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