/bootmem.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
29 #ifndef CONFIG_MEMORY_HOTPLUG
30 __initdata
31 #endif
32 ;
33 
34 static int numa_distance_cnt;
35 static u8 *numa_distance;
36 
37 static __init int numa_setup(char *opt)
38 {
39 	if (!opt)
40 		return -EINVAL;
41 	if (!strncmp(opt, "off", 3))
42 		numa_off = 1;
43 #ifdef CONFIG_NUMA_EMU
44 	if (!strncmp(opt, "fake=", 5))
45 		numa_emu_cmdline(opt + 5);
46 #endif
47 #ifdef CONFIG_ACPI_NUMA
48 	if (!strncmp(opt, "noacpi", 6))
49 		acpi_numa = -1;
50 #endif
51 	return 0;
52 }
53 early_param("numa", numa_setup);
54 
55 /*
56  * apicid, cpu, node mappings
57  */
58 s16 __apicid_to_node[MAX_LOCAL_APIC] = {
59 	[0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE
60 };
61 
62 int numa_cpu_node(int cpu)
63 {
64 	int apicid = early_per_cpu(x86_cpu_to_apicid, cpu);
65 
66 	if (apicid != BAD_APICID)
67 		return __apicid_to_node[apicid];
68 	return NUMA_NO_NODE;
69 }
70 
71 cpumask_var_t node_to_cpumask_map[MAX_NUMNODES];
72 EXPORT_SYMBOL(node_to_cpumask_map);
73 
74 /*
75  * Map cpu index to node index
76  */
77 DEFINE_EARLY_PER_CPU(int, x86_cpu_to_node_map, NUMA_NO_NODE);
78 EXPORT_EARLY_PER_CPU_SYMBOL(x86_cpu_to_node_map);
79 
80 void numa_set_node(int cpu, int node)
81 {
82 	int *cpu_to_node_map = early_per_cpu_ptr(x86_cpu_to_node_map);
83 
84 	/* early setting, no percpu area yet */
85 	if (cpu_to_node_map) {
86 		cpu_to_node_map[cpu] = node;
87 		return;
88 	}
89 
90 #ifdef CONFIG_DEBUG_PER_CPU_MAPS
91 	if (cpu >= nr_cpu_ids || !cpu_possible(cpu)) {
92 		printk(KERN_ERR "numa_set_node: invalid cpu# (%d)\n", cpu);
93 		dump_stack();
94 		return;
95 	}
96 #endif
97 	per_cpu(x86_cpu_to_node_map, cpu) = node;
98 
99 	set_cpu_numa_node(cpu, node);
100 }
101 
102 void numa_clear_node(int cpu)
103 {
104 	numa_set_node(cpu, NUMA_NO_NODE);
105 }
106 
107 /*
108  * Allocate node_to_cpumask_map based on number of available nodes
109  * Requires node_possible_map to be valid.
110  *
111  * Note: cpumask_of_node() is not valid until after this is done.
112  * (Use CONFIG_DEBUG_PER_CPU_MAPS to check this.)
113  */
114 void __init setup_node_to_cpumask_map(void)
115 {
116 	unsigned int node;
117 
118 	/* setup nr_node_ids if not done yet */
119 	if (nr_node_ids == MAX_NUMNODES)
120 		setup_nr_node_ids();
121 
122 	/* allocate the map */
123 	for (node = 0; node < nr_node_ids; node++)
124 		alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]);
125 
126 	/* cpumask_of_node() will now work */
127 	pr_debug("Node to cpumask map for %d nodes\n", nr_node_ids);
128 }
129 
130 static int __init numa_add_memblk_to(int nid, u64 start, u64 end,
131 				     struct numa_meminfo *mi)
132 {
133 	/* ignore zero length blks */
134 	if (start == end)
135 		return 0;
136 
137 	/* whine about and ignore invalid blks */
138 	if (start > end || nid < 0 || nid >= MAX_NUMNODES) {
139 		pr_warning("NUMA: Warning: invalid memblk node %d [mem %#010Lx-%#010Lx]\n",
140 			   nid, start, end - 1);
141 		return 0;
142 	}
143 
144 	if (mi->nr_blks >= NR_NODE_MEMBLKS) {
145 		pr_err("NUMA: too many memblk ranges\n");
146 		return -EINVAL;
147 	}
148 
149 	mi->blk[mi->nr_blks].start = start;
150 	mi->blk[mi->nr_blks].end = end;
151 	mi->blk[mi->nr_blks].nid = nid;
152 	mi->nr_blks++;
153 	return 0;
154 }
155 
156 /**
157  * numa_remove_memblk_from - Remove one numa_memblk from a numa_meminfo
158  * @idx: Index of memblk to remove
159  * @mi: numa_meminfo to remove memblk from
160  *
161  * Remove @idx'th numa_memblk from @mi by shifting @mi->blk[] and
162  * decrementing @mi->nr_blks.
