xref: /linux/arch/powerpc/mm/numa.c (revision c537b994505099b7197e7d3125b942ecbcc51eb6)
1 /*
2  * pSeries NUMA support
3  *
4  * Copyright (C) 2002 Anton Blanchard <anton@au.ibm.com>, IBM
5  *
6  * This program is free software; you can redistribute it and/or
7  * modify it under the terms of the GNU General Public License
8  * as published by the Free Software Foundation; either version
9  * 2 of the License, or (at your option) any later version.
10  */
11 #include <linux/threads.h>
12 #include <linux/bootmem.h>
13 #include <linux/init.h>
14 #include <linux/mm.h>
15 #include <linux/mmzone.h>
16 #include <linux/module.h>
17 #include <linux/nodemask.h>
18 #include <linux/cpu.h>
19 #include <linux/notifier.h>
20 #include <asm/sparsemem.h>
21 #include <asm/lmb.h>
22 #include <asm/system.h>
23 #include <asm/smp.h>
24 
25 static int numa_enabled = 1;
26 
27 static int numa_debug;
28 #define dbg(args...) if (numa_debug) { printk(KERN_INFO args); }
29 
30 int numa_cpu_lookup_table[NR_CPUS];
31 cpumask_t numa_cpumask_lookup_table[MAX_NUMNODES];
32 struct pglist_data *node_data[MAX_NUMNODES];
33 
34 EXPORT_SYMBOL(numa_cpu_lookup_table);
35 EXPORT_SYMBOL(numa_cpumask_lookup_table);
36 EXPORT_SYMBOL(node_data);
37 
38 static bootmem_data_t __initdata plat_node_bdata[MAX_NUMNODES];
39 static int min_common_depth;
40 static int n_mem_addr_cells, n_mem_size_cells;
41 
42 static void __cpuinit map_cpu_to_node(int cpu, int node)
43 {
44 	numa_cpu_lookup_table[cpu] = node;
45 
46 	dbg("adding cpu %d to node %d\n", cpu, node);
47 
48 	if (!(cpu_isset(cpu, numa_cpumask_lookup_table[node])))
49 		cpu_set(cpu, numa_cpumask_lookup_table[node]);
50 }
51 
52 #ifdef CONFIG_HOTPLUG_CPU
53 static void unmap_cpu_from_node(unsigned long cpu)
54 {
55 	int node = numa_cpu_lookup_table[cpu];
56 
57 	dbg("removing cpu %lu from node %d\n", cpu, node);
58 
59 	if (cpu_isset(cpu, numa_cpumask_lookup_table[node])) {
60 		cpu_clear(cpu, numa_cpumask_lookup_table[node]);
61 	} else {
62 		printk(KERN_ERR "WARNING: cpu %lu not found in node %d\n",
63 		       cpu, node);
64 	}
65 }
66 #endif /* CONFIG_HOTPLUG_CPU */
67 
68 static struct device_node * __cpuinit find_cpu_node(unsigned int cpu)
69 {
70 	unsigned int hw_cpuid = get_hard_smp_processor_id(cpu);
71 	struct device_node *cpu_node = NULL;
72 	const unsigned int *interrupt_server, *reg;
73 	int len;
74 
75 	while ((cpu_node = of_find_node_by_type(cpu_node, "cpu")) != NULL) {
76 		/* Try interrupt server first */
77 		interrupt_server = get_property(cpu_node,
78 					"ibm,ppc-interrupt-server#s", &len);
79 
80 		len = len / sizeof(u32);
81 
82 		if (interrupt_server && (len > 0)) {
83 			while (len--) {
84 				if (interrupt_server[len] == hw_cpuid)
85 					return cpu_node;
86 			}
87 		} else {
88 			reg = get_property(cpu_node, "reg", &len);
89 			if (reg && (len > 0) && (reg[0] == hw_cpuid))
90 				return cpu_node;
91 		}
92 	}
93 
94 	return NULL;
95 }
96 
97 /* must hold reference to node during call */
98 static const int *of_get_associativity(struct device_node *dev)
99 {
100 	return get_property(dev, "ibm,associativity", NULL);
101 }
102 
103 /* Returns nid in the range [0..MAX_NUMNODES-1], or -1 if no useful numa
104  * info is found.
