Linux Audio

Check our new training course

Linux kernel drivers training

Mar 31-Apr 9, 2025, special US time zones
Register
Loading...
v6.2
  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
 22int numa_off;
 23nodemask_t numa_nodes_parsed __initdata;
 24
 25struct pglist_data *node_data[MAX_NUMNODES] __read_mostly;
 26EXPORT_SYMBOL(node_data);
 27
 28static struct numa_meminfo numa_meminfo __initdata_or_meminfo;
 29static struct numa_meminfo numa_reserved_meminfo __initdata_or_meminfo;
 
 
 
 30
 31static int numa_distance_cnt;
 32static u8 *numa_distance;
 33
 34static __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}
 48early_param("numa", numa_setup);
 49
 50/*
 51 * apicid, cpu, node mappings
 52 */
 53s16 __apicid_to_node[MAX_LOCAL_APIC] = {
 54	[0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE
 55};
 56
 57int 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
 66cpumask_var_t node_to_cpumask_map[MAX_NUMNODES];
 67EXPORT_SYMBOL(node_to_cpumask_map);
 68
 69/*
 70 * Map cpu index to node index
 71 */
 72DEFINE_EARLY_PER_CPU(int, x86_cpu_to_node_map, NUMA_NO_NODE);
 73EXPORT_EARLY_PER_CPU_SYMBOL(x86_cpu_to_node_map);
 74
 75void 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
 97void 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 */
109void __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
125static 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 */
159void __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 */
172static 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 */
190int __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 */
196static 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 */
238int __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
258		/* preserve info for non-RAM areas above 'max_pfn': */
259		if (bi->end > high) {
260			numa_add_memblk_to(bi->nid, high, bi->end,
261					   &numa_reserved_meminfo);
262			bi->end = high;
263		}
264
265		/* and there's no empty block */
266		if (bi->start >= bi->end)
267			numa_remove_memblk_from(i--, mi);
268	}
269
270	/* merge neighboring / overlapping entries */
271	for (i = 0; i < mi->nr_blks; i++) {
272		struct numa_memblk *bi = &mi->blk[i];
273
274		for (j = i + 1; j < mi->nr_blks; j++) {
275			struct numa_memblk *bj = &mi->blk[j];
276			u64 start, end;
277
278			/*
279			 * See whether there are overlapping blocks.  Whine
280			 * about but allow overlaps of the same nid.  They
281			 * will be merged below.
282			 */
283			if (bi->end > bj->start && bi->start < bj->end) {
284				if (bi->nid != bj->nid) {
285					pr_err("node %d [mem %#010Lx-%#010Lx] overlaps with node %d [mem %#010Lx-%#010Lx]\n",
286					       bi->nid, bi->start, bi->end - 1,
287					       bj->nid, bj->start, bj->end - 1);
288					return -EINVAL;
289				}
290				pr_warn("Warning: node %d [mem %#010Lx-%#010Lx] overlaps with itself [mem %#010Lx-%#010Lx]\n",
291					bi->nid, bi->start, bi->end - 1,
292					bj->start, bj->end - 1);
293			}
294
295			/*
296			 * Join together blocks on the same node, holes
297			 * between which don't overlap with memory on other
298			 * nodes.
299			 */
300			if (bi->nid != bj->nid)
301				continue;
302			start = min(bi->start, bj->start);
303			end = max(bi->end, bj->end);
304			for (k = 0; k < mi->nr_blks; k++) {
305				struct numa_memblk *bk = &mi->blk[k];
306
307				if (bi->nid == bk->nid)
308					continue;
309				if (start < bk->end && end > bk->start)
310					break;
311			}
312			if (k < mi->nr_blks)
313				continue;
314			printk(KERN_INFO "NUMA: Node %d [mem %#010Lx-%#010Lx] + [mem %#010Lx-%#010Lx] -> [mem %#010Lx-%#010Lx]\n",
315			       bi->nid, bi->start, bi->end - 1, bj->start,
316			       bj->end - 1, start, end - 1);
317			bi->start = start;
318			bi->end = end;
319			numa_remove_memblk_from(j--, mi);
320		}
321	}
322
323	/* clear unused ones */
324	for (i = mi->nr_blks; i < ARRAY_SIZE(mi->blk); i++) {
325		mi->blk[i].start = mi->blk[i].end = 0;
326		mi->blk[i].nid = NUMA_NO_NODE;
327	}
328
329	return 0;
330}
331
332/*
333 * Set nodes, which have memory in @mi, in *@nodemask.
334 */
335static void __init numa_nodemask_from_meminfo(nodemask_t *nodemask,
336					      const struct numa_meminfo *mi)
337{
338	int i;
339
340	for (i = 0; i < ARRAY_SIZE(mi->blk); i++)
341		if (mi->blk[i].start != mi->blk[i].end &&
342		    mi->blk[i].nid != NUMA_NO_NODE)
343			node_set(mi->blk[i].nid, *nodemask);
344}
345
346/**
347 * numa_reset_distance - Reset NUMA distance table
348 *
349 * The current table is freed.  The next numa_set_distance() call will
350 * create a new one.
351 */
352void __init numa_reset_distance(void)
353{
354	size_t size = numa_distance_cnt * numa_distance_cnt * sizeof(numa_distance[0]);
355
356	/* numa_distance could be 1LU marking allocation failure, test cnt */
357	if (numa_distance_cnt)
358		memblock_free(numa_distance, size);
359	numa_distance_cnt = 0;
360	numa_distance = NULL;	/* enable table creation */
361}
362
363static int __init numa_alloc_distance(void)
364{
365	nodemask_t nodes_parsed;
366	size_t size;
367	int i, j, cnt = 0;
368	u64 phys;
369
370	/* size the new table and allocate it */
371	nodes_parsed = numa_nodes_parsed;
372	numa_nodemask_from_meminfo(&nodes_parsed, &numa_meminfo);
373
374	for_each_node_mask(i, nodes_parsed)
375		cnt = i;
376	cnt++;
377	size = cnt * cnt * sizeof(numa_distance[0]);
378
379	phys = memblock_phys_alloc_range(size, PAGE_SIZE, 0,
380					 PFN_PHYS(max_pfn_mapped));
381	if (!phys) {
382		pr_warn("Warning: can't allocate distance table!\n");
383		/* don't retry until explicitly reset */
384		numa_distance = (void *)1LU;
385		return -ENOMEM;
386	}
 