163  */
164 void __init numa_remove_memblk_from(int idx, struct numa_meminfo *mi)
165 {
166 	mi->nr_blks--;
167 	memmove(&mi->blk[idx], &mi->blk[idx + 1],
168 		(mi->nr_blks - idx) * sizeof(mi->blk[0]));
169 }
170 
171 /**
172  * numa_add_memblk - Add one numa_memblk to numa_meminfo
173  * @nid: NUMA node ID of the new memblk
174  * @start: Start address of the new memblk
175  * @end: End address of the new memblk
176  *
177  * Add a new memblk to the default numa_meminfo.
178  *
179  * RETURNS:
180  * 0 on success, -errno on failure.
181  */
182 int __init numa_add_memblk(int nid, u64 start, u64 end)
183 {
184 	return numa_add_memblk_to(nid, start, end, &numa_meminfo);
185 }
186 
187 /* Allocate NODE_DATA for a node on the local memory */
188 static void __init alloc_node_data(int nid)
189 {
190 	const size_t nd_size = roundup(sizeof(pg_data_t), PAGE_SIZE);
191 	u64 nd_pa;
192 	void *nd;
193 	int tnid;
194 
195 	/*
196 	 * Allocate node data.  Try node-local memory and then any node.
197 	 * Never allocate in DMA zone.
198 	 */
199 	nd_pa = memblock_alloc_nid(nd_size, SMP_CACHE_BYTES, nid);
200 	if (!nd_pa) {
201 		nd_pa = __memblock_alloc_base(nd_size, SMP_CACHE_BYTES,
202 					      MEMBLOCK_ALLOC_ACCESSIBLE);
203 		if (!nd_pa) {
204 			pr_err("Cannot find %zu bytes in any node (initial node: %d)\n",
205 			       nd_size, nid);
206 			return;
207 		}
208 	}
209 	nd = __va(nd_pa);
210 
211 	/* report and initialize */
212 	printk(KERN_INFO "NODE_DATA(%d) allocated [mem %#010Lx-%#010Lx]\n", nid,
213 	       nd_pa, nd_pa + nd_size - 1);
214 	tnid = early_pfn_to_nid(nd_pa >> PAGE_SHIFT);
215 	if (tnid != nid)
216 		printk(KERN_INFO "    NODE_DATA(%d) on node %d\n", nid, tnid);
217 
218 	node_data[nid] = nd;
219 	memset(NODE_DATA(nid), 0, sizeof(pg_data_t));
220 
221 	node_set_online(nid);
222 }
223 
224 /**
225  * numa_cleanup_meminfo - Cleanup a numa_meminfo
226  * @mi: numa_meminfo to clean up
227  *
228  * Sanitize @mi by merging and removing unnecessary memblks.  Also check for
229  * conflicts and clear unused memblks.
230  *
231  * RETURNS:
232  * 0 on success, -errno on failure.