105  */
106 static int of_node_to_nid_single(struct device_node *device)
107 {
108 	int nid = -1;
109 	const unsigned int *tmp;
110 
111 	if (min_common_depth == -1)
112 		goto out;
113 
114 	tmp = of_get_associativity(device);
115 	if (!tmp)
116 		goto out;
117 
118 	if (tmp[0] >= min_common_depth)
119 		nid = tmp[min_common_depth];
120 
121 	/* POWER4 LPAR uses 0xffff as invalid node */
122 	if (nid == 0xffff || nid >= MAX_NUMNODES)
123 		nid = -1;
124 out:
125 	return nid;
126 }
127 
128 /* Walk the device tree upwards, looking for an associativity id */
129 int of_node_to_nid(struct device_node *device)
130 {
131 	struct device_node *tmp;
132 	int nid = -1;
133 
134 	of_node_get(device);
135 	while (device) {
136 		nid = of_node_to_nid_single(device);
137 		if (nid != -1)
138 			break;
139 
140 	        tmp = device;
141 		device = of_get_parent(tmp);
142 		of_node_put(tmp);
143 	}
144 	of_node_put(device);
145 
146 	return nid;
147 }
148 EXPORT_SYMBOL_GPL(of_node_to_nid);
149 
150 /*
151  * In theory, the "ibm,associativity" property may contain multiple
152  * associativity lists because a resource may be multiply connected
153  * into the machine.  This resource then has different associativity
154  * characteristics relative to its multiple connections.  We ignore
155  * this for now.  We also assume that all cpu and memory sets have
156  * their distances represented at a common level.  This won't be
157  * true for hierarchical NUMA.
158  *
159  * In any case the ibm,associativity-reference-points should give
160  * the correct depth for a normal NUMA system.
161  *
162  * - Dave Hansen <haveblue@us.ibm.com>
163  */
164 static int __init find_min_common_depth(void)
165 {
166 	int depth;
167 	const unsigned int *ref_points;
168 	struct device_node *rtas_root;
169 	unsigned int len;
170 
171 	rtas_root = of_find_node_by_path("/rtas");
172 
173 	if (!rtas_root)
174 		return -1;
175 
176 	/*
177 	 * this property is 2 32-bit integers, each representing a level of
178 	 * depth in the associativity nodes.  The first is for an SMP
179 	 * configuration (should be all 0's) and the second is for a normal
180 	 * NUMA configuration.
181 	 */
182 	ref_points = get_property(rtas_root,
183 			"ibm,associativity-reference-points", &len);
184 
185 	if ((len >= 1) && ref_points) {
186 		depth = ref_points[1];
187 	} else {
188 		dbg("NUMA: ibm,associativity-reference-points not found.\n");
189 		depth = -1;
190 	}
191 	of_node_put(rtas_root);
192 
193 	return depth;
194 }
195 
196 static void __init get_n_mem_cells(int *n_addr_cells, int *n_size_cells)
197 {
198 	struct device_node *memory = NULL;
199 
200 	memory = of_find_node_by_type(memory, "memory");
201 	if (!memory)
202 		panic("numa.c: No memory nodes found!");
203 
204 	*n_addr_cells = prom_n_addr_cells(memory);
205 	*n_size_cells = prom_n_size_cells(memory);
206 	of_node_put(memory);
207 }
208 
209 static unsigned long __devinit read_n_cells(int n, const unsigned int **buf)
210 {
211 	unsigned long result = 0;
212 
213 	while (n--) {
214 		result = (result << 32) | **buf;
215 		(*buf)++;
216 	}
217 	return result;
218 }
219 
220 /*
221  * Figure out to which domain a cpu belongs and stick it there.