387
388	numa_distance = __va(phys);
389	numa_distance_cnt = cnt;
390
391	/* fill with the default distances */
392	for (i = 0; i < cnt; i++)
393		for (j = 0; j < cnt; j++)
394			numa_distance[i * cnt + j] = i == j ?
395				LOCAL_DISTANCE : REMOTE_DISTANCE;
396	printk(KERN_DEBUG "NUMA: Initialized distance table, cnt=%d\n", cnt);
397
398	return 0;
399}
400
401/**
402 * numa_set_distance - Set NUMA distance from one NUMA to another
403 * @from: the 'from' node to set distance
404 * @to: the 'to'  node to set distance
405 * @distance: NUMA distance
406 *
407 * Set the distance from node @from to @to to @distance.  If distance table
408 * doesn't exist, one which is large enough to accommodate all the currently
409 * known nodes will be created.
410 *
411 * If such table cannot be allocated, a warning is printed and further
412 * calls are ignored until the distance table is reset with
413 * numa_reset_distance().
414 *
415 * If @from or @to is higher than the highest known node or lower than zero
416 * at the time of table creation or @distance doesn't make sense, the call
417 * is ignored.
418 * This is to allow simplification of specific NUMA config implementations.
419 */
420void __init numa_set_distance(int from, int to, int distance)
421{
422	if (!numa_distance && numa_alloc_distance() < 0)
423		return;
424
425	if (from >= numa_distance_cnt || to >= numa_distance_cnt ||
426			from < 0 || to < 0) {
427		pr_warn_once("Warning: node ids are out of bound, from=%d to=%d distance=%d\n",
428			     from, to, distance);
429		return;
430	}
431
432	if ((u8)distance != distance ||
433	    (from == to && distance != LOCAL_DISTANCE)) {
434		pr_warn_once("Warning: invalid distance parameter, from=%d to=%d distance=%d\n",
435			     from, to, distance);
436		return;
437	}
438
439	numa_distance[from * numa_distance_cnt + to] = distance;
440}
441
442int __node_distance(int from, int to)
443{
444	if (from >= numa_distance_cnt || to >= numa_distance_cnt)
445		return from == to ? LOCAL_DISTANCE : REMOTE_DISTANCE;
446	return numa_distance[from * numa_distance_cnt + to];
447}
448EXPORT_SYMBOL(__node_distance);
449
450/*
451 * Sanity check to catch more bad NUMA configurations (they are amazingly
452 * common).  Make sure the nodes cover all memory.
453 */
454static bool __init numa_meminfo_cover_memory(const struct numa_meminfo *mi)
455{
456	u64 numaram, e820ram;
457	int i;
458
459	numaram = 0;
460	for (i = 0; i < mi->nr_blks; i++) {
461		u64 s = mi->blk[i].start >> PAGE_SHIFT;
462		u64 e = mi->blk[i].end >> PAGE_SHIFT;
463		numaram += e - s;
464		numaram -= __absent_pages_in_range(mi->blk[i].nid, s, e);
465		if ((s64)numaram < 0)
466			numaram = 0;
467	}
468
469	e820ram = max_pfn - absent_pages_in_range(0, max_pfn);
470
471	/* We seem to lose 3 pages somewhere. Allow 1M of slack. */
472	if ((s64)(e820ram - numaram) >= (1 << (20 - PAGE_SHIFT))) {
473		printk(KERN_ERR "NUMA: nodes only cover %LuMB of your %LuMB e820 RAM. Not used.\n",
474		       (numaram << PAGE_SHIFT) >> 20,
475		       (e820ram << PAGE_SHIFT) >> 20);
476		return false;
477	}
478	return true;
479}
480
481/*
482 * Mark all currently memblock-reserved physical memory (which covers the
483 * kernel's own memory ranges) as hot-unswappable.
484 */
485static void __init numa_clear_kernel_node_hotplug(void)
486{
487	nodemask_t reserved_nodemask = NODE_MASK_NONE;
488	struct memblock_region *mb_region;
489	int i;
490
491	/*
492	 * We have to do some preprocessing of memblock regions, to
493	 * make them suitable for reservation.
494	 *
495	 * At this time, all memory regions reserved by memblock are
496	 * used by the kernel, but those regions are not split up
497	 * along node boundaries yet, and don't necessarily have their
498	 * node ID set yet either.
499	 *
500	 * So iterate over all memory known to the x86 architecture,
501	 * and use those ranges to set the nid in memblock.reserved.
502	 * This will split up the memblock regions along node
503	 * boundaries and will set the node IDs as well.
504	 */
505	for (i = 0; i < numa_meminfo.nr_blks; i++) {
506		struct numa_memblk *mb = numa_meminfo.blk + i;
507		int ret;
508
509		ret = memblock_set_node(mb->start, mb->end - mb->start, &memblock.reserved, mb->nid);
510		WARN_ON_ONCE(ret);
511	}
512
513	/*
514	 * Now go over all reserved memblock regions, to construct a
515	 * node mask of all kernel reserved memory areas.
516	 *
517	 * [ Note, when booting with mem=nn[kMG] or in a kdump kernel,
518	 *   numa_meminfo might not include all memblock.reserved
519	 *   memory ranges, because quirks such as trim_snb_memory()
520	 *   reserve specific pages for Sandy Bridge graphics. ]
521	 */
522	for_each_reserved_mem_region(mb_region) {
523		int nid = memblock_get_region_node(mb_region);
524
525		if (nid != MAX_NUMNODES)
526			node_set(nid, reserved_nodemask);
527	}
528
529	/*
530	 * Finally, clear the MEMBLOCK_HOTPLUG flag for all memory
531	 * belonging to the reserved node mask.
532	 *
533	 * Note that this will include memory regions that reside
534	 * on nodes that contain kernel memory - entire nodes
535	 * become hot-unpluggable:
536	 */
537	for (i = 0; i < numa_meminfo.nr_blks; i++) {
538		struct numa_memblk *mb = numa_meminfo.blk + i;
539
540		if (!node_isset(mb->nid, reserved_nodemask))
541			continue;
542
543		memblock_clear_hotplug(mb->start, mb->end - mb->start);
544	}
545}
546
547static int __init numa_register_memblks(struct numa_meminfo *mi)
548{
 
549	int i, nid;
550
551	/* Account for nodes with cpus and no memory */
552	node_possible_map = numa_nodes_parsed;
553	numa_nodemask_from_meminfo(&node_possible_map, mi);
554	if (WARN_ON(nodes_empty(node_possible_map)))
555		return -EINVAL;
556
557	for (i = 0; i < mi->nr_blks; i++) {
558		struct numa_memblk *mb = &mi->blk[i];
559		memblock_set_node(mb->start, mb->end - mb->start,
560				  &memblock.memory, mb->nid);
561	}
562
563	/*
564	 * At very early time, the kernel have to use some memory such as
565	 * loading the kernel image. We cannot prevent this anyway. So any
566	 * node the kernel resides in should be un-hotpluggable.
567	 *
568	 * And when we come here, alloc node data won't fail.
569	 */
570	numa_clear_kernel_node_hotplug();
571
572	/*
573	 * If sections array is gonna be used for pfn -> nid mapping, check
574	 * whether its granularity is fine enough.
575	 */
576	if (IS_ENABLED(NODE_NOT_IN_PAGE_FLAGS)) {
577		unsigned long pfn_align = node_map_pfn_alignment();
578
579		if (pfn_align && pfn_align < PAGES_PER_SECTION) {
580			pr_warn("Node alignment %LuMB < min %LuMB, rejecting NUMA config\n",
581				PFN_PHYS(pfn_align) >> 20,
582				PFN_PHYS(PAGES_PER_SECTION) >> 20);
583			return -EINVAL;
584		}
585	}
 