233  */
234 int __init numa_cleanup_meminfo(struct numa_meminfo *mi)
235 {
236 	const u64 low = 0;
237 	const u64 high = PFN_PHYS(max_pfn);
238 	int i, j, k;
239 
240 	/* first, trim all entries */
241 	for (i = 0; i < mi->nr_blks; i++) {
242 		struct numa_memblk *bi = &mi->blk[i];
243 
244 		/* make sure all blocks are inside the limits */
245 		bi->start = max(bi->start, low);
246 		bi->end = min(bi->end, high);
247 
248 		/* and there's no empty or non-exist block */
249 		if (bi->start >= bi->end ||
250 		    !memblock_overlaps_region(&memblock.memory,
251 			bi->start, bi->end - bi->start))
252 			numa_remove_memblk_from(i--, mi);
253 	}
254 
255 	/* merge neighboring / overlapping entries */
256 	for (i = 0; i < mi->nr_blks; i++) {
257 		struct numa_memblk *bi = &mi->blk[i];
258 
259 		for (j = i + 1; j < mi->nr_blks; j++) {
260 			struct numa_memblk *bj = &mi->blk[j];
261 			u64 start, end;
262 
263 			/*
264 			 * See whether there are overlapping blocks.  Whine
265 			 * about but allow overlaps of the same nid.  They
266 			 * will be merged below.
267 			 */
268 			if (bi->end > bj->start && bi->start < bj->end) {
269 				if (bi->nid != bj->nid) {
270 					pr_err("NUMA: node %d [mem %#010Lx-%#010Lx] overlaps with node %d [mem %#010Lx-%#010Lx]\n",
271 					       bi->nid, bi->start, bi->end - 1,
272 					       bj->nid, bj->start, bj->end - 1);
273 					return -EINVAL;
274 				}
275 				pr_warning("NUMA: Warning: node %d [mem %#010Lx-%#010Lx] overlaps with itself [mem %#010Lx-%#010Lx]\n",
276 					   bi->nid, bi->start, bi->end - 1,
277 					   bj->start, bj->end - 1);
278 			}
279 
280 			/*
281 			 * Join together blocks on the same node, holes
282 			 * between which don't overlap with memory on other
283 			 * nodes.
284 			 */
285 			if (bi->nid != bj->nid)
286 				continue;
287 			start = min(bi->start, bj->start);
288 			end = max(bi->end, bj->end);
289 			for (k = 0; k < mi->nr_blks; k++) {
290 				struct numa_memblk *bk = &mi->blk[k];
291 
292 				if (bi->nid == bk->nid)
293 					continue;
294 				if (start < bk->end && end > bk->start)
295 					break;
296 			}
297 			if (k < mi->nr_blks)
298 				continue;
299 			printk(KERN_INFO "NUMA: Node %d [mem %#010Lx-%#010Lx] + [mem %#010Lx-%#010Lx] -> [mem %#010Lx-%#010Lx]\n",
300 			       bi->nid, bi->start, bi->end - 1, bj->start,
301 			       bj->end - 1, start, end - 1);
302 			bi->start = start;
303 			bi->end = end;
304 			numa_remove_memblk_from(j--, mi);
305 		}
306 	}
307 
308 	/* clear unused ones */
309 	for (i = mi->nr_blks; i < ARRAY_SIZE(mi->blk); i++) {
310 		mi->blk[i].start = mi->blk[i].end = 0;
311 		mi->blk[i].nid = NUMA_NO_NODE;
312 	}
313 
314 	return 0;
315 }
316 
317 /*
318  * Set nodes, which have memory in @mi, in *@nodemask.
319  */
320 static void __init numa_nodemask_from_meminfo(nodemask_t *nodemask,
321 					      const struct numa_meminfo *mi)
322 {
323 	int i;
324 
325 	for (i = 0; i < ARRAY_SIZE(mi->blk); i++)
326 		if (mi->blk[i].start != mi->blk[i].end &&
327 		    mi->blk[i].nid != NUMA_NO_NODE)
328 			node_set(mi->blk[i].nid, *nodemask);
329 }
330 
331 /**
332  * numa_reset_distance - Reset NUMA distance table
333  *
334  * The current table is freed.  The next numa_set_distance() call will
335  * create a new one.