222  * Return the id of the domain used.
223  */
224 static int __cpuinit numa_setup_cpu(unsigned long lcpu)
225 {
226 	int nid = 0;
227 	struct device_node *cpu = find_cpu_node(lcpu);
228 
229 	if (!cpu) {
230 		WARN_ON(1);
231 		goto out;
232 	}
233 
234 	nid = of_node_to_nid_single(cpu);
235 
236 	if (nid < 0 || !node_online(nid))
237 		nid = any_online_node(NODE_MASK_ALL);
238 out:
239 	map_cpu_to_node(lcpu, nid);
240 
241 	of_node_put(cpu);
242 
243 	return nid;
244 }
245 
246 static int __cpuinit cpu_numa_callback(struct notifier_block *nfb,
247 			     unsigned long action,
248 			     void *hcpu)
249 {
250 	unsigned long lcpu = (unsigned long)hcpu;
251 	int ret = NOTIFY_DONE;
252 
253 	switch (action) {
254 	case CPU_UP_PREPARE:
255 		numa_setup_cpu(lcpu);
256 		ret = NOTIFY_OK;
257 		break;
258 #ifdef CONFIG_HOTPLUG_CPU
259 	case CPU_DEAD:
260 	case CPU_UP_CANCELED:
261 		unmap_cpu_from_node(lcpu);
262 		break;
263 		ret = NOTIFY_OK;
264 #endif
265 	}
266 	return ret;
267 }
268 
269 /*
270  * Check and possibly modify a memory region to enforce the memory limit.
271  *
272  * Returns the size the region should have to enforce the memory limit.
273  * This will either be the original value of size, a truncated value,
274  * or zero. If the returned value of size is 0 the region should be
275  * discarded as it lies wholy above the memory limit.
276  */
277 static unsigned long __init numa_enforce_memory_limit(unsigned long start,
278 						      unsigned long size)
279 {
280 	/*
281 	 * We use lmb_end_of_DRAM() in here instead of memory_limit because
282 	 * we've already adjusted it for the limit and it takes care of
283 	 * having memory holes below the limit.
284 	 */
285 
286 	if (! memory_limit)
287 		return size;
288 
289 	if (start + size <= lmb_end_of_DRAM())
290 		return size;
291 
292 	if (start >= lmb_end_of_DRAM())
293 		return 0;
294 
295 	return lmb_end_of_DRAM() - start;
296 }
297 
298 /*
299  * Extract NUMA information from the ibm,dynamic-reconfiguration-memory
300  * node.  This assumes n_mem_{addr,size}_cells have been set.
301  */
302 static void __init parse_drconf_memory(struct device_node *memory)
303 {
304 	const unsigned int *lm, *dm, *aa;
305 	unsigned int ls, ld, la;
306 	unsigned int n, aam, aalen;
307 	unsigned long lmb_size, size;
308 	int nid, default_nid = 0;
309 	unsigned int start, ai, flags;
310 
311 	lm = get_property(memory, "ibm,lmb-size", &ls);
312 	dm = get_property(memory, "ibm,dynamic-memory", &ld);
313 	aa = get_property(memory, "ibm,associativity-lookup-arrays", &la);
314 	if (!lm || !dm || !aa ||
315 	    ls < sizeof(unsigned int) || ld < sizeof(unsigned int) ||
316 	    la < 2 * sizeof(unsigned int))
317 		return;
318 
319 	lmb_size = read_n_cells(n_mem_size_cells, &lm);
320 	n = *dm++;		/* number of LMBs */
321 	aam = *aa++;		/* number of associativity lists */
322 	aalen = *aa++;		/* length of each associativity list */
323 	if (ld < (n * (n_mem_addr_cells + 4) + 1) * sizeof(unsigned int) ||
324 	    la < (aam * aalen + 2) * sizeof(unsigned int))
325 		return;
326 
327 	for (; n != 0; --n) {
328 		start = read_n_cells(n_mem_addr_cells, &dm);
329 		ai = dm[2];
330 		flags = dm[3];
331 		dm += 4;
332 		/* 0x80 == reserved, 0x8 = assigned to us */
333 		if ((flags & 0x80) || !(flags & 0x8))
334 			continue;