586	if (!numa_meminfo_cover_memory(mi))
587		return -EINVAL;
588
589	/* Finally register nodes. */
590	for_each_node_mask(nid, node_possible_map) {
591		u64 start = PFN_PHYS(max_pfn);
592		u64 end = 0;
593
594		for (i = 0; i < mi->nr_blks; i++) {
595			if (nid != mi->blk[i].nid)
596				continue;
597			start = min(mi->blk[i].start, start);
598			end = max(mi->blk[i].end, end);
599		}
600
601		if (start >= end)
602			continue;
603
604		/*
605		 * Don't confuse VM with a node that doesn't have the
606		 * minimum amount of memory:
607		 */
608		if (end && (end - start) < NODE_MIN_SIZE)
609			continue;
610
611		alloc_node_data(nid);
612	}
613
614	/* Dump memblock with node info and return. */
615	memblock_dump_all();
616	return 0;
617}
618
619/*
620 * There are unfortunately some poorly designed mainboards around that
621 * only connect memory to a single CPU. This breaks the 1:1 cpu->node
622 * mapping. To avoid this fill in the mapping for all possible CPUs,
623 * as the number of CPUs is not known yet. We round robin the existing
624 * nodes.
625 */
626static void __init numa_init_array(void)
627{
628	int rr, i;
629
630	rr = first_node(node_online_map);
631	for (i = 0; i < nr_cpu_ids; i++) {
632		if (early_cpu_to_node(i) != NUMA_NO_NODE)
633			continue;
634		numa_set_node(i, rr);
635		rr = next_node_in(rr, node_online_map);
 
 
636	}
637}
638
639static int __init numa_init(int (*init_func)(void))
640{
641	int i;
642	int ret;
643
644	for (i = 0; i < MAX_LOCAL_APIC; i++)
645		set_apicid_to_node(i, NUMA_NO_NODE);
646
647	nodes_clear(numa_nodes_parsed);
648	nodes_clear(node_possible_map);
649	nodes_clear(node_online_map);
650	memset(&numa_meminfo, 0, sizeof(numa_meminfo));
651	WARN_ON(memblock_set_node(0, ULLONG_MAX, &memblock.memory,
652				  MAX_NUMNODES));
653	WARN_ON(memblock_set_node(0, ULLONG_MAX, &memblock.reserved,
654				  MAX_NUMNODES));
655	/* In case that parsing SRAT failed. */
656	WARN_ON(memblock_clear_hotplug(0, ULLONG_MAX));
657	numa_reset_distance();
658
659	ret = init_func();
660	if (ret < 0)
661		return ret;
662
663	/*
664	 * We reset memblock back to the top-down direction
665	 * here because if we configured ACPI_NUMA, we have
666	 * parsed SRAT in init_func(). It is ok to have the
667	 * reset here even if we did't configure ACPI_NUMA
668	 * or acpi numa init fails and fallbacks to dummy
669	 * numa init.
670	 */
671	memblock_set_bottom_up(false);
672
673	ret = numa_cleanup_meminfo(&numa_meminfo);
674	if (ret < 0)
675		return ret;
676
677	numa_emulation(&numa_meminfo, numa_distance_cnt);
678
679	ret = numa_register_memblks(&numa_meminfo);
680	if (ret < 0)
681		return ret;
682
683	for (i = 0; i < nr_cpu_ids; i++) {
684		int nid = early_cpu_to_node(i);
685
686		if (nid == NUMA_NO_NODE)
687			continue;
688		if (!node_online(nid))
689			numa_clear_node(i);
690	}
691	numa_init_array();
692
693	return 0;
694}
695
696/**
697 * dummy_numa_init - Fallback dummy NUMA init
698 *
699 * Used if there's no underlying NUMA architecture, NUMA initialization
700 * fails, or NUMA is disabled on the command line.
701 *
702 * Must online at least one node and add memory blocks that cover all
703 * allowed memory.  This function must not fail.
704 */
705static int __init dummy_numa_init(void)
706{
707	printk(KERN_INFO "%s\n",
708	       numa_off ? "NUMA turned off" : "No NUMA configuration found");
709	printk(KERN_INFO "Faking a node at [mem %#018Lx-%#018Lx]\n",
710	       0LLU, PFN_PHYS(max_pfn) - 1);
711
712	node_set(0, numa_nodes_parsed);
713	numa_add_memblk(0, 0, PFN_PHYS(max_pfn));
714
715	return 0;
716}
717
718/**
719 * x86_numa_init - Initialize NUMA
720 *
721 * Try each configured NUMA initialization method until one succeeds.  The
722 * last fallback is dummy single node config encompassing whole memory and
723 * never fails.
724 */
725void __init x86_numa_init(void)
726{
727	if (!numa_off) {
728#ifdef CONFIG_ACPI_NUMA
729		if (!numa_init(x86_acpi_numa_init))
730			return;
731#endif
732#ifdef CONFIG_AMD_NUMA
733		if (!numa_init(amd_numa_init))
734			return;
735#endif
736	}
737
738	numa_init(dummy_numa_init);
739}
740
741
742/*
743 * A node may exist which has one or more Generic Initiators but no CPUs and no
744 * memory.
745 *
746 * This function must be called after init_cpu_to_node(), to ensure that any
747 * memoryless CPU nodes have already been brought online, and before the
748 * node_data[nid] is needed for zone list setup in build_all_zonelists().
749 *
750 * When this function is called, any nodes containing either memory and/or CPUs
751 * will already be online and there is no need to do anything extra, even if
752 * they also contain one or more Generic Initiators.
753 */
754void __init init_gi_nodes(void)
755{
756	int nid;
 
 
 
 
 
 
 
 
 
 
 