336  */
337 void __init numa_reset_distance(void)
338 {
339 	size_t size = numa_distance_cnt * numa_distance_cnt * sizeof(numa_distance[0]);
340 
341 	/* numa_distance could be 1LU marking allocation failure, test cnt */
342 	if (numa_distance_cnt)
343 		memblock_free(__pa(numa_distance), size);
344 	numa_distance_cnt = 0;
345 	numa_distance = NULL;	/* enable table creation */
346 }
347 
348 static int __init numa_alloc_distance(void)
349 {
350 	nodemask_t nodes_parsed;
351 	size_t size;
352 	int i, j, cnt = 0;
353 	u64 phys;
354 
355 	/* size the new table and allocate it */
356 	nodes_parsed = numa_nodes_parsed;
357 	numa_nodemask_from_meminfo(&nodes_parsed, &numa_meminfo);
358 
359 	for_each_node_mask(i, nodes_parsed)
360 		cnt = i;
361 	cnt++;
362 	size = cnt * cnt * sizeof(numa_distance[0]);
363 
364 	phys = memblock_find_in_range(0, PFN_PHYS(max_pfn_mapped),
365 				      size, PAGE_SIZE);
366 	if (!phys) {
367 		pr_warning("NUMA: Warning: can't allocate distance table!\n");
368 		/* don't retry until explicitly reset */
369 		numa_distance = (void *)1LU;
370 		return -ENOMEM;
371 	}
372 	memblock_reserve(phys, size);
373 
374 	numa_distance = __va(phys);
375 	numa_distance_cnt = cnt;
376 
377 	/* fill with the default distances */
378 	for (i = 0; i < cnt; i++)
379 		for (j = 0; j < cnt; j++)
380 			numa_distance[i * cnt + j] = i == j ?
381 				LOCAL_DISTANCE : REMOTE_DISTANCE;
382 	printk(KERN_DEBUG "NUMA: Initialized distance table, cnt=%d\n", cnt);
383 
384 	return 0;
385 }
386 
387 /**
388  * numa_set_distance - Set NUMA distance from one NUMA to another
389  * @from: the 'from' node to set distance
390  * @to: the 'to'  node to set distance
391  * @distance: NUMA distance
392  *
393  * Set the distance from node @from to @to to @distance.  If distance table
394  * doesn't exist, one which is large enough to accommodate all the currently
395  * known nodes will be created.
396  *
397  * If such table cannot be allocated, a warning is printed and further
398  * calls are ignored until the distance table is reset with
399  * numa_reset_distance().
400  *
401  * If @from or @to is higher than the highest known node or lower than zero
402  * at the time of table creation or @distance doesn't make sense, the call
403  * is ignored.
404  * This is to allow simplification of specific NUMA config implementations.
405  */
406 void __init numa_set_distance(int from, int to, int distance)
407 {
408 	if (!numa_distance && numa_alloc_distance() < 0)
409 		return;
410 
411 	if (from >= numa_distance_cnt || to >= numa_distance_cnt ||
412 			from < 0 || to < 0) {
413 		pr_warn_once("NUMA: Warning: node ids are out of bound, from=%d to=%d distance=%d\n",
414 			    from, to, distance);
415 		return;
416 	}
417 
418 	if ((u8)distance != distance ||
419 	    (from == to && distance != LOCAL_DISTANCE)) {
420 		pr_warn_once("NUMA: Warning: invalid distance parameter, from=%d to=%d distance=%d\n",
421 			     from, to, distance);
422 		return;
423 	}
424 
425 	numa_distance[from * numa_distance_cnt + to] = distance;
426 }
427 
428 int __node_distance(int from, int to)
429 {
430 	if (from >= numa_distance_cnt || to >= numa_distance_cnt)
431 		return from == to ? LOCAL_DISTANCE : REMOTE_DISTANCE;
432 	return numa_distance[from * numa_distance_cnt + to];
433 }
434 EXPORT_SYMBOL(__node_distance);
435 
436 /*
437  * Sanity check to catch more bad NUMA configurations (they are amazingly
438  * common).  Make sure the nodes cover all memory.