335 		nid = default_nid;
336 		/* flags & 0x40 means associativity index is invalid */
337 		if (min_common_depth > 0 && min_common_depth <= aalen &&
338 		    (flags & 0x40) == 0 && ai < aam) {
339 			/* this is like of_node_to_nid_single */
340 			nid = aa[ai * aalen + min_common_depth - 1];
341 			if (nid == 0xffff || nid >= MAX_NUMNODES)
342 				nid = default_nid;
343 		}
344 		node_set_online(nid);
345 
346 		size = numa_enforce_memory_limit(start, lmb_size);
347 		if (!size)
348 			continue;
349 
350 		add_active_range(nid, start >> PAGE_SHIFT,
351 				 (start >> PAGE_SHIFT) + (size >> PAGE_SHIFT));
352 	}
353 }
354 
355 static int __init parse_numa_properties(void)
356 {
357 	struct device_node *cpu = NULL;
358 	struct device_node *memory = NULL;
359 	int default_nid = 0;
360 	unsigned long i;
361 
362 	if (numa_enabled == 0) {
363 		printk(KERN_WARNING "NUMA disabled by user\n");
364 		return -1;
365 	}
366 
367 	min_common_depth = find_min_common_depth();
368 
369 	if (min_common_depth < 0)
370 		return min_common_depth;
371 
372 	dbg("NUMA associativity depth for CPU/Memory: %d\n", min_common_depth);
373 
374 	/*
375 	 * Even though we connect cpus to numa domains later in SMP
376 	 * init, we need to know the node ids now. This is because
377 	 * each node to be onlined must have NODE_DATA etc backing it.
378 	 */
379 	for_each_present_cpu(i) {
380 		int nid;
381 
382 		cpu = find_cpu_node(i);
383 		BUG_ON(!cpu);
384 		nid = of_node_to_nid_single(cpu);
385 		of_node_put(cpu);
386 
387 		/*
388 		 * Don't fall back to default_nid yet -- we will plug
389 		 * cpus into nodes once the memory scan has discovered
390 		 * the topology.
391 		 */
392 		if (nid < 0)
393 			continue;
394 		node_set_online(nid);
395 	}
396 
397 	get_n_mem_cells(&n_mem_addr_cells, &n_mem_size_cells);
398 	memory = NULL;
399 	while ((memory = of_find_node_by_type(memory, "memory")) != NULL) {
400 		unsigned long start;
401 		unsigned long size;
402 		int nid;
403 		int ranges;
404 		const unsigned int *memcell_buf;
405 		unsigned int len;
406 
407 		memcell_buf = get_property(memory,
408 			"linux,usable-memory", &len);
409 		if (!memcell_buf || len <= 0)
410 			memcell_buf = get_property(memory, "reg", &len);
411 		if (!memcell_buf || len <= 0)
412 			continue;
413 
414 		/* ranges in cell */
415 		ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
416 new_range:
417 		/* these are order-sensitive, and modify the buffer pointer */
418 		start = read_n_cells(n_mem_addr_cells, &memcell_buf);
419 		size = read_n_cells(n_mem_size_cells, &memcell_buf);
420 
421 		/*
422 		 * Assumption: either all memory nodes or none will
423 		 * have associativity properties.  If none, then
424 		 * everything goes to default_nid.
425 		 */
426 		nid = of_node_to_nid_single(memory);
427 		if (nid < 0)
428 			nid = default_nid;
429 		node_set_online(nid);
430 
431 		if (!(size = numa_enforce_memory_limit(start, size))) {
432 			if (--ranges)
433 				goto new_range;
434 			else
435 				continue;
436 		}
437 
438 		add_active_range(nid, start >> PAGE_SHIFT,
439 				(start >> PAGE_SHIFT) + (size >> PAGE_SHIFT));
440 
441 		if (--ranges)
442 			goto new_range;
443 	}
444 
445 	/*
446 	 * Now do the same thing for each LMB listed in the ibm,dynamic-memory
447 	 * property in the ibm,dynamic-reconfiguration-memory node.