757
758	/*
759	 * Exclude this node from
760	 * bringup_nonboot_cpus
761	 *  cpu_up
762	 *   __try_online_node
763	 *    register_one_node
764	 * because node_subsys is not initialized yet.
765	 * TODO remove dependency on node_online
766	 */
767	for_each_node_state(nid, N_GENERIC_INITIATOR)
768		if (!node_online(nid))
769			node_set_online(nid);
770}
771
772/*
773 * Setup early cpu_to_node.
774 *
775 * Populate cpu_to_node[] only if x86_cpu_to_apicid[],
776 * and apicid_to_node[] tables have valid entries for a CPU.
777 * This means we skip cpu_to_node[] initialisation for NUMA
778 * emulation and faking node case (when running a kernel compiled
779 * for NUMA on a non NUMA box), which is OK as cpu_to_node[]
780 * is already initialized in a round robin manner at numa_init_array,
781 * prior to this call, and this initialization is good enough
782 * for the fake NUMA cases.
783 *
784 * Called before the per_cpu areas are setup.
785 */
786void __init init_cpu_to_node(void)
787{
788	int cpu;
789	u16 *cpu_to_apicid = early_per_cpu_ptr(x86_cpu_to_apicid);
790
791	BUG_ON(cpu_to_apicid == NULL);
792
793	for_each_possible_cpu(cpu) {
794		int node = numa_cpu_node(cpu);
795
796		if (node == NUMA_NO_NODE)
797			continue;
798
799		/*
800		 * Exclude this node from
801		 * bringup_nonboot_cpus
802		 *  cpu_up
803		 *   __try_online_node
804		 *    register_one_node
805		 * because node_subsys is not initialized yet.
806		 * TODO remove dependency on node_online
807		 */
808		if (!node_online(node))
809			node_set_online(node);
810
811		numa_set_node(cpu, node);
812	}
813}
814
815#ifndef CONFIG_DEBUG_PER_CPU_MAPS
816
817# ifndef CONFIG_NUMA_EMU
818void numa_add_cpu(int cpu)
819{
820	cpumask_set_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
821}
822
823void numa_remove_cpu(int cpu)
824{
825	cpumask_clear_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
826}
827# endif	/* !CONFIG_NUMA_EMU */
828
829#else	/* !CONFIG_DEBUG_PER_CPU_MAPS */
830
831int __cpu_to_node(int cpu)
832{
833	if (early_per_cpu_ptr(x86_cpu_to_node_map)) {
834		printk(KERN_WARNING
835			"cpu_to_node(%d): usage too early!\n", cpu);
836		dump_stack();
837		return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
838	}
839	return per_cpu(x86_cpu_to_node_map, cpu);
840}
841EXPORT_SYMBOL(__cpu_to_node);
842
843/*
844 * Same function as cpu_to_node() but used if called before the
845 * per_cpu areas are setup.
846 */
847int early_cpu_to_node(int cpu)
848{
849	if (early_per_cpu_ptr(x86_cpu_to_node_map))
850		return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
851
852	if (!cpu_possible(cpu)) {
853		printk(KERN_WARNING
854			"early_cpu_to_node(%d): no per_cpu area!\n", cpu);
855		dump_stack();
856		return NUMA_NO_NODE;
857	}
858	return per_cpu(x86_cpu_to_node_map, cpu);
859}
860
861void debug_cpumask_set_cpu(int cpu, int node, bool enable)
862{
863	struct cpumask *mask;
864
865	if (node == NUMA_NO_NODE) {
866		/* early_cpu_to_node() already emits a warning and trace */
867		return;
868	}
869	mask = node_to_cpumask_map[node];
870	if (!cpumask_available(mask)) {
871		pr_err("node_to_cpumask_map[%i] NULL\n", node);
872		dump_stack();
873		return;
874	}
875
876	if (enable)
877		cpumask_set_cpu(cpu, mask);
878	else
879		cpumask_clear_cpu(cpu, mask);
880
881	printk(KERN_DEBUG "%s cpu %d node %d: mask now %*pbl\n",
882		enable ? "numa_add_cpu" : "numa_remove_cpu",
883		cpu, node, cpumask_pr_args(mask));
884	return;
885}
886
887# ifndef CONFIG_NUMA_EMU
888static void numa_set_cpumask(int cpu, bool enable)
889{
890	debug_cpumask_set_cpu(cpu, early_cpu_to_node(cpu), enable);
891}
892
893void numa_add_cpu(int cpu)
894{
895	numa_set_cpumask(cpu, true);
896}
897
898void numa_remove_cpu(int cpu)
899{
900	numa_set_cpumask(cpu, false);
901}
902# endif	/* !CONFIG_NUMA_EMU */
903
904/*
905 * Returns a pointer to the bitmask of CPUs on Node 'node'.
906 */
907const struct cpumask *cpumask_of_node(int node)
908{
909	if ((unsigned)node >= nr_node_ids) {
910		printk(KERN_WARNING
911			"cpumask_of_node(%d): (unsigned)node >= nr_node_ids(%u)\n",
912			node, nr_node_ids);
913		dump_stack();
914		return cpu_none_mask;
915	}
916	if (!cpumask_available(node_to_cpumask_map[node])) {
917		printk(KERN_WARNING
918			"cpumask_of_node(%d): no node_to_cpumask_map!\n",
919			node);
920		dump_stack();
921		return cpu_online_mask;
922	}
923	return node_to_cpumask_map[node];
924}
925EXPORT_SYMBOL(cpumask_of_node);
926
927#endif	/* !CONFIG_DEBUG_PER_CPU_MAPS */
928
929#ifdef CONFIG_NUMA_KEEP_MEMINFO
930static int meminfo_to_nid(struct numa_meminfo *mi, u64 start)
931{
 
 
932	int i;
933
934	for (i = 0; i < mi->nr_blks; i++)
935		if (mi->blk[i].start <= start && mi->blk[i].end > start)
936			return mi->blk[i].nid;
937	return NUMA_NO_NODE;
938}
939
940int phys_to_target_node(phys_addr_t start)
941{
942	int nid = meminfo_to_nid(&numa_meminfo, start);
943
944	/*
945	 * Prefer online nodes, but if reserved memory might be
946	 * hot-added continue the search with reserved ranges.
947	 */
948	if (nid != NUMA_NO_NODE)
949		return nid;
950
951	return meminfo_to_nid(&numa_reserved_meminfo, start);
952}
953EXPORT_SYMBOL_GPL(phys_to_target_node);
954
955int memory_add_physaddr_to_nid(u64 start)
956{
957	int nid = meminfo_to_nid(&numa_meminfo, start);
958
959	if (nid == NUMA_NO_NODE)
960		nid = numa_meminfo.blk[0].nid;
961	return nid;
962}
963EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
964#endif
v4.6
 
  1/* Common code for 32 and 64-bit NUMA */
 
  2#include <linux/kernel.h>
  3#include <linux/mm.h>
  4#include <linux/string.h>
  5#include <linux/init.h>
  6#include <linux/bootmem.h>
  7#include <linux/memblock.h>
  8#include <linux/mmzone.h>
  9#include <linux/ctype.h>
 10#include <linux/module.h>
 11#include <linux/nodemask.h>
 12#include <linux/sched.h>
 13#include <linux/topology.h>
 14
 15#include <asm/e820.h>
 16#include <asm/proto.h>
 17#include <asm/dma.h>
 18#include <asm/acpi.h>
 19#include <asm/amd_nb.h>
 20
 21#include "numa_internal.h"
 22
 23int __initdata numa_off;
 24nodemask_t numa_nodes_parsed __initdata;
 25
 26struct pglist_data *node_data[MAX_NUMNODES] __read_mostly;
 27EXPORT_SYMBOL(node_data);
 28
 29static struct numa_meminfo numa_meminfo
 30#ifndef CONFIG_MEMORY_HOTPLUG
 31__initdata
 32#endif
 33;
 34
 35static int numa_distance_cnt;
 36static u8 *numa_distance;
 37
 38static __init int numa_setup(char *opt)
 39{
 40	if (!opt)
 41		return -EINVAL;
 42	if (!strncmp(opt, "off", 3))
 43		numa_off = 1;
 44#ifdef CONFIG_NUMA_EMU
 45	if (!strncmp(opt, "fake=", 5))
 46		numa_emu_cmdline(opt + 5);
 47#endif
 48#ifdef CONFIG_ACPI_NUMA
 49	if (!strncmp(opt, "noacpi", 6))
 50		acpi_numa = -1;
 51#endif
 