439  */
440 static bool __init numa_meminfo_cover_memory(const struct numa_meminfo *mi)
441 {
442 	u64 numaram, e820ram;
443 	int i;
444 
445 	numaram = 0;
446 	for (i = 0; i < mi->nr_blks; i++) {
447 		u64 s = mi->blk[i].start >> PAGE_SHIFT;
448 		u64 e = mi->blk[i].end >> PAGE_SHIFT;
449 		numaram += e - s;
450 		numaram -= __absent_pages_in_range(mi->blk[i].nid, s, e);
451 		if ((s64)numaram < 0)
452 			numaram = 0;
453 	}
454 
455 	e820ram = max_pfn - absent_pages_in_range(0, max_pfn);
456 
457 	/* We seem to lose 3 pages somewhere. Allow 1M of slack. */
458 	if ((s64)(e820ram - numaram) >= (1 << (20 - PAGE_SHIFT))) {
459 		printk(KERN_ERR "NUMA: nodes only cover %LuMB of your %LuMB e820 RAM. Not used.\n",
460 		       (numaram << PAGE_SHIFT) >> 20,
461 		       (e820ram << PAGE_SHIFT) >> 20);
462 		return false;
463 	}
464 	return true;
465 }
466 
467 /*
468  * Mark all currently memblock-reserved physical memory (which covers the
469  * kernel's own memory ranges) as hot-unswappable.
470  */
471 static void __init numa_clear_kernel_node_hotplug(void)
472 {
473 	nodemask_t reserved_nodemask = NODE_MASK_NONE;
474 	struct memblock_region *mb_region;
475 	int i;
476 
477 	/*
478 	 * We have to do some preprocessing of memblock regions, to
479 	 * make them suitable for reservation.
480 	 *
481 	 * At this time, all memory regions reserved by memblock are
482 	 * used by the kernel, but those regions are not split up
483 	 * along node boundaries yet, and don't necessarily have their
484 	 * node ID set yet either.
485 	 *
486 	 * So iterate over all memory known to the x86 architecture,
487 	 * and use those ranges to set the nid in memblock.reserved.
488 	 * This will split up the memblock regions along node
489 	 * boundaries and will set the node IDs as well.
490 	 */
491 	for (i = 0; i < numa_meminfo.nr_blks; i++) {
492 		struct numa_memblk *mb = numa_meminfo.blk + i;
493 		int ret;
494 
495 		ret = memblock_set_node(mb->start, mb->end - mb->start, &memblock.reserved, mb->nid);
496 		WARN_ON_ONCE(ret);
497 	}
498 
499 	/*
500 	 * Now go over all reserved memblock regions, to construct a
501 	 * node mask of all kernel reserved memory areas.
502 	 *
503 	 * [ Note, when booting with mem=nn[kMG] or in a kdump kernel,
504 	 *   numa_meminfo might not include all memblock.reserved
505 	 *   memory ranges, because quirks such as trim_snb_memory()
506 	 *   reserve specific pages for Sandy Bridge graphics. ]
507 	 */
508 	for_each_memblock(reserved, mb_region) {
509 		if (mb_region->nid != MAX_NUMNODES)
510 			node_set(mb_region->nid, reserved_nodemask);
511 	}
512 
513 	/*
514 	 * Finally, clear the MEMBLOCK_HOTPLUG flag for all memory
515 	 * belonging to the reserved node mask.
516 	 *
517 	 * Note that this will include memory regions that reside
518 	 * on nodes that contain kernel memory - entire nodes
519 	 * become hot-unpluggable:
520 	 */
521 	for (i = 0; i < numa_meminfo.nr_blks; i++) {
522 		struct numa_memblk *mb = numa_meminfo.blk + i;
523 
524 		if (!node_isset(mb->nid, reserved_nodemask))
525 			continue;
526 
527 		memblock_clear_hotplug(mb->start, mb->end - mb->start);
528 	}
529 }
530 
531 static int __init numa_register_memblks(struct numa_meminfo *mi)
532 {
533 	unsigned long uninitialized_var(pfn_align);
534 	int i, nid;
535 
536 	/* Account for nodes with cpus and no memory */
537 	node_possible_map = numa_nodes_parsed;
538 	numa_nodemask_from_meminfo(&node_possible_map, mi);
539 	if (WARN_ON(nodes_empty(node_possible_map)))
540 		return -EINVAL;
541 
542 	for (i = 0; i < mi->nr_blks; i++) {
543 		struct numa_memblk *mb = &mi->blk[i];
544 		memblock_set_node(mb->start, mb->end - mb->start,
545 				  &memblock.memory, mb->nid);
546 	}
547 
548 	/*
549 	 * At very early time, the kernel have to use some memory such as
550 	 * loading the kernel image. We cannot prevent this anyway. So any
551 	 * node the kernel resides in should be un-hotpluggable.