448 	 */
449 	memory = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
450 	if (memory)
451 		parse_drconf_memory(memory);
452 
453 	return 0;
454 }
455 
456 static void __init setup_nonnuma(void)
457 {
458 	unsigned long top_of_ram = lmb_end_of_DRAM();
459 	unsigned long total_ram = lmb_phys_mem_size();
460 	unsigned long start_pfn, end_pfn;
461 	unsigned int i;
462 
463 	printk(KERN_DEBUG "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
464 	       top_of_ram, total_ram);
465 	printk(KERN_DEBUG "Memory hole size: %ldMB\n",
466 	       (top_of_ram - total_ram) >> 20);
467 
468 	for (i = 0; i < lmb.memory.cnt; ++i) {
469 		start_pfn = lmb.memory.region[i].base >> PAGE_SHIFT;
470 		end_pfn = start_pfn + lmb_size_pages(&lmb.memory, i);
471 		add_active_range(0, start_pfn, end_pfn);
472 	}
473 	node_set_online(0);
474 }
475 
476 void __init dump_numa_cpu_topology(void)
477 {
478 	unsigned int node;
479 	unsigned int cpu, count;
480 
481 	if (min_common_depth == -1 || !numa_enabled)
482 		return;
483 
484 	for_each_online_node(node) {
485 		printk(KERN_DEBUG "Node %d CPUs:", node);
486 
487 		count = 0;
488 		/*
489 		 * If we used a CPU iterator here we would miss printing
490 		 * the holes in the cpumap.
491 		 */
492 		for (cpu = 0; cpu < NR_CPUS; cpu++) {
493 			if (cpu_isset(cpu, numa_cpumask_lookup_table[node])) {
494 				if (count == 0)
495 					printk(" %u", cpu);
496 				++count;
497 			} else {
498 				if (count > 1)
499 					printk("-%u", cpu - 1);
500 				count = 0;
501 			}
502 		}
503 
504 		if (count > 1)
505 			printk("-%u", NR_CPUS - 1);
506 		printk("\n");
507 	}
508 }
509 
510 static void __init dump_numa_memory_topology(void)
511 {
512 	unsigned int node;
513 	unsigned int count;
514 
515 	if (min_common_depth == -1 || !numa_enabled)
516 		return;
517 
518 	for_each_online_node(node) {
519 		unsigned long i;
520 
521 		printk(KERN_DEBUG "Node %d Memory:", node);
522 
523 		count = 0;
524 
525 		for (i = 0; i < lmb_end_of_DRAM();
526 		     i += (1 << SECTION_SIZE_BITS)) {
527 			if (early_pfn_to_nid(i >> PAGE_SHIFT) == node) {
528 				if (count == 0)
529 					printk(" 0x%lx", i);
530 				++count;
531 			} else {
532 				if (count > 0)
533 					printk("-0x%lx", i);
534 				count = 0;
535 			}
536 		}
537 
538 		if (count > 0)
539 			printk("-0x%lx", i);
540 		printk("\n");
541 	}
542 }
543 
544 /*
545  * Allocate some memory, satisfying the lmb or bootmem allocator where
546  * required. nid is the preferred node and end is the physical address of
547  * the highest address in the node.
548  *
549  * Returns the physical address of the memory.
550  */
551 static void __init *careful_allocation(int nid, unsigned long size,
552 				       unsigned long align,
553 				       unsigned long end_pfn)
554 {
555 	int new_nid;
556 	unsigned long ret = __lmb_alloc_base(size, align, end_pfn << PAGE_SHIFT);
557 
558 	/* retry over all memory */
559 	if (!ret)
560 		ret = __lmb_alloc_base(size, align, lmb_end_of_DRAM());
561 
562 	if (!ret)
563 		panic("numa.c: cannot allocate %lu bytes on node %d",
564 		      size, nid);
565 
566 	/*
567 	 * If the memory came from a previously allocated node, we must
568 	 * retry with the bootmem allocator.