 52	return 0;
 53}
 54early_param("numa", numa_setup);
 55
 56/*
 57 * apicid, cpu, node mappings
 58 */
 59s16 __apicid_to_node[MAX_LOCAL_APIC] = {
 60	[0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE
 61};
 62
 63int numa_cpu_node(int cpu)
 64{
 65	int apicid = early_per_cpu(x86_cpu_to_apicid, cpu);
 66
 67	if (apicid != BAD_APICID)
 68		return __apicid_to_node[apicid];
 69	return NUMA_NO_NODE;
 70}
 71
 72cpumask_var_t node_to_cpumask_map[MAX_NUMNODES];
 73EXPORT_SYMBOL(node_to_cpumask_map);
 74
 75/*
 76 * Map cpu index to node index
 77 */
 78DEFINE_EARLY_PER_CPU(int, x86_cpu_to_node_map, NUMA_NO_NODE);
 79EXPORT_EARLY_PER_CPU_SYMBOL(x86_cpu_to_node_map);
 80
 81void numa_set_node(int cpu, int node)
 82{
 83	int *cpu_to_node_map = early_per_cpu_ptr(x86_cpu_to_node_map);
 84
 85	/* early setting, no percpu area yet */
 86	if (cpu_to_node_map) {
 87		cpu_to_node_map[cpu] = node;
 88		return;
 89	}
 90
 91#ifdef CONFIG_DEBUG_PER_CPU_MAPS
 92	if (cpu >= nr_cpu_ids || !cpu_possible(cpu)) {
 93		printk(KERN_ERR "numa_set_node: invalid cpu# (%d)\n", cpu);
 94		dump_stack();
 95		return;
 96	}
 97#endif
 98	per_cpu(x86_cpu_to_node_map, cpu) = node;
 99
100	set_cpu_numa_node(cpu, node);
101}
102
103void numa_clear_node(int cpu)
104{
105	numa_set_node(cpu, NUMA_NO_NODE);
106}
107
108/*
109 * Allocate node_to_cpumask_map based on number of available nodes
110 * Requires node_possible_map to be valid.
111 *
112 * Note: cpumask_of_node() is not valid until after this is done.
113 * (Use CONFIG_DEBUG_PER_CPU_MAPS to check this.)
114 */
115void __init setup_node_to_cpumask_map(void)
116{
117	unsigned int node;
118
119	/* setup nr_node_ids if not done yet */
120	if (nr_node_ids == MAX_NUMNODES)
121		setup_nr_node_ids();
122
123	/* allocate the map */
124	for (node = 0; node < nr_node_ids; node++)
125		alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]);
126
127	/* cpumask_of_node() will now work */
128	pr_debug("Node to cpumask map for %d nodes\n", nr_node_ids);
129}
130
131static int __init numa_add_memblk_to(int nid, u64 start, u64 end,
132				     struct numa_meminfo *mi)
133{
134	/* ignore zero length blks */
135	if (start == end)
136		return 0;
137
138	/* whine about and ignore invalid blks */
139	if (start > end || nid < 0 || nid >= MAX_NUMNODES) {
140		pr_warning("NUMA: Warning: invalid memblk node %d [mem %#010Lx-%#010Lx]\n",
141			   nid, start, end - 1);
142		return 0;
143	}
144
145	if (mi->nr_blks >= NR_NODE_MEMBLKS) {
146		pr_err("NUMA: too many memblk ranges\n");
147		return -EINVAL;
148	}
149
150	mi->blk[mi->nr_blks].start = start;
151	mi->blk[mi->nr_blks].end = end;
152	mi->blk[mi->nr_blks].nid = nid;
153	mi->nr_blks++;
154	return 0;
155}
156
157/**
158 * numa_remove_memblk_from - Remove one numa_memblk from a numa_meminfo
159 * @idx: Index of memblk to remove
160 * @mi: numa_meminfo to remove memblk from
161 *
162 * Remove @idx'th numa_memblk from @mi by shifting @mi->blk[] and
163 * decrementing @mi->nr_blks.
164 */
165void __init numa_remove_memblk_from(int idx, struct numa_meminfo *mi)
166{
167	mi->nr_blks--;
168	memmove(&mi->blk[idx], &mi->blk[idx + 1],
169		(mi->nr_blks - idx) * sizeof(mi->blk[0]));
170}
171
172/**
 
 
 
 
 
 
 
 
 
 
 
 
 
173 * numa_add_memblk - Add one numa_memblk to numa_meminfo
174 * @nid: NUMA node ID of the new memblk
175 * @start: Start address of the new memblk
176 * @end: End address of the new memblk
177 *
178 * Add a new memblk to the default numa_meminfo.
179 *
180 * RETURNS:
181 * 0 on success, -errno on failure.
182 */
183int __init numa_add_memblk(int nid, u64 start, u64 end)
184{
185	return numa_add_memblk_to(nid, start, end, &numa_meminfo);
186}
187
188/* Allocate NODE_DATA for a node on the local memory */
189static void __init alloc_node_data(int nid)
190{
191	const size_t nd_size = roundup(sizeof(pg_data_t), PAGE_SIZE);
192	u64 nd_pa;
193	void *nd;
194	int tnid;
195
196	/*
197	 * Allocate node data.  Try node-local memory and then any node.
198	 * Never allocate in DMA zone.
199	 */
200	nd_pa = memblock_alloc_nid(nd_size, SMP_CACHE_BYTES, nid);
201	if (!nd_pa) {
202		nd_pa = __memblock_alloc_base(nd_size, SMP_CACHE_BYTES,
203					      MEMBLOCK_ALLOC_ACCESSIBLE);
204		if (!nd_pa) {
205			pr_err("Cannot find %zu bytes in node %d\n",
206			       nd_size, nid);
207			return;
208		}
209	}
210	nd = __va(nd_pa);
211
212	/* report and initialize */
213	printk(KERN_INFO "NODE_DATA(%d) allocated [mem %#010Lx-%#010Lx]\n", nid,
214	       nd_pa, nd_pa + nd_size - 1);
215	tnid = early_pfn_to_nid(nd_pa >> PAGE_SHIFT);
216	if (tnid != nid)
217		printk(KERN_INFO "    NODE_DATA(%d) on node %d\n", nid, tnid);
218
219	node_data[nid] = nd;
220	memset(NODE_DATA(nid), 0, sizeof(pg_data_t));
221
222	node_set_online(nid);
223}
224
225/**
226 * numa_cleanup_meminfo - Cleanup a numa_meminfo
227 * @mi: numa_meminfo to clean up
228 *
229 * Sanitize @mi by merging and removing unncessary memblks.  Also check for
230 * conflicts and clear unused memblks.
231 *
232 * RETURNS:
233 * 0 on success, -errno on failure.
234 */
235int __init numa_cleanup_meminfo(struct numa_meminfo *mi)
236{
237	const u64 low = 0;
238	const u64 high = PFN_PHYS(max_pfn);
239	int i, j, k;
240
241	/* first, trim all entries */
242	for (i = 0; i < mi->nr_blks; i++) {
243		struct numa_memblk *bi = &mi->blk[i];
244
245		/* make sure all blocks are inside the limits */
 
 
 
 
 
 
 
246		bi->start = max(bi->start, low);
247		bi->end = min(bi->end, high);
248
249		/* and there's no empty or non-exist block */
250		if (bi->start >= bi->end ||
251		    !memblock_overlaps_region(&memblock.memory,
252			bi->start, bi->end - bi->start))
 
 
 
 
 