552 	 *
553 	 * And when we come here, alloc node data won't fail.
554 	 */
555 	numa_clear_kernel_node_hotplug();
556 
557 	/*
558 	 * If sections array is gonna be used for pfn -> nid mapping, check
559 	 * whether its granularity is fine enough.
560 	 */
561 #ifdef NODE_NOT_IN_PAGE_FLAGS
562 	pfn_align = node_map_pfn_alignment();
563 	if (pfn_align && pfn_align < PAGES_PER_SECTION) {
564 		printk(KERN_WARNING "Node alignment %LuMB < min %LuMB, rejecting NUMA config\n",
565 		       PFN_PHYS(pfn_align) >> 20,
566 		       PFN_PHYS(PAGES_PER_SECTION) >> 20);
567 		return -EINVAL;
568 	}
569 #endif
570 	if (!numa_meminfo_cover_memory(mi))
571 		return -EINVAL;
572 
573 	/* Finally register nodes. */
574 	for_each_node_mask(nid, node_possible_map) {
575 		u64 start = PFN_PHYS(max_pfn);
576 		u64 end = 0;
577 
578 		for (i = 0; i < mi->nr_blks; i++) {
579 			if (nid != mi->blk[i].nid)
580 				continue;
581 			start = min(mi->blk[i].start, start);
582 			end = max(mi->blk[i].end, end);
583 		}
584 
585 		if (start >= end)
586 			continue;
587 
588 		/*
589 		 * Don't confuse VM with a node that doesn't have the
590 		 * minimum amount of memory:
591 		 */
592 		if (end && (end - start) < NODE_MIN_SIZE)
593 			continue;
594 
595 		alloc_node_data(nid);
596 	}
597 
598 	/* Dump memblock with node info and return. */
599 	memblock_dump_all();
600 	return 0;
601 }
602 
603 /*
604  * There are unfortunately some poorly designed mainboards around that
605  * only connect memory to a single CPU. This breaks the 1:1 cpu->node
606  * mapping. To avoid this fill in the mapping for all possible CPUs,
607  * as the number of CPUs is not known yet. We round robin the existing
608  * nodes.
609  */
610 static void __init numa_init_array(void)
611 {
612 	int rr, i;
613 
614 	rr = first_node(node_online_map);
615 	for (i = 0; i < nr_cpu_ids; i++) {
616 		if (early_cpu_to_node(i) != NUMA_NO_NODE)
617 			continue;
618 		numa_set_node(i, rr);
619 		rr = next_node_in(rr, node_online_map);
620 	}
621 }
622 
623 static int __init numa_init(int (*init_func)(void))
624 {
625 	int i;
626 	int ret;
627 
628 	for (i = 0; i < MAX_LOCAL_APIC; i++)
629 		set_apicid_to_node(i, NUMA_NO_NODE);
630 
631 	nodes_clear(numa_nodes_parsed);
632 	nodes_clear(node_possible_map);
633 	nodes_clear(node_online_map);
634 	memset(&numa_meminfo, 0, sizeof(numa_meminfo));
635 	WARN_ON(memblock_set_node(0, ULLONG_MAX, &memblock.memory,
636 				  MAX_NUMNODES));
637 	WARN_ON(memblock_set_node(0, ULLONG_MAX, &memblock.reserved,
638 				  MAX_NUMNODES));
639 	/* In case that parsing SRAT failed. */
640 	WARN_ON(memblock_clear_hotplug(0, ULLONG_MAX));
641 	numa_reset_distance();
642 
643 	ret = init_func();
644 	if (ret < 0)
645 		return ret;
646 
647 	/*
648 	 * We reset memblock back to the top-down direction
649 	 * here because if we configured ACPI_NUMA, we have
650 	 * parsed SRAT in init_func(). It is ok to have the
651 	 * reset here even if we did't configure ACPI_NUMA
652 	 * or acpi numa init fails and fallbacks to dummy
653 	 * numa init.