569 	 */
570 	new_nid = early_pfn_to_nid(ret >> PAGE_SHIFT);
571 	if (new_nid < nid) {
572 		ret = (unsigned long)__alloc_bootmem_node(NODE_DATA(new_nid),
573 				size, align, 0);
574 
575 		if (!ret)
576 			panic("numa.c: cannot allocate %lu bytes on node %d",
577 			      size, new_nid);
578 
579 		ret = __pa(ret);
580 
581 		dbg("alloc_bootmem %lx %lx\n", ret, size);
582 	}
583 
584 	return (void *)ret;
585 }
586 
587 static struct notifier_block __cpuinitdata ppc64_numa_nb = {
588 	.notifier_call = cpu_numa_callback,
589 	.priority = 1 /* Must run before sched domains notifier. */
590 };
591 
592 void __init do_init_bootmem(void)
593 {
594 	int nid;
595 	unsigned int i;
596 
597 	min_low_pfn = 0;
598 	max_low_pfn = lmb_end_of_DRAM() >> PAGE_SHIFT;
599 	max_pfn = max_low_pfn;
600 
601 	if (parse_numa_properties())
602 		setup_nonnuma();
603 	else
604 		dump_numa_memory_topology();
605 
606 	register_cpu_notifier(&ppc64_numa_nb);
607 	cpu_numa_callback(&ppc64_numa_nb, CPU_UP_PREPARE,
608 			  (void *)(unsigned long)boot_cpuid);
609 
610 	for_each_online_node(nid) {
611 		unsigned long start_pfn, end_pfn;
612 		unsigned long bootmem_paddr;
613 		unsigned long bootmap_pages;
614 
615 		get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
616 
617 		/* Allocate the node structure node local if possible */
618 		NODE_DATA(nid) = careful_allocation(nid,
619 					sizeof(struct pglist_data),
620 					SMP_CACHE_BYTES, end_pfn);
621 		NODE_DATA(nid) = __va(NODE_DATA(nid));
622 		memset(NODE_DATA(nid), 0, sizeof(struct pglist_data));
623 
624   		dbg("node %d\n", nid);
625 		dbg("NODE_DATA() = %p\n", NODE_DATA(nid));
626 
627 		NODE_DATA(nid)->bdata = &plat_node_bdata[nid];
628 		NODE_DATA(nid)->node_start_pfn = start_pfn;
629 		NODE_DATA(nid)->node_spanned_pages = end_pfn - start_pfn;
630 
631 		if (NODE_DATA(nid)->node_spanned_pages == 0)
632   			continue;
633 
634   		dbg("start_paddr = %lx\n", start_pfn << PAGE_SHIFT);
635   		dbg("end_paddr = %lx\n", end_pfn << PAGE_SHIFT);
636 
637 		bootmap_pages = bootmem_bootmap_pages(end_pfn - start_pfn);
638 		bootmem_paddr = (unsigned long)careful_allocation(nid,
639 					bootmap_pages << PAGE_SHIFT,
640 					PAGE_SIZE, end_pfn);
641 		memset(__va(bootmem_paddr), 0, bootmap_pages << PAGE_SHIFT);
642 
643 		dbg("bootmap_paddr = %lx\n", bootmem_paddr);
644 
645 		init_bootmem_node(NODE_DATA(nid), bootmem_paddr >> PAGE_SHIFT,
646 				  start_pfn, end_pfn);
647 
648 		free_bootmem_with_active_regions(nid, end_pfn);
649 
650 		/* Mark reserved regions on this node */
651 		for (i = 0; i < lmb.reserved.cnt; i++) {
652 			unsigned long physbase = lmb.reserved.region[i].base;
653 			unsigned long size = lmb.reserved.region[i].size;
654 			unsigned long start_paddr = start_pfn << PAGE_SHIFT;
655 			unsigned long end_paddr = end_pfn << PAGE_SHIFT;
656 
657 			if (early_pfn_to_nid(physbase >> PAGE_SHIFT) != nid &&
658 			    early_pfn_to_nid((physbase+size-1) >> PAGE_SHIFT) != nid)