253			numa_remove_memblk_from(i--, mi);
254	}
255
256	/* merge neighboring / overlapping entries */
257	for (i = 0; i < mi->nr_blks; i++) {
258		struct numa_memblk *bi = &mi->blk[i];
259
260		for (j = i + 1; j < mi->nr_blks; j++) {
261			struct numa_memblk *bj = &mi->blk[j];
262			u64 start, end;
263
264			/*
265			 * See whether there are overlapping blocks.  Whine
266			 * about but allow overlaps of the same nid.  They
267			 * will be merged below.
268			 */
269			if (bi->end > bj->start && bi->start < bj->end) {
270				if (bi->nid != bj->nid) {
271					pr_err("NUMA: node %d [mem %#010Lx-%#010Lx] overlaps with node %d [mem %#010Lx-%#010Lx]\n",
272					       bi->nid, bi->start, bi->end - 1,
273					       bj->nid, bj->start, bj->end - 1);
274					return -EINVAL;
275				}
276				pr_warning("NUMA: Warning: node %d [mem %#010Lx-%#010Lx] overlaps with itself [mem %#010Lx-%#010Lx]\n",
277					   bi->nid, bi->start, bi->end - 1,
278					   bj->start, bj->end - 1);
279			}
280
281			/*
282			 * Join together blocks on the same node, holes
283			 * between which don't overlap with memory on other
284			 * nodes.
285			 */
286			if (bi->nid != bj->nid)
287				continue;
288			start = min(bi->start, bj->start);
289			end = max(bi->end, bj->end);
290			for (k = 0; k < mi->nr_blks; k++) {
291				struct numa_memblk *bk = &mi->blk[k];
292
293				if (bi->nid == bk->nid)
294					continue;
295				if (start < bk->end && end > bk->start)
296					break;
297			}
298			if (k < mi->nr_blks)
299				continue;
300			printk(KERN_INFO "NUMA: Node %d [mem %#010Lx-%#010Lx] + [mem %#010Lx-%#010Lx] -> [mem %#010Lx-%#010Lx]\n",
301			       bi->nid, bi->start, bi->end - 1, bj->start,
302			       bj->end - 1, start, end - 1);
303			bi->start = start;
304			bi->end = end;
305			numa_remove_memblk_from(j--, mi);
306		}
307	}
308
309	/* clear unused ones */
310	for (i = mi->nr_blks; i < ARRAY_SIZE(mi->blk); i++) {
311		mi->blk[i].start = mi->blk[i].end = 0;
312		mi->blk[i].nid = NUMA_NO_NODE;
313	}
314
315	return 0;
316}
317
318/*
319 * Set nodes, which have memory in @mi, in *@nodemask.
320 */
321static void __init numa_nodemask_from_meminfo(nodemask_t *nodemask,
322					      const struct numa_meminfo *mi)
323{
324	int i;
325
326	for (i = 0; i < ARRAY_SIZE(mi->blk); i++)
327		if (mi->blk[i].start != mi->blk[i].end &&
328		    mi->blk[i].nid != NUMA_NO_NODE)
329			node_set(mi->blk[i].nid, *nodemask);
330}
331
332/**
333 * numa_reset_distance - Reset NUMA distance table
334 *
335 * The current table is freed.  The next numa_set_distance() call will
336 * create a new one.
337 */
338void __init numa_reset_distance(void)
339{
340	size_t size = numa_distance_cnt * numa_distance_cnt * sizeof(numa_distance[0]);
341
342	/* numa_distance could be 1LU marking allocation failure, test cnt */
343	if (numa_distance_cnt)
344		memblock_free(__pa(numa_distance), size);
345	numa_distance_cnt = 0;
346	numa_distance = NULL;	/* enable table creation */
347}
348
349static int __init numa_alloc_distance(void)
350{
351	nodemask_t nodes_parsed;
352	size_t size;
353	int i, j, cnt = 0;
354	u64 phys;
355
356	/* size the new table and allocate it */
357	nodes_parsed = numa_nodes_parsed;
358	numa_nodemask_from_meminfo(&nodes_parsed, &numa_meminfo);
359
360	for_each_node_mask(i, nodes_parsed)
361		cnt = i;
362	cnt++;
363	size = cnt * cnt * sizeof(numa_distance[0]);
364
365	phys = memblock_find_in_range(0, PFN_PHYS(max_pfn_mapped),
366				      size, PAGE_SIZE);
367	if (!phys) {
368		pr_warning("NUMA: Warning: can't allocate distance table!\n");
369		/* don't retry until explicitly reset */
370		numa_distance = (void *)1LU;
371		return -ENOMEM;
372	}
373	memblock_reserve(phys, size);
374
375	numa_distance = __va(phys);
376	numa_distance_cnt = cnt;
377
378	/* fill with the default distances */
379	for (i = 0; i < cnt; i++)
380		for (j = 0; j < cnt; j++)
381			numa_distance[i * cnt + j] = i == j ?
382				LOCAL_DISTANCE : REMOTE_DISTANCE;
383	printk(KERN_DEBUG "NUMA: Initialized distance table, cnt=%d\n", cnt);
384
385	return 0;
386}
387
388/**
389 * numa_set_distance - Set NUMA distance from one NUMA to another
390 * @from: the 'from' node to set distance
391 * @to: the 'to'  node to set distance
392 * @distance: NUMA distance
393 *
394 * Set the distance from node @from to @to to @distance.  If distance table
395 * doesn't exist, one which is large enough to accommodate all the currently
396 * known nodes will be created.
397 *
398 * If such table cannot be allocated, a warning is printed and further
399 * calls are ignored until the distance table is reset with
400 * numa_reset_distance().
401 *
402 * If @from or @to is higher than the highest known node or lower than zero
403 * at the time of table creation or @distance doesn't make sense, the call
404 * is ignored.
405 * This is to allow simplification of specific NUMA config implementations.
406 */
407void __init numa_set_distance(int from, int to, int distance)
408{
409	if (!numa_distance && numa_alloc_distance() < 0)
410		return;
411
412	if (from >= numa_distance_cnt || to >= numa_distance_cnt ||
413			from < 0 || to < 0) {
414		pr_warn_once("NUMA: Warning: node ids are out of bound, from=%d to=%d distance=%d\n",
415			    from, to, distance);
416		return;
417	}
418
419	if ((u8)distance != distance ||
420	    (from == to && distance != LOCAL_DISTANCE)) {
421		pr_warn_once("NUMA: Warning: invalid distance parameter, from=%d to=%d distance=%d\n",
422			     from, to, distance);
423		return;
424	}
425
426	numa_distance[from * numa_distance_cnt + to] = distance;
427}
428
429int __node_distance(int from, int to)
430{
431	if (from >= numa_distance_cnt || to >= numa_distance_cnt)
432		return from == to ? LOCAL_DISTANCE : REMOTE_DISTANCE;
433	return numa_distance[from * numa_distance_cnt + to];
434}
435EXPORT_SYMBOL(__node_distance);
436
437/*
438 * Sanity check to catch more bad NUMA configurations (they are amazingly
439 * common).  Make sure the nodes cover all memory.
440 */
441static bool __init numa_meminfo_cover_memory(const struct numa_meminfo *mi)
442{
443	u64 numaram, e820ram;
444	int i;
445
446	numaram = 0;
447	for (i = 0; i < mi->nr_blks; i++) {
448		u64 s = mi->blk[i].start >> PAGE_SHIFT;
449		u64 e = mi->blk[i].end >> PAGE_SHIFT;
450		numaram += e - s;
451		numaram -= __absent_pages_in_range(mi->blk[i].nid, s, e);
452		if ((s64)numaram < 0)
453			numaram = 0;
454	}
455
456	e820ram = max_pfn - absent_pages_in_range(0, max_pfn);
457
458	/* We seem to lose 3 pages somewhere. Allow 1M of slack. */
459	if ((s64)(e820ram - numaram) >= (1 << (20 - PAGE_SHIFT))) {
460		printk(KERN_ERR "NUMA: nodes only cover %LuMB of your %LuMB e820 RAM. Not used.\n",
461		       (numaram << PAGE_SHIFT) >> 20,
462		       (e820ram << PAGE_SHIFT) >> 20);
463		return false;
464	}
465	return true;
466}
467
468/*
469 * Mark all currently memblock-reserved physical memory (which covers the
470 * kernel's own memory ranges) as hot-unswappable.
471 */
472static void __init numa_clear_kernel_node_hotplug(void)
473{
474	nodemask_t reserved_nodemask = NODE_MASK_NONE;
475	struct memblock_region *mb_region;
476	int i;
477
478	/*
479	 * We have to do some preprocessing of memblock regions, to
480	 * make them suitable for reservation.
481	 *
482	 * At this time, all memory regions reserved by memblock are
483	 * used by the kernel, but those regions are not split up
484	 * along node boundaries yet, and don't necessarily have their
485	 * node ID set yet either.
486	 *
487	 * So iterate over all memory known to the x86 architecture,
488	 * and use those ranges to set the nid in memblock.reserved.
489	 * This will split up the memblock regions along node
490	 * boundaries and will set the node IDs as well.
491	 */
492	for (i = 0; i < numa_meminfo.nr_blks; i++) {
493		struct numa_memblk *mb = numa_meminfo.blk + i;
494		int ret;
495
496		ret = memblock_set_node(mb->start, mb->end - mb->start, &memblock.reserved, mb->nid);
497		WARN_ON_ONCE(ret);
498	}
499
500	/*
501	 * Now go over all reserved memblock regions, to construct a
502	 * node mask of all kernel reserved memory areas.
503	 *
504	 * [ Note, when booting with mem=nn[kMG] or in a kdump kernel,
505	 *   numa_meminfo might not include all memblock.reserved
506	 *   memory ranges, because quirks such as trim_snb_memory()
507	 *   reserve specific pages for Sandy Bridge graphics. ]
508	 */
509	for_each_memblock(reserved, mb_region) {
510		if (mb_region->nid != MAX_NUMNODES)
511			node_set(mb_region->nid, reserved_nodemask);
 