654 	 */
655 	memblock_set_bottom_up(false);
656 
657 	ret = numa_cleanup_meminfo(&numa_meminfo);
658 	if (ret < 0)
659 		return ret;
660 
661 	numa_emulation(&numa_meminfo, numa_distance_cnt);
662 
663 	ret = numa_register_memblks(&numa_meminfo);
664 	if (ret < 0)
665 		return ret;
666 
667 	for (i = 0; i < nr_cpu_ids; i++) {
668 		int nid = early_cpu_to_node(i);
669 
670 		if (nid == NUMA_NO_NODE)
671 			continue;
672 		if (!node_online(nid))
673 			numa_clear_node(i);
674 	}
675 	numa_init_array();
676 
677 	return 0;
678 }
679 
680 /**
681  * dummy_numa_init - Fallback dummy NUMA init
682  *
683  * Used if there's no underlying NUMA architecture, NUMA initialization
684  * fails, or NUMA is disabled on the command line.
685  *
686  * Must online at least one node and add memory blocks that cover all
687  * allowed memory.  This function must not fail.
688  */
689 static int __init dummy_numa_init(void)
690 {
691 	printk(KERN_INFO "%s\n",
692 	       numa_off ? "NUMA turned off" : "No NUMA configuration found");
693 	printk(KERN_INFO "Faking a node at [mem %#018Lx-%#018Lx]\n",
694 	       0LLU, PFN_PHYS(max_pfn) - 1);
695 
696 	node_set(0, numa_nodes_parsed);
697 	numa_add_memblk(0, 0, PFN_PHYS(max_pfn));
698 
699 	return 0;
700 }
701 
702 /**
703  * x86_numa_init - Initialize NUMA
704  *
705  * Try each configured NUMA initialization method until one succeeds.  The
706  * last fallback is dummy single node config encomapssing whole memory and
707  * never fails.
708  */
709 void __init x86_numa_init(void)
710 {
711 	if (!numa_off) {
712 #ifdef CONFIG_ACPI_NUMA
713 		if (!numa_init(x86_acpi_numa_init))
714 			return;
715 #endif
716 #ifdef CONFIG_AMD_NUMA
717 		if (!numa_init(amd_numa_init))
718 			return;
719 #endif
720 	}
721 
722 	numa_init(dummy_numa_init);
723 }
724 
725 static void __init init_memory_less_node(int nid)
726 {
727 	unsigned long zones_size[MAX_NR_ZONES] = {0};
728 	unsigned long zholes_size[MAX_NR_ZONES] = {0};
729 
730 	/* Allocate and initialize node data. Memory-less node is now online.*/
731 	alloc_node_data(nid);
732 	free_area_init_node(nid, zones_size, 0, zholes_size);
733 
734 	/*
735 	 * All zonelists will be built later in start_kernel() after per cpu
736 	 * areas are initialized.
737 	 */
738 }
739 
740 /*
741  * Setup early cpu_to_node.
742  *
743  * Populate cpu_to_node[] only if x86_cpu_to_apicid[],
744  * and apicid_to_node[] tables have valid entries for a CPU.
745  * This means we skip cpu_to_node[] initialisation for NUMA
746  * emulation and faking node case (when running a kernel compiled
747  * for NUMA on a non NUMA box), which is OK as cpu_to_node[]
748  * is already initialized in a round robin manner at numa_init_array,
749  * prior to this call, and this initialization is good enough
750  * for the fake NUMA cases.
751  *
752  * Called before the per_cpu areas are setup.
753  */
754 void __init init_cpu_to_node(void)
755 {
756 	int cpu;
757 	u16 *cpu_to_apicid = early_per_cpu_ptr(x86_cpu_to_apicid);
758 
759 	BUG_ON(cpu_to_apicid == NULL);
760 
761 	for_each_possible_cpu(cpu) {
762 		int node = numa_cpu_node(cpu);
763 
764 		if (node == NUMA_NO_NODE)
765 			continue;
766 
767 		if (!node_online(node))
768 			init_memory_less_node(node);
769 
770 		numa_set_node(cpu, node);
771 	}
772 }
773 
774 #ifndef CONFIG_DEBUG_PER_CPU_MAPS
775 
776 # ifndef CONFIG_NUMA_EMU
777 void numa_add_cpu(int cpu)
778 {
779 	cpumask_set_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
780 }
781 
782 void numa_remove_cpu(int cpu)
783 {
784 	cpumask_clear_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
785 }
786 # endif	/* !CONFIG_NUMA_EMU */
787 
788 #else	/* !CONFIG_DEBUG_PER_CPU_MAPS */
789 
790 int __cpu_to_node(int cpu)
791 {
792 	if (early_per_cpu_ptr(x86_cpu_to_node_map)) {
793 		printk(KERN_WARNING
794 			"cpu_to_node(%d): usage too early!\n", cpu);
795 		dump_stack();
796 		return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
797 	}
798 	return per_cpu(x86_cpu_to_node_map, cpu);
799 }
800 EXPORT_SYMBOL(__cpu_to_node);
801 
802 /*
803  * Same function as cpu_to_node() but used if called before the
804  * per_cpu areas are setup.