659 				continue;
660 
661 			if (physbase < end_paddr &&
662 			    (physbase+size) > start_paddr) {
663 				/* overlaps */
664 				if (physbase < start_paddr) {
665 					size -= start_paddr - physbase;
666 					physbase = start_paddr;
667 				}
668 
669 				if (size > end_paddr - physbase)
670 					size = end_paddr - physbase;
671 
672 				dbg("reserve_bootmem %lx %lx\n", physbase,
673 				    size);
674 				reserve_bootmem_node(NODE_DATA(nid), physbase,
675 						     size);
676 			}
677 		}
678 
679 		sparse_memory_present_with_active_regions(nid);
680 	}
681 }
682 
683 void __init paging_init(void)
684 {
685 	unsigned long max_zone_pfns[MAX_NR_ZONES];
686 	memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
687 	max_zone_pfns[ZONE_DMA] = lmb_end_of_DRAM() >> PAGE_SHIFT;
688 	free_area_init_nodes(max_zone_pfns);
689 }
690 
691 static int __init early_numa(char *p)
692 {
693 	if (!p)
694 		return 0;
695 
696 	if (strstr(p, "off"))
697 		numa_enabled = 0;
698 
699 	if (strstr(p, "debug"))
700 		numa_debug = 1;
701 
702 	return 0;
703 }
704 early_param("numa", early_numa);
705 
706 #ifdef CONFIG_MEMORY_HOTPLUG
707 /*
708  * Find the node associated with a hot added memory section.  Section
709  * corresponds to a SPARSEMEM section, not an LMB.  It is assumed that
710  * sections are fully contained within a single LMB.
711  */
712 int hot_add_scn_to_nid(unsigned long scn_addr)
713 {
714 	struct device_node *memory = NULL;
715 	nodemask_t nodes;
716 	int default_nid = any_online_node(NODE_MASK_ALL);
717 	int nid;
718 
719 	if (!numa_enabled || (min_common_depth < 0))
720 		return default_nid;
721 
722 	while ((memory = of_find_node_by_type(memory, "memory")) != NULL) {
723 		unsigned long start, size;
724 		int ranges;
725 		const unsigned int *memcell_buf;
726 		unsigned int len;
727 
728 		memcell_buf = get_property(memory, "reg", &len);
729 		if (!memcell_buf || len <= 0)
730 			continue;
731 
732 		/* ranges in cell */
733 		ranges = (len >> 2) / (n_mem_addr_cells + n_mem_size_cells);
734 ha_new_range:
735 		start = read_n_cells(n_mem_addr_cells, &memcell_buf);
736 		size = read_n_cells(n_mem_size_cells, &memcell_buf);
737 		nid = of_node_to_nid_single(memory);
738 
739 		/* Domains not present at boot default to 0 */
740 		if (nid < 0 || !node_online(nid))
741 			nid = default_nid;
742 
743 		if ((scn_addr >= start) && (scn_addr < (start + size))) {
744 			of_node_put(memory);
745 			goto got_nid;
746 		}
747 
748 		if (--ranges)		/* process all ranges in cell */
749 			goto ha_new_range;
750 	}
751 	BUG();	/* section address should be found above */
752 	return 0;
753 
754 	/* Temporary code to ensure that returned node is not empty */
755 got_nid:
756 	nodes_setall(nodes);
757 	while (NODE_DATA(nid)->node_spanned_pages == 0) {
758 		node_clear(nid, nodes);
759 		nid = any_online_node(nodes);
760 	}
761 	return nid;
762 }
763 #endif /* CONFIG_MEMORY_HOTPLUG */
764