 
512	}
513
514	/*
515	 * Finally, clear the MEMBLOCK_HOTPLUG flag for all memory
516	 * belonging to the reserved node mask.
517	 *
518	 * Note that this will include memory regions that reside
519	 * on nodes that contain kernel memory - entire nodes
520	 * become hot-unpluggable:
521	 */
522	for (i = 0; i < numa_meminfo.nr_blks; i++) {
523		struct numa_memblk *mb = numa_meminfo.blk + i;
524
525		if (!node_isset(mb->nid, reserved_nodemask))
526			continue;
527
528		memblock_clear_hotplug(mb->start, mb->end - mb->start);
529	}
530}
531
532static int __init numa_register_memblks(struct numa_meminfo *mi)
533{
534	unsigned long uninitialized_var(pfn_align);
535	int i, nid;
536
537	/* Account for nodes with cpus and no memory */
538	node_possible_map = numa_nodes_parsed;
539	numa_nodemask_from_meminfo(&node_possible_map, mi);
540	if (WARN_ON(nodes_empty(node_possible_map)))
541		return -EINVAL;
542
543	for (i = 0; i < mi->nr_blks; i++) {
544		struct numa_memblk *mb = &mi->blk[i];
545		memblock_set_node(mb->start, mb->end - mb->start,
546				  &memblock.memory, mb->nid);
547	}
548
549	/*
550	 * At very early time, the kernel have to use some memory such as
551	 * loading the kernel image. We cannot prevent this anyway. So any
552	 * node the kernel resides in should be un-hotpluggable.
553	 *
554	 * And when we come here, alloc node data won't fail.
555	 */
556	numa_clear_kernel_node_hotplug();
557
558	/*
559	 * If sections array is gonna be used for pfn -> nid mapping, check
560	 * whether its granularity is fine enough.
561	 */
562#ifdef NODE_NOT_IN_PAGE_FLAGS
563	pfn_align = node_map_pfn_alignment();
564	if (pfn_align && pfn_align < PAGES_PER_SECTION) {
565		printk(KERN_WARNING "Node alignment %LuMB < min %LuMB, rejecting NUMA config\n",
566		       PFN_PHYS(pfn_align) >> 20,
567		       PFN_PHYS(PAGES_PER_SECTION) >> 20);
568		return -EINVAL;
 
 
569	}
570#endif
571	if (!numa_meminfo_cover_memory(mi))
572		return -EINVAL;
573
574	/* Finally register nodes. */
575	for_each_node_mask(nid, node_possible_map) {
576		u64 start = PFN_PHYS(max_pfn);
577		u64 end = 0;
578
579		for (i = 0; i < mi->nr_blks; i++) {
580			if (nid != mi->blk[i].nid)
581				continue;
582			start = min(mi->blk[i].start, start);
583			end = max(mi->blk[i].end, end);
584		}
585
586		if (start >= end)
587			continue;
588
589		/*
590		 * Don't confuse VM with a node that doesn't have the
591		 * minimum amount of memory:
592		 */
593		if (end && (end - start) < NODE_MIN_SIZE)
594			continue;
595
596		alloc_node_data(nid);
597	}
598
599	/* Dump memblock with node info and return. */
600	memblock_dump_all();
601	return 0;
602}
603
604/*
605 * There are unfortunately some poorly designed mainboards around that
606 * only connect memory to a single CPU. This breaks the 1:1 cpu->node
607 * mapping. To avoid this fill in the mapping for all possible CPUs,
608 * as the number of CPUs is not known yet. We round robin the existing
609 * nodes.
610 */
611static void __init numa_init_array(void)
612{
613	int rr, i;
614
615	rr = first_node(node_online_map);
616	for (i = 0; i < nr_cpu_ids; i++) {
617		if (early_cpu_to_node(i) != NUMA_NO_NODE)
618			continue;
619		numa_set_node(i, rr);
620		rr = next_node(rr, node_online_map);
621		if (rr == MAX_NUMNODES)
622			rr = first_node(node_online_map);
623	}
624}
625
626static int __init numa_init(int (*init_func)(void))
627{
628	int i;
629	int ret;
630
631	for (i = 0; i < MAX_LOCAL_APIC; i++)
632		set_apicid_to_node(i, NUMA_NO_NODE);
633
634	nodes_clear(numa_nodes_parsed);
635	nodes_clear(node_possible_map);
636	nodes_clear(node_online_map);
637	memset(&numa_meminfo, 0, sizeof(numa_meminfo));
638	WARN_ON(memblock_set_node(0, ULLONG_MAX, &memblock.memory,
639				  MAX_NUMNODES));
640	WARN_ON(memblock_set_node(0, ULLONG_MAX, &memblock.reserved,
641				  MAX_NUMNODES));
642	/* In case that parsing SRAT failed. */
643	WARN_ON(memblock_clear_hotplug(0, ULLONG_MAX));
644	numa_reset_distance();
645
646	ret = init_func();
647	if (ret < 0)
648		return ret;
649
650	/*
651	 * We reset memblock back to the top-down direction
652	 * here because if we configured ACPI_NUMA, we have
653	 * parsed SRAT in init_func(). It is ok to have the
654	 * reset here even if we did't configure ACPI_NUMA
655	 * or acpi numa init fails and fallbacks to dummy
656	 * numa init.
657	 */
658	memblock_set_bottom_up(false);
659
660	ret = numa_cleanup_meminfo(&numa_meminfo);
661	if (ret < 0)
662		return ret;
663
664	numa_emulation(&numa_meminfo, numa_distance_cnt);
665
666	ret = numa_register_memblks(&numa_meminfo);
667	if (ret < 0)
668		return ret;
669
670	for (i = 0; i < nr_cpu_ids; i++) {
671		int nid = early_cpu_to_node(i);
672
673		if (nid == NUMA_NO_NODE)
674			continue;
675		if (!node_online(nid))
676			numa_clear_node(i);
677	}
678	numa_init_array();
679
680	return 0;
681}
682
683/**
684 * dummy_numa_init - Fallback dummy NUMA init
685 *
686 * Used if there's no underlying NUMA architecture, NUMA initialization
687 * fails, or NUMA is disabled on the command line.
688 *
689 * Must online at least one node and add memory blocks that cover all
690 * allowed memory.  This function must not fail.
691 */
692static int __init dummy_numa_init(void)
693{
694	printk(KERN_INFO "%s\n",
695	       numa_off ? "NUMA turned off" : "No NUMA configuration found");
696	printk(KERN_INFO "Faking a node at [mem %#018Lx-%#018Lx]\n",
697	       0LLU, PFN_PHYS(max_pfn) - 1);
698
699	node_set(0, numa_nodes_parsed);
700	numa_add_memblk(0, 0, PFN_PHYS(max_pfn));
701
702	return 0;
703}
704
705/**
706 * x86_numa_init - Initialize NUMA
707 *
708 * Try each configured NUMA initialization method until one succeeds.  The
709 * last fallback is dummy single node config encomapssing whole memory and
710 * never fails.
711 */
712void __init x86_numa_init(void)
713{
714	if (!numa_off) {
715#ifdef CONFIG_ACPI_NUMA
716		if (!numa_init(x86_acpi_numa_init))
717			return;
718#endif
719#ifdef CONFIG_AMD_NUMA
720		if (!numa_init(amd_numa_init))
721			return;
722#endif
723	}
724
725	numa_init(dummy_numa_init);
726}
727
728static __init int find_near_online_node(int node)
 