805  */
806 int early_cpu_to_node(int cpu)
807 {
808 	if (early_per_cpu_ptr(x86_cpu_to_node_map))
809 		return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
810 
811 	if (!cpu_possible(cpu)) {
812 		printk(KERN_WARNING
813 			"early_cpu_to_node(%d): no per_cpu area!\n", cpu);
814 		dump_stack();
815 		return NUMA_NO_NODE;
816 	}
817 	return per_cpu(x86_cpu_to_node_map, cpu);
818 }
819 
820 void debug_cpumask_set_cpu(int cpu, int node, bool enable)
821 {
822 	struct cpumask *mask;
823 
824 	if (node == NUMA_NO_NODE) {
825 		/* early_cpu_to_node() already emits a warning and trace */
826 		return;
827 	}
828 	mask = node_to_cpumask_map[node];
829 	if (!mask) {
830 		pr_err("node_to_cpumask_map[%i] NULL\n", node);
831 		dump_stack();
832 		return;
833 	}
834 
835 	if (enable)
836 		cpumask_set_cpu(cpu, mask);
837 	else
838 		cpumask_clear_cpu(cpu, mask);
839 
840 	printk(KERN_DEBUG "%s cpu %d node %d: mask now %*pbl\n",
841 		enable ? "numa_add_cpu" : "numa_remove_cpu",
842 		cpu, node, cpumask_pr_args(mask));
843 	return;
844 }
845 
846 # ifndef CONFIG_NUMA_EMU
847 static void numa_set_cpumask(int cpu, bool enable)
848 {
849 	debug_cpumask_set_cpu(cpu, early_cpu_to_node(cpu), enable);
850 }
851 
852 void numa_add_cpu(int cpu)
853 {
854 	numa_set_cpumask(cpu, true);
855 }
856 
857 void numa_remove_cpu(int cpu)
858 {
859 	numa_set_cpumask(cpu, false);
860 }
861 # endif	/* !CONFIG_NUMA_EMU */
862 
863 /*
864  * Returns a pointer to the bitmask of CPUs on Node 'node'.
865  */
866 const struct cpumask *cpumask_of_node(int node)
867 {
868 	if (node >= nr_node_ids) {
869 		printk(KERN_WARNING
870 			"cpumask_of_node(%d): node > nr_node_ids(%d)\n",
871 			node, nr_node_ids);
872 		dump_stack();
873 		return cpu_none_mask;
874 	}
875 	if (node_to_cpumask_map[node] == NULL) {
876 		printk(KERN_WARNING
877 			"cpumask_of_node(%d): no node_to_cpumask_map!\n",
878 			node);
879 		dump_stack();
880 		return cpu_online_mask;
881 	}
882 	return node_to_cpumask_map[node];
883 }
884 EXPORT_SYMBOL(cpumask_of_node);
885 
886 #endif	/* !CONFIG_DEBUG_PER_CPU_MAPS */
887 
888 #ifdef CONFIG_MEMORY_HOTPLUG
889 int memory_add_physaddr_to_nid(u64 start)
890 {
891 	struct numa_meminfo *mi = &numa_meminfo;
892 	int nid = mi->blk[0].nid;
893 	int i;
894 
895 	for (i = 0; i < mi->nr_blks; i++)
896 		if (mi->blk[i].start <= start && mi->blk[i].end > start)
897 			nid = mi->blk[i].nid;
898 	return nid;
899 }
900 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
901 #endif
902