 
 
 
 
 
 
 
 
 
 
 
 
729{
730	int n, val;
731	int min_val = INT_MAX;
732	int best_node = -1;
733
734	for_each_online_node(n) {
735		val = node_distance(node, n);
736
737		if (val < min_val) {
738			min_val = val;
739			best_node = n;
740		}
741	}
742
743	return best_node;
 
 
 
 
 
 
 
 
 
 
 
744}
745
746/*
747 * Setup early cpu_to_node.
748 *
749 * Populate cpu_to_node[] only if x86_cpu_to_apicid[],
750 * and apicid_to_node[] tables have valid entries for a CPU.
751 * This means we skip cpu_to_node[] initialisation for NUMA
752 * emulation and faking node case (when running a kernel compiled
753 * for NUMA on a non NUMA box), which is OK as cpu_to_node[]
754 * is already initialized in a round robin manner at numa_init_array,
755 * prior to this call, and this initialization is good enough
756 * for the fake NUMA cases.
757 *
758 * Called before the per_cpu areas are setup.
759 */
760void __init init_cpu_to_node(void)
761{
762	int cpu;
763	u16 *cpu_to_apicid = early_per_cpu_ptr(x86_cpu_to_apicid);
764
765	BUG_ON(cpu_to_apicid == NULL);
766
767	for_each_possible_cpu(cpu) {
768		int node = numa_cpu_node(cpu);
769
770		if (node == NUMA_NO_NODE)
771			continue;
 
 
 
 
 
 
 
 
 
 
772		if (!node_online(node))
773			node = find_near_online_node(node);
 
774		numa_set_node(cpu, node);
775	}
776}
777
778#ifndef CONFIG_DEBUG_PER_CPU_MAPS
779
780# ifndef CONFIG_NUMA_EMU
781void numa_add_cpu(int cpu)
782{
783	cpumask_set_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
784}
785
786void numa_remove_cpu(int cpu)
787{
788	cpumask_clear_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
789}
790# endif	/* !CONFIG_NUMA_EMU */
791
792#else	/* !CONFIG_DEBUG_PER_CPU_MAPS */
793
794int __cpu_to_node(int cpu)
795{
796	if (early_per_cpu_ptr(x86_cpu_to_node_map)) {
797		printk(KERN_WARNING
798			"cpu_to_node(%d): usage too early!\n", cpu);
799		dump_stack();
800		return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
801	}
802	return per_cpu(x86_cpu_to_node_map, cpu);
803}
804EXPORT_SYMBOL(__cpu_to_node);
805
806/*
807 * Same function as cpu_to_node() but used if called before the
808 * per_cpu areas are setup.
809 */
810int early_cpu_to_node(int cpu)
811{
812	if (early_per_cpu_ptr(x86_cpu_to_node_map))
813		return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
814
815	if (!cpu_possible(cpu)) {
816		printk(KERN_WARNING
817			"early_cpu_to_node(%d): no per_cpu area!\n", cpu);
818		dump_stack();
819		return NUMA_NO_NODE;
820	}
821	return per_cpu(x86_cpu_to_node_map, cpu);
822}
823
824void debug_cpumask_set_cpu(int cpu, int node, bool enable)
825{
826	struct cpumask *mask;
827
828	if (node == NUMA_NO_NODE) {
829		/* early_cpu_to_node() already emits a warning and trace */
830		return;
831	}
832	mask = node_to_cpumask_map[node];
833	if (!mask) {
834		pr_err("node_to_cpumask_map[%i] NULL\n", node);
835		dump_stack();
836		return;
837	}
838
839	if (enable)
840		cpumask_set_cpu(cpu, mask);
841	else
842		cpumask_clear_cpu(cpu, mask);
843
844	printk(KERN_DEBUG "%s cpu %d node %d: mask now %*pbl\n",
845		enable ? "numa_add_cpu" : "numa_remove_cpu",
846		cpu, node, cpumask_pr_args(mask));
847	return;
848}
849
850# ifndef CONFIG_NUMA_EMU
851static void numa_set_cpumask(int cpu, bool enable)
852{
853	debug_cpumask_set_cpu(cpu, early_cpu_to_node(cpu), enable);
854}
855
856void numa_add_cpu(int cpu)
857{
858	numa_set_cpumask(cpu, true);
859}
860
861void numa_remove_cpu(int cpu)
862{
863	numa_set_cpumask(cpu, false);
864}
865# endif	/* !CONFIG_NUMA_EMU */
866
867/*
868 * Returns a pointer to the bitmask of CPUs on Node 'node'.
869 */
870const struct cpumask *cpumask_of_node(int node)
871{
872	if (node >= nr_node_ids) {
873		printk(KERN_WARNING
874			"cpumask_of_node(%d): node > nr_node_ids(%d)\n",
875			node, nr_node_ids);
876		dump_stack();
877		return cpu_none_mask;
878	}
879	if (node_to_cpumask_map[node] == NULL) {
880		printk(KERN_WARNING
881			"cpumask_of_node(%d): no node_to_cpumask_map!\n",
882			node);
883		dump_stack();
884		return cpu_online_mask;
885	}
886	return node_to_cpumask_map[node];
887}
888EXPORT_SYMBOL(cpumask_of_node);
889
890#endif	/* !CONFIG_DEBUG_PER_CPU_MAPS */
891
892#ifdef CONFIG_MEMORY_HOTPLUG
893int memory_add_physaddr_to_nid(u64 start)
894{
895	struct numa_meminfo *mi = &numa_meminfo;
896	int nid = mi->blk[0].nid;
897	int i;
898
899	for (i = 0; i < mi->nr_blks; i++)
900		if (mi->blk[i].start <= start && mi->blk[i].end > start)
901			nid = mi->blk[i].nid;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
902	return nid;
903}
904EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
905#endif